Short-term load prediction method based on feature selection

Abstract

The invention provides a short-term load prediction method based on feature selection. The method comprises the following steps: acquiring first historical load data and corresponding historical influence factor data of a to-be-predicted bus; wherein the first historical load data comprises load data of various load types, and the historical influence factor data comprises data of various influence factors; according to the first historical load data and the historical influence factor data, determining feature importance of each influence factor for each load type and a correlation coefficient between each influence factor and each load type; screening out associated influence factors of each load type from a plurality of influence factors by utilizing the feature importance and the correlation coefficient; and predicting the load data of the to-be-predicted bus based on the associated influence factors. According to the invention, core features associated with load prediction can be accurately screened out, interference of redundant information is reduced, and the precision of load prediction is improved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of load prediction, in particular to a short-term load prediction method based on feature selection. BACKGROUND

[0002] With the rapid development of the energy internet and the large-scale grid connection of renewable energy, the operation environment of the power system is becoming increasingly complex. As the core link of power system dispatching, energy trading and risk control, short-term load prediction needs to accurately capture the dynamic characteristics and spatio-temporal coupling relationship of load. However, the load time series presents nonlinear, strong time-varying and high coupling characteristics, and traditional prediction methods are difficult to cope with such complex patterns.

[0003] With the rapid development of artificial intelligence technology, the mainstream methods in the field of short-term load prediction include neural networks, decision trees, extreme learning machines, deep learning models and their hybrid prediction methods. In the field of time series prediction, the self-attention mechanism of Transformer has attracted much attention due to its excellent long-distance dependence modeling capability. However, when using Transformer modeling for short-term load prediction, it is necessary to select appropriate influencing factors as input features according to artificial prior knowledge, which makes it difficult to accurately identify and select key features, and redundant information may interfere with load prediction, affecting the accuracy of load prediction. SUMMARY

[0004] The embodiments of the application provide a short-term load prediction method based on feature selection, which can accurately select core features associated with load prediction, reduce the interference of redundant information, and improve the accuracy of load prediction.

[0005] In a first aspect, the embodiments of the application provide a short-term load prediction method based on feature selection, comprising: obtaining first historical load data of a to-be-predicted bus and corresponding historical influencing factor data; wherein the first historical load data includes load data of multiple load types, and the historical influencing factor data includes data of multiple influencing factors; determining the feature importance of each influencing factor for each load type and the correlation coefficient between each influencing factor and each load type according to the first historical load data and the historical influencing factor data; selecting, from the multiple influencing factors, an associated influencing factor of each load type by using the feature importance and the correlation coefficient; predicting the load data of the to-be-predicted bus based on the associated influencing factor.

[0006] In a second aspect, the embodiments of the application provide a short-term load prediction device based on feature selection, comprising: The acquisition module is configured to acquire first historical load data of a bus to be predicted and corresponding historical influence factor data; wherein the first historical load data comprises load data of multiple load types, and the historical influence factor data comprises data of multiple influence factors; The determination module is configured to determine, according to the first historical load data and the historical influence factor data, a feature importance of each influence factor for each load type and a correlation coefficient between each influence factor and each load type. The screening module is configured to screen, by using the feature importance and the correlation coefficient, a relevant influence factor of each load type from the multiple influence factors. The prediction module is configured to predict load data of the bus to be predicted based on the relevant influence factor.

[0007] In a third aspect, an embodiment of the present application provides an electronic device, including a memory and a processor, the memory stores a computer program, and the processor implements the method in the first aspect or any possible implementation manner of the first aspect when executing the computer program.

[0008] In a fourth aspect, an embodiment of the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the method in the first aspect or any possible implementation manner of the first aspect.

[0009] In a fifth aspect, an embodiment of the present application provides a computer program product, which includes a computer program, and the computer program is executed by a processor to implement the method in the first aspect or any possible implementation manner of the first aspect.

[0010] Compared with the prior art, the embodiment of the present application has the following beneficial effects: The embodiment of the present application analyzes the relationship between each influence factor and each load type by two ways, to accurately find the relevant influence factor having strong correlation with each load type. By the load data of each load type in the first historical load data and the data of each influence factor in the historical influence factor data, the feature importance of the influence factor for the load prediction of each load type can be analyzed, and the nonlinear statistical dependence between each influence factor and each load type can be quantified respectively, to obtain the corresponding correlation coefficient, and capture the potential complex relationship that may be ignored. By comprehensively considering the above two results, the core feature having high prediction value or strong statistical correlation can be screened for each load type, that is, the relevant influence factor of each load type. By using the relevant influence factor to predict the load data of the bus to be predicted, only the relevant influence factor of each load type can be used in the prediction process, the interference of redundant information is reduced, and the accuracy and precision of the load prediction are improved. BRIEF DESCRIPTION OF DRAWINGS

[0011] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram of the industrial load curve provided in the embodiments of this application; Figure 2 This is a schematic diagram of the business load curve provided in an embodiment of this application; Figure 3 This is a schematic diagram of the residential load curve provided in the embodiments of this application; Figure 4 This is a schematic diagram of the bus load curve provided in an embodiment of this application; Figure 5 This is a flowchart illustrating the implementation of a feature-based short-term load forecasting method according to an embodiment of this application. Figure 6 This is a schematic diagram of the structure of the long-term prediction model provided in the embodiments of this application; Figure 7 This is a flowchart illustrating the implementation of a short-term load forecasting method based on feature selection, provided in another embodiment of this application. Figure 8 This is a schematic diagram illustrating the importance analysis of factors influencing industrial load provided in the embodiments of this application; Figure 9 This is a schematic diagram illustrating the correlation analysis of factors influencing industrial load provided in an embodiment of this application; Figure 10 This is a schematic diagram illustrating the importance analysis of factors affecting business load provided in an embodiment of this application; Figure 11 This is a schematic diagram illustrating the correlation analysis of factors influencing business load provided in an embodiment of this application; Figure 12 This is a schematic diagram illustrating the importance analysis of factors influencing residential load provided in an embodiment of this application; Figure 13 This is a schematic diagram illustrating the correlation analysis of factors influencing residential load provided in an embodiment of this application; Figure 14 These are schematic diagrams of bus load curves predicted by different methods provided in the embodiments of this application; Figure 15 This is a schematic diagram of the predicted industrial load provided in an embodiment of this application; Figure 16 This is a schematic diagram illustrating the predicted business load results provided in an embodiment of this application; Figure 17 This is a schematic diagram of the predicted residential load provided in an embodiment of this application; Figure 18 This is a schematic diagram of the predicted bus load provided in the embodiments of this application; Figure 19 This is a schematic diagram of the short-term load forecasting device based on feature selection provided in the embodiments of this application; Figure 20 This is a schematic diagram of the electronic device provided in the embodiments of this application. Detailed Implementation

[0013] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0014] The inventors have discovered that when performing load forecasting, it is necessary to select influencing factors as input features. However, feature selection is highly dependent on prior knowledge and is difficult to adaptively identify and filter key features, resulting in redundant information interference and affecting the accuracy of load forecasting.

[0015] As a key node in the regional power grid, the power system busbar exhibits typical characteristics of multi-source heterogeneous loads, mainly composed of industrial, commercial, and residential user loads. The differences in the operating characteristics of different load types significantly affect the spatiotemporal distribution of the busbar load.

[0016] (1) Industrial load characteristics: Industrial users implement refined electricity management based on electricity prices, and their typical daily load curves show the characteristics of "low during the day and high at night", such as Figure 1 As shown, this type of load exhibits a clear periodicity and is influenced by both production scheduling and electricity prices: production shifts are adjusted to reduce electricity consumption during peak electricity price periods, while full-load production is carried out during off-peak periods to optimize costs. Among them, the load of manufacturing and processing enterprises typically decreases by 40%-60% on weekends / holidays, while the load fluctuations of continuous production industries such as chemicals and steel are relatively mild.

[0017] (2) Characteristics of commercial load: Commercial load exhibits significant weekday periodicity and peak characteristics, such as Figure 2 As shown, its typical daily load curve exhibits a single-peak shape, with load generally declining during holidays, and the load during long holidays often dropping to 20%-30% of the weekday load. Different business types lead to different load characteristics: large shopping malls show a steeper peak load characteristic than office buildings, while the catering industry exhibits a single-peak pattern at midday.

[0018] (3) Characteristics of residential load: such as Figure 3 As shown, residential load exhibits a complex, multi-peak, and time-segmented characteristic with a clear periodicity. The morning peak load is mainly driven by cooking, commuting appliances, and heating demand; the midday load drops to 60%-70% of the daily baseline; and the evening peak includes electricity consumption for cooking, bathing, and entertainment activities. It is noteworthy that residential load shows significant regional differences, and with the emergence of new load types such as evening charging of electric vehicles, the load during traditional off-peak hours shows a significant upward trend.

[0019] (4) Comprehensive characteristics of bus load: such as Figure 4 As shown, the bus load exhibits weak periodicity and strong volatility, specifically manifested as drastic fluctuations in daily load factor and a significant standard deviation in weekly load factor. This characteristic is primarily due to the following factors: ① The superposition effect of heterogeneous loads offsets or weakens the original periodic characteristics of various load types; ② The complexity of the power system operating environment (including equipment failures, temporary maintenance, and random changes in user behavior) leads to sharp changes in the daily load curve; ③ Meteorological conditions such as heavy rain, snow, and extreme temperatures significantly impact power system operation. Furthermore, sudden events such as natural disasters, major social events, and power grid failures can trigger short-term load fluctuations, further amplifying the volatility characteristics of the bus load.

[0020] Based on the complex characteristics of the bus load described above, the embodiments of this application analyze the relationship between each influencing factor and each load type using two methods. By combining the results of these two methods, core features with high predictive value or strong statistical correlation can be selected for each load type, accurately identifying the influencing factors strongly correlated with each load type. Then, by using these influencing factors to predict the load data of the bus to be predicted, the prediction process can utilize only the influencing factors for each load type, reducing interference from redundant information and thus improving the accuracy and precision of load prediction.

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following description will be provided in conjunction with the accompanying drawings and specific embodiments.

[0022] like Figure 5 The flowchart illustrating the implementation of the feature selection-based short-term load forecasting method is detailed below: Step 501: Obtain the first historical load data and the corresponding historical influencing factor data of the bus to be predicted; wherein, the first historical load data includes load data of multiple load types, and the historical influencing factor data includes data of multiple influencing factors.

[0023] In this embodiment, the bus load may include loads of various types, such as industrial loads, commercial loads, and residential loads. Accordingly, the first historical load data may include historical load data of industrial loads, commercial loads, and residential loads.

[0024] Historical influencing factor data includes data on various factors that may affect changes in bus load. This can include meteorological factors such as temperature, humidity, wind speed, and air pressure; calendar rules such as holidays, weekday types, months, and quarters; and historical load data. To avoid interference between continuous data, one-hot encoding can be used to discretize the calendar rule information, ensuring that each category feature is represented independently, thereby enhancing the model's ability to capture differentiated features.

[0025] Optionally, outlier handling and missing value imputation can be performed on the first historical load data and historical influencing factors to improve data quality and thus improve the accuracy and reliability of the prediction results.

[0026] Here, the interquartile range method can be used to detect and remove outliers in bus load and meteorological data that deviate from the normal range. A linear interpolation method based on time-series characteristics is used to efficiently fill in missing data values.

[0027] Step 502: Based on the first historical load data and historical influencing factor data, determine the characteristic importance of each influencing factor for each load type, and the correlation coefficient between each influencing factor and each load type.

[0028] In this embodiment, since the bus load consists of multiple types of loads such as industrial, commercial, and residential, and the operating characteristics of different types of loads are significantly different, the influencing factors are analyzed separately according to the load type to avoid the mixed analysis from obscuring the correlation patterns of a single type.

[0029] Here, two independent methods are used to analyze the correlation between load type and influencing factors, so as to achieve complementary screening of multi-source heterogeneous data.

[0030] One approach is to quantify the influence weight and feature importance of each influencing factor on load forecasting for each load type using models such as gradient boosting trees. For example, a prediction model for each load type can be constructed using LightGBM with a gradient boosting tree architecture. By utilizing the feature importance in the model, influencing factors with significant predictive power can be identified, effectively extracting influencing factors that significantly affect load forecasting as features.

[0031] Another approach is to analyze the correlation between the data of each influencing factor and the load data of each load type to obtain the correlation coefficient between each influencing factor and each load type. For example, this can be obtained by calculating the Pearson correlation coefficient between each influencing factor and each load type. Furthermore, the maximum information coefficient can be used to quantify the nonlinear statistical dependence between features and loads, circumventing model assumption limitations and capturing potentially overlooked complex relationships.

[0032] Optionally, the first historical load data includes load data for industrial load, commercial load, and residential load. In this embodiment, based on the first historical load data and historical influencing factor data, the characteristic importance of each influencing factor for each load type is determined. This can be achieved by: inputting the industrial load data and historical influencing factor data from the first historical load data into a preset gradient enhancement model to perform influencing factor characteristic analysis, obtaining the characteristic importance of each influencing factor output by the gradient enhancement model for the industrial load; inputting the commercial load data and historical influencing factor data from the first historical load data into a preset gradient enhancement model to perform influencing factor characteristic analysis, obtaining the characteristic importance of each influencing factor output by the gradient enhancement model for the commercial load; and inputting the residential load data and historical influencing factor data from the first historical load data into a preset gradient enhancement model to perform influencing factor characteristic analysis, obtaining the characteristic importance of each influencing factor output by the gradient enhancement model for the residential load.

[0033] In this embodiment, the gradient boosting model (such as GBDT, XGBoost, etc.) is an ensemble learning algorithm. Its internal mechanism quantifies the contribution weight of input features (i.e., influencing factors) to the output target (i.e., load value) through iterative training, thereby directly outputting the feature importance of each influencing factor. The higher the feature importance value, the more critical the influencing factor is in explaining the changes in the corresponding load type.

[0034] Here, considering the different load characteristics and variations of industrial load, commercial load, and residential load, such as the fact that temperature, which is important to residential load, may be less important to industrial load, and that mixed analysis would mask this difference, we conduct characteristic importance analysis for each load type separately.

[0035] Taking industrial load as an example, when performing feature importance analysis, the industrial load data and the data of each influencing factor in the first historical load data can be used as inputs to a pre-defined gradient enhancement model. The gradient enhancement model learns the mapping relationship between industrial load and each influencing factor, and can calculate the contribution weight of each influencing factor in fitting the industrial load change process, i.e., the feature importance of each influencing factor to the industrial load. Similarly, the same applies to commercial load and residential load, which will not be elaborated upon here.

[0036] Optionally, in this embodiment, the correlation coefficient between each influencing factor and each load type is determined based on the first historical load data and historical influencing factor data. This can be achieved by: first, establishing a dataset for each load type and each influencing factor based on the first historical load data and historical influencing factor data; then, gridding each dataset based on the historical load data corresponding to the load type and the historical influencing factor data corresponding to the influencing factor in each dataset to obtain multiple grid structures for each dataset; calculating the mutual information of each grid structure for each dataset based on the multiple grid structures for each dataset; and finally, determining the correlation coefficient between the influencing factor and the load type for each dataset based on the maximum mutual information of each dataset.

[0037] In this embodiment, since the characteristics of industrial, commercial and residential loads are significantly different and the influencing factors have different mechanisms of action on different load types, each load type and each influencing factor combination is analyzed separately, a binary data set is constructed and its correlation strength is quantified.

[0038] Here, a one-to-one binary data set is constructed for each load type and each influencing factor to achieve effective analysis of the correlation between the two. For example, datasets of industrial load and temperature, industrial load and holidays, commercial load and temperature, residential load and temperature, etc.

[0039] Gridding is used to capture the distribution characteristics of load and influencing factors by dividing the range of variable values. For each binary data set, the range of load data is divided into multiple continuous intervals, and the range of influencing factor data is divided into multiple continuous intervals, forming a two-dimensional grid.

[0040] For example, for a dataset of residential load and temperature, the residential load ranges from 0 to 5000 kW and is divided into m intervals, and the temperature ranges from -10 to 40℃ and is divided into n intervals, forming an m×n two-dimensional grid, which is a matrix of load intervals × influencing factor intervals.

[0041] By adjusting the number of continuous intervals (i.e., the values ​​of m and n), different levels of mesh refinement can be obtained, generating various mesh structures. This is because different correlation modes (such as linear, nonlinear, and piecewise correlation) may be more easily captured at different mesh refinements.

[0042] Mutual information is a metric that measures the information shared by two variables; a higher value indicates a stronger correlation between them. Mutual information is primarily calculated based on the sample distribution after grid partitioning. By counting the number of samples within each grid, the corresponding mutual information can be calculated.

[0043] For a grid structure, the number of samples falling into both a load interval and an influencing factor interval is counted. Based on the sample size, the joint probability and marginal probability of each cell can be calculated. The joint probability of each cell is the number of samples in that cell divided by the total number of samples. The marginal probability of each cell includes the marginal probability of the load type and the marginal probability of the influencing factor. The marginal probability of the load type is the number of samples in the load interval corresponding to that cell divided by the total number of samples, and the marginal probability of the influencing factor is the number of samples in the influencing factor interval corresponding to that cell divided by the total number of samples.

[0044] The mutual information of the grid structure can be calculated using the joint probability and marginal probability of each cell. Specifically, the formula for calculating mutual information is: In the formula, Indicates the load type in the dataset. This represents the influencing factors in the dataset. This represents an interval corresponding to the load type in the grid structure. This represents an interval corresponding to an influencing factor in the grid structure. Mutual information representing the mesh structure, Representing an interval and interval The joint probability of the formed lattice Representing an interval and interval The intervals corresponding to the formed grid The marginal probability, Representing an interval and interval The intervals corresponding to the formed grid The marginal probability.

[0045] Since different grid structures have different abilities to capture the correlation between load and influencing factors, one grid structure may be closer to the actual correlation pattern between the two. Therefore, it is necessary to select the maximum value from the mutual information of multiple grid structures, that is, the maximum mutual information.

[0046] This involves finding the maximum normalized mutual information. The maximum possible mutual information is... The mutual information value can be normalized by dividing the maximum information coefficient metric for each grid structure by the maximum possible mutual information. Therefore, the formula for calculating the maximum information coefficient is: In the formula, Indicates the maximum information coefficient. Indicates variable parameters, .

[0047] Finally, the calculated maximum information coefficient can be used to determine the correlation coefficient between each influencing factor and each load type.

[0048] Step 503: Using feature importance and correlation coefficient, screen out the associated influencing factors for each load type from a variety of influencing factors.

[0049] In this embodiment, the associated influencing factors for each load type are screened out by feature importance and correlation coefficient. The screening results of the two methods can be integrated to retain the core features that show high predictive value or strong statistical correlation in different load types. While reducing the interference of redundant information, it also takes into account the model's generalization ability and interpretability, which has significant advantages over a single method.

[0050] Optionally, by utilizing feature importance and correlation coefficients, the associated influencing factors for each load type can be screened from multiple influencing factors. This can be achieved by standardizing the feature importance of each load type to obtain the standard importance of each load type; standardizing the correlation coefficient of each load type to obtain the standard correlation coefficient of each load type; then calculating the comprehensive impact value of each influencing factor on each load type based on the standard importance and standard correlation coefficient of each load type; and finally, based on the comprehensive impact value, screening the associated influencing factors for each load type from multiple influencing factors.

[0051] In this embodiment, considering that the value ranges of feature importance and correlation are different, the feature importance and correlation coefficient of each load type can be standardized to obtain the corresponding standard importance and standard correlation coefficient, so as to combine the two analysis results and screen the influencing factors.

[0052] Here, weights can be assigned to the standard importance and standard correlation coefficient, and the weighted value of each influencing factor can be calculated to obtain the comprehensive impact value, thus combining the results of the two analysis methods. For example, the formula for calculating the comprehensive impact value can be: In the formula, Indicating influencing factors The overall impact value, Weighting coefficients representing the importance of the criteria The weighting coefficient represents the standard correlation coefficient. Indicating influencing factors The importance of standards Indicating influencing factors The standard correlation coefficient, where, and The sum is 1. For example, the weighting coefficients... and You can set them all to 0.5, or... It is 0.4. The value is 0.6. The above weighting coefficient values ​​are for illustrative purposes only and are not intended as limitations.

[0053] Finally, multiple influencing factors can be ranked using their comprehensive impact values. Factors with comprehensive impact values ​​exceeding a preset threshold can be selected as related influencing factors. Alternatively, factors with the largest preset number of comprehensive impact values ​​can be selected as related influencing factors. For example, the factor with the largest comprehensive impact value can be selected, or the four factors with the largest comprehensive impact values ​​can be selected as related influencing factors.

[0054] Step 504: Based on the associated influencing factors, predict the load data of the bus to be predicted.

[0055] In this embodiment, the associated influencing factors corresponding to each load type are selected through the aforementioned steps. Therefore, when predicting the load data of the bus to be predicted, the associated influencing factors corresponding to each load type can be used to predict the corresponding load data, so as to improve the accuracy and precision of load prediction.

[0056] This application analyzes the relationship between each influencing factor and each load type using two methods to accurately identify the strongly correlated influencing factors for each load type. Specifically, by analyzing the load data for each load type in the first historical load data set and the data for each influencing factor in the historical influencing factor data set, the characteristic importance of each influencing factor for load forecasting of each load type can be analyzed. Furthermore, the nonlinear statistical dependence between each influencing factor and each load type can be quantified to obtain the corresponding correlation coefficient, capturing potentially overlooked complex relationships. Combining these two results, core features with high predictive value or strong statistical correlation can be selected for each load type, i.e., the correlated influencing factors for each load type. Then, by using these correlated influencing factors to predict the load data of the bus to be predicted, the prediction process can utilize only the correlated influencing factors for each load type, reducing the interference of redundant information and thus improving the accuracy and precision of load forecasting.

[0057] The inventors have also discovered that short-term bus load forecasting requires accurately capturing the dynamic characteristics and spatiotemporal coupling relationships of heterogeneous loads from multiple sources, including industrial, commercial, and residential loads. The random fluctuations of industrial loads, the seasonal patterns of commercial loads, and the intraday periodic variations of residential loads result in load time series exhibiting nonlinear, highly time-varying, and highly coupled characteristics. Traditional forecasting methods struggle to handle such complex patterns. Single models cannot simultaneously leverage the advantages of local feature extraction and global time series modeling, leading to a significant trade-off between computational efficiency and prediction accuracy. Furthermore, existing models have limited ability to capture long-term dependencies, lack effective coordination mechanisms for multi-source loads, and fail to adequately utilize complementary information from different load types.

[0058] Due to the complex characteristics of bus loads, traditional forecasting methods, neglecting the differences among multi-source loads, struggle to accurately capture the underlying patterns of load changes, resulting in suboptimal forecasting performance. Direct forecasting methods fail to fully consider the cumulative effects of heterogeneous loads, random disturbances, and the impact of unforeseen events. Therefore, in the actual planning and operation of power systems, it is essential to adopt more accurate forecasting methods that comprehensively consider these factors to improve the accuracy of bus load forecasting and provide more reliable and efficient decision support for power system management.

[0059] In the embodiments of this application, based on the unique characteristics of each load type, the load data of each load type is predicted by utilizing the associated influencing factors corresponding to each load type, thereby achieving accurate prediction.

[0060] In some embodiments, predicting the load data of the bus to be predicted based on related influencing factors can be achieved by first obtaining the second historical load data of the bus to be predicted for the period preceding the predicted period, and the influencing factor data for the predicted period; then obtaining the related influencing factor data corresponding to each load type from the influencing factor data for the predicted period; finally, determining the load prediction data for each load type in the predicted period based on the load data for each load type in the second historical load data, the related influencing factor data corresponding to each load type, and the preset load prediction model for each load type; wherein, the load prediction model for each load type is trained based on the historical load data for each load type, the historical data of the related influencing factors corresponding to each load type, and the historical load data for the corresponding prediction period.

[0061] In this embodiment, since the load characteristics of each load type are significantly different, the load data can be predicted separately for each load type by utilizing the identified related influencing factors for each load type, thereby improving the accuracy and precision of load prediction for each load type.

[0062] Here, the time period to be predicted can be 1 hour, 2 hours, 4 hours, 24 hours, 48 ​​hours, or 72 hours, etc. There are no restrictions here, and it can be set as needed. Correspondingly, the length of the time period preceding the time period to be predicted can be the same as the length of the time period to be predicted.

[0063] Since the associated influencing factors differ for each load type, we can directly obtain all the influencing factor data and then extract the associated influencing factor data for each load type. This ensures that when forecasting load for each load type, we only focus on the corresponding associated influencing factors, eliminate the interference of redundant factors, reduce input noise, and improve forecasting efficiency and accuracy.

[0064] For each load type, a corresponding load forecasting model is trained. Different load forecasting models are used to capture the relationship between different load types and their corresponding influencing factors, achieving targeted load forecasting. Specifically, key hyperparameters such as the learning rate are optimized through grid search, and the root mean square error of the validation set is used as the evaluation metric to select the optimal hyperparameter combination, ensuring a balance between forecasting accuracy and computational efficiency.

[0065] Correspondingly, when forecasting the period to be forecasted, load forecasting models for each load type are used to forecast the load data for each load type separately, focusing on the patterns of their own load types, avoiding cross-interference, and further improving the accuracy and precision of load forecasting.

[0066] Optionally, to improve the accuracy of short-term bus load forecasting, the long-term forecasting model proposed in this application can be used for feature screening of related influencing factors and load forecasting. This long-term forecasting model is an improvement on the transformer model, forming a new bus load transformer (BLformer) model, the specific structure of which is as follows: Figure 6 As shown.

[0067] The first layer is the feature selection layer, which can be implemented using a lightweight gradient boosting machine (LightGBM) and maximum mutual information coefficient (MIC). This layer is used to implement the step of selecting related influencing factors as described in the previous embodiments. It explores the nonlinear relationship between bus load and external features such as meteorological parameters (temperature, humidity, etc.) and holidays, selecting the features that contribute most to load prediction, reducing the input dimensionality, retaining key driving factors of load changes, and providing high-information-density time-series feature input for subsequent deep learning modules.

[0068] It is important to note that this feature filtering layer can be included in the load forecasting model for each load type. Alternatively, it can be set as a separate module, independent of the load forecasting model for each load type. In this case, the BLformer model does not include this feature filtering layer, and the BLformer model is only used for load forecasting.

[0069] The second layer consists of instance normalization and block segmentation. Instance normalization performs local standardization on the load to adapt to its time-varying characteristics and eliminate dimensional differences. Block segmentation divides the continuous time series into fixed-length windows, preserving local patterns of short-term load fluctuations while laying the foundation for multi-scale time series feature extraction in subsequent layers. This block segmentation strategy balances local sensitivity with global modeling capabilities.

[0070] The third layer is a dilated convolutional layer. By adjusting the kernel dilation rate, it captures long-term cyclical patterns (annual / quarterly load trends) and cross-temporal correlation features (differences in patterns between holidays and weekdays) with lower computational cost. Compared to standard convolution, the sparse connectivity of dilated convolution significantly improves the propagation efficiency of temporal features, making it particularly suitable for scenarios with long-term fluctuations (such as seasonal electricity consumption habits), while avoiding the overfitting risk associated with fully connected layers.

[0071] The fourth layer is a projection + location encoding layer. The projection layer compresses the multi-channel block features into a unified space, eliminating dimensional differences between channels. The location encoding assigns a time-sensitive global identifier to each block, explicitly incorporating absolute time information (day of the week, season, etc.) and relative temporal relationships (time intervals between adjacent blocks).

[0072] The fifth layer is a multi-head sparse attention Transformer encoder. The multi-head attention mechanism computes the correlations at different time scales in parallel, and the sparse strategy reduces computational complexity by limiting the attention window. The multi-layer Transformer encoder optimizes implicit temporal patterns through self-attention iteration, and is particularly good at capturing complex nonlinear temporal dependencies in bus loads, outputting a temporal representation matrix.

[0073] The sixth layer is a flattening + linear output layer. The flattening operation transforms the high-dimensional feature mapping into a one-dimensional vector, and the linear output layer maps the feature space to the prediction space to generate the final prediction result.

[0074] In some embodiments, see Figure 7 The flowchart shown illustrates the implementation of the feature-based short-term load forecasting method. After determining the load forecast data for each load type in the forecast period based on the load data of each load type in the second historical load data, the data of the associated influencing factors corresponding to each load type, and the preset load forecasting model for each load type, the bus load forecast data for the bus to be forecasted can be calculated based on the preset weights of each load type in the bus to be forecasted and the load forecast data of each load type. The preset weights are determined based on the proportion of the historical load data of each load type in the corresponding bus load data.

[0075] In this embodiment, the load forecast data for each load type is weighted to obtain the total bus load, which is the bus load forecast data corresponding to the bus to be predicted during the forecast period. Compared with directly forecasting the total bus load as a whole, this embodiment can effectively solve the inherent defect of traditional aggregate forecasting methods that ignore the heterogeneity of different load types, and achieve accurate prediction of bus load data.

[0076] Here, the preset weight for each load type can be determined based on the weight of historical load data for each load type in the bus to be predicted.

[0077] Optionally, since the contribution of each load type to the bus load varies at different times, the weight of each load type can be dynamically allocated based on the contribution patterns at different times.

[0078] Here, the average proportion of each load type in the bus load of each time period in the historical load data can be calculated to obtain the initial weight of each load type in each time period. Then, by exponentially smoothing and normalizing the initial weights, the preset weight of each load type in each time period can be obtained.

[0079] The formula for calculating the initial weights can be: In the formula, Indicates time Load type Initial weights, load type For industrial load Commercial load and residential load One of them, Indicates the total period of historical data. Indicates the first in historical data One cycle, Indicates the first Each cycle time Load type The load value, Indicates the first Each cycle time The load value of the busbar load.

[0080] The formula for calculating the preset weights can be: In the formula, Indicates time Load type Preset weights, This represents all load types. Indicates time Load type The initial weights, This represents a smoothing factor, which allows for flexible control over the sensitivity of weights to changes in historical proportions. The optimal value is selected through grid search, with the goal of minimizing the mean absolute percentage error (MAPE) of the validation set.

[0081] Correspondingly, time The predicted value of the bus load can be In the formula, Indicates time The predicted value of the bus load, Indicates time Load type The load forecast value, This represents the scaling factor.

[0082] The formula for calculating the scaling factor is as follows: In the formula, Indicates time The historical average actual bus load of the busbar; Indicates time The average load is obtained by weighting all load types using their corresponding preset weights. , Indicates time Load type The historical average load.

[0083] Busbar loads mainly include industrial loads. Commercial load and residential load Three load types, then In the formula, Indicates time Industrial load The corresponding preset weights, Indicates time Commercial load The corresponding preset weights, Indicates time Residential load The corresponding preset weights, the three satisfy .

[0084] The above method allows for dynamic weighting of load forecasts for industrial, commercial, and residential loads, enabling accurate prediction of the bus load at the corresponding time.

[0085] Furthermore, since the bus load includes not only industrial, commercial, and residential loads, but also power output, the aforementioned proportional factors and dynamic weighting methods can be used to compensate for the remaining parts of the bus load using these three load types, thus achieving accurate prediction of the bus load.

[0086] In some embodiments, the short-term load forecasting method provided in this application is verified using actual operating data from a certain region in 2020. The sampling frequency of the actual operating data is 15 minutes. The dataset contains multi-dimensional load information: industrial, commercial, residential, and bus load data, and simultaneously collects key meteorological parameters such as temperature, air pressure, relative humidity, rainfall, and solar radiation intensity, providing comprehensive data support for the construction of the load forecasting model.

[0087] To evaluate the model's prediction accuracy, three statistical indicators were selected: mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE).

[0088] Feature contribution analysis of influencing factors: Historical load information, meteorological data, and calendar rules were selected as initial input features. Load information included historical data for the previous four hours of industrial, commercial, and residential loads; meteorological data included temperature, air pressure, humidity, rainfall, and solar radiation intensity; calendar rules included hourly time periods, weekday types, seasons, and holiday information.

[0089] To avoid interference between continuous data, one-hot encoding is used to discretize the calendar rule information, ensuring that each category feature is represented independently, thereby enhancing the model's ability to capture differentiated features. The correspondence between feature types and encodings is shown in Table 1.

[0090] Table 1 Feature Type and Coding Correspondence Table

[0091] To explore the dominant factors in load forecasting accuracy, the LightGBM and MIC algorithms were used together to analyze the contribution of load and meteorological features. The two algorithms exhibited different sensitivities to load and meteorological features. Combining multiple algorithms to evaluate feature importance allows for a more comprehensive understanding of the feature contribution mechanism, thereby improving forecasting accuracy.

[0092] For industrial load characteristics, see Figure 8 and Figure 9 Historical industrial load and holiday information showed a significant correlation in both algorithms. LightGBM was more sensitive to specific time-period characteristics, while MIC was more sensitive to rainfall and temperature (especially high temperatures). Based on the comprehensive evaluation results of the two algorithms, the industrial load prediction model ultimately incorporated historical industrial load, temperature, rainfall, and holiday characteristics.

[0093] For commercial load characteristics, see Figure 10 and Figure 11 Both algorithms demonstrate that historical business load, holiday information, temperature, and weekend information have a significant impact. LightGBM shows some sensitivity to seasonal information, while MIC does not exhibit similar characteristics. Based on the comprehensive evaluation results, the actual impact of seasonal variations on business activity is limited, and the business load forecasting model ultimately incorporates historical business load, temperature, holiday, and weekend characteristics.

[0094] For residential load characteristics, see Figure 12 and Figure 13The analysis results of the two algorithms were highly consistent, both indicating a significant correlation with historical resident load, holidays, weekends, and seasonal information. Based on the comprehensive evaluation results, the resident load prediction model ultimately incorporated historical resident load, holiday, weekend, and seasonal characteristics.

[0095] Based on this, the input characteristics of the load forecasting models for the three load types are obtained, as shown in Table 2.

[0096] Table 2 Feature Input Table

[0097] To achieve optimal prediction performance, the model hyperparameters were optimized using a grid search algorithm, with root mean square error (RMSE) used as the evaluation metric. The final parameter combinations are shown in Table 3.

[0098] Table 3 Parameter Selection and Optimization Range

[0099] To verify the predictive performance of the model (BLformer model) in the short-term load forecasting method provided in this application, three mainstream time series forecasting models were selected as benchmark comparison models, mainly including Informer, Autoformer, and Fedformer. A coupled forecasting paradigm was adopted, that is, industrial, commercial, and residential loads are predicted independently, and the bus load forecasting result is obtained by load superposition. All comparison models have undergone hyperparameter optimization, and the forecasting duration is uniformly set to 24-hour rolling forecasting.

[0100] like Figure 14 As shown, BLformer's predicted curve most closely approximates the true value distribution, outperforming the baseline model in both trend following and peak / valley feature capture. Its predicted trajectory exhibits good synchronization with the true curve, demonstrating excellent time-following ability. In contrast, Informer and Autoformer show significant prediction bias during load abrupt changes, while Fedformer, although improving overall accuracy, still suffers from phase shift issues. Specifically, during periods of sharp load increases / decreases, the baseline model shows a significantly larger prediction error compared to the proposed method (BLformer).

[0101] Furthermore, based on the model prediction accuracy comparison results shown in Table 4, the BLformer model proposed in this application demonstrates significant advantages in predicting industrial, commercial, residential, and bus loads. Specifically, compared to mainstream time series forecasting models such as Informer, Autoformer, and Fedformer, BLformer achieves the best performance in all three key evaluation metrics (RMSE, MAE, and MAPE).

[0102] For example, in industrial load forecasting, BLformer achieved an RMSE of 5.43, a MAE of 4.35, and a MAPE of 5.56%, which were on average 27.6%, 23.9%, and 29.8% lower than the suboptimal Fedformer model, respectively. This validates the effectiveness of BLformer in capturing complex industrial load patterns.

[0103] Table 4 Performance Comparison of Load Forecasting Models

[0104] To further validate the effectiveness of each module in the BLformer model, ablation experiments were conducted. The results, as shown in Table 5, demonstrate that the collaborative integration of modules in BLformer achieves simultaneous improvements in prediction accuracy and computational efficiency. Compared to the baseline model PatchTST, the full version of BLformer shows significant enhancements across all three metrics: RMSE (8.28 vs. 15.61), MAE (6.68 vs. 11.05), and MAPE (4.16% vs. 6.01%), representing improvements of 46.9%, 39.5%, and 30.8%, respectively. These results confirm the effectiveness of the module design.

[0105] Sparse Attention Module (SparseATT): This module significantly reduces computational complexity. In bus load prediction, the version using only this module reduced training time by 9.0% (38.5 vs. 42.3 minutes) and decreased RMSE by 25.9% (11.57 vs. 15.61), demonstrating its efficiency in modeling long-term dependencies.

[0106] FeatherSelection module: This module demonstrates excellent noise suppression capabilities. In bus load forecasting, this module improved the MAE index by 36.2% (7.06 vs. 11.05) and reduced the computation of redundant features by 30% through dynamic pruning.

[0107] Dilated Convolutional Module (DCNN): This module effectively captures sudden changes in training load. However, using it alone increases training time by 11.3% (47.1 vs. 42.3 minutes), highlighting the necessity of co-optimization of modules.

[0108] Integration performance: When all modules are integrated, the training time of BLformer is 40.2 minutes, which is 4.9% shorter than the baseline model and 14.6% faster than the version using only DCNN. Among them, sparseATT reduces the space complexity of traditional attention mechanisms, and FeatherSelection reduces the forward propagation computation through adaptive feature pruning.

[0109] Table 5 Comparison of Model Prediction Accuracy

[0110] Experimental results show that BLformer, by organically combining sparse attention mechanisms, dynamic feature selection, and dilated convolution, achieves accurate modeling of complex load characteristics while maintaining a lightweight model. This provides an efficient and effective solution for practical power system forecasting tasks.

[0111] This embodiment also verifies the performance advantages of the coupled prediction method through experimental analysis. First, Figures 15 to 17 The forecast results of the coupled forecasting model for industrial load, commercial load, and residential load are presented respectively. Subsequently, Figure 18 A comparative analysis of the coupled forecasting method and the direct total load forecasting method is presented. Experimental results show that coupled forecasting significantly outperforms direct forecasting in both forecasting accuracy and stability. This improvement is mainly due to the fact that the coupled forecasting method can fully consider the interaction between industrial, commercial, and residential loads, as well as the characteristic differences of various load types. By constructing a coupling mechanism that can capture the inherent correlation between different load types, the method in this application effectively improves the overall forecasting performance.

[0112] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0113] The following are device embodiments of this application. For details not described in detail, please refer to the corresponding method embodiments described above. Figure 19 The diagram shows a schematic of a short-term load forecasting device based on feature selection provided in an embodiment of this application. For ease of explanation, only the parts relevant to the embodiment of this application are shown: like Figure 19 As shown, the short-term load forecasting device 190 based on feature selection includes: The acquisition module 1901 is used to acquire the first historical load data and the corresponding historical influencing factor data of the bus to be predicted; wherein, the first historical load data includes load data of multiple load types, and the historical influencing factor data includes data of multiple influencing factors; The determination module 1902 is used to determine the characteristic importance of each influencing factor for each load type, and the correlation coefficient between each influencing factor and each load type, based on the first historical load data and historical influencing factor data. The screening module 1903 is used to screen out the associated influencing factors for each load type from a variety of influencing factors by utilizing feature importance and correlation coefficient; The prediction module 1904 is used to predict the load data of the bus to be predicted based on the associated influencing factors.

[0114] In some embodiments, the screening module 1903 is specifically used to: standardize the feature importance of each load type to obtain the standard importance of each load type; standardize the correlation coefficient of each load type to obtain the standard correlation coefficient of each load type; calculate the comprehensive impact value of each influencing factor on each load type based on the standard importance and standard correlation coefficient of each load type; and screen out the associated influencing factors of each load type from multiple influencing factors based on the comprehensive impact value.

[0115] In some embodiments, the prediction module 1904 is specifically configured to: acquire second historical load data of the bus to be predicted for the period preceding the predicted period, and influencing factor data for the predicted period; acquire data of associated influencing factors corresponding to each load type from the influencing factor data for the predicted period; determine load prediction data for each load type within the predicted period based on the load data for each load type in the second historical load data, the data of associated influencing factors corresponding to each load type, and a preset load prediction model for each load type; wherein the load prediction model for each load type is trained based on the historical load data for each load type, the historical data of associated influencing factors corresponding to each load type, and the historical load data for the corresponding prediction period.

[0116] In some embodiments, the prediction module 1904 is further configured to: calculate the bus load prediction data of the bus to be predicted based on the preset weight of each load type in the bus to be predicted and the load prediction data of each load type; wherein the preset weight is determined based on the proportion of the historical load data of each load type in the bus to be predicted in the corresponding bus load data.

[0117] In some embodiments, the first historical load data includes load data for industrial load, commercial load, and residential load; the determining module 1902 is specifically configured to: input the load data of industrial load and historical influencing factor data from the first historical load data into a preset gradient enhancement model, perform influencing factor feature analysis, and obtain the characteristic importance of each influencing factor output by the gradient enhancement model to the industrial load; input the load data of commercial load and historical influencing factor data from the first historical load data into a preset gradient enhancement model, perform influencing factor feature analysis, and obtain the characteristic importance of each influencing factor output by the gradient enhancement model to the commercial load; input the load data of residential load and historical influencing factor data from the first historical load data into a preset gradient enhancement model, perform influencing factor feature analysis, and obtain the characteristic importance of each influencing factor output by the gradient enhancement model to the residential load.

[0118] In some embodiments, the determining module 1902 is specifically configured to: establish a data set for each load type and each influencing factor based on the first historical load data and historical influencing factor data; grid each data set based on the historical load data corresponding to the load type and the historical influencing factor data corresponding to the influencing factor in each data set to obtain multiple grid structures for each data set; calculate the mutual information of each grid structure for each data set based on the multiple grid structures for each data set; and determine the correlation coefficient between the influencing factor and the load type corresponding to each data set based on the maximum mutual information of each data set.

[0119] In some embodiments, the acquisition module 1901 is further configured to: perform outlier processing and missing value imputation on the first historical load data and historical influencing factors.

[0120] Figure 20 This is a schematic diagram of the electronic device provided in an embodiment of this application. For example... Figure 20 As shown, the electronic device 2000 of this embodiment includes a processor 2001 and a memory 2002. The memory 2002 stores a computer program 2003. When the processor 2001 executes the computer program 2003, it implements the steps in the various method embodiments described above. Alternatively, when the processor 2001 executes the computer program 2003, it implements the functions of each module / unit in the various device embodiments described above.

[0121] For example, computer program 2003 may be divided into one or more modules / units, which are stored in memory 2002 and executed by processor 2001 to complete this application. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 2003 in electronic device 2000.

[0122] Electronic device 2000 may include, but is not limited to, processor 2001 and memory 2002. Those skilled in the art will understand that... Figure 20 This is merely an example of electronic device 2000 and does not constitute a limitation on electronic device 2000. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device may also include input / output devices, network access devices, buses, etc.

[0123] The processor 2001 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0124] The memory 2002 can be an internal storage unit of the electronic device 2000, such as a hard disk or RAM of the electronic device 2000. The memory 2002 can also be an external storage device of the electronic device 2000, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the electronic device 2000. Furthermore, the memory 2002 can include both internal and external storage units of the electronic device 2000. The memory 2002 is used to store computer programs and other programs and data required by the electronic device. The memory 2002 can also be used to temporarily store data that has been output or will be output.

[0125] For the sake of simplicity and clarity, only the above-described functional modules / units are used as examples. In practical applications, the functions described above can be assigned to different functional modules / units as needed. These modules / units can be implemented in hardware, software, or a combination of both.

[0126] This invention also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the methods described in the above-described method embodiments.

[0127] This invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the methods described in the above-described method embodiments.

[0128] Computer programs include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. Computer-readable media can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0129] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0130] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A short-term load forecasting method based on feature selection, characterized in that, include: Obtain the first historical load data and the corresponding historical influencing factor data of the bus to be predicted; wherein, the first historical load data includes load data of multiple load types, and the historical influencing factor data includes data of multiple influencing factors; Based on the first historical load data and the historical influencing factor data, determine the characteristic importance of each influencing factor for each load type, and the correlation coefficient between each influencing factor and each load type; Using the importance of the aforementioned features and the correlation coefficients, the associated influencing factors for each load type are screened from the multiple influencing factors; Based on the aforementioned influencing factors, the load data of the bus to be predicted is forecasted.

2. The short-term load forecasting method based on feature selection according to claim 1, characterized in that, Using the aforementioned feature importance and correlation coefficient, the associated influencing factors for each load type are screened from the multiple influencing factors, including: The characteristic importance of each load type is standardized to obtain the standard importance of each load type; The correlation coefficients for each load type were standardized to obtain the standard correlation coefficients for each load type. Calculate the comprehensive impact value of each influencing factor on each load type based on the standard importance and standard correlation coefficient of each load type; Based on the comprehensive impact value, the associated influencing factors for each load type are selected from a variety of influencing factors.

3. The short-term load forecasting method based on feature selection according to claim 1, characterized in that, Based on the aforementioned influencing factors, the load data of the bus to be predicted is predicted, including: Obtain the second historical load data of the bus to be predicted for the period preceding the predicted period, as well as the influencing factor data for the predicted period. From the data on influencing factors for the period to be predicted, obtain the data on the associated influencing factors corresponding to each load type; Based on the load data of each load type in the second historical load data, the data of the associated influencing factors corresponding to each load type, and the preset load prediction model for each load type, the load prediction data for each load type in the period to be predicted is determined respectively. The load forecasting model for each load type is trained based on historical load data for each load type, historical data of the associated influencing factors for each load type, and historical load data for the corresponding forecast period.

4. The short-term load forecasting method based on feature selection according to claim 3, characterized in that, After determining the load forecast data for each load type within the forecast period based on the load data for each load type in the second historical load data, the data of the associated influencing factors corresponding to each load type, and the preset load forecast model for each load type, the method further includes: Based on the preset weight of each load type in the busbar to be predicted and the load prediction data of each load type, calculate the busbar load prediction data of the busbar to be predicted; The preset weights are determined based on the proportion of historical load data for each load type in the corresponding bus load data.

5. The short-term load forecasting method based on feature selection according to any one of claims 1 to 4, characterized in that, The first historical load data includes load data for industrial load, commercial load, and residential load; Based on the first historical load data and the historical influencing factor data, determine the characteristic importance of each influencing factor for each load type, including: The industrial load data and the historical influencing factor data in the first historical load data are input into a preset gradient improvement model to perform influencing factor feature analysis, and the characteristic importance of each influencing factor output by the gradient improvement model to the industrial load is obtained. The load data of commercial load in the first historical load data and the historical influencing factor data are input into a preset gradient boosting model to perform influencing factor feature analysis, and the characteristic importance of each influencing factor output by the gradient boosting model to the commercial load is obtained. The load data of residents' load in the first historical load data and the historical influencing factor data are input into a preset gradient improvement model to perform influencing factor feature analysis, and the characteristic importance of each influencing factor output by the gradient improvement model to the residents' load is obtained.

6. The short-term load forecasting method based on feature selection according to any one of claims 1 to 4, characterized in that, Based on the first historical load data and the historical influencing factor data, determine the correlation coefficient between each influencing factor and each load type, including: Based on the first historical load data and the historical influencing factor data, establish a data set for each load type and each influencing factor; Based on the historical load data corresponding to the load type and the historical influencing factor data corresponding to the influencing factors in each dataset, each dataset is gridded to obtain multiple grid structures for each dataset; Based on the multiple grid structures for each dataset, calculate the mutual information for each grid structure of each dataset; Based on the maximum mutual information of each dataset, the correlation coefficient between the influencing factors and the load type corresponding to each dataset is determined.

7. The short-term load forecasting method based on feature selection according to any one of claims 1 to 4, characterized in that, Before determining the characteristic importance of each influencing factor for each load type, and the correlation coefficient between each influencing factor and each load type, based on the first historical load data and the historical influencing factor data, the method further includes: Outlier handling and missing value imputation are performed on the first historical load data and the historical influencing factors.

8. A short-term load forecasting device based on feature selection, characterized in that, include: The acquisition module is used to acquire the first historical load data and the corresponding historical influencing factor data of the bus to be predicted; wherein, the first historical load data includes load data of multiple load types, and the historical influencing factor data includes data of multiple influencing factors; The determination module is used to determine, based on the first historical load data and the historical influencing factor data, the characteristic importance of each influencing factor for each load type, and the correlation coefficient between each influencing factor and each load type; A screening module is used to select the associated influencing factors for each load type from the multiple influencing factors by utilizing the importance of the features and the correlation coefficient; The prediction module is used to predict the load data of the bus to be predicted based on the associated influencing factors.

9. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images