A power load prediction method based on multi-scale decomposition guidance and token collaborative modeling

Abstract

The present application relates to a kind of power load prediction method based on multiscale decomposition guide and token collaborative modeling, belong to power load prediction technical field.The method includes: obtaining historical power load data, constructs multivariate historical power load time series data and carries out reversible normalization processing, time-frequency decomposition and multiscale enhancement, obtains enhanced load seasonal representation and carries out token feature enhancement and collaborative feature extraction, obtains load seasonal prediction feature, to generate load seasonal prediction result and load trend prediction result and carry out fusion and residual correction, obtain normalized load prediction result and carry out anti-normalization processing and continuity correction, obtain the power load prediction result of future multiple time steps.The present application is aimed at solving the technical problems that the existing technology has difficulty in effectively extracting multiscale short-term disturbance, the correlation between different load variables is not modeled enough, and the prediction result in long prediction interval is not continuous with the historical load sequence boundary.

Description

Technical Field

[0001] This invention relates to a power load forecasting method based on multi-scale decomposition guidance and token collaborative modeling, belonging to the field of power load forecasting technology. Background Technology

[0002] Power load forecasting is a crucial technology in power system operation and dispatching, generation planning, distribution network operation control, and electricity demand management. Accurately predicting power load changes over a future period helps the power system formulate dispatching strategies in advance, reduce the risk of supply-demand imbalances, and improve the safety and economy of grid operation. With the widespread deployment of smart meters, load monitoring terminals, and distribution automation equipment, the power system can collect continuous historical power load data from multiple load monitoring nodes, power supply areas, or electrical equipment. This data is typically presented as a multivariate time series, with different variables corresponding to load changes at different monitoring nodes, power supply areas, or electrical equipment.

[0003] In real-world power load forecasting scenarios, historical power load data exhibits significant non-stationarity and multi-scale variation characteristics. On one hand, power load is influenced by factors such as residential electricity consumption habits, industrial production activities, commercial operation patterns, weather conditions, holiday schedules, and unforeseen events, leading to substantial differences in the load mean, variance, and distribution across different time periods. On the other hand, power load sequences typically contain multiple components simultaneously, including long-term trends, daily or weekly fluctuations, and localized short-term disturbances. These components are coupled with each other, increasing the difficulty of accurately extracting load variation patterns from the model. Furthermore, there are often correlations in electricity consumption behavior and load transfer relationships among multiple load monitoring nodes, power supply areas, or electrical equipment. If only a single node or device is modeled independently, the mutual influence between different load variables may be overlooked; conversely, if all load variables are modeled uniformly, they may be subject to interference from redundant variables and noise.

[0004] Existing power load forecasting methods typically include statistical forecasting methods, machine learning methods, and deep learning methods. Traditional statistical methods often rely on linear or fixed-period assumptions, making it difficult to adequately adapt to the nonlinear variations and non-stationary distributions in complex power load data. Deep learning methods based on recurrent neural networks, convolutional neural networks, Transformer structures, or linear models have improved load sequence modeling capabilities to some extent, but still have the following shortcomings: First, some methods fail to adequately distinguish between trend components and periodic fluctuation components in power load, leading to interference between long-term variation information and local load fluctuation information; second, some methods do not adequately utilize load detail fluctuations at different time scales, making it difficult to effectively characterize multi-scale changes such as daily cycles, weekly cycles, and short-term disturbances; third, some methods do not adequately model the correlations between multiple load monitoring nodes, power supply areas, or electrical equipment, affecting the multivariate load forecasting effect; fourth, some methods lack frequency domain-level load cycle feature enhancement mechanisms, making it difficult to effectively extract frequency response information from power load; fifth, within a long forecast interval, there may be boundary discontinuities between the starting position of the forecast sequence and the end of the historical load sequence, thus affecting the stability of the power load forecasting results.

[0005] Therefore, it is necessary to propose a forecasting method for power load forecasting scenarios, which can handle non-stationary distribution, multi-scale periodic fluctuations, long-term trend changes, correlations between load variables, frequency domain response characteristics, and discontinuities in forecasting boundaries for historical power load data from multiple nodes, regions, or devices, thereby improving the accuracy and stability of power load forecasting for multiple future time steps. Summary of the Invention

[0006] The purpose of this invention is to provide a power load forecasting method based on multi-scale decomposition guidance and token collaborative modeling. This method aims to solve the technical problems of existing power load forecasting methods when processing historical power load data from multiple load monitoring nodes, power supply areas, or electrical equipment. These problems include significant non-stationary distribution, coupling of long-term trends with periodic load fluctuations, difficulty in effectively extracting multi-scale short-term disturbances, insufficient modeling of correlations between different load variables, and discontinuity between forecast results and historical load sequence boundaries within long forecast intervals. Consequently, these methods result in insufficient accuracy and stability of power load forecasts for multiple future time steps.

[0007] To achieve the above objectives, the technical solution of the present invention is: a power load forecasting method based on multi-scale decomposition guidance and token collaborative modeling, comprising the following steps:

[0008] Step 1: Acquire historical power load data of multiple load monitoring nodes, power supply areas, and electrical equipment in the target power system over continuous time to construct multivariate historical power load time series data and perform reversible normalization processing to obtain the normalized power load input sequence. Then, perform time-frequency decomposition to obtain load trend components, load seasonality components, and multi-scale high-frequency detail components. Generate adaptive weights based on the energy characteristics of the multi-scale high-frequency detail components, and perform multi-scale enhancement on the load seasonality components based on the adaptive weights to obtain the enhanced load seasonality representation.

[0009] Step 2: Divide the enhanced load seasonality representation into multiple local time segments and map the local time segments into load seasonality token representations; perform token feature enhancement on the load seasonality token representations, extract short-term load fluctuation features between adjacent local time segments, obtain the enhanced load seasonality token representations, and perform time-dependent modeling, channel-dependent modeling, cross-variable mixing, and frequency domain token fusion to obtain load seasonality prediction features;

[0010] Step 3: Generate a seasonal load forecast result based on the seasonal load forecast characteristics, and generate a load trend forecast result based on the load trend component; perform adaptive fusion on the seasonal load forecast result and the load trend forecast result to obtain a basic load forecast result; generate a residual correction term based on the energy ratio of the load trend component and the seasonal load component, and superimpose the residual correction term onto the basic load forecast result to obtain a normalized load forecast result;

[0011] Step 4: Perform inverse normalization on the normalized load forecast result to obtain the initial power load forecast result; perform continuity correction on the starting part of the initial power load forecast result based on the local anchor point at the end of the historical power load input sequence to obtain the power load forecast result for multiple future time steps.

[0012] Optionally, Step 1 specifically includes:

[0013] Step 1.1: Define multivariate historical electricity load time series data as ,in, Indicates the sample batch size. Indicates the length of historical power load input. This indicates the number of load monitoring nodes, power supply areas, or electrical equipment. Each element in the table represents the power load value of the corresponding node, user, or device at the corresponding historical moment;

[0014] The sample mean and sample standard deviation are calculated for the multivariate historical electricity load time series data along the time dimension. Reversible normalization is then performed based on the sample mean and sample standard deviation, expressed as:

[0015]

[0016] In the formula, This represents the normalized power load input sequence. This represents the sample mean of the historical electricity load input sequence over the time dimension. This represents the sample standard deviation of the historical power load input sequence over the time dimension;

[0017] Step 1.2: Perform time-frequency decomposition on the normalized power load input sequence to obtain low-frequency approximate components and high-frequency detail components at multiple scales, expressed as:

[0018]

[0019] In the formula, This represents the time-frequency decomposition operator acting on the normalized power load input sequence. Represents the low-frequency approximate component. Indicates the first High-frequency detail components at each scale Indicates the number of decomposition scales;

[0020] The high-frequency detail components at each scale are set to zero, and an inverse transform is performed based on the low-frequency approximation components to obtain the load trend components, expressed as follows:

[0021]

[0022] In the formula, This indicates the relationship with the time-frequency decomposition operator. The corresponding inverse transform operator, Indicates load trend components, This represents the set of zero components corresponding to the dimensions of high-frequency detail components at each scale; the load seasonality component is obtained based on the difference between the normalized power load input sequence and the load trend component. ;

[0023] Step 1.3: Perform energy calculations on the high-frequency detail components at the multiple scales, and generate adaptive weights based on the energy characteristics of the high-frequency detail components at different scales. The expression is as follows:

[0024]

[0025]

[0026] In the formula, This represents the operation of averaging the elements of the tensor within the parentheses. Indicates the first Energy values ​​of high-frequency detail components at each scale; Indicates the first Energy values ​​of high-frequency detail components at each scale; Indicates the first Adaptive weights corresponding to each scale; Indicates the summation index. ; A constant that prevents division by zero;

[0027] The high-frequency detail components at different scales are length-aligned and then weighted and fused according to the adaptive weights to obtain a multi-scale load seasonal refinement representation. The expression is:

[0028]

[0029] In the formula, Indicates the length alignment operator;

[0030] The multi-scale load seasonality refinement representation is fused with the load seasonality component to obtain the enhanced load seasonality representation. ,in, This represents the fusion coefficient.

[0031] Optionally, Step 2 specifically includes:

[0032] Step 2.1: Divide the obtained enhanced load seasonality representation into multiple local load time segments, and map the local load time segments into load seasonality token representations. , among which, This represents the local load time segment embedding mapping. Indicates periodic position encoding;

[0033] Step 2.2: Perform token feature enhancement on the load seasonality token representation. Extract short-term fluctuation features between adjacent load time segments through local convolution of the token dimension, and extract load feature interaction information within the token through feedforward transformation of the feature dimension to obtain the enhanced load token representation. , among which, This indicates a token feature enhancement mapping. Indicates the learnable gating coefficient;

[0034] Step 2.3: Perform channel-dependent modeling on the enhanced load token representation. Select the load channel patterns to participate in the modeling based on the energy characteristics of the corresponding channels of different load monitoring nodes, power supply areas, or electrical equipment, and perform attention modeling on the selected load channel patterns to obtain the channel-enhanced load token representation. ,in, This represents a channel dependency modeling mapping;

[0035] Step 2.4: Perform time-dependent modeling on the channel-enhanced load token representation. Select the time patterns to participate in the modeling based on the low-frequency representation or mode selection results in the load time dimension, and perform attention modeling on the selected time patterns to obtain the time-enhanced load token representation. ,in, This represents a time-dependent modeling mapping;

[0036] Step 2.5: Perform cross-variable mixing on the time-enhanced load token representation, interacting information between different load variables at the variable dimensions corresponding to load monitoring nodes, power supply areas, or electrical equipment to obtain the cross-variable mixed load token representation. ,in, Represents a cross-variable mixed mapping;

[0037] Step 2.6: Perform frequency domain token enhancement on the cross-variable hybrid load token representation, extracting frequency domain response features in the time and variable channel dimensions of the load token to obtain the frequency domain enhanced load token representation. ,in, This indicates a frequency domain token enhancement mapping. Indicates the frequency domain enhancement gating coefficient;

[0038] Step 2.7: Generate a fusion gating based on the periodic characteristics of the power load input sequence and the energy ratio of the seasonal load component to the trend load component. Then, dynamically fuse the intervariate mixed load token representation and the frequency-domain enhanced load token representation based on the fusion gating to obtain the seasonal load prediction features. The expression is:

[0039]

[0040] In the formula, Represents dynamic feature fusion mapping. This represents a fusion gating system generated from the periodic characteristics of electricity load and the energy ratio of the seasonal component to the trend component of the load. This represents the dynamic fusion gating coefficient.

[0041] Optionally, Step 3 specifically includes:

[0042] Step 3.1: The load seasonality prediction result is jointly generated by the flattening mapping prediction unit and the token mixing prediction unit. The flattening mapping prediction unit is used to directly map the load seasonality prediction features to the prediction length. The token mixing prediction unit is used to perform hybrid modeling of the token relationships between different local load time segments and the internal feature relationships of the tokens. The outputs of the flattening mapping prediction unit and the token mixing prediction unit are fused through learnable mixing coefficients to obtain the load seasonality prediction result. The expression is:

[0043]

[0044] In the formula, Represents the flattening mapping prediction unit. Indicates a token-mixed prediction unit. Represents the learnable mixing coefficients;

[0045] Step 3.2: The load trend prediction result maps the load trend component on the historical input length to the prediction length through a trend mapping branch, thus obtaining the load trend prediction result within the future prediction interval. ,in, Indicates the trend mapping branch;

[0046] Step 3.3: Adaptively fuse the seasonal load forecast results and the trend load forecast results to obtain the basic load forecast results. ,in, This represents the fusion weights corresponding to the seasonal load forecast results. This indicates the fusion weights corresponding to the load trend forecast results;

[0047] Step 3.4: Calculate the load trend energy ratio based on the load trend component and the load seasonality component, and generate the residual prediction term based on the load trend energy ratio and the normalized power load input sequence. The expression is as follows:

[0048]

[0049]

[0050] In the formula, Indicates the energy ratio of load trend. This indicates an operation that calculates the variance of the tensor within the parentheses along the time dimension. This indicates the learnable residual gain. Indicates the residual prediction branch, Represents the residual prediction term;

[0051] Step 3.5: Superimpose the residual prediction term onto the basic load prediction result to obtain the normalized load prediction result. ,in, This indicates an energy-constrained superposition operation.

[0052] Optionally, Step 4 specifically includes:

[0053] Step 4.1: Using the sample mean and sample standard deviation saved during the reversible normalization process, perform inverse normalization on the obtained normalized load forecast results to obtain the initial power load forecast results. ;

[0054] Step 4.2: Calculate the local anchor point value based on one or more time steps at the end of the historical power load input sequence. The expression is:

[0055]

[0056] In the formula, Indicates the local anchor value. This indicates the length of the historical window used to calculate local anchor point values. Indicates the first The power load values ​​corresponding to all samples and all load monitoring nodes, power supply areas or electrical equipment at each historical moment;

[0057] Step 4.3: Calculate the prediction starting offset based on the local anchor point value and the starting position of the initial power load prediction result. ,in, This represents the predicted load value corresponding to all samples and all load monitoring nodes, power supply areas, or electrical equipment at the first prediction time step of the initial power load prediction results.

[0058] Step 4.4: Based on the predicted initial offset, perform continuity correction on the initial portion of the initial power load prediction result to obtain the final power load prediction result, expressed as:

[0059]

[0060] In the formula, This indicates the final power load forecast result in the [number]th [year]. The predicted load values ​​for all samples and all load monitoring nodes, power supply areas, or electrical equipment at each prediction time step. This indicates the initial power load forecast result at the [number]th [year]. The predicted load value corresponding to each prediction time step. Indicates the prediction step size index. Indicates the predicted length. Indicates the first The continuity correction coefficient corresponding to each prediction time step.

[0061] The beneficial effects of this invention are:

[0062] 1. This invention reduces the non-stationary distribution differences between different load objects and different time intervals by reversibly normalizing the historical power load data of multiple load monitoring nodes, power supply areas or electrical equipment, so that the subsequent model can learn the power load change pattern more stably.

[0063] 2. This invention decomposes the power load input sequence into load trend components and load seasonality components through time-frequency decomposition, reducing the mutual interference between long-term load change trends and periodic load fluctuations such as daily and weekly cycles, and improving the pertinence of power load trend prediction and periodic fluctuation prediction.

[0064] 3. This invention enhances the ability to express short-term disturbances, local fluctuations and periodic changes in load at different time scales by using energy adaptive weighting of multi-scale high-frequency detail components, which is beneficial to improving the load prediction accuracy in complex power consumption scenarios.

[0065] 4. This invention improves the collaborative modeling capability of multivariable power load data by tokenizing local load time segments, time-dependent modeling, and channel-dependent modeling, while simultaneously capturing the load change relationship over continuous time and the correlation between different load monitoring nodes, power supply areas, or electrical equipment.

[0066] 5. This invention further extracts the periodicity and frequency response features in power load data through a frequency domain token fusion mechanism, enabling the model to better adapt to daily, weekly, and other electricity consumption cycle changes.

[0067] 6. This invention enables the forecasting process to dynamically adjust the forecasting basis according to the trend and periodic characteristics of different power load input samples by adaptively fusing the seasonal load forecasting results with the load trend forecasting results, thereby improving the adaptability of the load forecasting results for multiple future time steps.

[0068] 7. This invention reduces local prediction bias and boundary discontinuities between the end of the historical power load sequence and the beginning of the prediction sequence by residual correction and continuity correction, thereby improving the accuracy and stability of long-term power load prediction results. Attached Figure Description

[0069] Figure 1 This is a flowchart of the steps of the present invention. Detailed Implementation

[0070] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0071] Example 1: As Figure 1 As shown, a power load forecasting method based on multi-scale decomposition guidance and token collaborative modeling includes the following steps:

[0072] Step 1: Acquire historical power load data of multiple load monitoring nodes, power supply areas, and electrical equipment in the target power system over continuous time to construct multivariate historical power load time series data and perform reversible normalization processing to obtain the normalized power load input sequence. Then, perform time-frequency decomposition to obtain load trend components, load seasonality components, and multi-scale high-frequency detail components. Generate adaptive weights based on the energy characteristics of the multi-scale high-frequency detail components, and perform multi-scale enhancement on the load seasonality components based on the adaptive weights to obtain the enhanced load seasonality representation.

[0073] Step 1.1: Obtain historical power load data over continuous time from multiple load monitoring nodes, multiple power supply areas, or multiple electrical devices in the target power system, and construct multivariate historical power load time series data in chronological order. Define multivariate historical power load time series data as... ,in, Indicates the sample batch size. Indicates the length of historical power load input. This indicates the number of load monitoring nodes, power supply areas, or electrical equipment. Each element in the table represents the power load value of the corresponding node, user, or device at the corresponding historical moment;

[0074] The sample mean and sample standard deviation of the multivariate historical electricity load time series data are calculated along the time dimension, as expressed by:

[0075]

[0076]

[0077] In the formula, This represents the sample mean of the historical electricity load input sequence over the time dimension. This represents the sample standard deviation of the historical electricity load input sequence over the time dimension. This represents a constant that prevents division by zero. Indicates the first The power load values ​​corresponding to all samples and all load monitoring nodes, power supply areas or electrical equipment at each historical moment;

[0078] Reversible normalization is performed based on the sample mean and sample standard deviation to reduce the distributional differences between different samples. The expression is as follows:

[0079]

[0080] In the formula, This represents the normalized power load input sequence;

[0081] Step 1.2: Perform time-frequency decomposition on the normalized power load input sequence to separate the long-term trend, periodic fluctuations, and local disturbances in the power load data, obtaining low-frequency approximate components and high-frequency detail components at multiple scales. The expression is as follows:

[0082]

[0083] In the formula, This represents the time-frequency decomposition operator acting on the normalized power load input sequence. This represents the low-frequency approximate component, used to characterize the long-term variation trend of electricity load. Indicates the first High-frequency detail components at each scale are used to characterize the electrical load at the 1st... Periodic fluctuations or local disturbances on a time scale Indicates the number of decomposition scales;

[0084] The high-frequency detail components at each scale are set to zero, and an inverse transform is performed based on the low-frequency approximation components to obtain the load trend components, expressed as follows:

[0085]

[0086] In the formula, This indicates the relationship with the time-frequency decomposition operator. The corresponding inverse transform operator, Indicates load trend components, This represents the set of zero components corresponding to the dimensions of high-frequency detail components at each scale; the load seasonality component is obtained based on the difference between the normalized power load input sequence and the load trend component. , Used to characterize periodic fluctuations and local variations in electrical load;

[0087] Step 1.3: Perform energy calculations on the high-frequency detail components at the multiple scales, and generate adaptive weights based on the energy characteristics of the high-frequency detail components at different scales. The expression is as follows:

[0088]

[0089]

[0090] In the formula, This represents the operation of averaging the elements of the tensor within the parentheses. Indicates the first The energy value of each high-frequency detail component at each scale is used to measure the intensity of power load fluctuation information at that scale. Indicates the first Energy values ​​of high-frequency detail components at each scale; Indicates the first Adaptive weights corresponding to each scale; Indicates the summation index. ; A constant that prevents division by zero;

[0091] The high-frequency detail components at different scales are length-aligned and then weighted and fused according to the adaptive weights to obtain a multi-scale load seasonal refinement representation. The expression is:

[0092]

[0093] In the formula, The length alignment operator is used to align high-frequency detail components of different scales to the same time length as the normalized power load input sequence. In this embodiment, the length alignment operator includes at least one of upsampling, interpolation, or truncation.

[0094] The multi-scale load seasonality refinement representation is fused with the load seasonality component to obtain the enhanced load seasonality representation. ,in, This represents the fusion coefficient.

[0095] Step 2: Divide the enhanced load seasonality representation into multiple local time segments and map the local time segments into load seasonality token representations; perform token feature enhancement on the load seasonality token representations, extract short-term load fluctuation features between adjacent local time segments, obtain the enhanced load seasonality token representations, and perform time-dependent modeling, channel-dependent modeling, cross-variable mixing, and frequency domain token fusion to obtain load seasonality prediction features;

[0096] Step 2.1: Divide the obtained enhanced load seasonality representation into multiple local load time segments according to a preset segment length and sliding step size, and map the local load time segments into load seasonality token representations. ,in, This represents a local load time segment embedding map, used to convert load segments in continuous time into token features. This indicates a periodic location code, used to characterize daily, weekly, or other periodic location information in power load data;

[0097] Step 2.2: Perform token feature enhancement on the load seasonality token representation. Extract short-term fluctuation features between adjacent load time segments through local convolution of the token dimension, and extract load feature interaction information within the token through feedforward transformation of the feature dimension to obtain the enhanced load token representation. ,in, This indicates a token feature enhancement mapping. This represents the learnable gating coefficient, used to control the degree of influence of the token feature enhancement result on the original load seasonal token representation;

[0098] Step 2.3: Perform channel dependency modeling on the enhanced load token representation to capture the correlation between different variables. Select load channel patterns to participate in the modeling based on the energy characteristics of the corresponding channels of different load monitoring nodes, power supply areas, or electrical equipment. Then, perform attention modeling on the selected load channel patterns to obtain the channel-enhanced load token representation. ,in, This represents a channel dependency modeling mapping used to capture the correlations between different load monitoring nodes, power supply areas, or electrical equipment.

[0099] Step 2.4: Perform time-dependent modeling on the channel-enhanced load token representation to capture the long-term dependencies between different time tokens. Select the time patterns to participate in the modeling based on the low-frequency representation or pattern selection results in the load time dimension, and perform attention modeling on the selected time patterns to capture the long-term dependencies of the power load in continuous time, thus obtaining the time-enhanced load token representation. ,in, This represents a time-dependent modeling mapping;

[0100] Step 2.5: Perform cross-variable mixing on the time-enhanced load token representation, interacting information between different load variables at the variable dimensions corresponding to load monitoring nodes, power supply areas, or electrical equipment to obtain the cross-variable mixed load token representation. ,in, Represents a cross-variable mixed mapping;

[0101] Step 2.6: Perform frequency domain token enhancement on the cross-variable hybrid load token representation. Extract frequency domain response features in the time and variable channel dimensions of the load token to characterize the periodic fluctuations and frequency variations in the power load data, thus obtaining the frequency domain enhanced load token representation. ,in, This indicates a frequency domain token enhancement mapping. Indicates the frequency domain enhancement gating coefficient;

[0102] Step 2.7: Generate a fusion gating based on the periodic characteristics of the power load input sequence and the energy ratio of the seasonal load component to the trend load component. Then, dynamically fuse the intervariate mixed load token representation and the frequency-domain enhanced load token representation based on the fusion gating to obtain the seasonal load prediction features. The expression is:

[0103]

[0104] In the formula, Represents dynamic feature fusion mapping. This represents a fusion gating system generated from the periodic characteristics of electricity load and the energy ratio of the seasonal component to the trend component of the load. This represents the dynamic fusion gating coefficient.

[0105] Step 3: Generate a seasonal load forecast result based on the seasonal load forecast characteristics, and generate a load trend forecast result based on the load trend component; perform adaptive fusion on the seasonal load forecast result and the load trend forecast result to obtain a basic load forecast result; generate a residual correction term based on the energy ratio of the load trend component and the seasonal load component, and superimpose the residual correction term onto the basic load forecast result to obtain a normalized load forecast result;

[0106] Step 3.1: The load seasonality prediction result is jointly generated by the flattening mapping prediction unit and the token mixing prediction unit. The flattening mapping prediction unit is used to directly map the load seasonality prediction features to the prediction length. The token mixing prediction unit is used to perform hybrid modeling of the token relationships between different local load time segments and the internal feature relationships of the tokens. The outputs of the flattening mapping prediction unit and the token mixing prediction unit are fused through learnable mixing coefficients to obtain the load seasonality prediction result. The expression is:

[0107]

[0108] In the formula, Represents the flattening mapping prediction unit. Indicates a token-mixed prediction unit. This represents the learnable mixing coefficients used to fuse the outputs of the flattening prediction unit and the token mixing prediction unit;

[0109] Step 3.2: The load trend prediction result maps the load trend component on the historical input length to the prediction length through a trend mapping branch, thus obtaining the load trend prediction result within the future prediction interval. ,in, Indicates the trend mapping branch;

[0110] Step 3.3: Adaptively fuse the seasonal load forecast results and the trend load forecast results to obtain the basic load forecast results. ,in, This represents the fusion weights corresponding to the seasonal load forecast results. This indicates the fusion weights corresponding to the load trend forecast results;

[0111] Step 3.4: Calculate the load trend energy ratio based on the load trend component and the load seasonality component, and generate the residual prediction term based on the load trend energy ratio and the normalized power load input sequence. The expression is as follows:

[0112]

[0113]

[0114] In the formula, Indicates the energy ratio of load trend. This indicates an operation that calculates the variance of the tensor within the parentheses along the time dimension. The variance representing the load trend component. The variance representing the seasonal component of the load. This indicates the learnable residual gain. Indicates the residual prediction branch, Represents the residual prediction term;

[0115] Step 3.5: Superimpose the residual prediction term onto the basic load prediction result to obtain the normalized load prediction result. ,in, This indicates an energy constraint superposition operation, used to control the residual correction magnitude based on the base load forecast results and residual forecast terms, thereby reducing local deviations in power load forecasting.

[0116] Step 4: Perform inverse normalization on the normalized load forecast result to obtain the initial power load forecast result; perform continuity correction on the starting part of the initial power load forecast result based on the local anchor point at the end of the historical power load input sequence to obtain the power load forecast result for multiple future time steps.

[0117] Step 4.1: Using the sample mean and sample standard deviation saved during the reversible normalization process, perform inverse normalization on the obtained normalized load forecast results to obtain the initial power load forecast results. ;

[0118] Step 4.2: Calculate the local anchor point value based on one or more time steps at the end of the historical power load input sequence. The expression is:

[0119]

[0120] In the formula, Indicates the local anchor value. This indicates the length of the historical window used to calculate local anchor point values;

[0121] Step 4.3: Calculate the prediction starting offset based on the local anchor point value and the starting position of the initial power load prediction result. ,in, This represents the predicted load value corresponding to all samples and all load monitoring nodes, power supply areas, or electrical equipment at the first prediction time step of the initial power load prediction results.

[0122] Step 4.4: Based on the predicted initial offset, perform continuity correction on the initial portion of the initial power load prediction result to obtain the final power load prediction result, expressed as:

[0123]

[0124] In the formula, This indicates the final power load forecast result in the [number]th [year]. The predicted load values ​​for all samples and all load monitoring nodes, power supply areas, or electrical equipment at each prediction time step. This indicates the initial power load forecast result at the [number]th [year]. The predicted load value corresponding to each prediction time step. Indicates the prediction step size index. Indicates the predicted length. Indicates the first The continuity correction coefficient corresponding to each prediction time step.

[0125] Optionally, in this embodiment, the continuity correction coefficient is determined by a learnable correction parameter and a prediction length factor, so that the beginning part of the prediction sequence moves closer to the local anchor point at the end of the historical power load input sequence, and the influence of the local anchor point on the power load prediction result is gradually reduced as the prediction step size increases.

[0126] Optionally, in specific implementation, historical power load data of multiple load monitoring nodes or multiple electrical devices within the target power supply area are first acquired over a continuous period of time. This historical power load data includes the power load values ​​corresponding to each load monitoring node or electrical device at different historical moments. The historical power load data is then constructed into a multivariate historical power load time series data according to chronological order, and divided into training, validation, and test sets. Simultaneously, the historical power load input length and prediction length are set.

[0127] During the model training phase, historical power load sequences from the training set are input into the model and sequentially processed through reversible normalization, time-frequency decomposition, multi-scale load seasonality enhancement, local load time segmentation and token embedding, load token collaborative modeling, load trend and load seasonality fusion, residual correction, inverse normalization, and continuity correction to obtain power load prediction results for multiple future time steps. Based on the error between the power load prediction results and the actual power load values, the model parameters are trained and optimized.

[0128] During the model application phase, the historical power load sequence of the target power supply area over the most recent period is input into the trained model to obtain regional power load prediction results for multiple future time steps. These prediction results can be used to assist in load management, power generation planning, distribution network operation scheduling, and electricity demand analysis within the power supply area.

[0129] Furthermore, to verify the effectiveness of the present invention, this experiment uses the publicly available electricity load dataset Electricity. Electricity contains historical electricity load records of multiple electricity users over continuous time, which can be used to verify the feasibility and prediction effect of the method of the present invention in multivariate electricity load prediction scenarios.

[0130] Specifically, the Electricity power load dataset is divided into training, validation, and test sets according to time sequence. The model takes historical power load sequences as input and, after reversible normalization, time-frequency decomposition, multi-scale load seasonality enhancement, token collaborative modeling, load trend and load seasonality fusion, residual correction, inverse normalization, and continuity correction, outputs power load prediction results for multiple future time steps.

[0131] Furthermore, this experiment uses mean absolute error (MAE) and root mean square error (RMSE) as evaluation indicators for prediction performance. The smaller the values ​​of MAE and RMSE, the smaller the error between the prediction result and the actual power load value, and the better the prediction effect. Table 1 shows the basic information of the power load dataset used in this experiment.

[0132] Table 1. Basic information of the dataset used in the embodiments of the present invention.

[0133]

[0134] Furthermore, to verify the effectiveness of the key modules in this invention, ablation experiments were conducted on the complete method and several variants that removed the key modules. The experimental results are shown in Table 2.

[0135] Table 2 Ablation experiment results of the key modules of this invention on the power load dataset.

[0136]

[0137] As can be seen from Table 2, compared with the variant that removes key technology modules, the complete method achieves lower MAE and RMSE on the power load dataset, indicating that the time-frequency decomposition, multi-scale load seasonality enhancement, time-dependent modeling and channel-dependent modeling, frequency domain token fusion, and token hybrid prediction unit in this invention can jointly improve the power load prediction performance.

[0138] Furthermore, to verify the effectiveness of the method of this invention compared with existing prediction methods, this embodiment also selects DLinear, PatchTST, FEDformer, iTransformer, TimesNet, and TimeMixer as comparison methods and conducts experiments on the Electricity power load dataset. Here, DLinear represents a time series prediction method based on decomposing linear structures, PatchTST represents a time series prediction method based on time segment token modeling, FEDformer represents a Transformer prediction method based on frequency domain enhancement, iTransformer represents a Transformer prediction method based on inverted variable token modeling, TimesNet represents a prediction method based on time variation modeling, and TimeMixer represents a prediction method based on a multi-scale time mixture structure. The experimental results are shown in Table 3.

[0139] Table 3. Average experimental results comparing the method of this invention with existing prediction methods.

[0140]

[0141] As shown in Table 3, on the Electricity power load dataset, this invention achieves lower average MAE and average RMSE. Compared with existing prediction methods such as DLinear, PatchTST, FEDformer, iTransformer, TimesNet, and TimeMixer, this invention achieves better results in terms of overall average error. This indicates that this invention, through techniques such as reversible normalization, time-frequency decomposition, multi-scale load seasonality enhancement, token collaborative modeling, frequency domain token fusion, load trend and load seasonality fusion, and residual correction, can effectively improve the accuracy and stability of power load prediction.

[0142] It should be noted that the above experiments are only used to illustrate the feasibility and effectiveness of the method of the present invention in power load forecasting scenarios, and are not intended to limit the application scope of the present invention. The present invention can also adjust the input length, prediction length, decomposition scale, token length, sliding step size, embedding dimension, number of network layers, and training parameters according to different power load forecasting scenarios.

[0143] In summary, this invention addresses historical power load data from multiple load monitoring nodes, power supply areas, or electrical equipment in a power system. Through reversible normalization, time-frequency decomposition, multi-scale load seasonality enhancement, token collaborative modeling, load trend and load seasonality fusion, residual correction, and continuity correction, it enables the prediction of power load for multiple future time steps. This invention can be applied to scenarios such as power dispatching, load management, electricity demand analysis, and distribution operation auxiliary decision-making.

[0144] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A power load forecasting method based on multi-scale decomposition guidance and token collaborative modeling, characterized in that, The method includes the following steps: Step 1: Acquire historical power load data of multiple load monitoring nodes, power supply areas, and electrical equipment in the target power system over continuous time to construct multivariate historical power load time series data and perform reversible normalization processing to obtain the normalized power load input sequence. Then, perform time-frequency decomposition to obtain load trend components, load seasonality components, and multi-scale high-frequency detail components. Generate adaptive weights based on the energy characteristics of the multi-scale high-frequency detail components, and perform multi-scale enhancement on the load seasonality components based on the adaptive weights to obtain the enhanced load seasonality representation. Step 2: Divide the enhanced load seasonality representation into multiple local time segments and map the local time segments into load seasonality token representations; perform token feature enhancement on the load seasonality token representations, extract short-term load fluctuation features between adjacent local time segments, obtain the enhanced load seasonality token representations, and perform time-dependent modeling, channel-dependent modeling, cross-variable mixing, and frequency domain token fusion to obtain load seasonality prediction features; Step 3: Generate a seasonal load forecast result based on the seasonal load forecast characteristics, and generate a load trend forecast result based on the load trend component; perform adaptive fusion on the seasonal load forecast result and the load trend forecast result to obtain a basic load forecast result; generate a residual correction term based on the energy ratio of the load trend component and the seasonal load component, and superimpose the residual correction term onto the basic load forecast result to obtain a normalized load forecast result; Step 4: Perform inverse normalization on the normalized load forecast result to obtain the initial power load forecast result; perform continuity correction on the starting part of the initial power load forecast result based on the local anchor point at the end of the historical power load input sequence to obtain the power load forecast result for multiple future time steps.

2. The power load forecasting method based on multi-scale decomposition guidance and token collaborative modeling according to claim 1, characterized in that, Step 1 specifically refers to: Step 1.1: Define multivariate historical electricity load time series data as ,in, Indicates the sample batch size. Indicates the length of historical power load input. This indicates the number of load monitoring nodes, power supply areas, or electrical equipment. Each element in the table represents the power load value of the corresponding node, user, or device at the corresponding historical moment; The sample mean and sample standard deviation are calculated for the multivariate historical electricity load time series data along the time dimension. Reversible normalization is then performed based on the sample mean and sample standard deviation, expressed as: ； In the formula, This represents the normalized power load input sequence. This represents the sample mean of the historical electricity load input sequence over the time dimension. This represents the sample standard deviation of the historical power load input sequence over the time dimension. Step 1.2: Perform time-frequency decomposition on the normalized power load input sequence to obtain low-frequency approximate components and high-frequency detail components at multiple scales, expressed as: ； In the formula, This represents the time-frequency decomposition operator acting on the normalized power load input sequence. Represents the low-frequency approximate component. Indicates the first High-frequency detail components at each scale Indicates the number of decomposition scales; The high-frequency detail components at each scale are set to zero, and an inverse transform is performed based on the low-frequency approximation components to obtain the load trend components, expressed as follows: ； In the formula, This indicates the relationship with the time-frequency decomposition operator. The corresponding inverse transform operator, Indicates load trend components, This represents the set of zero components corresponding to the dimensions of high-frequency detail components at each scale; the load seasonality component is obtained based on the difference between the normalized power load input sequence and the load trend component. ; Step 1.3: Perform energy calculations on the high-frequency detail components at the multiple scales, and generate adaptive weights based on the energy characteristics of the high-frequency detail components at different scales. The expression is as follows: ； ； In the formula, This represents the operation of averaging the elements of the tensor within the parentheses. Indicates the first Energy values ​​of high-frequency detail components at each scale; Indicates the first Energy values ​​of high-frequency detail components at each scale; Indicates the first Adaptive weights corresponding to each scale; Indicates the summation index. ; A constant that prevents division by zero; The high-frequency detail components at different scales are length-aligned and then weighted and fused according to the adaptive weights to obtain a multi-scale load seasonal refinement representation. The expression is: ； In the formula, Indicates the length alignment operator; The multi-scale load seasonality refinement representation is fused with the load seasonality component to obtain the enhanced load seasonality representation. ,in, This represents the fusion coefficient.

3. The power load forecasting method based on multi-scale decomposition guidance and token collaborative modeling according to claim 2, characterized in that, Step 2 specifically includes: Step 2.1: Divide the obtained enhanced load seasonality representation into multiple local load time segments, and map the local load time segments into load seasonality token representations. ,in, This represents the local load time segment embedding mapping. Indicates periodic position encoding; Step 2.2: Perform token feature enhancement on the load seasonality token representation. Extract short-term fluctuation features between adjacent load time segments through local convolution of the token dimension, and extract load feature interaction information within the token through feedforward transformation of the feature dimension to obtain the enhanced load token representation. ,in, This indicates a token feature enhancement mapping. Indicates the learnable gating coefficient; Step 2.3: Perform channel-dependent modeling on the enhanced load token representation. Select the load channel patterns to participate in the modeling based on the energy characteristics of the corresponding channels of different load monitoring nodes, power supply areas, or electrical equipment, and perform attention modeling on the selected load channel patterns to obtain the channel-enhanced load token representation. ,in, This represents a channel dependency modeling mapping; Step 2.4: Perform time-dependent modeling on the channel-enhanced load token representation. Select the time patterns to participate in the modeling based on the low-frequency representation or mode selection results in the load time dimension, and perform attention modeling on the selected time patterns to obtain the time-enhanced load token representation. ,in, This represents a time-dependent modeling mapping; Step 2.5: Perform cross-variable mixing on the time-enhanced load token representation, interacting information between different load variables at the variable dimensions corresponding to load monitoring nodes, power supply areas, or electrical equipment to obtain the cross-variable mixed load token representation. ,in, Represents a cross-variable mixed mapping; Step 2.6: Perform frequency domain token enhancement on the cross-variable hybrid load token representation, extracting frequency domain response features in the time and variable channel dimensions of the load token to obtain the frequency domain enhanced load token representation. ,in, This indicates a frequency domain token enhancement mapping. Indicates the frequency domain enhancement gating coefficient; Step 2.7: Generate a fusion gating based on the periodic characteristics of the power load input sequence and the energy ratio of the seasonal load component to the trend load component. Then, dynamically fuse the intervariate mixed load token representation and the frequency-domain enhanced load token representation based on the fusion gating to obtain the seasonal load prediction features. The expression is: ； In the formula, Represents dynamic feature fusion mapping. This represents a fusion gating system generated from the periodic characteristics of electricity load and the energy ratio of the seasonal component and the trend component of the load. This represents the dynamic fusion gating coefficient.

4. The power load forecasting method based on multi-scale decomposition guidance and token collaborative modeling according to claim 3, characterized in that, Step 3 specifically refers to: Step 3.1: The load seasonality prediction result is jointly generated by the flattening mapping prediction unit and the token mixing prediction unit. The flattening mapping prediction unit is used to directly map the load seasonality prediction features to the prediction length. The token mixing prediction unit is used to perform hybrid modeling of the token relationships between different local load time segments and the internal feature relationships of the tokens. The outputs of the flattening mapping prediction unit and the token mixing prediction unit are fused through learnable mixing coefficients to obtain the load seasonality prediction result. The expression is: ； In the formula, Represents the flattening mapping prediction unit. Indicates a token-mixed prediction unit. Represents the learnable mixing coefficients; Step 3.2: The load trend prediction result maps the load trend component on the historical input length to the prediction length through a trend mapping branch, thus obtaining the load trend prediction result within the future prediction interval. ,in, Indicates a trend mapping branch; Step 3.3: Adaptively fuse the seasonal load forecast results and the trend load forecast results to obtain the basic load forecast results. ,in, This represents the fusion weights corresponding to the seasonal load forecast results. This indicates the fusion weights corresponding to the load trend forecast results; Step 3.4: Calculate the load trend energy ratio based on the load trend component and the load seasonality component, and generate the residual prediction term based on the load trend energy ratio and the normalized power load input sequence. The expression is as follows: ； ； In the formula, Indicates the energy ratio of load trend. This indicates an operation that calculates the variance of the tensor within the parentheses along the time dimension. This indicates the learnable residual gain. Indicates the residual prediction branch, Represents the residual prediction term; Step 3.5: Superimpose the residual prediction term onto the basic load prediction result to obtain the normalized load prediction result. ,in, This indicates an energy-constrained superposition operation.

5. The power load forecasting method based on multi-scale decomposition guidance and token collaborative modeling according to claim 4, characterized in that, Step 4 specifically refers to: Step 4.1: Using the sample mean and sample standard deviation saved during the reversible normalization process, perform inverse normalization on the obtained normalized load forecast results to obtain the initial power load forecast results. ; Step 4.2: Calculate the local anchor point value based on one or more time steps at the end of the historical power load input sequence. The expression is: ； In the formula, Indicates the local anchor value. This indicates the length of the historical window used to calculate local anchor point values. Indicates the first The power load values ​​corresponding to all samples and all load monitoring nodes, power supply areas or electrical equipment at each historical moment; Step 4.3: Calculate the prediction starting offset based on the local anchor point value and the starting position of the initial power load prediction result. ,in, This represents the predicted load value corresponding to all samples and all load monitoring nodes, power supply areas, or electrical equipment at the first prediction time step of the initial power load prediction results. Step 4.4: Based on the predicted initial offset, perform continuity correction on the initial portion of the initial power load prediction result to obtain the final power load prediction result, expressed as: ； In the formula, This indicates the final power load forecast result in the [number]th [year]. The predicted load values ​​for all samples and all load monitoring nodes, power supply areas, or electrical equipment at each prediction time step. This indicates the initial power load forecast result in the first... The predicted load value corresponding to each prediction time step. Indicates the prediction step size index. Indicates the predicted length. Indicates the first The continuity correction coefficient corresponding to each prediction time step.

Images