Distribution network line loss prediction method based on maximum mutual information coefficient and deep learning

Abstract

The present invention provides a distribution network line loss prediction method based on maximum mutual information coefficient and deep learning. The method uses actual distribution network historical line loss data and climate and meteorological data as characteristic variables to calculate the maximum mutual information coefficient between the characteristic variables and the line loss at the time to be predicted; a distribution network line loss prediction model consisting of an input layer, a multi-layer LSTM unit, a Dropout layer and an output layer is constructed; a maximum mutual information coefficient value greater than M The characteristic variables of the distribution network are used as a feature set. Within the feature set, features are selected in descending order according to the maximum mutual information coefficient value as the input of the distribution network line loss prediction model. The optimal input feature set is constructed in combination with the model prediction error to achieve accurate prediction of distribution network line loss. The distribution network line loss prediction method of the present invention can quantitatively analyze the correlation between the distribution network's historical line loss characteristics at multiple times, climate and meteorological characteristics, and the line loss at the time to be predicted, construct a complete and effective model input feature set, reduce the complexity of model training, and improve the prediction accuracy of the line loss prediction model.

Description

Technical Field

[0001] The present invention belongs to the field of steady-state analysis of power systems and relates to a distribution network line loss prediction method. More specifically, it relates to a distribution network line loss prediction method based on maximum mutual information coefficient and deep learning. Background Art

[0002] Distribution network line loss represents the energy loss caused by the power system during the power distribution process. It is an important comprehensive indicator for measuring the operating status, management quality, and economic benefits of power networks. It can directly reflect the planning, design, and management and operation level of the power network, and directly guide the operation, maintenance, transformation, upgrading, and planning and scheduling of the power network. Accurate prediction of distribution network line loss can effectively guide grid management personnel in regulating grid operations and promptly determine the safety status of power lines. With the rapid advancement of smart grid construction in recent years and the increasing maturity of technologies such as big data and artificial intelligence, grid operation measurement data has grown exponentially, providing effective support for the accurate prediction of distribution network line loss. However, due to the massive and multi-source nature of distribution network line loss data, traditional prediction methods based on physical mechanisms cannot effectively analyze the nonlinear hidden connections between multi-source massive measurement data and distribution network line loss.

[0003] In-depth analysis of power systems based on data-driven methods is a core strategic requirement for current smart grid development. Unlike traditional analysis methods based on physical models, data-driven methods deeply explore the spatiotemporal correlations of massive amounts of historical grid data to establish efficient data-driven models. This avoids complex and difficult physical modeling and effectively improves the efficiency and accuracy of analytical modeling. Currently, while distribution network line loss prediction methods based on intelligent algorithms such as deep learning can effectively mine nonlinear hidden connections between operating parameters in historical grid data, the multi-source and massive nature of grid operating parameter data makes training distribution network prediction models based on deep learning time-consuming and labor-intensive. How to achieve rapid and accurate prediction of distribution network line losses based on data-driven methods is a significant research topic. Summary of the Invention

[0004] The purpose of the present invention is to provide a distribution network line loss prediction method based on maximum mutual information coefficient and deep learning, which can effectively construct the input feature set of the line loss prediction model and improve the accuracy and speed of the line loss prediction model.

[0005] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:

[0006] A distribution network line loss prediction method based on maximum mutual information coefficient and deep learning, the steps of this method are as follows:

[0007] (1) Obtain the historical data of distribution network line loss and the corresponding climate and meteorological data, construct characteristic variables based on the historical data of distribution network line loss in different periods before the prediction time and the climate and meteorological data in different periods before the prediction time, use the line loss data at the prediction time as the output feature, and calculate the maximum mutual information coefficient of each characteristic variable and the output feature respectively;

[0008] (2) Establish a distribution network line loss prediction model consisting of an input layer, a multi-layer LSTM unit, a Dropout layer, and an output layer; select n feature variables with a maximum mutual information coefficient value greater than M as a feature set, and then select feature variables in the feature set in descending order according to the maximum mutual information coefficient value as the input of the distribution network line loss prediction model. The number of selected features is recorded as m, and the initial value is 1;

[0009] (3) According to the constructed distribution network line loss prediction model and input features, iteratively train and test the line loss prediction model to obtain the model prediction error and determine whether m is less than n. If so, m=m+1, and return to continue iterative training and testing the line loss prediction model. Otherwise, terminate the model training.

[0010] (4) According to the obtained model prediction error, the optimal input feature set of the distribution network line loss prediction model is determined, data is collected according to the optimal input feature set, and input into the distribution network line loss prediction model to realize the distribution network line loss prediction.

[0011] As a preferred implementation, in step (1), the data sampling interval is T, and k points are sampled per day; the line loss data s*T time before the prediction time and the climate and meteorological data v*T time before the prediction time are respectively taken to construct feature variables, and the line loss data at the prediction time is used as the output feature, and the maximum mutual information coefficient of each feature variable and the output feature is respectively calculated, where s and v are natural numbers, s=1~7*k, v=0~7*k.

[0012] As a preferred embodiment, the mutual information value between the feature variable and the output feature is first calculated, and then the maximum mutual information coefficient between the feature variable and the output feature is calculated based on the mutual information value. When calculating the mutual information value, the calculation is simplified based on the following method:

[0013] The characteristic variables a and b are transformed into a scatter plot with a and b as the coordinate axes, and the a coordinate axis is divided into segment A and the b coordinate axis is divided into segment B, that is, multiple small squares are obtained; the probability of the coordinate point (a, b) falling into each small square is calculated to obtain the joint probability p(a, b), and then the mutual information value between a and b is calculated.

[0014] As a preferred implementation, in the distribution network line loss prediction model, the input layer is composed of normalized feature variables, and the number of input layer nodes is the dimension of the feature variables.

[0015] As a preferred implementation, the LSTM unit is composed of a forget gate, an input gate, and an output gate.

[0016] As a preferred implementation, the output layer is composed of a fully connected layer, and the number of nodes in the output layer is the dimension of the output features.

[0017] As a preferred embodiment, the LSTM unit is provided with two layers, the number of LSTM unit nodes in the first layer is 85, the number of LSTM unit nodes in the second layer is 100, and the Dropout probability value of the Dropout layer is 0.1.

[0018] As a preferred implementation, the value of the feature set selection criterion M is 0.3 times the maximum value of the maximum mutual information coefficients of all feature variables and output features.

[0019] As a preferred implementation, the climate and meteorological data include temperature values, wind speed values, and precipitation values.

[0020] As a preferred implementation method, the acquired distribution network line loss historical data and the corresponding climate and meteorological data are divided into training set data and test set data. The training set data is used to quantitatively analyze the maximum mutual information coefficient between the characteristic variables and the output features and participate in model training. The test set data is used to perform accuracy testing on the line loss prediction model after each model training to obtain the model prediction error.

[0021] As a preferred implementation, the model prediction error adopts the root mean square error.

[0022] The present invention proposes a distribution network line loss prediction method based on maximum mutual information coefficient and deep learning. This method avoids complex and difficult physical modeling. The maximum mutual information coefficient is used to quantitatively analyze the correlation between the multi-time line loss characteristics of the actual distribution network, the climate and meteorological characteristics, and the line loss at the time to be predicted. Compared with linear correlation analysis, the quantitative correlation analysis based on the maximum mutual information coefficient is more comprehensive and can dig out the nonlinear correlation between input features and output features. Further, based on the quantitative analysis results of the correlation, the input feature set is selected, which can effectively remove the adverse effects of invalid features and improve data quality. Based on the selected input feature set and combined with a deep LSTM neural network, a distribution network line loss prediction model is constructed. The model training time and difficulty are greatly reduced, and the prediction accuracy is further improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a schematic flow chart of the method of the present invention.

[0024] Figure 2 It is a schematic diagram of the line loss prediction model structure of the present invention.

[0025] Figure 3 This is a graph of the test error results of the line loss prediction model based on different input features of the present invention.

[0026] Figure 4 3 is a schematic diagram of line loss prediction results based on real data of an actual distribution network in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is described in detail below with reference to the accompanying drawings and implementation examples. It should be understood that the specific implementation examples described herein are only used to explain the present invention and are not intended to limit the invention.

[0028] Figure 1 The following is a flow chart of the method of the present invention, which introduces the basic steps of the method of the present invention.

[0029] 1) The present embodiment was validated based on real-world line loss historical data and climate data from an actual distribution network. Data was sampled at a 3-hour interval, with 8 points sampled per day. Line loss and climate data were collected from 485 days between 2017 and 2018, totaling 3,880 points. Data from points 1-3,200 served as the training set for the distribution network line loss prediction model, while data from points 3201-3,880, totaling 680 points, served as the test set for the model. Quantitative correlation analysis was performed based on the model training set. The input features involved in the quantitative analysis were line loss s*3 hours before the predicted time (s ranged from 1 to 56) and temperature, wind speed, and precipitation v*3 hours before the predicted time (v ranged from 0 to 56). This analysis involved correlation analysis of line loss-related features within seven days before the predicted time. The output feature was the line loss at the predicted time. The simulation experiments were conducted on a 64-bit computer with a 2.9GHz Intel Core i5 processor. A deep LSTM neural network was implemented using the MATLAB Deep Learning Toolbox.

[0030] The model input and output feature data involved in the quantitative analysis of correlation are constructed as follows:

[0031] X t =(x t ,x t+1 ,...x t+3143 ),t=1,2,...,56

[0032] X t1 =(x t1 ,x t1+1 ,...x t1+3143 ),t1=1,2,...,57 (1)

[0033] Y=(y t2 ,yt2+1 ,...y t2+3143 ),t2=57

[0034] Where, X t is the historical line loss value; X t1 is a meteorological climate characteristic data set, which includes temperature, wind speed, and precipitation values; Y is the model output feature, i.e., the line loss value at the time to be predicted; t, t1, and t2 are time points;

[0035] The quantitative analysis of correlation is based on the maximum mutual information coefficient, and its specific calculation method is as follows:

[0036] 11) For two given characteristic variables a and b (in this embodiment, a is X t 、X t1 , where X t1 Take the temperature value, wind speed value, and precipitation value respectively, and b is Y), and first calculate the mutual information value between them. The calculation formula is as follows:

[0037]

[0038] Where p(a,b) is the joint probability between feature variables a and b;

[0039] 12) To simplify the calculation process of the joint probability p(a, b), the characteristic variables a and b are transformed into a scatter plot with a and b as the coordinate axes. The a coordinate axis is divided into segment A and the b coordinate axis is divided into segment B. This will result in multiple small squares. By calculating the probability of the coordinate point (a, b) falling into each small square, the joint probability p(a, b) is obtained, and then the mutual information value between a and b is calculated. The calculation formula is as follows:

[0040]

[0041] 13) Based on I(a;b) obtained in step 12), the maximum mutual information coefficient (MIC) between variables a and b is calculated as follows:

[0042]

[0043] Where I(a,b) is the mutual information between variables a and b; A and B are the number of segments along the a and b directions, i.e., the grid distribution. The values ​​of A and B increase from 1 and satisfy A*B≤Q. The size of Q is 0.6 of the total amount of data.

[0044] 2) Based on the deep LSTM neural network theory, a multi-layer deep neural network distribution network line loss prediction model with input layer, hidden layer, Dropout layer and output layer is constructed. The schematic diagram of the line loss prediction model structure is shown in the figure. Figure 2 The specific construction method is as follows:

[0045] 21) The input layer of the distribution network line loss prediction model consists of normalized input features. The number of input layer nodes is the dimension of the input features:

[0046] 22) The hidden layer of the distribution network line loss prediction model is composed of multiple layers of LSTM units. The LSTM unit contains a forget gate, an input gate, and an output gate. The calculation formula is:

[0047] f t =σ(W f *[h t-1 ,x t ]+b f )

[0048] i t =σ(W i *[h t-1 ,x t ]+b i ) (5)

[0049] o t =σ(W o *[h t-1 ,x t ]+b o )

[0050] Where, f t 、i t 、o t Respectively represent the calculation results of the forget gate, input gate and output gate at time t; W f 、W i 、W o Represents the weight matrices of the forget gate, input gate, and output gate respectively; b f 、b i 、b o Represent the bias terms of the forget gate, input gate, and output gate respectively; σ is the sigmoid activation function;

[0051] 23) The Dropout layer of the distribution network line loss prediction model is used to prevent overfitting of the model. The output layer is composed of a fully connected layer. The number of nodes in the output layer is the dimension of the output feature. In this embodiment, the output feature dimension is 1, which is the line loss value at the time to be predicted.

[0052] 24) The model prediction error is calculated using the root mean square error (RMSE), which is calculated as follows:

[0053]

[0054] Where yi ' is the predicted value; yi is the true value; z is the total number of samples; 3) According to the quantitative analysis results of the correlation obtained in step 1), the maximum mutual information coefficient between the feature variable and the output feature is 0.5699, that is, M is 0.1710. Feature variables with a maximum mutual information coefficient greater than M are selected to form the feature set, and a total of 47 feature variables are finally selected;

[0055] 31) To verify the effectiveness of the selected feature set, a test was conducted based on the constructed line loss prediction model; the input features were selected from all feature variables (a total of 227 feature variables within 7 days), feature variables with a maximum mutual information coefficient greater than M (a total of 47 feature variables), and feature variables with a maximum mutual information coefficient less than M (a total of 180 feature variables);

[0056] 32) The prediction model structure (number of LSTM layers and nodes) corresponding to different input features was determined through multiple experiments (to ensure the minimum RMSE prediction error). The number of model training iterations was 250, and the initial learning rate was 0.01.

[0057] 33) The prediction model structure determined by different input features was tested. The test results of all feature variables, feature variables with maximum mutual information coefficient greater than M, and feature variables with maximum mutual information coefficient less than M as input features of the prediction model are as follows (taking the average value of 5 experimental tests): average prediction error (RMSE): 0.087, 0.021, 0.131; model iteration training time: 136 seconds, 77 seconds, 117 seconds.

[0058] The test results directly prove the effectiveness of constructing the input features of the line loss prediction model based on the maximum mutual information coefficient. It can deeply explore the correlation between historical line losses at multiple times, meteorological and climate characteristics and line losses at the time to be predicted, and thus effectively remove the adverse effects of invalid features, reduce the dimension of input features, reduce the complexity of model iterative training, and improve the accuracy of line loss prediction.

[0059] 4) To further construct the optimal input feature set for the line loss prediction model, features are selected from the selected feature set in descending order according to the maximum mutual information coefficient value as the input of the distribution network line loss prediction model. The number of selected features is denoted as m, and the initial value is 1. The maximum mutual information coefficient value is greater than or equal to 0, and the larger the maximum mutual information coefficient value between two sets of sequence data, the higher the correlation between the two sets of sequence data.

[0060] According to the distribution network line loss prediction model and input features obtained in steps 2) and 3), the line loss prediction model is iteratively trained based on the training set, and the prediction accuracy is tested on the test set to obtain the model prediction error. It is then determined whether m is less than n. If so, m = m + 1 and the process returns to step 3). If not, the process proceeds to step 5).

[0061] The model prediction error test results based on different input features are as follows Figure 3 As shown;

[0062] 5) The optimal input feature set is the input feature corresponding to the minimum RMSE value (0.0141) obtained in step 4). In this example, the optimal number of input features is 6, which are the historical line losses 8*3, 1*3, 7*3, 2*3, and 6*3 hours before the predicted time, and the historical temperature 1*3 hours before the predicted time.

[0063] Based on the optimal feature input feature historical data, combined with the multi-layer deep neural network line loss prediction model obtained in step 2), the line loss prediction results obtained based on the first 12 days of sampling data on the test set are as follows: Figure 4 As shown in the figure, the maximum number of iterations for line loss prediction model training is 250, the model iteration training time is 56 seconds, and the model parameters after training are as follows: number of input layer nodes: 6, 2 LSTM layers in total, number of nodes in LSTM layer 1: 85, number of nodes in LSTM layer 2: 100, the Dropout probability value of the Dropout layer is 0.1, and the number of output features is 1, which is the line loss value at the time to be predicted.

[0064] The proposed method for predicting line losses in distribution networks, based on maximum mutual information coefficient and deep learning, provides a quantitative reference standard for selecting input features for line-loss prediction models and effectively constructs the optimal input feature set for these models. By processing massive amounts of historical distribution network and meteorological climate data from multiple sources, the correlation between these multi-source features and the output of the line-loss prediction model is quantified. This effectively removes the adverse effects of invalid features, significantly reduces the dimensionality of the input features, and lowers the iterative training complexity of the line-loss prediction model, further improving the accuracy of line-loss predictions.

Claims

1. A distribution network line loss prediction method based on maximum mutual information coefficient and deep learning, characterized in that: The method comprises the following steps: (1) Obtain the historical data of distribution network line loss and the corresponding climate and meteorological data, construct characteristic variables based on the historical data of distribution network line loss in different periods before the prediction time and the climate and meteorological data in different periods before the prediction time, use the line loss data at the prediction time as the output feature, and calculate the maximum mutual information coefficient between each characteristic variable and the output feature; (2) Establish a distribution network line loss prediction model consisting of an input layer, a multi-layer LSTM unit, a Dropout layer, and an output layer; select the maximum mutual information coefficient value greater than M of n The feature variables are taken as the feature set, and then the feature variables are selected in descending order according to the maximum mutual information coefficient value as the input of the distribution network line loss prediction model. The number of selected features is recorded as m , the initial value is 1; (3) According to the constructed distribution network line loss prediction model and input features, iteratively train and test the line loss prediction model to obtain the model prediction error and judge m Is it less than n , if so, then m = m +1, return to continue iterative training and testing of the line loss prediction model, otherwise end model training; (4) According to the obtained model prediction error, the optimal input feature set of the distribution network line loss prediction model is determined. Data is collected according to the optimal input feature set and input into the distribution network line loss prediction model to realize the distribution network line loss prediction.

2. The method according to claim 1, characterized in that In step (1), the data sampling interval is T , one-day sampling k points; take the points before the prediction time s * T Line loss data before the time, before the prediction time v * T The climate and meteorological data before the time are used to construct the characteristic variables, and the line loss data at the prediction time is used as the output feature. The maximum mutual information coefficient between each characteristic variable and the output feature is calculated respectively, where s 、 v is a natural number, s =1~7* k , v =0~7* k .

3. The method according to claim 1, characterized in that First, calculate the mutual information value between the feature variable and the output feature, and then calculate the maximum mutual information coefficient between the feature variable and the output feature based on the mutual information value. When calculating the mutual information value, simplify the calculation based on the following method: The characteristic variable a and b Transformed into a 、 b For the scatter plot with the coordinate axis, a The coordinate axis is divided into A part, b The coordinate axis is divided into B Segment, that is, multiple small squares are obtained; Calculate the coordinates of each small square ( a,b ) falls into the probability, and the joint probability is obtained p ( a,b ), and then calculate a and b The mutual information value between .

4. The method according to claim 1, wherein In the distribution network line loss prediction model, the input layer is composed of normalized feature variables, and the number of input layer nodes is the dimension of the feature variables.

5. The method according to claim 1, wherein The LSTM unit consists of a forget gate, an input gate and an output gate.

6. The method according to claim 1, characterized in that The output layer is composed of a fully connected layer, and the number of nodes in the output layer is the dimension of the output feature.

7. The method according to claim 1, characterized in that The LSTM unit is provided with two layers, the number of LSTM unit nodes in the first layer is 85, the number of LSTM unit nodes in the second layer is 100, and the Dropout probability value of the Dropout layer is 0.

1.

8. The method according to claim 1, characterized in that described M The value of is 0.3 times the maximum value of the maximum mutual information coefficients of all feature variables and output features.

9. The method according to claim 1, characterized in that The climate and meteorological data include temperature values, wind speed values, and precipitation values.

10. The method according to claim 1, characterized in that The acquired distribution network line loss historical data and the corresponding climate and meteorological data are divided into training set data and test set data. The training set data is used to quantitatively analyze the maximum mutual information coefficient between the characteristic variables and the output features and participate in model training. The test set data is used to perform accuracy testing on the line loss prediction model after each model training to obtain the model prediction error; the model prediction error adopts the root mean square error.

Images