A regional electricity consumption prediction method and terminal based on a mixture model

Abstract

The application discloses a regional power consumption prediction method and terminal based on a hybrid model, and the method comprises the following steps: obtaining historical power consumption data of a target region; performing data preprocessing on the historical power consumption data to obtain an original data set; performing data processing on the original data set by an empirical mode decomposition model to obtain a combined data set; constructing a power consumption prediction model based on a convolutional neural network model and a double echo state network model, and training the power consumption prediction model according to the combined data set; and predicting future power consumption of the target region by the trained power consumption prediction model. The application effectively fuses the spatial feature extraction capability of the convolutional neural network and the time sequence feature prediction capability of the echo state network, so that the power consumption prediction model can effectively learn the spatial features and the time features from the historical power consumption data. Compared with a single prediction model, the power consumption prediction model of the application effectively improves the prediction precision.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of power prediction, in particular to a regional power consumption prediction method based on a hybrid model and a terminal. BACKGROUND

[0002] With the development of smart grids and current measurement infrastructure, short-term power consumption prediction is attracting more and more attention. In fact, the prediction of future power load is a key method to avoid energy waste and develop effective power management strategies. Therefore, balancing power production and consumption is an urgent task, and its implementation largely depends on the means and methods of power production planning. Among them, accurate prediction of power demand is a key factor in optimizing power generation and consumption, saving energy and determining energy prices. However, the integration of highly nonlinear and seasonal energy combination scenarios, including meteorological environment, into the existing power grid increases the uncertainty of power generation and brings additional challenges to accurate prediction. Early prediction methods based on machine learning are prone to data overfitting or poor prediction accuracy. SUMMARY

[0003] The technical problem to be solved by the application is to provide a regional power consumption prediction method based on a hybrid model and a terminal, which can realize accurate prediction of regional power consumption.

[0004] To solve the above technical problems, the technical scheme adopted by the application is:

[0005] A regional power consumption prediction method based on a hybrid model, comprising:

[0006] acquiring historical power consumption data of a target region;

[0007] performing data preprocessing on the historical power consumption data to obtain an original data set;

[0008] performing data processing on the original data set by an empirical mode decomposition model to obtain a combined data set;

[0009] constructing a power consumption prediction model based on a convolutional neural network model and a double echo state network model, and training the power consumption prediction model according to the combined data set;

[0010] predicting future power consumption of the target region by the trained power consumption prediction model.

[0011] To solve the above technical problems, another technical scheme adopted by the application is:

[0012] A regional power consumption prediction terminal based on a hybrid model comprises a memory, a processor and a computer program stored on the memory and running on the processor, and each step in the above-mentioned regional power consumption prediction method based on a hybrid model is implemented when the processor executes the computer program.

[0013] The present application has the beneficial effects that: the historical power consumption data of a target region is acquired, and an original data set is constructed according to the historical power consumption data, and the original data set is processed by an empirical mode decomposition model, so that the original data set is screened, random errors are avoided, and the accuracy of subsequent model training is improved. At the same time, the power consumption prediction model is constructed based on a convolutional neural network model and a dual echo state network model, effectively fusing the spatial feature extraction ability of the convolutional neural network and the time sequence feature prediction ability of the echo state network, so that the power consumption prediction model can effectively learn the spatial features and time features from the historical power consumption data. Compared with a single prediction model, the power consumption prediction model of the present application effectively improves the prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 A step flowchart of a regional power consumption prediction method based on a hybrid model provided by the embodiment of the present application is provided;

[0015] Figure 2 Another flowchart of a regional power consumption prediction method based on a hybrid model provided by the embodiment of the present application is provided;

[0016] Figure 3 An architecture schematic diagram of a dual echo state network model provided by the embodiment of the present application is provided;

[0017] Figure 4 A structure schematic diagram of a regional power consumption prediction terminal based on a hybrid model provided by the embodiment of the present application is provided;

[0018] Label explanation:

[0019] 100. A regional power consumption prediction terminal based on a hybrid model; 101. a memory; 102. a processor. DETAILED DESCRIPTION

[0020] To explain the technical content, purposes and effects of the present application in detail, the following will be explained in combination with the embodiments and the drawings.

[0021] The embodiment of the present application provides a regional power consumption prediction method based on a hybrid model, comprising:

[0022] Acquiring historical power consumption data of a target region;

[0023] Performing data preprocessing according to the historical power consumption data to obtain an original data set;

[0024] data processing on the original data set by an empirical mode decomposition model to obtain a combined data set;

[0025] An electricity consumption prediction model is constructed based on a convolutional neural network model and a dual echo state network model, and the electricity consumption prediction model is trained according to the combined data set;

[0026] The future electricity consumption of the target area is predicted by the trained electricity consumption prediction model.

[0027] From the above description, the beneficial effects of the present application are that the historical electricity consumption data of the target area is obtained, and an original data set is constructed according to the historical electricity consumption data, and the original data set is processed by an empirical mode decomposition model (EMD) to filter the data of the original data set, avoid random errors, and improve the accuracy of subsequent model training. At the same time, the present application constructs an electricity consumption prediction model based on a convolutional neural network model (CNN) and a dual echo state network model (DESN), effectively fusing the spatial feature extraction capability of the convolutional neural network and the time sequence feature prediction capability of the echo state network, so that the electricity consumption prediction model can effectively learn the spatial features and time features from the historical electricity consumption data. Compared with a single prediction model, the electricity consumption prediction model of the present application effectively improves the prediction accuracy.

[0028] Further, the data preprocessing according to the historical electricity consumption data to obtain an original data set comprises:

[0029] The abnormal values in the historical electricity consumption data are removed by a moving average method, and the historical electricity consumption data is normalized to obtain standard data;

[0030] The abnormal values in the standard data are removed by an interquartile range method to obtain an original data set.

[0031] From the above description, the historical electricity consumption data may contain abnormal values with sudden energy consumption peaks, which will make it difficult for subsequent models to predict future energy consumption. Therefore, when constructing the original data set, the abnormal values in the historical electricity consumption data should be removed as much as possible to reduce the random errors of the original data set and improve the training accuracy of the subsequent model.

[0032] Further, the data processing on the original data set by an empirical mode decomposition model to obtain a combined data set comprises:

[0033] The original data set is decomposed by an empirical mode decomposition model to obtain an initial feature set, and the correlation coefficient of each feature component in the initial feature set with the original data set is calculated based on a Pearson correlation coefficient method;

[0034] Select the feature components with the correlation coefficient greater than the preset threshold as a target feature set;

[0035] Merge the target feature set with the original data set to obtain a combined data set.

[0036] As can be seen from the above description, by decomposing the original data set through the EMD model, an initial feature set composed of the decomposed feature components is obtained, and then the decomposed feature components are analyzed for correlation based on the Pearson correlation coefficient method, and the feature components with greater correlation with the original data set are selected as the input data of the subsequent power prediction model, which can effectively avoid the occurrence of random errors and improve the prediction accuracy of the model.

[0037] Further, the power prediction model includes a plurality of hybrid prediction models with different reserve pool spectral radii and a fully connected layer;

[0038] The power prediction model is constructed based on the convolutional neural network model and the dual echo state network model, and the power prediction model is trained according to the combined data set, which includes:

[0039] The convolutional neural network model is fused with a plurality of dual echo state network models with different reserve pool spectral radii to obtain a plurality of hybrid prediction models;

[0040] The combined data set is processed through the hybrid prediction model to obtain the output data of each hybrid prediction model;

[0041] The output data of each hybrid prediction model is fused through the fully connected layer to obtain target data;

[0042] The hybrid prediction model and the fully connected layer are iteratively optimized according to the combined data set and the target data.

[0043] As can be seen from the above description, the CNN model can effectively extract the spatial features of the data, and the DESN model can effectively extract and process the forward and backward dependent features of the data in the time sequence, so the hybrid prediction model (CDESN) constructed based on the CNN model and the DESN model can effectively learn the spatial features and the time features of the data, thereby accurately predicting the future data. At the same time, since the state of the echo state network with a larger reserve pool spectral radius is unstable, and the network memory capacity of the echo state network with a smaller spectral radius is weak, a plurality of CDESN models with different reserve pool spectral radii are constructed, and finally the output data of each CDESN model is fused through the fully connected layer for processing, so as to eliminate the errors caused by different spectral radii and improve the prediction accuracy of the model.

[0044] Further, the dual echo state network model comprises an encoding echo state network, a decoding echo state network and a latent connection layer.

[0045] The combined data set is respectively input into the convolutional neural network model in each hybrid prediction model to extract initial data features.

[0046] The combined data set is respectively input into the convolutional neural network model in each hybrid prediction model to extract initial data features.

[0047] The initial data features are input into the encoding echo state network for feature dimension reduction to obtain reduced data features.

[0048] The reduced data features are input into the decoding echo state network through the latent connection layer for data decoding to obtain output data of each hybrid prediction model.

[0049] As can be seen from the above description, the DESN model of the present application comprises two parts of an encoding network and a decoding network, that is, one ESN network is used for encoding and one ESN network is used for decoding. The encoding part maps the combined data set to a latent space of a preset dimension, thereby realizing feature dimension reduction, and the decoding part reconstructs the reduced features from the latent space and brings them back to the original space for output. Compared with the traditional ESN model, the DESN model of the present application can effectively learn the high-level feature representation of data, thereby better capturing the internal structural features of data.

[0050] Further, the initial data features are input into the encoding echo state network for feature dimension reduction to obtain reduced data features.

[0051]

[0052] wherein x(t+1) represents the state of the reservoir of the encoding echo state network at t+1, f ε represents the activation function of the encoding echo state network, W in represents the input weight of the encoding echo state network, u(t+1) represents the initial data features input into the reservoir of the encoding echo state network at t+1, represents the weight of the reservoir of the encoding echo state network, x(t) represents the state of the reservoir of the encoding echo state network at t, b ε represents the bias vector of the encoding echo state network, z(t+1) represents the reduced data features output from the reservoir of the encoding echo state network at t+1, represents the input weight of the latent connection layer.

[0053] As can be seen from the above description, the state of the reservoir pool of the encoding echo state network depends on the initial data features of the input, the state at the previous time and the reservoir pool weight of the current encoding echo network, and the reduced dimension data features of the encoding echo state network are obtained based on the state of the reservoir pool at the current time and the input weight of the potential connection layer. The input weight of the potential connection layer is iteratively optimized, so as to force the reservoir pool weight of the encoding echo state network to be optimized, thereby completing the optimization training of the encoding echo state network.

[0054] Further, the reduced dimension data features are input into the decoding echo state network through the potential connection layer for data decoding to obtain the output data of each hybrid prediction model, and the output data is specifically:

[0055]

[0056] Wherein, represents the state of the reservoir pool of the decoding echo state network at t+1 time, f D represents the activation function of the decoding echo state network, represents the output weight of the potential connection layer, and z(t+1) represents the reduced dimension data features output by the reservoir pool of the encoding echo state network at t+1 time, represents the weight of the reservoir pool of the decoding echo state network, represents the state of the reservoir pool of the decoding echo state network at t time, b D represents the bias vector of the decoding echo state network, represents the output data of the reservoir pool of the decoding echo state network at t+1 time, W out represents the output weight of the decoding echo state network.

[0057] As can be seen from the above description, the decoding echo state network decodes the reduced dimension data features based on the reservoir pool weight thereof to determine the state of the reservoir pool of the current decoding echo state network. Then, the output data is reconstructed by combining the state of the reservoir pool of the current decoding echo state network and the output weight obtained by the current training. The potential connection layer serves as a bridge connecting the encoding network and the decoding network, effectively adjusts the weight parameters between the two networks, so that the model parameters are optimized.

[0058] Further, the historical power consumption data is normalized to obtain standard data, and the standard data is specifically:

[0059]

[0060] Denormalized(a)=Normalized(a)*(max(a)-min(a))+min(a);

[0061] Normalized(a) represents standard data, x(a) represents historical power consumption data, min(a) and max(a) represent the minimum and maximum values of historical power consumption data respectively, and Denormalized(a) represents converting the standard data to the original numerical size.

[0062] As can be seen from the above description, since different types of historical power consumption data can have different dimensions and dimension units, the historical power consumption data is limited within a certain range by normalizing the historical power consumption data, thereby eliminating the adverse effects caused by singular sample data.

[0063] Further, the abnormal values in the standard data are removed by the interquartile range method to obtain the original data set, specifically:

[0064] The upper quartile and the lower quartile of the standard data are calculated.

[0065] The difference between the upper quartile and the lower quartile is calculated to obtain the interquartile range.

[0066] The abnormal value range is determined according to the upper quartile, the lower quartile and the interquartile range.

[0067] The abnormal values in the standard data are determined according to the abnormal value range and removed to obtain the original data set.

[0068] As can be seen from the above description, after data normalization, the interquartile range method can effectively identify abnormal values with large deviations, thereby removing abnormal values in the standard data that are ignored due to omission errors or deviation from the normal statistical distribution of the standard data.

[0069] Another embodiment of the present application provides a regional power consumption prediction terminal based on a hybrid model, which comprises a memory, a processor and a computer program stored in the memory and running on the processor, and the processor implements each step of the above-mentioned regional power consumption prediction method based on a hybrid model when executing the computer program.

[0070] From the above description, the beneficial effects of the present application are that the historical power consumption data of the target area is obtained, and the original data set is constructed according to the historical power consumption data, and the original data set is processed through an empirical mode decomposition model (EMD), so that the original data set is screened, random errors are avoided, and the accuracy of subsequent model training is improved. At the same time, the power consumption prediction model is constructed based on a convolutional neural network model (CNN) and a double echo state network model (DESN), effectively fusing the spatial feature extraction ability of the convolutional neural network and the time sequence feature prediction ability of the echo state network, so that the power consumption prediction model can effectively learn the spatial features and time features from the historical power consumption data. Compared with a single prediction model, the power consumption prediction model of the present application effectively improves the prediction accuracy.

[0071] Embodiments of the present application provide a region power consumption prediction method and terminal based on a hybrid model, which can be applied to power load prediction of a power system, and realize accurate prediction of region power consumption to help power system operators to reasonably plan and manage loads and power generation. The following will be described through specific embodiments:

[0072] Please refer to Figure 1 and Figure 2 Embodiments of the present application are:

[0073] A region power consumption prediction method based on a hybrid model, comprising:

[0074] S1, obtaining historical power consumption data of a target area.

[0075] In some embodiments, the historical power consumption data includes historical power consumption and related data corresponding to the historical power consumption. The historical power consumption is the power consumption of the target area in each week and each period of the day. The related data includes power consumption analysis data and meteorological data affecting power consumption. The analysis data includes power consumption trend, seasonal fluctuation trend, holiday power consumption effect, etc. The meteorological data includes temperature, etc.

[0076] S2, data preprocessing according to the historical power consumption data to obtain an original data set.

[0077] Specifically, the step S2 comprises:

[0078] S21, removing outliers in the historical power consumption data through a moving average method, and normalizing the historical power consumption data to obtain standard data.

[0079] The normalization of the historical power consumption data to obtain standard data is specifically:

[0080]

[0081] Denormalized(a) = Normalized(a) * (max(a) - min(a)) + min(a) (2)

[0082] wherein, Normalized(a) represents standard data, x(a) represents historical power consumption data, min(a) and max(a) represent minimum value and maximum value of the historical power consumption data respectively, and Denormalized(a) represents converting the standard data into the original numerical size, and the historical power consumption data can be normalized to the range of 0-1 through the above normalization processing formula.

[0083] S22, removing outliers in the standard data through the interquartile range method to obtain an original data set.

[0084] It should be noted that the interquartile range refers to the difference between the fourth percentile of data (i.e., the upper quartile) and the third percentile of data (i.e., the lower quartile). Therefore, the interquartile range of data will follow a 25% decomposition point.

[0085] In some embodiments, the upper quartile, the lower quartile and the interquartile range of data are calculated based on the interquartile range method, and it can be determined which data are considered outliers according to a preset determination rule.

[0086] In an alternative implementation, the step S22 is specifically:

[0087] S221, calculating the upper quartile and the lower quartile of the standard data.

[0088] S222, calculating the difference between the upper quartile and the lower quartile to obtain the interquartile range.

[0089] Specifically, the interquartile range IQR = Q3 - Q1, wherein Q3 represents the upper quartile and Q1 represents the lower quartile.

[0090] S223, determining an outlier range according to the upper quartile, the lower quartile and the interquartile range.

[0091] S224, determining outliers in the standard data according to the outlier range and removing the outliers to obtain an original data set.

[0092] In some embodiments, data less than Q1 - 1.5IQR or greater than Q3 + 1.5IQR can be determined as outliers based on the upper quartile Q3, the lower quartile Q1 and the interquartile range IQR.

[0093] S3, performing data processing on the original data set through an empirical mode decomposition model to obtain a combined data set.

[0094] Specifically, as shown in Figure 2 step S3 includes:

[0095] S31, data decomposition is performed on the original data set by an empirical mode decomposition model to obtain an initial feature set.

[0096] In some embodiments, data decomposition is performed on the original data set X s by an EMD model to obtain an initial feature set X, X={imf1, imf2, imf3, ……imf n-1 n}, where imf i and Res respectively represent an IMF component (Intrinsic Mode Functions) and a residual of decomposition, and n represents the total number of time series components of decomposition.

[0097] S33, a correlation coefficient of each feature component in the initial feature set and the original data set is calculated based on a Pearson correlation coefficient method.

[0098] As a filtering feature selection method, the Pearson correlation coefficient has strong universality and low complexity, and has a strong advantage in processing large-scale data sets, and can eliminate a large number of irrelevant features in a short time. Therefore, the Pearson correlation coefficient method is used to select and filter the initial feature set obtained by EMD decomposition, and the calculation formula of the Pearson correlation coefficient is specifically:

[0099]

[0100] where γ represents the correlation coefficient, that is, the correlation between different feature components and the original data set. x i and y i respectively represent different feature components and the original data set, and respectively represent the average values of different feature components and the original data set, and n represents the sample number.

[0101] S34, a feature component with a correlation coefficient greater than a preset threshold is selected as a target feature set.

[0102] In some embodiments, if the correlation coefficient r i of a feature component is greater than a preset threshold t, the feature component is added to the target feature set X f ; if the correlation coefficient r i of a feature component is less than or equal to the preset threshold t, the feature component is removed, thereby filtering the feature components with a correlation less than the threshold in the initial feature set.

[0103] S35, merging the target feature set and the original data set to obtain a combined data set.

[0104] In some embodiments, the combined data set Data={X s , X f}.

[0105] Empirical Mode Decomposition (EMD) is a signal processing technique that can decompose non-stationary signals into a series of Intrinsic Mode Functions (IMF). The present application uses EMD to decompose the collected original data set to extract its inherent patterns and trends. EMD can decompose signals of different frequencies step by step according to the characteristics of the data itself, thereby obtaining several periodic and trend orthogonal signals. Currently, the general prediction method based on EMD is to establish a prediction model for each decomposed subsequence, and then superimpose the output of each prediction model to obtain the final prediction result. Although this method improves the prediction accuracy, it introduces multiple random errors. At this time, if a prediction model is established for each subsequence, there will be some high-frequency noise components in the decomposed subsequences, which will lead to large prediction errors when modeling these subsequences, affecting the overall prediction accuracy. At the same time, since multiple prediction models are established, the overall complexity of the model will increase. The EMD model of the present application performs correlation analysis on the decomposed subsequences, thereby selecting components with high correlation with the original sequence as input features of the model, to avoid random errors and improve the accuracy of the subsequent prediction model.

[0106] S4, constructing an electricity consumption prediction model based on a convolutional neural network model and a dual echo state network model, and training the electricity consumption prediction model according to the combined data set.

[0107] S5, predicting the future electricity consumption of the target area by the trained electricity consumption prediction model.

[0108] In some embodiments, based on the trained electricity consumption prediction model, the electricity consumption of the target area on the same day can be predicted, i.e., the short-term energy consumption of the target area is predicted.

[0109] Please refer to Figure 2 and Figure 3 , embodiment two of the present application is:

[0110] A regional electricity consumption prediction method based on a hybrid model, which is different from embodiment one in that the specific steps of step S4 are limited.

[0111] Wherein, the electricity consumption prediction model comprises a plurality of hybrid prediction models with different reserve pool spectral radii and a fully connected layer, and the hybrid prediction model is represented as C-DESN, and the step S4 comprises:

[0112] S41, fuse the convolutional neural network model with multiple dual echo state network models with different reservoir spectral radii to obtain multiple hybrid prediction models.

[0113] Specifically, the dual echo state network model (DESN) of the application comprises two echo state networks (ESN), one of which is used for encoding mapping of data, and the other is used for decoding mapping of data, and a trainable code layer is introduced between the two ESNs for generating an optimized latent representation of the input. Therefore, the dual echo state network model comprises an encoding echo state network, a decoding echo state network and a latent connection layer. As shown in the figure, the dual echo state network model (DESN) also comprises an input layer and an output layer, the input layer being connected in sequence with the encoding echo state network (hereinafter referred to as the encoding part), the latent connection layer, the decoding echo state network (hereinafter referred to as the decoding part) and the output layer. Among them, the encoding part maps the original data to an m-dimensional latent space (m is less than the dimension of the original input data), the decoding part reconstructs the input from the latent space and brings it back to the original space, and the latent connection layer is a trainable code layer. Figure 3

[0114] Among them, ESN is a special type of recurrent neural network (RNN), which replaces the hidden layer in RNN with a reservoir. Compared with RNN model, ESN model uses a fixed recurrent neural network as a storage and uses a linear learning method to simply modify the output weight matrix, which effectively improves the calculation efficiency.

[0115] Each ESN model comprises a memory, an input layer and an output layer. If the number of input neurons is represented as K, the number of storage neurons is represented as N, and the number of output neurons is represented as L, W in , W, W back , and W out represent the weights of the input matrix, the storage matrix, the feedback matrix and the output matrix respectively, then the sizes of the input matrix, the storage matrix, the feedback matrix and the output matrix are N×K, N×N, N×L and L×(K+N) in turn. The ESN model is trained through a supervised learning process, and the training process mainly includes updating the reservoir state and converting the weight matrix of the reservoir to the weight matrix W out of the output layer.

[0116] Let u=u(n) represent the external input of the ESN model, x=x(n) represent the storage state of the ESN model, and y=y(n) represent the output vector of the ESN model. For each ESN model, the calculation process of the input signal from the input layer to the reservoir is as follows:

[0117] x(t+1)=f(W in u(t+1), Wx(t)+W​back y(t)) (4)

[0118] The output of the ESN model at t+1 is:

[0119] y(t+1)=f out (W out x(t+1)) (5)

[0120] where f denotes the sigmoid function within the reserve layer, and f out denotes the output layer activation function.

[0121] Thus, the DESN model can be represented as:

[0122] z=f ε (W ε u+b ε ) (6)

[0123]

[0124] where formula (6) describes the calculation method of the encoding part in the DESN model, f ε denotes the activation function of the encoding part, and W ε denotes the weight matrix of the encoding part. Formula (7) describes the calculation method of the decoding part in the DESN model, f D denotes the activation function of the decoding part, and W D denotes the weight matrix of the decoding part. Formula (8) describes the loss function of the DESN model which is calculated by the norm between the original input u and the reconstructed input .

[0125] Based on the architecture setting of the DESN model of the present application, it has the advantage of keeping most of the weights in the network fixed when training the network to reconstruct the sequence passed as input.

[0126] S42, respectively, the combined data set is input into the convolutional neural network model in each of the hybrid prediction models to extract initial data features.

[0127] Specifically, the step S42 includes:

[0128] S421, the combined data set is input into the convolutional neural network model in each of the hybrid prediction models to extract initial data features.

[0129] In some embodiments, the convolutional neural network model (CNN) includes two 1-dimensional convolutional layers, one pooling layer, and one Flatten layer. The corresponding initial data features are extracted from the lag sequence by the CNN model.

[0130] In some embodiments, since the initial data features extracted by the CNN model are two-dimensional data, and the ESN model is for processing time series, the initial data features need to be converted from two-dimensional data to one-dimensional data before being input into the encoding ESN.

[0131] S422, inputting the initial data features into the encoding echo state network for feature dimension reduction to obtain reduced dimension data features.

[0132] The step S422 specifically includes:

[0133]

[0134]

[0135] wherein x(t+1) represents the state of the reservoir of the encoding echo state network at t+1, f ε represents the activation function of the encoding echo state network, W in represents the input weight of the encoding echo state network, u(t+1) represents the input data of the reservoir of the encoding echo state network at t+1, represents the weight of the reservoir of the encoding echo state network, x(t) represents the state of the reservoir of the encoding echo state network at t, b ε represents the bias vector of the encoding echo state network, z(t+1) represents the reduced dimension data features output by the reservoir of the encoding echo state network at t+1, represents the input weight of the latent connection layer.

[0136] S423, inputting the reduced dimension data features through the latent connection layer into the decoding echo state network for data decoding to obtain the output data of each mixed prediction model.

[0137] The step S423 specifically includes:

[0138]

[0139] wherein, represents the state of the reservoir of the decoding echo state network at t+1, f D represents the activation function of the decoding echo state network, represents the output weight of the latent connection layer, z(t+1) represents the reduced dimension data features output by the reservoir of the encoding echo state network at t+1, represents the weight of the reservoir of the decoding echo state network, represents the state of the reservoir of the decoding echo state network at t, b Da bias vector representing a decoding echo state network, output data of a reservoir pool of the decoding echo state network at t+1, out an output weight of the decoding echo state network.

[0140] It should be noted that the DESN model of the present application uses two separate ESN models, so that the encoding process and the decoding process of the DESN model are decoupled, thereby allowing the weights are responsible for learning the input dynamics and their encoding, while the weights are used for decoding and input reconstruction. A latent connection layer is introduced as a bridge between the encoding part and the decoding part, and the weights in the latent connection layer are trained to force the encoding part to generate an optimized encoding representation. The number of neurons in the latent connection layer is fixed at m to control the number of encoding features.

[0141] S43, fusing the output data of each of the hybrid prediction models through a full connection layer to obtain target data.

[0142] S44, iteratively optimizing the hybrid prediction models and the full connection layer according to the combined data set and the target data.

[0143] As Figure 2 shown, the present application constructs an original data set according to historical power consumption data, and decomposes the original data set through an EMD model to obtain an initial feature set composed of IMF components and a residual Res. The correlation between each feature component in the initial feature set and the original data set is determined based on the Pearson correlation coefficient method to select feature components with greater correlation with the original data set as a target feature set. The original data set and the target feature set are merged to obtain a combined data set. At the same time, the combined data set is input into a C-DESN model obtained by fusing a CNN model and a DESN model with different reservoir pool spectral radii, and the prediction results of each C-DESN model are fused through a full connection layer to obtain the final power consumption prediction result. The C-DESN model and the full connection layer are iteratively optimized based on the prediction result to obtain an optimal power consumption prediction model.

[0144] Please refer to Figure 4 , embodiment three of the present application is:

[0145] A regional power consumption prediction terminal 100 based on a hybrid model, comprising a memory 101, a processor 102, and a computer program stored on the memory 101 and running on the processor 102, wherein the processor 102 implements each step of the regional power consumption prediction method based on a hybrid model according to the computer program.

[0146] To sum up, the application provides a regional power consumption prediction method and terminal based on a hybrid model, historical power consumption data of a target region is obtained, and an original data set is constructed according to the historical power consumption data, data processing is performed on the original data set by an EMD model to obtain a target feature set with relatively large correlation with the original data set, and the original data set and the target feature set are merged to obtain a combined data set, so that random errors are avoided and the accuracy of subsequent model training is improved. At the same time, the combined data set is input into a C-DESN model obtained by fusing a CNN model and a DESN model with different reserve pool spectral radii for processing, and the prediction results of each C-DESN model are fused by a full connection layer to obtain a final prediction result. The C-DESN model and the full connection layer are iteratively optimized based on the prediction result, so as to obtain an optimal power consumption prediction model. The power consumption prediction model combines the CNN model and the DESN model, can not only effectively learn spatial features from the combined data set, but also extract time features of data. In addition, a double reserve pool network is introduced into the DESN model to perform encoding and decoding operations, so that the model can better learn high-level feature representation of data and better capture the internal structure of data. Compared with a single prediction model, the power consumption prediction model effectively improves the prediction accuracy.

[0147] The above is only an embodiment of the application, and does not limit the patent scope of the application, and any equivalent transformation or direct or indirect application in the related technical field based on the content of the specification and drawings of the application is also included in the patent protection scope of the application.

Claims

1. A method for predicting regional electricity consumption based on a hybrid model, characterized in that: include: Obtain historical electricity consumption data for the target area; Performing data preprocessing on the historical electricity consumption data to obtain an original data set; Processing the original data set by an empirical mode decomposition model to obtain a combined data set; Building a power prediction model based on a convolutional neural network model and a double echo state network model, and training the power prediction model according to the combined data set; Predicting the future electricity consumption of the target area using the trained electricity prediction model; The power prediction model includes a hybrid prediction model with multiple reserve pools of different spectrum radius and a fully connected layer; The constructing of the power prediction model based on the convolutional neural network model and the double echo state network model, and training the power prediction model according to the combined data set includes: The convolutional neural network model is fused with multiple double echo state network models with different reservoir spectral radii to obtain multiple hybrid prediction models; Performing feature learning and processing on the combined data set through the hybrid prediction model respectively to obtain output data of each hybrid prediction model; The output data of each hybrid prediction model are fused through a fully connected layer to obtain target data; The hybrid prediction model and the fully connected layer are iteratively optimized according to the combined data set and the target data.

2. The method according to claim 1, characterized in that The performing data preprocessing according to the historical electricity consumption data to obtain the original data set includes: Eliminating abnormal values ​​in the historical electricity consumption data by a moving average method, and normalizing the historical electricity consumption data to obtain standard data; The outliers in the standard data were eliminated by the interquartile range method to obtain the original data set.

3. The method according to claim 1, characterized in that The step of processing the original data set by using an empirical mode decomposition model to obtain a combined data set includes: Performing data decomposition on the original data set using an empirical mode decomposition model to obtain an initial feature set; Calculate the correlation coefficient between each feature component in the initial feature set and the original data set based on the Pearson correlation coefficient method; Select the feature components whose correlation coefficient is greater than a preset threshold as the target feature set; The target feature set is merged with the original data set to obtain a combined data set.

4. The method according to claim 1, wherein The dual echo state network model includes an encoding echo state network, a decoding echo state network and a potential connection layer; The step of performing feature learning and processing on the combined data set through the hybrid prediction model to obtain output data of each hybrid prediction model comprises: Inputting the combined data set into the convolutional neural network model in each of the hybrid prediction models to extract initial data features; Inputting the initial data features into the coded echo state network to perform feature dimensionality reduction to obtain reduced-dimensional data features; The dimension-reduced data features are input into the decoding echo state network through the potential connection layer to perform data decoding, thereby obtaining output data of each hybrid prediction model.

5. The method according to claim 4, characterized in that The initial data features are input into the coded echo state network to perform feature dimensionality reduction, and the reduced dimensionality data features are obtained as follows: ； ； in, represents the state of the coded echo state network's reserve pool at time t+1, represents the activation function of the coded echo state network, represents the input weights of the coded echo state network, represents the input data of the reserve pool of the coded echo state network at time t+1, represents the weight of the reservoir of the coded echo state network, represents the state of the coded echo state network's reserve pool at time t, represents the bias vector of the coded echo state network, Represents the dimensionality reduction data features output by the reserve pool of the coded echo state network at time t+1, Represents the input weights of the latent connection layer.

6. The method according to claim 4, characterized in that The dimension-reduced data features are input into the decoding echo state network through the potential connection layer for data decoding, and the output data of each hybrid prediction model is obtained as follows: ； ； in, represents the state of the decoded echo state network's reserve pool at time t+1, represents the activation function of the decoded echo state network, represents the output weight of the potential connection layer, Represents the dimensionality reduction data features output by the reserve pool of the coded echo state network at time t+1, represents the weight of the reservoir of the decoding echo state network, represents the state of the decoded echo state network's reserve pool at time t, represents the bias vector of the decoded echo state network, represents the output data of the reserve pool of the decoded echo state network at time t+1, represents the output weights of the decoded echo state network.

7. The method according to claim 2, characterized in that The historical electricity consumption data is normalized to obtain standard data as follows: ； ； in, Indicates standard data, Represents historical electricity consumption data, and Respectively represent the minimum and maximum values ​​of historical electricity consumption data, Indicates converting standard data to its original numerical value.

8. The method according to claim 2, characterized in that The outliers in the standard data are eliminated by the interquartile range method to obtain the original data set: Calculate the upper quartile and the lower quartile of the standard data; Calculate the difference between the upper quartile and the lower quartile to obtain the interquartile range; Determine an outlier range according to the upper quartile, the lower quartile, and the interquartile range; Outliers in the standard data are determined according to the outlier range and removed to obtain an original data set.

9. A regional electricity consumption prediction terminal based on a hybrid model, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the processor implements the various steps in the method for predicting regional electricity consumption based on a hybrid model as described in any one of claims 1 to 8.

Images