CN110110887A - To the prediction technique of low-voltage platform area line loss per unit - Google Patents

Abstract

This application proposes a prediction method for the line loss rate of the low-voltage station area, including obtaining the line loss data of the low-voltage station area, cleaning the obtained data, and eliminating the abnormal line loss data; performing K-Means on the cleaned data according to the characteristics Clustering, calculate the silhouette coefficient corresponding to each K value, and select the number of clusters whose silhouette coefficient is closest to 1 as the optimal number of clusters; normalize or standardize the clustered feature data, and determine the training set based on the processed data and a test set; construct a convolutional neural network model, use the obtained data matrix to train the model, and use the trained model to predict the data. The prediction model of the line loss rate of the low-voltage station area is established through the convolutional neural network based on K-Means clustering and deep learning theory, which not only considers the relationship between the total power supply, station area capacity, total users and other characteristic data and the current line loss rate In addition, historical line loss data such as the line loss rate of the previous month, the line loss rate of the same period, and the cumulative line loss rate of the same period were used to improve the accuracy of the forecast.

Description

Prediction Method of Line Loss Ratio in Low Voltage Station Area

technical field

The invention belongs to the field of data processing, in particular to a method for predicting the line loss rate of a low-voltage station area.

Background technique

Line loss rate is an important concern in the daily management of power grid enterprises. Line loss will be caused in the links of power generation, transmission, distribution, and power consumption, so the line loss rate is an important economic indicator in the power system. The current prediction methods for line loss include support vector machines, random forests, neural networks, etc. and their improved algorithms. Fast and accurate forecasting of rates.

Contents of the invention

In order to solve the shortcomings and deficiencies in the prior art, the present invention proposes a prediction method for the line loss rate of the low-voltage station area, and establishes a prediction method for the line loss rate of the low-voltage station area based on the convolutional neural network of K-Means clustering and deep learning theory. The prediction model can improve the accuracy of forecasting due to the addition of historical line loss data such as the line loss rate of the previous month, the line loss rate of the same period, and the cumulative line loss rate of the same period.

Specifically, the forecasting method includes:

Obtain the line loss data of the low-voltage station area from the database, clean the obtained data, and eliminate the abnormal line loss data;

Carry out K-Means clustering on the cleaned data according to the characteristics, calculate the silhouette coefficient corresponding to each K value, and select the cluster number with the silhouette coefficient closest to 1 as the optimal cluster number;

Normalize or standardize the clustered feature data, and determine the training set and test set based on the processed data;

Construct a convolutional neural network model, use the obtained data matrix to train the model, and use the trained model to predict the data.

Optionally, obtain the line loss data of the low-voltage station area from the database, perform data cleaning on the obtained data, and eliminate the abnormal line loss data, including:

Obtain the line loss data of the low-voltage station area from the database according to the written SQL script statement;

The retrieved data contains many useless attributes, empty attributes such as department numbers, etc., delete these attributes, and retain useful attributes such as T_PPQ, L_LLR, L_LM_LLR, L_PERIOD_LLR, ACCU_PERIOD_LLR, TG_CAP, TOTAL_USER_NUM; (Note: T_PPQ is the total power supply, L_LLR is Line loss rate, L_LM_LLR is the line loss rate of last month, L_PERIOD_LLR is the line loss rate of the same period, ACCU_PERIOD_LLR is the cumulative line loss rate of the same period, TG_CAP is the capacity of the station area, TOTAL_USER_NUM is the total number of users, the following are expressed in letters);

Eliminate abnormal line loss data from the processing results of the previous step, such as: T_PPQ<100, TOTAL_USER_NUM=0, L_PERIOD_LLR=0, L_LM_LLR=0, ACCU_PERIOD_LLR=0, and organize and save the eliminated data in a table.

Optionally, perform K-Means clustering on the cleaned data according to features, calculate the silhouette coefficient corresponding to each K value, and select the number of clusters whose silhouette coefficient is closest to 1 as the optimal number of clusters including:

Take out the two eigenvalues of T_PPQ and TOTAL_USER_NUM from the processing results of the previous step, perform K-Means clustering on all stations, calculate the silhouette coefficients corresponding to each K value, and select the number of clusters with the silhouette coefficient closest to 1 as the optimal number of clusters;

Optionally, normalize or standardize the clustered feature data, and determine the training set and test set based on the processed data, including:

To cluster the station area into K categories, K prediction models need to be established according to the K categories, and L_LM_LLR, L_PERIOD_LLR, ACCU_PERIOD_LLR, and TG_CAP are selected as feature input values in each prediction model, and standardized processing is performed before input;

In each prediction model, L_LLR is selected as the label value, and the line loss rate can be divided into four categories according to the line loss reference data of a city power grid: qualified line loss rate (0<L_LLR<3.5), unqualified line loss rate (3.5 <=L_LLR<10), unqualified line loss rate (10<=L_LLR<100), abnormal line loss rate (L_LLR<0), One-Hot encoding processing before input;

Divide feature values and label values together into training set and test set, and the split ratio is three-quarters and one-fourth.

Optionally, construct a convolutional neural network model, use the obtained data matrix to train the model, and use the trained model to predict the data, including:

Set the filters of the first layer of the convolutional neural network to 32, the filters of the second layer to 64, and the size of the convolution kernels of the two layers to 1×1, and add pooling after the two layers of convolution Layer, the pooling layer adopts the maximum pooling, the size is 1×1, and finally add a layer of Dropout, the Dropout ratio is set to 0.2, and the boundary is zero-filled, that is, the padding parameter is set to the same;

The fully connected layer adopts a one-layer structure, the number of neurons in the output layer is 2×2×64, and a softmax layer is added to make the output of the neural network into a probability distribution. The output after softmax regression processing is:

Where y ₁ , y ₂ ,...,y _n are the output of the original neural network, and y′ _i is the probability distribution of the output;

The loss function uses the cross-entropy loss function. Given two probability distributions p and q, the cross-entropy of p represented by q is:

The activation function selects the Relu function: F(x)=max(0,x), where x is the weighted sum of each neuron node, the optimizer is selected as an adaptive learning optimization algorithm, and the gradient is gradually reduced through backpropagation until the final convergence ;

Train the model and save the model, send 6294 training data of class 0 and 8009 training data of class 1 to the two models for training;

Calculate the output value of each neuron forward, calculate the error term of each neuron in reverse, calculate the gradient of each weight according to the corresponding error term, and optimize the model parameters.

The beneficial effects brought by the technical scheme provided by the invention are:

Aiming at the problem that the current power grid power system cannot quickly and accurately predict the line loss rate, the invention proposes to combine power data, and establish a prediction model for the line loss rate of low-voltage station areas based on convolutional neural networks based on K-Means clustering and deep learning theory. This method not only fully considers the relationship between characteristic data such as total power supply, station area capacity, and total users, and the current line loss rate, but also uses historical data such as the line loss rate of the previous month, the same period, and the cumulative line loss rate of the same period Line loss data to improve prediction accuracy.

Description of drawings

In order to illustrate the technical solution of the present invention more clearly, the accompanying drawings that need to be used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. Ordinary technicians can also obtain other drawings based on these drawings on the premise of not paying creative work.

Fig. 1 is a schematic flow chart of the method for predicting the line loss rate of the low-voltage platform area proposed by the embodiment of the present application;

Fig. 2 is a schematic diagram of the contour coefficient corresponding to each K value proposed in the embodiment of the present application;

Fig. 3 is a schematic diagram of the number and proportion of data of class 0 and class 1 proposed in the embodiment of the present application;

Fig. 4 is a schematic diagram of the number of steps and the accuracy rate of the type 0 test data proposed in the embodiment of the present application;

Fig. 5 is a schematic diagram of the number of steps and the accuracy rate of the test data of type 1 proposed in the embodiment of the present application.

Detailed ways

In order to make the structure and advantages of the present invention clearer, the structure of the present invention will be further described below in conjunction with the accompanying drawings.

Embodiment one

In order to solve the shortcomings and deficiencies in the prior art, the present invention proposes a prediction method for the line loss rate of the low-voltage station area, and establishes a prediction method for the line loss rate of the low-voltage station area based on the convolutional neural network of K-Means clustering and deep learning theory. The prediction model can improve the accuracy of forecasting due to the addition of historical line loss data such as the line loss rate of the previous month, the line loss rate of the same period, and the cumulative line loss rate of the same period.

This embodiment proposes a prediction method for the line loss rate in the low-voltage station area, as shown in Figure 1, the prediction method includes:

11. Obtain the line loss data of the low-voltage station area from the database, clean the obtained data, and eliminate the abnormal line loss data;

12. Perform K-Means clustering on the cleaned data according to the characteristics, calculate the silhouette coefficient corresponding to each K value, and select the cluster number with the silhouette coefficient closest to 1 as the optimal cluster number;

13. Normalize or standardize the clustered feature data, and determine the training set and test set based on the processed data;

14. Construct a convolutional neural network model, use the obtained data matrix to train the model, and use the trained model to predict the data.

In implementation, the specific content of step 11 includes:

111. Obtain the line loss data of the low-voltage station area from the database according to the written SQL script statement;

112. The retrieved data contains a lot of useless attributes, empty attributes such as department numbers, etc., delete these attributes, and retain useful attributes such as T_PPQ, L_LLR, L_LM_LLR, L_PERIOD_LLR, ACCU_PERIOD_LLR, TG_CAP, TOTAL_USER_NUM; (Note: T_PPQ is the total power supply, L_LLR is the line loss rate, L_LM_LLR is the line loss rate of the previous month, L_PERIOD_LLR is the line loss rate of the same period, ACCU_PERIOD_LLR is the cumulative line loss rate of the same period, TG_CAP is the capacity of the station area, TOTAL_USER_NUM is the total number of users, all of which are expressed in letters below);

113. Eliminate abnormal line loss data such as: T_PPQ<100, TOTAL_USER_NUM=0, L_PERIOD_LLR=0, L_LM_LLR=0, ACCU_PERIOD_LLR=0 from the processing results of the previous step, and organize and save the eliminated data in a table.

The specific content of step 12 includes:

Take out the two eigenvalues of T_PPQ and TOTAL_USER_NUM from the processing results of the previous step, perform K-Means clustering on all stations, calculate the silhouette coefficients corresponding to each K value, and select the number of clusters with the silhouette coefficient closest to 1 as the optimal The number of clusters; the silhouette coefficients corresponding to each K value are shown in Figure 2. It can be seen from Fig. 2 that when K=2, the silhouette coefficient is closer to 1, so the cleaned low-pressure station area data are divided into two categories through K-Mean clustering, which are recorded as 0 and 1 respectively. The number and proportion of data of category 0 and category 1 are shown in Figure 3.

The specific content of step 13 includes:

131. To cluster the station area into K categories, K forecast models need to be established according to K categories, and L_LM_LLR, L_PERIOD_LLR, ACCU_PERIOD_LLR, TG_CAP are selected as feature input values in each forecast model, and standardized processing is performed before input ;

132. In each prediction model, L_LLR is selected as the label value, and the line loss rate can be divided into four categories according to the reference data of a city’s power grid line loss: qualified line loss rate (0<L_LLR<3.5), unqualified line loss rate (3.5<=L_LLR<10), unqualified line loss rate (10<=L_LLR<100), abnormal line loss rate (L_LLR<0), One-Hot encoding processing before input;

133. Divide feature values and label values into a training set and a test set, and the split ratio is three quarters and one quarter.

Before performing K-Means clustering, the number of clusters K needs to be determined first, and then the silhouette coefficients corresponding to each K value need to be calculated. Silhouette coefficient calculation formula:

where a(i) is the average distance between sample x and other points in the cluster; b(i) is the average distance between sample x and all points in the nearest cluster;

Calculate the silhouette coefficient corresponding to each K value, and the calculation formula of the silhouette coefficient can be analyzed to show that S(i) is close to 1, which means that the sample clustering is reasonable, and the number of clusters with a silhouette coefficient close to 1 is selected as the optimal number of clusters K;

The clustering number K of the sample data is determined; then K-Means clustering can be performed. The distance metric for clustering is the square of the Euclidean distance:

Where x and y represent two different samples, and n represents the dimension of the sample (the number of features).

The specific implementation method is: to cluster the station area into 2 categories, then the root needs to establish 2 prediction models, select L_LM_LLR, L_PERIOD_LLR, ACCU_PERIOD_LLR, TG_CAP as the feature input value in each prediction model, and perform standardized processing before input ;

In each prediction model, L_LLR is selected as the label value, and the line loss rate can be divided into four categories according to the line loss reference data of a city power grid: qualified line loss rate (0<L_LLR<3.5), unqualified line loss rate (3.5 <=L_LLR<10), unqualified line loss rate (10<=L_LLR<100), abnormal line loss rate (L_LLR<0), One-Hot encoding processing before input;

The 8393 low-voltage station areas in category 0 and the 10679 low-voltage station areas in category 1 were divided into training set and test set respectively, and the first three-quarters and the last quarter of the training set were used as the test set, and Convert the training set and test set data into a matrix form for input into the convolutional neural network model; the specific data quantity is shown in Table 1;

Table 1: Quantity of each data

The specific content of step 14 includes:

141. Set the filters of the first layer of the convolutional neural network to 32, the filters of the second layer to 64, and the size of the convolution kernels of the two layers to 1×1, and add them after the two layers of convolution Pooling layer, the pooling layer adopts the maximum pooling, the size is 1×1, and finally add a layer of Dropout, the Dropout ratio is set to 0.2, and the boundary is zero-filled, that is, the padding parameter is set to the same;

142. The fully connected layer adopts a one-layer structure, the number of neurons in the output layer is 2×2×64, and a softmax layer is added to make the neural network output into a probability distribution. The output after softmax regression processing is:

Where y ₁ , y ₂ ,...,y _n are the output of the original neural network, and y′ _i is the probability distribution of the output;

143. The loss function adopts the cross entropy loss function. Given two probability distributions p and q, the cross entropy of p is represented by q as:

144. The activation function selects the Relu function: F(x)=max(0,x), where x is the weighted sum of each neuron node, the optimizer is selected as an adaptive learning optimization algorithm, and the gradient is gradually reduced through back propagation until final convergence;

145. Train the model and save the model, send 6294 training data of class 0 and 8009 training data of class 1 to the two models for training;

146. Calculate the output value of each neuron forward, calculate the error term of each neuron in reverse, calculate the gradient of each weight according to the corresponding error term, and optimize the model parameters.

In implementation, this part is divided into two parts: model training and testing.

1. Training part

Train the model and save the model, send 6294 training data of class 0 and 8009 training data of class 1 to the two models in matrix for training; calculate the output value of each neuron forward; calculate each neuron in reverse the error term;

Calculate the gradient of each weight according to the corresponding error term, optimize the model parameters, adjust the number of hidden units, learning rate, and Drop ratio of each layer, whether to use average pooling, whether to add convolutional layers, whether to Add Dropout layer, etc.;

Save the two trained models for testing purposes.

2. Test part

Use two saved models to make predictions respectively;

Send the 2099 test data of category 0 and the 2670 test data of category 1 into the two models in a matrix for one-by-one prediction;

Send 2099 test data of class 0 into the trained model, record it as 1 if the prediction is accurate, and record 0 if the prediction is wrong, and save the above 1 and 0 data, and send 2670 test data of class 1 into the trained model Do the same after the model;

Calculate the accuracy of the two prediction models by using the 1 and 0 data saved in class 0 and class 1. After calculation, the accuracy rate of the 2099 test data tests of class 0 is about 81%, and the test data test accuracy of 2670 test data of class 1 The accuracy rate is about 82%. The number of steps and the accuracy rate of the 0-type and 1-type test data are shown in Figure 4 and Figure 5.

The present invention proposes a combination of electric power data and a convolutional neural network based on K-Means clustering and deep learning theory to establish a prediction model for line loss rate in low-voltage station areas. Perform data cleaning on the acquired data, and perform K-Means clustering according to characteristics after removing abnormal line loss data. Station areas with similar characteristics should have similar line loss rates. Before K-Means clustering, the optimal number of clusters K is determined by the silhouette coefficient, and then clustered into K clusters, then normalized and standardized methods are used to process the data, and finally K convolutional neural network models are respectively established for prediction.

The serial numbers in the above embodiments are for description only, and do not represent the sequence of the components during assembly or use.

The above description is only an embodiment of the present invention, and is not intended to limit the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention Inside.

Claims (5)

1. the prediction technique of pair low-voltage platform area line loss per unit, which is characterized in that the prediction technique includes:

Line loss data in low-voltage platform area are obtained from database, the data of acquisition are subjected to data cleansing, reject wherein abnormal line loss Data；

The data cleaned are subjected to K-Means cluster by feature, calculate the corresponding silhouette coefficient of each K value, and select profile Cluster numbers of the coefficient closest to 1 are optimum clustering number；

Characteristic after normalization or standardization cluster, based on treated, data determine training set and test set；

Convolutional neural networks model is constructed, model is trained using data matrix is obtained, uses trained model logarithm According to being predicted.

2. the prediction technique according to claim 1 to low-voltage platform area line loss per unit, which is characterized in that described from database Low-voltage platform area line loss data are obtained, the data of acquisition are subjected to data cleansing, reject wherein abnormal line loss data, comprising:

The line loss data of low-voltage platform area are obtained from database according to the SQL script sentence write；

The data of taking-up include many useless attribute, null attribute such as industry numbers etc., are deleted these attributes, and T_ is retained The useful category such as PPQ, L_LLR, L_LM_LLR, L_PERIOD_LLR, ACCU_PERIOD_LLR, TG_CAP, TOTAL_USER_NUM Property；

The line loss data of rejecting abnormalities from the processing result of previous step, and the data preparation after rejecting is protected in a table It deposits.

3. the prediction technique according to claim 1 to low-voltage platform area line loss per unit, which is characterized in that the data that will be cleaned K-Means cluster is carried out by feature, calculates the corresponding silhouette coefficient of each K value, and selects cluster numbers of the silhouette coefficient closest to 1 Include: for optimum clustering number

Two characteristic values of T_PPQ, TOTAL_USER_NUM are taken out from the processing result of previous step, and K- is carried out to all areas Means cluster, calculate the corresponding silhouette coefficient of each K value, and select silhouette coefficient closest to 1 cluster numbers be optimum cluster Number.

4. the prediction technique according to claim 1 to low-voltage platform area line loss per unit, which is characterized in that the normalization or mark Characteristic after standardization cluster, based on treated, data determine training set and test set, comprising:

Platform area is clustered as K classification, then needs to establish K prediction model according to K classification, is chosen in each prediction model L_LM_LLR, L_PERIOD_LLR, ACCU_PERIOD_LLR, TG_CAP are characterized input value, quasi- in input advance rower Change processing, the value of K are positive integer；

Choosing L_LLR in each prediction model and being used as is label value, can be by line loss per unit according to certain alternating current net wire loss reference data Be divided into four classes: qualified line loss per unit owes qualified line loss per unit, unqualified line loss per unit, abnormal line loss per unit, and One-Hot is carried out before input Coded treatment；

Characteristic value and label value are divided into training set and test set together, and primary contract is 3/4ths and a quarter.

5. the prediction technique according to claim 1 to low-voltage platform area line loss per unit, which is characterized in that the building convolution mind Through network model, model is trained using data matrix is obtained, data are predicted using trained model, is wrapped It includes:

32 are set by the Filter of the first layer of convolutional neural networks, the Filter of the second layer is set as 64, two layers volume Product core size is both configured to 1 × 1 size, and pond layer is separately added into after two layers of convolution, and pond layer is using maximum pond, size 1 × 1, one layer of Dropout is finally added, Drop ratio is set as 0.2, carries out the i.e. setting padding parameter of zero padding to boundary and is same；

Full articulamentum uses 1 layer of structure, output layer neuron number 2 × 2 × 64, then plus one layer softmax layers make nerve net Network output becomes a probability distribution, through the output after softmax recurrence processing are as follows:

Wherein y₁,y₂,…,y_nIt is the output of original neural network, y '_iIt is the probability distribution of output；

Loss function gives two probability distribution p and q, the cross entropy of p is indicated by q using cross entropy loss function Are as follows:

Activation primitive chooses Relu function: F (x)=max (0, x), and wherein x is the weighted sum of each neuron node, optimizer choosing It is selected as adaptive learning optimization algorithm, gradient decline is gradually carried out by backpropagation until convergence；

8009 training datas of 6294 training datas of 0 class and 1 class are respectively fed to two by training pattern and preservation model Model is trained；

The each neuron output value of forward calculation, the error term of each neuron of retrospectively calculate calculate every according to corresponding error term The gradient of a weight, optimizes model parameter.