A power load prediction method based on time-frequency enhancement and multi-level wavelet modeling

Abstract

The application provides a power load prediction method based on time-frequency enhancement and multi-level wavelet modeling, which comprises the following steps: constructing a multivariate historical power load time sequence and generating an instance normalized sequence according to a variable channel; performing time domain enhancement and spectrum enhancement on the instance normalized sequence to obtain an enhanced feature sequence; performing level discrete wavelet decomposition on the enhanced feature sequence to obtain a level-1 approximate coefficient sequence and a detail coefficient sequence of each level wavelet coefficient; performing independent resolution branch modeling based on the coefficient sequences obtained by the decomposition, modeling and predicting the coefficient sequences corresponding to each resolution branch, obtaining a predicted approximate coefficient sequence and each level predicted detail sequence, and then performing multi-level wavelet reconstruction, and performing reverse normalization processing on the reconstruction result to obtain a power load prediction result. The method can simultaneously enhance the time domain and frequency domain features and combine multi-resolution modeling, thereby reducing the calculation complexity and improving the accuracy and stability of the long-time sequence prediction of the power load.

Description

Technical Field

[0001] This invention relates to the field of data analysis and time series forecasting technology, and in particular to a power load forecasting method based on time-frequency enhancement and multi-level wavelet modeling. Background Technology

[0002] Time series forecasting is a crucial technique for estimating future series from historical data, and it has wide applications in fields such as power load forecasting, energy management, and weather forecasting. Power load data, in particular, exhibits significant periodicity, multi-scale variations, and complex fluctuations such as sudden peaks and slumps. With the continuous accumulation of power system operation data, power load data is characterized by long time spans, multiple variables, and high sampling frequencies, making the need for modeling for long-term time series forecasting increasingly prominent.

[0003] For the long-term time-series forecasting of electricity load, researchers have conducted extensive studies focusing on long-dependency modeling, sequence decomposition modeling, and computational efficiency optimization. Early methods primarily employed statistical models. However, with the growth of data scale and the improvement of computing power, deep learning-based methods have gradually become mainstream. Models such as recurrent neural networks, convolutional neural networks, and Transformers have demonstrated strong fitting ability and predictive performance in complex nonlinear scenarios. Nevertheless, research shows that on some benchmark time-series forecasting datasets, simple linear models can achieve predictive performance equal to or even surpass that of some complex Transformer structures.

[0004] Therefore, methods based on the MLP-Mixer concept have gradually gained attention in recent years. These methods mix features in the time and channel dimensions and have achieved good results on some benchmark datasets. However, these methods also have certain limitations: when the historical window length is large, directly mixing in the long-term dimension leads to a rapid increase in computational complexity; on the other hand, some methods use time-domain decomposition based on moving averages to process multi-scale sequences, which is difficult to fully characterize the complex seasonal structure and sudden peaks and drops commonly found in power loads, and direct mixing of multi-scale sequences can easily introduce information interference.

[0005] In addition, some studies have attempted to enhance modeling capabilities by introducing frequency domain analysis or wavelet transform. For example, Fourier transform or wavelet decomposition can be used to extract information in the time and frequency domains to improve the model's ability to represent multi-scale features. However, existing technical solutions still suffer from difficulties in fully utilizing multi-resolution information, require multiple transformations between the time and frequency domains during processing, leading to excessive computational complexity and the potential for cross-scale information interference.

[0006] Therefore, it is necessary to provide a method for long-term power load forecasting to improve the forecasting accuracy and generalization ability of multivariate long-term power load forecasting. Summary of the Invention

[0007] The purpose of this invention is to provide a power load forecasting method based on time-frequency enhancement and multi-level wavelet modeling, so as to improve the forecasting accuracy and generalization ability of multivariable power load long-term forecasting.

[0008] In a first aspect, the power load prediction method based on time-frequency enhancement and multi-level wavelet modeling provided by this invention includes: constructing a multivariate historical power load time series, and performing instance normalization processing according to variable channels to generate an instance normalized sequence; performing time-domain enhancement and spectral enhancement on the instance normalized sequence to obtain an enhanced feature sequence; and performing... The first-order discrete wavelet decomposition yields the second-order discrete wavelet decomposition. The first-order approximation coefficient sequence and the detail coefficient sequence of each wavelet coefficient; based on the first-order approximation coefficient sequence. Independent resolution branch modeling is performed on the approximate coefficient sequence and the detailed coefficient sequence of wavelet coefficients at each level. The coefficient sequence corresponding to each resolution branch is modeled and predicted to obtain the predicted approximate coefficient sequence and the predicted detailed coefficient sequence at each level. Multi-level wavelet reconstruction is performed based on the predicted approximate coefficient sequence and the predicted detailed coefficient sequence at each level, and the reconstruction results are inversely normalized to obtain the power load prediction results.

[0009] The beneficial effects of the power load forecasting method based on time-frequency enhancement and multi-level wavelet modeling provided by this invention are as follows: The multi-level wavelet decomposition design represents the input power load time series as wavelet coefficient sequences of different resolutions, enabling the simultaneous extraction of low-frequency trend information and high-frequency detail information. Independent resolution branches are designed for processing wavelet coefficient sequences of different resolutions, avoiding information interference and loss, and improving the efficiency of multi-scale feature utilization. It can coordinate time-domain and frequency-domain feature enhancement with multi-resolution modeling, improving the accuracy and stability of long-term power load forecasting while reducing computational complexity.

[0010] In one possible embodiment, the instance normalization processing by variable channel employs reversible instance normalization, which includes: calculating the mean and standard deviation for each variable channel of each sample in the multivariate historical power load time series along the length dimension of the input sequence, and performing mean-removal and standard-deviation normalization on the input features of the multivariate historical power load time series by variable channel; when performing inverse normalization processing, the same mean and standard deviation as in the normalization processing are used to perform inverse transformation to recover the corresponding scale and offset information.

[0011] In another possible embodiment, the enhanced feature sequence is obtained by performing temporal and spectral enhancement on the instance normalized sequence, including: applying a transpose function to perform dimensionality transformation on the instance normalized sequence; performing cross-convolution on the dimensionally transformed sequence using two one-dimensional convolutional layers to obtain two convolutional outputs; processing the two convolutional outputs using a one-dimensional convolutional layer to obtain a temporal enhanced feature sequence; performing a fast Fourier transform on the temporal enhanced feature sequence to obtain a frequency domain representation; calculating the frequency domain power spectrum based on the frequency domain representation; determining an adaptive threshold based on a learnable threshold parameter; constructing a frequency mask and filtering frequency domain features; weighting the frequency domain representation with global complex learnable weights to obtain global spectral features; weighting the frequency domain features filtered by the mask with high-frequency complex learnable weights to obtain local spectral features; and fusing the global and local spectral features and performing an inverse fast Fourier transform to obtain the final enhanced feature sequence.

[0012] In other possible embodiments, each resolution branch includes a patch partitioning and embedding mapping module, a patch dimension mixing submodule, an embedding dimension mixing submodule, and a prediction head module; the resolution branch performs prediction processing including processing the first... The approximation coefficient sequence and the detail coefficient sequence of wavelet coefficients at each level are subjected to independent patching, embedding, block dimension mixing, embedding dimension mixing, and prediction head mapping.

[0013] The processing performed in the patch partitioning and embedding mapping module includes: setting the patch length and step size; performing patch partitioning after extending the end of the coefficient sequence to be partitioned according to the step size to obtain the patched output; and performing patch partitioning according to the set embedding dimension. Map the patched output to The patch embedding result is obtained from the 3D feature space.

[0014] The patch dimension blending submodule performs feature blending on the patch embedding results at the patch dimension to obtain patch dimension blended features. This includes batch normalization, dimension permutation, two-layer linear mapping, and activation function processing. The processing follows the formula: , ,in, This indicates mixed features across patch dimensions. This indicates the result of patch embedding. Indicates batch normalization, Indicates dimension permutation. This represents the features after batch normalization and dimension replacement processing of the patch embedding results. This represents the GELU activation function. This represents a linear mapping layer for patch-dimension hybrid submodules. The tensor space representing the mixed features of patch dimensions. Indicates the number of variables. Indicates the first The number of patches after patch division in each resolution branch.

[0015] The embedding dimension blending submodule performs feature blending on the patch dimension blending features to obtain embedded dimension blending features. The process includes: obtaining normalized features through dimension permutation and batch normalization, and then applying two layers of linear mapping and... The function further processes the normalized features and then concatenates the result with the normalized features. The processing procedure satisfies the following formula: , , Indicates normalized features, Represents a linear mapping layer for embedding dimension-mixed submodules. This represents the GELU activation function. This indicates a feature that is a mixture of embedded dimensions.

[0016] The prediction head module applies a flattening function to flatten the embedded dimension hybrid features, and performs a linear transformation on the flattened features to obtain the prediction wavelet coefficients for the corresponding resolution.

[0017] The multi-level wavelet reconstruction and inverse normalization process for the predicted wavelet coefficient sequence includes: reconstructing the predicted wavelet coefficient sequence in the order of inverse wavelet transform, and performing inverse normalization on the reconstruction result to obtain the power load prediction result. The inverse normalization process satisfies the following formula: ,in, express Time of the first The prediction length for each variable channel is The power load forecast results Indicates after refactoring Time of the first Predicted values ​​for each variable channel, Indicates the first The standard deviation of each variable channel along the length dimension of the input sequence. Indicates the first The mean of each variable channel along the length dimension of the input sequence.

[0018] Secondly, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described power load forecasting method based on time-frequency enhancement and multi-level wavelet modeling.

[0019] For the beneficial effects of the second aspect mentioned above, please refer to the description of the first aspect mentioned above. Attached Figure Description

[0020] Figure 1A flowchart illustrating a power load forecasting method based on time-frequency enhancement and multi-level wavelet modeling, provided for an embodiment of the present invention;

[0021] Figure 2 A schematic diagram of the method architecture for a power load forecasting method based on time-frequency enhancement and multi-level wavelet modeling provided in an embodiment of the present invention;

[0022] Figure 3 This is a schematic diagram of the structure of an interactive convolutional block and an adaptive spectral block provided in an embodiment of the present invention;

[0023] Figure 4 This is a schematic diagram of an electronic device structure provided in an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions in the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention. Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by those skilled in the art. The terms "comprising" and similar expressions used herein mean that the element or object preceding the word covers the element or object listed following the word and its equivalents, but do not exclude other elements or objects.

[0025] This embodiment provides a power load forecasting method based on time-frequency enhancement and multi-level wavelet modeling.

[0026] See Figure 1 and Figure 2 Power load forecasting methods based on time-frequency enhancement and multi-level wavelet modeling include:

[0027] S101: Construct a multivariate historical power load time series and generate instance normalized sequences according to variable channels.

[0028] In one possible embodiment, the instance normalization process by variable channel adopts reversible instance normalization, which includes: calculating the mean and standard deviation for each variable channel of each sample in the multivariate historical power load time series along the length dimension of the input sequence, and performing mean-removal and standard-deviation normalization on the input features of the multivariate historical power load time series by variable channel.

[0029] In a specific embodiment, given historical window data of a multivariate electricity load time series, a multivariate historical electricity load time series is constructed as follows: ,in, Represents a multivariate historical electricity load time series. , Indicates the number of variables. This indicates the length of the historical window. The corresponding prediction target is a future window with a length of [missing information]. The sequence, specifically in the form of... .

[0030] To mitigate the impact of non-stationarity in the electricity load time series, instance normalization is performed for each variable channel separately: Let Time of the first The mean of each variable channel within the historical window is The standard deviation is ,in, This represents a small constant to prevent the result from being zero. Using multivariate historical electricity load time series as input features, instance normalization (mean removal and standard deviation division) is performed on each variable channel of the input features. The normalization result is... , express Time of the first Instance normalization results for each variable channel. (Combined) The instance normalization results of each variable channel yield the instance normalization sequence. .

[0031] S102: Perform temporal and spectral enhancement on the instance normalized sequence to obtain the enhanced feature sequence.

[0032] In one possible embodiment, performing temporal and spectral enhancement on the instance normalized sequence to obtain an enhanced feature sequence includes: performing temporal feature extraction and interactive enhancement on the instance normalized sequence to obtain a temporal enhanced feature sequence; and performing global filtering and adaptive noise suppression on the temporal enhanced feature sequence in the frequency domain to obtain the final enhanced feature sequence.

[0033] For example, the transpose function is applied to transform the dimension of the instance normalized sequence; two one-dimensional convolutional layers are used to perform cross-convolution on the dimension-transformed sequence to obtain two convolutional outputs; a one-dimensional convolutional layer is used to process the two convolutional outputs to obtain a temporal enhanced feature sequence; a fast Fourier transform is performed on the temporal enhanced feature sequence to obtain a frequency domain representation; the frequency domain power spectrum is calculated based on the frequency domain representation, and an adaptive threshold is determined based on a learnable threshold parameter, thereby constructing a frequency mask and filtering frequency domain features; global complex learnable weights are used to weight the frequency domain representation to obtain global spectral features; high-frequency complex learnable weights are used to weight the filtered frequency domain features to obtain local spectral features; the global spectral features and local spectral features are fused and then subjected to inverse fast Fourier transform to obtain the final enhanced feature sequence.

[0034] See the instruction manual appendix Figure 3 In one specific embodiment, interactive convolutional blocks (ICBs) are set to perform temporal feature extraction and interactive enhancement on the instance normalized sequence to obtain a temporal enhanced feature sequence. The specific processing includes the following sub-steps: First, the transpose function is applied to perform dimensionality transformation on the instance normalized sequence: , This represents the dimensionality transformation result of the instance normalized sequence, where Transpose represents the transpose function. Let B represent the transposed temporal data dimension, and B represent the batch size. Then, the ICB contains two parallel one-dimensional convolutional layers (Conv1 and Conv2). Conv1 performs a linear combination of different variable channels at the same time step to capture global features, while Conv2 observes whether local temporal patterns exist several time steps before and after to model a wider range of contextual dependencies. The two one-dimensional convolutional layers are used to perform interactive convolutions on the dimensionality-transformed sequence to obtain the outputs of the two convolutional layers. , , This represents the GELU activation function. This indicates element-wise multiplication. Finally, the outputs of the two convolutional layers are added element-wise and then processed through the final one-dimensional convolutional layer Conv3 to obtain the temporal enhanced feature sequence. : .

[0035] It's important to note that Conv1 and Conv2 can select different kernel sizes to extract temporal features from different receptive fields, depending on the specific dataset and scenario. Conv3 fuses the interactive features and maps the feature dimensions back to the input feature dimensions. The output dimensions of Conv1 and Conv2 must be consistent for cross-multiplication, and ideally, Conv1 and Conv2 should handle different receptive fields to provide interactive scoring value. The output dimension of Conv3 must return to the input feature dimensions to connect to the subsequent ASB and backbone network. For example, a 1×1 convolution Conv1 and a 3×1 convolution Conv2 can be configured.

[0036] An adaptive spectrum block (ASB) is used to perform global filtering and adaptive noise suppression on the time-domain enhanced feature sequence in the frequency domain to obtain the final enhanced feature sequence. The specific processing includes the following sub-steps: First, a real-number fast Fourier transform is performed on the time-domain enhanced feature sequence to obtain the frequency domain representation. , Represents the Fast Fourier Transform of real numbers. This represents the number of frequency points corresponding to the real number after the Fast Fourier Transform. This represents a complex tensor space. Then, the frequency domain power spectrum is calculated to identify the dominant frequency components: P represents the power spectrum in the frequency domain. In the frequency domain representation, the first The complex spectral amplitudes corresponding to each variable channel Represents the power spectrum in the frequency domain The real tensor space in which it resides; a learnable threshold parameter is introduced, and an adaptive threshold is determined based on the frequency domain power spectrum. Construct a binary mask The binary mask is broadcast along the channel dimension and then multiplied element-wise with the frequency domain representation to achieve adaptive threshold filtering, thus obtaining the filtered frequency domain features. Among them, the learnable threshold parameter is used as part of the running parameters in network training, and is jointly optimized with the frequency domain complex weighted parameters during training. This allows the threshold setting to adaptively adjust the retention range of the dominant frequency component according to the spectral energy distribution of the training data, thereby reducing manual parameter tuning and improving the adaptive capability of computational processing.

[0037] Furthermore, globally complex learnable weights are employed. The global spectral features are obtained by element-wise weighting of the frequency domain representation. High-frequency complex learnable weights are used The local spectral features are obtained by element-wise weighting of the filtered frequency domain features. : , .

[0038] Finally, the global and local spectral features are fused and subjected to inverse fast Fourier transform to obtain the final enhanced feature sequence. : ,in, This represents the inverse fast Fourier transform. Indicates enhanced feature sequence The real tensor space in which it resides.

[0039] S103: Perform on the enhanced feature sequence The first-order discrete wavelet decomposition yields the second-order discrete wavelet decomposition. The sequence of approximate coefficients and the sequence of detailed coefficients of wavelet coefficients at each level.

[0040] In one possible embodiment, the enhanced feature sequence is performed. The first-order discrete wavelet decomposition retains only the final first-order discrete wavelet decomposition. Approximation coefficient sequence and Level 1 to Level 2 Complete detail coefficient sequence of wavelet coefficients : , Indicates the wavelet basis type, Denotes the decomposition series, and It is an adjustable hyperparameter. The value can be set based on the data length or experience. For example, it can be set in the range of 2 to 4 based on experience to avoid excessive decomposition of information due to too large a decomposition level and insufficient multi-scale decomposition due to too small a decomposition level.

[0041] In one specific embodiment, wavelet basis type Selected from at least one of the following wavelet families: Daubechies wavelet, Symlets wavelet, Coiflets wavelet, or Biorthogonal wavelet.

[0042] S104: Based on the first Independent resolution branch modeling is performed on the approximate coefficient sequence and the detail coefficient sequence of wavelet coefficients at each level. The coefficient sequence corresponding to each resolution branch is modeled and predicted to obtain the predicted approximate coefficient sequence and the predicted detail coefficient sequence at each level.

[0043] In one possible embodiment, each resolution branch includes a patch partitioning and embedding mapping module, a patch dimension mixing submodule, an embedding dimension mixing submodule, and a prediction head module; the resolution branch performs prediction processing including processing the first... The approximation coefficient sequence and the detail coefficient sequence of wavelet coefficients at each level are subjected to independent patching, embedding, block dimension mixing, embedding dimension mixing, and prediction head mapping.

[0044] The processing performed in the patch partitioning and embedding mapping module includes: setting the patch length and step size; performing patch partitioning after extending the end of the coefficient sequence to be partitioned according to the step size to obtain the patched output; and performing patch partitioning according to the set embedding dimension. Map the patched output to The patch embedding result is obtained from the 3D feature space.

[0045] The patch dimension blending submodule performs feature blending on the patch embedding results at the patch dimension to obtain patch dimension blended features. This includes batch normalization, dimension permutation, two-layer linear mapping, and activation function processing. The processing follows the formula: , ,in, This indicates mixed features across patch dimensions. This indicates the result of patch embedding. Indicates batch normalization, Indicates dimension permutation. This represents the features after batch normalization and dimension replacement processing of the patch embedding results. This represents the GELU activation function. This represents a linear mapping layer for patch-dimension hybrid submodules. This represents the real tensor space containing the patch-dimension hybrid features. Indicates the number of variables. Indicates the first The number of patches formed after the coefficient sequence corresponding to each resolution branch is divided into patches.

[0046] The embedding dimension blending submodule performs feature blending on the patch dimension blending features to obtain embedded dimension blending features. The process includes: obtaining normalized features through dimension permutation and batch normalization, and then applying two layers of linear mapping and... The function further processes the normalized features and then concatenates the result with the normalized features. The processing follows the formula: , , Indicates normalized features, Represents a linear mapping layer for embedding dimension-mixed submodules. This represents the GELU activation function. This indicates a feature that is a mixture of embedded dimensions.

[0047] The prediction head module applies a flattening function to flatten the embedded dimension mixed features, and then performs a linear transformation on the flattened features to obtain the prediction wavelet coefficients for the corresponding resolution.

[0048] In one specific embodiment, each wavelet coefficient sequence is modeled as the input to an independent resolution branch, and the total number of resolution branches is... Let the input of the arbitrary resolution branch be... , Indicates the input number The wavelet coefficient sequence of the nth resolution branch can specifically be the nth resolution branch. Approximation coefficient sequence Or the detail coefficient sequence of any first-order wavelet coefficients , Indicates the input number The coefficient sequence length of each resolution branch. Each resolution branch includes a patch partitioning and embedding mapping module, a patch dimension mixing submodule, an embedding dimension mixing submodule, and a prediction head module.

[0049] The processing performed by the patch partitioning and embedding mapping module includes: for the first... Each resolution branch has a patch length of [number]. Step size is , Used to control the local sensing range of a single patch. Used to control the sliding window's movement interval and patch density, according to and After end-padding the wavelet coefficient sequence, it is segmented using a sliding window. The number of patches obtained from the segmentation is... : After patching, a patch tensor is obtained. : , This represents the real tensor space where the patch tensor resides. Mapping the patch tensor to... The dimensional embedding space yields patch embedding results, i.e., patch-level feature representations. : Embedding indicates the embedding of a mapping function. The real tensor space representing the patch embedding result, where, This represents the patch embedding dimension, which is an even number selected between 32 and 128 through a hyperparameter search.

[0050] The processing performed by the patch dimension hybrid submodule includes: in the... In each resolution branch, for First, batch normalization and dimension permutation are performed. Then, two layers of linear mapping and activation functions are used to blend the patch dimensions to obtain mixed patch dimensions features. The process can be expressed by the following formula: , ,in, This indicates mixed features across patch dimensions. This indicates the result of patch embedding. Indicates batch normalization, Indicates dimension permutation. This represents the GELU activation function. This represents a linear mapping layer for patch-dimension hybrid submodules. Represents the real tensor space of patch-dimension mixed features. Indicates the number of variables. Indicates the first The number of patches corresponding to each resolution branch.

[0051] The processing performed by the embedding dimension blending submodule includes: performing inverse permutation and batch normalization on the patch dimension blending features to obtain normalized features; then processing the normalized features in the embedding dimension through two layers of linear mapping and activation functions, and concatenating the result with the normalized features to achieve blending with residuals to obtain the embedding dimension blending features. The processing can be expressed by the formula: , , Indicates normalized features, Represents a linear mapping layer for embedding dimension-mixed submodules. This represents the GELU activation function. This indicates a feature that is a mixture of embedded dimensions.

[0052] The processing performed by the prediction head module includes: applying a flattening function to flatten the embedded dimension mixture features, and performing a linear transformation on the flattened features to obtain the prediction wavelet coefficients for the corresponding resolution. The processing can be expressed by the following formula: , , The flattened feature is represented by the flattened function. The real tensor space representing the flattened features. Indicates the corresponding number The predicted wavelet coefficients for each resolution branch, where Linear represents a linear function. Indicates the first The real tensor space of the predicted wavelet coefficients corresponding to each resolution branch. Indicates the first The length of the predicted wavelet coefficients corresponding to each resolution branch.

[0053] Each resolution branch is processed independently to obtain the result compared to the first... Predicted wavelet coefficient sequences corresponding to the approximate coefficient sequence and the detail coefficient sequence of each level of wavelet coefficients , Indicates the relationship with the first The predicted wavelet coefficients corresponding to the first-order approximation coefficient sequence Indicates the relationship with the first The predicted wavelet coefficients corresponding to the detail coefficient sequence of the wavelet coefficients.

[0054] In a preferred embodiment, the resolution branch includes a patch partitioning and embedding mapping module, a Mixer module, and a prediction head module, wherein the Mixer module is composed of the patch dimension mixing submodule and the embedding dimension mixing submodule described in the above embodiments. In one possible embodiment, multiple Mixer modules are stacked in the resolution branch to improve the modeling capability of the resolution branch.

[0055] S105: Based on the predicted approximation coefficient sequence and the predicted detail coefficient sequence at each level, perform multi-level wavelet reconstruction, and perform inverse normalization processing on the reconstruction results to obtain the power load prediction results.

[0056] In one possible embodiment, multi-level wavelet reconstruction and inverse normalization of the predicted wavelet coefficients includes: reconstructing the predicted wavelet coefficient sequence in the order of inverse wavelet transform to obtain the reconstruction result. : , Represents the refactoring function. This indicates the wavelet basis type. The reconstructed results are then subjected to inverse normalization to obtain the power load forecast. The inverse normalization process satisfies the following formula: ,in, express Time of the first The prediction length for each variable channel is The power load forecast results Indicates after refactoring Time of the first Predicted values ​​for each variable channel, Indicates the first The standard deviation of each variable channel along the length dimension of the input sequence. Indicates the first The mean of each variable channel along the length dimension of the input sequence.

[0057] For example, when performing inverse normalization, the same mean and standard deviation as in normalization are used to perform an inverse transformation to recover the corresponding scale and offset information.

[0058] This invention provides a power load forecasting method based on time-frequency enhancement and multi-level wavelet modeling. It extracts and fuses time-domain features under different receptive fields using interactive convolutional blocks to suppress noise and enhance local and long-range dependency modeling capabilities. Furthermore, it maps time-domain features to the frequency domain using adaptive spectral blocks, constructs an adaptive threshold mask based on spectral energy distribution to select dominant frequencies, and performs weighted filtering using global and local learnable frequency domain weights before inverse transformation back to the time domain to obtain spectral enhancement features. Subsequently, it employs multi-level discrete wavelet decomposition to decompose the enhanced power load sequence into final-level approximation coefficients and multi-level detail coefficients, maintaining the relative independence of information at different frequency scales through branched modeling. In each resolution branch, it utilizes overlapping patch embedding and a mixer network to efficiently model long-sequence dependencies, outputting wavelet coefficient prediction results for future time periods, and obtaining the final predicted sequence through multi-level wavelet reconstruction. This invention can synergistically enhance time-domain and frequency-domain features and combine them with multi-resolution modeling, improving the accuracy and stability of long-term power load forecasting while reducing computational complexity.

[0059] Specifically, by using multi-level wavelet decomposition, the input power load time series is represented as a wavelet coefficient sequence with different resolutions, enabling the simultaneous extraction of low-frequency trend information and high-frequency detail information. Simultaneously, interactive convolutional blocks are designed to enhance local and global dependency representations and suppress noise disturbances in the time domain, while adaptive spectral blocks are designed to enhance frequency components in the frequency domain by using adaptive power spectrum thresholding and combining learnable frequency domain weights. Compared to existing modeling or decomposition schemes, this approach is more effective in characterizing complex changes and non-stationary features such as sudden spikes in the data.

[0060] Independent resolution branches were designed for processing wavelet coefficient sequences of different resolutions to avoid information interference and loss, thereby improving the efficiency of multi-scale feature utilization.

[0061] The design employs a block-based mechanism to extract local temporal patterns, and achieves cross-block global information interaction and feature dimension mixing through block-based blending and embedding blending, thereby balancing the ability to model local details and long-distance dependencies.

[0062] The design of modular processing and branched structure for long sequence modeling effectively reduces computational complexity and resource overhead compared to existing technologies that directly perform global modeling over long time dimensions, making it more suitable for deployment in real-world systems.

[0063] In other embodiments of this application, an electronic device is disclosed, such as... Figure 4 As shown, the electronic device 300 may include: one or more processors 301; a memory 302; a display 303; one or more application programs (not shown); and one or more computer programs 304. These devices can be connected via one or more communication buses 305. The one or more computer programs 304 are stored in the memory and configured to be executed by the one or more processors 301. The one or more computer programs 304 include instructions that can be used to perform actions such as... Figure 1 And the steps in the corresponding embodiments.

[0064] Through the above description of the embodiments, those skilled in the art will clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device, and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0065] In the embodiments of this application, the functional units can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0066] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, essentially, or the parts that contribute to the prior art, or all or part of the technical solutions, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as flash memory, portable hard disk, read-only memory, random access memory, magnetic disk, or optical disk.

[0067] The above description is merely a specific implementation of the embodiments of this application, but the protection scope of the embodiments of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in the embodiments of this application should be covered within the protection scope of the embodiments of this application. Therefore, the protection scope of the embodiments of this application should be determined by the protection scope of the claims.

Claims

1. A power load forecasting method based on time-frequency enhancement and multi-level wavelet modeling, characterized in that, include: Construct a multivariate historical power load time series and perform instance normalization processing according to variable channels to generate an instance normalized sequence; The process of enhancing the instance normalized sequence by performing temporal and spectral enhancements to obtain an enhanced feature sequence includes: applying a transpose function to transform the dimension of the instance normalized sequence; performing cross-convolution on the dimension-transformed sequence using two one-dimensional convolutional layers to obtain two convolutional outputs; processing the two convolutional outputs using a one-dimensional convolutional layer to obtain a temporal enhanced feature sequence; performing a fast Fourier transform on the temporal enhanced feature sequence to obtain a frequency domain representation; calculating the frequency domain power spectrum based on the frequency domain representation; determining an adaptive threshold based on a learnable threshold parameter; constructing a frequency mask and filtering frequency domain features; weighting the frequency domain representation with global complex learnable weights to obtain global spectral features; weighting the frequency domain features filtered by the mask with high-frequency complex learnable weights to obtain local spectral features; and fusing the global and local spectral features and performing an inverse fast Fourier transform to obtain the final enhanced feature sequence. The process involves setting interactive convolutional blocks to extract temporal features from the instance normalized sequence and perform interactive enhancement to obtain a temporally enhanced feature sequence. The specific steps include: First, applying a transpose function to transform the dimension of the instance normalized sequence. , This represents the dimensionality transformation result of the instance normalized sequence, where Transpose represents the transpose function. This represents the dimension of the transposed time-series data. B Let C represent the batch size, C represent the number of variables, and L represent the history window length. Then, the interactive convolutional block contains two parallel one-dimensional convolutional layers, Conv1 and Conv2. Conv1 performs a linear combination of different variable channels at the same time step to capture global features, while Conv2 observes whether local temporal patterns exist several time steps before and after the transformation to model a wider range of contextual dependencies. The two one-dimensional convolutional layers are used to perform interactive convolution on the transformed sequence to obtain the outputs of the two convolutional layers. , , This represents the GELU activation function. This indicates element-wise multiplication; finally, the outputs of the two convolutional layers are added element-wise and then processed through the final one-dimensional convolutional layer Conv3 to obtain the temporal enhanced feature sequence. : ; An adaptive spectral block is used to perform global filtering and adaptive noise suppression on the time-domain enhanced feature sequence in the frequency domain to obtain the final enhanced feature sequence. The specific processing includes the following sub-steps: First, a real-number fast Fourier transform is performed on the time-domain enhanced feature sequence to obtain the frequency domain representation. , Represents the Fast Fourier Transform of real numbers. This represents the number of frequency points corresponding to the real number after the Fast Fourier Transform. The complex tensor space is represented; then, the frequency domain power spectrum is calculated to identify the dominant frequency components: P represents the power spectrum in the frequency domain. In the frequency domain representation, the first The complex spectral amplitudes corresponding to each variable channel Represents the power spectrum in the frequency domain The real tensor space in which it resides; a learnable threshold parameter is introduced, and an adaptive threshold is determined based on the frequency domain power spectrum. Construct a binary mask The binary mask is broadcast along the channel dimension and then multiplied element-wise with the frequency domain representation to achieve adaptive threshold filtering, thus obtaining the filtered frequency domain features. ; Perform on the enhanced feature sequence The first-order discrete wavelet decomposition yields the second-order discrete wavelet decomposition. The sequence of approximate coefficients at each level and the sequence of detailed coefficients of wavelet coefficients at each level; Based on the Independent resolution branch modeling is performed on the approximate coefficient sequence and the detail coefficient sequence of wavelet coefficients at each level. The coefficient sequence corresponding to each resolution branch is modeled and predicted to obtain the predicted approximate coefficient sequence and the predicted detail coefficient sequence at each level. Based on the predicted approximation coefficient sequence and the predicted detail coefficient sequence at each level, multi-level wavelet reconstruction is performed, and the reconstruction results are inversely normalized to obtain the power load prediction results.

2. The method according to claim 1, characterized in that, The instance normalization process based on variable channels employs reversible instance normalization, which includes: calculating the mean and standard deviation for each variable channel of each sample in the multivariate historical power load time series along the length dimension of the input sequence, and performing mean-removal and standard-deviation-reduction normalization on the input features of the multivariate historical power load time series based on variable channels; When performing inverse normalization, the same mean and standard deviation as in normalization are used to perform an inverse transformation to recover the corresponding scale and offset information.

3. The method according to claim 1, characterized in that, Each resolution branch includes a patch partitioning and embedding mapping module, a patch dimension mixing submodule, an embedding dimension mixing submodule, and a prediction head module; The resolution branch performs prediction processing, including the first... The approximation coefficient sequence and the detail coefficient sequence of wavelet coefficients at each level are subjected to independent patching, embedding, block dimension mixing, embedding dimension mixing, and prediction head mapping.

4. The method according to claim 3, characterized in that, The processes performed in the patch partitioning and embedding mapping module include: Set the patch length and step size, and perform patching after extending the end of the coefficient sequence to be divided according to the step size to obtain the patched output; Based on the set embedding dimension Map the patched output to The patch embedding result is obtained from the 3D feature space.

5. The method according to claim 4, characterized in that, The patch dimension blending submodule performs feature blending processing on the patch embedding result at the patch dimension to obtain patch dimension blended features, including batch normalization, dimension permutation, two-layer linear mapping, and activation function processing. The processing process satisfies the following formula: , ,in, This indicates mixed features across patch dimensions. This indicates the result of patch embedding. Indicates batch normalization, Indicates dimension permutation. This represents the features after batch normalization and dimension replacement processing of the patch embedding results. This represents the GELU activation function. This represents a linear mapping layer for patch-dimension hybrid submodules. The tensor space representing the mixed features of patch dimensions. Indicates the number of variables. Indicates the first The number of patches after patch division in each resolution branch.

6. The method according to claim 5, characterized in that, The embedding dimension blending submodule performs feature blending processing on the patch dimension blending features in the embedding dimension to obtain embedded dimension blending features. The processing includes: obtaining normalized features through dimension permutation and batch normalization, and then performing two-layer linear mapping and... The function further processes the normalized features and then concatenates the result with the normalized features. The processing procedure satisfies the following formula: , , Represents normalized features. Represents a linear mapping layer for embedding dimension-mixed submodules. This represents the GELU activation function. This indicates a feature that is a mixture of embedded dimensions.

7. The method according to claim 6, characterized in that, The prediction head module applies a flattening function to flatten the embedded dimension hybrid features, and performs a linear transformation on the flattened features to obtain the prediction wavelet coefficients for the corresponding resolution.

8. The method according to claim 1, characterized in that, The multi-level wavelet reconstruction and inverse normalization process for the predicted wavelet coefficient sequence includes: reconstructing the predicted wavelet coefficient sequence in the order of inverse wavelet transform, and performing inverse normalization on the reconstruction result to obtain the power load prediction result. The inverse normalization process satisfies the following formula: ,in, express Time of the first The prediction length for each variable channel is The power load forecast results Indicates after refactoring Time of the first Predicted values ​​for each variable channel, Indicates the first The standard deviation of each variable channel along the length dimension of the input sequence. Indicates the first The mean of each variable channel along the length dimension of the input sequence.

9. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is executed by the processor, it implements the power load forecasting method based on time-frequency enhancement and multi-level wavelet modeling as described in any one of claims 1 to 8.

Images