CN109659933B - Electric energy quality prediction method for power distribution network with distributed power supply based on deep learning model - Google Patents

Abstract

A method for predicting the power quality of a power distribution network with distributed power supplies based on a deep learning model comprises the following steps: acquiring historical data of the power quality, wherein the historical data comprises variable data of power quality influence factors and power quality index data; preprocessing historical data of the power quality, including data normalization, time sequence conversion, training data and evaluation data segmentation; the method comprises the following steps of memorizing a neural network at long time of electric energy quality, namely determining an LSTM prediction model, constructing the LSTM prediction model, initializing parameters, and determining the LSTM prediction model based on training data; a prediction model performance evaluation based on the evaluation data; and predicting the power quality index of the system in the future period. The invention has the advantages that: 1. the electric energy quality of a power distribution network containing DGs is effectively predicted; 2. higher accuracy can be obtained in processing the time series prediction problem; 3. the influence of each characteristic input variable in the prediction model is fully considered.

Description

Electric energy quality prediction method for power distribution network with distributed power supply based on deep learning model

Technical Field

The invention relates to a method for predicting the power quality of a power distribution network containing a distributed power supply based on a deep learning model, belonging to the field of electrical engineering and power quality.

Background

The electric energy as a product to be brought to the market should also be requested for quality as other commodities. The indexes reflecting the quality condition of the electric energy mainly comprise voltage deviation, harmonic waves, voltage flicker and fluctuation, voltage dip and interruption, frequency deviation, harmonic waves and inter-harmonic waves, three-phase voltage unbalance and the like. With the development of green energy and smart grid technology, it is a future trend that Distributed Generation (DG) such as photovoltaic power generation and wind power generation is widely connected to a power distribution network, and the Distributed Generation (DG) is a beneficial supplement of a traditional power distribution network. However, the problem of power quality in the distribution network containing the DGs is more serious due to the large fluctuation, randomness and multiple control modes. If the electric energy quality situation of the DG-containing power distribution network in the future period can be accurately predicted, the system can perform effective early warning, prevention and improvement schemes according to the electric energy quality situation perception result, and the safety and the reliability of the system are correspondingly improved.

With the rapid development of deep learning technology in recent years, the potential of the application of the technology in the prediction field is also widely concerned by the wide researchers. A Long Short-Term Memory neural network (LSTM) is an improved time cycle neural network and belongs to the category of deep learning models. Because the long-term and short-term dependence information of the memory time sequence is the default behavior of the LSTM, the LSTM is suitable for processing and predicting interval and delay events in the time sequence, the time correlation of time sequence data is fully considered, and the prediction performance is outstanding. The deep learning technology based on the LSTM model is applied to power quality prediction, and effective basis can be provided for subsequent power quality early warning, treatment and decision improvement.

Currently, research results for predicting the power quality of a distribution network containing DGs are few. The patent with the application number of 201310122440.X provides a wind power quality trend prediction method, a Monte Carlo simulation algorithm is adopted to predict the power quality trend, but the optimal solution of the Monte Carlo simulation algorithm has a close relation with the data sampling amount, and the prediction effect is difficult to achieve the optimal effect when the sampling amount does not meet certain requirements; the patent with the application number of 201810644850.3 provides a power quality prediction method for a power distribution network with a distributed power supply based on a support vector machine, but the support vector machine cannot well consider the correlation of a training set in time, so that the prediction accuracy is greatly limited; a small amount of research applies LSTM to power system short-term load prediction and grid disturbance prediction: for example, patent with application number CN201810054530.2 proposes a depth LSTM-based power grid load prediction method; the patent with the application number of CN201710101569.0 provides a power grid disturbance prediction method and device based on an LSTM # RTRBM deep learning model; while more research and application on LSTM prediction is mainly applied in financial terms: for example, patent with application number CN201711395406.4 proposes a stock forecasting method based on ARMA-LSTM model; the patent with application number CN201810169957.7 proposes a transaction index abnormality monitoring method based on a deep learning model LSTM.

Disclosure of Invention

The invention provides a DG-containing power distribution network power quality prediction method based on an LSTM deep learning model, aiming at overcoming the problem that the potential power quality problem of a system cannot be effectively perceived, pre-warned and processed due to the insufficient power quality prediction capability of the existing DG-containing power distribution network.

The method comprehensively considers the factors of increasing influence of the electric energy quality of the DG-containing power distribution network system such as the climate environment, the system load and the like, and reasonably predicts the electric energy quality index of the system in the future period by utilizing the good time series memory capacity based on the LSTM deep learning model.

The invention discloses a DG-containing power distribution network electric energy quality prediction method based on a long-time and short-time memory neural network, namely an LSTM deep learning model, as shown in the attached figure 1, the process comprises the following steps:

1. acquiring historical data of electric energy quality: the method comprises the steps that N groups of input variable value data vectors and output variable value data vectors in a 24-hour time section are obtained by reading an electric energy quality historical data set X of a DG-containing target power distribution network, wherein the electric energy quality historical data set X comprises historical data values of electric energy quality index items and historical data values of influence factors of the electric energy quality index items in corresponding time periods, and the electric energy quality historical data set X is used for training and evaluating a system electric energy quality LSTM deep learning prediction model; the input variable values comprise time values, temperature values, illumination values and system electrical load values, and the output variable values comprise power quality index items such as voltage deviation, frequency deviation, three-phase unbalance and harmonic distortion; grouping and pairing the input variable value data and the output variable value data vectors by time marks;

2. preprocessing historical data of power quality: preprocessing the historical data set X of the system power quality obtained in the step 1, wherein the preprocessing comprises data normalization, time series conversion and data segmentation;

step 201, data normalization: because the numerical difference between the input variable items influencing the system power quality is large and has different dimensions, in order to ensure that all variable items can equally act on the system power quality prediction, the input and output ranges of the nonlinear activation function in the prediction model are considered, and the prediction neuron is prevented from falling into a saturation state, the system power quality historical data acquired in the step 1 needs to be subjected to normalization processing;

aiming at index item values of temperature, illumination, system power load and power quality, selecting the maximum value and the minimum value of each variable item in all historical data vector groups obtained in the step 1, and normalizing the maximum value and the minimum value to a [0,1] numerical interval through a formula (1):

in the formula: x is any variable data value before normalization processing, x_minIs the minimum value of x homogeneous variable terms in all historical data, x_maxThe value is the maximum value in all historical data of the x homogeneous variable items, and x' is the normalized data value corresponding to the x value;

after the processing is finished, obtaining a system power quality normalization historical data set Xc;

step 202, time series conversion: in the LSTM deep learning algorithm, common time sequence data cannot meet the requirements of supervised prediction learning; in the process of realizing a multivariable prediction target based on supervised learning, extracting time series values of all variables and time series values of a prediction target in a data set with a determined time length to form a supervised learning data frame; after the processing is finished, the system power quality normalization historical data set Xc is converted into a supervised learning historical data set S_h；

Step 203, data segmentation: in order to determine internal proper parameters of the LSTM-based deep learning electric energy quality prediction model, proper amount of historical data needs to be provided for prediction model training; in order to reasonably evaluate the effectiveness of the determined prediction model, proper amount of historical data needs to be provided for performance evaluation of the prediction model;

the normalized and time-series-converted supervised learning history data set S acquired in step 202_hPerforming segmentation into training data subsets S_tAnd test data subset S_c(ii) a In order to enable the prediction model to be fully learned and the performance of the prediction model to be accurately evaluated, the training data subset S is divided into a plurality of training data subsets_tAnd test data subset S_cNeeds to be reasonably distributed to respectively occupy the supervised learning historical data set S _h70% and 30% of the complete data contained therein;

3. determining a power quality LSTM prediction model: constructing an LSTM power quality prediction model framework according to the determined input and output variables, and based on a training data subset S_tPerforming supervised learning and determining internal parameters of the LSTM prediction model, including structural parameters and training parameters; the structural parameters comprise the number of model unit cells and the internal structural parameters of the unit cells, and the training parameters comprise the number Z of training cycles and the batch size P of batch training;

step 301, constructing an electric energy quality LSTM prediction model: constructing a system electric energy quality LSTM prediction model framework by taking a time value, a temperature value, an illumination value and a system electric load value as input variables and taking voltage deviation, frequency deviation, three-phase unbalance and harmonic distortion rate as output variables;

the LSTM prediction model consists of a plurality of LSTM unit cells; the internal structure of each unit cell is: for the LSTM unit cell at time t, the input x is given_tAnd the last unit cell state quantity C_t-1The information will go through three processes in the transmission process, and then its output h is obtained_t；

Process 1 is a forgetting process: this part determines how much information of the last LSTM unit cell is retained, which implements the mechanism and outputs f_tThe following equation (2) yields:

where σ is a Sigmoid function, i.e., an S-type function, for compressing the input continuous real values to a numerical range [0, 1%]To (c) to (d);

is input of [ h_t-1,x_t]Relative to f_tThe weight coefficient matrix and the bias term of (1); wherein h is_t-1For the output of the last LSTM unit cell, [ h ]_t-1,x_t]Is formed by_t-1And x_tInputting the combined LSTM unit cells;

process 2 is an input process: this part determines the actual input of the current LSTM unit cells, the mechanism of realization being characterized jointly by equations (3), (4) and (5); wherein i_tRepresenting the output of Sigmoid function for controlling C_tThe information adding amount is large; c_tIndicating the state information of the present LSTM unit cell,

represents the new information input:

in the formula (I), the compound is shown in the specification,

and

respectively represent [ h_t-1,x_t]Relative to i_tAnd

the weight coefficient matrix and the bias term of (1); tanh is a hyperbolic tangent activation function for compressing an input quantity to a numerical range of [ -1,1]To (c) to (d);

process 3 is an output process: this section determines the output of the current LSTM unit cell, the mechanism of realization being characterized by equations (6) and (7) together; wherein o is_tFor controlling the extent to which the cell state is filtered:

h_t＝o_t*tanh(C_t) (7)

in the formula (I), the compound is shown in the specification,

respectively represent [ h_t-1,x_t]Relative to o_tThe weight coefficient matrix and the bias term of (1);

step 302, initializing the structure parameters of the prediction model: according to the requirements of a supervised learning algorithm, the determination requirement of the number of LSTM unit cells is greater than the characteristic number of the prediction model variable; aiming at the structural parameters of each LSTM cell unit, performing parameter initialization on a weight coefficient matrix in deep learning by adopting a 0.02 × randn (num _ params) method, wherein randn () represents a standard normal distribution function, and num _ params represents the number of parameters; for the bias matrix, all the bias matrix are set to 0 for initialization;

step 303, training the data subset S_tReading: on the basis of completing LSTM electric energy quality prediction model framework construction and prediction model structure parameter initialization, the electric energy quality LSTM prediction model is determined, and sufficient training based on effective training data must be carried out on a prediction network; therefore, the training obtained in step 203 is readTraining data subset S_tAs training data for the LSTM prediction model;

step 304, setting of training cycle number Z and batch size P: in the LSTM prediction model training, too long training period number can cause the prediction model to be over-fitted, and too short training period number can cause the prediction model not to well reflect effective information contained in a training data set, so that the training data subset S is required to be used_tReasonably setting the training period number Z according to the data quantity; the training speed of LSTM supervised learning can be accelerated by adopting batch training, but the overlarge batch size in the batch training can occupy overlarge computer memory and reduce prediction precision, so that the method also needs to be based on S_tReasonably setting the batch size P according to the data volume in the step (1);

step 305, determining a power quality LSTM prediction model: the LSTM prediction model frame constructed in the steps 301 to 304, the set model structure parameters, the training parameters and the read training data subset S_tThe training process of the power quality LSTM prediction model can be implemented:

the first step is as follows: calculating the output value of each LSTM unit cell according to a forward propagation calculation method;

the second step is that: inversely calculating an error term for each LSTM unit cell, the error term being related not only to the output process described in step 301 but also to the hidden layer at the previous and subsequent time points;

the third step: calculating a gradient of each weight according to the corresponding error term;

the fourth step: according to the corresponding error items and the gradient of each weight, updating the weight removal value by applying a gradient optimization algorithm;

after the iterative training process is completed, all the parameter values involved in the structure of the LSTM prediction model are recorded and retained, including [ h ] described in step 301_t-1,x_t]And the bias matrix parameters

4. And (3) performance evaluation of a prediction model: according to the electric energy quality LSTM prediction model determined in the step 3, before the electric energy quality LSTM prediction model is really put into an actual DG-containing distribution network system for electric energy quality prediction, effect test and performance evaluation of the prediction model are carried out;

step 401, testing the data subset S_cReading: the performance evaluation of the power quality LSTM prediction model must be based on sufficient tests based on historical data, and therefore, the test data subset S obtained in step 203 needs to be read_cAs performance evaluation data of the LSTM prediction model; subset of test data S_cThe historical data in (1) fall into two categories: the system power quality influence factor data is used as input variable type, and the power quality index item data is used as output data type;

step 402, based on the test data subset S_cThe electric energy quality prediction: performing an electric energy quality LSTM prediction process based on deep learning according to the electric energy quality LSTM prediction model determined in the step 3 and the system electric energy quality influence factor data read in the step 401 as the input variable category to obtain a corresponding system electric energy quality prediction value;

step 403, predicting model performance evaluation: comparing the predicted power quality value obtained in step 402 and output as LSTM prediction model, and the test data subset S read in step 401_cThe electric energy quality index item data is used as the output data type to realize the prediction performance evaluation of the obtained LSTM prediction model;

in order to effectively evaluate the prediction accuracy of the LSTM prediction model of the power quality, the root mean square error index epsilon of the LSTM prediction model is considered through a formula (8)_RMSE：

In the formula: y is_jFor testing a subset S of data_cThe original actual historical index value of the power quality,

n is a subset S of the test data for the predicted power quality value obtained in step 402_cThe number of data sets used for the evaluation test; j is the data group number for testing;

ε_RMSEreflects the average deviation degree of the predicted value relative to the observed value, and the value of the deviation degree is related to n; however, when the data quantity n is determined, epsilon_RMSEThe smaller the value of (A), the higher the prediction accuracy of the reflection prediction model is, and the better the prediction performance is;

5. predicting the power quality index of the system in the future period: based on the test evaluation in the step 4, an LSTM prediction model with good performance is confirmed, and reasonable prediction of the electric energy quality index item value of the system in the future period can be carried out on the premise that various input variable values influencing the electric energy quality of the system in the future period are available;

step 501, inputting variable data set reading in future time period: acquiring and reading an input variable data set of a DG-containing target power distribution network system in a certain period in the future, wherein the input variable data set comprises a temperature value, an illumination value, a system power load value and a time value;

step 502, predicting the electric energy quality in the future period: based on the input variable data set read in step 501 in a certain future period and the LSTM prediction model determined in step 3, performing power quality prediction on a DG-containing target power distribution network to obtain prediction data of a power quality index item of the system in the corresponding future period;

step 503, inverse normalization processing of the prediction data: the electric energy quality index item prediction data obtained in step 502 belongs to normalized data, and in order to obtain system electric energy quality prediction data with actual physical significance, inverse normalization processing needs to be performed according to a formula (9):

x＝x′*(x_max-x_min)+x_min (9)。

the invention has the following beneficial effects: 1. the electric energy quality of the distribution network containing the DGs is effectively predicted by utilizing the long-time memory neural network LSTM regression function; 2. the long-time memory neural network LSTM can extract the relevance between data on a longer time scale, and can obtain higher precision in processing time sequence prediction problems; 3. the influence of each characteristic input variable in the prediction model is fully considered by using the LSTM supervised learning method.

Drawings

FIG. 1 is a flow chart of an embodiment of the method of the present invention.

Fig. 2 is a topology diagram of a 13-node power distribution network with DG access.

Fig. 3a to 3b are schematic diagrams of power quality prediction based on an LSTM deep learning model, where fig. 3a is an internal structure diagram of an LSTM unit, and fig. 3b is an overall structure diagram of the LSTM power quality prediction model.

Fig. 4 is a graph comparing the predicted value and the actual value of the voltage index.

FIG. 5 is a graph of relative error values for predicted values of voltage indicators.

Detailed Description

The present invention will be described in further detail with reference to examples and drawings, but the present invention is not limited to these examples. In the embodiment, a general block diagram of a method for predicting the power quality of a power distribution network with distributed generation based on a deep learning model is shown in fig. 1, and the method comprises the following steps:

1. acquiring historical data of electric energy quality: the method comprises the steps that N groups of input variable value data vectors and output variable value data vectors in a 24-hour time section are obtained by reading an electric energy quality historical data set X of a DG-containing target power distribution network, wherein the electric energy quality historical data set X comprises historical data values of electric energy quality index items and historical data values of influence factors of the electric energy quality index items in corresponding time periods, and the electric energy quality historical data set X is used for training and evaluating a system electric energy quality LSTM deep learning prediction model; the input variable values comprise time values, temperature values, illumination values and system electrical load values, and the output variable values comprise power quality index items such as voltage deviation, frequency deviation, three-phase unbalance and harmonic distortion; grouping and pairing the input variable value data and the output variable value data vectors by time marks;

taking a 13-node 10.5kV DG-containing power distribution network with a topological structure as shown in figure 2 as an example for simulation, the power quality scene change of a target power distribution network is realized by changing the temperature, the illumination intensity and the system power load, and the power quality scene change is obtained92 groups of data vectors of 2208 groups of data vectors containing 8 variable data of time values, temperature values, illumination values, system electrical load values, voltage deviation values, frequency deviation values, three-phase unbalance degrees and harmonic distortion rates in a 24-hour time section form a system power quality historical data set matrix X_(2208×8)；

2. Preprocessing historical data of power quality: preprocessing the historical data set X of the system power quality obtained in the step 1, wherein the preprocessing comprises data normalization, time series conversion and data segmentation;

step 201, data normalization: because the numerical difference between the input variable items influencing the system power quality is large and has different dimensions, in order to ensure that all variable items can equally act on the system power quality prediction, the input and output ranges of the nonlinear activation function in the prediction model are considered, and the prediction neuron is prevented from falling into a saturation state, the system power quality historical data acquired in the step 1 needs to be subjected to normalization processing;

selecting the maximum value and the minimum value of each variable item in all the historical data vector groups obtained in the step 1 aiming at each index item value of temperature, illumination, system power load and power quality, and normalizing the maximum value and the minimum value to a numerical value interval of [0,1] through a formula (1);

after the processing is finished, obtaining a system power quality normalization historical data set Xc;

in an embodiment, the system power quality historical data set matrix X obtained in step 1 is used_(2208×8)Respectively carrying out normalization processing on each characteristic value column according to a formula (1) to obtain a system power quality normalization historical data set Xc; in the specific embodiment, taking the voltage deviation value data in the 1 st group 24-hour time segment in 92 groups of data as an example, the data comparison before and after the normalization processing is given; taking a 'voltage deviation' index item in a plurality of index items of the electric energy quality as an example, the numerical values before and after normalization are compared, and the numerical values are shown in table 1:

TABLE 1 group 1 Voltage offset value data normalization before and after comparison

Time period	Before normalization (kV)	After normalization processing
			1	0.045	0.256917
2	0.0499	0.353755
			3	0.0381	0.120553
4	0.052	0.395257
			5	0.0395	0.148221
6	0.0485	0.326087
			7	0.0483	0.322134
8	0.0486	0.328063
			9	0.0617	0.586957
10	0.0571	0.496047
			11	0.0713	0.77668
12	0.0641	0.634387
			13	0.07	0.750988
14	0.0563	0.480237
			15	0.0697	0.745059
16	0.0702	0.754941
			17	0.0745	0.839921
18	0.0723	0.796443
			19	0.0768	0.885375
20	0.0792	0.932806
			21	0.0643	0.63834
22	0.0664	0.679842
			23	0.0787	0.922925
24	0.0639	0.630435

Step 202, time series conversion: in the LSTM deep learning algorithm, common time sequence data cannot meet the requirements of supervised prediction learning; in the process of realizing a multivariable prediction target based on supervised learning, extracting time series values of all variables and time series values of a prediction target in a data set with a determined time length to form a supervised learning data frame; after the processing is finished, the power quality of the system is normalized by the historical numberThe data set Xc is converted into a supervised learning history data set S_h；

In the embodiment, a packaged series _ to _ super () function in a time series processing tool in LSTM deep learning is called to convert the power quality normalized historical data set Xc into a supervised learning historical data set S_h(ii) a The function of the series _ to _ superimposed () function is to rearrange the data of the Xc into a data form with multivariable input corresponding to one or more electric energy quality index item variables output;

step 203, data segmentation: in order to determine internal proper parameters of the LSTM-based deep learning electric energy quality prediction model, proper amount of historical data needs to be provided for prediction model training; in order to reasonably evaluate the effectiveness of the determined prediction model, proper amount of historical data needs to be provided for performance evaluation of the prediction model;

the normalized and time-series-converted supervised learning history data set S acquired in step 202_hPerforming segmentation into training data subsets S_tAnd test data subset S_c(ii) a In order to enable the prediction model to be fully learned and the performance of the prediction model to be accurately evaluated, the training data subset S is divided into a plurality of training data subsets_tAnd test data subset S_cNeeds to be reasonably distributed to respectively occupy the supervised learning historical data set S _h70% and 30% of the complete data contained therein;

in the examples, according to S_t：S_h＝0.7：1，S_c：S_h0.3: 1, distributing training data and test data; thus, the training data subset S_tIncluding 0.7 × 92 × 24 hours of data, test data subset S_cThe data of 672 hours is contained in the table, and the table is 0.3 multiplied by 92 multiplied by 24;

3. determining a power quality LSTM prediction model: constructing an LSTM power quality prediction model framework according to the determined input and output variables, and based on a training data subset S_tPerforming supervised learning and determining internal parameters of the LSTM prediction model, including structural parameters and training parameters; wherein the structural parameters includeThe number of unit cells of the model and the internal structure parameters of the unit cells, wherein the training parameters comprise a training period number Z and a batch size P according to batch training;

step 301, constructing an electric energy quality LSTM prediction model: constructing a system electric energy quality LSTM prediction model framework by taking a time value, a temperature value, an illumination value and a system electric load value as input variables and taking voltage deviation, frequency deviation, three-phase unbalance and harmonic distortion rate as output variables;

the LSTM prediction model consists of a plurality of LSTM unit cells; the internal structure of each unit cell is: for the t-th LSTM unit cell, the input x is given_tAnd the last unit cell state quantity C_t-1The information will go through three processes in the transmission process, and then its output h is obtained_t；

Process 1 is a forgetting process: this part determines how much information of the last LSTM unit cell is retained, which implements the mechanism and outputs f_tThe formula (2) is used for obtaining;

process 2 is an input process: this part determines the actual input of the current LSTM unit cells, the mechanism of realization being characterized jointly by equations (3), (4) and (5);

process 3 is an output process: this section determines the output of the current LSTM unit cell, the mechanism of realization being characterized by equations (6) and (7) together;

in an example, as shown in FIG. 3, FIG. 3a shows the internal structure of the LSTM unit cell and the internal correlation between the LSTM unit cells at the time; FIG. 3b shows the working principle of the LSTM-based power quality prediction model, where { x₁，x₂，…，x_tCharacterization of the input of LSTM unit cells at different times, { A₁，A₂，…，A_mCharacterization of the respective Unit cells of LSTM, { h_t，…，h_mCharacterizing the output of the LSTM unit cell;

step 302, initializing the structure parameters of the prediction model: according to the requirements of a supervised learning algorithm, the selection requirement of the number of LSTM unit cells is greater than the characteristic number of the prediction model variable; aiming at the structural parameters of each LSTM cell unit, performing parameter initialization on a weight coefficient matrix in deep learning by adopting a 0.02 × randn (num _ params) method, wherein randn () represents a standard normal distribution function, and num _ params represents the number of parameters; for the bias matrix, all the bias matrix are set to 0 for initialization;

in an embodiment, the initial weighting coefficient matrix is compared

The method of 0.02 × randm (num _ params) is adopted for deep learning parameter initialization, and the bias matrix is used

All of the internal elements of (a) are set to 0;

step 303, training the data subset S_tReading: on the basis of completing LSTM electric energy quality prediction model framework construction and prediction model structure parameter initialization, the electric energy quality LSTM prediction model is determined, and sufficient training based on effective training data must be carried out on a prediction network; therefore, the training data subset S obtained in step 203 is read_tAs training data for the LSTM prediction model;

step 304, setting of training cycle number Z and batch size P: in the LSTM prediction model training, too long training period number can cause the prediction model to be over-fitted, and too short training period number can cause the prediction model not to well reflect effective information contained in a training data set, so that the training data subset S is required to be used_tReasonably setting the training period number Z according to the data quantity; the training speed of LSTM supervised learning can be accelerated by adopting batch training, but the overlarge batch size in the batch training can occupy overlarge computer memory and reduce prediction precision, so that the method also needs to be based on S_tReasonably setting the batch size P according to the data volume in the step (1);

in an embodiment, the subset S is based on training data_tReasonably selecting the training period number Z as 500 and the batch size P as 64;

step 305, determining a power quality LSTM prediction model: the LSTM prediction model frame constructed according to the steps 301 to 304, the set model structure parameters, the training parameters and the read trainingTraining data subset S_tThe training process of the power quality LSTM prediction model can be implemented:

the first step is as follows: calculating the output value of each LSTM unit cell according to a forward propagation calculation method;

the second step is that: inversely calculating an error term for each LSTM unit cell, the error term being related not only to the output process described in step 301 but also to the hidden layer at the previous and subsequent time points;

the third step: calculating a gradient of each weight according to the corresponding error term;

the fourth step: according to the corresponding error items and the gradient of each weight, updating the weight removal value by applying a gradient optimization algorithm;

after the iterative training process is completed, all the parameter values involved in the structure of the LSTM prediction model are recorded and retained, including [ h ] described in step 301_t-1,x_t]And the bias matrix parameters

4. And (3) performance evaluation of a prediction model: according to the electric energy quality LSTM prediction model determined in the step 3, before the electric energy quality LSTM prediction model is really put into an actual DG-containing distribution network system for electric energy quality prediction, effect test and performance evaluation of the prediction model are carried out;

step 401, testing the data subset S_cReading: the performance evaluation of the power quality LSTM prediction model must be based on sufficient tests based on historical data, and therefore, the test data subset S obtained in step 203 needs to be read_cAs performance evaluation data of the LSTM prediction model; subset of test data S_cThe historical data in (1) fall into two categories: the system power quality influence factor data is used as input variable type, and the power quality index item data is used as output data type;

step 402, based on the test data subset S_cQuality of electric energy ofAnd (3) prediction: performing an electric energy quality LSTM prediction process based on deep learning according to the electric energy quality LSTM prediction model determined in the step 3 and the system electric energy quality influence factor data read in the step 401 as the input variable category to obtain a corresponding system electric energy quality prediction value;

step 403, predicting model performance evaluation: comparing the predicted power quality value obtained in step 402 and output as LSTM prediction model, and the test data subset S read in step 401_cThe electric energy quality index item data is used as the output data type to realize the prediction performance evaluation of the obtained LSTM prediction model;

in order to effectively evaluate the prediction accuracy of the LSTM prediction model of the power quality, the root mean square error index epsilon of the LSTM prediction model is considered through a formula (8)_RMSE；

ε_RMSEReflects the average deviation degree of the predicted value relative to the observed value, and the value of the deviation degree is related to n; however, when the data quantity n is determined, epsilon_RMSEThe smaller the value of (A), the higher the prediction accuracy of the reflection prediction model is, and the better the prediction performance is;

in an embodiment, the subset of tested data S_cThe method comprises the evaluation test of 672 hours of data to obtain the root mean square error index epsilon_RMSEIs 5.6; the value is less than [10,15 ]]This epsilon_RMSEThe general value range shows that the electric energy quality LSTM prediction model obtained in the step 3 has good prediction effect and high accuracy;

5. predicting the power quality index of the system in the future period: based on the test evaluation in the step 4, an LSTM prediction model with good performance is confirmed, and reasonable prediction of the electric energy quality index item value of the system in the future period can be carried out on the premise that various input variable values influencing the electric energy quality of the system in the future period are available;

step 501, inputting variable data set reading in future time period: acquiring and reading an input variable data set of a DG-containing target power distribution network system in a certain period in the future, wherein the input variable data set comprises a temperature value, an illumination value, a system power load value and a time value;

in an embodiment, a history-divided data set S is obtained and read_hIn the 24-hour time zone of the next dayThe input variable data set comprises a temperature value, an illumination intensity value, a system electrical load value and a time value; specific numerical values are shown in table 2;

step 502, predicting the electric energy quality in the future period: based on the input variable data set read in step 501 in a certain future period and the LSTM prediction model determined in step 3, performing power quality prediction on a DG-containing target power distribution network to obtain prediction data of a power quality index item of the system in the corresponding future period;

step 503, inverse normalization processing of the prediction data: the electric energy quality index item prediction data obtained in the step 502 belongs to normalized data, and in order to obtain system electric energy quality prediction data with actual physical significance, inverse normalization processing needs to be carried out according to a formula (9);

in an embodiment, according to the prediction processes of step 502 and step 503, taking the voltage deviation index item in the power quality as an example, the voltage deviation index item prediction data in a 24-hour time section in the future of the system corresponding to the input variable data set in the prediction model is obtained; the ratio of the predicted value to the actual value of the voltage deviation index after inverse normalization in a 24-hour time section in the future is shown in table 2:

TABLE 2 comparison of predicted and actual voltage deviation values over a 24 hour time period in the future day

Fig. 4 shows that the predicted value and the actual value of the voltage deviation within 24 hours of the future day of the distribution network system model containing the DG in the embodiment are compared numerically; wherein the solid line represents actual value data of the system voltage deviation, and the dotted line represents predicted value data of the system voltage deviation; it can be seen that, the system voltage deviation index item prediction data can well track the actual number of the system except that the actual value tracked at the end point has obvious fluctuationAccordingly, the prediction effect is better; FIG. 5 shows the distribution of relative error values between the predicted values and the actual values of the voltage deviation indicator items of the system in 24 hours a day, which represents the deviation degree of the predicted values from the actual values, and the root mean square error indicator epsilon is calculated_RMSEThe prediction accuracy is higher for 5.

The embodiment analyzes and shows that the method provided by the invention can realize the electric energy quality prediction of the DG-containing power distribution network, controls the prediction error in a smaller range, provides effective basis for further electric energy quality control and decision improvement of the DG-containing target power distribution network system, and further discovers and solves the potential hidden danger of the electric energy quality as early as possible.

The embodiments described in this specification are merely illustrative of implementations of the inventive concept and the scope of the present invention should not be considered limited to the specific forms set forth in the embodiments but rather by the equivalents thereof as may occur to those skilled in the art upon consideration of the present inventive concept.

Claims (1)

1. A method for predicting the power quality of a power distribution network with distributed power supplies based on a deep learning model comprises the following steps:

step 1, acquiring historical data of power quality: the method comprises the steps that an electric energy quality historical data set X of a DG-containing target power distribution network is read, the electric energy quality historical data set X comprises electric energy quality index item historical data values and historical data values of influence factors of the electric energy quality index item historical data values in corresponding time periods, N groups of input variable value data vectors and output variable value data vectors in each group in a 24-hour time section are obtained, and the N groups of input variable value data vectors and the output variable value data vectors are used for training and evaluating a system electric energy quality LSTM deep learning prediction model; the input variable values comprise time values, temperature values, illumination values and system electrical load values, and the output variable values comprise power quality index items such as voltage deviation, frequency deviation, three-phase unbalance and harmonic distortion; grouping and pairing the input variable value data and the output variable value data vectors by time marks;

step 2, preprocessing historical data of power quality: preprocessing the historical data set X of the system power quality obtained in the step 1, wherein the preprocessing comprises data normalization, time series conversion and data segmentation;

step 201, data normalization: because the numerical difference between the input variable items influencing the system power quality is large and has different dimensions, in order to ensure that all variable items can equally act on the system power quality prediction, the input and output ranges of the nonlinear activation function in the prediction model are considered, and the prediction neuron is prevented from falling into a saturation state, the system power quality historical data acquired in the step 1 needs to be subjected to normalization processing;

aiming at index item values of temperature, illumination, system power load and power quality, selecting the maximum value and the minimum value of each variable item in all historical data vector groups obtained in the step 1, and normalizing the maximum value and the minimum value to a [0,1] numerical interval through a formula (1):

in the formula: x is any variable data value before normalization processing, x_minIs the minimum value of x homogeneous variable terms in all historical data, x_maxThe value is the maximum value in all historical data of the x homogeneous variable items, and x' is the normalized data value corresponding to the x value;

after the processing is finished, obtaining a system power quality normalization historical data set Xc;

step 202, time series conversion: in the LSTM deep learning algorithm, common time sequence data cannot meet the requirements of supervised prediction learning; in the process of realizing a multivariable prediction target based on supervised learning, extracting time series values of all variables and time series values of a prediction target in a data set with a determined time length to form a supervised learning data frame; after the processing is finished, the system power quality normalization historical data set Xc is converted into a supervised learning historical data set S_h；

Step 203, data segmentation: in order to determine internal proper parameters of the LSTM-based deep learning electric energy quality prediction model, proper amount of historical data needs to be provided for prediction model training; in order to reasonably evaluate the effectiveness of the determined prediction model, proper amount of historical data needs to be provided for performance evaluation of the prediction model;

the normalized and time-series-converted supervised learning history data set S acquired in step 202_hPerforming segmentation into training data subsets S_tAnd test data subset S_c(ii) a In order to enable the prediction model to be fully learned and the performance of the prediction model to be accurately evaluated, the training data subset S is divided into a plurality of training data subsets_tAnd test data subset S_cNeeds to be reasonably distributed to respectively occupy the supervised learning historical data set S_h70% and 30% of the complete data contained therein;

step 3, determining an electric energy quality LSTM prediction model: constructing an LSTM power quality prediction model framework according to the determined input and output variables, and based on a training data subset S_tPerforming supervised learning and determining internal parameters of the LSTM prediction model, including structural parameters and training parameters; the structural parameters comprise the number of model unit cells and the internal structural parameters of the unit cells, and the training parameters comprise the number Z of training cycles and the batch size P of batch training;

step 301, constructing an electric energy quality LSTM prediction model: constructing a system electric energy quality LSTM prediction model framework by taking a time value, a temperature value, an illumination value and a system electric load value as input variables and taking voltage deviation, frequency deviation, three-phase unbalance and harmonic distortion rate as output variables;

the LSTM prediction model consists of a plurality of LSTM unit cells; the internal structure of each unit cell is: for the LSTM unit cell at time t, the input x is given_tAnd the last unit cell state quantity C_t-1The information will go through three processes in the transmission process, and then its output h is obtained_t；

Process 1 is a forgetting process: this part determines how much information of the last LSTM unit cell is retained, which implements the mechanism and outputs f_tThe following equation (2) yields:

where σ is a Sigmoid function, i.e., an S-type function, for compressing the input continuous real values to a numerical range [0, 1%]To (c) to (d);

is input of [ h_t-1,x_t]Relative to f_tThe weight coefficient matrix and the bias term of (1); wherein h is_t-1For the output of the last LSTM unit cell, [ h ]_t-1,x_t]Is formed by_t-1And x_tInputting the combined LSTM unit cells;

process 2 is an input process: this part determines the actual input of the current LSTM unit cells, the mechanism of realization being characterized jointly by equations (3), (4) and (5); wherein i_tRepresenting the output of Sigmoid function for controlling C_tThe information adding amount is large; c_tIndicating the state information of the present LSTM unit cell,

represents the new information input:

in the formula (I), the compound is shown in the specification,

and

respectively represent [ h_t-1,x_t]Relative to i_tAnd

the weight coefficient matrix and the bias term of (1); tanh is a hyperbolic tangent activation function for compressing an input quantity to a numerical range of [ -1,1]To (c) to (d);

process 3 is an output process: this section determines the output of the current LSTM unit cell, the mechanism of realization being characterized by equations (6) and (7) together; wherein o is_tFor controlling the extent to which the cell state is filtered:

h_t＝o_t*tanh(C_t) (7)

in the formula (I), the compound is shown in the specification,

respectively represent [ h_t-1,x_t]Relative to o_tThe weight coefficient matrix and the bias term of (1);

step 302, initializing the structure parameters of the prediction model: according to the requirements of a supervised learning algorithm, the determination requirement of the number of LSTM unit cells is greater than the characteristic number of the prediction model variable; aiming at the structural parameters of each LSTM cell unit, performing parameter initialization on a weight coefficient matrix in deep learning by adopting a 0.02 × randn (num _ params) method, wherein randn () represents a standard normal distribution function, and num _ params represents the number of parameters; for the bias matrix, all the bias matrix are set to 0 for initialization;

step 303, training the data subset S_tReading: on the basis of completing LSTM electric energy quality prediction model framework construction and prediction model structure parameter initialization, the electric energy quality LSTM prediction model is determined, and sufficient training based on effective training data must be carried out on a prediction network; therefore, the training data subset S obtained in step 203 is read_tAs training data for the LSTM prediction model;

step 304, setting of training cycle number Z and batch size P: in the LSTM prediction model training, too long training period number can cause the prediction model to be over-fitted, and too short training period number can cause the prediction model not to well reflect effective information contained in a training data set, so that the training data subset S is required to be used_tReasonably setting the training period number Z according to the data quantity; the training speed of LSTM supervised learning can be accelerated by adopting batch training, but the overlarge batch size in the batch training can occupy overlarge computer memory and reduce prediction precision, so that the method also needs to be based on S_tReasonably setting the batch size P according to the data volume in the step (1);

step 305, determining a power quality LSTM prediction model: the LSTM prediction model frame constructed in the steps 301 to 304, the set model structure parameters, the training parameters and the read training data subset S_tThe training process of the power quality LSTM prediction model can be implemented:

the first step is as follows: calculating the output value of each LSTM unit cell according to a forward propagation calculation method;

the second step is that: inversely calculating an error term for each LSTM unit cell, the error term being related not only to the output process described in step 301 but also to the hidden layer at the previous and subsequent time points;

the third step: calculating a gradient of each weight according to the corresponding error term;

the fourth step: according to the corresponding error items and the gradient of each weight, updating the weight removal value by applying a gradient optimization algorithm;

after the iterative training process is completed, all the parameter values involved in the structure of the LSTM prediction model are recorded and retained, including [ h ] described in step 301_t-1,x_t]And the bias matrix parameters

Step 4, predicting model performance evaluation: according to the electric energy quality LSTM prediction model determined in the step 3, before the electric energy quality LSTM prediction model is really put into an actual DG-containing distribution network system for electric energy quality prediction, effect test and performance evaluation of the prediction model are carried out;

step 401, testing the data subset S_cReading: the performance evaluation of the power quality LSTM prediction model must be based on sufficient tests based on historical data, and therefore, the test data subset S obtained in step 203 needs to be read_cAs performance evaluation data of the LSTM prediction model; subset of test data S_cThe historical data in (1) fall into two categories: the system power quality influence factor data is used as input variable type, and the power quality index item data is used as output data type;

step 402, based on the test data subset S_cThe electric energy quality prediction: performing an electric energy quality LSTM prediction process based on deep learning according to the electric energy quality LSTM prediction model determined in the step 3 and the system electric energy quality influence factor data read in the step 401 as the input variable category to obtain a corresponding system electric energy quality prediction value;

step 403, predicting model performance evaluation: comparing the predicted power quality value obtained in step 402 and output as LSTM prediction model, and the test data subset S read in step 401_cThe electric energy quality index item data is used as the output data type to realize the prediction performance evaluation of the obtained LSTM prediction model;

in order to effectively evaluate the prediction accuracy of the LSTM prediction model of the power quality, the root mean square error index epsilon of the LSTM prediction model is considered through a formula (8)_RMSE：

In the formula: y is_jFor testing a subset S of data_cThe original actual historical index value of the power quality,

n is a subset S of the test data for the predicted power quality value obtained in step 402_cThe number of data sets used for the evaluation test; j is the data group number for testing;

ε_RMSEreflects the average deviation degree of the predicted value relative to the observed value, and the value of the deviation degree is related to n; however, when the data quantity n is determined, epsilon_RMSEThe smaller the value of (A), the higher the prediction accuracy of the reflection prediction model is, and the better the prediction performance is;

and 5, predicting the power quality index of the system in the future period: based on the test evaluation in the step 4, an LSTM prediction model with good performance is confirmed, and reasonable prediction of the electric energy quality index item value of the system in the future period can be carried out on the premise that various input variable values influencing the electric energy quality of the system in the future period are available;

step 501, inputting variable data set reading in future time period: acquiring and reading an input variable data set of a DG-containing target power distribution network system in a certain period in the future, wherein the input variable data set comprises a temperature value, an illumination value, a system power load value and a time value;

step 502, predicting the electric energy quality in the future period: based on the input variable data set read in step 501 in a certain future period and the LSTM prediction model determined in step 3, performing power quality prediction on a DG-containing target power distribution network to obtain prediction data of a power quality index item of the system in the corresponding future period;

step 503, inverse normalization processing of the prediction data: the electric energy quality index item prediction data obtained in step 502 belongs to normalized data, and in order to obtain system electric energy quality prediction data with actual physical significance, inverse normalization processing needs to be performed according to a formula (9):

x＝x′*(x_max-x_min)+x_min (9)。

Images