Low-voltage distribution area load prediction method and device

Abstract

The invention discloses a load prediction method and device for a low-voltage distribution area, and the method comprises the following steps: collecting historical climate data and historical load data of the low-voltage distribution area, the historical climate data comprising historical average daily temperature, average daily wind speed, average daily precipitation and average daily humidity; the historical load data comprises previous daily average load electric quantity data; processing the historical climate data and the historical load data to obtain a data set; constructing long-term and short-term time sequence network models of cross-dimension attention, and training the long-term and short-term time sequence network models of the cross-dimension attention by using the data set; and predicting future load data of the low-voltage distribution area by using the trained cross-dimension attention long-term and short-term time sequence network models. According to the low-voltage distribution area load prediction method provided by the invention, the precision of load prediction is remarkably improved by fusing climatic factors, optimizing a signal decomposition algorithm and innovating a prediction model structure.

Description

Technical Field

[0001] The present invention relates to the field of power systems, and particularly to a method and device for load forecasting in a low-voltage distribution substation area. Background Art

[0002] Load forecasting is a key link to ensure the balance of power supply. Affected by weather changes, emergency response requirements in special situations, etc., the loads such as cooling and heating electricity consumption and industrial electricity consumption have strong volatility. With the continuous growth of the installed capacity of new energy and the gradual increase in the proportion of new energy access to the power grid, it has brought great challenges to the accurate forecasting of loads, timely ensuring the power supply to users, and improving the reliability of user power supply.

[0003] At present, the forecasting of power loads is mostly based on historical data in the past, which rarely considers the regional climate environment of the network. According to the analysis of the actual load curve, climate factors, industrial load conditions, etc. have great impacts on load forecasting. Accurately forecasting the power load on the user side can not only timely adjust the power generation plan and guide enterprises to conduct power dispatching to ensure the stability and reliability of power supply, but also optimize the allocation of power resources and improve the overall operation efficiency of the power system. Therefore, developing a power load forecasting technology that comprehensively considers various influencing factors has important practical significance and broad application prospects. Summary of the Invention

[0004] Aiming at the deficiencies of the above-mentioned prior art, the technical problem to be solved by the present invention is: to provide a method and device for load forecasting in a low-voltage distribution substation area that can accurately forecast the power load on the user side.

[0005] To solve the above technical problem, a technical solution adopted by the present invention is: to provide a method for load forecasting in a low-voltage distribution substation area, including the following steps:

[0006] Collect historical climate data and historical load data of the low-voltage distribution substation area, where the historical climate data includes historical daily average temperature, daily average wind speed, daily average precipitation, and daily average humidity, and the historical load data includes historical daily average load power data;

[0007] Process the historical climate data and historical load data to obtain a data set;

[0008] Construct a long-term and short-term time series network model with cross-dimensional attention and use the data set to train the long-term and short-term time series network model with cross-dimensional attention;

[0009] Use the trained long-term and short-term time series network model with cross-dimensional attention to forecast the future load data of the low-voltage distribution substation area.

[0010] Further, in the step of processing the historical climate data and historical load data to obtain a data set, the following sub-steps are included:

[0011] Preprocess the historical climate data;

[0012] Decompose the historical daily average load power data by using the successive variational mode decomposition algorithm optimized by the Harris hawk optimization algorithm to obtain several continuous periodic components;

[0013] Arrange the decomposed components and the preprocessed historical climate data in chronological order and integrate them to obtain a multi-dimensional data set.

[0014] Further, in the step of preprocessing the historical climate data, the following sub-steps are included:

[0015] Use the method of finding the mean of neighboring numbers to fill in the missing data in the historical climate data to obtain the data after filling;

[0016] Use the method of maximum-minimum normalization to normalize the data after filling to obtain the preprocessed historical climate data.

[0017] Further, in the step of decomposing the historical daily average load power data by using the successive variational mode decomposition algorithm optimized by the Harris hawk optimization algorithm to obtain several continuous periodic components, the following sub-steps are included:

[0018] Set the objective function of the Harris hawk optimization algorithm, and the objective function is the minimum envelope entropy;

[0019] Set the population size, maximum number of iterations of the Harris hawk optimization algorithm and the range of the parameters to be optimized, and optimize the parameters. The parameter to be optimized is the maximum penalty factor of the successive variational mode decomposition;

[0020] Take the optimal value of the maximum penalty factor obtained by the Harris hawk optimization as the decomposition parameter of the successive variational mode decomposition, and decompose the data sequence composed of the historical daily average load power data to obtain several continuous periodic components.

[0021] Further, the method of optimizing the maximum penalty factor by using the Harris hawk optimization algorithm specifically includes:

[0022] Initialize the maximum penalty factor and substitute the initial value of the maximum penalty factor into the successive variational mode decomposition algorithm to decompose the historical daily average load power data;

[0023] Calculate the objective function value according to the number of decomposition layers and components;

[0024] Determine whether the current objective function value is the minimum value during the iteration process. If it is the minimum value, take the corresponding maximum penalty factor as the result of the current optimization. Otherwise, update the prey escape energy and the escape probability;

[0025] Adjust the search strategy according to the absolute value of the prey escape energy, and update the maximum penalty factor according to the adjusted search strategy;

[0026] Determine whether the maximum number of iterations is reached. If the maximum number of iterations is not reached, substitute the latest output maximum penalty factor into the successive variational mode decomposition algorithm to decompose the daily average load power data in the past period, and return to the step of calculating the objective function value according to the number of decomposition layers and components. Otherwise, terminate the iteration and take the maximum penalty factor corresponding to the minimum objective function value during the iteration as the final optimization result.

[0027] Furthermore, the objective function is expressed as:

[0028]

[0029] In Equation (1), F obj represents the objective function, K represents the number of decomposition layers, hilbert represents the Hilbert transform, imf(k) represents the k-th component after the decomposition of the original signal, abs represents taking the absolute value, sum represents summation,.* represents dot multiplication operation, and amp(k) represents the amplitude of the k-th imf component.

[0030] Furthermore, the specific structure of the long-term and short-term time series network model of the cross-dimensional attention includes:

[0031] A convolutional layer for extracting local features of the input data. The convolutional layer is a one-dimensional convolutional layer;

[0032] A long short-term memory network for modeling the local features extracted by the convolutional layer to achieve the superposition of long-term and short-term memory of feature information;

[0033] A cross-attention module for taking the output of the last time step of the long short-term memory network to form a matrix, calculating the weight scores between time steps through matrix operations, and performing weighted fusion on the features according to the weight scores;

[0034] An autoregressive unit for receiving the output of the cross-attention module, capturing the linear correlation relationship in multi-dimensional data, and correcting the predicted result;

[0035] A fully connected layer for mapping the features output by the cross-attention module to the prediction result and adding it to the output of the autoregressive unit to obtain the final prediction value.

[0036] Furthermore, the calculation process of the cross-attention module includes:

[0037] Take the output of the last time step of the long and short time series network to form a first matrix;

[0038] Take the outputs of all time steps of the long and short time series network to form a second matrix and a third matrix respectively, and the information contained in the second matrix and the third matrix is the same;

[0039] Transpose the third matrix to obtain a transposed matrix, and calculate the cross-weight score according to the first matrix and the transposed matrix;

[0040] Normalize the cross-weight score to obtain a weight matrix;

[0041] Multiply the weight matrix by the second matrix to calculate the output matrix of the last time step after cross-attention processing.

[0042] Furthermore, when training the long-term and short-term time series network model of cross-dimensional attention using the dataset, the mean squared error is used as the loss function for training, and its expression is:

[0043]

[0044] In formula (2), MSE train represents the loss function for training the long-term and short-term time series network model of cross-dimensional attention, n represents the length of the training data, and p i represents the actual load data in the training data, is the load data predicted by the long-term and short-term time series network model of cross-dimensional attention during the training process.

[0045] To solve the above technical problems, another technical solution adopted by the present invention is: to provide a load prediction device for a low-voltage distribution substation area, including:

[0046] A data acquisition module for collecting historical climate data and historical load data of the low-voltage distribution substation area. The historical climate data includes historical daily average temperature, daily average wind speed, daily average precipitation, and daily average humidity, and the historical load data includes historical daily average load power data;

[0047] A data processing module for processing the historical climate data and historical load data to obtain a dataset;

[0048] A model construction and training module for constructing a long-term and short-term time series network model of cross-dimensional attention and training the long-term and short-term time series network model of cross-dimensional attention using the dataset;

[0049] A prediction module, which is used to predict the future load data of a low-voltage distribution substation area by using a trained long-term and short-term time series network model with cross-dimensional attention.

[0050] The load prediction method and device for a low-voltage distribution substation area of the present invention have at least the following beneficial effects: By integrating historical climate data and historical load data, the present invention constructs a multi-dimensional data set, effectively capturing the correlation between climate fluctuations and electricity consumption behavior. Compared with traditional methods that only rely on historical loads, this solution can accurately reflect the impact of extreme weather and seasonal changes on electricity demand, is applicable to the prediction of volatile loads in scenarios with high penetration of new energy, and the prediction results are more in line with the actual power supply demand; The historical load data is decomposed by the successive variational mode decomposition algorithm optimized by the Harris hawk algorithm, overcoming the defect that the traditional successive variational mode decomposition relies on manual experience for parameter tuning. The decomposed IMF components have both low noise interference and high physical significance, providing clear multi-scale feature inputs for subsequent models and reducing the risk of overfitting; The designed cross-dimensional attention mechanism solves the problem that traditional models are insufficient in capturing local mutations or long-term trends by dynamically weighted fusion of long-term and short-term time series features output by the long and short time series networks. Combined with the compensation and correction of the linear relationship by the autoregressive unit, the model can simultaneously depict the non-linear dynamics and periodic laws of load data, and can reduce the prediction error in complex fluctuation scenarios. Description of the Drawings

[0051] The drawings described herein are used to provide a further understanding of the present application, and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation to the present application. In the drawings:

[0052] Figure 1 It is a flowchart of an implementation manner of the load prediction method for a low-voltage distribution substation area of the present invention.

[0053] Figure 2 It is a curve of the climate data of the low-voltage distribution substation area collected in the past year in this implementation manner.

[0054] Figure 3 It is Figure 1 The flowchart of step S200 in

[0055] Figure 4 It is Figure 2 The flowchart of step S220 in

[0056] Figure 5 It is a curve graph of the change of the objective function value of the maximum penalty factor iteratively optimized by the Harris hawk optimization algorithm.

[0057] Figure 6 It is a curve graph of each IMF component obtained by decomposing the daily average load power data in the past period by the successive variational mode decomposition algorithm optimized by the Harris hawk optimization algorithm.

[0058] Figure 7 It is a structural diagram of a long - term and short - term time - series network model for cross - dimensional attention.

[0059] Figure 8 It is a comparison chart of the daily average load power data and the real data of the low - voltage distribution transformer area in the future for a certain period predicted.

[0060] Figure 9 It is a structural block diagram of an embodiment of the load prediction device for the low - voltage distribution transformer area of the present invention. Specific embodiments

[0061] The present invention will be further described below with reference to the accompanying drawings.

[0062] Please refer to Figure 1 , which is a flowchart of an embodiment of the load prediction method for the low - voltage distribution transformer area of the present invention. This embodiment specifically includes the following steps:

[0063] S100. Collect data.

[0064] Specifically, collect the historical climate data and historical load data of the low - voltage distribution transformer area. The historical climate data includes historical daily average temperature, daily average wind speed, daily average precipitation, and daily average humidity. The historical load data includes past daily average load power data. Please refer to Figure 2 , which is the climate data curve of the past year of the low - voltage distribution transformer area collected in this embodiment.

[0065] S200. Process data.

[0066] Specifically, process the historical climate data and historical load data to obtain a data set.

[0067] Please refer to Figure 3 , this step S200 includes the following sub - steps:

[0068] S210. Pre - process the historical climate data.

[0069] Specifically, first use the method of finding the mean of neighboring numbers to fill in the missing data in the historical climate data to obtain the data after filling; then use the maximum - minimum normalization method to normalize the data after filling to obtain the pre - processed historical climate data. The normalization formula is:

[0070]

[0071] Among them, y represents the normalized data, X max represents the maximum value in the historical climate data, X minRepresents the minimum value in the historical climate data. For example, when normalizing the historical daily average temperature, X max represents the maximum value in the historical daily average temperature, X min represents the minimum value in the historical daily average temperature.

[0072] S220. Decompose the historical daily average load power data.

[0073] Specifically, use the successive variational mode decomposition algorithm optimized by the Harris hawk optimization algorithm to decompose the historical daily average load power data to obtain several continuous periodic components. Please refer to Figure 4 , this step S220 includes the following sub-steps:

[0074] S221. Set the objective function.

[0075] Specifically, set the objective function of the Harris hawk optimization algorithm, and the objective function is the minimum envelope entropy. The objective function is as follows:

[0076]

[0077] where, F obj represents the objective function, K represents the decomposition layer number, hilbert represents the Hilbert transform, imf(k) represents the k-th component after the decomposition of the original signal, abs represents taking the absolute value, sum represents summation,.* represents dot multiplication operation, and amp(k) represents the amplitude of the k-th imf component.

[0078] S222. Optimize the maximum penalty factor.

[0079] Specifically, set the population size, maximum iteration number of the Harris hawk optimization algorithm and the range of the parameter to be optimized, and optimize the parameter. The parameter to be optimized is the maximum penalty factor of the successive variational mode decomposition. Please refer to Figure 5, which is a curve graph showing the change of the objective function value of the maximum penalty factor obtained by iteratively optimizing using the Harris hawk optimization algorithm. The method for optimizing the maximum penalty factor using the Harris hawk optimization algorithm is specifically as follows: Initialize the maximum penalty factor and substitute the initial value of the maximum penalty factor into the successive variational mode decomposition algorithm to decompose the daily average load power data of previous periods. Calculate the objective function value according to the number of decomposition layers and components; Determine whether the current objective function value is the minimum value during the iteration. If it is the minimum value, take the corresponding maximum penalty factor as the current optimization result. Otherwise, update the prey escape energy and escape probability. Adjust the search strategy according to the absolute value of the prey escape energy, and update the maximum penalty factor according to the adjusted search strategy. When it is judged whether the absolute value of the prey escape energy is greater than or equal to 1, if the absolute value of the prey escape energy is greater than or equal to 1, enter the exploration stage and update the maximum penalty factor. Otherwise, enter the exploitation stage and update the maximum penalty factor. Determine whether the maximum number of iterations is reached. If the maximum number of iterations is not reached, substitute the latest output maximum penalty factor into the successive variational mode decomposition algorithm to decompose the daily average load power data of previous periods, and return to the step of calculating the objective function value according to the number of decomposition layers and components. Otherwise, terminate the iteration and take the maximum penalty factor corresponding to the minimum objective function value during the iteration as the final optimization result.

[0080] S223. Substitute the optimal value of the maximum penalty factor for decomposition.

[0081] Specifically, take the optimal value of the maximum penalty factor obtained by the Harris hawk optimization as the decomposition parameter of the successive variational mode decomposition, and decompose the data sequence composed of the daily average load power data of previous periods to obtain several continuous periodic components imf. Please refer to Figure 6 , which is a curve graph showing each imf component obtained by decomposing the daily average load power data of previous periods using the successive variational mode decomposition algorithm optimized by the Harris hawk optimization algorithm.

[0082] S230. Integrate the data.

[0083] Specifically, arrange the decomposed components and the preprocessed historical climate data in chronological order and integrate them to obtain a multi-dimensional data set.

[0084] S300. Construct and train the model.

[0085] Specifically, construct a long-term and short-term time series network model with cross-dimensional attention and train the long-term and short-term time series network model with cross-dimensional attention using the data set. Please refer to Figure 7 , The specific structure of the long-term and short-term time series network model with cross-dimensional attention includes a convolutional layer, a long short-term memory network (LSTM), a cross-attention module, an autoregressive unit, and a fully connected layer. Among them:

[0086] The convolutional layer is used to extract local features of the input data, and the convolutional layer is a one-dimensional convolutional layer. In order to ensure that the extracted information is not lost, in this embodiment, the number of convolutional layer kernels is selected to be 1. The data dimension is (batch_size, length, dim), where batch_size represents the batch processing size of the network, length represents the length of the input sequence, and dim represents the dimension of the multi-dimensional input data.

[0087] The long short-term memory network is used to model the local features extracted by the convolutional layer to achieve the superposition of long-term and short-term memories of feature information. The long short-term memory network is composed of multiple long short-term memory units, and each unit includes an input gate, a forget gate, and an output gate. Multiple long short-term memory units achieve the superposition of long-term and short-term memories of feature information. The output data dimension is (batch_size, length, hidden_dim), where hidden_dim represents the dimension of the hidden layer in the long short-term memory network, that is, the number of long short-term memory units.

[0088] The cross-attention module is used to take the output of the last time step of the long short-term memory network to form a matrix, calculate the weight scores between time steps through matrix operations, and perform weighted fusion on the features according to the weight scores. The calculation process of the cross-attention module includes: taking the output of the last time step of the long short-term memory network to form a first matrix; taking the outputs of all time steps of the long short-term memory network to form a second matrix and a third matrix respectively, and the second matrix and the third matrix contain the same information; transposing the third matrix to obtain a transposed matrix, and calculating the cross-weight scores according to the first matrix and the transposed matrix; normalizing the cross-weight scores to obtain a weight matrix; multiplying the weight matrix by the second matrix to calculate the output matrix of the last time step after being processed by the cross-attention module.

[0089] The calculation formula of the cross-weight score is as follows:

[0090]

[0091] Among them, Score_C represents the cross-weight score, Q represents the first matrix, H represents the third matrix, and H T represents the transposed matrix of the third matrix, represents the element-wise multiplication of matrices.

[0092] The cross-weight scores are sized (batch_size, 1, length). When normalizing the cross-weight scores to obtain the weight matrix, the specific operation is as follows: perform a softmax function on them to normalize the cross-weight scores into a weight matrix weights with a range of 0-1, and the size of weights is (batch_size, 1, length).

[0093] The calculation formula for the output matrix at the last time step after being processed by the cross-attention module is as follows:

[0094]

[0095] Among them, Q' represents the output matrix at the last time step after being processed by the cross-attention module, and V represents the second matrix.

[0096] The autoregressive unit is used to receive the output of the cross-attention module, capture the linear correlation in multi-dimensional data, and correct the predicted results.

[0097] The fully connected layer is used to map the features output by the cross-attention module to the prediction results, and add them to the output of the autoregressive unit to obtain the final predicted value.

[0098] When training the long-term and short-term time series network model of cross-dimensional attention using the dataset, the mean squared error is used as the loss function for training, and its expression is:

[0099]

[0100] Among them, MSE train represents the loss function for training the long-term and short-term time series network model of cross-dimensional attention, n represents the length of the training set, p i represents the actual load data in the training set, represents the load data predicted by the long-term and short-term time series network model of cross-dimensional attention during the training process.

[0101] In order to evaluate the final performance of the long-term and short-term time series network model of cross-dimensional attention after training and understand its generalization ability on other data, as a preferred implementation, a certain proportion of the data ranked in the front in the dataset is divided into the training set, and the data ranked in the back is divided into the test set. In this implementation, 80% of the data sorted by time is used as the training set, and 20% of the data sorted by time is used as the test set. After completing the training of the long-term and short-term time series network model of cross-dimensional attention using the training set, the trained model is then tested using the test set.

[0102] S400 predicts future load data.

[0103] Specifically, a long-term and short-term time series network model with trained cross-dimensional attention is used to predict the future load data of the low-voltage distribution transformer area. Please refer to Figure 8 , which is a comparison chart of the daily average load power data and the real data of the low-voltage distribution transformer area predicted for a future period of time.

[0104] Please refer to Figure 9 , which is a structural block diagram of an embodiment of the load prediction device for the low-voltage distribution transformer area of the present invention. The load prediction device for the low-voltage distribution transformer area of this embodiment is used to implement the load prediction method for the low-voltage distribution transformer area described in the above embodiment. Specifically, the load prediction device for the low-voltage distribution transformer area of this embodiment includes a data acquisition module 100, a data processing module 200, a model construction and training module 300, and a prediction module 400. Among them:

[0105] The data acquisition module 100 is used to collect the historical climate data and historical load data of the low-voltage distribution transformer area. The historical climate data includes historical daily average temperature, daily average wind speed, daily average precipitation, and daily average humidity. The historical load data includes historical daily average load power data.

[0106] The data processing module 200 is used to process the historical climate data and historical load data to obtain a data set. The data processing module 200 is also used to preprocess the historical climate data; the successive variational mode decomposition algorithm optimized by the Harris hawk optimization algorithm is used to decompose the historical daily average load power data to obtain several continuous periodic components; the decomposed components and the preprocessed historical climate data are arranged in chronological order and integrated to obtain a multi-dimensional data set; a certain proportion of the data in the front of the data set is divided into a training set, and the data in the back is divided into a test set.

[0107] The model construction and training module 300 is used to construct a long-term and short-term time series network model with cross-dimensional attention and train the long-term and short-term time series network model with cross-dimensional attention using the dataset. The long-term and short-term time series network model with cross-dimensional attention includes a convolutional layer, a long short-term memory network, a cross-attention module, an autoregressive unit, and a fully connected layer. Among them: The convolutional layer is used to extract local features of the input data, and the convolutional layer is a one-dimensional convolutional layer. In order to ensure that the extracted information is not lost, in this embodiment, the number of convolutional layer kernels is selected to be 1. The long short-term memory network is used to model the local features extracted by the convolutional layer to achieve the superposition of long-term and short-term memory of feature information. The long short-term memory network is composed of multiple long short-term memory units, and each unit includes an input gate, a forget gate, and an output gate. Multiple long short-term memory units achieve the superposition of long-term and short-term memory of feature information. The cross-attention module is used to take the output of the last time step of the long short-term memory network to form a matrix, calculate the weight scores between time steps through matrix operations, and perform weighted fusion on the features according to the weight scores. The autoregressive unit is used to receive the output of the cross-attention module, capture the linear correlation relationship in multi-dimensional data, and correct the predicted results. The fully connected layer is used to map the features output by the cross-attention module to the predicted results and add them to the output of the autoregressive unit to obtain the final predicted value.

[0108] The prediction module 400 is used to predict the future load data of the low-voltage distribution substation area using the trained long-term and short-term time series network model with cross-dimensional attention.

[0109] By integrating historical climate data and historical load data, the present invention constructs a multi-dimensional dataset, effectively capturing the correlation between climate fluctuations and electricity consumption behavior. Compared with the traditional method that only relies on historical loads, this solution can accurately reflect the impact of extreme weather and seasonal changes on electricity demand, is applicable to the prediction of volatile loads in scenarios with high penetration of new energy, and the prediction results are more in line with the actual power supply demand; the successive variational mode decomposition algorithm optimized by the Harris hawk algorithm is used to decompose historical load data, overcoming the defect of the traditional successive variational mode decomposition that relies on manual experience for parameter tuning. The decomposed IMF components have both low noise interference and high physical significance, providing clear multi-scale feature inputs for the subsequent model and reducing the risk of overfitting; the designed cross-dimensional attention mechanism solves the problem that traditional models are insufficient in capturing local mutations or long-term trends by dynamically weighted fusion of long-term and short-term time series features output by the long short-term memory network. Combined with the compensation and correction of the linear relationship by the autoregressive unit, the model can simultaneously depict the non-linear dynamics and periodic laws of load data, and can reduce the prediction error in complex fluctuation scenarios.

[0110] The above content only expresses the preferred embodiments of the present invention. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all fall within the protection scope of the present invention. Therefore, the protection scope of the present invention patent shall be subject to the appended claims.

Claims

1. A load forecasting method for a low-voltage distribution substation area, characterized in that, Including the following steps: Collect historical climate data and historical load data of the low-voltage distribution substation area. The historical climate data includes historical daily average temperature, daily average wind speed, daily average precipitation, and daily average humidity. The historical load data includes historical daily average load power data; Process the historical climate data and historical load data to obtain a data set; Construct a long-term and short-term time series network model with cross-dimensional attention and use the data set to train the long-term and short-term time series network model with cross-dimensional attention; Use the trained long-term and short-term time series network model with cross-dimensional attention to predict the future load data of the low-voltage distribution substation area.

2. The load forecasting method for a low-voltage distribution substation area according to claim 1, characterized in that, In the step of processing the historical climate data and historical load data to obtain a data set, the following sub-steps are included: Preprocess the historical climate data; Use the successive variational mode decomposition algorithm optimized by the Harris hawk optimization algorithm to decompose the historical daily average load power data to obtain several continuous periodic components; Arrange the decomposed components and the preprocessed historical climate data in chronological order and integrate them to obtain a multi-dimensional data set.

3. The load forecasting method for low-voltage distribution substations according to claim 2, wherein, In the step of preprocessing the historical climate data, the following sub-steps are included: Use the method of finding the mean of neighboring numbers to fill in the missing data in the historical climate data to obtain the filled data; Use the maximum-minimum normalization method to normalize the filled data to obtain the preprocessed historical climate data.

4. The load forecasting method for low-voltage distribution substation area according to claim 2, wherein, In the step of using the successive variational mode decomposition algorithm optimized by the Harris hawk optimization algorithm to decompose the historical daily average load power data to obtain several continuous periodic components, the following sub-steps are included: Set the objective function of the Harris hawk optimization algorithm, and the objective function is the minimum envelope entropy; Set the population size, maximum number of iterations of the Harris hawk optimization algorithm, and the range of the parameters to be optimized. The parameter to be optimized is the maximum penalty factor of the successive variational mode decomposition, and optimize the parameters; Use the optimal value of the maximum penalty factor obtained by the Harris hawk optimization as the decomposition parameter of the successive variational mode decomposition, and decompose the data sequence composed of the historical daily average load power data to obtain several continuous periodic components.

5. The load forecasting method for a low-voltage power distribution substation area according to claim 4, wherein The method of using the Harris hawk optimization algorithm to optimize the maximum penalty factor specifically includes: Initialize the maximum penalty factor and substitute the initial value of the maximum penalty factor into the successive variational mode decomposition algorithm to decompose the historical daily average load power data; Calculate the objective function value according to the number of decomposition layers and components; Judge whether the current objective function value is the minimum value in the iterative process. If it is the minimum value, use the corresponding maximum penalty factor as the current optimization result. Otherwise, update the prey escape energy and escape probability; Adjust the search strategy according to the absolute value of the prey escape energy, and update the maximum penalty factor according to the adjusted search strategy; Determine whether the maximum number of iterations is reached. If the maximum number of iterations is not reached, substitute the latest output maximum penalty factor into the successive variational mode decomposition algorithm to decompose the historical daily average load power data, and return to the step of calculating the objective function value according to the number of decomposed layers and components. Otherwise, terminate the iteration, and take the maximum penalty factor corresponding to the minimum objective function value in the iteration as the final optimization result.

6. The load forecasting method for a low-voltage distribution substation area according to claim 5, wherein, The objective function is expressed as: In formula (1), F obj represents the objective function, K represents the decomposition level, hilbert represents the Hilbert transform, imf(k) represents the k-th component after the decomposition of the original signal, abs represents taking the absolute value, sum represents summation,.* represents the dot product operation, and amp(k) represents the amplitude of the k-th imf component.

7. The load forecasting method for a low-voltage distribution substation area according to claim 1, wherein, The specific structure of the long-term and short-term time series network model with cross-dimensional attention includes: A convolutional layer for extracting local features of the input data, and the convolutional layer is a one-dimensional convolutional layer; A long short-term memory network for modeling the local features extracted by the convolutional layer to achieve the superposition of long-term and short-term memory of feature information; A cross-attention module for taking the output of the last time step of the long short-term memory network to form a matrix, calculating the weight scores between time steps through matrix operations, and performing weighted fusion on the features according to the weight scores; An autoregressive unit for receiving the output of the cross-attention module, capturing the linear correlation relationship in multi-dimensional data, and correcting the predicted result; A fully connected layer for mapping the features output by the cross-attention module to the predicted result and adding it to the output of the autoregressive unit to obtain the final predicted value.

8. The load forecasting method for a low-voltage distribution substation area according to claim 7, wherein, The calculation process of the cross-attention module includes: Taking the output of the last time step of the long short-term memory network to form a first matrix; Taking the outputs of all time steps of the long short-term memory network to form a second matrix and a third matrix respectively, and the second matrix and the third matrix contain the same information; Transposing the third matrix to obtain a transposed matrix, and calculating the cross-weight score according to the first matrix and the transposed matrix; Normalizing the cross-weight score to obtain a weight matrix; Multiplying the weight matrix by the second matrix to calculate the output matrix of the last time step after cross-attention processing.

9. The load forecasting method for a low-voltage distribution substation area according to claim 1, characterized in that, When training the long-term and short-term time series network model with cross-dimensional attention using the dataset, the mean squared error is used as the training loss function, and its expression is: In formula (II), MSE train represents the loss function for training the long-term and short-term time series network model of cross-dimensional attention. n represents the length of the training data, and p i represents the actual load data in the training data, is the load data predicted by the long-term and short-term time series network model of cross-dimensional attention during the training process.

10. A load forecasting device for a low-voltage distribution substation, characterized in that, Including: A data acquisition module for collecting historical climate data and historical load data of the low-voltage distribution substation area. The historical climate data includes historical daily average temperature, daily average wind speed, daily average precipitation, and daily average humidity, and the historical load data includes historical daily average load power data; A data processing module for processing the historical climate data and historical load data to obtain a dataset; A model construction and training module for constructing a long-term and short-term time series network model with cross-dimensional attention and training the long-term and short-term time series network model with cross-dimensional attention using the dataset; A prediction module for predicting the future load data of the low-voltage distribution substation area using the trained long-term and short-term time series network model with cross-dimensional attention.

Images