A short-term bus load prediction method, device, equipment and medium

Abstract

The application discloses a short-term bus load prediction method, device, equipment and medium, and belongs to the power grid prediction field.The method is: collecting original bus load data of a to-be-predicted period, performing multi-scale time sequence segmentation on the original bus load data according to a preset scale segmentation number, and obtaining a plurality of load time sequences representing load evolution characteristics under different scales; decomposing each load time sequence into a seasonal component and a trend component; inputting the seasonal component and the trend component under each scale into a pre-trained first multi-scale fusion time sequence network, performing bidirectional cross-scale feature interaction through the first multi-scale fusion time sequence network, and obtaining joint features under each scale; performing time sequence reconstruction according to the joint features under each scale, and outputting a final short-term bus load prediction result.The application can improve the accuracy of short-term bus load prediction.

Description

Technical Field

[0001] This application relates to the field of power grid forecasting, and in particular to a method, apparatus, equipment and medium for short-term bus load forecasting. Background Technology

[0002] Short-term bus load forecasting is a fundamental task in power system operation and dispatch planning, and its accuracy directly impacts the safe operation and economic benefits of the power grid. With the deepening reform of the power market and the continuous increase in the proportion of renewable energy integration, accurate bus load forecasting provides crucial decision-making support for power grid dispatching agencies, including the rational scheduling of unit start-up and shutdown, optimization of power load distribution, and improvement of reactive power compensation and voltage control response capabilities. Especially when dealing with sudden load fluctuations and seasonal peak electricity demand, high-precision forecasting results can effectively reduce power grid operation risks and ensure power supply reliability, making it a core component in realizing intelligent control of the new power system.

[0003] However, existing short-term bus load forecasting technologies still have significant limitations in practical applications. On the one hand, traditional forecasting models often employ single-scale modeling, making it difficult to simultaneously capture the intertwined long-term growth trends and complex short-term random disturbances in load sequences, resulting in insufficient ability to characterize non-stationary load features. On the other hand, existing multi-scale models often lack adaptive parameter adjustment mechanisms during feature fusion, and the interactions between scales are mostly unidirectional information transmissions, failing to achieve deep coupling between local details and the global context. These shortcomings collectively lead to significant deviations in the forecasting results of existing technologies under dynamic operating conditions with varying load characteristics, making it difficult to meet the increasingly refined grid dispatching requirements. Summary of the Invention

[0004] This application provides a method, apparatus, equipment, and medium for short-term bus load forecasting, which can solve the problem of low accuracy in short-term bus load forecasting in the prior art.

[0005] Some embodiments of this application provide a short-term bus load forecasting method, including: The original bus load data for the period to be predicted is collected, and the original bus load data is divided into multiple scale time series according to the preset number of scale segments to obtain several load time series that characterize the load evolution characteristics at different scales. The number of scale segments is determined by a competitive optimization method that integrates differential evolution strategies. Each load time series is decomposed into a seasonal component characterizing the periodicity of the load and a trend component characterizing the long-term evolution of the load. The seasonal and trend components at each scale are input into a pre-trained first multi-scale fusion temporal network. The first multi-scale fusion temporal network performs bidirectional cross-scale feature interaction to align and map the periodic patterns and evolution trends at different scales, thereby obtaining joint features at each scale. Based on the joint features at each scale, time series reconstruction is performed to output the final short-term bus load forecast results.

[0006] Compared with existing technologies, the above embodiments have the following advantages: Firstly, this application utilizes a competitive optimization method based on an integrated differential evolution strategy to adaptively determine the number of scale segments matching the load fluctuation pattern, enabling multi-scale time series segmentation to conform to actual load change patterns and avoiding feature loss or redundancy caused by inappropriate scale selection. Subsequently, the load time series at each scale is decomposed into seasonal and trend components, achieving decoupling of load cycle patterns and long-term evolution trends, reducing the impact of non-stationarity on prediction accuracy. Based on this, a multi-scale fusion time series network is used to perform bidirectional cross-scale feature interaction, allowing short-term fluctuation information and long-term trend information to mutually correct and supplement each other at different scales, forming more consistent and discriminative joint features. Finally, time series reconstruction is performed based on the joint features, and the resulting prediction results can simultaneously consider local changes and overall trends, thereby overcoming the problem of large prediction deviations in complex load scenarios in existing technologies and significantly improving the accuracy of short-term bus load prediction.

[0007] Furthermore, the step of decomposing each load time series into a seasonal component characterizing the periodicity of the load and a trend component characterizing the long-term evolution of the load includes: Perform a Fast Fourier Transform on each of the aforementioned load time series to obtain the corresponding frequency domain sequence; Based on a preset low-pass filtering threshold, the low-frequency components in each frequency domain sequence are retained while the high-frequency components are suppressed to obtain the corresponding trend frequency domain components. Perform an inverse Fourier transform on each of the aforementioned trend frequency domain components to obtain each of the aforementioned trend components in the time domain; Subtracting the corresponding trend component from each of the load time series yields the corresponding seasonal component.

[0008] Compared to existing technologies, the above embodiments have the following advantages: By performing Fast Fourier Transform on the load time series at various scales and separating low-frequency and high-frequency components based on a low-pass filter threshold, frequency domain decoupling of the long-term load trend and periodic fluctuations is achieved. On the one hand, the trend component obtained by the inverse transform of the low-frequency component can stably characterize the overall evolution trend of the load, avoiding interference from short-term random disturbances on trend modeling; on the other hand, the seasonal component obtained through the residual method completely preserves the periodic fluctuation characteristics. Compared with the direct time-domain decomposition method, this frequency-domain filtering-based decomposition method has the advantages of high computational efficiency and clear component boundaries, effectively reducing the non-stationarity of the load series and providing more physically meaningful and discriminative input features for subsequent multi-scale feature fusion.

[0009] Furthermore, the step of performing bidirectional cross-scale feature interaction through the first multi-scale fusion temporal network to align and map the periodic patterns and evolution trends at different scales, and obtaining joint features at each scale, includes: Using the seasonal component at the lowest scale as the starting seasonal feature, the seasonal feature containing the load cycle pattern at the previous scale is upsampled and aligned according to the path from low to high scale to obtain the corresponding upsampled feature. The upsampled feature is then fused with the seasonal component at the current scale to obtain the seasonal feature at the current scale after being corrected by the long-cycle feature. Using the trend component at the highest scale as the starting trend feature, the trend feature containing the load evolution trend at the previous scale is downsampled and compressed according to the path from high to low scale to obtain the corresponding downsampled feature. The downsampled feature is then fused with the trend component at the current scale to obtain the trend feature at the current scale that incorporates instantaneous fluctuation information. The seasonal features and trend features at the same scale are fused across channels to obtain the joint features at each scale.

[0010] Compared to existing technologies, the above embodiments have the following advantages: By constructing a bidirectional cross-scale feature interaction mechanism, collaborative modeling of load cycle patterns and evolution trends at different time scales is achieved. Specifically, seasonal components are upsampled and fused step-by-step from low to high scales, allowing short-cycle fluctuations to be corrected against a long-cycle background; trend components are downsampled and fused step-by-step from high to low scales, enabling long-term evolution information to constrain the instability of short-term predictions. Furthermore, through cross-channel fusion at the same scale, deep coupling of cycle and trend information is achieved. This bidirectional interaction structure overcomes the limitations of unidirectional transmission in traditional multi-scale models, significantly enhancing the consistency and complementarity of information across scales, and improving the expressive power of joint features for complex load change patterns.

[0011] Furthermore, the original bus load data is divided into multiple time series segments according to a preset number of scale segments to obtain several load time series characterizing the load evolution characteristics at different scales, including: A sampling operator for the number of scale segments is generated by using preset rules; wherein each sampling operator corresponds to a scale. The original bus load data is resampled using each of the sampling operators to generate multiple load subsequences with different sampling frequencies; Each load subsequence is truncated by a sliding window, and the truncated sequences are normalized in dimension to obtain multiple load time series that are interconnected in the time dimension and have different scales.

[0012] Compared to existing technologies, the above embodiments offer the following advantages: By generating corresponding sampling operators for different scales and resampling the original bus load data, multiple load subsequences with different sampling frequencies but consistent temporal correlation are constructed, providing a clear data source foundation for multi-scale modeling. Combining sliding window truncation and dimensionality normalization not only ensures the alignment of time series at different scales in the time dimension but also eliminates the amplitude bias problem caused by differences in sampling frequencies. Compared to directly extracting multi-scale features from the same sequence, this approach achieves scale decoupling at the data level, helping the model learn load evolution characteristics at different time resolutions, thereby improving the stability and fusionability of multi-scale features.

[0013] Furthermore, the number of scale segments is determined through a competitive optimization method integrating differential evolution strategies, including: Construct an initial parameter combination population containing multiple different combinations of first training parameters; wherein, the first training parameter combination includes the scale segmentation number; The initial parameter combination population is iteratively updated using the differential evolution strategy until the preset iteration termination condition is met, resulting in the final parameter combination population. Train an untrained second multi-scale fusion temporal network according to each second training parameter combination in the final parameter combination population, and obtain the first training result corresponding to each second training parameter combination. The number of scale segments is determined based on the combination of second training parameters corresponding to the optimal first training result.

[0014] Compared to existing technologies, the above embodiments have the following advantages: By introducing a competitive optimization method with an integrated differential evolution strategy, the number of scale segments is incorporated into the learnable training parameter space, achieving adaptive determination of the multi-scale structure. Compared to manually setting or empirically selecting the number of scales, this method can automatically select the optimal scale configuration based on actual prediction performance, avoiding the problems of insufficient feature representation due to too few scales or redundant noise introduced by too many scales. Simultaneously, through competitive evaluation of model training results under different parameter combinations, the determination of the number of scale segments is directly linked to prediction accuracy, ensuring the matching between the multi-scale modeling structure and the characteristics of the load data from a mechanistic perspective.

[0015] Further, the iterative update of the initial parameter combination population using the differential evolution strategy includes: During each iteration update, several third training parameter combinations are selected from the current parameter combination population to generate mutation combinations; By performing cross-operations between the mutation combination and each of the third training parameter combinations in the current parameter combination population, several experimental combinations can be obtained. The second multi-scale fusion temporal network is trained by each of the aforementioned experimental combinations to obtain the second training result corresponding to each of the aforementioned experimental combinations. The second multi-scale fusion temporal network is trained by each of the third training parameter combinations to obtain the third training result corresponding to each of the third training parameter combinations. Based on the second training result and the third training result, each of the experimental combinations competes with each of the third training parameter combinations, and the losing third training parameter combination is replaced with the winning experimental combination.

[0016] Compared to existing technologies, the above embodiments have the following advantages: By introducing an iterative mechanism into the differential evolution process, the parameter search process possesses continuous exploration and the ability to weed out inferior solutions. In each iteration, the experimental combination and the original parameter combination are compared and trained under the same network structure, effectively avoiding the premature convergence problem caused by relying on only a single evaluation path. This competitive update strategy can gradually eliminate parameter combinations with poor prediction performance and retain solutions with better generalization ability, thereby improving the number of scale segments and the stability and global optimality of related training parameter search, providing a reliable parameter foundation for the long-term operation of the model.

[0017] Furthermore, after outputting the final short-term bus load forecast result, the following are also included: Collect the actual measurement results corresponding to the short-term bus load forecast results, and calculate the error between the actual measurement results and the short-term bus load forecast results; Determine whether the error is within a preset range; If it is within the preset range, then update the model parameters of the first multi-scale fusion temporal network; If the network is not within the preset range, the training parameters of the first multi-scale fusion temporal network are updated using the differential evolution strategy, and the first multi-scale fusion temporal network is retrained based on the updated training parameters.

[0018] Compared to existing technologies, the above embodiments have the following advantages: By introducing actual measurement results as feedback after the prediction output and dynamically adjusting model parameters or training parameters based on the prediction error, the prediction model possesses online adaptive optimization capabilities. When the prediction error is within a reasonable range, only fine-tuning of the model parameters is performed, which helps maintain model stability; when the error exceeds the threshold, parameter re-optimization based on a differential evolution strategy is triggered, preventing continuous performance degradation of the model when load characteristics change. This closed-loop update mechanism enables the model to continuously evolve with changes in the load operating environment, improving the robustness and reliability of short-term bus load forecasting in long-term application scenarios.

[0019] Another embodiment of this application provides a short-term bus load forecasting device, including: a data processing module, a sequence decomposition module, a feature processing module, and a forecasting module; The data processing module is used to collect the original bus load data of the period to be predicted, and to perform multi-scale time series segmentation on the original bus load data according to the preset number of scale segments to obtain several load time series that characterize the load evolution characteristics at different scales. The number of scale segments is determined by a competitive optimization method that integrates differential evolution strategies. The sequence decomposition module is used to decompose each load time series into a seasonal component that characterizes the periodicity of the load and a trend component that characterizes the long-term evolution trend of the load. The feature processing module is used to input the seasonal components and trend components at each scale into a pre-trained first multi-scale fusion temporal network, and perform bidirectional cross-scale feature interaction through the first multi-scale fusion temporal network to align and map the periodic patterns and evolution trends at different scales, thereby obtaining joint features at each scale. The prediction module is used to perform time-series reconstruction based on the joint features at each scale and output the final short-term bus load prediction result.

[0020] Another embodiment of this application also provides a terminal device, including: a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the steps of the short-term bus load forecasting method of this application.

[0021] Another embodiment of this application also provides a computer-readable storage medium item, including: a stored computer program, which, when the computer program is running, controls the device where the computer-readable storage medium is located to perform the steps of the short-term bus load forecasting method of this application. Attached Figure Description

[0022] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments 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 from these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating a short-term bus load forecasting method provided in some embodiments of this application; Figure 2 This is a comparison chart of short-line bus load prediction results provided in some embodiments of this application; Figure 3 This is a comparison chart of the main error indicators in a short-term bus load forecasting task provided in some embodiments of this application; Figure 4 This is a comparison chart of convergence curves for iterative updates of training parameters provided in some embodiments of this application; Figure 5 This is a comparison chart of prediction error distribution provided in some embodiments of this application; Figure 6 This is a schematic diagram of the structure of a short-term bus load prediction device provided in some embodiments of this application. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0026] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0027] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0028] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0029] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0030] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0031] Existing short-term bus load forecasting technologies still have significant limitations in practical applications. On the one hand, traditional forecasting models often employ single-scale modeling, making it difficult to simultaneously capture the intertwined long-term growth trends and complex short-term random disturbances in load sequences, resulting in insufficient ability to characterize non-stationary load features. On the other hand, existing multi-scale models often lack adaptive parameter adjustment mechanisms during feature fusion, and the interactions between scales are mostly unidirectional information transmissions, failing to achieve deep coupling between local details and the global context. These shortcomings collectively lead to significant deviations in the forecasting results of existing technologies under dynamic operating conditions with varying load characteristics, making it difficult to meet the increasingly refined grid dispatching requirements.

[0032] Please refer to Figure 1 To address the problem of low accuracy in short-term bus load forecasting in existing technologies, this application provides a short-term bus load forecasting method, comprising the following steps S101 to S104: S101: Collect the original bus load data for the period to be predicted, and perform multi-scale time series segmentation on the original bus load data according to the preset number of scale segments to obtain several load time series that characterize the load evolution characteristics at different scales. The number of scale segments is determined by a competitive optimization method that integrates differential evolution strategies.

[0033] Preferably, in some embodiments of this application, the original bus load data includes: bus voltage, current, active and reactive power, frequency, switch status, power factor, meteorological characteristics (temperature, humidity, wind speed, solar irradiance, etc.), time period information (hour, weekday, holiday identifiers) and regional load statistics collected from the power system bus and related equipment.

[0034] Preferably, in some embodiments of this application, before performing multi-scale time-series segmentation on the original bus load data, the method further includes: repairing and smoothing missing values, outliers, and measurement noise in the collected data using interpolation, bidirectional imputation, and moving average filtering; normalizing or standardizing data of different dimensions to ensure that feature values ​​are distributed within a uniform range; and ensuring the consistency of the original bus load data in the time dimension through timestamp alignment and synchronous sampling, thereby forming a complete, continuous, high-quality bus load time series. ,in, The total length of the time series. Indicates at time The multidimensional observation vector, The number of input feature dimensions.

[0035] Furthermore, in some embodiments of this application, the step of performing multi-scale time-series segmentation on the original bus load data according to a preset number of scale segments to obtain several load time series characterizing the load evolution characteristics at different scales includes: A sampling operator for the number of scale segments is generated by using preset rules; wherein each sampling operator corresponds to a scale. The original bus load data is resampled using each of the sampling operators to generate multiple load subsequences with different sampling frequencies; Each load subsequence is truncated by a sliding window, and the truncated sequences are normalized in dimension to obtain multiple load time series that are interconnected in the time dimension and have different scales.

[0036] Preferably, in some embodiments of this application, the set corresponding to the sampling operator is: ,in, , Representing the The sampling step size (i.e., sampling operator, whose unit can be minutes or number of sampling points) corresponding to the scale. This represents the number of scale segments, and also the maximum scale level.

[0037] Preferably, in some embodiments of this application, the first The steps for obtaining load time series at the scale include: , ; in, For the first The first at the level scale Each sample point represents the average result of the bus load time series within this scale window; This is the floor symbol.

[0038] The above formulas can be used to obtain load time series at different scales. ,in, , For the first Number of data points in the load time series at the level scale.

[0039] This application constructs multiple load subsequences with different sampling frequencies but consistent temporal correlation by generating corresponding sampling operators for different scales and resampling the original bus load data, thus providing a clear data source foundation for multi-scale modeling. Combining sliding window truncation and dimensionality normalization not only ensures the alignment of time series at different scales in the time dimension but also eliminates the amplitude bias problem caused by differences in sampling frequencies. Compared to directly extracting multi-scale features from the same sequence, this approach achieves scale decoupling at the data level, helping the model learn the load evolution characteristics at different time resolutions, thereby improving the stability and fusionability of multi-scale features.

[0040] Furthermore, in some embodiments of this application, the number of scale segments is determined by a competitive optimization method integrating differential evolution strategies, including: Construct an initial parameter combination population containing multiple different combinations of first training parameters; wherein, the first training parameter combination includes the scale segmentation number; The initial parameter combination population is iteratively updated using the differential evolution strategy until the preset iteration termination condition is met, resulting in the final parameter combination population. Train an untrained second multi-scale fusion temporal network according to each second training parameter combination in the final parameter combination population, and obtain the first training result corresponding to each second training parameter combination. The number of scale segments is determined based on the combination of second training parameters corresponding to the optimal first training result.

[0041] Preferably, in some embodiments of this application, all training parameter combinations in the embodiments of this application include: the number of scale segments, the model parameters of the multi-scale fusion temporal network, and the hyperparameters in the training process of the multi-scale fusion temporal network.

[0042] Preferably, in some embodiments of this application, all training parameter combinations in the embodiments of this application are ,in, These are the model parameters for a multi-scale fusion temporal network. The learning rate during the training process of a multi-scale fusion temporal network; For regularization coefficients during the training process of multi-scale fusion temporal networks; These are the weighting coefficients of the loss function during the training of a multi-scale fusion temporal network; The number of scale segments also represents the number of multi-scale layers in a multi-scale fusion temporal network.

[0043] Preferably, in some embodiments of this application, the loss function during the training process of the multi-scale fusion temporal network is... ,include: ; in, Indicates predicted value Compared with the actual load value The mean square error between them; Indicates the parameter regularization term; This indicates a predictive smoothing constraint.

[0044] Preferably, in some embodiments of this application, constructing an initial parameter combination population containing multiple different first training parameter combinations includes: obtaining an initial template parameter combination. ,right Adding a perturbation term yields the initial parameter combination population. ,in, ,in, For the first The first training parameter combination, This is a disturbance term.

[0045] Preferably, in some embodiments of this application, after obtaining the training results corresponding to any combination of training parameters, the fitness of the training parameter combination is evaluated based on the training results using the following formula. : ; in, This is the normalized root mean square error; Mean absolute percentage error; Indicates the training time during the training process; These represent the weight coefficients corresponding to each objective. .

[0046] Preferably, in some embodiments of this application, determining the number of scale segments based on the second training parameter combination corresponding to the optimal first training result includes: calculating the fitness corresponding to each second training parameter combination according to the fitness calculation function mentioned in the above preferred embodiments, and selecting the number of scale segments in the second training parameter combination with the optimal fitness as the basis for subsequent scale division.

[0047] This application introduces a competitive optimization method based on an integrated differential evolution strategy, incorporating the number of scale segments into a learnable training parameter space to achieve adaptive determination of the multi-scale structure. Compared to manually setting or empirically selecting the number of scales, this method can automatically select the optimal scale configuration based on actual prediction performance, avoiding the problems of insufficient feature representation due to too few scales or redundant noise introduced by too many scales. Furthermore, through competitive evaluation of model training results under different parameter combinations, the determination of the number of scale segments is directly linked to prediction accuracy, ensuring the matching between the multi-scale modeling structure and the characteristics of the load data from a mechanistic perspective.

[0048] Furthermore, in some embodiments of this application, the iterative update of the initial parameter combination population using the differential evolution strategy includes: During each iteration update, several third training parameter combinations are selected from the current parameter combination population to generate mutation combinations; By performing cross-operations between the mutation combination and each of the third training parameter combinations in the current parameter combination population, several experimental combinations can be obtained. The second multi-scale fusion temporal network is trained by each of the aforementioned experimental combinations to obtain the second training result corresponding to each of the aforementioned experimental combinations. The second multi-scale fusion temporal network is trained by each of the third training parameter combinations to obtain the third training result corresponding to each of the third training parameter combinations. Based on the second training result and the third training result, each of the experimental combinations competes with each of the third training parameter combinations, and the losing third training parameter combination is replaced with the winning experimental combination.

[0049] Preferably, in some embodiments of this application, the calculation formula for the variant combination is: ; in, Representing the Next iteration update; For the first The iteration update yields the... Item variation combination; , , The first The combination of three third training parameters selected during the next iteration update; The coefficient of variation is 1. .

[0050] Preferably, in some embodiments of this application, the calculation formula for the test combination is: ; in, For the first The iteration update yields the... Test combination; For the cross operator; For the first The first iteration update The third training parameter combination; This represents the crossover probability.

[0051] Preferably, in some embodiments of this application, each of the experimental combinations competes with each of the third training parameter combinations using the following formula: ; in, A function for calculating fitness; For the first The first iteration update The third training parameter combination.

[0052] By introducing an iterative mechanism into the differential evolution process, the parameter search process possesses continuous exploration and elimination capabilities. In each iteration, the experimental combination and the original parameter combination are compared and trained under the same network structure, effectively avoiding premature convergence caused by relying on a single evaluation path. This competitive update strategy can gradually eliminate parameter combinations with poor prediction performance and retain solutions with better generalization ability, thereby improving the number of scale segments and the stability and global optimum of related training parameter search, providing a reliable parameter foundation for long-term model operation.

[0053] S102: Decompose each load time series into a seasonal component that characterizes the periodicity of the load and a trend component that characterizes the long-term evolution of the load.

[0054] Furthermore, in some embodiments of this application, the step of decomposing each load time series into a seasonal component characterizing the periodicity of the load and a trend component characterizing the long-term evolution trend of the load includes: Perform a Fast Fourier Transform on each of the aforementioned load time series to obtain the corresponding frequency domain sequence; Based on a preset low-pass filtering threshold, the low-frequency components in each frequency domain sequence are retained while the high-frequency components are suppressed to obtain the corresponding trend frequency domain components. Perform an inverse Fourier transform on each of the aforementioned trend frequency domain components to obtain each of the aforementioned trend components in the time domain; Subtracting the corresponding trend component from each of the load time series yields the corresponding seasonal component.

[0055] Preferably, in some embodiments of this application, the calculation formula for the trend frequency domain component is as follows: ; in, For the first The trend-related frequency domain components corresponding to the scale; It is the Fourier transform function; This is a low-pass filter function; This is the inverse Fourier transform function.

[0056] Preferably, in some embodiments of this application, the formula for calculating the seasonal component is: ,in, For the first The seasonal component corresponding to the scale.

[0057] This application achieves frequency domain decoupling of long-term load trends and periodic fluctuations by performing Fast Fourier Transform (FFT) on load time series at various scales and separating low-frequency and high-frequency components based on low-pass filtering thresholds. On the one hand, the trend components obtained by inverse transform of low-frequency components can stably characterize the overall evolution trend of the load, avoiding interference from short-term random disturbances on trend modeling; on the other hand, the seasonal components obtained through residual methods fully preserve the periodic fluctuation characteristics. Compared with direct time-domain decomposition methods, this frequency-domain filtering-based decomposition method has the advantages of high computational efficiency and clear component boundaries, effectively reducing the non-stationarity of the load series and providing more physically meaningful and discriminative input features for subsequent multi-scale feature fusion.

[0058] S103: Input the seasonal components and trend components at each scale into the pre-trained first multi-scale fusion temporal network, and perform bidirectional cross-scale feature interaction through the first multi-scale fusion temporal network to align and map the periodic patterns and evolution trends at different scales, thereby obtaining the joint features at each scale.

[0059] Furthermore, in some embodiments of this application, the step of performing bidirectional cross-scale feature interaction through the first multi-scale fusion temporal network to align and map the periodic patterns and evolution trends at different scales, and to obtain joint features at each scale, includes: Using the seasonal component at the lowest scale as the starting seasonal feature, the seasonal feature containing the load cycle pattern at the previous scale is upsampled and aligned according to the path from low to high scale to obtain the corresponding upsampled feature. The upsampled feature is then fused with the seasonal component at the current scale to obtain the seasonal feature at the current scale after being corrected by the long-cycle feature. Using the trend component at the highest scale as the starting trend feature, the trend feature containing the load evolution trend at the previous scale is downsampled and compressed according to the path from high to low scale to obtain the corresponding downsampled feature. The downsampled feature is then fused with the trend component at the current scale to obtain the trend feature at the current scale that incorporates instantaneous fluctuation information. The seasonal features and trend features at the same scale are fused across channels to obtain the joint features at each scale.

[0060] Preferably, in some embodiments of this application, the calculation formula for the seasonal characteristic includes: ; in, and The first Level and First Seasonal characteristics corresponding to the scale; It is a bottom-up fusion function that can be implemented by convolutional blocks or attention layers; For upsampling or scale matching operations; when hour, .

[0061] Preferably, in some embodiments of this application, the calculation formula for the trend feature includes: ; in, and The first Level and First Trend characteristics corresponding to the level scale; It is a top-down fusion function; This indicates a downsampling or dimensionality reduction operation.

[0062] Preferably, in some embodiments of this application, the calculation formula for the joint feature includes: ; in, For the first Joint features corresponding to the level scale; and The linear transformation weights are respectively for seasonality and trend characteristics; For scale-specific bias terms; This represents a non-linear activation function.

[0063] This application constructs a bidirectional, cross-scale feature interaction mechanism to achieve collaborative modeling of load cycle patterns and evolution trends across different time scales. Specifically, seasonal components are progressively upsampled and fused from low to high scales, allowing short-cycle fluctuations to be corrected against a long-cycle background; trend components are progressively downsampled and fused from high to low scales, enabling long-term evolution information to constrain the instability of short-term predictions. Furthermore, cross-channel fusion at the same scale achieves deep coupling of cycle and trend information. This bidirectional interactive structure overcomes the limitations of unidirectional transmission in traditional multi-scale models, significantly enhancing the consistency and complementarity of information across scales and improving the expressive power of joint features for complex load change patterns.

[0064] S104: Perform time-series reconstruction based on the joint features at each scale, and output the final short-term bus load forecast result.

[0065] Preferably, in some embodiments of this application, the step of performing time-series reconstruction based on the joint features at each scale and outputting the final short-term bus load prediction result includes: inputting the joint features corresponding to each scale into the corresponding predictor to obtain the prediction result of each predictor, and fusing the prediction results output by each predictor to obtain the final short-term bus load prediction result.

[0066] Preferably, in some embodiments of this application, the joint features corresponding to each scale are input into the corresponding predictor to obtain the prediction result of each predictor, including: the prediction result output by each predictor is... ,in, For the first The prediction results output by the predictor corresponding to the level scale; For the first Predictors corresponding to the scale; This is the set of network parameters for the predictor.

[0067] Preferably, in some embodiments of this application, the prediction results output by each predictor are fused to obtain the final short-term bus load prediction result, including: Each prediction result is weighted and fused using the following formula: ; in, This is the final short-term bus load forecast result; For the first The level scale corresponds to the fusion weights of the predictor. .

[0068] Furthermore, in some embodiments of this application, after outputting the final short-term bus load forecast result, the method further includes: Collect the actual measurement results corresponding to the short-term bus load forecast results, and calculate the error between the actual measurement results and the short-term bus load forecast results; Determine whether the error is within a preset range; If it is within the preset range, then update the model parameters of the first multi-scale fusion temporal network; If the network is not within the preset range, the training parameters of the first multi-scale fusion temporal network are updated using the differential evolution strategy, and the first multi-scale fusion temporal network is retrained based on the updated training parameters.

[0069] Preferably, in some embodiments of this application, the formula for calculating the error is as follows: ,in, for Time error, for Short-time bus load value at time t. for The short-term bus load forecast at any given time.

[0070] Preferably, in some embodiments of this application, determining whether the error is within a preset range includes: Calculate the error at each moment within the prediction period, calculate the mean and variance of the error, and set independent performance degradation thresholds for the mean and variance of the error respectively; The Kullback-Leibler divergence (KL divergence) is calculated using the following formula: And set a corresponding performance degradation threshold for KL divergence: ; in, The first in the original bus load data Probability distribution of the data items; For the previous time window corresponding to the original bus load data Probability distribution of the data items; This represents the probability distribution of various data points in the original bus load data. This represents the probability distribution of various data points within the previous time window relative to the original bus load data. This represents the number of data items in the bus load data. When either the mean or variance of the error is greater than the corresponding performance degradation threshold, and the KL divergence is less than or equal to the corresponding performance degradation threshold, the preset interval is determined to be satisfied. When the KL divergence is greater than the corresponding performance degradation threshold, it is determined that the preset interval is not met.

[0071] Preferably, in some embodiments of this application, the step of updating the model parameters of the first multi-scale fusion temporal network if it falls within the preset interval includes: The model parameters are updated using the following formula: ; in, These are the model parameters to be updated for the current first multi-scale fusion temporal network; These are the updated model parameters for the first multi-scale fusion temporal network. To fine-tune the learning rate; This is the training loss function for the local window (e.g., mean squared error, MSE). The formula represents the use of the latest bus load data. Perform a small number of gradient descent updates on the model to quickly correct short-term error shifts; This is the measured value of the busbar load.

[0072] Preferably, in some embodiments of this application, when the network is not within the preset range, the training parameters of the first multi-scale fusion temporal network are updated by the differential evolution strategy, and the first multi-scale fusion temporal network is retrained based on the updated training parameters, including: reselecting the optimal combination of training parameters in accordance with the above preferred embodiments, and updating the first multi-scale fusion temporal network with the optimal combination of training parameters.

[0073] By incorporating actual measurement feedback after the prediction output and dynamically adjusting model parameters or training parameters based on the prediction error, the prediction model possesses online adaptive optimization capabilities. When the prediction error is within a reasonable range, only fine-tuning of the model parameters is performed, which helps maintain model stability. When the error exceeds a threshold, parameter re-optimization based on a differential evolution strategy is triggered, preventing continuous performance degradation when load characteristics change. This closed-loop update mechanism enables the model to continuously evolve with changes in the load operating environment, improving the robustness and reliability of short-term bus load forecasting in long-term application scenarios.

[0074] To further illustrate the effectiveness of the short-term bus load forecasting method provided in the embodiments of this application, the following will be combined with... Figures 2 to 5 To provide further explanation.

[0075] refer to Figure 2 The paper compares the short-term bus load prediction results of the proposed method with those of typical deep learning models such as LSTM (Long Short-Term Memory) and TCN (Temporal Convolutional Network) within the same time period. As shown in the figure, the prediction curve of the proposed method closely matches the actual load curve, accurately tracking load change trends and effectively suppressing hysteresis bias at abrupt changes. In contrast, LSTM and TCN methods exhibit some prediction hysteresis or amplitude bias at peak-to-valley transitions, indicating that the proposed method has superior fitting ability and stability in capturing the time-varying characteristics and nonlinear laws of bus load.

[0076] refer to Figure 3 This paper presents a comparison of the main error metrics of the proposed method with those of LSTM, GRU (Gated Recurrent Unit), TCN, and Transformer models in short-term bus load forecasting tasks, including root mean square error (RMSE) and mean absolute percentage error (MAPE). The results show that the RMSE and MAPE values ​​of the proposed method are significantly lower than those of other baseline models. Specifically, the RMSE is reduced by approximately 20% compared to the traditional LSTM model, and the MAPE is reduced by approximately 25%, indicating a significant advantage in prediction accuracy. These results validate the effectiveness of the competitive optimization strategy based on ensemble differential evolution in parameter optimization and model robustness improvement.

[0077] refer to Figure 4The convergence curve of the fitness function of the competitive optimization algorithm based on ensemble differential evolution in this application during the model training phase is shown. It can be seen that as the number of evolutionary iterations increases, the fitness value gradually rises and stabilizes after approximately the 25th generation, indicating that the model parameters gradually approach the global optimum during the competitive optimization process. The curve converges smoothly without significant oscillations, demonstrating that the competitive optimization mechanism proposed in this application maintains high stability and global search capability while maintaining convergence speed, providing a reliable automated hyperparameter tuning method for model training.

[0078] Figure 5 The paper presents the prediction error distribution of the proposed method compared to models such as LSTM, GRU, and Transformer on the test set. As shown in the box plots, the proposed method exhibits the lowest median error and the most concentrated distribution, with a significantly smaller error fluctuation range compared to other models, indicating more stable prediction results and fewer outliers. In contrast, the error distribution of traditional models is more dispersed, with some samples showing significant bias. This application achieves error suppression and enhanced stability through multi-scale fusion and competitive optimization, further demonstrating the robustness and reliability of the method under complex load fluctuation scenarios.

[0079] In summary, the short-term bus load forecasting method provided in this application has the following advantages over existing technologies: First, this application utilizes a competitive optimization method based on an integrated differential evolution strategy to adaptively determine the number of scale segments matching the load fluctuation pattern, ensuring that multi-scale time series segmentation conforms to actual load change patterns and avoiding feature loss or redundancy caused by inappropriate scale selection. Subsequently, the load time series at each scale is decomposed into seasonal and trend components, achieving decoupling of load cycle patterns and long-term evolution trends, reducing the impact of non-stationarity on forecast accuracy. Based on this, a multi-scale fusion time series network is used to perform bidirectional cross-scale feature interaction, enabling short-term fluctuation information and long-term trend information to mutually correct and supplement each other at different scales, forming more consistent and discriminative joint features. Finally, time series reconstruction is performed based on the joint features, resulting in forecasts that simultaneously consider local changes and overall trends, thereby overcoming the problem of large prediction deviations in complex load scenarios in existing technologies and significantly improving the accuracy of short-term bus load forecasting.

[0080] like Figure 6 As shown, based on the above method embodiments, an embodiment of this application provides a short-term bus load prediction device, including: a data processing module 201, a sequence decomposition module 202, a feature processing module 203, and a prediction module 204. The data processing module 201 is used to collect the original bus load data of the period to be predicted, and to perform multi-scale time series segmentation on the original bus load data according to the preset number of scale segments to obtain several load time series that characterize the load evolution characteristics at different scales. The number of scale segments is determined by a competitive optimization method that integrates differential evolution strategies. The sequence decomposition module 202 is used to decompose each load time series into a seasonal component that characterizes the periodicity of the load and a trend component that characterizes the long-term evolution trend of the load. The feature processing module 203 is used to input the seasonal components and trend components at each scale into a pre-trained first multi-scale fusion temporal network, and perform bidirectional cross-scale feature interaction through the first multi-scale fusion temporal network to align and map the periodic patterns and evolution trends at different scales, thereby obtaining joint features at each scale. The prediction module 204 is used to perform time-series reconstruction based on the joint features at each scale and output the final short-term bus load prediction result.

[0081] Further, in some embodiments of this application, the sequence decomposition module 202 includes: a Fourier transform unit, a filtering unit, an inverse transform unit, and a calculation unit; the sequence decomposition module 202 is used to decompose each load time series into a seasonal component characterizing the periodicity of the load and a trend component characterizing the long-term evolution trend of the load, including: The Fourier transform unit is used to perform a fast Fourier transform on each of the load time series to obtain the corresponding frequency domain sequence. The filtering unit is used to retain the low-frequency components and suppress the high-frequency components in each frequency domain sequence according to a preset low-pass filtering threshold, so as to obtain the corresponding trend frequency domain components. The inverse transform unit is used to perform an inverse Fourier transform on each of the trend frequency domain components to obtain each of the trend components in the time domain. The calculation unit is used to subtract the corresponding trend component from each of the load time series to obtain the corresponding seasonal component.

[0082] Further, in some embodiments of this application, the feature processing module 203 includes: a first fusion unit, a second fusion unit, and a third fusion unit; the feature processing module 203 is used to perform bidirectional cross-scale feature interaction through the first multi-scale fusion temporal network, so as to align and map the periodic patterns and evolution trends at different scales, and obtain joint features at each scale, including: The first fusion unit is used to take the seasonal component at the lowest scale as the starting seasonal feature, and upsample and align the seasonal feature containing the load cycle pattern at the previous scale according to the path from low to high scale to obtain the corresponding upsampled feature. The upsampled feature is then fused with the seasonal component at the current scale to obtain the seasonal feature at the current scale after being corrected by the long-cycle feature. The second fusion unit is used to take the trend component at the highest scale as the starting trend feature, and downsample and compress the trend feature containing the load evolution trend at the previous scale according to the path from high to low scale to obtain the corresponding downsampled feature. The downsampled feature is then fused with the trend component at the current scale to obtain the trend feature at the current scale that incorporates instantaneous fluctuation information. The third fusion unit is used to perform cross-channel fusion of the seasonal features and the trend features at the same scale to obtain the joint features at each scale.

[0083] Further, in some embodiments of this application, the data processing module 201 includes: an operator generation unit, a sampling unit, and a normalization unit; the data processing module 201 is used to perform multi-scale time-series segmentation on the original bus load data according to a preset number of scale segments to obtain several load time series characterizing the load evolution characteristics at different scales, including: The operator generation unit is used to generate sampling operators for the number of scale segments according to preset rules; wherein each sampling operator corresponds to a scale; The sampling unit is used to resample the original bus load data using each of the sampling operators to generate multiple load subsequences with different sampling frequencies. The normalization unit is used to truncate each of the load subsequences through a sliding window and perform dimensional normalization on each truncated sequence to obtain multiple load time series that are correlated in the time dimension and have different scales.

[0084] Furthermore, in some embodiments of this application, the number of scale segments is determined by a competitive optimization method integrating differential evolution strategies, including: Construct an initial parameter combination population containing multiple different combinations of first training parameters; wherein, the first training parameter combination includes the scale segmentation number; The initial parameter combination population is iteratively updated using the differential evolution strategy until the preset iteration termination condition is met, resulting in the final parameter combination population. Train an untrained second multi-scale fusion temporal network according to each second training parameter combination in the final parameter combination population, and obtain the first training result corresponding to each second training parameter combination. The number of scale segments is determined based on the combination of second training parameters corresponding to the optimal first training result.

[0085] Furthermore, in some embodiments of this application, the iterative update of the initial parameter combination population using the differential evolution strategy includes: During each iteration update, several third training parameter combinations are selected from the current parameter combination population to generate mutation combinations; By performing cross-operations between the mutation combination and each of the third training parameter combinations in the current parameter combination population, several experimental combinations can be obtained. The second multi-scale fusion temporal network is trained by each of the aforementioned experimental combinations to obtain the second training result corresponding to each of the aforementioned experimental combinations. The second multi-scale fusion temporal network is trained by each of the third training parameter combinations to obtain the third training result corresponding to each of the third training parameter combinations. Based on the second training result and the third training result, each of the experimental combinations competes with each of the third training parameter combinations, and the losing third training parameter combination is replaced with the winning experimental combination.

[0086] Furthermore, in some embodiments of this application, after outputting the final short-term bus load forecast result, the method further includes: Collect the actual measurement results corresponding to the short-term bus load forecast results, and calculate the error between the actual measurement results and the short-term bus load forecast results; Determine whether the error is within a preset range; If it is within the preset range, then update the model parameters of the first multi-scale fusion temporal network; If the network is not within the preset range, the training parameters of the first multi-scale fusion temporal network are updated using the differential evolution strategy, and the first multi-scale fusion temporal network is retrained based on the updated training parameters.

[0087] In summary, the short-term bus load forecasting device provided in this application has the following advantages over existing technologies: Firstly, this application utilizes a competitive optimization method based on an integrated differential evolution strategy to adaptively determine the number of scale segments matching the load fluctuation pattern, enabling multi-scale time series segmentation to conform to actual load change patterns and avoiding feature loss or redundancy caused by improper scale selection. Subsequently, the load time series at each scale is decomposed into seasonal and trend components, achieving decoupling of load cycle patterns and long-term evolution trends, reducing the impact of non-stationarity on forecast accuracy. Based on this, a multi-scale fusion time series network performs bidirectional cross-scale feature interaction, allowing short-term fluctuation information and long-term trend information to mutually correct and supplement each other at different scales, forming more consistent and discriminative joint features. Finally, time series reconstruction is performed based on the joint features, resulting in forecasts that simultaneously consider local changes and overall trends, thereby overcoming the problem of large prediction deviations in complex load scenarios in existing technologies and significantly improving the accuracy of short-term bus load forecasting.

[0088] It is understood that the above-described device embodiments correspond to the method embodiments of this application, and can implement the short-term bus load forecasting method provided by any of the above-described method embodiments of this application.

[0089] It should be noted that the device embodiments described above are merely illustrative, and some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided in this application, the connection relationships between modules indicate that they have communication connections, which can specifically be implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0090] Based on the above embodiments of the short-term bus load forecasting method, another embodiment of this application provides a terminal device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the short-term bus load forecasting method of any embodiment of this application.

[0091] For example, in this embodiment, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete this application. The one or more module units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the terminal device.

[0092] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0093] The processor 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. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0094] Based on the above-described method embodiments, another embodiment of this application provides a computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the short-term bus load forecasting method described in any of the above-described method embodiments of this application.

[0095] The modules / units integrated in the device / terminal equipment, if implemented as software functional units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the 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.

Claims

1. A short-term bus load forecasting method, characterized in that, include: The original bus load data for the period to be predicted is collected, and the original bus load data is divided into multiple scale time series according to the preset number of scale segments to obtain several load time series that characterize the load evolution characteristics at different scales. The number of scale segments is determined by a competitive optimization method that integrates differential evolution strategies. Each load time series is decomposed into a seasonal component characterizing the periodicity of the load and a trend component characterizing the long-term evolution of the load. The seasonal and trend components at each scale are input into a pre-trained first multi-scale fusion temporal network. The first multi-scale fusion temporal network performs bidirectional cross-scale feature interaction to align and map the periodic patterns and evolution trends at different scales, thereby obtaining joint features at each scale. Based on the joint features at each scale, time series reconstruction is performed to output the final short-term bus load forecast results.

2. The short-term bus load forecasting method as described in claim 1, characterized in that, The process of decomposing each load time series into a seasonal component characterizing the periodicity of the load and a trend component characterizing the long-term evolution of the load includes: Perform a Fast Fourier Transform on each of the aforementioned load time series to obtain the corresponding frequency domain sequence; Based on a preset low-pass filtering threshold, the low-frequency components in each frequency domain sequence are retained while the high-frequency components are suppressed to obtain the corresponding trend frequency domain components. Perform an inverse Fourier transform on each of the aforementioned trend frequency domain components to obtain each of the aforementioned trend components in the time domain; Subtracting the corresponding trend component from each of the load time series yields the corresponding seasonal component.

3. The short-term bus load forecasting method as described in claim 1, characterized in that, The step of performing bidirectional cross-scale feature interaction through the first multi-scale fusion temporal network to align and map the periodic patterns and evolution trends at different scales, and obtaining joint features at each scale, includes: Using the seasonal component at the lowest scale as the starting seasonal feature, the seasonal feature containing the load cycle pattern at the previous scale is upsampled and aligned according to the path from low to high scale to obtain the corresponding upsampled feature. The upsampled feature is then fused with the seasonal component at the current scale to obtain the seasonal feature at the current scale after being corrected by the long-cycle feature. Using the trend component at the highest scale as the starting trend feature, the trend feature containing the load evolution trend at the previous scale is downsampled and compressed according to the path from high to low scale to obtain the corresponding downsampled feature. The downsampled feature is then fused with the trend component at the current scale to obtain the trend feature at the current scale that incorporates instantaneous fluctuation information. The seasonal features and trend features at the same scale are fused across channels to obtain the joint features at each scale.

4. The short-term bus load forecasting method as described in claim 1, characterized in that, The original bus load data is divided into multiple time series segments according to a preset number of scales to obtain several load time series that characterize the load evolution characteristics at different scales, including: A sampling operator for the number of scale segments is generated by using preset rules; wherein each sampling operator corresponds to a scale. The original bus load data is resampled using each of the sampling operators to generate multiple load subsequences with different sampling frequencies; Each load subsequence is truncated by a sliding window, and the truncated sequences are normalized in dimension to obtain multiple load time series that are interconnected in the time dimension and have different scales.

5. The short-term bus load forecasting method as described in claim 1, characterized in that, The number of scale segments is determined through a competitive optimization method integrating differential evolution strategies, including: Construct an initial parameter combination population containing multiple different combinations of first training parameters; wherein, the first training parameter combination includes the scale segmentation number; The initial parameter combination population is iteratively updated using the differential evolution strategy until the preset iteration termination condition is met, resulting in the final parameter combination population. Train an untrained second multi-scale fusion temporal network according to each second training parameter combination in the final parameter combination population, and obtain the first training result corresponding to each second training parameter combination. The number of scale segments is determined based on the combination of second training parameters corresponding to the optimal first training result.

6. The short-term bus load forecasting method as described in claim 5, characterized in that, The iterative update of the initial parameter combination population using the differential evolution strategy includes: During each iteration update, several third training parameter combinations are selected from the current parameter combination population to generate mutant combinations; By performing cross-operations between the mutation combination and each of the third training parameter combinations in the current parameter combination population, several experimental combinations can be obtained; The second multi-scale fusion temporal network is trained by each of the aforementioned experimental combinations to obtain the second training result corresponding to each of the aforementioned experimental combinations. The second multi-scale fusion temporal network is trained by each of the third training parameter combinations to obtain the third training result corresponding to each of the third training parameter combinations. Based on the second training result and the third training result, each of the experimental combinations competes with each of the third training parameter combinations, and the losing third training parameter combination is replaced with the winning experimental combination.

7. The short-term bus load forecasting method as described in claim 1, characterized in that, After outputting the final short-term bus load forecast result, the system also includes: Collect the actual measurement results corresponding to the short-term bus load forecast results, and calculate the error between the actual measurement results and the short-term bus load forecast results; Determine whether the error is within a preset range; If it is within the preset range, then update the model parameters of the first multi-scale fusion temporal network; If the network is not within the preset range, the training parameters of the first multi-scale fusion temporal network are updated using the differential evolution strategy, and the first multi-scale fusion temporal network is retrained based on the updated training parameters.

8. A short-term bus load forecasting device, characterized in that, include: Data processing module, sequence decomposition module, feature processing module, and prediction module; The data processing module is used to collect the original bus load data of the period to be predicted, and to perform multi-scale time series segmentation on the original bus load data according to the preset number of scale segments to obtain several load time series that characterize the load evolution characteristics at different scales. The number of scale segments is determined by a competitive optimization method that integrates differential evolution strategies. The sequence decomposition module is used to decompose each load time series into a seasonal component that characterizes the periodicity of the load and a trend component that characterizes the long-term evolution trend of the load. The feature processing module is used to input the seasonal components and trend components at each scale into a pre-trained first multi-scale fusion temporal network, and perform bidirectional cross-scale feature interaction through the first multi-scale fusion temporal network to align and map the periodic patterns and evolution trends at different scales, thereby obtaining joint features at each scale. The prediction module is used to perform time-series reconstruction based on the joint features at each scale and output the final short-term bus load prediction result.

9. A terminal device, characterized in that, The system includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements a short-term bus load forecasting 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 includes a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform a short-term bus load forecasting method as described in any one of claims 1 to 7.

Images