Multi-region power load prediction method, device, equipment, medium and program product

Abstract

The invention discloses a multi-region power load prediction method, device and equipment, a medium and a program product, and the method comprises the steps: collecting historical load time series data of a plurality of regions in a power system, carrying out the preprocessing of the historical load time series data, obtaining candidate load time series data, and carrying out the prediction of the candidate load time series data; performing depth feature extraction and multi-dimensional feature analysis on the candidate load time sequence data, performing feature fusion on the depth feature and the multi-dimensional feature matrix to obtain a target load feature, performing clustering analysis on a plurality of regions based on the target load feature, determining a region cluster corresponding to each region, and performing clustering analysis on the region clusters; constructing a load prediction model corresponding to each region cluster according to the target load characteristics, and performing power load prediction on the power system based on the load prediction model corresponding to each region cluster; according to the method, the load characteristics of each region are effectively mined, so that the method can better adapt to load changes of different regions, and accurate power load prediction is provided for different regions in a power system.

Description

Technical Field

[0001] The present invention relates to the field of power load forecasting, and particularly to a multi-region power load forecasting method, device, equipment, medium and program product. Background Art

[0002] In recent years, the wide access of distributed energy has provided users with clean green electricity. However, with the continuous increase of its penetration rate, the traditional distribution network has been unable to cope with the integration of distributed energy, frequently experiencing problems such as power flow over-limit, reverse power flow, and protection maloperation, which seriously affect the stability of the distribution network. At the same time, profound changes have also occurred on the demand side of electricity. The load is no longer just passively accepting the one-way regulation of the power grid traditionally, but is gradually developing to be able to interact with the power grid bidirectionally, and the control methods are more diversified. In addition, the development and coordinated operation technology of series-parallel combined flexible energy storage devices on the distribution network side are also continuously evolving, posing higher requirements for the flexibility of the power grid. Therefore, the traditional distribution network is difficult to adapt to and meet the complexity and diversity requirements of future power grids in power control. At the same time, in a distributed energy storage system, the power and energy demands as well as the user electricity consumption behaviors in different regions are not the same, which brings challenges to the resource scheduling optimization of the distributed energy storage system. Therefore, to achieve the safe and stable power supply of distributed energy and the reasonable scheduling of resources, it is necessary to reasonably predict the electricity load of each region and then determine the energy demand. Currently, the existing distributed load forecasting methods often use a single algorithm to forecast the power load for different regions, thus ignoring the uniqueness and differences of load data in different regions of the power system, resulting in a low accuracy of power load forecasting in the power system and seriously affecting the power resource scheduling efficiency. Summary of the Invention

[0003] The main objective of the present invention is to provide a multi-region power load forecasting method, device, equipment, medium and program product, aiming to solve the technical problem that the existing technology ignores the uniqueness and differences of load data in different regions of the power system, resulting in a low accuracy of power load forecasting in the power system and seriously affecting the power resource scheduling efficiency.

[0004] To achieve the above objective, the present invention provides a multi-region power load forecasting method, and the method includes the following steps: Collect historical load time series data of multiple regions in the power system, and preprocess the historical load time series data to obtain candidate load time series data; Perform deep feature extraction on the candidate load time series data to obtain a deep feature matrix for each region; Perform multi-dimensional feature analysis on the candidate load time series data to obtain a multi-dimensional feature matrix for each region; Fuse the depth feature matrix with the multi-dimensional feature matrix to obtain the target load feature, and perform clustering analysis on the multiple regions based on the target load feature to determine the region clusters corresponding to each region; Construct a load prediction model corresponding to each region cluster according to the target load feature; Perform power load forecasting on the power system based on the load prediction models corresponding to each region cluster.

[0005] Optionally, the extracting the depth feature matrix of each region from the candidate load time series data includes: Perform bidirectional long short-term dependence analysis on the candidate load time series data to obtain the forward hidden state and the backward hidden state at each time step, and splice the forward hidden state and the backward hidden state to obtain the bidirectional dependence feature at each time step; Based on the bidirectional hidden state, perform a linear transformation to construct a query matrix, a key matrix, and a value matrix; Determine the attention feature at each time step according to the query matrix, the key matrix, and the value matrix: ; wherein, represents the attention feature at time step , represents the query matrix, represents the matrix, represents the value matrix, represents the transpose, represents the dimension of the key matrix, represents the activation function; Splice the bidirectional dependence feature at each time step with the attention feature to obtain the initial depth feature: ; wherein, represents the initial depth feature at time step , represents the feature splicing operation, represents time step 's bidirectional dependence feature; Construct a depth feature matrix based on the initial depth feature at each time step: ; wherein, represents the depth feature matrix, represents the time dimension, represents the feature dimension.

[0006] Optionally, the performing multi-dimensional feature analysis on the candidate load time series data to obtain the multi-dimensional feature matrix of each region: Calculate the mean, standard deviation, and coefficient of variation of the candidate load time series data: ; Where, represents the mean of the candidate load time series data, represents the standard deviation of the candidate load time series data, represents the coefficient of variation; Conduct global load fluctuation analysis based on the mean, the standard deviation, and the coefficient of variation to obtain basic statistical features; Conduct time fluctuation analysis based on the candidate load time series data to obtain time pattern features, where the time pattern features include peak-valley difference, period mean ratio, and daily load rate: ; ; ; Where, represents the peak-valley difference of the load within a day, represents the maximum value of the load during a day, represents the minimum value of the load within a day, represents the ratio of the load means in the morning and evening periods, represents the average load value in the morning period, represents the average load value in the evening period, represents the daily load rate, represents the average value of the load within a day; Conduct short-term fluctuation analysis based on the candidate load time series data to obtain volatility features, where the volatility features include lag autocorrelation coefficient: ; Where, represents the lag autocorrelation coefficient, represents the th load value at a time step, and the load value is a data point in the candidate load time series data, represents the total number of time steps, represents the lag step, represents the data mean of the candidate load time series data; Combine the basic statistical features, the time pattern features, and the volatility features into a multi-dimensional feature matrix.

[0007] Optionally, the method of fusing the deep feature matrix with the multi-dimensional feature matrix to obtain target load features and performing clustering analysis on the multiple regions based on the target load features to determine the region clusters corresponding to the regions includes: Fuse the depth feature matrix with the multi-dimensional feature matrix to obtain the target load feature: ; Among them, represents the target load feature, represents the multi-dimensional feature matrix; Determine the feature similarity between regions based on the target load feature: ; Among them, represents region and region The feature similarity between them, represents the feature dimension of the target load feature, represents region The th feature, represents region The th feature; Perform clustering analysis on the multiple regions according to the feature similarity to determine the region clusters corresponding to each region.

[0008] Optionally, the constructing of the load prediction model corresponding to each region cluster according to the target load feature includes: Construct an initial prediction model corresponding to each region cluster, and the objective function of the initial prediction model is as follows: ; Among them, represents the objective function, represents the loss function, represents the number of samples, represents the true value of the th sample, represents the th sample after passing through decision trees represents the decision tree ensemble, represents the regularization term part for preventing overfitting, represents the th decision tree complexity function; Normalize the target load feature: ; Among them, represents the th feature of the th sample after normalization, represents the th sample before normalization, the a feature, and respectively represent the maximum and minimum values of the th feature; Evaluate the importance of each feature in the normalized target load feature, and screen the normalized target load feature based on the importance evaluation result to obtain the target training feature. The importance evaluation refers to the following formula: ; where represents the gain value of the feature in the decision tree , represents the importance evaluation result; Train the initial prediction model based on the target training feature to obtain a load prediction model.

[0009] Optionally, the preprocessing of the historical load time series data to obtain candidate load time series data includes: Calculate the standard score of each data point in the historical load time series data: ; where represents the standard score of the data point , represents the mean of the historical load time series data, represents the standard deviation of the historical load time series data; Mark the data points with standard scores higher than the anomaly threshold in the historical load time series data as anomaly data points, and perform data cleaning on the anomaly data points; Fill in the missing data in the historical load time series data after data cleaning to obtain the initial load time series data. The missing data is filled in according to the following formula: ; where represents the length of the data missing segment, represents the missing position of the current missing value in the missing segment, represents the known data point after the missing segment, represents the known data point before the missing segment; Perform normalization processing on the initial load time series data to obtain candidate load time series data. The normalization processing refers to the following formula: ; where represents the normalized data point, represents the minimum data point in the initial load time series data, Represents the maximum value data point.

[0010] In addition, to achieve the above object, the present invention also provides a multi - region power load forecasting device, which includes: A data processing module, configured to collect historical load time - series data of multiple regions in the power system, and pre - process the historical load time - series data to obtain candidate load time - series data; A deep feature analysis module, configured to perform deep feature extraction on the candidate load time - series data to obtain a deep feature matrix for each region; A multi - dimensional feature analysis module, configured to perform multi - dimensional feature analysis on the candidate load time - series data to obtain a multi - dimensional feature matrix for each region; A feature clustering analysis module, configured to fuse the deep feature matrix and the multi - dimensional feature matrix to obtain target load features, and perform clustering analysis on the multiple regions based on the target load features to determine the region clusters corresponding to each region; A prediction model construction module, configured to construct a load prediction model corresponding to each region cluster according to the target load features; A power load forecasting module, configured to perform power load forecasting on the power system based on the load prediction models corresponding to each region cluster.

[0011] In addition, to achieve the above object, the present application also provides a multi - region power load forecasting device, which includes: a memory, a processor, and a computer program stored on the memory and executable on the processor, and the computer program is configured to implement the steps of the multi - region power load forecasting method as described above.

[0012] In addition, to achieve the above object, the present application also provides a computer - readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the multi - region power load forecasting method as described above.

[0013] In addition, to achieve the above object, the present application also provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, it implements the steps of the multi - region power load forecasting method as described above.

[0014] The present invention collects historical load time series data of multiple regions in a power system, preprocesses the historical load time series data to obtain candidate load time series data, performs deep feature extraction on the candidate load time series data to obtain a deep feature matrix for each region, performs multi-dimensional feature analysis on the candidate load time series data to obtain a multi-dimensional feature matrix for each region, fuses the deep feature matrix and the multi-dimensional feature matrix to obtain a target load feature, and performs clustering analysis on the multiple regions based on the target load feature to determine the region clusters corresponding to each region. A load prediction model corresponding to each region cluster is constructed based on the target load feature, and power load prediction is performed on the power system based on the load prediction model corresponding to each region cluster. Since the present invention performs deep feature analysis and multi-dimensional feature analysis on the load time series data of each region respectively, it realizes in-depth excavation of the load characteristics of each region, effectively avoids the problem that potential features and differences of different regions are ignored, improves the prediction accuracy, and performs clustering analysis based on the target load feature obtained by fusing deep features and multi-dimensional features to cluster regions of the same type, provides a reasonable load prediction model for regions with the same features, effectively improves the power load prediction efficiency, more accurately captures and reflects the load change trends of each region, greatly improves the accuracy of power system load prediction, and provides accurate prediction result support for power resource scheduling. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, for those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0016] Figure 1 It is a schematic structural diagram of a multi-region power load prediction device for the hardware operating environment involved in the embodiment solution of the present invention; Figure 2 It is a schematic flowchart of the first embodiment of the multi-region power load prediction method of the present invention; Figure 3 It is a schematic flowchart of the deep feature extraction process in an embodiment of the multi-region power load prediction method of the present invention; Figure 4 It is a schematic flowchart of the multi-dimensional feature analysis process in an embodiment of the multi-region power load prediction method of the present invention; Figure 5 It is a schematic flowchart of the load prediction model construction process in an embodiment of the multi-region power load prediction method of the present invention; Figure 6 It is a schematic block diagram of the first embodiment of the multi-region power load prediction device of the present invention.

[0017] The realization, functional features and advantages of the present invention will be further described with reference to the embodiments and the accompanying drawings. Specific embodiments

[0018] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0019] Refer to Figure 1 , Figure 1 It is a schematic structural diagram of a multi-region power load forecasting device for the hardware operating environment involved in the embodiment solution of the present invention.

[0020] As Figure 1 shown, the multi-region power load forecasting device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display) and an input unit such as a keyboard (Keyboard). Optionally, the user interface 1003 may further include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a wireless-fidelity (WI-FI) interface). The memory 1005 may be a high-speed random access memory (Random Access Memory, RAM), or a stable non-volatile memory (Non-Volatile Memory, NVM), such as a disk memory. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0021] Those skilled in the art can understand that Figure 1 the structure shown in

[0022] As Figure 1 shown, the memory 1005, as a computer-readable storage medium, may include an operating system, a network communication module, a user interface module, and a multi-region power load forecasting program.

[0023] In Figure 1In the multi - region power load forecasting device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the multi - region power load forecasting device of the present invention can be arranged in the multi - region power load forecasting device. The multi - region power load forecasting device calls the multi - region power load forecasting program stored in the memory 1005 through the processor 1001 and executes the multi - region power load forecasting method provided by the embodiments of the present invention.

[0024] Embodiments of the present invention provide a multi - region power load forecasting method. Refer to Figure 2 , Figure 2 which is a schematic flowchart of the first embodiment of the multi - region power load forecasting method of the present invention.

[0025] In this embodiment, the multi - region power load forecasting method includes the following steps: Step S10: Collect historical load time - series data of multiple regions in the power system, and pre - process the historical load time - series data to obtain candidate load time - series data.

[0026] It should be noted that this embodiment can be applied to the power load forecasting of a distributed power supply system including multiple power - consuming regions. By performing in - depth feature analysis and multi - dimensional feature analysis on the load time - series data of each region respectively, this embodiment can effectively mine the load characteristics of each region, effectively avoid the problem that potential features and differences of different regions are ignored, improve the prediction accuracy. By fusing the deep features and multi - dimensional features, and performing clustering analysis based on the fused target load features, it can cluster the regions of the same type, provide a reasonable load forecasting model for regions with the same features, effectively improve the power load forecasting efficiency, more accurately capture and reflect the load change trends of each region, greatly improve the accuracy of power system load forecasting, and provide accurate prediction result support for power resource scheduling.

[0027] It should be understood that the execution subject of this embodiment can be a computing service device with data processing, network communication, and program running functions, such as a tablet computer, a personal computer, a mobile phone, etc., or a terminal electronic device capable of implementing the above functions. Hereinafter, a multi - region power load forecasting device (referred to as a forecasting device for short) will be used as an example to illustrate this embodiment and the following embodiments.

[0028] It should be noted that the power system can be a distributed multi - area power supply system. The multi - area power supply system emphasizes the characteristic that the power supply system covers multiple geographical or functional areas, and each area has independent electricity consumption demands and power supply management. Such a system can include a central power station or distributed power generation units, and transmits electricity to each electricity consumption area through the power grid.

[0029] It should be explained that in this embodiment, short - term load forecasting can be performed on the power system. Short - term load forecasting of the power system refers to the prediction of the change in the load demand of the power system in the next few hours, one day, one week or a shorter period. It is an important link in the operation and dispatching of the power system, and is of great significance for ensuring the safe and stable operation of the power grid, optimizing the allocation of power resources, and improving the quality of power supply services. Short - term load forecasting of the power system needs to consider various factors, including historical load data, weather changes, holidays, economic activities, social events, etc. Through comprehensive analysis of these factors, a mathematical model can be established or machine learning algorithms can be used to predict future load demands. The results of short - term load forecasting can provide an important basis for power system dispatching, helping dispatchers reasonably arrange power generation plans and adjust the operation mode of the power grid to cope with load changes.

[0030] It can be understood that in the first step of load forecasting in this embodiment, the prediction device can collect historical load time - series data of each region. These data are the basis for subsequent analysis, so the integrity and accuracy of the data must be ensured. After collection, pre - processing is performed on the historical load time - series data, such as data cleaning and filling of missing data for the historical load time - series data.

[0031] In some embodiments, when the prediction device collects data, non - uniformly sampled data is resampled to a fixed interval (such as 15 min): ; where represents the set of time steps within the time window, used to describe the data distribution within a specific time window, represents the current time point, represents the half - width of the time window.

[0032] Furthermore, in order to improve data quality and thus improve prediction efficiency and accuracy, step S10 above may include: Step S101: Calculate the standard score of each data point in the historical load time - series data; Step S102: Mark the data points with a standard score higher than the anomaly threshold in the historical load time - series data as anomaly data points, and perform data cleaning on the anomaly data points; Step S103: Fill in the missing data in the historical load time - series data after data cleaning to obtain the initial load time - series data; Step S104: Normalize the initial load time series data to obtain candidate load time series data.

[0033] It should be noted that after the historical load time series data is collected, the prediction device can perform data cleaning work to process the outliers in the data, such as sudden increases, sudden decreases, or values that are significantly unreasonable, to ensure the reliability of the data. For missing data, the prediction device can use appropriate methods for supplementation, such as interpolation or speculation based on historical data, while ensuring that the time points of all data are aligned for subsequent analysis. Finally, to eliminate the dimensional differences between data, the prediction device can perform data standardization processing to make the data comparable and consistent.

[0034] It should be noted that the Z-score method can be used to clean the outliers in the data. The data points with are marked as outliers and cleaned. The standard score calculation refers to the following formula: where represents the standard score of the data point , represents the mean of the historical load time series data, represents the standard deviation of the historical load time series data, is the outlier threshold.

[0035] When there are missing values, linear interpolation can be used for filling. The missing data is filled according to the following formula: ; where represents the length of the data missing segment, represents the missing position of the current missing value in the missing segment, represents the known data point after the missing segment, represents the known data point before the missing segment.

[0036] When the prediction device performs data scaling, the min-max normalization method can be used to scale the values to [0, 1]. The normalization process refers to the following formula: ; where represents the normalized data point, represents the minimum data point in the initial load time series data, represents the maximum data point.

[0037] After the above processing, the normalized, aligned, and complete time series data can be obtained. represents the The preprocessed load value at each time point: ; Step S20: performing deep feature extraction on the candidate load time series data to obtain a deep feature matrix of each region.

[0038] In some embodiments, the prediction device can construct a deep feature extraction network, perform deep feature extraction on the candidate load time series data based on the deep feature extraction network, and obtain the deep feature matrix of each region. The deep feature extraction network can be constructed based on a bidirectional long short-term memory network (Bidirectional Long Short-Term Memory, BiLSTM) and an attention mechanism module (Attention). The deep feature extraction network can automatically learn the high-order time series dependencies and nonlinear patterns of the load data, making up for the limitations of traditional feature engineering. We use a multi-level recurrent neural network architecture, combined with a bidirectional long short-term memory network and an attention mechanism, to fully explore the deep features of the load data.

[0039] Furthermore, in order to accurately extract the deep features in the load data and explore the potential load characteristics of each area, refer to Figure 3 , Figure 3 This is a schematic diagram of a deep feature extraction process in an embodiment. The above step S20 may include: Step S201: performing a bidirectional long-term and short-term dependency analysis on the candidate load time series data, obtaining a forward hidden state and a reverse hidden state of each time step, and concatenating the forward hidden state with the reverse hidden state to obtain a bidirectional dependency feature of each time step; Step S202: Performing linear transformation based on the bidirectional hidden state to construct a query matrix, a key matrix and a value matrix; Step S203: determining the attention feature of each time step according to the query matrix, the key matrix and the value matrix; Step S204: concatenate the bidirectional dependency features and attention features of each time step to obtain initial depth features; Step S205: construct a depth feature matrix based on the initial depth features of each time step.

[0040] In some embodiments, the prediction device may receive the candidate load time series data obtained after preprocessing through the input layer of the deep feature extraction network, and may fuse the category features (such as region, season, etc.) as auxiliary input, and learn the long-term and short-term dependencies through the bidirectional long short-term memory network, referring to the following formula: ; The prediction device assigns weights to the load data at each time step through the self-attention mechanism module to enhance the weights at key time points, referring to the following formula: ; Wherein, represents the attention feature at time step , represents the query matrix, represents the matrix, represents the value matrix, represents the transpose, represents the dimension of the key matrix, represents the activation function.

[0041] The prediction device concatenates the bidirectional dependence features and the attention features at each time step: ; Wherein, represents the initial depth feature at time step , represents the feature concatenation operation, represents the bidirectional dependence feature at time step .

[0042] The prediction device constructs a depth feature matrix based on the initial depth features at each time step: ; Wherein, represents the depth feature matrix, represents the time dimension, represents the feature dimension.

[0043] In some embodiments, the prediction device can train the depth feature extraction network using a loss function that combines MAE and smooth loss (Huber Loss) to balance the robustness and accuracy of the depth feature extraction network: ; Wherein, represents the loss value, represents the weight coefficient, represents the mean error function, represents the smooth loss function, represents the predicted value, represents the true value.

[0044] The dynamic feature selection of the depth feature extraction network can dynamically adjust the input features based on gradient-weighted analysis (Grad-CAM) or feature importance scores.

[0045] Regularization of the deep feature extraction network can use Dropout and Layer Normalization to prevent overfitting.

[0046] Step S30: Perform multi-dimensional feature analysis on the candidate load time series data to obtain the multi-dimensional feature matrix of each region.

[0047] It should be noted that feature engineering is a key step in load forecasting, which directly affects the accuracy and generalization ability of the forecasting model. In this embodiment, multi-dimensional feature extraction and deep feature extraction are respectively performed on the candidate load time series data.

[0048] In some embodiments, the prediction device can perform multi-dimensional feature extraction from three aspects: basic statistical features, time pattern features, and volatility features. Basic statistical features include the mean, standard deviation, coefficient of variation, etc. of the load, which can reflect the overall level and fluctuation of the load. Time pattern features focus on the change rules of the load within a day, within a week, and seasonally, such as peak and valley periods, differences between weekdays and weekends, and seasonal load changes. Volatility features describe the volatility and periodicity of the load through volatility, autocorrelation features, and periodic intensity indicators at multiple time scales.

[0049] Further, in order to accurately extract the multi-dimensional features in the load data of each region, refer to Figure 4 , Figure 4 As shown in the schematic diagram of the multi-dimensional feature analysis process in an embodiment, the above step S30 may include: Step S301: Calculate the mean, standard deviation, and coefficient of variation of the candidate load time series data; Step S302: Perform global load fluctuation analysis based on the mean, the standard deviation, and the coefficient of variation to obtain basic statistical features; Step S303: Perform time fluctuation analysis based on the candidate load time series data to obtain time pattern features; Step S304: Perform short-term fluctuation analysis based on the candidate load time series data to obtain volatility features; Step S305: Combine the basic statistical features, the time pattern features, and the volatility features into a multi-dimensional feature matrix.

[0050] It should be noted that multi-dimensional feature analysis is a key step in load forecasting, which directly affects the accuracy and generalization ability of the forecasting model. In addition to using the deep feature extraction network for feature extraction, the prediction device can also perform feature extraction from three aspects: basic statistical features, time pattern features, and volatility features.

[0051] It should be noted that the basic statistical features may include the mean, standard deviation, coefficient of variation, etc. of the load, which can reflect the overall level and fluctuation of the load, and are calculated with reference to the following formula: ; Among them, represents the mean of the candidate load time series data, represents the standard deviation of the candidate load time series data, represents the coefficient of variation.

[0052] The time pattern features focus on the variation rules of the load within a day, within a week, and seasonally, such as peak and valley periods, differences between weekdays and weekends, seasonal load changes, etc. In some embodiments, the time pattern features include the peak-valley difference, period mean ratio, and daily load rate, and are calculated with reference to the following formula: ; ; ; Among them, represents the peak-valley difference of the load within a day, represents the maximum value of the load within a day, represents the minimum value of the load within a day, represents the mean ratio of the load in the morning and evening periods, represents the average load value in the morning period (e.g., 7:00 - 9:00), represents the average load value in the evening period (e.g., 23:00 - 5:00), represents the daily load rate, represents the average value of the load within a day.

[0053] The volatility features describe the volatility and periodicity of the load through the volatility rate, autocorrelation features, and periodic intensity index on multiple time scales. The autocorrelation features can be obtained based on the lag autocorrelation coefficient, and the lag autocorrelation coefficient is calculated with reference to the following formula: ; Among them, represents the lag autocorrelation coefficient, represents the th load value at the time step, and the load value is a data point in the candidate load time series data, represents the total number of time steps, represents the lag step, represents the data mean of the candidate load time series data.

[0054] Finally, the extracted basic statistical features, deep feature matrix, and original load data are concatenated to form the input data of the model. For example, all features can be combined into a feature matrix F, where each row corresponds to a time point or time window: ; Step S40: Perform feature fusion on the deep feature matrix and the multi-dimensional feature matrix to obtain target load features, and perform cluster analysis on the multiple regions based on the target load features to determine the region clusters corresponding to each region.

[0055] It can be understood that, in order to classify regions with similar load characteristics, this embodiment can perform cluster analysis on multiple regions in the power system. For example, the hierarchical clustering method is adopted, with the Ward minimum variance method as the linkage criterion, and the similarity is measured based on the Euclidean distance.

[0056] Furthermore, in order to accurately aggregate regions of the same type and thus improve the prediction efficiency, the above step S40 may include: Step S401: Perform feature fusion on the deep feature matrix and the multi-dimensional feature matrix to obtain target load features; Step S402: Determine the feature similarity between each region based on the target load features; Step S403: Perform cluster analysis on the multiple regions according to the feature similarity to determine the region clusters corresponding to each region.

[0057] It should be noted that in this embodiment, by fusing the deep feature matrix and the multi-dimensional feature matrix, the feature similarity between each region is calculated based on the target load features, so as to effectively identify regions with similar load characteristics, providing a basis for the construction of subsequent differential prediction models. The feature fusion refers to the following formula: ; where, represents the target load feature, represents the multi-dimensional feature matrix.

[0058] The calculation of the feature similarity between each region refers to the following formula: ; where, represents the region and the region the feature similarity between them, represents the feature dimension of the target load feature, represents the region the th feature, represents the region the feature

[0059] In some embodiments, during the clustering process, the prediction device can dynamically adjust the number of clusters to ensure the rationality and accuracy of clustering. The prediction device can also optimize the feature weights to improve the stability and reliability of the clustering results. Finally, the quality of the clustering results is evaluated through stability verification to ensure the effectiveness and repeatability of clustering.

[0060] In some embodiments, K data points are randomly selected as the initial clustering centers, denoted as For each sample , calculate its distance from each clustering center and assign it to the cluster corresponding to the nearest clustering center. For each cluster , calculate the mean of all samples in the cluster as the new clustering center: ; where is the number of samples in cluster .

[0061] Step S50: Construct a load prediction model corresponding to each regional cluster according to the target load characteristics.

[0062] In some embodiments, the prediction device can construct a differentiated prediction model according to the load characteristics of each category. The model selection strategy is to select an appropriate prediction algorithm based on category features, such as XGBoost, LightGBM, CatBoost, etc. These algorithms have advantages in processing complex data and capturing non-linear relationships. In some embodiments, the prediction device can specifically optimize the model parameters according to the characteristics of each category area to improve the prediction accuracy. In terms of feature input, the category features are used as a supplement to the model input, and feature importance analysis is used to determine which features have the greatest impact on the prediction results. At the same time, the prediction device can adopt a dynamic feature selection mechanism to dynamically adjust the input features according to the performance of the model and the changes in the data to ensure the stability and adaptability of the prediction model.

[0063] Furthermore, in order to improve the load prediction accuracy and accurately capture and reflect the load change trends in each region, referring to Figure 5 , Figure 5 As shown in FIG. Step S501: Construct an initial prediction model corresponding to each regional cluster; Step S502: Normalize the target load characteristics; Step S503: Evaluate the importance of each feature in the normalized target load feature, and screen the normalized target load feature based on the importance evaluation result to obtain the target training feature; Step S504: Train the initial prediction model based on the target training feature to obtain a load prediction model.

[0064] It can be understood that, in order to more accurately predict the load changes in each region, this embodiment adopts the method of separately modeling and separately extracting features for each region. Since there are significant differences in the load data characteristics of different regions, it is difficult for a single prediction model to comprehensively and accurately reflect the load change laws of all regions. Therefore, in this embodiment, multiple load models are constructed according to the load data characteristics of each region. These models can fully consider the uniqueness of each region, so as to more accurately capture and reflect the load change trends of each region.

[0065] In specific implementation, the prediction device can perform feature data preparation, including historical power load data and feature data of the target load feature: Power load data: , where represents the load value at the i th time point; Feature data: , where is the feature vector at the i th time point. In the data preprocessing stage, it is necessary to perform normalization processing on the feature data to eliminate the dimension difference: ; Among them, represents the th feature of the th normalized sample, represents the th feature of the th sample before normalization, and respectively represent the maximum and minimum values of the th feature.

[0066] The prediction device can construct an initial prediction model based on XGBoost, and use the feature importance evaluation function of XGBoost to screen the features that have the most influence on power load prediction. The importance evaluation refers to the following formula: ; Among them, represents the gain value of the feature in the decision tree , represents the importance evaluation result.

[0067] XGBoost is an ensemble learning algorithm based on gradient boosting, which improves the learning efficiency of the model by constructing multiple weak learners (usually decision trees) and combining them into a strong learner. Its goal is to minimize the loss function, and the objective function of the initial prediction model is as follows: ; where, represents the objective function, represents the loss function, represents the number of samples, represents the -th true value of the sample, represents the -th predicted value of the sample after passing through decision trees, represents the set of decision trees, represents the regularization term part used to prevent overfitting, represents the -th complexity function of the decision tree.

[0068] is the predicted value of the i -th sample, which is obtained by summing the predicted values of T trees: ; In each round of iteration, XGBoost optimizes the loss function by calculating the first-order derivative and the second-order derivative: ; ; For the squared error loss function: ; ; Step S60: Perform power load forecasting on the power system based on the load forecasting models corresponding to each regional cluster.

[0069] In some embodiments, the prediction device can evaluate the model performance using metrics such as root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE), and their calculation formulas are as follows: ; ; ; where, represents the actual value, represents the model predicted value. After calculating the prediction error of the model, the best combination of hyperparameters is selected through cross-validation to ensure the generalization ability of the model.

[0070] In some embodiments, the prediction device may use the XGBoost model with optimized parameters to predict the test set: ; wherein, represents the prediction result of the test set, represents the prediction result of the th decision tree, represents the feature input of the test set.

[0071] Subsequently, the prediction result is compared with the actual load data to analyze the accuracy of the model. At the same time, the prediction result can also be displayed through a visualization tool to help judge the performance of the model.

[0072] In this embodiment, historical load time series data of multiple regions in the power system is collected, and the historical load time series data is preprocessed to obtain candidate load time series data. Deep feature extraction is performed on the candidate load time series data to obtain the deep feature matrix of each region. Multidimensional feature analysis is performed on the candidate load time series data to obtain the multidimensional feature matrix of each region. The deep feature matrix and the multidimensional feature matrix are subjected to feature fusion to obtain the target load feature. Cluster analysis is performed on the multiple regions based on the target load feature to determine the region clusters corresponding to each region. A load prediction model corresponding to each region cluster is constructed based on the target load feature. Power load prediction of the power system is performed based on the load prediction model corresponding to each region cluster. Since in this embodiment, deep feature analysis and multidimensional feature analysis are respectively performed on the load time series data of each region, in-depth mining of the load characteristics of each region is realized, effectively avoiding the problem that potential features and differences of different regions are ignored, improving the prediction accuracy. By performing feature fusion on the deep feature and the multidimensional feature, and performing cluster analysis based on the fused target load feature, clustering of regions of the same type is realized, a reasonable load prediction model is provided for regions with the same features, effectively improving the power load prediction efficiency, more accurately capturing and reflecting the load change trends of each region, greatly improving the accuracy of power system load prediction, and providing accurate prediction result support for power resource scheduling.

[0073] In addition, an embodiment of the present invention also proposes a computer-readable storage medium, on which a multi-region power load prediction program is stored. When the multi-region power load prediction program is executed by a processor, the steps of the multi-region power load prediction method described above are implemented.

[0074] The computer-readable storage medium provided by the present application can be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or components, or any combination of the above. More specific examples of the computer-readable storage medium may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM) or flash memory, optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the above. In this embodiment, the computer-readable storage medium can be any tangible medium that contains or stores a program, and this program can be used by or in combination with an instruction execution system, device, or component. The program code contained on the computer-readable storage medium can be transmitted by any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination of the above.

[0075] The above computer-readable storage medium can be included in the multi-region power load forecasting device; it can also exist separately without being assembled into the multi-region power load forecasting device.

[0076] In addition, an embodiment of the present invention also proposes a computer program product, including a multi-region power load forecasting program. When the multi-region power load forecasting program is executed by a processor, it implements the steps of the multi-region power load forecasting method described above.

[0077] The specific implementation manner of the computer program product of the present invention is basically the same as that of the above embodiments of the multi-region power load forecasting method, and will not be elaborated here.

[0078] Refer to Figure 6 , Figure 6 , which is the structural block diagram of the first embodiment of the multi-region power load forecasting device of the present invention.

[0079] As Figure 6 shown, the multi-region power load forecasting device proposed by the embodiment of the present invention includes: A data processing module 10, configured to collect historical load time series data of multiple regions in the power system and preprocess the historical load time series data to obtain candidate load time series data; The deep feature analysis module 20 is configured to perform deep feature extraction on the candidate load time series data to obtain the deep feature matrices of each region; The multi-dimensional feature analysis module 30 is configured to perform multi-dimensional feature analysis on the candidate load time series data to obtain the multi-dimensional feature matrices of each region; The feature clustering analysis module 40 is configured to perform feature fusion on the deep feature matrix and the multi-dimensional feature matrix to obtain the target load features, and perform clustering analysis on the multiple regions based on the target load features to determine the region clusters corresponding to each region; The prediction model construction module 50 is configured to construct a load prediction model corresponding to each region cluster according to the target load features; The power load prediction module 60 is configured to perform power load prediction on the power system based on the load prediction models corresponding to each region cluster.

[0080] In this embodiment, by collecting the historical load time series data of multiple regions in the power system, preprocessing the historical load time series data to obtain candidate load time series data, performing deep feature extraction on the candidate load time series data to obtain the deep feature matrices of each region, performing multi-dimensional feature analysis on the candidate load time series data to obtain the multi-dimensional feature matrices of each region, performing feature fusion on the deep feature matrix and the multi-dimensional feature matrix to obtain the target load features, performing clustering analysis on the multiple regions based on the target load features to determine the region clusters corresponding to each region, constructing a load prediction model corresponding to each region cluster according to the target load features, and performing power load prediction on the power system based on the load prediction models corresponding to each region cluster; since this embodiment performs deep feature analysis and multi-dimensional feature analysis on the load time series data of each region respectively, thereby realizing in-depth mining of the load characteristics of each region, effectively avoiding the problem that the potential features and differences of different regions are ignored, improving the prediction accuracy, performing feature fusion on the deep features and multi-dimensional features, performing clustering analysis based on the fused target load features, realizing clustering of regions of the same type, providing a reasonable load prediction model for regions with the same features, effectively improving the power load prediction efficiency, more accurately capturing and reflecting the load change trends of each region, greatly improving the accuracy of power system load prediction, and providing accurate prediction result support for power resource scheduling.

[0081] The multi-region power load forecasting device provided by this application adopts the multi-region power load forecasting method in the above-mentioned embodiment, and can solve the technical problem of multi-region power load forecasting. Compared with the prior art, the beneficial effects of the multi-region power load forecasting device provided by this application are the same as those of the multi-region power load forecasting method provided by the above-mentioned embodiment, and other technical features in the multi-region power load forecasting device are the same as the features disclosed in the method of the above-mentioned embodiment, which will not be elaborated here.

[0082] It should be understood that the above is only for illustration and does not constitute any limitation to the technical solution of the present invention. In specific applications, those skilled in the art can set according to needs, and the present invention does not limit this.

[0083] It should be noted that the above-described work process is only illustrative and does not limit the protection scope of the present invention. In actual applications, those skilled in the art can select some or all of them according to actual needs to achieve the purpose of the solution of this embodiment, and there is no limitation here.

[0084] In addition, for the technical details not described in detail in this embodiment, reference can be made to the multi-region power load forecasting method provided in any embodiment of the present invention, which will not be elaborated here.

[0085] It should be noted that in this article, the terms "including", "comprising" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or further includes elements inherent to such process, method, article or system. Without more limitations, the element defined by the statement "including one..." does not exclude the existence of another identical element in the process, method, article or system including that element.

[0086] The serial numbers of the above-mentioned embodiments of the present invention are only for description and do not represent the superiority or inferiority of the embodiments.

[0087] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory / random access memory, magnetic disk, optical disk), and includes several instructions for causing a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in various embodiments of the present invention.

[0088] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention accordingly. Any equivalent structure or equivalent process transformation made by using the content of the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A multi-region power load forecasting method, characterized in that, The multi-region power load forecasting method includes: Collect historical load time series data of multiple regions in the power system, and preprocess the historical load time series data to obtain candidate load time series data; Perform deep feature extraction on the candidate load time series data to obtain the deep feature matrix of each region; Perform multi-dimensional feature analysis on the candidate load time series data to obtain the multi-dimensional feature matrix of each region; Fuse the deep feature matrix and the multi-dimensional feature matrix to obtain the target load feature, and perform clustering analysis on the multiple regions based on the target load feature to determine the region clusters corresponding to each region; Construct a load forecasting model corresponding to each region cluster according to the target load feature; Perform power load forecasting on the power system based on the load forecasting models corresponding to each region cluster.

2. The multi-region power load forecasting method according to claim 1, wherein The performing deep feature extraction on the candidate load time series data to obtain the deep feature matrix of each region includes: Perform bidirectional long short-term dependence analysis on the candidate load time series data to obtain the forward hidden state and backward hidden state of each time step, and splice the forward hidden state and the backward hidden state to obtain the bidirectional dependence feature of each time step; Perform linear transformation based on the bidirectional hidden state to construct a query matrix, a key matrix, and a value matrix; Determine the attention feature of each time step according to the query matrix, the key matrix, and the value matrix: ； Among them, represents the time step of the attention feature, represents the query matrix, represents the matrix, represents the value matrix, represents the transpose, represents the dimension of the key matrix, represents the activation function; Splice the bidirectional dependence feature of each time step and the attention feature to obtain the initial deep feature: ； Among them, represents the initial depth feature at the time step , represents the feature concatenation operation, represents the bidirectional dependence feature at the time step . Construct a deep feature matrix based on the initial deep feature of each time step: ； Among them, represents the depth feature matrix, represents the time dimension, represents the feature dimension.

3. The multi-region power load forecasting method according to claim 2, wherein, The performing multi-dimensional feature analysis on the candidate load time series data to obtain the multi-dimensional feature matrix of each region: Calculate the mean, standard deviation, and coefficient of variation of the candidate load time series data; ； Among them, represents the mean of the candidate load time series data, represents the standard deviation of the candidate load time series data, represents the coefficient of variation; Perform global load fluctuation analysis according to the mean, the standard deviation, and the coefficient of variation to obtain the basic statistical feature; Perform time fluctuation analysis based on the candidate load time series data to obtain the time pattern feature, and the time pattern feature includes peak-valley difference, period mean ratio, and daily load rate: ； ； ； Among them, represents the peak-valley difference of the load within a day, represents the maximum value of the load in a day, represents the minimum value of the load within a day, represents the ratio of the load mean values in the morning and evening periods, represents the average load value in the morning period, represents the average load value in the evening period, represents the daily load factor, represents the average value of the load within a day; Perform short-term fluctuation analysis based on the candidate load time series data to obtain the volatility feature, and the volatility feature includes lag autocorrelation coefficient: ； in, represents the lagged autocorrelation coefficient, Indicates The load value of a time step is a data point in the candidate load time series data. represents the total number of time steps, represents the number of lag steps, represents the data mean of candidate load time series data; Combine the basic statistical feature, the time pattern feature, and the volatility feature into a multi-dimensional feature matrix.

4. The multi-region power load forecasting method according to claim 3, wherein, The fusing the deep feature matrix and the multi-dimensional feature matrix to obtain the target load feature, and performing clustering analysis on the multiple regions based on the target load feature to determine the region clusters corresponding to each region includes: Fuse the deep feature matrix and the multi-dimensional feature matrix to obtain the target load feature: ； Among them, represents the target load characteristic, represents the multi-dimensional feature matrix; Determine the feature similarity between regions based on the target load feature: ； Among them, represents the area and the area The feature similarity between them, represents the feature dimension of the target load feature, represents the area the th feature, represents the area the th feature; Perform clustering analysis on the multiple regions according to the feature similarity to determine the region clusters corresponding to each region.

5. The multi-region power load forecasting method according to claim 4, characterized in that The constructing a load forecasting model corresponding to each region cluster according to the target load feature includes: Construct an initial forecasting model corresponding to each region cluster, and the objective function of the initial forecasting model is as follows: ； Among them, represents the objective function, represents the loss function, represents the number of samples, represents the true value of the th sample, After the th sample passes through the decision tree set, represents the regularization term part used to prevent overfitting, represents the complexity function of the Perform normalization processing on the target load feature: ； Among them, represents the -th feature of the -th normalized sample, represents the -th feature of the -th sample before normalization, and respectively represent the maximum and minimum values of the -th feature; Evaluate the importance of each feature in the normalized target load feature, and filter the normalized target load feature based on the importance evaluation result to obtain the target training feature. The importance evaluation refers to the following formula: ； Among them, represents the gain value of the feature in the decision tree, representing the importance evaluation result; Train the initial prediction model based on the target training feature to obtain a load prediction model.

6. The multi-region power load forecasting method according to any one of claims 1 to 5, characterized in that The preprocessing of the historical load time series data to obtain candidate load time series data includes: Calculate the standard score of each data point in the historical load time series data: ； wherein, represents the standardized score of the data point, represents the mean of the historical load time series data, and represents the standard deviation of the historical load time series data; Mark the data points with a standard score higher than the anomaly threshold in the historical load time series data as anomaly data points, and perform data cleaning on the anomaly data points; Perform missing data filling on the historical load time series data after data cleaning to obtain the initial load time series data. The missing data filling is performed with reference to the following formula: ； Among them, represents the length of the data missing segment, represents the missing position of the current missing value in the missing segment, represents the known data points after the missing segment, represents the known data points before the missing segment; Perform normalization processing on the initial load time series data to obtain candidate load time series data. The normalization processing refers to the following formula: ； Among them, represents the normalized data points, represents the minimum data point in the initial load time series data, represents the maximum data point.

7. A multi-region power load forecasting device, characterized in that, The multi-region power load prediction device includes: A data processing module, configured to collect historical load time series data of multiple regions in a power system, and preprocess the historical load time series data to obtain candidate load time series data; A deep feature analysis module, configured to perform deep feature extraction on the candidate load time series data to obtain a deep feature matrix for each region; A multi-dimensional feature analysis module, configured to perform multi-dimensional feature analysis on the candidate load time series data to obtain a multi-dimensional feature matrix for each region; A feature clustering analysis module, configured to fuse the deep feature matrix and the multi-dimensional feature matrix to obtain a target load feature, and perform clustering analysis on the multiple regions based on the target load feature to determine the region clusters corresponding to each region; A prediction model construction module, configured to construct a load prediction model corresponding to each region cluster according to the target load feature; A power load prediction module, configured to perform power load prediction on the power system based on the load prediction models corresponding to each region cluster.

8. A multi - region power load forecasting device, characterized in that, The multi-region power load prediction device includes: a memory, a processor, and a multi-region power load prediction program stored on the memory and executable on the processor. The multi-region power load prediction program is configured to implement the multi-region power load prediction method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a multi-region power load prediction program, and when the multi-region power load prediction program is executed by a processor, it implements the multi-region power load prediction method according to any one of claims 1 to 6.

10. A computer program product, characterized in that, The computer program product includes a multi-region power load prediction program, and when the multi-region power load prediction program is executed by a processor, it implements the steps of the multi-region power load prediction method according to any one of claims 1 to 6.

Images