CN112564098B - High-proportion photovoltaic power distribution network voltage prediction method based on time convolution neural network - Google Patents
High-proportion photovoltaic power distribution network voltage prediction method based on time convolution neural network Download PDFInfo
- Publication number
- CN112564098B CN112564098B CN202011387165.0A CN202011387165A CN112564098B CN 112564098 B CN112564098 B CN 112564098B CN 202011387165 A CN202011387165 A CN 202011387165A CN 112564098 B CN112564098 B CN 112564098B
- Authority
- CN
- China
- Prior art keywords
- voltage
- time
- data
- feature
- model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- H—ELECTRICITY
- H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
- H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
- H02J3/00—Circuit arrangements for ac mains or ac distribution networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/11—Complex mathematical operations for solving equations, e.g. nonlinear equations, general mathematical optimization problems
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- H—ELECTRICITY
- H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
- H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
- H02J3/00—Circuit arrangements for ac mains or ac distribution networks
- H02J3/12—Circuit arrangements for ac mains or ac distribution networks for adjusting voltage in ac networks by changing a characteristic of the network load
- H02J3/16—Circuit arrangements for ac mains or ac distribution networks for adjusting voltage in ac networks by changing a characteristic of the network load by adjustment of reactive power
-
- H—ELECTRICITY
- H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
- H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
- H02J2203/00—Indexing scheme relating to details of circuit arrangements for AC mains or AC distribution networks
- H02J2203/10—Power transmission or distribution systems management focussing at grid-level, e.g. load flow analysis, node profile computation, meshed network optimisation, active network management or spinning reserve management
-
- H—ELECTRICITY
- H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
- H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
- H02J2203/00—Indexing scheme relating to details of circuit arrangements for AC mains or AC distribution networks
- H02J2203/20—Simulating, e g planning, reliability check, modelling or computer assisted design [CAD]
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02E—REDUCTION OF GREENHOUSE GAS [GHG] EMISSIONS, RELATED TO ENERGY GENERATION, TRANSMISSION OR DISTRIBUTION
- Y02E40/00—Technologies for an efficient electrical power generation, transmission or distribution
- Y02E40/30—Reactive power compensation
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Mathematical Physics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Optimization (AREA)
- Mathematical Analysis (AREA)
- Computational Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Pure & Applied Mathematics (AREA)
- Software Systems (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Power Engineering (AREA)
- Health & Medical Sciences (AREA)
- Algebra (AREA)
- Databases & Information Systems (AREA)
- Operations Research (AREA)
- Supply And Distribution Of Alternating Current (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a high-proportion photovoltaic power distribution network voltage prediction method based on a time convolution neural network, which comprises the following steps of: step 1, carrying out data preprocessing on original load data: based on multiple time scales, performing normalization processing on the voltage time sequence data by adopting a maximum and minimum interval scaling method to obtain a complete voltage sequence; step 2, constructing an input feature vector set: performing feature screening based on an extreme gradient lifting tree algorithm of a decision tree, constructing a training sample set, outputting each feature weight, and screening out different feature subsets by combining the weight and the voltage prediction model condition; and 3, establishing a voltage prediction framework based on the photovoltaic power distribution network with high proportion, training a time convolution network prediction model, and obtaining a voltage prediction result. The method combines the extracted characteristics with time and inputs the characteristics into different channels of the time convolution neural network model to obtain a prediction result, so that the aim of remarkably improving the voltage prediction precision of the power distribution network is fulfilled.
Description
Technical Field
The invention relates to the field of high-proportion photovoltaic power distribution network voltage prediction, in particular to a high-proportion photovoltaic power distribution network voltage prediction method based on a time convolution neural network.
Background
Under the dual pressure of increasingly worsening environmental problems and shortage of traditional energy, the grid-connected capacity of new photovoltaic energy is rapidly increased, the distributed photovoltaic access proportion of users of a part of low-voltage distribution networks is higher, the problems of voltage fluctuation, out-of-limit and the like are more serious, the fluctuation and the intermittent change of environmental factors such as solar radiation, weather, temperature and the like are obvious, a lot of problems are brought to large-scale grid connection of distributed photovoltaic, the risk of out-of-limit and fluctuation of the high-proportion distributed photovoltaic distribution networks is aggravated, and the problem is difficult to solve only by means of traditional voltage regulation. In addition, the node voltage fluctuation is more obvious due to the high-proportion sudden output change of the distributed photovoltaic, and the photovoltaic property rights of users belong to users, so that the uncertainty of the operation of the power grid is further increased.
Along with the improvement of the intelligent degree of the power grid, the electric meters with the communication function are used in a large amount in the power distribution network, so that a large amount of data are accumulated in the running of the power distribution network. Considering that photovoltaic power generation has certain regularity, some potentially valuable information is hidden in the huge data volume, and the data value is not effectively utilized at present and needs to be explored by means of an artificial intelligence technology. Therefore, from the data driving perspective, the voltage variation trend is predicted, and a brand new idea is provided for solving the problem.
In conclusion, due to the fact that the historical operation data volume of the power grid is huge, and the voltage prediction belongs to the time series prediction category and has certain regularity, the potential key information of the global reactive voltage data can be deeply mined by preprocessing the big data according to the historical data formed in the power grid information operation process, and a reliable data set is provided for the research of the intelligent prediction technology of the high-proportion distributed photovoltaic distribution network voltage. The voltage prediction research is developed by an artificial intelligence deep learning theory from the data driving angle, a new power distribution network voltage prediction method with higher precision is explored, the technical problems occurring in the development process of a power distribution network and a distributed source are solved in time, and the competitiveness of the power distribution network in the aspects of operation, control, optimization and the like is improved.
However, due to the black box property of the deep learning model structure, the feature of voltage prediction needs to be screened, the complexity of the deep learning model is reduced, overfitting is prevented, and a relatively ideal prediction result can be obtained. In addition, the voltage prediction data form must be analyzed in depth to find a prediction method that can satisfy the actual condition of the voltage data.
Disclosure of Invention
In order to make up for the defects in the prior art, the invention provides the voltage prediction method of the high-proportion photovoltaic power distribution network based on the time convolution neural network, so as to achieve the purpose of improving the reactive voltage prediction result of the photovoltaic power distribution network.
The invention adopts the following technical scheme: a high-proportion photovoltaic power distribution network voltage prediction method based on a time convolution neural network comprises the following steps:
and 3, establishing a voltage prediction framework based on the photovoltaic power distribution network with a high proportion, training a time convolution neural network TCN prediction model, connecting the extracted feature vector to a neural network with a full-connection hidden layer, and outputting a voltage prediction value.
Further, in step 1, in order to eliminate the influence of the voltage and power data dimension, the raw data is subjected to non-dimensionalization processing, so that the values are placed in the (0,1) interval, and the normalization formula of the preprocessing is as follows:
in the formula (1), v * (t) is the reactive voltage data after normalization, v (t) is the voltage time series signal, v max And v min Respectively the maximum value and the minimum value of the reactive voltage signal.
Further, the step 2 specifically includes three steps:
step 21: firstly, according to data preprocessing in step 1, obtaining node voltage and net power data of a complete and dimensionless power distribution network, sampling the data through a sliding time window to construct sample characteristics, setting the length of the sliding window to be H, setting the sliding step length to be 1 time step length, and respectively obtaining node voltage characteristic vectors V according to a time sequence i And node net power feature vector P i And a corresponding label y i The following were used:
in the formula (2), t is determined along with the selection of the whole data set at a certain time point, i is a sample number, and V is i And P i Is a feature vector with dimension H, each element in the vector is historical data of previous H time points, y i Voltage data of a sample label with a value of 1 hour later;
step 22: after the discrete time variable is subjected to one-hot encoding processing, a time characteristic vector T of a prediction time point is constructed i (ii) a Wherein i is the sample number, T i Represents y i Time information of the corresponding predicted point;
step 23: node voltage and net power vector V obtained by sliding window i 、P i And the temporal feature vector T i Performing series connection to obtain the final input feature vector x of the ith sample i Comprises the following steps:
x i =[V i ,P i ,T i ](3)
feature vector x i And label y i The training sample set which jointly forms the XGboost algorithm is as follows:
in the formula (4), n is the number of training samples;
the concrete expression of the penalty function L of all tree models of the XGboost algorithm is as follows:
in the formula (5), n is the number of samples, y i For the value of the tag of the ith sample,predicting an output value for the ith sample model, wherein in the regression problem, l is generally a square error function; wherein omega (b) m ) The calculation formula of the complexity term of the tree model, namely the model regularization term, is as follows:
in the formula (6), T is the number of leaf nodes of a single tree model, w is a leaf node output vector, and gamma and lambda are parameters for controlling the weight of the regularization term and can be adjusted before algorithm training;
applying a forward distribution algorithm calculation idea, initializing a model, and performing M rounds of circular calculation, wherein a decision tree model is optimized each time, and the newly added tree model minimizes a loss function L; in the t-th round of calculation, the objective function Obj to be minimized (t) The calculation formula is as follows:
in the formula (7), b t The tree model is trained for the t-th round,the predicted output value of the additive model is obtained for the first t-1 round, C is a constant which is the sum of the complexities of the tree models obtained for the first t-1 round, i.e.Ω(b t ) Complexity of the tree model obtained for the t-th round;
calculating the Gain when the tree model divides the nodes according to the formula (7), and determining the optimal division point which enables the Gain to be maximum by performing characteristic traversal, wherein the Gain calculation formula is as follows:
in the formula (8), the first two terms respectively represent gains generated after the left sub-tree and the right sub-tree are segmented for the node, and the last term is the gain when the node is not segmented; the tree model structure can be optimized through the gain, and then a learning model enabling the loss function to be smaller is obtained.
Further, the step 3 specifically includes three steps:
constructing a voltage prediction model based on TCN, wherein the voltage prediction model comprises an input layer, an output layer and a residual error module;
step 31: constructing an input feature set of voltage prediction: the extracted feature vectors obtained from step 2: a node vector, a net power, and a time vector; in order to extract the historical data features, combining a historical voltage sequence v and a node net power feature vector sequence p with equal sliding window length into an H x 2-dimensional matrix from top to bottom, wherein H is the length of a moving time window, inputting the H into a double-channel convolution layer of the TCN, and then inputting the H into a residual error module of the TCN to perform feature extraction operation;
step 32: each residual module is used for replacing a convolutional layer, namely, the characteristic extraction work of historical data is carried out through the stacking of the residual modules; the receiving field of the TCN is determined by the network depth N, the convolution kernel size k and the expansion rate d, N residual modules of the TCN are stacked in a connection mode of input and output, each module comprises two one-dimensional expansion causal convolution layers with the same expansion rate and convolution kernel size, the expansion rate d is increased by an exponential function with the base of 2 along with the deepening of the residual modules, and the historical input features are fully extracted;
step 33: the output extracted in step 32 is a feature matrix, which is set as E, and the dimension of the feature matrix is the length of the input sequence and the number of convolution kernels; aiming at the problem of voltage point prediction, data of the last moment T of a time sequence is output, the last time point value of the sequence extracted by each convolution kernel is taken as a feature vector F dimension as the number of the convolution kernels, the feature vector is taken as a feature vector finally extracted by a TCN network, the feature vector is connected with the time feature T in series and is connected to a neural network with a fully-connected hidden layer for output, the dimension of an output layer is 1, and an output voltage predicted value y is obtained i 。
Compared with the closest prior art, the invention has the following beneficial effects:
1. according to the method, a power distribution network reactive voltage prediction model considering photovoltaic voltage regulation is constructed, so that the improvement of the quality of electric energy and the improvement of the safety and stability of the operation of a power grid are facilitated, energy conservation and loss reduction can be realized through reactive compensation and other modes, and the economical efficiency and reliability of the operation are improved;
2. on the basis of constructing a high-proportion photovoltaic prediction model, the method uses a time convolution neural network to train the model, adopts a structure of causal convolution and expansion convolution, effectively solves the problem of insufficient extraction of historical data features, has the inherent advantages of parallel computing capacity, small occupied memory, difficulty in gradient disappearance or explosion and the like, adopts a network structure connected by residual errors to deepen the network, adopts a form of expansion convolution, enables the network to feel farther historical information, improves the extraction capability of the network on deep historical information features, and further improves the voltage prediction precision.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a diagram of a simulation system architecture employed in an application example of the present invention;
FIG. 3 is a diagram of a time convolutional neural network used in the present invention;
FIG. 4 is an exploded view of the XGboost algorithm feature screening applied to the present invention;
FIG. 5 is a voltage diagram of the prediction node 16 at a time scale of 3h in an application example of the present invention;
fig. 6 is a voltage diagram of the prediction node 16 on a 6h time scale in an application example of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
Example 1
As shown in fig. 1, the flow of the high-proportion photovoltaic power distribution network voltage prediction method based on the time convolution neural network described in the present invention is shown in fig. 1, and specifically includes the following steps:
step 1: data preprocessing of raw load data
The step 1 is as follows:
historical operation data of a key node of the power distribution network in the last year are selected, the time scale of the predicted voltage is 1h, namely s is 1, and a rolling prediction mode is adopted to construct a training feature set. In order to facilitate the training of a time convolution neural network prediction model and the feature extraction, the voltage time sequence data is subjected to maximum and minimum normalization processing, so that the original data is positioned in a [0,1] interval, and the normalization processing formula is as shown in formula (1):
in the formula (1), v * (t) is the reactive voltage data after normalization processing, v (t) is a voltage time series signal, v max And v min Respectively the maximum value and the minimum value of the reactive voltage signal;
step 2: constructing a set of input feature vectors
The step 2 comprises the following three steps:
step 21: firstly, according to data preprocessing in step 1, obtaining node voltage and net power data of a complete and dimensionless power distribution network, sampling the data through a sliding time window to construct sample characteristics, setting the length of the sliding window to be H, setting the sliding step length to be 1 time step length, and respectively obtaining node voltage characteristic vectors V according to a time sequence i And node net power feature vector P i And a corresponding label y i The following:
in the formula (2), t is determined along with the selection of the whole data set at a certain time point, i is a sample number, and V is i And P i Is a feature vector with dimension H, each element in the vector is historical data of the previous H time points, y i Voltage data of a sample label with a value of 1 hour later;
step 22: after the discrete time variable is subjected to one-hot encoding processing, a time characteristic vector T of a prediction time point is constructed i (ii) a Wherein i is the sample number, T i Represents y i Time information of the corresponding predicted point;
step 23: node voltage and net power vector V obtained by sliding window i 、P i And the temporal feature vector T i Performing series connection to obtain the final input feature vector x of the ith sample i Comprises the following steps:
x i =[V i ,P i ,T i ] (3)
feature vector x i And label y i The training sample set which jointly forms the XGboost algorithm is as follows:
in the formula (4), n is the number of training samples;
the specific expression of the loss function L of all tree models of the XGboost algorithm is as follows:
in the formula (5), n is the number of samples, y i For the value of the tag of the ith sample,predicting an output value for the ith sample model, wherein in the regression problem, l is generally a square error function; wherein omega (b) m ) The calculation formula of the complexity term of the tree model, namely the model regularization term, is as follows:
in the formula (6), T is the number of leaf nodes of a single tree model, w is a leaf node output vector, and gamma and lambda are parameters for controlling the weight of the regularization term and can be adjusted before algorithm training;
applying a forward distribution algorithm calculation idea, initializing a model, and performing M rounds of circular calculation, wherein a decision tree model is optimized each time, and the newly added tree model minimizes a loss function L; in the t-th round of calculation, the objective function Obj to be minimized (t) The calculation formula is as follows:
in the formula (7), b t The tree model is trained for the t-th round,the additive model prediction output value is obtained for the first t-1 round, C is a constant whose value is the sum of the complexity of the tree models obtained for the first t-1 round, i.e.Ω(b t ) Complexity of the tree model obtained for the t-th round;
calculating the Gain when the tree model divides the nodes according to the formula (7), and determining the optimal division point which enables the Gain to be maximum by performing characteristic traversal, wherein the Gain calculation formula is as follows:
in the equation (8), the first two terms represent gains generated after the left sub-tree and the right sub-tree are segmented for the node, respectively, and the last term represents a gain when the node is not segmented. The tree model structure can be optimized through the gain, and then a learning model enabling the loss function to be smaller is obtained. Finally, obtaining a screened feature vector subset: a node vector, a net power, and a time vector;
and step 3: establishment of voltage prediction framework based on photovoltaic power distribution network with high proportion
The step 3 specifically comprises three steps:
constructing a voltage prediction model based on TCN, wherein the voltage prediction model comprises an input layer, an output layer and a residual error module;
step 31: constructing an input feature set of voltage prediction: the extracted feature vectors obtained from step 2: a node vector, a net power, and a time vector; in order to extract the historical data features, combining a historical voltage sequence v and a node net power feature vector sequence p with equal sliding window length into an H x 2-dimensional matrix from top to bottom, wherein H is the length of a moving time window, inputting the H into a double-channel convolution layer of the TCN, and then inputting the H into a residual error module of the TCN to perform feature extraction operation;
step 32: each residual module is used for replacing a convolutional layer, namely, the characteristic extraction work of historical data is carried out through the stacking of the residual modules; the receiving field of the TCN is determined by the network depth N, the convolution kernel size k and the expansion rate d, N residual modules of the TCN are stacked in a connection mode of input and output, each module comprises two one-dimensional expansion causal convolution layers with the same expansion rate and convolution kernel size, the expansion rate d is increased by an exponential function with the base of 2 along with the deepening of the residual modules, and the historical input features are fully extracted;
step 33: the output extracted in step 32 is a feature matrix, which is set as E, and the dimension of the feature matrix is the length of the input sequence and the number of convolution kernels; aiming at the problem of voltage point prediction, data of the last moment T of a time sequence is output, the last time point value of the sequence extracted by each convolution kernel is taken as a feature vector F dimension as the number of the convolution kernels, the feature vector is taken as a feature vector finally extracted by a TCN network, the feature vector is connected with the time feature T in series and is connected to a neural network with a fully-connected hidden layer for output, the dimension of an output layer is 1, and an output voltage predicted value y is obtained i 。
Example 2
1) Building network model containing high-proportion photovoltaic power distribution network
Figure 2 of the accompanying drawings is an IEEE33 node power distribution system, which comprises 3 photovoltaic power sources, wherein the nodes 5,14 and 28 are provided with photovoltaic power sources, and the capacity of the photovoltaic power sources is shown in table 1.
TABLE 1 nodal photovoltaic Power supply parameters
Mounting |
5 | 14 | 28 |
Active power/ |
25 | 16 | 45 |
Reactive power/ |
8 | 5 | 5 |
2) Historical voltage data analysis of power distribution network
The bus 1 of the system is taken as a balance node, the node voltage approximately fluctuates in the range of 220-240V, a preprocessed voltage sequence of the node 16 in about 10 days is obtained according to the formula (1), then the preprocessed data is subjected to characteristic analysis of the XGboost algorithm, and the frequency proportion of each characteristic used for splitting the decision tree, namely the characteristic weight (the sum of all weights is 1), is output, as shown in FIG. 4.
3) Voltage prediction for photovoltaic power distribution network with high proportion
And respectively placing the voltage characteristic vector v and the node net power characteristic vector p with the same sliding window length in two channels of the convolutional layer, inputting the voltage characteristic vector v and the node net power characteristic vector p into a residual error module of the TCN, fully extracting the characteristics, outputting the characteristic matrix extracted by the residual error module of the TCN, connecting the characteristic matrix with the time characteristic T in series, and connecting the characteristic matrix to a neural network with a fully-connected hidden layer for outputting to obtain a final predicted value. In order to show the prediction effect more clearly, the result is compared with the prediction effect of the LSTM, BPNN and SVM models, and the voltage variation curve of the prediction node 16 is obtained 3 hours before and 6 hours before under different time scales of different models, as shown in fig. 5 and 6.
Claims (3)
1. The high-proportion photovoltaic power distribution network voltage prediction method based on the time convolution neural network is characterized by comprising the following steps of:
step 1, performing data preprocessing on original load data: based on multiple time scales, performing normalization processing on the voltage time sequence data by adopting a maximum and minimum interval scaling method to obtain a complete voltage sequence;
step 2, constructing an input feature vector set: performing feature screening based on an XGboost algorithm of a decision tree, constructing a training sample set, outputting each feature weight, and screening out different feature subsets by combining the weight and the voltage prediction model condition;
step 3, establishing a voltage prediction framework based on the photovoltaic power distribution network with high proportion: training a time convolution neural network TCN prediction model, connecting the extracted feature vector to a neural network with a fully-connected hidden layer, and outputting a voltage prediction value;
the step 3 specifically comprises three steps:
constructing a voltage prediction model based on TCN, wherein the voltage prediction model comprises an input layer, an output layer and a residual error module;
step 31: constructing an input feature set of voltage prediction: the extracted feature vectors obtained from step 2: a node vector, a net power, and a time vector; in order to extract the historical data features, combining a historical voltage sequence v and a node net power feature vector sequence p with equal sliding window length into an H x 2-dimensional matrix from top to bottom, wherein H is the length of a moving time window, inputting the H into a double-channel convolution layer of the TCN, and then inputting the H into a residual error module of the TCN to perform feature extraction operation;
step 32: each residual module is used for replacing a convolutional layer, namely, the characteristic extraction work of historical data is carried out through the stacking of the residual modules; the receiving field of the TCN is determined by the network depth N, the convolution kernel size k and the expansion rate d, N residual modules of the TCN are stacked in a connection mode of input and output, each module comprises two one-dimensional expansion causal convolution layers with the same expansion rate and convolution kernel size, the expansion rate d is increased by an exponential function with the base of 2 along with the deepening of the residual modules, and the historical input features are fully extracted;
step 33: the output extracted in step 32 is a feature matrix, which is set as E, and the dimension of the feature matrix is the length of the input sequence and the number of convolution kernels; aiming at the problem of voltage point prediction, data of the last moment T of a time sequence is output, the last time point value of the sequence extracted by each convolution kernel is taken as a feature vector F dimension as the number of the convolution kernels, the feature vector is taken as a feature vector finally extracted by a TCN network, the feature vector is connected with the time feature T in series and is connected to a neural network with a fully-connected hidden layer for output, the dimension of an output layer is 1, and an output voltage predicted value y is obtained i 。
2. The method for predicting the voltage of the high-proportion photovoltaic power distribution network based on the time convolution neural network according to claim 1, wherein in the step 1, in order to eliminate the influence of the voltage and power data dimension, the raw data is subjected to non-dimensionalization processing, so that the values are placed in a (0,1) interval, and the normalization formula of the preprocessing is as follows:
in the formula (1), v * (t) is the reactive voltage data after normalization, v (t) is the voltage time series signal, v max And v min Respectively the maximum value and the minimum value of the reactive voltage signal.
3. The method for predicting the voltage of the high-proportion photovoltaic power distribution network based on the time convolution neural network as claimed in claim 1, wherein the step 2 specifically comprises three steps:
step 21: firstly, according to data preprocessing in step 1, obtaining node voltage and net power data of a complete and dimensionless power distribution network, sampling the data through a sliding time window to construct sample characteristics, setting the length of the sliding window to be H, setting the sliding step length to be 1 time step length, and respectively obtaining node electricity according to a time sequencePressure eigenvector V i And node net power feature vector P i And a corresponding label y i The following:
in the formula (2), t is determined along with the selection of the whole data set at a certain time point, i is a sample number, and V is i And P i Is a feature vector with dimension H, each element in the vector is historical data of previous H time points, y i Voltage data of a sample label with a value of 1 hour later;
step 22: after the independent thermal coding processing is carried out on the discrete time variables, a time characteristic vector T of a prediction time point is constructed i (ii) a Wherein i is the sample number, T i Represents y i Time information of the corresponding predicted point;
step 23: node voltage and net power vector V obtained by sliding window i 、P i And the temporal feature vector T i Performing series connection to obtain the final input feature vector x of the ith sample i Comprises the following steps:
x i =[V i ,P i ,T i ] (3)
feature vector x i And label y i The training sample set which jointly forms the XGboost algorithm is as follows:
in the formula (4), n is the number of training samples;
the specific expression of the loss function L of all tree models of the XGboost algorithm is as follows:
in the formula (5), n is the number of samples, y i For the value of the tag of the ith sample,predicting an output value for the ith sample model, wherein in the regression problem, l is generally a square error function; wherein omega (b) m ) The calculation formula of the complexity term of the tree model, namely the model regularization term, is as follows:
in the formula (6), T is the number of leaf nodes of a single tree model, w is a leaf node output vector, and gamma and lambda are parameters for controlling the weight of the regularization term and can be adjusted before algorithm training;
applying a forward distribution algorithm calculation idea, initializing a model, and performing M rounds of circular calculation, wherein a decision tree model is optimized each time, and the newly added tree model minimizes a loss function L; in the t-th round of calculation, the objective function Obj to be minimized (t) The calculation formula is as follows:
in the formula (7), b t The tree model is trained for the t-th round,the predicted output value of the additive model is obtained for the first t-1 round, C is a constant which is the sum of the complexities of the tree models obtained for the first t-1 round, i.e.Ω(b t ) Complexity of the tree model obtained for the t-th round;
calculating the Gain when the tree model divides the nodes according to the formula (7), and determining the optimal division point which enables the Gain to be maximum by performing characteristic traversal, wherein the Gain calculation formula is as follows:
in the formula (8), the first two terms respectively represent gains generated after the left sub-tree and the right sub-tree are segmented for the node, the last term is the gain when the node is not segmented, the tree model structure can be optimized through the gains, and then the learning model enabling the loss function to be smaller is obtained.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011387165.0A CN112564098B (en) | 2020-12-02 | 2020-12-02 | High-proportion photovoltaic power distribution network voltage prediction method based on time convolution neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202011387165.0A CN112564098B (en) | 2020-12-02 | 2020-12-02 | High-proportion photovoltaic power distribution network voltage prediction method based on time convolution neural network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN112564098A CN112564098A (en) | 2021-03-26 |
CN112564098B true CN112564098B (en) | 2022-08-30 |
Family
ID=75047305
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202011387165.0A Active CN112564098B (en) | 2020-12-02 | 2020-12-02 | High-proportion photovoltaic power distribution network voltage prediction method based on time convolution neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN112564098B (en) |
Families Citing this family (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112926692B (en) * | 2021-04-09 | 2023-05-09 | 四川翼飞视科技有限公司 | Target detection device, method and storage medium based on non-uniform mixed convolution |
CN112966882B (en) * | 2021-04-15 | 2023-07-28 | 四川中电启明星信息技术有限公司 | Power distribution network scheduling method based on space-time global matching |
CN113296947B (en) * | 2021-05-24 | 2023-05-23 | 中山大学 | Resource demand prediction method based on improved XGBoost model |
CN113723670B (en) * | 2021-08-12 | 2023-12-08 | 浙江大学 | Photovoltaic power generation power short-term prediction method with variable time window |
CN113420872B (en) * | 2021-08-24 | 2021-12-14 | 江苏智臻能源科技有限公司 | Fine-grained photovoltaic load prediction method based on time convolution network |
CN113705885B (en) * | 2021-08-26 | 2024-05-17 | 南京理工大学 | VMD, XGBoost and TCN fused power distribution network voltage prediction method and system |
CN113705579B (en) * | 2021-08-27 | 2024-03-15 | 河海大学 | Automatic image labeling method driven by visual saliency |
CN114169603A (en) * | 2021-12-04 | 2022-03-11 | 湖北第二师范学院 | XGboost-based regional primary school entrance academic degree prediction method and system |
CN114784883B (en) * | 2022-06-16 | 2022-10-11 | 国网江西省电力有限公司电力科学研究院 | Distributed photovoltaic grid-connected optimization method and system based on node voltage constraint |
CN117150374A (en) * | 2023-10-27 | 2023-12-01 | 国网江西省电力有限公司电力科学研究院 | Transformer fault identification method and system based on feature importance sorting |
Family Cites Families (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108896914B (en) * | 2018-05-11 | 2020-08-04 | 浙江工业大学 | Gradient lifting tree modeling and prediction method for health condition of lithium battery |
CN111367349A (en) * | 2018-12-26 | 2020-07-03 | 株洲中车时代电气股份有限公司 | Photovoltaic MPPT control method and system based on prediction model |
CN110956312B (en) * | 2019-11-15 | 2022-11-15 | 国网浙江省电力有限公司电力科学研究院 | Photovoltaic power distribution network voltage prediction method based on EMD-CNN deep neural network |
CN112016251B (en) * | 2020-09-02 | 2023-01-31 | 哈尔滨工程大学 | Nuclear power device fault diagnosis method and system |
-
2020
- 2020-12-02 CN CN202011387165.0A patent/CN112564098B/en active Active
Also Published As
Publication number | Publication date |
---|---|
CN112564098A (en) | 2021-03-26 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN112564098B (en) | High-proportion photovoltaic power distribution network voltage prediction method based on time convolution neural network | |
CN110059878B (en) | Photovoltaic power generation power prediction model based on CNN LSTM and construction method thereof | |
CN111008728A (en) | Method for predicting short-term output of distributed photovoltaic power generation system | |
CN110738344B (en) | Distributed reactive power optimization method and device for load prediction of power system | |
Guo et al. | Study on short-term photovoltaic power prediction model based on the Stacking ensemble learning | |
CN102855412A (en) | Wind electric power prediction method and device thereof | |
CN113282122A (en) | Commercial building energy consumption prediction optimization method and system | |
CN108599172B (en) | Transmission and distribution network global load flow calculation method based on artificial neural network | |
CN108183488B (en) | High-permeability distributed energy system intelligent pressure regulating method based on cluster division | |
CN109978253B (en) | Electric power system short-term load prediction method based on incremental learning | |
CN111915092A (en) | Ultra-short-term wind power prediction method based on long-time and short-time memory neural network | |
CN113746087B (en) | Power grid transient stability sample controllable generation and evaluation method and system based on CTGAN | |
CN111461921B (en) | Load modeling typical user database updating method based on machine learning | |
CN114462718A (en) | CNN-GRU wind power prediction method based on time sliding window | |
CN115907195A (en) | Photovoltaic power generation power prediction method, system, electronic device and medium | |
CN111242355A (en) | Photovoltaic probability prediction method and system based on Bayesian neural network | |
CN115496263A (en) | Short-term photovoltaic power generation prediction method based on space-time genetic attention network | |
CN113283179A (en) | Short-term load prediction method based on multi-target LSTM integrated network | |
CN117154690A (en) | Photovoltaic power generation power prediction method and system based on neural network | |
CN105404142A (en) | Aluminum electrolytic multi-parameter control method based on BP neural network and MBFO algorithm | |
CN117748495A (en) | Photovoltaic power station short-term generation power prediction method based on residual neural network | |
Aishwarya et al. | Prediction of time series data using GA-BPNN based hybrid ANN model | |
CN113962454A (en) | LSTM energy consumption prediction method based on dual feature selection and particle swarm optimization | |
CN113422371A (en) | Distributed power supply local voltage control method based on graph convolution neural network | |
CN113393119A (en) | Stepped hydropower short-term scheduling decision method based on scene reduction-deep learning |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |