A Power Load Forecasting Method Based on the SLSTM-RMTN Model

Abstract

The present invention relates to a power load forecasting method based on the SLSTM-RMTN model. For the target power system, it is designed to start from the STL decomposition of data values, and for the trend component, seasonal component, and residual component respectively, screen the target sensors corresponding to each component. Train the target network to be trained, which consists of a long short-term memory network, residual connection, and polynomial expansion, to obtain the power load forecasting models under each component. Then, in the prediction application, for the sampled data, execute the predictions of the power load forecasting models under each component respectively and perform fusion to complete the power load forecasting of the target power system. The designed solution can accurately capture the time series characteristics of power load data in the application, reduce the data dimension, significantly improve the accuracy and generalization ability of the model, and can optimize the allocation of power resources and balance the supply and demand relationship.

Description

Technical Field

[0001] The present invention relates to a power load forecasting method based on an SLSTM-RMTN model, belonging to the technical field of power load forecasting. Background Art

[0002] With the development of the economy and the improvement of people's living standards, the demand for electricity continues to grow, and the importance of power load forecasting becomes increasingly prominent. Forecasting models have always been an important topic that researchers are concerned about, and a lot of efforts have been made for this. For example, using a long short-term memory (LSTM) network for time series forecasting can effectively capture the time-varying characteristics of power loads, thereby improving the forecasting performance. In addition, load forecasting can also help balance the power supply and demand relationship and optimize resource allocation. However, existing traditional methods have problems of low forecasting accuracy and insufficient generalization ability when facing complex seasonal and trend load changes. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a power load forecasting method based on an SLSTM-RMTN model, which can more accurately capture the time series characteristics of load data, so as to better achieve the reasonable coordination of power consumption and power generation.

[0004] The present invention adopts the following technical solutions to solve the above technical problems: The present invention designs a power load forecasting method based on an SLSTM-RMTN model, and performs the following steps A to D to obtain a power load forecasting model corresponding to the target power system, and then performs steps to step , and conducts power load forecasting for the target power system;

[0005] Step A. Obtain the data values of each preset primary sensor at each sampling moment within a preset historical period corresponding to the target power system, as well as the data values of the power load, and perform data preprocessing and update, and then enter step B;

[0006] Step B. For the data values of each primary sensor and the data values of the power load at each sampling moment respectively, perform STL decomposition to obtain the data values of the trend component, the data values of the seasonal component, and the data values of the residual component corresponding to the data values, and then enter step C;

[0007] Step C. For the trend component, the seasonal component, and the residual component respectively, screen out each primary sensor with a correlation value with the power load greater than a preset correlation threshold under the component to form each target sensor corresponding to the component, and then enter step D;

[0008] Step D. For the trend component, seasonal component, and residual component respectively, using the data values at a preset number of consecutive sampling moments corresponding to each target sensor of the component and the power load corresponding to each component as inputs, and combining the data value at the th sampling moment of the power load corresponding to the component as the output, train the network to be trained for the preset target to obtain the power load prediction model for the corresponding component of the target power system. Then, obtain the power load prediction models for the trend component, seasonal component, and residual component of the target power system respectively, and combine them to form the power load prediction model corresponding to the target power system; consecutive sampling moments, and use the data value at the th sampling moment of the power load corresponding to the component as the output, train the network to be trained for the preset target to obtain the power load prediction model for the corresponding component of the target power system. Then, obtain the power load prediction models for the trend component, seasonal component, and residual component of the target power system respectively, and combine them to form the power load prediction model corresponding to the target power system;

[0009] Step . Aggregate the target sensors corresponding to the trend component, seasonal component, and residual component respectively, and based on consecutive sampling moments in the historical time direction starting from the current sampling moment to form each target sampling moment, collect the data values of all target sensors and the data value of the power load corresponding to each target sampling moment of the target power system, and then enter Step ;

[0010] Step . For the data values of all target sensors and the data value of the power load corresponding to each target sampling moment respectively, perform STL decomposition to obtain the data values of the trend component, seasonal component, and residual component corresponding to the data values, and then enter Step ;

[0011] Step . Using the next adjacent sampling moment in the future time direction from the current sampling moment as the prediction moment, for the trend component, seasonal component, and residual component respectively, using the data values of each target sensor corresponding to the component and the data values of the power load corresponding to each component at each target sampling moment as inputs, apply the power load prediction model for the corresponding component of the target power system to obtain the data values of the power load corresponding to the component at the prediction moment. Then, obtain the data values of the trend component, seasonal component, and residual component of the power load corresponding to the prediction moment, and then enter Step ;

[0012] Step . For the data values of the trend component, seasonal component, and residual component of the power load corresponding to the prediction moment, perform an addition process to obtain the power load of the target power system corresponding to the prediction moment.

[0013] As a preferred technical solution of the present invention: in step A, after obtaining the data values of each preset primary sensor and the data value of the power load corresponding to each sampling moment within the preset historical period of the target power system, each primary sensor and the power load are respectively used as preprocessing objects, and the following steps A1 to A3 are executed to perform preprocessing and update on the data values corresponding to each sampling moment of the preprocessing object;

[0014] Step A1. For the data values corresponding to each sampling moment of the preprocessing object, perform denoising update, and then proceed to step A2;

[0015] Step A2. According to the preset normal value range corresponding to the preprocessing object, determine whether there are abnormal values in the data values corresponding to each sampling moment of the preprocessing object. If so, replace and update the abnormal values with the average value of each normal value, otherwise do not perform any processing, and then proceed to step A3;

[0016] Step A3. Perform normalization update on the data values corresponding to each sampling moment of the preprocessing object.

[0017] As a preferred technical solution of the present invention: in step B, for the data values of each primary sensor and the data value of the power load corresponding to each sampling moment, according to the following formula:

[0018]

[0019] Perform STL decomposition to obtain the data values corresponding to the trend component data values and the seasonal component data values and the residual component data values , where, represents the data value corresponding to the th sampling moment, represents the data value of the trend component corresponding to the data value corresponding to the th sampling moment, represents the data value of the seasonal component corresponding to the data value corresponding to the th sampling moment, represents the data value of the residual component corresponding to the data value corresponding to the th sampling moment.

[0020] As a preferred technical solution of the present invention: in step C, for the trend component, seasonal component, and residual component respectively, execute the following steps C1 to C2;

[0021] Step C1. For each primary sensor respectively, according to the following formula:

[0022]

[0023] Obtain the correlation value under the corresponding component between the primary selected sensor and the power load , where , represents the number of primary selected sensors, represents the th data value sequence at each sampling moment under the corresponding component of the th primary selected sensor, represents the th data value at the th sampling moment in , represents the power load th data value sequence at each sampling moment under the corresponding component, represents the th data value at the th sampling moment in , , represents the number of sampling moments within the preset historical period, represents the distribution probability value of represents the distribution probability value of represents the joint distribution probability value of and

[0024] Step C2. Screen and obtain each primary selected sensor with a correlation value greater than the preset correlation threshold, and form each target sensor corresponding to the component.

[0025] As a preferred technical solution of the present invention: the target network to be trained in step D is sequentially connected in series with an input layer, a polynomial layer, a stacked long short-term memory network, a composite fully connected layer, a Dropout layer, a fusion layer, a second fully connected layer, and a linear transformation layer from the input end to the output end. Among them, the stacked long short-term memory network includes two long short-term memory networks sequentially connected in series from the input end to the output end. The composite fully connected layer includes a first fully connected layer and an activation layer sequentially connected in series from the input end to the output end. The input end of the fusion layer is simultaneously connected to the output end of the input layer. The fusion layer adds the output of the input layer and the output of the Dropout layer, and outputs the addition result to the input end of the second fully connected layer.

[0026] As a preferred technical solution of the present invention: the polynomial layer performs the following operations on the input data received and forwarded by the input layer ; represents the number of target sensors corresponding to the component plus 1;

[0027] First, according to the preset highest expansion order corresponding to the target network to be trained , according to the following formula:

[0028]

[0029] obtain , represents the input data with respect to the expansion terms of each order of the highest expansion order ;

[0030] Then, based on , , , according to the following formula:

[0031]

[0032] obtain , and further obtain , and output it to the stacked long short-term memory network, where represents the output of the th layer in the polynomial layer corresponding to the th sampling moment, represents the number of monomials in the polynomial layer based on , , represents the weight vector of the th layer in the polynomial layer, represents the th weight in is the parameter to be trained in the polynomial layer.

[0033] As a preferred technical solution of the present invention: in the stacked long short-term memory network, the output of the polynomial layer is received by the first long short-term memory network in sequence, and according to , obtain , and output it to the second long short-term memory network in sequence. The output of the first long short-term memory network in sequence is received by the second long short-term memory network in sequence, and according to , obtain , and output it to the composite fully connected layer, where represents the function of the first long short-term memory network in sequence, represents the function of the second long short-term memory network in sequence;

[0034] In the composite fully connected layer, first, the output of the stacked long short-term memory network is received by the first fully connected layer, according to the following formula:

[0035]

[0036] obtain , and output it to the activation layer, where represents the first fully connected layer function, represents the output of the th layer in the first fully connected layer corresponding to the th sampling moment, ;

[0037] Then the activation layer receives the output of the first fully connected layer , according to the following formula:

[0038]

[0039] obtain , and output it to the Dropout layer, where represents the activation layer function, represents the output of the th layer in the activation layer corresponding to the th sampling moment.

[0040] As a preferred technical solution of the present invention: the Dropout layer receives the output of the composite fully connected layer , according to the following formula:

[0041]

[0042] obtain , and output it to the fusion layer, where represents the inactivation layer function, ;

[0043] The fusion layer receives the output of the inactivation layer , and receives the input data forwarded by the input layer, according to the following formula:

[0044]

[0045] perform fusion processing to obtain , and output it to the second fully connected layer, where , represents the dimension of the output of the fusion layer, represents the output of the th layer in the fusion layer corresponding to the th sampling moment.

[0046] As a preferred technical solution of the present invention: the second fully connected layer receives the output of the fusion layer , according to the following formula:

[0047]

[0048] obtain , and output it to the linear transformation layer; where represents the second fully connected layer function; represents the output of the th layer in the second fully connected layer corresponding to the th sampling moment;

[0049] The linear transformation layer receives the output of the second fully connected layer , and based on , according to the following formula:

[0050]

[0051] obtain output, where represents the output of the linear transformation layer corresponding to the th sampling moment, and represent the weight matrix and bias term of the linear transformation layer.

[0052] As a preferred technical solution of the present invention: in the process of training the preset target network to be trained according to the Adam optimizer in step D, for each iteration respectively, first according to the following MSE loss function,

[0053]

[0054] obtain the loss structure , where represents the actual value of the th iteration of the target network to be trained corresponding to the th sample, represents the predicted value of the th iteration of the target network to be trained corresponding to the th sample, represents the number of samples for training, represents the th iteration of the loss value;

[0055] Then, for the data values of the coefficients of each polynomial layer in the target network to be trained and the data values of the weights in the stacked long short-term memory network, update according to the following formula;

[0056]

[0057] where represents the corresponding The data value of the next iteration, indicating the corresponding updated data value of the next iteration, representing the gradient of and

[0058] A power load forecasting method based on the SLSTM-RMTN model according to the present invention, compared with the prior art by adopting the above technical solution, has the following technical effects:

[0059] (1) A power load forecasting method based on the SLSTM-RMTN model designed by the present invention starts from the STL decomposition of data values for the target power system. For the trend component, seasonal component, and residual component respectively, the corresponding target sensors are selected. The target network to be trained composed of a long short-term memory network, residual connection, and polynomial expansion is trained to obtain the power load forecasting model under each component. Then, in the prediction application, for the sampled data, the predictions of the power load forecasting models under each component are respectively executed and fused to complete the power load forecasting of the target power system. In the application of the design scheme, the time series characteristics of the power load data can be accurately captured, the data dimension can be reduced, the accuracy and generalization ability of the model are significantly improved, and the power resource allocation can be optimized and the supply-demand relationship can be balanced. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 is a schematic structural diagram of the power load forecasting method based on the SLSTM-RMTN model designed by the present invention;

[0061] Figure 2 is a schematic architecture diagram of the target network to be trained in the design of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0062] The following further details the specific embodiments of the present invention in conjunction with the accompanying drawings of the specification.

[0063] In view of the deficiencies of the prior art, the present invention specifically designs a power load forecasting method based on the SLSTM-RMTN model. In practical applications, as Figure 1 shown, the specific design first performs the following steps A to D to obtain the power load forecasting model corresponding to the target power system.

[0064] Step A. Obtain the data values of each preset primary sensor at each sampling moment within the preset historical period corresponding to the target power system, as well as the data value of the power load. Then, using each primary sensor and the power load as preprocessing objects respectively, perform the following steps A1 to A3, preprocess and update the data values at each sampling moment corresponding to the preprocessing objects, and then proceed to Step B.

[0065] Step A1. For the data values at each sampling moment corresponding to the preprocessing object, perform denoising update, and then proceed to Step A2.

[0066] Step A2. According to the preset normal value range corresponding to the preprocessing object, determine whether there are any outliers among the data values at each sampling moment corresponding to the preprocessing object. If so, replace and update the outliers with the average value of each normal value; otherwise, do not perform any processing, and then proceed to Step A3.

[0067] Step A3. Perform normalization update on the data values at each sampling moment corresponding to the preprocessing object to make the data values have the same scale range.

[0068] Step B. For the data values of each primary sensor and the data value of the power load at each sampling moment respectively, according to the following formula:

[0069]

[0070] Perform STL (Seasonal and Trend decomposition using Loess) decomposition to obtain the data values corresponding to the trend component data values and the seasonal component data values and the residual component data values , where represents the data value corresponding to the th sampling moment, represents the data value of the trend component corresponding to the data value at the th sampling moment, represents the data value of the seasonal component corresponding to the data value at the th sampling moment, represents the data value of the residual component corresponding to the data value at the th sampling moment, and then proceed to Step C.

[0071] Step C. For the trend component, seasonal component, and residual component respectively, screen and obtain each primary sensor with a correlation value greater than the preset correlation threshold with the power load under the component, and form each target sensor corresponding to the component, and then proceed to Step D.

[0072] In practical applications, in the above step C, for the trend component, seasonal component, and residual component respectively, the following steps C1 to C2 are executed.

[0073] Step C1. For each preliminary selected sensor respectively, according to the following formula:

[0074]

[0075] Obtain the correlation value between the preliminary selected sensor and the power load under the corresponding component , where , represents the number of preliminary selected sensors, represents the data value sequence of each sampling moment under the corresponding component of the th preliminary selected sensor, represents the data value at the th sampling moment in represents the power load the data value sequence of each sampling moment under the corresponding component, represents the data value at the th sampling moment in represents the mutual value between and represents "Mutual Information", that is, mutual information, which is used to measure the and mutual dependence relationship; based on , represents the number of sampling moments in the preset historical period, represents the distribution probability value of represents the distribution probability value of represents the joint distribution probability value of and

[0076] Step C2. Screen and obtain each preliminary selected sensor with a correlation value greater than the preset correlation threshold to form each target sensor corresponding to the component.

[0077] Step D. For the trend component, seasonal component, and residual component respectively, using the data values of the preset number of consecutive sampling moments of each target sensor corresponding to the component and the power load respectively corresponding to the component as the input, combined with the Taking the data value at a sampling moment as the output, training is performed on the network to be trained for a preset target, obtaining a power load forecasting model for the corresponding component of the target power system. Furthermore, power load forecasting models for the trend component, seasonal component, and residual component corresponding to the target power system are obtained, and they are combined to form the power load forecasting model corresponding to the target power system.

[0078] In the actual application of the above step D, such as Figure 2 shown, the network to be trained for the target sequentially connects an input layer, a polynomial layer, a stacked long short-term memory network, a composite fully connected layer, a Dropout layer (random inactivation layer), a fusion layer, a second fully connected layer, and a linear transformation layer from the input end to the output end. Among them, the stacked long short-term memory network includes two long short-term memory networks sequentially connected in series from the input end to the output end. The composite fully connected layer includes a first fully connected layer and an activation layer sequentially connected in series from the input end to the output end. The input end of the fusion layer is simultaneously connected to the output end of the input layer. The fusion layer adds the output of the input layer and the output of the Dropout layer (random inactivation layer), and outputs the added result to the input end of the second fully connected layer.

[0079] In the specific data processing process of the network to be trained for the target, the polynomial layer performs the following operations on the input data received and forwarded by the input layer , , and does the following operations; represents the number of target sensors corresponding to the component plus 1;

[0080] First, according to the preset highest expansion term corresponding to the network to be trained for the target , according to the following formula:

[0081]

[0082] obtain , represents the input data about the highest expansion term of each expansion term.

[0083] Then, based on , , , according to the following formula:

[0084]

[0085] obtain , and further obtain , and output it to the stacked long short-term memory network, where represents the output of the th layer in the polynomial layer corresponding to the th sampling moment, Denote the number of monomials in the polynomial layer based on , , and the number of monomials in the polynomial layer, Denote the weight vector of the -th layer in the polynomial layer, Denote The -th weight in , and is the parameter to be trained in the polynomial layer.

[0086] In the stacked long short-term memory network, the output of the polynomial layer is received by the first long short-term memory network in sequence , and according to , obtain , and output it to the second long short-term memory network in sequence. The output of the first long short-term memory network in sequence is received by the second long short-term memory network in sequence , and according to , obtain , and output it to the composite fully connected layer. Among them, Denote the function of the first long short-term memory network in sequence, Denote the function of the second long short-term memory network in sequence.

[0087] Here, in the design of the stacked long short-term memory network, the first long short-term memory network in sequence extracts short-term time dependencies within the time step range, analyzes the feature sequence at each time step, and passes it to the second long short-term memory network in sequence through the hidden state for higher-level processing. This non-linear transformation of the input features can capture complex time dependencies. The second long short-term memory network in sequence extracts long-term dependencies over the entire time series and outputs the hidden states of all time steps, thereby retaining the dynamic information of the entire time series and providing a richer feature representation for subsequent prediction tasks.

[0088] In the composite fully connected layer, first, the output of the stacked long short-term memory network is received by the first fully connected layer , and according to the following formula:

[0089]

[0090] Obtain , and output it to the activation layer. Among them, Denote the function of the first fully connected layer, Denote the output of the -th layer in the first fully connected layer corresponding to the -th sampling moment, .

[0091] Then, the output of the first fully connected layer is received by the activation layer , and according to the following formula:

[0092]

[0093] Obtain , and output it to the Dropout layer (random inactivation layer), where represents the activation layer function, represents the th layer in the activation layer corresponding to the output at the th sampling moment.

[0094] After the calculation of the first fully connected layer in the composite fully connected layer, introduce the activation function tanh to map the input data to the range of [-1, 1]. Its non-linear characteristics enable the model to better capture complex input-output mapping relationships.

[0095] The Dropout layer (random inactivation layer) receives the output of the composite fully connected layer , and according to the following formula:

[0096]

[0097] Obtain , and output it to the fusion layer, where represents the inactivation layer function, , and here the random inactivation technology is adopted. By randomly discarding some neurons during training, the generalization ability of the model is improved.

[0098] The fusion layer receives the output of the inactivation layer , and also receives the input data forwarded by the input layer , and according to the following formula:

[0099]

[0100] Perform fusion processing to obtain , and output it to the second fully connected layer, where , represents the dimension of the output of the fusion layer, represents the th layer in the fusion layer corresponding to the output at the th sampling moment.

[0101] The second fully connected layer receives the output of the fusion layer , and according to the following formula:

[0102]

[0103] Obtain , and output it to the linear transformation layer; where represents the second fully connected layer function; Indicates the output corresponding to the th layer in the second fully connected layer at the th sampling moment.

[0104] The linear transformation layer receives the output of the second fully connected layer , and based on , according to the following formula:

[0105]

[0106] obtains the output, where represents the output of the linear transformation layer corresponding to the th sampling moment, and represent the weight matrix and bias term of the linear transformation layer.

[0107] And in actual applications, before training the target network to be trained, the model parameters include the number of neurons in the stacked LSTM network, the polynomial coefficients in the residual multi-dimensional Taylor network, and other related parameters, and perform random initialization. The purpose of random initialization is to avoid the model parameters falling into local minima and improve the training efficiency, and the initialization of the weights is usually carried out through normal distribution or uniform distribution. This initialization method can ensure that the gradient propagation in the initial stage of network training is relatively stable, which helps to accelerate the model convergence.

[0108] During the process of training the target network to be trained, according to the Adam optimizer, during the process of training the target network to be trained, for each iteration in turn, first according to the following MSE loss function,

[0109]

[0110] obtains the loss structure , where represents the actual value of the th iteration of the target network to be trained for the th sample, represents the predicted value of the th iteration of the target network to be trained for the th sample, represents the number of samples used for training, represents the th iteration of the loss value.

[0111] Then, for the data values of the coefficients of each polynomial layer in the target network to be trained and the data values of the weights in the stacked long short-term memory network, update them according to the following formula;

[0112]

[0113] Among them, represents the data value corresponding to the th iteration, represents the updated data value corresponding to the th iteration, represents the gradient of , represents the learning rate.

[0114] During the specific training process, the parameters of the model (including the weights of the LSTM network and the polynomial coefficients in RMTN) are updated through the Adam optimizer and the backpropagation algorithm. The Adam optimizer dynamically adjusts the learning rate according to the gradient and historical gradient of each parameter to accelerate the convergence of the model and improve the prediction accuracy. During the backpropagation process, the gradient is reversely transmitted layer by layer from the output layer to the input layer, and the network parameters are calculated and updated in turn to achieve the gradual optimization of the model.

[0115] Based on the execution of the above steps A to D, a power load prediction model corresponding to the target power system is obtained. Then, in the application, further design and execute steps to for power load prediction of the target power system.

[0116] Step . Aggregate the respective target sensors corresponding to the trend component, seasonal component, and residual component, and based on the consecutive sampling moments in the historical time direction starting from the current sampling moment, form each target sampling moment, collect the data values of all the respective target sensors corresponding to each target sampling moment of the target power system, as well as the data value of the power load, and then enter step .

[0117] Step . For the data values of all the respective target sensors corresponding to each target sampling moment, as well as the data value of the power load, respectively perform STL (Seasonal and Trend decomposition using Loess) decomposition to obtain the data values of the trend component, seasonal component, and residual component corresponding to the data values, and then enter step .

[0118] Step . Taking the next adjacent sampling moment in the future time direction from the current sampling moment as the prediction moment, respectively for the trend component, seasonal component, and residual component, using the data values corresponding to each target sensor at each target sampling moment for the component, and the data values of the power load corresponding to each component as inputs, applying the power load prediction model for the corresponding component of the target power system, obtaining the data values of the power load corresponding to the component at the prediction moment, and further obtaining the data values of the power load corresponding to the trend component, seasonal component, and residual component at the prediction moment, and then entering step .

[0119] Step . For the data values of the power load corresponding to the trend component, seasonal component, and residual component at the prediction moment, perform an addition process to obtain the power load of the target power system at the corresponding prediction moment.

[0120] The above technical solution designs a power load prediction method based on the SLSTM-RMTN model. For the target power system, starting from the STL decomposition of the data values, respectively for the trend component, seasonal component, and residual component, screening the target sensors corresponding to each component, training the target network to be trained composed of long short-term memory network, residual connection, and polynomial expansion, obtaining the power load prediction model for each component, and then in the prediction application, respectively performing the prediction of the power load prediction model for each component on the sampling data and fusing them to complete the power load prediction of the target power system; the designed solution can accurately capture the time series characteristics of the power load data, reduce the data dimension, significantly improve the accuracy and generalization ability of the model, and can optimize the power resource allocation and balance the supply and demand relationship.

[0121] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those of ordinary skill in the art, various changes can be made without departing from the gist of the present invention.

Claims

1. A power load forecasting method based on the SLSTM-RMTN model, characterized in that: Perform the following steps A to D to obtain the power load forecasting model corresponding to the target power system, and then perform steps i to iv to conduct power load forecasting for the target power system; Step A. Obtain the data values of each preset primary sensor at each sampling moment within the preset historical period corresponding to the target power system, as well as the data values of the power load, and perform data preprocessing and update, then proceed to Step B; Step B. For the data values of each primary sensor and the data values of the power load at each sampling moment, perform STL decomposition to obtain the data values of the trend component, seasonal component, and residual component corresponding to the data values, then proceed to Step C; Step C. For the trend component, seasonal component, and residual component respectively, screen out the primary sensors with a correlation value greater than the preset correlation threshold with the power load under the component to form the respective target sensors corresponding to the component, then proceed to Step D; Step D. For the trend component, seasonal component, and residual component respectively, use the respective target sensors corresponding to the component and the data values of the preset number m of consecutive sampling moments of the power load corresponding to the component as inputs, and combine the data value of the (m + 1)-th sampling moment corresponding to the power load under the component as the output to train the target network to be trained, obtain the power load forecasting model corresponding to the component of the target power system, and further obtain the power load forecasting models corresponding to the trend component, seasonal component, and residual component of the target power system respectively, and combine them to form the power load forecasting model corresponding to the target power system; Step i. Summarize the respective target sensors corresponding to the trend component, seasonal component, and residual component, and based on m consecutive sampling moments in the historical time direction starting from the current sampling moment, form each target sampling moment, collect the data values of all the respective target sensors and the data values of the power load corresponding to the target power system at each target sampling moment, then proceed to Step ii; Step ii. For the data values of all the respective target sensors and the data values of the power load at each target sampling moment, perform STL decomposition to obtain the data values of the trend component, seasonal component, and residual component corresponding to the data values, then proceed to Step iii; Step iii. Taking the next adjacent sampling moment in the future time direction from the current sampling moment as the forecasting moment, for the trend component, seasonal component, and residual component respectively, use the respective target sensors corresponding to the component at each target sampling moment and the data values of the power load corresponding to the component as inputs, apply the power load forecasting model corresponding to the component of the target power system to obtain the data values of the power load corresponding to the component at the forecasting moment, and further obtain the data values of the trend component, seasonal component, and residual component corresponding to the power load at the forecasting moment, then proceed to Step iv; Step iv. Perform an addition operation on the data values of the trend component, seasonal component, and residual component corresponding to the power load at the forecasting moment to obtain the power load of the target power system corresponding to the forecasting moment.

2. The power load prediction method based on the SLSTM-RMTN model according to claim 1, wherein: In the said step A, after obtaining the data values of each preset primary sensor and the data value of the power load corresponding to each sampling moment within the preset historical period for the target power system, each primary sensor and the power load are respectively used as the preprocessing objects, and the following steps A1 to A3 are executed to perform preprocessing and update on the data values corresponding to each sampling moment of the preprocessing objects; Step A1. For the data values corresponding to each sampling moment of the preprocessing object, perform denoising update, and then proceed to step A2; Step A2. According to the preset normal value range corresponding to the preprocessing object, determine whether there are outliers in the data values corresponding to each sampling moment of the preprocessing object. If so, replace and update the outliers with the average value of each normal value, otherwise do not perform any processing, and then proceed to step A3; Step A3. Perform normalization update on the data values corresponding to each sampling moment of the preprocessing object.

3. The power load forecasting method based on the SLSTM-RMTN model according to claim 1, characterized in that: In the said step B, for the data values of each primary sensor and the data value of the power load corresponding to each sampling moment respectively, according to the following formula: X(t) = T(t) + S(t) + R(t) Perform STL decomposition to obtain the data value T(t) of the trend component, the data value S(t) of the seasonal component, and the data value R(t) of the residual component corresponding to the data value X(t), where X(t) represents the data value corresponding to the t-th sampling moment, T(t) represents the data value of the trend component corresponding to the data value corresponding to the t-th sampling moment, S(t) represents the data value of the seasonal component corresponding to the data value corresponding to the t-th sampling moment, and R(t) represents the data value of the residual component corresponding to the data value corresponding to the t-th sampling moment.

4. A power load forecasting method based on the SLSTM-RMTN model according to claim 1, characterized in that: In the said step C, for the trend component, the seasonal component, and the residual component respectively, execute the following steps C1 to C2; Step C1. For each primary sensor respectively, according to the following formula: Obtain the correlation value MI(U a ,U b ) under the corresponding component between the primary selected sensor and the power load, where 1 ≤ b ≤ B, B represents the number of primary selected sensors, U b represents the data value sequence at each sampling moment under the corresponding component of the b-th primary selected sensor, U b (t) represents the data value at the t-th sampling moment in U b , U a represents the data value sequence at each sampling moment under the corresponding component of the power load a, U a (t) represents the data value at the t-th sampling moment in U a , MI(U a ,U b ) represents the mutual information value between U a and U b . Based on 1 ≤ t ≤ T, T represents the number of sampling moments within the preset historical period, P(U a (t)) represents the distribution probability value of U a , P(U b (t)) represents the distribution probability value of U b , P(U a (t),U b (t)) represents the joint distribution probability value of U a and U b ; then go to step C2; Step C2. Screen and obtain each primary sensor with a correlation value greater than the preset correlation threshold to form each target sensor corresponding to the component.

5. The power load forecasting method based on the SLSTM-RMTN model according to claim 1, characterized in that: The target network to be trained in the said step D is successively connected in series with an input layer, a polynomial layer, a stacked long short-term memory network, a composite fully connected layer, a Dropout layer, a fusion layer, a second fully connected layer, and a linear transformation layer from the input end to the output end direction. Among them, the stacked long short-term memory network includes two long short-term memory networks successively connected in series from the input end to the output end direction. The composite fully connected layer includes a first fully connected layer and an activation layer successively connected in series from the input end to the output end direction. The input end of the fusion layer is simultaneously connected to the output end of the input layer. The fusion layer adds the output of the input layer and the output of the Dropout layer and outputs the added result to the input end of the second fully connected layer.

6. The power load forecasting method based on the SLSTM-RMTN model according to claim 5, wherein: The polynomial layer performs the following operations on the input data X(t) = x1(t), …, x Q (t), where 1 ≤ t ≤ m; Q represents the number of target sensors corresponding to the component plus 1; First, according to the preset highest expansion term M corresponding to the target network to be trained, according to the following formula: I(t) = [1, x1(t)…x Q (t), x1(t)x2(t)…(x Q (t)) 2 ,…, x1(t)…x M (t)…(x Q (t)) M T ​ Obtain I(t), where I(t) represents the expansion of the input data x1(t), …, x Q (t) for each term of the highest expansion order M; Then, according to W p =[w 1,p ,…,w n,p ,…,w N(Q,M),p , 1 ≤ p ≤ P, 1 ≤ n ≤ N(Q, M), according to the following formula: l p (t) = W p ·I(t) Obtain l p (t), and then obtain L(t) = [l1(t), …, l p (t), …, l P (t)] T , and output it to the stacked long short-term memory network, where l p (t) represents the output of the p-th layer in the polynomial layer corresponding to the t-th sampling moment, N(Q, M) represents the number of monomials in the polynomial layer based on Q and M, W p represents the weight vector of the p-th layer in the polynomial layer, w n,p represents the n-th weight in W p , and w n,p is a parameter to be trained in the polynomial layer.

7. The power load forecasting method based on the SLSTM-RMTN model according to claim 6, characterized in that: In the stacked long short-term memory network, the output L(t) of the polynomial layer is received by the first long short-term memory network in sequence. According to H1(t) = LSTM1(L(t)), H1(t) is obtained and output to the second long short-term memory network in sequence. The second long short-term memory network in sequence receives the output H1(t) of the first long short-term memory network in sequence. According to H2(t) = LSTM2(H1(t)), H2(t) is obtained and output to the composite fully connected layer, where LSTM1( ) represents the function of the first long short-term memory network in sequence, and LSTM2( ) represents the function of the second long short-term memory network in sequence; In the composite fully connected layer, first, the output H2(t) of the stacked long short-term memory network is received by the first fully connected layer, and according to the following formula: [h1(t),…,h p (t),…,h P (t)] = FC1(H2(t)) Obtain [h1(t),…,h p (t),…,h P (t)], and output them to the activation layer, where FC1() represents the first fully connected layer function, and h p (t) represents the output of the p-th layer in the first fully connected layer corresponding to the t-th sampling moment, 1 ≤ t ≤ m; Then the activation layer receives the output [h1(t), …, h p (t), …, h P (t)] of the first fully-connected layer, according to the following formula: tanh[h1(t),…,h p (t),…,h P (t)]=[b1(t),…,b p (t),…,b P (t)] Obtain [b1(t), …, b p (t), …, b P (t)], and output it to the Dropout layer, where tanh[] represents the activation layer function, and b p (t) represents the output of the p-th layer in the activation layer corresponding to the t-th sampling moment.

8. The power load forecasting method based on the SLSTM-RMTN model according to claim 7, characterized in that: The Dropout layer receives the output [b1(t), …, b p (t), …, b P (t)] of the composite fully connected layer, according to the following formula: G(t) = Dropout[b1(t),…,b p (t),…,b P (t)] G(t) is obtained and output to the fusion layer, where Dropout[] represents the function of the inactivation layer, and 1 ≤ t ≤ m; The fusion layer receives the output G(t) of the inactivation layer and the input data X(t) forwarded by the input layer. According to the following formula: Z(t) = G(t) + X(t) = [z1(t), …, z l (t), …, z L (t)] Perform fusion processing to obtain [z1(t), …, z l (t), …, z L (t)], and output it to the second fully connected layer, where 1 ≤ l ≤ L, L represents the dimension of the output of the fusion layer, and z l (t) represents the output of the l-th layer in the fusion layer corresponding to the t-th sampling moment.

9. The method for predicting electric load based on the SLSTM-RMTN model according to claim 8, characterized in that: The second fully connected layer receives the output [z1(t), …, z l (t), …, Z L (t)] of the fusion layer, according to the following formula: R(t) = FC2[z1(t),…,z l (t),…,z L (t)] = [r1(t),…,r l (t),…,r L (t)] Obtain [r1(t), …, r l (t), …, r L (t)], and output it to the linear transformation layer; where FC2( ) represents the second fully connected layer function; r l (t) represents the output of the l-th layer in the second fully connected layer corresponding to the t-th sampling moment; The linear transformation layer receives the output [r1(t), …, r l (t), …, r L (t)] of the second fully connected layer, and based on 1 ≤ t ≤ m, according to the following formula: Obtain an output, where represents the output of the linear transformation layer corresponding to the (m + 1)-th sampling moment, and Φ and Θ represent the weight matrix and bias term of the linear transformation layer, respectively.

10. The power load forecasting method based on the SLSTM-RMTN model according to claim 1, characterized in that: In the process of training the network to be trained for the preset target according to the Adam optimizer in step D, for each iteration in sequence, first, according to the following MSE loss function, Obtain the loss structure L k , where y j,k represents the actual value of the j-th sample of the target network to be trained corresponding to the k-th iteration, represents the predicted value of the j-th sample of the target network to be trained corresponding to the k-th iteration, J represents the number of samples used for training, and L k represents the loss value at the l-th iteration; Then, for the data values of the coefficients of each polynomial layer in the network to be trained for the target and the data values of each weight in the stacked long short-term memory network, they are updated according to the following formula; Among them, α k represents the data value corresponding to the k-th iteration, and α k+1 represents the updated data value corresponding to the (k + 1)-th iteration. represents the gradient of L k with respect to α k , and η represents the learning rate.

Images