CN115310782A - Power consumer demand response potential evaluation method and device based on neural turing machine - Google Patents

Abstract

The invention provides a method and a device for evaluating demand response potential of a power consumer based on a Neural Training Machine (NTM), wherein the method comprises the following steps: grouping users by adopting a mean shift method to obtain user group data sets with different response characteristics; based on the user group data set, executing an attention mechanism AM in a gating cycle unit GRU, and extracting the demand response behavior characteristics of the user in different meteorological environments; and evaluating the demand response load adjustment quantity of the user at the given real-time electricity price based on the extracted demand response behavior characteristics through a neural turing machine model. The real-time demand response evaluation of the user side is realized, the meteorological environment influence and the response behavior characteristics of the user are considered, the accuracy and the speed of potential evaluation are improved, a reliable evaluation result is provided for the subsequent electricity price optimization decision, and the real-time accurate demand response management of the user side is facilitated to be realized.

Description

Power consumer demand response potential assessment method and device based on neural turing machine

Technical Field

The invention relates to the field of power consumer real-time demand response potential evaluation, in particular to a method and a device for evaluating power consumer demand response potential based on a neural turing machine.

Background

Accurate potential assessment is the key to the demand response aggregation management of the virtual power plant on the users. At present, the research can be mainly divided into methods such as fuzzy evaluation, scene matching, load modeling, uncertainty analysis, transfer learning, neural algorithm exploration and the like. The flexibility of both fuzzy evaluation and scene matching in practical applications is limited. In the aspect of user load modeling, the early consumer psychological model and the early price elasticity model do not aim at the energy utilization characteristics of different users, and the fuzzy overall modeling influences the accuracy of an evaluation result. Therefore, some scholars study the classification modeling of detailed load devices. And carrying out classification modeling on the resident load equipment to help the new energy to be consumed. There are documents that classify loads in large commercial buildings such as air conditioners, electric vehicles and the like and establish time distribution models, and a bottom-up demand response potential evaluation method is proposed. The model may represent the energy usage characteristics of a particular load over a period of time. However, the demand response characteristics of different types of users are difficult to embody, and the demand response characteristics cannot be directly applied to response evaluation of the virtual power plant.

The neural algorithm can better improve the accuracy of the potential evaluation aiming at the characteristics of the user. However, how to extract and evaluate response features of users of different categories on the premise of ensuring evaluation efficiency is still an important challenge in making power price decisions in an actual virtual power plant. There is literature that models uncertainty in demand response environments with a deep learning model based on Recurrent Neural Networks (RNNs). A Long-Short Term Memory algorithm (LSTM) is helpful for improving the problems of gradient extinction, explosion and the like of the RNN. However, the LSTM algorithm has numerous parameters and complex calculation, and cannot meet the requirement of a virtual power plant for high-efficiency real-time evaluation of a large number of users. Moreover, there is still room for improvement in terms of accuracy. The advent of the neuroleptic machine helped to ameliorate the above problems. On one hand, it selects a gated round robin Unit (GRU) as the core controller. When the actual data with simple input and output dimensions are processed, the calculation amount can be obviously reduced, and the calculation efficiency is improved. On the other hand, the external memory matrix can more comprehensively retain a great deal of information such as the user and the meteorological environment. The Attention Mechanism (AM) can accurately locate effective information from massive information, so that long-distance dependency relationship can be conveniently captured, and accuracy of demand response mining is improved. At present, neural turing machines are mostly applied to the field of language modeling, are excellent in the aspects of solving the problems of voice recognition, handwritten character generation, natural language processing and the like, and have a larger research space in the aspect of evaluating the demand response potential of a virtual power plant.

Disclosure of Invention

In order to solve the existing defects, the invention provides the neural-turing-based power user demand response potential evaluation method and device, which realize the real-time demand response evaluation of the user side, consider the meteorological environment influence and the response behavior characteristics of the user, improve the accuracy and speed of the potential evaluation, provide a reliable evaluation result for the subsequent electricity price optimization decision, and are beneficial to realizing the real-time accurate demand response management of the user side.

A power consumer demand response potential assessment method based on a neural turing machine comprises the following steps:

grouping users by adopting a mean shift method to obtain user group data sets with different response characteristics;

based on the user group data set, executing an attention mechanism AM in a gating cycle unit GRU, and extracting the demand response behavior characteristics of the user in different meteorological environments;

and evaluating the demand response load adjustment quantity of the user at the given real-time electricity price based on the extracted demand response behavior characteristics through a neural turing machine model.

Further, grouping users by using a mean shift method to obtain user group data sets with different response characteristics, specifically comprising the following steps:

step 1.1, expanding and establishing a storage matrix data set required by training by a user historical data set;

step 1.2, in unmarked users, randomly selecting a user C as a central point, wherein the characteristic vector of the user C is omega _c,t (ii) a Determining the radius r of a sliding window, wherein the r is used for representing the characteristic vector floating range;

step 1.3, according to the characteristic vector omega of the user i in the area _i,t Calculating the drift vector S of the center point C in the region _h ，C _k Set of user points representing distances from the center C smaller than the radius r:

C _k (ω _c,t )＝{x:(ω _i,t -ω _c,t ) ^T (ω _i,t -ω _c,t )＜r ² }

a gaussian kernel function G (ω) is introduced to measure the actual offset contribution of each sample user point:

wherein h is the bandwidth of a Gaussian function;

after introducing the Gaussian kernel function, the vector S is drifted _h (ω _c,t ) Comprises the following steps:

wherein NU is the total number of users;

step 1.4, endowing different weight coefficients for each sample user point, and finally writing the drift vector into the following form:

wherein, the calculation formula of the weight coefficient is as follows:

step 1.5, updating the position of the central point, and drifting:

ω _c,t :＝ω _c,t +S _h

step 1.6, repeating the iterative moving process of the step 1.2-1.5 until the offset vector is smaller than a preset value to determine the central point of the current user group, and classifying the users in the current window radius r into a cluster C;

and step 1.7, traversing users which are not grouped, repeating the steps 1.2-1.6 until user points in the area are all marked, completing user grouping, and respectively updating the user group data set and the storage matrix data set of each group.

Further, step 1.1 specifically includes:

historical response count of user i during period tAccording to the inclusion of real-time electricity prices lambda _dri,t Environment data C _i,t Demand response load d _i,t Using historical data set N _i Represents:

N _i ＝{(λ _dri,1 ,C _i,1 ,d _i,1 ),(λ _dri,2 ,C _i,2 ,d _i,2 ),…,(λ _dri,t-1 ,C _i,t-1 ,d _i,t-1 ),(λ _dri,t ,C _i,t ,d _i,t )}

the environment data comprises the temperature and the humidity of the user at the current moment;

expansion is carried out on the historical data set to obtain a storage matrix data set R _i ：

R _i ＝{(ω _i,1 ,d _i,1 ),(ω _i,2 ,d _i,2 ),…,(ω _i,t-1 ,d _i,t-1 ),(ω _i,t ,d _i,t )}

Wherein, in the time period t, the feature vector omega _i,t The system consists of real-time price and environmental data of the current time period, and incentive electricity price, demand response and environmental meteorological data of the past L time periods:

ω _i,t ＝{λ _dri,t-L ,C _i,t-L ,d _i,t-L ,…,λ _dri,t-2 ,C _i,t-2 ,d _i,t-2 ,λ _dri,t-1 ,C _i,t-1 ,d _i,t-1 ,λ _dri,t ,C _i,t }。

further, based on the user group data set, an attention mechanism AM is executed in the gated loop unit GRU, and demand response behavior characteristics of the user in different meteorological environments are extracted, which specifically includes the following steps:

step 2.1: for GRU unit at t, input x _t For storing the user feature vector omega in the matrix data set _i,t Simultaneously reading the memory matrix M _t Memory vector m from the last training period _i,t-1 ；

Step 2.2: compute update Gate z _i,t And a reset gate r _i,t Updating the door z _i,t Used for determining the storage ratio of the user history information and the current input information, and resettingDoor r _i,t History information used to decide to forget:

z _i,t ＝σ(W ^z x _i,t +U ^z m _i,t-1 )

r _i,t ＝σ(W ^r x _i,t +U ^r m _i,t-1 )

where σ (·) is the activation function; w is a group of ^z 、W ^r And U ^z 、U ^r Respectively different weight coefficient matrixes;

step 2.3: combining the actually input x according to the calculation result of the reset gate _i,t Generating the latest characteristic content of the user in the current time period

Wherein, tanh (-) is an activation function, and W and U are weight coefficient matrixes;

step 2.4: calculating to obtain the latest transmission vector h _i,t Update the transmission matrix and combine h _i,t Attention focusing unit passed to period t:

step 2.5: performing content attention gathering, inputting h given by a controller _t And the memory matrix M _t Performing content-based attention clustering, measuring similarity by cosine distance, calculating cosine distance between input and each memory segment in memory matrix, and normalizing to obtain similarity weight

The calculation formula is as follows:

wherein, beta _t Representing parameters generated in the GRU processing process, wherein j is the sequence number of a memory vector except the ith user in the memory matrix;

step 2.6: according to the parameter g output by the controller _t Judging the degree of concentration of attention needs based on the content to obtain an interpolation vector

g _t Between 0 and 1, the closer to 1, the greater the degree of attention clustering based on the content:

step 2.7: through weight conversion, different memory segments are focused under different conditions, the purpose of accurately recalling a certain moment is realized, and s output by the GRU is output according to the controller _t Performing attention gathering based on positions, considering the influence of meteorological data of the external environment when the user actually responds, performing linear combination on each element of the whole weight vector again, giving higher weight to the memory vector under the similar environment, and obtaining a position vector m through position attention gathering _t (i) The calculation formula is as follows:

wherein s is _(i-j),t For a period t, a weight sequence of each memory vector in the memory matrix and the memory vector of the user i is related;

step 2.8: according to the parameter gamma output by the controller _t And performing exponential operation on the feature vectors and then normalizing to increase the discrimination:

wherein, γ _t To weight the sharpening parameter, m _i,t The memory vector of the user i in the t time period finally obtained after updating is the extracted demand response behavior characteristic;

step 2.9: for a given training time period T, the steps 2.1-2.8 are circulated until the extraction of the demand response behavior characteristics in the time period T of the user group is completed, and the memory matrix M is updated _t 。

Furthermore, the method comprises the following steps of evaluating the demand response load adjustment quantity of the user under the given real-time electricity price through a neural turing machine model based on the extracted demand response behavior characteristics:

step 3.1: reading original input data information from a user group data set, and respectively training neural turing models of different groups of users at different time periods based on a user response characteristic extraction process combining GRU and an attention mechanism to obtain the neural turing models of different groups of users at different time periods;

step 3.2: when actual real-time potential evaluation is carried out, selecting the neural turing machine model of the group of users corresponding to the time interval obtained by training in the step 3.1, inputting planned real-time electricity price in the first 15-25 minutes of the arrival of the evaluation time interval, and obtaining the evaluated user response result through the neural turing machine model;

step 3.3: after the user demand response time period is finished, acquiring the actual response result of the user under the condition of stimulating the electricity price in real time, and updating the user group data set and the storage matrix data set;

step 3.4: calculating an evaluation deviation rate according to an actual value and an evaluation value of user demand response, returning the evaluation deviation rate to the GRU unit, performing model feedback correction, and obtaining effect evaluation of the neural turing model on user demand response potential evaluation;

the root mean square error RMSE is used as a loss function of the GRU network, and the calculation formula is as follows:

in the formula, d and

respectively representing an actual value and a predicted value of user response, wherein n is total training days;

evaluating the evaluation effect of the neural turing machine model by using the average absolute percentage error MAPE, wherein the evaluation effect is better when the error is smaller, and the calculation formula of the average absolute percentage error is as follows:

further, step 3.1 specifically includes:

(1) The GRU processes input contents from users in the same group, wherein the input contents comprise two parts, one part is a user characteristic vector containing information such as real-time electricity price and environmental meteorological data, the other part is a memory vector from a storage matrix, and the GRU extracts a transmission characteristic vector of a user for demand response;

(2) The GRU transmits the transmission characteristic vector and the results of the parameters generated in the middle to the reading and writing head, wherein the parameters comprise beta _t 、g _t 、s _t 、γ _t For subsequent information attention gathering calculation and the read-write head to update and erase itself;

(3) The write head erases and writes the memory matrix, and updates the latest transmission vector from the GRU unit at the current moment; the write head performs write operation to modify the content of the memory matrix, the write operation mainly comprises two steps of clearing and writing, and based on the output of the controller, the write head generates an elimination vector e _t And an additional vector a _t Elimination vector e _t Each value in the range of 0 to 1, by varying degreesThe scaling of (2) weakens and eliminates different memory contents, and the additional vector performs memory superposition on the original basis to complete the writing of a new memory matrix:

M _t (i)＝M _t-1 (i)[1-m _t (i)e _t ]

M _t (i)＝M _t (i)+m _t (i)a _t

(4) The attention mechanism calculates and updates a memory vector and a memory matrix according to the latest transmission vector;

(5) Reading contents from the memory matrix by using a reading head, wherein the read contents are the latest memory characteristic content part in the t time period of the user, and returning information to the GRU unit of the controller;

(6) According to the latest memory vector content, the controller GRU calculates the user load adjustment amount through the full connection layer and outputs the result;

(7) And (5) updating the total memory matrix according to the real-time input, output and memory matrix, and circulating the steps (1) - (6) until the model training in all time periods is completed.

A neural turing machine-based power consumer demand response potential assessment device comprises:

the user group data set acquisition module is used for grouping users by adopting a mean shift method to obtain user group data sets with different response characteristics;

the demand response behavior feature extraction module is used for executing an attention mechanism AM in the gated loop unit GRU based on the user group data set and extracting demand response behavior features of the user in different meteorological environments;

and the demand response potential evaluation module is used for evaluating the demand response load adjustment amount of the user under the given real-time electricity price based on the extracted demand response behavior characteristics through the neural turing model.

Further, the user group data set obtaining module uses a mean shift method to group users to obtain user group data sets with different response characteristics, and the method specifically includes:

step 1.1, expanding and establishing a storage matrix data set required by training by a user historical data set;

step 1.2, in unmarked users, randomly selecting a user C as a central point, wherein the characteristic vector of the user C is omega _c,t (ii) a Determining a radius r of a sliding window, wherein the r is used for representing a characteristic vector floating range;

step 1.3, according to the characteristic vector omega of the user i in the area _i,t Calculating the drift vector S of the central point C in the area _h ，C _k Set of user points representing distances from the center C smaller than the radius r:

C _k (ω _c,t )＝{x:(ω _i,t -ω _c,t ) ^T (ω _i,t -ω _c,t )＜r ² }

a gaussian kernel function G (ω) is introduced to measure the actual offset contribution of each sample user point:

wherein h is the bandwidth of a Gaussian function;

after introducing the Gaussian kernel function, the vector S is drifted _h (ω _c,t ) Comprises the following steps:

wherein NU is the total number of users;

step 1.4, endowing each sample user point with different weight coefficients, and writing the drift vector into the following form:

wherein, the calculation formula of the weight coefficient is as follows:

step 1.5, updating the position of the central point, and drifting:

ω _c,t :＝ω _c,t +S _h

step 1.6, repeating the iterative moving process of the step 1.2-1.5 until the offset vector is smaller than a preset value to determine the central point of the current user group, and classifying the users in the current window radius r into a cluster C;

and step 1.7, traversing users which are not grouped, repeating the steps 1.2-1.6 until user points in the area are all marked, completing user grouping, and respectively updating the user group data set and the storage matrix data set of each group.

Further, step 1.1 specifically includes:

in the period t, the historical response data of the user i contains the real-time electricity price lambda _dri,t Environment data C _i,t Demand response load d _i,t Using historical data set N _i Represents:

N _i ＝{(λ _dri,1 ,C _i,1 ,d _i,1 ),(λ _dri,2 ,C _i,2 ,d _i,2 ),…,(λ _dri,t-1 ,C _i,t-1 ,d _i,t-1 ),(λ _dri,t ,C _i,t ,d _i,t )}

the environment data comprises the temperature and the humidity of the user at the current moment;

expansion is carried out on the historical data set to obtain a storage matrix data set R _i ：

R _i ＝{(ω _i,1 ,d _i,1 ),(ω _i,2 ,d _i,2 ),…,(ω _i,t-1 ,d _i,t-1 ),(ω _i,t ,d _i,t )}

Wherein, in the t period, the characteristic vector omega _i,t The system consists of real-time price and environmental data of the current time period, and incentive electricity price, demand response and environmental meteorological data of the past L time periods:

ω _i,t ＝{λ _dri,t-L ,C _i,t-L ,d _i,t-L ,…,λ _dri,t-2 ,C _i,t-2 ,d _i,t-2 ,λ _dri,t-1 ,C _i,t-1 ,d _i,t-1 ,λ _dri,t ,C _i,t }。

further, the demand response behavior feature extraction module executes an attention mechanism AM in the gated loop unit GRU based on the user group data set, and extracts demand response behavior features of the user in different meteorological environments, specifically including:

step 2.1: for GRU unit at t, input x _t For storing the user feature vector omega in the matrix data set _i,t Simultaneously reading the memory matrix M _t The memory vector m from the last training period _i,t-1 ；

Step 2.2: compute update gate z _i,t And a reset gate r _i,t Updating the door z _i,t For determining the storage ratio of the user history information and the current input information, resetting the gate r _i,t History information used to decide to forget:

z _i,t ＝σ(W ^z x _i,t +U ^z m _i,t-1 )

r _i,t ＝σ(W ^r x _i,t +U ^r m _i,t-1 )

where σ (·) is the activation function; w ^z 、W ^r And U ^z 、U ^r Respectively different weight coefficient matrixes;

step 2.3: combining the actually input x according to the calculation result of the reset gate _i,t Generating the latest characteristic content of the user in the current time period

Where, tanh (-) is the activation function, W and U are the weighting coefficient matrix;

step 2.4: calculating to obtain the latest transmission vector h _i,t Update the transmission matrix and sum h _i,t Attention focusing unit passed to period t:

step 2.5: performing content attention gathering, inputting h given by a controller _t And the memory matrix M _t Performing content-based attention clustering, measuring similarity by cosine distance, calculating cosine distance between input and each memory segment in memory matrix, and normalizing to obtain similarity weight

The calculation formula is as follows:

wherein, beta _t Representing parameters generated in the GRU processing process, wherein j is the sequence number of a memory vector except the ith user in the memory matrix;

step 2.6: interpolation modification

According to the parameter g output by the controller _t Judging the degree of concentration needed based on the content attention to obtain an interpolation vector

g _t Between 0 and 1, the closer to 1, the greater the degree of attention clustering based on the content:

step 2.7: performing position attention focusing:

the attention based on the position is gathered into the re-linear combination of the whole weight vector elements, namely, different memory segments are focused under different conditions through weight transformation, the purpose of accurately memorizing a certain moment is realized, and the aim of accurately memorizing the moment is realized according to s output by the GRU of the controller _t Performing attention gathering based on position, namely considering the influence of weather data of the external environment when the user actually responds, performing linear combination on each element of the whole weight vector again, giving higher weight to the memory vector under the similar environment, and obtaining a position vector m by position attention gathering _t (i) The calculation formula is as follows:

wherein s is _(i-j),t For a period t, memorizing a weight sequence of each memory vector in the matrix and the memory vector of the user i;

step 2.8: carrying out weight sharpening:

according to the parameter gamma output by the controller _t And performing exponential operation on the feature vectors and then normalizing to increase the discrimination:

wherein, gamma is _t To weight the sharpening parameter, m _i,t The memory vector of the user i in the t time period finally obtained after updating is the extracted demand response behavior characteristic;

step 2.9: for a given training time period T, the steps 2.1-2.8 are circulated until the extraction of the demand response behavior characteristics in the time period T of the user group is completed, and the memory matrix M is updated _t 。

Further, the demand response behavior feature extraction module evaluates the demand response load adjustment amount of the user at a given real-time electricity price through the neural turing machine model based on the extracted demand response behavior feature, and specifically includes:

step 3.1: reading original input data information from a user group data set, and respectively carrying out neural turing machine model training of different groups of users at different time periods based on a user response characteristic extraction process combining GRU and an attention mechanism;

step 3.2: when actual real-time potential evaluation is carried out, selecting a neural turing machine model of a time interval corresponding to the group of users, inputting planned real-time electricity prices in the first 15-25 minutes of the arrival of the evaluation time interval, and acquiring an evaluated user response result by the neural turing machine model, =;

step 3.3: after the user demand response time period is finished, acquiring the actual response result of the user under the condition of real-time excitation electricity price, and updating a user group data set and a storage matrix data set;

step 3.4: calculating an evaluation deviation rate according to an actual value and an evaluation value of user demand response, returning the evaluation deviation rate to the GRU unit, performing model feedback correction, and obtaining effect evaluation of the neural turing model on user demand response potential evaluation;

the root mean square error RMSE is used as a loss function of the GRU network, and the calculation formula is as follows:

in the formula, d and

respectively representing the actual value and the predicted value of the user response, wherein n is the total training days;

evaluating the evaluation effect of the neural turing machine model by using the average absolute percentage error MAPE, wherein the evaluation effect is better when the error is smaller, and the calculation formula of the average absolute percentage error is as follows:

further, step 3.1 specifically includes:

(1) The GRU processes input contents from users in the same group, wherein the input contents comprise two parts, one part is a user characteristic vector containing information such as real-time electricity price and environmental meteorological data, and the other part is a memory vector from a storage matrix;

(2) The GRU transmits the transmission characteristic vector and the results of the parameters generated in the middle to the reading and writing head, wherein the parameters comprise beta _t 、g _t 、s _t 、γ _t For subsequent information attention gathering calculation and the read-write head to update and erase itself;

(3) The read-write head erases and writes the memory matrix, and updates the latest transmission vector from the GRU unit at the current moment; the write head performs write operation to modify the content of the memory matrix, the write operation mainly comprises two steps of clearing and writing, and based on the output of the controller, the write head generates an elimination vector e _t And an additional vector a _t Eliminating the vector e _t Each value range is 0-1, and different memory contents are weakened and eliminated through scaling in different degrees. Additional vector a _t The range is not limited to 0-1, and the memory superposition is carried out on the original basis to complete the writing of a new memory matrix:

M _t (i)＝M _t-1 (i)[1-m _t (i)e _t ]

M _t (i)＝M _t (i)+m _t (i)a _t

(4) The attention mechanism calculates and updates a memory vector and a memory matrix according to the latest transmission vector;

(5) Reading contents from the memory matrix by using a reading head, wherein the read contents are the latest memory characteristic content part in the t time period of the user, and returning information to the GRU unit of the controller;

(6) According to the latest memory vector content, the controller GRU calculates the user load adjustment amount through the full connection layer and outputs the result;

(7) And (5) updating the total memory matrix according to the real-time input, output and memory matrix, and circulating the steps (1) - (6) until the model training in all time periods is completed.

The technical scheme provided by the invention has the beneficial effects that:

(1) On the aspect of an energy company, the neural turing machine-based power user demand response potential assessment method can help the energy company to more scientifically formulate a real-time dynamic price and demand response incentive policy based on the obtained high-precision response assessment value, and the conclusion obtained by analyzing and processing the user response detail result can be used for extracting user behavior characteristics, adjusting, perfecting and scientifically assessing the ongoing energy efficiency project of the energy company, and more reasonably designing and implementing future projects and allocating funds;

(2) On the aspect of resident users, the power consumer demand response potential assessment method based on the neural turing machine can be combined with meteorological environment conditions, user own power utilization habits, response capability and preference of dynamic prices, and assesses the load adjustment amount of a user under the real-time power price, so that the subsequent optimization power price formulation is guaranteed, the power utilization energy demand is met, the user energy cost is reduced, the comfort level of the user is guaranteed, and the power consumption economic benefit of residents can be greatly improved. In addition, the method deeply excavates the behavior habit characteristics of the user, and provides possibility for solving the self energy use habit of the user;

(3) On the social level, the neural-turing-based power consumer demand response potential assessment method can provide references of user behavior characteristic rules for policy makers, power grid dispatchers, efficient research teachers and students, salespeople and the like, can effectively promote strategic making meeting future development requirements of the society, can be applied to actual power energy dispatching and power consumer response management, provides references for exploring user behavior characteristics for researchers, prompts energy sales managers to fully excavate user-side response potential, mobilizes user-side response new energy smooth consumption requirements on the whole, ensures the safety and stability of power system operation, and effectively improves the intelligent social energy interaction level and the development of management technology.

Drawings

FIG. 1 is a flowchart illustrating an embodiment of a method for assessing a demand response potential of a power consumer based on a neural turing machine;

FIG. 2 is a schematic illustration of a demand response feature extraction process incorporating GRUs and attention mechanisms in accordance with the present invention;

FIG. 3 is a graph comparing the deviation of the present invention with a recurrent neural network RNN, a long short term memory artificial neural network LSTM, and a simple GRU model.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, but not all, embodiments of the present invention. All other embodiments, which can be obtained by a person skilled in the art without any inventive step based on the embodiments of the present invention, are within the scope of the present invention.

As shown in fig. 1, the embodiment of the present invention provides a neural turing machine-based power consumer demand response potential evaluation method, which includes mean shift user clustering grouping, power consumer response behavior feature extraction, and neural turing machine potential evaluation. The method specifically comprises the following steps:

step 1: grouping users by adopting a mean shift method to obtain user group data sets with different response characteristics, and specifically comprising the following steps:

and 1.1, expanding and establishing a storage matrix data set required by training by using a user historical data set.

In the period t, the historical response data of the user i contains the real-time electricity price lambda _dri,t Environment data C _i,t Demand response load d _i,t Etc. using historical data set N _i Represents:

N _i ＝{(λ _dri,1 ,C _i,1 ,d _i,1 ),(λ _dri,2 ,C _i,2 ,d _i,2 ),…,(λ _dri,t-1 ,C _i,t-1 ,d _i,t-1 ),(λ _dri,t ,C _i,t, d _i,t )}

the environment data includes the temperature and humidity of the user at the current moment.

It is considered that the response of the user in the current period is not only related to the current period, but also influenced by the price, the response amount and the like of the adjacent period. Thus, the memory matrix R is derived from the expansion of the historical data set _i ：

R _i ＝{(ω _i,1 ,d _i,1 ),(ω _i,2 ,d _i,2 ),…,(ω _i,t-1 ,d _i,t-1 ),(ω _i,t ,d _i,t )}

Wherein, t time period, the feature vector omega _i,t The system consists of real-time price and environmental data of the current time period, and incentive electricity price, demand response and environmental meteorological data of the past L time periods:

ω _i,t ＝{λ _dri,t-L ,C _i,t-L ,d _i,t-L ,…,λ _dri,t-2 ,C _i,t-2 ,d _i,t-2 ,λ _dri,t-1 ,C _i,t-1 ,d _i,t-1 ,λ _dri,t ,C _i,t }

step 1.2, in unmarked users, randomly selecting a user C as a central point, wherein the characteristic vector of the user C is omega _c,t (ii) a The radius r of the sliding window, i.e. the feature vector floating range, is determined.

Step 1.3, according to the characteristic vector omega of the user i in the area _i,t Calculating the drift vector S of the center point C in the region _h ，C _k Set of user points representing distances from the center C smaller than the radius r:

C _k (ω _c,t )＝{x:(ω _i,t -ω _c,t ) ^T (ω _i,t -ω _c,t )＜r ² }

in actual calculations, since the total mean offset contribution of each sample user point to the sample is different, a gaussian kernel function G (ω) is introduced to measure the actual offset contribution of each sample user point.

Where h is the bandwidth of the gaussian function.

After introducing the Gaussian kernel function, the vector S is drifted _h (ω _c,t ) Comprises the following steps:

wherein NU is the total number of users.

Step 1.4, considering that the actual demand response capability of each user is different, considering that the importance of different users is different, giving different weight coefficients to each sample user point, and finally writing the drift vector into the following form:

wherein, the calculation formula of the weight coefficient is as follows:

and step 1.5, updating the position of the central point and carrying out drifting.

ω _c,t ＝ω _c,t +S _h

And 1.6, repeating the iterative moving process of the steps 1.2-1.5 until the offset vector is smaller than a preset value (namely convergence occurs) and no higher density occurs, determining the central point of the current user group, and classifying the users in the current window radius r into a cluster C.

And step 1.7, traversing users which are not grouped, repeating the steps 1.2-1.6 until user points in the area are all marked, completing user grouping, and respectively updating the user group data set and the storage matrix data set of each group. .

And 2, step: based on the user group data set, an attention mechanism AM is executed in the gated loop unit GRU, and demand response behavior characteristics of the user in different meteorological environments are extracted, as shown in fig. 2, step 2 specifically includes the following steps:

step 2.1: for GRU unit at t, input x _t For storing the user feature vector omega in the matrix data set _i,t Simultaneously reading the memory matrix M _t Memory vector m from the last training period _i,t-1 。

Step 2.2: compute update gate z _i,t And a reset gate r _i,t The refresh gate determines the storage ratio of the user history information and the current input information, and the reset gate determines the history information to be forgotten.

z _i,t ＝σ(W ^z x _i,t +U ^z m _i,t-1 )

r _i,t ＝σ(W ^r x _i,t +U ^r m _i,t-1 )

Wherein, sigma (·) is an activation function, and the sigmoid function is adopted in the invention; w ^z W ^r And U ^z U ^r Respectively different weight coefficient matrices.

Step 2.3: combining the actually input x according to the calculation result of the reset gate _i,t Generating the latest characteristic content of the user in the current time period

Wherein, tanh (-) is an activation function; w and U are weight coefficient matrices.

Step 2.4: calculating to obtain the latest transmission vector h _i,t Update the transmission matrix and combine h _i,t To the attention focusing unit for the period t.

Step 2.5: content-based addressing is performed. H given by input controller _t And the memory matrix M _t Content-based attention gathering is performed. The invention adopts cosine distance to measure similarity, and obtains similarity weight by calculating cosine distance between input and each memory segment (characteristic vector) in the memory matrix and then normalizing

The calculation formula is as follows:

wherein, beta _t Represents the parameters generated in the GRU processing process, and j is the memory vector sequence number of the memory matrix except the ith user.

Step 2.6: interpolation modification (interpolation)

According to the parameter g output by the controller _t Judging the degree of concentration needed based on the content attention to obtain an interpolation vector

g _t Between 0 and 1, the closer to 1, the greater the degree of attention focusing on the content.

Step 2.7: location-based addressing is performed. Base ofThe attention at the position is gathered into the re-linear combination of the whole weight vector elements, namely, different memory segments are focused under different conditions through weight transformation, and the purpose of accurately recalling a certain moment (position) is realized. S according to the output of the controller GRU _t And performing attention gathering based on positions, namely considering the influence of weather data of the external environment when the user actually responds, performing linear combination on each element of the whole weight vector again, and giving higher weight to the memory vector in the similar environment. In the present invention, location attention gathering can better utilize weather data of the environment where the user is located. According to different response scenes, different response behavior habits of the user are focused with emphasis, and the accuracy and the efficiency of demand response mining are improved. The position attention is gathered to obtain a position vector m _t (i) The calculation formula is as follows:

wherein s is _(i-j),t For the period t, the weight sequence of each memory vector in the memory matrix and the memory vector of the user i is related.

Step 2.8: and carrying out weight sharpening. In practical applications, the response habits of the same user typically remain stable for a period of time. To prevent the difference between the feature vectors from being insignificant, the parameter γ is output from the controller _t And performing exponential operation on the characteristic vectors and then normalizing, so as to increase the discrimination and improve the positioning speed of the characteristic information.

Wherein, γ _t To weight the sharpening parameter, m _i,t And extracting the obtained response behavior characteristic information for the finally obtained memory vector of the user i in the t period after updating.

According to the model disclosed by the invention, an attention mechanism is added into a traditional GRU for improvement, so that the calculation efficiency, the evaluation accuracy and the long-time sequence processing capability of response characteristic extraction are effectively improved. Compared with the common neural algorithms such as LSTM and the like, the algorithm core GRU used by the method can improve the real-time calculation efficiency when processing low-dimensional input and output, and is suitable for quickly evaluating the actual demand of demand response potential of a plurality of users; in addition, the improved GRU utilizes an external memory matrix to simultaneously store meteorological data and user response characteristic information, and then uses an attention mechanism to accurately position information in mass data, so that long-distance dependency can be conveniently captured, and the real-time demand response load prediction accuracy of a user under different electricity prices is improved.

Step 2.9: for a given training time period T, the steps 2.1-2.8 are circulated until the extraction of the demand response behavior characteristics in the time period T of the user group is completed, and the memory matrix M is updated _t The overall process is shown in figure 1.

And step 3: through a neural turing machine model, based on the extracted demand response behavior characteristics, evaluating the demand response load adjustment amount of the user at a given real-time electricity price, wherein the step 3 specifically comprises the following steps:

step 3.1: original input data information is read from a user group data set, and training of neural turing machines of different groups of users in different time periods is respectively carried out based on a user response characteristic extraction process combining GRU and an attention mechanism.

(1) The GRU processes input content from the same group of users. The input content comprises two parts, wherein one part is a user characteristic vector containing information such as real-time electricity price and environmental meteorological data, the other part is a memory vector from a storage matrix, and the GRU extracts a transmission characteristic vector of a user for carrying out demand response.

(2) The GRU transmits the transmission characteristic vector and the intermediate generated parameters to the read-write head, wherein the parameters comprise beta _t 、g _t 、s _t 、γ _t And the like, which is used for subsequent information attention gathering calculation and the read-write head to carry out self updating and erasing work.

(3) The write head erases and writes the memory matrix. And updating the latest transmission vector from the GRU unit at the current moment.

The write head performs a write operation to modify the memory matrix contents,the write operation mainly comprises two steps of clearing and writing. Based on the output of the controller, the write head generates an erasure vector e _t And an additional vector a _t . Elimination vector e _t Each value range is 0-1, and different memory contents are weakened and eliminated through scaling in different degrees. Additional vector a _t The range is not limited to 0-1, and the memory superposition is carried out on the original basis to complete the writing of the new memory matrix.

M _t (i)＝M _t-1 (i)[1-m _t (i)e _t ]

M _t (i)＝M _t (i)+m _t (i)a _t

Wherein M is _t (i) Intermediate memory matrix for user i during t period, M _t (i) The memory matrix for user i is updated for the write head for a period t.

(4) The attention mechanism calculates and updates the memory vector and the memory matrix according to the latest transmission vector.

(5) And reading the content from the memory matrix by using the reading head, wherein the read content is the latest memory characteristic content part in the t time period of the user, and returning information to the controller GRU unit.

(6) And according to the latest memory vector content, the controller GRU calculates the user load adjustment quantity through the full connection layer and outputs the result.

(7) And (5) updating the total memory matrix according to the real-time input, output and memory matrix, and circulating (1) - (6) until model training in all time periods is completed.

Step 3.2: when actual real-time potential evaluation is carried out, selecting the neural turing machine model of the group of users in the corresponding time period, inputting planned real-time electricity prices in the first 15-25 minutes of the arrival of the evaluation time period, and obtaining the evaluated user response result, namely the load adjustment quantity of the user in the time period through the neural turing machine model.

Step 3.3: and after the user demand response time period is finished, acquiring the actual response result of the user, namely the load adjustment amount, and updating the user group data set and the storage matrix data set under the condition of stimulating the electricity price in real time.

Step 3.4: and calculating an evaluation deviation rate according to the actual value and the evaluation value of the user demand response, returning the evaluation deviation rate to the GRU unit, performing model feedback correction, and obtaining the effect evaluation of the neural turing model on the user demand response potential evaluation.

The Root Mean Square Error (RMSE) is used as a loss function of the GRU network, and is calculated as:

in the formula, d and

respectively representing the actual value and the predicted value of the user response, and n is the total training days.

Evaluating the evaluation effect of the neural turing machine model by using Mean Absolute Percentage Error (MAPE), wherein the evaluation effect is better when the error is smaller, and the calculation formula of the Mean absolute percentage error is as follows:

example verification

1. Description of the drawings: all the steps can be written by Python3.7, and in order to realize more convenience and more efficient calculation, the method is suggested to be realized on a TensorFlow2.2 platform, and a Keras version is 2.3.1.

2. Example verification content

In order to compare the response evaluation effect of the NTM provided by the invention, a Recurrent Neural Network (RNN), a Long-Short Term Memory artificial Neural Network (LSTM) and a simple GRU model are further established for comparison. 4 representative sessions were selected for continuous training, and MAPE pairs of the training results of the four algorithms are shown in fig. 3. Generally speaking, the neural turing machine can rapidly achieve good prediction accuracy when the training times are less. The overall estimation deviation of the RNN is large, and the error is obvious in partial time period and is kept at about 15%. The LSTM can achieve MAPE reduction to about 8% in 5000 times of training, the GRU final deviation is close to the LSTM, but the training speed is faster than the LSTM. The accuracy and the training efficiency of the neural turing machine are high, and when the training times exceed 5000 times, the MAPE can be reduced to about 2%. The result shows that the attention mechanism in the neural turing machine can help the GRU to accurately extract the user characteristics, and the algorithm performance is improved from the two aspects of accuracy and efficiency.

The embodiment of the invention also provides a device for evaluating the demand response potential of the power consumer based on the neural turing machine, which comprises:

the user group data set acquisition module is used for grouping users by adopting a mean shift method to obtain user group data sets with different response characteristics;

the demand response behavior feature extraction module is used for executing an attention mechanism AM in a gated cyclic unit GRU based on the user group data set and extracting demand response behavior features of the user in different meteorological environments;

and the demand response potential evaluation module is used for evaluating the demand response load adjustment amount of the user under the given real-time electricity price based on the extracted demand response behavior characteristics through the neural turing model.

The user group data set obtaining module adopts a mean shift method to group users to obtain user group data sets with different response characteristics, and the method specifically comprises the following steps:

step 1.1, expanding and establishing a storage matrix data set required by training by a user historical data set;

step 1.2, in unmarked users, randomly selecting a user C as a central point, wherein the characteristic vector of the user C is omega _c,t (ii) a Determining the radius r of a sliding window, wherein the r is used for representing the characteristic vector floating range;

step 1.3, according to the characteristic vector omega of the user i in the area _i,t Calculating the drift vector S of the center point C in the region _h ，C _k Set of user points representing distances from the center C smaller than the radius r:

C _k (ω _c,t )＝{x:(ω _i,t -ω _c,t ) ^T (ω _i,t -ω _c,t )＜r ² }

a gaussian kernel function G (ω) is introduced to measure the actual offset contribution of each sample user point:

wherein h is the bandwidth of a Gaussian function;

after introducing the Gaussian kernel function, the vector S is drifted _h (ω _c,t ) Comprises the following steps:

wherein NU is the total number of users;

step 1.4, endowing different weight coefficients for each sample user point, and finally writing the drift vector into the following form:

wherein, the calculation formula of the weight coefficient is as follows:

step 1.5, updating the position of the central point, and drifting:

ω _c,t :＝ω _c,t +S _h

step 1.6, repeating the iterative moving process of the step 1.2-1.5 until the offset vector is smaller than a preset value to determine the central point of the current user group, and classifying the users in the current window radius r into a cluster C;

and step 1.7, traversing users which are not grouped, repeating the steps 1.2-1.6 until user points in the area are all marked, completing user grouping, and respectively updating the user group data set and the storage matrix data set of each group.

Wherein, step 1.1 specifically includes:

in the period t, the historical response data of the user i contains the real-time electricity price lambda _dri,t Environment data C _i,t Demand response load d _i,t Using historical data set N _i Represents:

N _i ＝{(λ _dri,1 ,C _i,1 ,d _i,1 ),(λ _dri,2 ,C _i,2 ,d _i,2 ),…,(λ _dri,t-1 ,C _i,t-1 ,d _i,t-1 ),(λ _dri,t ,C _i,t ,d _i,t )}

the environment data comprises the temperature and the humidity of the user at the current moment;

expansion is carried out on the historical data set to obtain a storage matrix data set R _i ：

R _i ＝{(ω _i,1 ,d _i,1 ),(ω _i,2 ,d _i,2 ),…,(ω _i,t-1 ,d _i,t-1 ),(ω _i,t ,d _i,t )}

Wherein, in the time period t, the feature vector omega _i,t The system consists of real-time price and environmental data of the current time period, and incentive electricity price, demand response and environmental meteorological data of past L time periods:

ω _i,t ＝{λ _dri,t-L ,C _i,t-L ,d _i,t-L ,…,λ _dri,t-2 ,C _i,t-2 ,d _i,t-2 ,λ _dri,t-1 ,C _i,t-1 ,d _i,t-1 ,λ _dri,t ,C _i,t }。

the demand response behavior feature extraction module executes an attention mechanism AM in a gated loop unit GRU based on the user group data set, and extracts demand response behavior features of the user in different meteorological environments, and specifically includes:

step 2.1: for GRU unit at t, input x _t For storing user characteristics in a matrix data setVector omega _i,t Simultaneously reading the memory matrix M _t Memory vector m from the last training period _i,t-1 ；

Step 2.2: compute update gate z _i,t And a reset gate r _i,t Updating the door z _i,t For determining the storage ratio of the user history information and the current input information, resetting the gate r _i,t History information used to decide to forget:

z _i,t ＝σ(W ^z x _i,t +U ^z m _i,t-1 )

r _i,t ＝σ(W ^r x _i,t +U ^r m _i,t-1 )

where σ (·) is the activation function; w ^z 、W ^r And U ^z 、U ^r Respectively different weight coefficient matrixes;

step 2.3: combining the actually input x according to the calculation result of the reset gate _i,t Generating the latest characteristic content of the user in the current time period

Where, tanh (-) is the activation function, W and U are the weighting coefficient matrix;

step 2.4: calculating to obtain the latest transmission vector h _i,t Update the transmission matrix and combine h _i,t Attention focusing unit passed to period t:

step 2.5: for concentration of content, given by input controllerh _t And the memory matrix M _t Performing content-based attention clustering, measuring similarity by cosine distance, calculating cosine distance between input and each memory segment in memory matrix, and normalizing to obtain similarity weight

The calculation formula is as follows:

wherein, beta _t Representing parameters generated in the GRU processing process, wherein j is a memory vector sequence number except the ith user in the memory matrix;

step 2.6: according to the parameter g output by the controller _t Judging the degree of concentration needed based on the content attention to obtain an interpolation vector

g _t Between 0 and 1, the closer to 1, the greater the degree of attention clustering based on the content:

step 2.7: through weight conversion, different memory segments are focused under different conditions, the purpose of accurately recalling a certain moment is realized, and s output by the GRU is output according to the controller _t Performing attention gathering based on positions, considering the influence of meteorological data of the external environment when the user actually responds, performing linear combination on each element of the whole weight vector again, giving higher weight to the memory vector under the similar environment, and obtaining a position vector m through position attention gathering _t (i) The calculation formula is as follows:

wherein s is _(i-j),t For a period t, a weight sequence of each memory vector in the memory matrix and the memory vector of the user i is related;

step 2.8: according to the parameter gamma output by the controller _t And carrying out exponential operation on the characteristic vectors and then normalizing to increase the discrimination:

wherein, γ _t To weight the sharpening parameter, m _i,t The memory vector of the user i in the t period finally obtained after updating, namely the extracted demand response behavior characteristic;

step 2.9: for a given training time period T, the steps 2.1-2.8 are circulated until the extraction of the demand response behavior characteristics in the time period T of the user group is completed, and the memory matrix M is updated _t 。

The demand response behavior feature extraction module is used for evaluating the demand response load adjustment quantity of the user under the given real-time electricity price through a neural turing model based on the extracted demand response behavior feature, and specifically comprises the following steps:

step 3.1: reading original input data information from a user group data set, and respectively carrying out neural turing machine model training of different groups of users at different time periods based on a user response characteristic extraction process combining GRU and an attention mechanism;

step 3.2: when actual real-time potential evaluation is carried out, selecting a neural turing machine model of the group of users in a corresponding time period, inputting a planned real-time electricity price in the first 15-25 minutes of the arrival of the evaluation time period, and obtaining an evaluated user response result through the neural turing machine model;

step 3.3: after the user demand response time period is finished, acquiring the actual response result of the user under the condition of stimulating the electricity price in real time, and updating the user group data set and the storage matrix data set;

step 3.4: calculating an evaluation deviation rate according to an actual value and an evaluation value of the user demand response, returning the evaluation deviation rate to the GRU unit, performing model feedback correction, and obtaining effect evaluation of the neural turing model on the user demand response potential evaluation;

the root mean square error RMSE is used as a loss function of the GRU network, and the calculation formula is as follows:

in the formula, d and

respectively representing an actual value and a predicted value of user response, wherein n is total training days;

the evaluation effect of the neural turing machine model is evaluated by using the average absolute percentage error MAPE, the evaluation effect is better when the error is smaller, and the calculation formula of the average absolute percentage error is as follows:

wherein, step 3.1 specifically includes:

(1) The GRU processes input contents from users in the same group, wherein the input contents comprise two parts, one part is a user characteristic vector containing real-time electricity price and environmental meteorological data information, and the other part is a memory vector from a storage matrix;

(2) The GRU transmits the transmission characteristic vector and the parameter generated in the middle to the reading and writing head, wherein the parameter comprises beta _t 、g _t 、s _t 、γ _t For subsequent information attention gathering calculation and the read-write head to update and erase itself;

(3) The write head erases and writes the memory matrix, and updates the latest transmission vector from the GRU unit at the current moment; the write head performs write operation to modify the content of the memory matrix, the write operation mainly includes two operations of erasing and writingA step in which the write head generates an erasure vector e based on the output of the controller _t And an additional vector a _t Eliminating the vector e _t Each value range is 0-1, and the vector a is added by weakening and eliminating different memory contents through scaling of different degrees _t And (3) performing memory superposition on the original basis to finish the writing of a new memory matrix:

M _t (i)＝M _t-1 (i)[1-m _t (i)e _t ]

M _t (i)＝M _t (i)+m _t (i)a _t

(4) The attention mechanism calculates and updates a memory vector and a memory matrix according to the latest transmission vector;

(5) Reading contents from the memory matrix by using a reading head, wherein the read contents are the latest memory characteristic content part of the user t time period, and returning information to a GRU unit of the controller;

(6) According to the latest memory vector content, the controller GRU calculates the user load adjustment amount through the full connection layer and outputs the result;

(7) And (5) updating the total memory matrix according to the real-time input, output and memory matrix, and circulating the steps (1) - (6) until the model training in all time periods is completed.

The embodiment of the invention also provides a new energy station sending-out line differential protection system based on instantaneous value integration, which comprises: a computer-readable storage medium and a processor;

the computer readable storage medium is used for storing executable instructions;

the processor is configured to read executable instructions stored in the computer-readable storage medium, and execute the new energy station outgoing line differential protection method based on instantaneous value integration according to the first aspect.

An embodiment of the present invention further provides a non-transitory computer-readable storage medium, on which a computer program is stored, where the computer program, when executed by a processor, implements the new energy station outgoing line differential protection method based on instantaneous value integration according to the first aspect.

As will be appreciated by one skilled in the art, embodiments of the present invention may be provided as a method, system, or computer program product. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.

The present invention is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

Compared with the algorithms such as K-means clustering and the like, the mean shift clustering algorithm does not need to stipulate the number of clusters for the mean shift clustering algorithm in advance, but classifies the mean shift clustering algorithm by means of natural data driving, and is more suitable for the requirement of independently exploring user energy characteristics in the open market; then, the GRU and the attention mechanism are combined to extract the user response behavior characteristics, the GRU can efficiently and quickly process input data, the attention mechanism and the memory module are introduced to quickly position the response behavior in a similar environment based on environmental meteorological data, the capture capability of the GRU algorithm on long-distance dependency is enhanced, and the demand response load adjustment quantity of a user under real-time electricity price is accurately and efficiently mined; and finally, based on the response behavior characteristics, finishing the real-time demand response potential evaluation of the user by using the trained neural turing machine model to obtain the load amount actively adjusted by the user under different real-time electricity prices. The method can effectively improve the accuracy of the user demand response potential evaluation, reduce the calculation burden when a power grid company processes a large number of small users, effectively improve the real-time evaluation efficiency, provide a reliable basis for the later real-time electricity price optimizing decision, facilitate the user real-time demand response management according to the environment condition, and further promote the safe and stable operation of the power system.

Finally, it should be noted that: the above embodiments are only for illustrating the technical solutions of the present invention and not for limiting the same, and although the present invention is described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: modifications and equivalents may be made to the embodiments of the invention without departing from the spirit and scope of the invention, which is to be covered by the claims.

Claims (12)

1. A power consumer demand response potential assessment method based on a neural turing machine is characterized by comprising the following steps:

grouping users by adopting a mean shift method to obtain user group data sets with different response characteristics;

based on the user group data set, executing an attention mechanism AM in a gating cycle unit GRU, and extracting the demand response behavior characteristics of the user in different meteorological environments;

and evaluating the demand response load adjustment quantity of the user at the given real-time electricity price based on the extracted demand response behavior characteristics through a neural turing machine model.

2. The neural turing machine-based power consumer demand response potential assessment method according to claim 1, characterized in that users are grouped by a mean shift method to obtain a user group data set with different response characteristics, specifically comprising the steps of:

step 1.1, expanding and establishing a storage matrix data set required by training by a user historical data set;

step 1.2, in unmarked users, randomly selecting a user C as a central point, wherein the characteristic vector of the user C is omega _c,t (ii) a Determining the radius r of a sliding window, wherein the r is used for representing the characteristic vector floating range;

step 1.3, according to the characteristic vector omega of the user i in the area _i,t Calculating the drift vector S of the center point C in the region _h ，C _k Set of user points representing distances from the center C smaller than the radius r:

C _k (ω _c,t )＝{x:(ω _i,t -ω _c,t ) ^T (ω _i,t -ω _c,t )＜r ² }

a gaussian kernel function G (ω) is introduced to measure the actual offset contribution of each sample user point:

wherein h is the bandwidth of a Gaussian function;

after the introduction of the gaussian kernel function,drift vector S _h (ω _c,t ) Comprises the following steps:

wherein NU is the total number of users;

step 1.4, endowing each sample user point with different weight coefficients, and writing the drift vector into the following form:

wherein, the calculation formula of the weight coefficient is as follows:

step 1.5, updating the position of the central point, and drifting:

ω _c,t :＝ω _c,t +S _h

step 1.6, repeating the iterative moving process of the step 1.2-1.5 until the offset vector is smaller than a preset value to determine the central point of the current user group, and classifying the users in the current window radius r into a cluster C;

and step 1.7, traversing users which are not grouped, repeating the steps 1.2-1.6 until user points in the area are marked, completing user grouping, and respectively updating the user group data set and the storage matrix data set of each group.

3. The neural turing machine-based power consumer demand response potential assessment method of claim 2, wherein: step 1.1 specifically comprises:

in the period t, the historical response data of the user i contains the real-time electricity price lambda _dri,t Environment data C _i,t Demand response load d _i,t Using historical data set N _i Represents:

N _i ＝{(λ _dri,1 ,C _i,1 ,d _i,1 ),(λ _dri,2 ,C _i,2 ,d _i,2 ),…,(λ _dri,t-1 ,C _i,t-1 ,d _i,t-1 ),(λ _dri,t ,C _i,t ,d _i,t )}

the environment data comprises the temperature and the humidity of the user at the current moment;

expansion is carried out on the historical data set to obtain a storage matrix data set R _i ：

R _i ＝{(ω _i,1 ,d _i,1 ),(ω _i,2 ,d _i,2 ),…,(ω _i,t-1 ,d _i,t-1 ),(ω _i,t ,d _i,t )}

Wherein, in the time period t, the feature vector omega _i,t The system consists of real-time price and environmental data of the current time period, and incentive electricity price, demand response and environmental meteorological data of the past L time periods:

ω _i,t ＝{λ _dri,t-L ,C _i,t-L ,d _i,t-L ,…,λ _dri,t-2 ,C _i,t-2 ,d _i,t-2 ,λ _dri,t-1 ,C _i,t-1 ,d _i,t-1 ,λ _dri,t ,C _i,t }。

4. the method for assessing the demand response potential of the electric power user based on the Neuropter's machine as claimed in claim 2, wherein based on the user group data set, an attention mechanism AM is executed in the gated loop unit GRU to extract the demand response behavior characteristics of the user in different meteorological environments, and the method specifically comprises the following steps:

step 2.1: for GRU unit at t, input x _t For storing the user feature vector omega in the matrix data set _i,t Simultaneously reading the memory matrix M _t Memory vector m from the last training period _i,t-1 ；

Step 2.2: compute update Gate z _i,t And a reset gate r _i,t Updating the door z _i,t For determining the storage ratio of the user history information and the current input information, resetting the gate r _i,t History information used to decide to forget:

z _i,t ＝σ(W ^z x _i,t +U ^z m _i,t-1 )

r _i,t ＝σ(W ^r x _i,t +U ^r m _i,t-1 )

where σ (·) is the activation function; w ^z 、W ^r And U ^z 、U ^r Respectively different weight coefficient matrixes;

step 2.3: combining the actually input x according to the calculation result of the reset gate _i,t Generating the latest characteristic content of the user in the current time period

Where, tanh (-) is the activation function, W and U are the weighting coefficient matrix;

step 2.4: calculating to obtain the latest transmission vector h _i,t Update the transmission matrix and combine h _i,t Attention focusing unit passed to period t:

step 2.5: performing content attention gathering, inputting h given by the controller _t And the memory matrix M _t Performing content-based attention clustering, measuring similarity by cosine distance, calculating cosine distance between input and each memory segment in memory matrix, and normalizing to obtain similarity weight

The calculation formula is as follows:

wherein, beta _t Representing parameters generated in the GRU processing process, wherein j is the sequence number of a memory vector except the ith user in the memory matrix;

step 2.6: according to the parameter g output by the controller _t Judging the degree of concentration needed based on the content attention to obtain an interpolation vector

g _t Between 0 and 1, the closer to 1, the greater the degree of attention clustering based on the content:

step 2.7: through weight conversion, different memory segments are focused under different conditions, the purpose of accurately recalling a certain moment is realized, and s output by the GRU is output according to the controller _t Performing attention gathering based on positions, considering the influence of meteorological data of the external environment when the user actually responds, performing linear combination on each element of the whole weight vector again, giving higher weight to the memory vector under the similar environment, and obtaining a position vector m through position attention gathering _t (i) The calculation formula is as follows:

wherein s is _(i-j),t For a period of t, memorize each in the matrixA weight sequence of the memory vector and the memory vector of the user i;

step 2.8:

according to the parameter gamma output by the controller _t And performing exponential operation on the feature vectors and then normalizing to increase the discrimination:

wherein, γ _t To weight the sharpening parameter, m _i,t The memory vector of the user i in the t period finally obtained after updating, namely the extracted demand response behavior characteristic;

step 2.9: for a given training time period T, the steps 2.1-2.8 are circulated until the extraction of the demand response behavior characteristics in the time period T of the user group is completed, and the memory matrix M is updated _t 。

5. The neural turing machine-based power consumer demand response potential assessment method of claim 4, wherein the method comprises the following steps: through a neural turing machine model, based on the extracted demand response behavior characteristics, evaluating the demand response load adjustment amount of a user under a given real-time electricity price, and specifically comprising the following steps:

step 3.1: reading original input data information from user group data in a centralized manner, and respectively training neural turing machines of different groups of users in different time periods based on a user response characteristic extraction process combining GRU and an attention mechanism to obtain neural turing machines of different groups of users in different time periods;

step 3.2: when actual real-time potential evaluation is carried out, selecting the neural turing machine model of the group of users corresponding to the time interval obtained by training in the step 3.1, inputting planned real-time electricity price in the first 15-25 minutes of the arrival of the evaluation time interval, and obtaining the evaluated user response result, namely the load adjustment quantity of the user in the time interval through the neural turing machine model;

step 3.3: after the user demand response time period is finished, acquiring the actual response result of the user, namely the load adjustment amount, and updating the user group data set and the storage matrix data set under the condition of stimulating the electricity price in real time;

step 3.4: calculating an evaluation deviation rate according to an actual value and an evaluation value of user demand response, returning the evaluation deviation rate to the GRU unit, performing model feedback correction, and obtaining effect evaluation of the neural turing model on user demand response potential evaluation;

the root mean square error RMSE is used as a loss function of the GRU network, and the calculation formula is as follows:

in the formula, d and

respectively representing an actual value and a predicted value of user response, wherein n is total training days;

evaluating the evaluation effect of the neural turing machine model by using the average absolute percentage error MAPE, wherein the evaluation effect is better when the error is smaller, and the calculation formula of the average absolute percentage error is as follows:

6. the neural turing machine-based power consumer demand response potential assessment method according to claim 5, wherein the step 3.1 specifically comprises:

(1) The GRU processes input contents from users in the same group, wherein the input contents comprise two parts, one part is a user characteristic vector containing information such as real-time electricity price and environmental meteorological data, the other part is a memory vector from a storage matrix, and the GRU extracts a transmission characteristic vector of a user for demand response;

(2) The GRU transmits the transmission characteristic vector and the results of the parameters generated in the middle to the reading and writing head, wherein the parameters comprise beta _t 、g _t 、s _t 、γ _t For after useContinuing information attention gathering calculation and the read-write head to carry out self updating and erasing work;

(3) The write head erases and writes the memory matrix, and updates the latest transmission vector from the GRU unit at the current moment; the write head performs write operation to modify the content of the memory matrix, the write operation mainly comprises two steps of clearing and writing, and based on the output of the controller, the write head generates an elimination vector e _t And an additional vector a _t Elimination vector e _t Each value range is 0-1, and different memory contents are weakened and eliminated through scaling of different degrees. Additional vector a _t The range is not limited to 0-1, and the memory superposition is carried out on the original basis to complete the writing of a new memory matrix:

M _t (i)＝M _t-1 (i)[1-m _t (i)e _t ]

M _t (i)＝M _t (i)+m _t (i)a _t

(4) The attention mechanism calculates and updates a memory vector and a memory matrix according to the latest transmission vector;

(5) Reading contents from the memory matrix by using a reading head, wherein the read contents are the latest memory characteristic content part in the t time period of the user, and returning information to the GRU unit of the controller;

(6) According to the latest memory vector content, the controller GRU calculates the user load adjustment amount through a full connection layer and outputs a result;

(7) And (5) updating the total memory matrix according to the real-time input, output and memory matrix, and circulating the steps (1) - (6) until the model training in all time periods is completed.

7. A device for assessing demand response potential of a power consumer based on a neural turing machine is characterized by comprising:

the user group data set acquisition module is used for grouping users by adopting a mean shift method to obtain user group data sets with different response characteristics;

the demand response behavior feature extraction module is used for executing an attention mechanism AM in the gated loop unit GRU based on the user group data set and extracting demand response behavior features of the user in different meteorological environments;

and the demand response potential evaluation module is used for evaluating the demand response load adjustment quantity of the user under the given real-time electricity price based on the extracted demand response behavior characteristics through a neural Turing machine model.

8. The device for evaluating the demand response potential of the power consumer based on the neural turing machine, according to claim 7, wherein the consumer group data set obtaining module uses a mean shift method to group the consumers to obtain consumer group data sets with different response characteristics, and specifically comprises:

step 1.1, a storage matrix data set required by training is established by expanding a user historical data set;

step 1.2, in unmarked users, randomly selecting a user C as a central point, wherein the characteristic vector of the user C is omega _c,t (ii) a Determining the radius r of a sliding window, wherein the r is used for representing the characteristic vector floating range;

step 1.3, according to the characteristic vector omega of the user i in the area _i,t Calculating the drift vector S of the center point C in the region _h ，C _k Set of user points representing distances from the center C smaller than the radius r:

C _k (ω _c,t )＝{x:(ω _i,t -ω _c,t ) ^T (ω _i,t -ω _c,t )＜r ² }

a gaussian kernel function G (ω) is introduced to measure the actual offset contribution of each sample user point:

wherein h is the bandwidth of a Gaussian function;

introducing GaussAfter kernel function, drift vector S _h (ω _c,t ) Comprises the following steps:

wherein NU is the total number of users;

step 1.4, endowing different weight coefficients for each sample user point, and finally writing the drift vector into the following form:

wherein, the calculation formula of the weight coefficient is as follows:

step 1.5, updating the position of the central point, and drifting:

ω _c,t :＝ω _c,t +S _h

step 1.6, repeating the iterative moving process of the step 1.2-1.5 until the offset vector is smaller than a preset value to determine the central point of the current user group, and classifying the users in the current window radius r into a cluster C;

and step 1.7, traversing users which are not grouped, repeating the steps 1.2-1.6 until user points in the area are all marked, completing user grouping, and respectively updating the user group data set and the storage matrix data set of each group.

9. The device for assessing the demand response potential of the electric power user based on the neural turing machine, according to claim 8, wherein the step 1.1 specifically comprises:

in the period t, the historical response data of the user i contains the real-time electricity price lambda _dri,t Environment data C _i,t Demand response load d _i,t Using historical data set N _i Represents:

N _i ＝{(λ _dri,1 ,C _i,1 ,d _i,1 ),(λ _dri,2 ,C _i,2 ,d _i,2 ),…,(λ _dri,t-1 ,C _i,t-1 ,d _i,t-1 ),(λ _dri,t ,C _i,t ,d _i,t )}

the environment data comprises the temperature and the humidity of the user at the current moment;

expansion is carried out on the historical data set to obtain a storage matrix data set R _i ：

R _i ＝{(ω _i,1 ,d _i,1 ),(ω _i,2 ,d _i,2 ),…,(ω _i,t-1 ,d _i,t-1 ),(ω _i,t ,d _i,t )}

Wherein, in the time period t, the feature vector omega _i,t The system consists of real-time price and environmental data of the current time period, and incentive electricity price, demand response and environmental meteorological data of the past L time periods:

ω _i,t ＝{λ _dri,t-L ,C _i,t-L ,d _i,t-L ,…,λ _dri,t-2 ,C _i,t-2 ,d _i,t-2 ,λ _dri,t-1 ,C _i,t-1 ,d _i,t-1 ,λ _dri,t ,C _i,t }。

10. the device for assessing the demand response potential of the electric power user based on a neuroturing machine as claimed in claim 9, wherein the demand response behavior feature extraction module executes an attention mechanism AM in the gated loop unit GRU based on the user group data set to extract the demand response behavior feature of the user in different meteorological environments, and specifically comprises:

step 2.1: for GRU unit at t, input x _t For storing the user feature vector omega in the matrix data set _i,t Simultaneously reading the memory matrix M _t Memory vector m from the last training period _i,t-1 ；

Step 2.2: compute update Gate z _i,t And a reset gate r _i,t Updating the door z _i,t For determining the storage ratio of the user history information and the current input information, resetting the gate r _i,t History information used to decide to forget:

z _i,t ＝σ(W ^z x _i,t +U ^z m _i,t-1 )

r _i,t ＝σ(W ^r x _i,t +U ^r m _i,t-1 )

where σ (·) is the activation function; w is a group of ^z 、W ^r And U ^z 、U ^r Respectively different weight coefficient matrixes;

step 2.3: combining the actually input x according to the calculation result of the reset gate _i,t Generating the latest characteristic content of the user in the current time period

Where, tanh (-) is the activation function, W and U are the weighting coefficient matrix;

step 2.4: calculating to obtain the latest transmission vector h _i,t Update the transmission matrix and combine h _i,t Attention focusing unit passed to period t:

step 2.5: performing content attention gathering, inputting h given by the controller _t And the memory matrix M _t Performing content-based attention clustering, measuring similarity by cosine distance, calculating cosine distance between input and each memory segment in memory matrix, and normalizing to obtain similarity weight

The calculation formula is as follows:

wherein, beta _t Representing parameters generated in the GRU processing process, wherein j is a memory vector sequence number except the ith user in the memory matrix;

step 2.6: according to the parameter g output by the controller _t Judging the degree of concentration needed based on the content attention to obtain an interpolation vector

g _t Between 0 and 1, the closer to 1, the greater the degree of attention clustering based on the content:

step 2.7: through weight conversion, different memory segments are focused under different conditions, the purpose of accurately recalling a certain moment is realized, and s output by the GRU is output according to the controller _t Performing attention gathering based on positions, considering the influence of meteorological data of the external environment when the user actually responds, performing linear combination on each element of the whole weight vector again, giving higher weight to the memory vector under the similar environment, and obtaining a position vector m through position attention gathering _t (i) The calculation formula is as follows:

wherein s is _(i-j),t For a period t, a weight sequence of each memory vector in the memory matrix and the memory vector of the user i is related;

step 2.8: according to the parameter gamma output by the controller _t And performing exponential operation on the feature vectors and then normalizing to increase the discrimination:

wherein, gamma is _t To weight the sharpening parameter, m _i,t The memory vector of the user i in the t period finally obtained after updating, namely the extracted demand response behavior characteristic;

step 2.9: for a given training time period T, the steps 2.1-2.8 are circulated until the extraction of the demand response behavior characteristics in the time period T of the user group is completed, and the memory matrix M is updated _t 。

11. The device for evaluating the demand response potential of the power consumer based on the neural turing machine, according to claim 10, wherein the demand response behavior feature extraction module evaluates the demand response load adjustment amount of the power consumer at a given real-time electricity price based on the extracted demand response behavior feature through the neural turing machine model, and specifically comprises:

step 3.1: reading original input data information from a user group data set, and respectively training neural turing machines of different groups of users at different time periods based on a user response characteristic extraction process combining GRU and an attention mechanism;

step 3.2: when actual real-time potential evaluation is carried out, selecting a neural turing machine model of the group of users in a corresponding time period, inputting a planned real-time electricity price in the first 15-25 minutes of the arrival of the evaluation time period, and obtaining an evaluated user response result through the neural turing machine model;

step 3.3: after the user demand response time period is finished, acquiring the actual response result of the user under the condition of real-time excitation electricity price, and updating a user group data set and a storage matrix data set;

step 3.4: calculating an evaluation deviation rate according to an actual value and an evaluation value of user demand response, returning the evaluation deviation rate to the GRU unit, performing model feedback correction, and obtaining effect evaluation of the neural turing model on user demand response potential evaluation;

the root mean square error RMSE is used as a loss function of the GRU network, and the calculation formula is as follows:

in the formula, d and

respectively representing the actual value and the predicted value of the user response, wherein n is the total training days;

evaluating the evaluation effect of the neural turing machine model by using the average absolute percentage error MAPE, wherein the evaluation effect is better when the error is smaller, and the calculation formula of the average absolute percentage error is as follows:

12. the device for assessing the demand response potential of the electric power user based on the neural turing machine, according to claim 11, wherein the step 3.1 specifically comprises:

(1) The GRU processes input contents from users in the same group, wherein the input contents comprise two parts, one part is a user characteristic vector containing real-time electricity price and environmental meteorological data information, and the other part is a memory vector from a storage matrix;

(2) The GRU transmits the transmission characteristic vector and the parameter generated in the middle to the reading and writing head, wherein the parameter comprises beta _t 、g _t 、s _t 、γ _t For subsequent information attention gathering calculation and the read-write head to update and erase itself;

(3) The write head erases, writes, and updates the memory matrixThe latest transmission vector from the GRU unit at the previous moment; the write head performs write operation to modify the content of the memory matrix, the write operation mainly comprises two steps of clearing and writing, and based on the output of the controller, the write head generates an elimination vector e _t And an additional vector a _t Elimination vector e _t Each value range is 0-1, and the vector a is added by weakening and eliminating different memory contents through scaling of different degrees _t And (3) performing memory superposition on the original basis to finish the writing of a new memory matrix:

M _t (i)＝M _t-1 (i)[1-m _t (i)e _t ]

M _t (i)＝M _t (i)+m _t (i)a _t

(4) The attention mechanism calculates and updates a memory vector and a memory matrix according to the latest transmission vector;

(5) Reading contents from the memory matrix by using a reading head, wherein the read contents are the latest memory characteristic content part in the t time period of the user, and returning information to the GRU unit of the controller;

(6) According to the latest memory vector content, the controller GRU calculates the user load adjustment amount through the full connection layer and outputs the result;

(7) And (5) updating the total memory matrix according to the real-time input, output and memory matrix, and circulating the steps (1) - (6) until the model training in all time periods is completed.

Images