A method for predicting the construction scale of distribution network expansion capacity in a timely manner

Abstract

A method for predicting the construction scale of the time series of the supporting capacity of the distribution network business expansion includes the following steps: extracting the time series influencing factor data and historical construction capacity data of similar business expansion supporting capacity construction projects; preprocessing the data set and dividing it into a training set and a test set; constructing a distribution network business expansion supporting capacity time series construction scale prediction model based on an interpretable recurrent neural network, the prediction model includes an input layer, a linear layer, a nonlinear feature extraction layer and a linear regression layer; applying corresponding penalty constraints to the weight parameters of the prediction model to obtain an improved prediction model; inputting the data of the training set into the improved prediction model for training, and inputting the data of the test set into the trained prediction model to obtain the prediction result; judging whether the deviation between the prediction result and the true value is within the allowable range, and if so, outputting the prediction result. This design improves the flexibility of the prediction algorithm structure while ensuring the interpretability and prediction accuracy of the prediction result.

Description

Technical Field

[0001] The present invention relates to the field of distribution networks, and in particular to a method for predicting the time-sequence construction scale of distribution network industry expansion supporting capacity, which is mainly suitable for improving prediction accuracy. Background Art

[0002] The business expansion supporting power grid project is a very important project in the power supply enterprise. It aims to meet the customers' new installation and capacity increase electricity demand. The power supply enterprise is responsible for the construction of new, expanded and renovated distribution network projects with voltage levels of 35KV and below. The existing business expansion supporting project capacity construction is mainly based on the expert experience method. The experts formulate the construction capacity of the business expansion supporting projects of the distribution network in different cities based on actual experience and familiarity with the forecast object. The results obtained by different experts are not exactly the same. There is a large degree of subjective arbitrariness and lack of a scientific measurement system. In addition, the construction scale of the business expansion supporting projects of the distribution network in various cities has a certain degree of uncertainty. It is difficult to fully consider the timing characteristics of the construction of the business expansion supporting capacity based on the subjective conjecture of experts alone, which may easily lead to inadequate execution of the electricity application and access link for customers in various industries. Therefore, it is urgent to build a business expansion supporting capacity timing construction scale prediction model to scientifically predict the construction timing and planned scale of the business expansion supporting project capacity for customers in different industries in various cities and improve the prediction accuracy. Summary of the invention

[0003] The purpose of the present invention is to overcome the defects and problems of low prediction accuracy in the prior art and to provide a method for predicting the construction scale of the expansion capacity of the distribution network industry with high prediction accuracy.

[0004] To achieve the above objectives, the technical solution of the present invention is: a method for predicting the construction scale of the expansion capacity of the distribution network industry, comprising the following steps:

[0005] S1. Conduct a correlation analysis on the influencing indicators of the timing and scale of the construction of supporting capacity for business expansion, and find that the indicators with the greatest impact on supporting projects for business expansion are timing influencing factors and historical construction capacity;

[0006] S2. Screen out similar business expansion supporting capacity construction projects from the massive historical business expansion supporting capacity construction scale database, and extract the time series influencing factor data and historical construction capacity data of similar business expansion supporting capacity construction projects;

[0007] S3, first pre-processing the time series influencing factor data and historical construction capacity data extracted in step S2, then normalizing the data set, and then dividing the data set into a training set and a test set;

[0008] S4. Construct a distribution network industry expansion capacity time series construction scale prediction model based on an interpretable recurrent neural network, which includes an input layer, a linear layer, a nonlinear feature extraction layer and a linear regression layer;

[0009] The input layer is used to input the time series influencing factor data and historical construction capacity data of similar business expansion supporting capacity construction projects;

[0010] The linear layer is used to linearly combine the data of the input layer;

[0011] The nonlinear feature extraction layer uses Legendre polynomial to extract nonlinear features and uses it as input of the linear regression layer;

[0012] The linear regression layer is used to linearly combine the nonlinear features to form the final prediction result;

[0013] S5, applying corresponding penalty constraints to the weight parameters of the prediction model to obtain an improved prediction model;

[0014] S6, inputting the data of the training set in step S3 into the improved prediction model for training to obtain a trained prediction model, and inputting the data of the test set in step S3 into the trained prediction model to obtain a prediction result;

[0015] S7. Determine whether the deviation between the predicted result and the true value is within the allowable range. If so, output the predicted result; otherwise, return to step S4.

[0016] In step S1, the influencing indicators include load density, load utilization, average household distribution transformer capacity, electricity sales, daily maximum load and historical construction capacity; the time series influencing factors include load density, average household distribution transformer capacity and daily maximum load.

[0017] In step S2, similar business expansion supporting capacity construction projects are screened from a massive database of historical business expansion supporting capacity construction scales using the K-Means clustering algorithm.

[0018] In step S3, the interquartile range method is first used to detect outliers, and then the outliers are removed from the data set. Then the quadratic interpolation method is used to fill the data gaps. The Min-Max normalization method is then used to normalize the data set so that the data of different dimensions are distributed between [0-1]. Then the data set is divided into 65% as a training set and 35% as a test set.

[0019] In step S4, the input of the input layer is the feature vector x at the current moment t and the memory value vector h of the prediction model at the previous moment t-1 , sort the typical features of the time series influencing factor data and the historical construction capacity data to obtain a feature vector x -t ,Right now:

[0020]

[0021] Predict the memory value vector h of the previous moment of the model t-1 for:

[0022]

[0023] In the formula, is the value of the i-th eigenvector at time t, d is the number of eigenvectors, is the value of the i-th hidden layer neuron at time t-1, and k is the number of hidden layer neurons;

[0024] The linear layer consists of k nodes, each of which performs linear weighting on the time series influencing factor data and the historical construction capacity data, and uses the linear weighted value of each node as the input of the corresponding nonlinear feature extraction layer, namely:

[0025]

[0026] In the formula, β j is the projection weight of the input data of the jth node of the linear layer, is the linear weighted value of the jth node in the linear layer at time t;

[0027] The nonlinear feature extraction layer is composed of k sub-networks. The input and output of each sub-network are single variables. The sub-networks are not connected to each other. The sub-networks form a mapping relationship with the nodes of the linear layer. The sub-network performs single-variable activation processing on the output of each node of the linear layer and outputs the corresponding activation value. And use it as the input of the corresponding linear regression layer; the univariate activation processing process is as follows:

[0028]

[0029] In the formula, g j (·) is the activation function corresponding to the jth node;

[0030] The linear regression layer consists of a single node. The input of the linear regression layer is the activation value of all nodes in the nonlinear feature extraction layer. The output is the value after linear transformation and linear combination of this layer, that is, the linear combination of the output activation function:

[0031]

[0032] In the formula, y t is the prediction result, γ k is the scaling parameter and μ is the shift parameter.

[0033] In step S5, the corresponding penalty constraint is imposed on the weight parameters of the prediction model as follows:

[0034] (1) Regularize the projection weights and scaling parameters:

[0035]

[0036]

[0037] In the formula, β ij is the projection weight between the i-th neuron in the input layer and the j-th neuron in the linear layer, γ j is the scaling parameter of the jth neuron; T1 and T2 are the regularization strengths of the parameters, T1 ≥ 1, T2 ≥ 0;

[0038] (2) Implement roughness penalty constraints on the activation function:

[0039] ∫[g″ j (z)] 2 dF j (z)≤T3

[0040] In the formula, z is the training value of the feature variable, F j (·) is the cumulative distribution function of the j-th projection variable; T3 is the regularization strength of the parameter, T3 ≥ 0;

[0041] (3) Implement orthogonal constraints on projection weights:

[0042] β T β=I k

[0043] In the formula, I k is a k-order all-one matrix;

[0044] (4) Standardization constraints:

[0045] ∫g j (z)dF j (z)=0

[0046] ∫g j (z) 2 dF j (z)=1.

[0047] In step S7, the mean absolute error, root mean square error, coefficient of variation and coefficient of determination are used to comprehensively evaluate the prediction results:

[0048]

[0049]

[0050]

[0051]

[0052] In the formula, O j is the actual value of the construction scale of the distribution network industry expansion supporting capacity of the i-th sample, is the forecast value of the construction scale of the expansion capacity of the distribution network industry of the i-th sample, N is the number of samples, MAE is the mean absolute error, RMSE is the root mean square error, CV is the coefficient of variation, R 2 is the determination coefficient.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] In a method for predicting the construction scale of the supporting capacity of a distribution network according to the present invention, based on an interpretable recurrent neural network algorithm, data such as load density, average household distribution transformer capacity, daily maximum load, and historical construction capacity that affect the construction scale of the supporting capacity of the distribution network are used as characteristic parameters for predicting the construction scale of the supporting capacity of the distribution network. By preprocessing the characteristic parameters, penalty constraints, error estimation, and algorithm structure adjustment, the inherent laws of various types of factors and the construction timing and planned scale of the supporting capacity of the distribution network are excavated to achieve the deduction of the dynamic time series change process of the construction scale of the supporting capacity of the distribution network. In addition, in order to improve the sparsity and smoothing performance of the interpretable recurrent neural network, corresponding penalty constraints are imposed on the weight parameters of the network structure, so that the interpretable recurrent neural network improves the flexibility of the prediction algorithm structure while pursuing the interpretability and prediction accuracy of the prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 The present invention is a flow chart of a method for predicting the construction scale of the time sequence of the expansion capacity of the distribution network industry.

[0056] Figure 2 It is a schematic diagram of the distribution network industry expansion supporting capacity time series construction scale prediction model based on the interpretable recurrent neural network in the present invention.

[0057] Figure 3 It is a curve chart of the construction scale of the business expansion supporting capacity timing in the embodiment of the present invention. DETAILED DESCRIPTION

[0058] The present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0059] See also Figure 1 , Figure 2 A method for predicting the construction scale of distribution network expansion supporting capacity timing, comprising the following steps:

[0060] S1. Conduct a correlation analysis on the influencing indicators of the timing and scale of the construction of supporting capacity for business expansion, and find that the indicators with the greatest impact on supporting projects for business expansion are timing influencing factors and historical construction capacity;

[0061] S2. Screen out similar business expansion supporting capacity construction projects from the massive historical business expansion supporting capacity construction scale database, and extract the time series influencing factor data and historical construction capacity data of similar business expansion supporting capacity construction projects;

[0062] S3, first pre-processing the time series influencing factor data and historical construction capacity data extracted in step S2, then normalizing the data set, and then dividing the data set into a training set and a test set;

[0063] S4. Construct a distribution network industry expansion capacity time series construction scale prediction model based on an interpretable recurrent neural network, which includes an input layer, a linear layer, a nonlinear feature extraction layer and a linear regression layer;

[0064] The input layer is used to input the time series influencing factor data and historical construction capacity data of similar business expansion supporting capacity construction projects;

[0065] The linear layer is used to linearly combine the data of the input layer;

[0066] The nonlinear feature extraction layer uses Legendre polynomial to extract nonlinear features and uses it as input of the linear regression layer;

[0067] The linear regression layer is used to linearly combine the nonlinear features to form the final prediction result;

[0068] S5, applying corresponding penalty constraints to the weight parameters of the prediction model to obtain an improved prediction model;

[0069] S6, inputting the data of the training set in step S3 into the improved prediction model for training to obtain a trained prediction model, and inputting the data of the test set in step S3 into the trained prediction model to obtain a prediction result;

[0070] S7. Determine whether the deviation between the predicted result and the true value is within the allowable range. If so, output the predicted result; otherwise, return to step S4.

[0071] In step S1, the influencing indicators include load density, load utilization, average household distribution transformer capacity, electricity sales, daily maximum load and historical construction capacity; the time series influencing factors include load density, average household distribution transformer capacity and daily maximum load.

[0072] In step S2, similar business expansion supporting capacity construction projects are screened from a massive database of historical business expansion supporting capacity construction scales using the K-Means clustering algorithm.

[0073] In step S3, the interquartile range method is first used to detect outliers, and then the outliers are removed from the data set. Then the quadratic interpolation method is used to fill the data gaps. The Min-Max normalization method is then used to normalize the data set so that the data of different dimensions are distributed between [0-1]. Then the data set is divided into 65% as a training set and 35% as a test set.

[0074] In step S4, the input of the input layer is the feature vector x at the current moment t and the memory value vector h of the prediction model at the previous moment t-1 , sort the typical features of the time series influencing factor data and the historical construction capacity data to obtain a feature vector x t ,Right now:

[0075]

[0076] Predict the memory value vector h of the previous moment of the model t-1 for:

[0077]

[0078] In the formula, is the value of the i-th eigenvector at time t, d is the number of eigenvectors, is the value of the i-th hidden layer neuron at time t-1, and k is the number of hidden layer neurons;

[0079] The linear layer consists of k nodes, each of which performs linear weighting on the time series influencing factor data and the historical construction capacity data, and uses the linear weighted value of each node as the input of the corresponding nonlinear feature extraction layer, namely:

[0080]

[0081] In the formula, β j is the projection weight of the input data of the jth node of the linear layer, is the linear weighted value of the jth node in the linear layer at time t;

[0082] The nonlinear feature extraction layer is composed of k sub-networks. The input and output of each sub-network are single variables. The sub-networks are not connected to each other. The sub-networks form a mapping relationship with the nodes of the linear layer. The sub-network performs single-variable activation processing on the output of each node of the linear layer and outputs the corresponding activation value. And use it as the input of the corresponding linear regression layer; the univariate activation processing process is as follows:

[0083]

[0084] In the formula, g j (·) is the activation function corresponding to the jth node;

[0085] The linear regression layer consists of a single node. The input of the linear regression layer is the activation value of all nodes in the nonlinear feature extraction layer. The output is the value after linear transformation and linear combination of this layer, that is, the linear combination of the output activation function:

[0086]

[0087] In the formula, y t is the prediction result, γ k is the scaling parameter and μ is the shift parameter.

[0088] In step S5, the corresponding penalty constraint is imposed on the weight parameters of the prediction model as follows:

[0089] (1) Regularize the projection weights and scaling parameters:

[0090]

[0091]

[0092] In the formula, β ij is the projection weight between the i-th neuron in the input layer and the j-th neuron in the linear layer, γ j is the scaling parameter of the jth neuron; T1 and T2 are the regularization strengths of the parameters, T1 ≥ 1, T2 ≥ 0;

[0093] (2) Implement roughness penalty constraints on the activation function:

[0094] ∫[g″ j (z)] 2 dF j (z)≤T3

[0095] In the formula, z is the training value of the feature variable, Fj(·) is the cumulative distribution function of the j-th projection variable; T3 is the regularization strength of the parameter, T3 ≥ 0;

[0096] (3) Implement orthogonal constraints on projection weights:

[0097] β T β=I k

[0098] In the formula, I k is a k-order all-one matrix;

[0099] (4) Standardization constraints:

[0100] ∫g j (z)dF j (z)=0

[0101] ∫g j (z) 2 dF j (z)=1.

[0102] In step S7, the mean absolute error, root mean square error, coefficient of variation and coefficient of determination are used to comprehensively evaluate the prediction results:

[0103]

[0104]

[0105]

[0106]

[0107] In the formula, O j is the actual value of the construction scale of the distribution network industry expansion supporting capacity of the i-th sample, is the forecast value of the construction scale of the expansion capacity of the distribution network industry of the i-th sample, N is the number of samples, MAE is the mean absolute error, RMSE is the root mean square error, CV is the coefficient of variation, R 2 is the determination coefficient.

[0108] The principle of the present invention is described as follows:

[0109] The load density, average household distribution and transformation capacity, daily maximum load and other time-series influencing factors data and historical construction capacity data related to the business expansion and supporting projects are used as input features of the prediction method. Through the feature transformation and feature extraction of the designed prediction method, various influencing factors are quantitatively analyzed, and the inherent laws of the historical capacity construction timing and planned scale of business expansion and supporting facilities for various types of customers in different industries are excavated, so as to realize the monthly deduction of the timing and scale of the future business expansion and supporting capacity construction of distribution networks in various cities.

[0110] The interpretability of the prediction model and the complexity of the prediction model usually conflict with each other, but the more complex the prediction model is, the higher its prediction accuracy is. Therefore, in order to improve the accuracy of the prediction of the interpretable recurrent neural network algorithm without losing the interpretability of the algorithm, this design improves the sparsity and smoothing performance of the neural network by imposing corresponding penalty constraints on the weight parameters of the interpretable recurrent neural network structure, while pursuing the flexibility of network prediction accuracy and maintaining its good interpretability. In order to improve the sparsity of the network, the projection weights and scaling parameters are regularized; in order to improve the smoothing performance of the network, the roughness penalty constraint is imposed on the activation function; in order to reduce the impact of multicollinearity factors on model prediction and avoid the occurrence of exactly the same projection weights, orthogonal constraints need to be imposed on the projection weights; in order to improve the sparsity of the network, the projection weights and scaling parameters are processed, which may cause the activation function to be unrecognizable. Therefore, in order to enhance the identifiability of the network, appropriate normalization constraints are imposed.

[0111] Example:

[0112] See also Figure 1 A method for predicting the construction scale of distribution network expansion supporting capacity timing, comprising the following steps:

[0113] S1. Conduct a correlation analysis on the influencing indicators of the timing and scale of the construction of supporting capacity for business expansion, and find that the indicators with the greatest impact on supporting projects for business expansion are timing influencing factors and historical construction capacity;

[0114] The influencing indicators include load density, load utilization, average household distribution and transformation capacity, power sales, daily maximum load and historical construction capacity;

[0115] The SPSS software was used to conduct correlation analysis on the influencing indicators of the construction scale of the supporting capacity of business expansion, and the results are shown in Table 1;

[0116] Table 1. Correlation analysis results of indicators affecting the construction scale of business expansion supporting capacity timing

[0117]

[0118] Analysis of the Pearson correlation coefficient in Table 1 shows that the historical construction capacity is the factor that has the greatest impact on the prediction results, and the correlation coefficient between load utilization and electricity sales is the smallest. Therefore, the load density, average household distribution capacity, daily maximum load and other time-series influencing factor data and historical construction capacity data are selected as the input of the industry expansion supporting capacity time-series construction scale prediction algorithm;

[0119] S2. Screen out similar business expansion supporting capacity construction projects from the massive historical business expansion supporting capacity construction scale database through K-Means clustering algorithm, and extract the time series influencing factor data and historical construction capacity data of similar business expansion supporting capacity construction projects;

[0120] S3, first pre-processing the time series influencing factor data and historical construction capacity data extracted in step S2, then normalizing the data set, and then dividing the data set into a training set and a test set;

[0121] First, the interquartile range method is used to detect outliers, and then the outliers are removed from the data set. Then, the quadratic interpolation method is used to fill the data gaps. The Min-Max normalization method is used to normalize the data set so that the data of different dimensions are distributed between [0-1]. Then, the data set is divided into 65% as the training set and 35% as the test set.

[0122] S4. Construct a distribution network expansion capacity time series construction scale prediction model based on an interpretable recurrent neural network, see Figure 2 ,The prediction model includes an input layer, a linear layer, a nonlinear feature extraction layer and a linear regression layer;

[0123] The input layer is used to input the time series influencing factor data and historical construction capacity data of similar business expansion supporting capacity construction projects;

[0124] The input of the input layer is the feature vector x at the current moment t and the memory value vector h of the prediction model at the previous moment t-1 , sort the typical features of the time series influencing factor data and the historical construction capacity data to obtain a feature vector x t ,Right now:

[0125]

[0126] Predict the memory value vector h of the previous moment of the model t-1 for:

[0127]

[0128] In the formula, is the value of the i-th eigenvector at time t (i.e., the value of the i-th input neuron), d is the number of eigenvectors, is the value of the i-th hidden layer neuron at time t-1, and k is the number of hidden layer neurons;

[0129] The linear layer is used to linearly combine the data of the input layer;

[0130] The linear layer consists of k nodes, each of which performs linear weighting on the time series influencing factor data and the historical construction capacity data, and uses the linear weighted value of each node as the input of the corresponding nonlinear feature extraction layer, namely:

[0131]

[0132] In the formula, β j is the projection weight of the input data of the jth node of the linear layer, is the linear weighted value of the jth node in the linear layer at time t;

[0133] The nonlinear feature extraction layer uses Legendre polynomial to extract nonlinear features and uses it as input of the linear regression layer;

[0134] The nonlinear feature extraction layer is composed of k sub-networks. The input and output of each sub-network are single variables. The sub-networks are not connected to each other. The sub-networks and the nodes of the linear layer form a mapping relationship. In order to increase the nonlinearity of the neural network, the sub-network generally adopts a nonlinear activation function. The sub-network performs single-variable activation processing on the output of each node in the linear layer and outputs the corresponding activation value. And use it as the input of the corresponding linear regression layer; the univariate activation processing process is as follows:

[0135]

[0136] In the formula, g j (·) is the activation function corresponding to the jth node;

[0137] The linear regression layer is used to linearly combine the nonlinear features to form the final prediction result;

[0138] The linear regression layer (i.e., the output layer of the neural network) consists of a single node and generally uses a linear activation function. The input of the linear regression layer is the activation value of all nodes in the nonlinear feature extraction layer, and the output is the value after linear transformation and linear combination of this layer, that is, the linear combination of the output activation function:

[0139]

[0140] In the formula, y t is the prediction result, γ k is the scaling parameter, μ is the shift parameter;

[0141] S5, applying corresponding penalty constraints to the weight parameters of the prediction model to obtain an improved prediction model;

[0142] The corresponding penalty constraints are imposed on the weight parameters of the prediction model as follows:

[0143] (1) Regularize the projection weights and scaling parameters:

[0144]

[0145]

[0146] In the formula, β ijis the projection weight between the i-th neuron in the input layer and the j-th neuron in the linear layer, γ j is the scaling parameter of the jth neuron; T1 and T2 are the regularization strengths of the parameters, T1 ≥ 1, T2 ≥ 0;

[0147] (2) Implement roughness penalty constraints on the activation function:

[0148] ∫[g″ j (z)] 2 dF j (z)≤T3

[0149] In the formula, z is the training value of the feature variable, F j (·) is the cumulative distribution function of the j-th projection variable; T3 is the regularization strength of the parameter, T3 ≥ 0;

[0150] (3) Implement orthogonal constraints on projection weights:

[0151] β T β=I k

[0152] In the formula, I k is a k-order all-one matrix;

[0153] (4) Standardization constraints:

[0154] ∫g j (z)dF j (z)=0

[0155] ∫g j (z) 2 dF j (z) = 1;

[0156] S6, inputting the data of the training set in step S3 into the improved prediction model for training to obtain a trained prediction model, and inputting the data of the test set in step S3 into the trained prediction model to obtain a prediction result;

[0157] S7, determine whether the deviation between the predicted result and the true value is within the allowable range, if so, output the predicted result; otherwise, return to step S4;

[0158] The mean absolute error, root mean square error, coefficient of variation and coefficient of determination were used to comprehensively evaluate the prediction results:

[0159]

[0160]

[0161]

[0162]

[0163] In the formula, O j is the actual value of the construction scale of the distribution network industry expansion supporting capacity of the i-th sample, is the forecast value of the construction scale of the expansion capacity of the distribution network industry for the i-th sample, N is the number of samples; MAE is the mean absolute error, which measures the closeness of the forecast result to the actual value; RMSE is the root mean square error, which measures the concentration of data around the best fit line; CV is the coefficient of variation, which describes the fitness of the forecast model. The lower its value, the smaller the residual relative to the forecast result, indicating that the model fits better; R 2 is the determination coefficient. The larger its value is, the higher the fitting accuracy is.

[0164] Taking a residential community in a certain city as an example, the K-Means clustering algorithm is used to screen similar community business expansion supporting capacity construction projects from the massive historical business expansion supporting capacity construction database, and the time series influencing factor data and historical construction capacity data of similar business expansion supporting capacity construction projects are extracted. Based on the above-mentioned distribution network business expansion supporting capacity time series construction scale prediction method based on interpretable recurrent neural network, a simulation analysis is carried out with 12 months in a year as the business expansion supporting capacity construction period, such as Figure 3 shown.

[0165] The overall prediction performance of the method was evaluated by using the four commonly used indicators in statistics. The MAE was 52.08 kVA, the RMSE was 58.40 kVA, the CV was 5.47%, and the R 2 The results show that the proposed method for predicting the construction scale of distribution network expansion capacity based on interpretable recurrent neural network has a high prediction accuracy.

Claims

1. A method for predicting the construction scale of distribution network expansion capacity timing, characterized in that: The following steps are involved: S1. Conduct a correlation analysis on the influencing indicators of the timing and scale of the construction of supporting capacity for business expansion, and find that the indicators with the greatest impact on supporting projects for business expansion are timing influencing factors and historical construction capacity; S2. Screen out similar business expansion supporting capacity construction projects from the massive historical business expansion supporting capacity construction scale database, and extract the time series influencing factor data and historical construction capacity data of similar business expansion supporting capacity construction projects; S3, first pre-processing the time series influencing factor data and historical construction capacity data extracted in step S2, then normalizing the data set, and then dividing the data set into a training set and a test set; S4. Construct a distribution network industry expansion capacity time series construction scale prediction model based on an interpretable recurrent neural network, which includes an input layer, a linear layer, a nonlinear feature extraction layer and a linear regression layer; The input layer is used to input the time series influencing factor data and historical construction capacity data of similar business expansion supporting capacity construction projects; The linear layer is used to linearly combine the data of the input layer; The nonlinear feature extraction layer uses Legendre polynomial to extract nonlinear features and uses it as input of the linear regression layer; The linear regression layer is used to linearly combine the nonlinear features to form the final prediction result; S5, applying corresponding penalty constraints to the weight parameters of the prediction model to obtain an improved prediction model; The corresponding penalty constraints are imposed on the weight parameters of the prediction model as follows: (1) Regularize the projection weights and scaling parameters: In the formula, β ij is the projection weight between the i-th neuron in the input layer and the j-th neuron in the linear layer, γ j is the scaling parameter of the jth neuron; T1 and T2 are the regularization strengths of the parameters, T1 ≥ 1, T2 ≥ 0; (2) Implement roughness penalty constraints on the activation function: ∫[g" j (z)] 2 dF j (z)≤T3 In the formula, z is the training value of the feature variable, F j (·) is the cumulative distribution function of the j-th projection variable; T3 is the regularization strength of the parameter, T3 ≥ 0; (3) Implement orthogonal constraints on projection weights: β T β＝I k In the formula, I k is a k-order all-one matrix; (4) Standardization constraints: ∫g j (z)dF j (2)=0 ∫g j (z) 2 dF j (z)=1 S6, inputting the data of the training set in step S3 into the improved prediction model for training to obtain a trained prediction model, and inputting the data of the test set in step S3 into the trained prediction model to obtain a prediction result; S7. Determine whether the deviation between the predicted result and the true value is within the allowable range. If so, output the predicted result; otherwise, return to step S4.

2. A method for predicting the scale of construction of distribution network expansion capacity in a time sequence according to claim 1, characterized in that: In step S1, the influencing indicators include load density, load utilization, average household distribution transformer capacity, electricity sales, daily maximum load and historical construction capacity; the time series influencing factors include load density, average household distribution transformer capacity and daily maximum load.

3. A method for predicting the scale of construction of distribution network expansion capacity in a time sequence according to claim 1, characterized in that: In step S2, similar business expansion supporting capacity construction projects are screened from a massive database of historical business expansion supporting capacity construction scales using the K-Means clustering algorithm.

4. A method for predicting the scale of construction of distribution network expansion capacity in a time sequence according to claim 1, characterized in that: In step S3, the interquartile range method is first used to detect outliers, and then the outliers are removed from the data set. Then the quadratic interpolation method is used to fill the data gaps. The Min-Max normalization method is then used to normalize the data set so that the data of different dimensions are distributed between [0-1]. Then the data set is divided into 65% as a training set and 35% as a test set.

5. The method for predicting the construction scale of distribution network expansion capacity in a time sequence according to claim 1 is characterized by: In step S4, the input of the input layer is the feature vector x at the current moment t and the memory value vector h of the prediction model at the previous moment t-1 , sort the typical features of the time series influencing factor data and the historical construction capacity data to obtain a feature vector x t ,Right now: Predict the memory value vector h of the previous moment of the model t-1 for: In the formula, is the value of the i-th eigenvector at time t, d is the number of eigenvectors, is the value of the i-th hidden layer neuron at time t-1, and k is the number of hidden layer neurons; The linear layer consists of k nodes, each of which performs linear weighting on the time series influencing factor data and the historical construction capacity data, and uses the linear weighted value of each node as the input of the corresponding nonlinear feature extraction layer, namely: In the formula, β j is the projection weight of the input data of the jth node of the linear layer, is the linear weighted value of the jth node in the linear layer at time t; The nonlinear feature extraction layer is composed of k sub-networks. The input and output of each sub-network are single variables. The sub-networks are not connected to each other. The sub-networks form a mapping relationship with the nodes of the linear layer. The sub-network performs single-variable activation processing on the output of each node of the linear layer and outputs the corresponding activation value. And use it as the input of the corresponding linear regression layer; The univariate activation process is as follows: In the formula, g j (·) is the activation function corresponding to the jth node; The linear regression layer consists of a single node. The input of the linear regression layer is the activation value of all nodes in the nonlinear feature extraction layer. The output is the value after linear transformation and linear combination of this layer, that is, the linear combination of the output activation function: In the formula, y t is the prediction result, γ k is the scaling parameter and μ is the shift parameter.

6. A method for predicting the scale of construction of distribution network expansion capacity in a time sequence according to claim 1, characterized in that: In step S7, the mean absolute error, root mean square error, coefficient of variation and coefficient of determination are used to comprehensively evaluate the prediction results: In the formula, O i is the actual value of the construction scale of the distribution network industry expansion supporting capacity of the i-th sample, is the forecast value of the construction scale of the expansion capacity of the distribution network industry of the i-th sample, N is the number of samples, MAE is the mean absolute error, RMSE is the root mean square error, CV is the coefficient of variation, R 2 is the determination coefficient.

Images