Power load prediction method based on SLSTM-RMTN model

Abstract

The invention relates to a power load prediction method based on an SLSTM-RMTN model, and the method comprises the steps: taking the STL decomposition of a data value as a starting point for a target power system, and screening target sensors corresponding to a trend component, a season component and a residual component respectively, a target to-be-trained network composed of a long-short-term memory network, residual connection and polynomial expansion is trained, a power load prediction model under each component is obtained, then under prediction application, prediction of the power load prediction model under each component is executed for sampling data, fusion is carried out, and a prediction result is obtained. Power load prediction of the target power system is completed; when the design scheme is applied, the time sequence characteristics of the power load data can be accurately captured, the data dimension is reduced, the precision and generalization ability of the model are remarkably improved, the power resource configuration can be optimized, and the supply-demand relationship can be balanced.

Description

Technical Field

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

[0002] With the development of economy and the improvement of people's living standards, the demand for electricity continues to grow, and the importance of power load forecasting has become more and more prominent. Forecasting models have always been an important topic of concern for researchers, and a lot of efforts have been made to this end. For example, using long short-term memory (LSTM) networks for time series forecasting can effectively capture the time-varying characteristics of power load, thereby improving forecasting performance. In addition, load forecasting can also help balance the relationship between power supply and demand and optimize resource allocation. However, existing traditional methods have problems with 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 the SLSTM-RMTN model, which can more accurately capture the time series characteristics of load data, thereby better realizing the reasonable coordination of power consumption and power generation.

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

[0005] Step A. Obtaining the data values ​​of each preset primary sensor and the data value of the power load at each sampling time in the preset historical period corresponding to the target power system, and performing data preprocessing and updating, and then proceeding to step B;

[0006] Step B. Perform STL decomposition on the data values ​​of each primary sensor and the data value of the power load at each sampling time, obtain the data values ​​corresponding to the trend component, the seasonal component, and the residual component, and then proceed to step C;

[0007] Step C. For the trend component, the seasonal component, and the residual component, respectively, the preliminary selected sensors whose correlation values ​​with the power load under the component are greater than the preset correlation threshold are screened to form the target sensors corresponding to the component, and then proceed to step D;

[0008] Step D. For the trend component, seasonal component, and residual component, the target sensors corresponding to the components and the power loads are respectively corresponding to the preset quantities under the components. The data value of the continuous sampling moment is input, combined with the corresponding component of the power load, The data value at each sampling moment is output, and the network to be trained is trained for the preset target to obtain the power load forecasting model under the corresponding component of the target power system, and then the power load forecasting model under the corresponding trend component, seasonal component, and residual component of the target power system is obtained, and the power load forecasting model corresponding to the target power system is combined to form the power load forecasting model corresponding to the target power system;

[0009] step . Summarize the target sensors corresponding to the trend component, seasonal component, and residual component, and based on the historical time direction from the current sampling time The target sampling moments are composed of continuous sampling moments, and the data values ​​of all target sensors and the data values ​​of the power load of the target power system corresponding to each target sampling moment are collected, and then the step ;

[0010] step . Perform STL decomposition on the data values ​​of all target sensors and the data values ​​of power load at each target sampling time, obtain the data values ​​corresponding to the trend component, the seasonal component, and the residual component, and then enter step ;

[0011] step . Take 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, take the data values ​​of each target sensor corresponding to the component at each target sampling moment, and the data values ​​of the corresponding components of the power load as input, apply the power load prediction model under the corresponding components of the target power system, obtain the data values ​​of the corresponding components of the power load at the prediction moment, and then obtain the data values ​​of the trend component, seasonal component, and residual component of the power load at the prediction moment, and then enter the step ;

[0012] step . For the data values ​​of the trend component, seasonal component, and residual component corresponding to the power load at the forecast time, an addition process is performed to obtain the power load of the target power system corresponding to the forecast time.

[0013] As a preferred technical solution of the present invention: in the step A, after obtaining the data values ​​of each preset primary sensor and the data values ​​of the power load at each sampling time in the preset historical period corresponding to the target power system, each of the primary sensors and the power load is used as a preprocessing object, and the following steps A1 to A3 are performed to preprocess and update the data values ​​of the preprocessing objects corresponding to each sampling time;

[0014] Step A1. Execute denoising update for the data value of the preprocessing object corresponding to each sampling time, and then proceed to step A2;

[0015] Step A2. According to the preset normal value range corresponding to the preprocessed object, determine whether there are abnormal values ​​in the data values ​​corresponding to each sampling time of the preprocessed object, and if so, replace and update the abnormal value with the average value of each normal value, otherwise, do nothing, and then proceed to step A3;

[0016] Step A3: normalize and update the data value of the preprocessing object corresponding to each sampling time.

[0017] As a preferred technical solution of the present invention: in the step B, the data values ​​of each primary selected sensor and the data value of the power load at each sampling time are respectively calculated according to the following formula:

[0018]

[0019] Perform STL decomposition to obtain data values The data values ​​corresponding to the trend components , seasonal component data values , the data value of the residual component ,in, Indicates the corresponding The data value at the sampling moment, Indicates the corresponding The data value of the trend component corresponding to the data value at each sampling moment, Indicates the corresponding The data value of the seasonal component corresponding to the data value at the sampling time, Indicates the corresponding The data value of the residual component corresponding to the data value at each sampling moment.

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

[0021] Step C1. For each pre-selected sensor, use the following formula:

[0022]

[0023] Obtain the correlation value of the corresponding component between the primary selected sensor and the power load ,in, , Indicates the number of primary selected sensors, Indicates The data value sequence at each sampling time under the corresponding component of the primary selected sensor, express Middle The data value at the sampling moment, Indicates power load The data value sequence at each sampling time under the corresponding component, express Middle The data value at the sampling moment, express and The mutuality value between , Indicates the number of sampling moments in the preset historical period. express The distribution probability value of express The distribution probability value of express and The joint distribution probability value of ; then enter step C2;

[0024] Step C2: Screen and obtain the preliminary selected sensors whose correlation values ​​are greater than a preset correlation threshold value to form the target sensors corresponding to the components.

[0025] As a preferred technical solution of the present invention: the target network to be trained in step D is connected in series in sequence from the input end to the output end, including 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, wherein the stacked long short-term memory network includes two long short-term memory networks connected in series in sequence from the input end to the output end, the composite fully connected layer includes a first fully connected layer and an activation layer connected in series in sequence 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 receives and forwards the input data received by the input layer. , , do the following; The number of target sensors corresponding to the component is increased by 1;

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

[0028]

[0029] get , Represents input data About the highest expansion order The expansion of each order of ;

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

[0031]

[0032] get , and then obtain , and output to the stacked long short-term memory network, where represents the first Layer corresponds to The output at each sampling time, Indicates based on , The number of monomials in the polynomial layer of , represents the first The weight vector of the layer, express Middle weights, are the parameters 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 first long short-term memory network receives the output of the polynomial layer ,according to ,get , and output to the second long short-term memory network in sequence, and the second long short-term memory network in sequence receives the output of the first long short-term memory network in sequence ,according to ,get , and output to the composite fully connected layer, where represents the first long short-term memory network function in sequence, Represents the sequential second LSTM network function;

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

[0035]

[0036] get , and output to the activation layer, where represents the first fully connected layer function, Indicates the first fully connected layer Layer corresponds to The output at each sampling time, ;

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

[0038]

[0039] get , and output to the Dropout layer, where represents the activation layer function, Indicates the activation layer Layer corresponds to The output at each 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] get , and output to the fusion layer, where represents the deactivation layer function, ;

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

[0044]

[0045] Perform fusion processing to obtain , and output to the second fully connected layer, where , represents the dimension of the fusion layer output, Indicates the first Layer corresponds to The output at each 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] get , and output to the linear transformation layer; where Represents the second fully connected layer function; Indicates the second fully connected layer Layer corresponds to The output at each 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] get Output, where Indicates that the linear transformation layer corresponds to the The output at each sampling time, and Represents 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, firstly according to the following MSE loss function,

[0053]

[0054] Get the loss structure ,in, Indicates that the target network to be trained corresponds to Next iteration The actual value of the samples, Indicates that the target network to be trained corresponds to Next iteration The predicted value of samples, represents the number of samples used for training, Indicates The loss value at the iteration;

[0055] Then, the data values ​​of the coefficients of the polynomial layer in the target network to be trained and the data values ​​of the weights in the stacked long short-term memory network are updated according to the following formulas;

[0056]

[0057] in, Indicates the corresponding The data value of the iteration, Indicates the corresponding The updated data value of the iteration, express right The gradient of Represents the learning rate.

[0058] The power load forecasting method based on the SLSTM-RMTN model described in the present invention adopts the above technical solution and has the following technical effects compared with the prior art:

[0059] (1) The present invention designs a power load forecasting method based on the SLSTM-RMTN model. For the target power system, the design starts with the STL decomposition of the data value, selects the target sensors corresponding to the trend component, seasonal component and residual component respectively, and trains the target to-be-trained network composed of the long short-term memory network, residual connection and polynomial expansion to obtain the power load forecasting model under each component. Then, in the prediction application, the prediction of the power load forecasting model under each component is executed for the sampled data, and the prediction is integrated to complete the power load forecasting of the target power system. In the application, the design scheme 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 optimize the allocation of power resources and balance the supply and demand relationship. BRIEF DESCRIPTION OF THE DRAWINGS

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

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

[0062] The specific implementation modes of the present invention will be further described in detail below in conjunction with the accompanying drawings.

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

[0064] Step A. Obtain the data values ​​of each preset preliminary sensor and the data value of the power load at each sampling time in the preset historical period corresponding to the target power system, and then use each preliminary sensor and the power load as the preprocessing object, respectively, and execute the following steps A1 to A3, preprocess and update the data value of the preprocessing object corresponding to each sampling time, and then enter step B.

[0065] Step A1. Perform denoising and updating on the data values ​​of the preprocessing object corresponding to each sampling time, and then proceed to step A2.

[0066] Step A2. According to the preset normal value range corresponding to the preprocessed object, determine whether there are abnormal values ​​in the data values ​​corresponding to each sampling time of the preprocessed object. If yes, replace and update the abnormal value with the average value of each normal value. Otherwise, do nothing and then enter step A3.

[0067] Step A3: normalize and update the data values ​​of the preprocessed object at each sampling time so that 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 time, the following formula is used:

[0069]

[0070] Perform STL (Seasonal and Trend decomposition using Loess) decomposition to obtain data values The data values ​​corresponding to the trend components , seasonal component data values , the data value of the residual component ,in, Indicates the corresponding The data value at the sampling moment, Indicates the corresponding The data value of the trend component corresponding to the data value at each sampling moment, Indicates the corresponding The data value of the seasonal component corresponding to the data value at the sampling time, Indicates the corresponding The data value of the residual component corresponding to the data value at each sampling moment, and then enters step C.

[0071] Step C. For the trend component, the seasonal component, and the residual component, respectively, the preliminary selected sensors whose correlation values ​​with the power load under the component are greater than the preset correlation threshold are screened to form the target sensors corresponding to the component, and then proceed to step D.

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

[0073] Step C1. For each pre-selected sensor, use the following formula:

[0074]

[0075] Obtain the correlation value of the corresponding component between the primary selected sensor and the power load ,in, , Indicates the number of primary selected sensors, Indicates The data value sequence at each sampling time under the corresponding component of the primary selected sensor, express Middle The data value at the sampling moment, Indicates power load The data value sequence at each sampling time under the corresponding component, express Middle The data value at the sampling moment, express and The mutuality value between Represents "Mutual Information", that is, mutual information, which is used to measure and interdependence; based on , Indicates the number of sampling moments in the preset historical period. express The distribution probability value of express The distribution probability value of express and The joint distribution probability value of ; then enter step C2.

[0076] Step C2: Screen and obtain the preliminary selected sensors whose correlation values ​​are greater than a preset correlation threshold value to form the target sensors corresponding to the components.

[0077] Step D. For the trend component, seasonal component, and residual component, the target sensors corresponding to the components and the power loads are respectively corresponding to the preset quantities under the components. The data value of the continuous sampling moment is input, combined with the corresponding component of the power load, The data value at each sampling moment is output, and the network to be trained is trained for the preset target to obtain the power load forecasting model under the corresponding component of the target power system, and then the power load forecasting model of the target power system corresponding to the trend component, seasonal component and residual component is obtained, which are combined to form the power load forecasting model corresponding to the target power system.

[0078] In practical applications, the above step D is Figure 2 As shown, the target network to be trained is connected in series from the input end to the output end in sequence, including 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. The stacked long short-term memory network includes two long short-term memory networks 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 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, and the fusion layer adds the output of the input layer and the output of the Dropout layer (random inactivation layer), and outputs the addition result to the input end of the second fully connected layer.

[0079] In the specific data processing process of the target network to be trained, the polynomial layer receives and forwards the input data received by the input layer. , , do the following; The number of target sensors corresponding to the component is increased by 1;

[0080] First, according to the preset maximum expansion order corresponding to the target network to be trained , according to the following formula:

[0081]

[0082] get , Represents input data About the highest expansion order The expansion of each term of .

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

[0084]

[0085] get , and then obtain , and output to the stacked long short-term memory network, where represents the first Layer corresponds to The output at each sampling time, Indicates based on , The number of monomials in the polynomial layer of , represents the first The weight vector of the layer, express Middle weights, are the parameters to be trained in the polynomial layer.

[0086] In the stacked LSTM network, the first LSTM network receives the output of the polynomial layer. ,according to ,get , and output to the second long short-term memory network in sequence, and the second long short-term memory network in sequence receives the output of the first long short-term memory network in sequence ,according to ,get , and output to the composite fully connected layer, where represents the first long short-term memory network function in sequence, Represents the sequential second LSTM network function.

[0087] In the design of the stacked LSTM network here, the first LSTM network extracts short-term temporal dependencies within the time step range, analyzes the feature sequence at each time step, and passes it to the second LSTM network through the hidden state for higher-level processing. This nonlinear transformation of the input features can capture complex temporal dependencies. The second LSTM network 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 richer feature expressions for subsequent prediction tasks.

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

[0089]

[0090] get , and output to the activation layer, where represents the first fully connected layer function, Indicates the first fully connected layer Layer corresponds to The output at each sampling time, .

[0091] The activation layer then receives the output of the first fully connected layer , according to the following formula:

[0092]

[0093] get , and output to the Dropout layer (random inactivation layer), where represents the activation layer function, Indicates the activation layer Layer corresponds to The output at each sampling moment.

[0094] The activation function tanh is introduced after the calculation of the first fully connected layer in the composite fully connected layer to map the input data to the range of [-1,1]. Its nonlinear characteristics enable the model to better capture the complex input-output mapping relationship.

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

[0096]

[0097] get , and output to the fusion layer, where represents the deactivation layer function, ,Here the random dropout technique is used to improve the generalization ability of the model by randomly discarding some neurons during training.

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

[0099]

[0100] Perform fusion processing to obtain , and output to the second fully connected layer, where , represents the dimension of the fusion layer output, Indicates the first Layer corresponds to The output at each sampling moment.

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

[0102]

[0103] get , and output to the linear transformation layer; where Represents the second fully connected layer function; Indicates the second fully connected layer Layer corresponds to The output at each 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] get Output, where Indicates that the linear transformation layer corresponds to the The output at each sampling time, and Represents the weight matrix and bias term of the linear transformation layer.

[0107] And in the practical application of the above step D, before training the target network to be trained, the model parameters including the number of neurons in the stacked LSTM network, the polynomial coefficients in the residual multidimensional Taylor network and other related parameters are randomly initialized. The purpose of random initialization is to avoid the model parameters from falling into the local minimum and improve the training efficiency, and the weights are usually initialized by normal distribution or uniform distribution. This initialization method can ensure that the gradient propagation of the network is relatively stable in the early stage of training, which helps to accelerate the convergence of the model.

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

[0109]

[0110] Get the loss structure ,in, Indicates that the target network to be trained corresponds to Next iteration The actual value of the samples, Indicates that the target network to be trained corresponds to Next iteration The predicted value of samples, represents the number of samples used for training, Indicates The loss value at the iteration.

[0111] Then, the data values ​​of the coefficients of the polynomial layer in the target network to be trained and the data values ​​of the weights in the stacked long short-term memory network are updated according to the following formulas;

[0112]

[0113] in, Indicates the corresponding The data value of the iteration, Indicates the corresponding The updated data value of the iteration, express right The gradient of Represents the learning rate.

[0114] In the specific training process, the parameters of the model (including the weights of the LSTM network and the polynomial coefficients in the RMTN) are updated through the Adam optimizer and the back-propagation 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. In the back-propagation process, the gradient is transmitted backward from the output layer to the input layer layer by 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, the power load forecasting model corresponding to the target power system is obtained, and then in the application, the execution steps are further designed. To step , power load forecasting is carried out for the target power system.

[0116] step . Summarize the target sensors corresponding to the trend component, seasonal component, and residual component, and based on the historical time direction from the current sampling time The target sampling moments are composed of continuous sampling moments, and the data values ​​of all target sensors and the data values ​​of the power load of the target power system corresponding to each target sampling moment are collected, and then the step .

[0117] step . Perform STL (Seasonal and Trend decomposition using Loess) decomposition on the data values ​​of all target sensors and the data values ​​of power load at each target sampling time, and obtain the data values ​​corresponding to the trend component, the seasonal component, and the residual component, and then enter step .

[0118] step . Take 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, take the data values ​​of each target sensor corresponding to the component at each target sampling moment, and the data values ​​of the corresponding components of the power load as input, apply the power load prediction model under the corresponding components of the target power system, obtain the data values ​​of the corresponding components of the power load at the prediction moment, and then obtain the data values ​​of the trend component, seasonal component, and residual component of the power load at the prediction moment, and then enter the step .

[0119] step . For the data values ​​of the trend component, seasonal component, and residual component corresponding to the power load at the forecast time, an addition process is performed to obtain the power load of the target power system corresponding to the forecast time.

[0120] The above technical scheme designs a power load forecasting method based on the SLSTM-RMTN model. For the target power system, the design starts with the STL decomposition of the data value, and selects the target sensors corresponding to the trend component, seasonal component, and residual component respectively. The target to-be-trained network composed of the 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, the prediction of the power load forecasting model under each component is executed for the sampled data, and the prediction is integrated to complete the power load forecasting of the target power system. In the application, the design scheme 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 allocation of power resources and balance the supply and demand relationship.

[0121] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. A power load forecasting method based on the SLSTM-RMTN model, characterized by: Execute the following steps A to D to obtain the power load forecasting model corresponding to the target power system, and then execute step To step , power load forecasting for the target power system; Step A. Obtaining the data values ​​of each preset primary sensor and the data value of the power load at each sampling time in the preset historical period corresponding to the target power system, and performing data preprocessing and updating, and then proceeding to step B; Step B. Perform STL decomposition on the data values ​​of each primary sensor and the data value of the power load at each sampling time, obtain the data values ​​corresponding to the trend component, the seasonal component, and the residual component, and then proceed to step C; Step C. For the trend component, the seasonal component, and the residual component, respectively, the preliminary selected sensors whose correlation values ​​with the power load under the component are greater than the preset correlation threshold are screened to form the target sensors corresponding to the component, and then proceed to step D; Step D. For the trend component, seasonal component, and residual component, the target sensors corresponding to the components and the power loads are respectively corresponding to the preset quantities under the components. The data value of the continuous sampling moment is input, combined with the corresponding component of the power load, The data value at each sampling moment is output, and the target network to be trained is trained to obtain the power load forecasting model under the corresponding component of the target power system, and then the power load forecasting model under the corresponding trend component, seasonal component, and residual component of the target power system is obtained, and the power load forecasting model corresponding to the target power system is combined to form the power load forecasting model corresponding to the target power system; step . Summarize the target sensors corresponding to the trend component, seasonal component, and residual component, and based on the historical time direction from the current sampling time The target sampling moments are composed of continuous sampling moments, and the data values ​​of all target sensors and the data values ​​of the power load of the target power system corresponding to each target sampling moment are collected, and then the step ; step . Perform STL decomposition on the data values ​​of all target sensors and the data values ​​of power load at each target sampling time, obtain the data values ​​corresponding to the trend component, the seasonal component, and the residual component, and then enter step ; step . Take 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, take the data values ​​of each target sensor corresponding to the component at each target sampling moment, and the data values ​​of the corresponding components of the power load as input, apply the power load prediction model under the corresponding components of the target power system, obtain the data values ​​of the corresponding components of the power load at the prediction moment, and then obtain the data values ​​of the trend component, seasonal component, and residual component of the power load at the prediction moment, and then enter the step ; step . For the data values ​​of the trend component, seasonal component, and residual component corresponding to the power load at the forecast time, an addition process is performed to obtain the power load of the target power system corresponding to the forecast time.

2. According to claim 1, a power load forecasting method based on the SLSTM-RMTN model is characterized in that: In the step A, after obtaining the data values ​​of each pre-selected sensor and the data values ​​of the power load at each sampling time in the preset historical period corresponding to the target power system, each pre-selected sensor and the power load are respectively used as pre-processing objects, and the following steps A1 to A3 are performed to pre-process and update the data values ​​of the pre-processing objects corresponding to each sampling time; Step A1. Execute denoising update for the data value of the preprocessing object corresponding to each sampling time, and then proceed to step A2; Step A2. According to the preset normal value range corresponding to the preprocessed object, determine whether there are abnormal values ​​in the data values ​​corresponding to each sampling time of the preprocessed object, and if so, replace and update the abnormal value with the average value of each normal value, otherwise, do nothing, and then proceed to step A3; Step A3: normalize and update the data value of the preprocessing object corresponding to each sampling time.

3. According to the method for power load forecasting based on the SLSTM-RMTN model of claim 1, it is characterized by: In step B, the data values ​​of the primary selected sensors and the data values ​​of the power load at each sampling time are respectively calculated according to the following formula: ； Perform STL decomposition to obtain data values The data values ​​corresponding to the trend components , seasonal component data values , the data value of the residual component ,in, Indicates the corresponding The data value at the sampling moment, Indicates the corresponding The data value of the trend component corresponding to the data value at each sampling moment, Indicates the corresponding The data value of the seasonal component corresponding to the data value at the sampling time, Indicates the corresponding The data value of the residual component corresponding to the data value at each sampling moment.

4. According to claim 1, a method for predicting power load based on the SLSTM-RMTN model is characterized by: In the step C, the following steps C1 to C2 are performed for the trend component, the seasonal component, and the residual component respectively; Step C1. For each pre-selected sensor, use the following formula: ； Obtain the correlation value of the corresponding component between the primary selected sensor and the power load ,in, , Indicates the number of primary selected sensors, Indicates The data value sequence at each sampling time under the corresponding component of the primary selected sensor, express Middle The data value at the sampling moment, Indicates power load The data value sequence at each sampling time under the corresponding component, express Middle The data value at the sampling moment, express and The mutuality value between , Indicates the number of sampling moments in the preset historical period. express The distribution probability value of express The distribution probability value of express and The joint distribution probability value of ; then enter step C2; Step C2: Screen and obtain the preliminary selected sensors whose correlation values ​​are greater than a preset correlation threshold value to form the target sensors corresponding to the components.

5. According to claim 1, a method for predicting power load based on the SLSTM-RMTN model is characterized in that: The target network to be trained in step D is connected in series in sequence from the input end to the output end, including 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. The stacked long short-term memory network includes two long short-term memory networks connected in series in sequence from the input end to the output end, the composite fully connected layer includes a first fully connected layer and an activation layer connected in series in sequence 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.

6. According to claim 5, a method for predicting power load based on the SLSTM-RMTN model is characterized in that: The polynomial layer receives and forwards the input data from the input layer. , , do the following; The number of target sensors corresponding to the component is increased by 1; First, according to the preset maximum expansion order corresponding to the target network to be trained , according to the following formula: ； get , Represents input data About the highest expansion order The expansion of each order of ; Then according to , , , according to the following formula: ； get , and then obtain , and output to the stacked long short-term memory network, where represents the first Layer corresponds to The output at each sampling time, Indicates based on , The number of monomials in the polynomial layer of , represents the first The weight vector of the layer, express Middle weights, are the parameters to be trained in the polynomial layer.

7. According to claim 6, a method for predicting power load based on the SLSTM-RMTN model is characterized by: In the stacked long short-term memory network, the first long short-term memory network receives the output of the polynomial layer. ,according to ,get , and outputs to the second long short-term memory network in sequence, and the second long short-term memory network in sequence receives the output of the first long short-term memory network in sequence ,according to ,get , and output to the composite fully connected layer, where represents the first long short-term memory network function in sequence, Represents the sequential second long short-term memory network function; In the composite fully connected layer, the first fully connected layer first receives the output of the stacked long short-term memory network. , according to the following formula: ； get , and output to the activation layer, where represents the first fully connected layer function, Indicates the first fully connected layer Layer corresponds to The output at each sampling time, ; The activation layer then receives the output of the first fully connected layer , according to the following formula: ； get , and output to the Dropout layer, where represents the activation layer function, Indicates the activation layer Layer corresponds to The output at each sampling moment.

8. According to claim 7, a method for predicting power load based on the SLSTM-RMTN model is characterized by: The Dropout layer receives the output of the composite fully connected layer , according to the following formula: ； get , and output to the fusion layer, where represents the deactivation layer function, ; The fusion layer receives the output of the deactivation layer , and receive input data forwarded by the input layer , according to the following formula: ； Perform fusion processing to obtain , and output to the second fully connected layer, where , represents the dimension of the fusion layer output, Indicates the first Layer corresponds to The output at each sampling moment.

9. The power load forecasting method based on the SLSTM-RMTN model according to claim 8 is characterized by: The second fully connected layer receives the output of the fusion layer , according to the following formula: ； get , and output to the linear transformation layer; where Represents the second fully connected layer function; Indicates the second fully connected layer Layer corresponds to The output at each sampling moment; The linear transformation layer receives the output of the second fully connected layer , and based on , according to the following formula: ； get Output, where Indicates that the linear transformation layer corresponds to the The output at each sampling time, and Represents the weight matrix and bias term of the linear transformation layer.

10. The power load forecasting method based on the SLSTM-RMTN model according to claim 1 is characterized in that: In the step D, according to the Adam optimizer, during the training of the preset target network to be trained, for each iteration, firstly according to the following MSE loss function, ； Get the loss structure ,in, Indicates that the target network to be trained corresponds to Next iteration The actual value of the samples, Indicates that the target network to be trained corresponds to Next iteration The predicted value of samples, represents the number of samples used for training, Indicates The loss value at the iteration; Then, the data values ​​of the coefficients of the polynomial layer in the target network to be trained and the data values ​​of the weights in the stacked long short-term memory network are updated according to the following formulas; ； in, Indicates the corresponding The data value of the iteration, Indicates the corresponding The updated data value of the iteration, express right The gradient of Represents the learning rate.

Images