A long-term power load forecasting method, system, device and medium

Abstract

The application relates to the technical field of power load prediction, and discloses a long-term power load prediction method, system, device and medium. The method comprises the following steps: acquiring historical power load, electricity price and meteorological data in a target area; extracting the global evolution trend, non-stationary change characteristics and the synchronous change relationship among the historical power load, electricity price and meteorological data, as a time-domain baseline prediction representation; decomposing the time-domain baseline prediction representation into a low-frequency approximate component and a plurality of high-frequency detail components, introducing a periodic nonlinear activation function with periodic inductive bias based on a channel attention mechanism, modeling the coupling relationship among the historical power load, electricity price and meteorological data in each frequency component and the periodic fluctuation existing in each frequency component, respectively, performing channel attention weighting on each frequency component, and obtaining a frequency-domain baseline prediction representation; and combining the time-domain baseline prediction representation and the frequency-domain baseline prediction representation to predict the long-term power load of the target area.

Description

Technical Field

[0001] This invention relates to the field of power load forecasting technology, and in particular to a long-term power load forecasting method, system, equipment and medium. Background Technology

[0002] Load forecasting, as a fundamental component of power system analysis and operation control, provides crucial information for grid planning, unit combination, economic dispatch, energy storage operation, reserve capacity allocation, and power trading decision support. Based on different time scales, power load forecasting can generally be categorized into ultra-short-term, short-term, medium-term, and long-term load forecasting. Ultra-short-term and short-term load forecasting focus more on real-time dispatch, intraday regulation, and short-cycle operation control, while long-term load forecasting focuses on load change trends over a longer future period, and is of great significance for grid planning, energy resource allocation, and long-term dispatch decisions.

[0003] Currently, deep learning-based methods have been applied in the field of power load forecasting. These methods first preprocess the input load data to reduce the impact of non-stationary data distribution on the forecast results; then, they extract the local variation features and time dependencies of the load data, and obtain the frequency characteristics of the load data through frequency domain transformation and attention mechanisms; finally, by fusing the time domain and frequency domain features, the power load forecast results are output.

[0004] However, actual long-term power loads typically exhibit nonlinearity, multi-scale periodicity, and non-stationarity; they are also dynamically coupled with various external variables such as temperature, humidity, air pressure, wind speed, solar irradiance, and electricity prices. Existing methods only achieve power load forecasting through conventional time-series dependencies and simple frequency domain transformations and attention weighting in frequency domain processing. These methods lack the ability to model the aforementioned complex periodic fluctuations and multivariate coupling relationships, resulting in low accuracy in power load forecasting. Summary of the Invention

[0005] The purpose of this invention is to provide a long-term power load forecasting method, system, device and medium that can solve the problem of low forecasting accuracy in existing methods.

[0006] To address the aforementioned technical problems, this invention provides a long-term power load forecasting method, comprising: Acquire historical power load, electricity price, and meteorological data within the target area to form a multivariate time series; Extract the global evolution trend, non-stationary change characteristics, and synchronous change relationship between historical power load, electricity price and meteorological data of multivariate time series as the time domain baseline prediction representation; Multivariate time series are decomposed into a low-frequency approximation component and multiple high-frequency detail components by multi-level discrete wavelet transform. The low-frequency approximation component is used to characterize the long-term trend and slow change characteristics of the multivariate time series, while the multiple high-frequency detail components are used to characterize the local fluctuations, rapid disturbances and periodic changes of the multivariate time series at different time scales. Based on the channel attention mechanism, by introducing a periodic nonlinear activation function with periodic inductive bias, the coupling relationship between historical power load, electricity price and meteorological data in the low-frequency approximation component and each high-frequency detail component, as well as the periodic fluctuations therein, are modeled to perform channel attention weighting on the low-frequency approximation component and the corresponding high-frequency detail component, so as to obtain the frequency domain baseline prediction representation. By combining time-domain baseline prediction representation and frequency-domain baseline prediction representation, the long-term power load of the target area is predicted.

[0007] Furthermore, after decomposing the multivariate time series into a low-frequency approximate component and multiple high-frequency detail components through multi-level discrete wavelet transform, the method further includes: For the low-frequency approximation component and each high-frequency detail component, multiple parallel learnable filters are used to enhance the frequency characteristics. The enhanced low-frequency approximation component and each high-frequency detail component are subjected to soft thresholding to suppress redundant disturbances with amplitudes less than a preset threshold and retain key frequency features, thereby obtaining the processed low-frequency approximation component and multiple high-frequency detail components.

[0008] Furthermore, the channel attention mechanism, by introducing a periodic nonlinear activation function with a periodic inductive bias, models the coupling relationship between historical power load, electricity price, and meteorological data in the low-frequency approximation component and each high-frequency detail component, as well as the periodic fluctuations therein, to perform channel attention weighting on the low-frequency approximation component and the corresponding high-frequency detail component, thereby obtaining a frequency domain baseline prediction representation, including: For the processed low-frequency approximation component and each high-frequency detail component, the historical power load, electricity price and meteorological data are linearly mixed across variables through convolutional layers, and then global average pooling is performed to extract the corresponding channel global features. Channel attention weights corresponding to the global features of each channel are generated through a fully connected network containing periodic nonlinear activation functions. The attention weight of each channel is multiplied element-wise with the corresponding processed low-frequency approximation component or high-frequency detail component to obtain the channel-weighted frequency features. By combining the low-frequency approximation components and the channel-weighted frequency features corresponding to each high-frequency detail component, a frequency domain baseline prediction representation is obtained.

[0009] Furthermore, the step of combining the channel-weighted frequency features corresponding to the low-frequency approximation component and each high-frequency detail component to obtain the frequency domain baseline prediction representation includes: For the channel-weighted frequency features corresponding to the low-frequency approximation component and each high-frequency detail component, a learnable transformation matrix is ​​used for feature mapping. The frequency domain features obtained after feature mapping are converted back to the time domain by inverse discrete wavelet transform, and the features are fused by preset learnable fusion weights of low-frequency approximation components and each high-frequency detail component to obtain the frequency domain baseline prediction representation.

[0010] Furthermore, the method of combining time-domain baseline prediction representation and frequency-domain baseline prediction representation to predict the long-term power load of the target area includes: Based on the dynamic gating mechanism, a shared latent representation is generated by the gating network according to the characteristics of the multivariate time series and the time-domain baseline prediction representation and the frequency-domain baseline prediction representation, so as to obtain the dynamic fusion weight of the time-domain baseline prediction representation and the frequency-domain baseline prediction representation. By utilizing dynamic fusion weights, the time-domain baseline prediction representation and the frequency-domain baseline prediction representation are fused, and the long-term power load of the target area is predicted.

[0011] Furthermore, the extraction of the global evolution trend, non-stationary change characteristics, and synchronous change relationships between historical power load, electricity price, and meteorological data of the multivariate time series as a time-domain baseline prediction representation includes: Multivariate time series are expanded into one-dimensional vectors according to channel and time dimensions; A time-domain baseline prediction representation is obtained by mapping a one-dimensional vector through a two-layer feedforward neural network.

[0012] Furthermore, the periodic nonlinear activation function that introduces a periodic inductive bias is: function.

[0013] The present invention also provides a long-term power load forecasting system, comprising: The sequence acquisition module is used to acquire historical power load, electricity price and meteorological data within the target area to form a multivariate time series. The time-domain feature extraction module is used to extract the global evolution trend, non-stationary change characteristics, and synchronous change relationship between historical power load, electricity price and meteorological data of multivariate time series as a time-domain baseline prediction representation; The frequency decomposition module is used to decompose a multivariate time series into a low-frequency approximation component and multiple high-frequency detail components through multi-level discrete wavelet transform. The low-frequency approximation component is used to characterize the long-term trend and slow change characteristics of the multivariate time series, while the multiple high-frequency detail components are used to characterize the local fluctuations, rapid disturbances and periodic changes of the multivariate time series at different time scales. The frequency domain feature extraction module is used to model the coupling relationship between historical power load, electricity price and meteorological data in the low-frequency approximate component and each high-frequency detail component, as well as the periodic fluctuations therein, based on the channel attention mechanism by introducing a periodic nonlinear activation function with periodic inductive bias. This allows for channel attention weighting of the low-frequency approximate component and the corresponding high-frequency detail component to obtain the frequency domain baseline prediction representation. The load forecasting module is used to forecast the long-term power load of a target area by combining time-domain baseline forecasting representation and frequency-domain baseline forecasting representation.

[0014] Embodiments of the present invention also provide a computer device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the above-described long-term power load forecasting method.

[0015] Embodiments of the present invention also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described long-term power load forecasting method.

[0016] The long-term power load forecasting method provided by this invention has at least the following beneficial effects: This invention, in predicting electricity load, not only considers historical electricity load within the target area but also incorporates electricity prices and various meteorological data to construct a multivariate time series. Then, it extracts time-domain and frequency-domain feature representations respectively. In the frequency-domain feature extraction process, the low-frequency trend component and multiple high-frequency detail components in the multivariate time series are first explicitly decomposed into multiple scales. This ensures that the time-domain and frequency-domain branches take into account both macroscopic load evolution patterns and microscopic non-stationary disturbance characteristics, while also fully considering the dynamic coupling relationships of the aforementioned multi-source variables in different frequency spaces, thereby improving the prediction accuracy of long-term electricity load.

[0017] Furthermore, by combining the cross-variable attention feature cross-mechanism of periodic inductive bias and introducing the periodic nonlinear activation function into the channel attention mechanism, the ability to express periodic features can be introduced in the channel weight generation process. This allows for better identification of periodic fluctuations and their coupling relationships among multiple variables, and enhances the ability to extrapolate the inherent strong periodicity of the load sequence through nonlinear periodic evolution. This avoids the attenuation or phase drift of periodic patterns under ultra-long-term forecasting, thereby further improving the forecasting accuracy of long-term power load. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 A flowchart illustrating a long-term power load forecasting method provided by the present invention. Figure 1 ; Figure 2 This is a schematic diagram of the structure of a long-term multivariate load prediction model provided by the present invention; Figure 3 A schematic diagram of a time-domain branch provided by the present invention; Figure 4 A schematic diagram of a frequency domain branch structure provided by the present invention; Figure 5 A flowchart illustrating a long-term power load forecasting method provided by the present invention. Figure 2 . Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0020] The technical solutions provided by the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] Long-term multivariate load forecasting not only needs to utilize the trend, periodicity, and volatility information in historical load sequences, but also needs to comprehensively consider the impact of multiple external variables such as electricity price, temperature, humidity, air pressure, wind speed, and solar irradiance on load changes. Compared to short-term load forecasting, long-term load forecasting needs to maintain forecast accuracy over a longer forecasting step, and forecasting errors are more likely to accumulate and amplify during multi-step output processes. Actual power load sequences typically exhibit nonlinearity, multi-scale periodicity, non-stationarity, and local abrupt disturbances, making it difficult for single time-domain models or traditional frequency-domain models to simultaneously account for long-term evolution trends and short-term high-frequency disturbances.

[0022] This invention provides a long-term power load forecasting method, which includes two parts: time-domain branch forecasting and frequency-domain branch forecasting. It can make full use of time-domain trend information and frequency-domain multi-scale periodic information, solve the problem that existing long-term load forecasting methods are insufficient in modeling complex periodic fluctuations and multivariate coupling relationships, and improve the forecasting accuracy of long-term power load.

[0023] One embodiment of the present invention relates to a long-term power load forecasting method. The specific process of the long-term power load forecasting method in this embodiment can be as follows: Figure 1 As shown, it includes: Step 101: Obtain historical power load, electricity price, and meteorological data within the target area to form a multivariate time series. Step 102: Extract the global evolution trend, non-stationary change characteristics, and synchronous change relationship between historical power load, electricity price and meteorological data of the multivariate time series as the time domain baseline prediction representation; Step 103: Decompose the multivariate time series into a low-frequency approximation component and multiple high-frequency detail components through multi-level discrete wavelet transform; wherein, the low-frequency approximation component is used to characterize the long-term trend and slow change characteristics of the multivariate time series, and the multiple high-frequency detail components are used to characterize the local fluctuations, rapid disturbances and periodic change characteristics of the multivariate time series at different time scales. Step 104: Based on the channel attention mechanism, by introducing a periodic nonlinear activation function with periodic inductive bias, the coupling relationship between historical power load, electricity price and meteorological data in the low-frequency approximation component and each high-frequency detail component, as well as the periodic fluctuations therein, are modeled to perform channel attention weighting on the low-frequency approximation component and the corresponding high-frequency detail component, so as to obtain the frequency domain baseline prediction representation. Step 105: Combine the time-domain baseline prediction representation and the frequency-domain baseline prediction representation to predict the long-term power load of the target area.

[0024] The following is a detailed description of the implementation details of the long-term power load forecasting method in this embodiment. The following content is only for the convenience of understanding and is not necessary for implementing this solution.

[0025] This embodiment uses a constructed long-term multivariate load forecasting model to predict the long-term electricity load of a target area using multivariate time series data. The overall structure of this long-term multivariate load forecasting model is as follows: Figure 2 As shown, it mainly includes a data input layer, a normalization layer, a time-frequency dual-branch prediction layer (containing parallel time-domain and frequency-domain branches), a dynamic fusion layer, and a result output layer.

[0026] In its implementation, this long-term multivariate load prediction model is trained using a multivariate load dataset constructed in the following manner: Suppose the preprocessed multivariate time series is as follows: ; in, Indicates the total length of the time series. Indicates the first Multivariate observations at each time step This indicates the number of variables (including load, electricity price, and various meteorological characteristics, such as temperature, humidity, air pressure, wind speed, solar radiation, etc.).

[0027] A sliding time window is used to construct training samples. Let the length of the historical input sequence be... L The predicted sequence length is H , Model input for each training sample and the corresponding true prediction label They are represented as follows: ; ; Therefore, a training dataset can be constructed: ; in, This indicates the number of training samples.

[0028] Using the aforementioned long-term multivariate load forecasting model, after obtaining the multivariate time series data for the target area, and addressing the issues of varying external variables' dimensions and seasonal distribution shifts, the multivariate time series data are first normalized at the sample level. For the input samples... , No. Mean of each variable channel over time and standard deviation They are respectively: ; ; in, Indicates the first The variable in the first... The value of each time step.

[0029] The normalized input sequence can be represented as: ; in, To prevent tiny constants with a denominator of zero, and These are learnable affine transformation parameters. This processing method allows different samples to be normalized based on their own statistical characteristics, thereby reducing the impact of non-stationary distribution variations on model training and prediction results.

[0030] Then, the multivariate time series is expanded into a one-dimensional vector according to the channel and time dimensions, and the one-dimensional vector is mapped through a two-layer feedforward neural network to obtain the time-domain baseline prediction representation.

[0031] In specific implementation, according to, for example Figure 3 The temporal prediction branch structure shown obtains the temporal baseline prediction representation: This branch first takes the input sequence Expanded into a one-dimensional vector according to the channel and time dimensions: ; in, This indicates a vectorization operation.

[0032] Then it is fed into a lightweight two-layer feedforward neural network, and the mapping process is as follows: ; in, and These are the weight matrix and bias term of the first layer of the neural network, respectively. This is a non-linear activation function. This activation function uses... function: ; The prediction output of the time-domain branch is: ; in, and These are the weight matrix and bias term of the second layer of the neural network, respectively. This represents the prediction result output by the time-domain branch.

[0033] This time-domain prediction branch can directly learn the synchronous changes between load, electricity price and meteorological variables in the original time space, providing a stable time-domain benchmark prediction result for subsequent fusion with the frequency-domain prediction branch.

[0034] Next, the multivariate time series is decomposed into a low-frequency approximate component and multiple high-frequency detail components by multi-level discrete wavelet transform, and the frequency domain baseline prediction representation is obtained based on this.

[0035] Specifically, for the low-frequency approximation component and each high-frequency detail component, multiple parallel learnable filters are used to enhance the frequency features. The enhanced low-frequency approximation component and each high-frequency detail component are then subjected to soft thresholding to suppress redundant disturbances with amplitudes less than a preset threshold and retain key frequency features, thus obtaining the processed low-frequency approximation component and multiple high-frequency detail components.

[0036] For the processed low-frequency approximation component and each high-frequency detail component, historical power load, electricity price, and meteorological data are linearly mixed across variables through convolutional layers, followed by global average pooling to extract the corresponding channel global features. A fully connected network containing periodic nonlinear activation functions is used to generate channel attention weights corresponding to each channel's global features. Each channel attention weight is then multiplied element-wise with the corresponding processed low-frequency approximation component or high-frequency detail component to obtain the channel-weighted frequency features.

[0037] For the channel-weighted frequency features corresponding to the low-frequency approximation component and each high-frequency detail component, a learnable transformation matrix is ​​used for feature mapping. The frequency domain features obtained after feature mapping are transformed back to the time domain using inverse discrete wavelet transform, and feature fusion is performed using the preset learnable fusion weights of the low-frequency approximation component and each high-frequency detail component to obtain the frequency domain baseline prediction representation.

[0038] In practical implementation, the above content is achieved through methods such as... Figure 4 The frequency domain prediction branch network FENet structure shown is implemented. (By...) Figure 4 It can be seen that the frequency domain branch uses the normalized multivariate input sequence As input, the original sequence is first decomposed into low-frequency approximate components and multiple high-frequency detail components through multi-level discrete wavelet transform; then, the multi-frequency decomposition enhancement module performs feature enhancement and noise suppression on each frequency component; next, the adaptive multi-channel attention module models the coupling relationship between different variable channels; finally, the adaptive multi-frequency fusion module fuses the prediction results corresponding to different frequency components to obtain the frequency domain branch prediction result.

[0039] The implementation process of multi-level discrete wavelet decomposition is as follows: To explicitly separate different frequency components in the input sequence, this invention employs multi-level discrete wavelet transform on the normalized input sequence. Decompose the wavelet basis functions. The number of decomposition layers is Then the multi-level discrete wavelet decomposition process can be expressed as: ; in, This represents the low-frequency approximate component obtained from the final decomposition layer, used to characterize the long-term trend and slow-change characteristics in the load sequence; Indicates the first High-frequency detail components are used to characterize local fluctuations, rapid disturbances, and periodic changes at different time scales. Because the multi-level discrete wavelet transform downsamples layer by layer along the time dimension during the decomposition process, different frequency components have different time lengths.

[0040] Therefore, in this embodiment, feature enhancement and variable modeling are performed on each frequency component according to different wavelet decomposition scales. In the adaptive multi-frequency fusion stage, inverse discrete wavelet transform is used to reconstruct the enhanced low-frequency approximate component and the high-frequency detail components of each layer layer by layer according to the hierarchical relationship of wavelet decomposition, so that the multi-scale frequency features are restored to a unified time length, thereby ensuring the dimensionality consistency in the frequency domain prediction result output process.

[0041] Through the above multi-level discrete wavelet decomposition, the original multivariate time series is divided into multiple subsequences with different frequency characteristics, thus providing a foundation for subsequent frequency component enhancement, variable coupling modeling, and multi-frequency fusion.

[0042] The implementation process of the multi-frequency decomposition enhancement module is as follows: Since different frequency components contribute differently to the long-term prediction results, and the frequency components obtained by decomposition may contain redundant perturbations and noise, this embodiment further constructs a multi-frequency decomposition enhancement module to perform feature enhancement and noise suppression on each frequency component.

[0043] For any frequency component ,use Enhance its frequency features using a parallel learnable filter: ; in, Indicates the first The weights of a learnable filter, , This represents element-wise multiplication. This represents the enhanced frequency component.

[0044] Subsequently, the enhanced frequency components are subjected to soft thresholding to suppress redundant perturbations with small amplitudes and preserve key frequency characteristics. ; ; in, Represents the soft threshold function. For threshold parameters, This represents the frequency components after enhancement and denoising.

[0045] This module enables the model to adaptively select key frequency information relevant to the prediction task, while reducing the impact of high-frequency noise and short-term abnormal disturbances on long-term prediction results.

[0046] The implementation process of the adaptive multi-channel attention module is as follows: Electricity load is dynamically coupled with variables such as electricity price, temperature, humidity, wind speed, and solar irradiance. Different variables contribute differently to the prediction results at different time scales and prediction step sizes. To model the dynamic correlations among multiple variables, this embodiment constructs an adaptive multi-channel attention module.

[0047] For the frequency components after multi-frequency decomposition and enhancement Firstly, adopt Convolution performs linear mixing across variables: ; in, express Convolution operations are used to enhance the information exchange between different variable channels.

[0048] Then, to Perform global average pooling to extract global features from each channel: ; in, This indicates a global average pooling operation.

[0049] Furthermore, channel attention weights are generated using a fully connected network incorporating periodic nonlinear activation functions: ; in, and For learnable weight matrix, The periodic nonlinear activation function that introduces a periodic inductive bias can be expressed as: ; in, This is a learnable parameter used to adjust the frequency of the periodic nonlinear components. This activation function can introduce periodic feature representation capabilities during the channel weight generation process, enabling the model to better identify the periodic fluctuations and their coupling relationships in load, electricity price, and meteorological variables, and helping to improve the model's ability to extrapolate periodic evolution characteristics.

[0050] Finally, the channel attention weights are multiplied element-wise by the input frequency components to obtain the channel-weighted frequency features: ; in, This represents the frequency characteristics after adaptive multichannel attention processing.

[0051] This module enables the model to adaptively enhance key variable features and suppress redundant variable information based on the importance of each variable in different frequency components, thereby improving the ability to model multivariate coupling relationships.

[0052] The implementation process of the adaptive multi-frequency fusion module is as follows: Different frequency components play different roles in long-term forecasting. Low-frequency components typically contain long-term trends and seasonal variations, while high-frequency components contain information on local disturbances, short-term fluctuations, and abrupt changes. To adaptively fuse the forecast results of different frequency components, this embodiment constructs an adaptive multi-frequency fusion module.

[0053] Frequency characteristics after channel attention processing First, a learnable transformation matrix is ​​used for feature mapping: ; in, For the first The learnable transformation matrix corresponding to each frequency component This represents the frequency characteristics after mapping.

[0054] Subsequently, the frequency domain features are transformed back to the time domain using inverse discrete wavelet transform to obtain the time domain prediction results for each frequency component: ; in, This represents the inverse discrete wavelet transform. Learnable fusion weights corresponding to different frequency components are introduced. The output of the frequency domain prediction branch can then be expressed as: ; in, This represents the prediction result output by the frequency domain prediction branch. Through this fusion method, the model can adaptively adjust the contribution of low-frequency trend information and high-frequency disturbance information in the final prediction based on different prediction samples and different prediction step sizes.

[0055] Finally, based on the obtained time-domain and frequency-domain baseline prediction representations, a shared implicit representation is generated through a gating network using a dynamic gating mechanism, taking into account the characteristics of the multivariate time series and the time-domain and frequency-domain baseline prediction representations. This yields a dynamic fusion weight for the time-domain and frequency-domain baseline prediction representations. Using this dynamic fusion weight, the time-domain and frequency-domain baseline prediction representations are fused, and the long-term power load of the target area is predicted.

[0056] In its implementation, the time-domain prediction branch can directly extract the overall evolution trend and synchronous change relationships of variables in the original sequence, while the frequency-domain prediction branch can extract multi-scale periodic features and local perturbation features; the two are complementary. To fully utilize the prediction results from the time and frequency domains, this invention employs a dynamic gating mechanism to adaptively fuse the prediction outputs of the two branches.

[0057] Let the time-domain branch output be The frequency domain branch output is The fusion weights generated by the dynamic gating network are The final normalized prediction result is: ; in, This is used to characterize the contribution of the frequency domain prediction result to the final output; This indicates the degree of contribution of the time-domain prediction result to the final output.

[0058] Dynamic gating weights can be adaptively generated by the gating network based on input sequence features or bi-branch hidden features: ; in, This represents a shared latent representation composed of input sequences or branch features. and These are the learnable parameters of the gated network. for function.

[0059] Through the aforementioned dynamic gating fusion method, the model can adaptively adjust the weights of time-domain and frequency-domain predictions based on the load variation characteristics, frequency composition, and multivariate coupling degree of different samples, thereby improving the accuracy and robustness of long-term prediction scenarios.

[0060] And due to type output In a normalized space, to obtain prediction results with practical physical meaning, inverse normalization is required. Based on the sample statistics of the above multivariate time series normalization formula, the inverse normalization process can be expressed as: ; in, Indicates the first The output variable in the future... The final predicted value at each time step.

[0061] Through this inverse normalization process, the model output can be restored to the original physical dimensions, resulting in the final long-term multivariate prediction results.

[0062] In summary, this embodiment constructs a long-term multivariate load forecasting method that integrates time-domain and frequency-domain dual-branch networks. This achieves collaborative modeling of the original time-domain evolution law, multi-scale frequency characteristics, and multivariate dynamic coupling relationships, improving the accuracy and robustness of long-term load, electricity price, and related meteorological variable forecasts. It effectively solves the problem that existing load forecasting methods are susceptible to high-frequency disturbances, error accumulation, and insufficient characterization of multivariate coupling relationships in long-term forecasting scenarios, which leads to a decrease in forecast accuracy.

[0063] The time-domain branch employs a lightweight multilayer perceptron (MLP) to directly capture the overall trend of variables in the original multivariate sequence, establishing a stable baseline prediction representation. The frequency-domain branch integrates a frequency enhancement network (FENet) to perform multi-scale frequency domain decomposition, feature enhancement, multivariate coupling modeling, and multi-frequency fusion, providing refined modeling of long-term trends, periodic fluctuations, high-frequency disturbances, and multivariate coupling relationships in the load sequence. Finally, a dynamic gating mechanism adaptively fuses the prediction outputs of the two branches based on the feature representations extracted from both branches, and outputs the long-term multivariate load prediction result after inverse normalization. This invention fully utilizes time-domain trend information and frequency-domain multi-scale periodic information, solving the problem of insufficient modeling capability for complex periodic fluctuations and multivariate coupling relationships in existing long-term load forecasting methods, and improving the accuracy and robustness of long-term power load, electricity price, and related meteorological variable predictions.

[0064] In some embodiments, the long-term power load forecasting method of the present invention can be implemented as follows: Figure 5 The process shown is as follows: Step S1: Obtain historical load, electricity price, and meteorological characteristic data within the target area to form raw multivariate time series data.

[0065] Step S2 involves preprocessing the original multivariate time series data, including data cleaning, outlier handling, missing value completion, time scale unification, and sample construction, to obtain the dataset.

[0066] Step S3: Based on the preset historical input sequence length L and predicted sequence length H, construct the multivariate time series as input samples and predicted labels.

[0067] Step S4: Perform Reversible Instance Normalization (RevIN) on the input samples to obtain the normalized input sequence.

[0068] By performing sample-level normalization on multivariate time series data that include historical load, electricity price, and meteorological characteristics, the impact of non-stationary distribution changes on the prediction results can be reduced.

[0069] Step S5: Input the normalized input sequence into the time domain branch and obtain the time domain prediction baseline result through a lightweight multilayer perceptron.

[0070] The time-domain prediction branch directly extracts the overall evolution trend, synchronous change relationship of variables and non-stationary change characteristics from the original multivariate sequence through a lightweight multilayer perceptron (MLP), and outputs the time-domain baseline prediction results.

[0071] Step S6: Input the normalized input sequence into the frequency domain branch, and obtain the frequency domain prediction result through multi-level discrete wavelet decomposition, multi-frequency decomposition enhancement, adaptive multi-channel attention and adaptive multi-frequency fusion.

[0072] The independent frequency domain prediction branch performs multi-frequency decomposition on the input sequence through multi-level discrete wavelet transform (MDWT), separating the long-term low-frequency trend component and multi-level high-frequency detail components. The multi-frequency decomposition enhancement module applies learnable filtering and soft thresholding to each component. The adaptive multi-channel attention module with fused periodic inductive bias characterizes the coupling relationship of multiple variables in different frequency domains. The adaptive multi-frequency fusion module restores the time domain and outputs the frequency domain prediction result.

[0073] Step S7: Through a data-driven dynamic gating fusion mechanism, the time-domain prediction results and frequency-domain prediction results are adaptively weighted and fused to obtain the fused prediction results in the normalized space.

[0074] Step S8: Perform denormalization on the fused normalized prediction results to obtain long-term multivariate prediction results restored to the original physical dimensions.

[0075] Through a feature-driven dynamic gating mechanism, the time-domain prediction results and frequency-domain prediction results are adaptively weighted and cross-fused based on the feature representations extracted by the two branches. After inverse normalization, the final long-term multivariate prediction results are obtained.

[0076] Compared with the prior art, the long-term power load forecasting method of the present invention has the following advantages: First, this invention proposes a decoupled long-term forecasting architecture with parallel time and frequency domain branches. The time domain branch utilizes a lightweight MLP to quickly lock the global baseline trend of historical evolution in multivariate time series, while the frequency domain branch explicitly separates multi-scale frequency components through a frequency enhancement network, extracting periodic features, high-frequency disturbances, and local abrupt changes. Parallel modeling of these two branches can simultaneously consider both macroscopic load evolution patterns and microscopic non-stationary disturbance characteristics, thereby improving the accuracy of forecasting complex load sequences.

[0077] Second, this invention designs a multi-frequency decomposition enhancement mechanism based on multi-level discrete wavelet decomposition and soft thresholding filtering. This mechanism not only solves the problem of characterizing frequency components at different scales after multi-level wavelet decomposition, but also adaptively selects sub-frequency bands through parallel learnable filters and soft thresholding denoising mechanisms, effectively suppressing random high-frequency noise disturbances while preserving details of short-term extreme load changes.

[0078] Third, this invention pioneered a cross-variable attention feature crossover mechanism combining periodic inductive bias. For the first time, this invention introduces the Snake periodic nonlinear activation function with learnable frequency parameters into wavelet domain cross-variable channel attention. This innovative combination enables the model to accurately capture the nonlinear periodic cascading effects between electricity price fluctuations, multi-source meteorological characteristics, and system load, completely overcoming the industry's technical bottleneck of periodic decay and phase shift in long-step load forecasting.

[0079] Fourth, a feature-driven dynamic gating branch-level adaptive adjustment mechanism was constructed. Based on the complementary feature representations formed by the time-domain branch and the frequency-domain branch, this mechanism adaptively allocates the contribution weights of different prediction branches under different samples, different prediction step sizes, and different operating scenarios, realizing the weighted fusion of time-frequency prediction results at the sample level and step size level, enabling the model to flexibly adapt to changes in prediction uncertainty under different power consumption scenarios and ultra-long prediction step sizes.

[0080] It can be seen that the present invention establishes a parallel architecture that combines time-frequency decoupling, periodic evolution extrapolation capability, and cross-variable dynamic modeling capability, which has important industrial application value.

[0081] The steps of the various methods described above are only for clarity. In practice, they can be combined into one step or some steps can be split into multiple steps. As long as they include the same logical relationship, they are all within the protection scope of this invention. Adding insignificant modifications or introducing insignificant designs to the algorithm or process, without changing the core design of the algorithm and process, are also within the protection scope of this invention.

[0082] Another embodiment of the present invention relates to a long-term power load forecasting system, which includes: The sequence acquisition module is used to acquire historical power load, electricity price and meteorological data within the target area to form a multivariate time series. The time-domain feature extraction module is used to extract the global evolution trend, non-stationary change characteristics, and synchronous change relationship between historical power load, electricity price and meteorological data of multivariate time series as a time-domain baseline prediction representation; The frequency decomposition module is used to decompose a multivariate time series into a low-frequency approximation component and multiple high-frequency detail components through multi-level discrete wavelet transform. The low-frequency approximation component is used to characterize the long-term trend and slow change characteristics of the multivariate time series, while the multiple high-frequency detail components are used to characterize the local fluctuations, rapid disturbances and periodic changes of the multivariate time series at different time scales. The frequency domain feature extraction module is used to model the coupling relationship between historical power load, electricity price and meteorological data in the low-frequency approximate component and each high-frequency detail component, as well as the periodic fluctuations therein, based on the channel attention mechanism by introducing a periodic nonlinear activation function with periodic inductive bias. This allows for channel attention weighting of the low-frequency approximate component and the corresponding high-frequency detail component to obtain the frequency domain baseline prediction representation. The load forecasting module is used to forecast the long-term power load of a target area by combining time-domain baseline forecasting representation and frequency-domain baseline forecasting representation.

[0083] It is not difficult to see that this embodiment is a system embodiment corresponding to the above method embodiments, and this embodiment can be implemented in conjunction with the above method embodiments. The relevant technical details and technical effects mentioned in the above embodiments are still valid in this embodiment, and will not be repeated here to reduce repetition. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the above embodiments.

[0084] It is worth mentioning that all modules involved in this embodiment are logical modules. In practical applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. Furthermore, to highlight the innovative aspects of this invention, this embodiment does not introduce units that are not closely related to solving the technical problem proposed by this invention; however, this does not mean that other units are absent from this embodiment.

[0085] Another embodiment of the present invention relates to a computer device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the long-term power load forecasting methods of the above embodiments.

[0086] The memory and processor are connected via a bus, which can include any number of interconnecting buses and bridges, connecting various circuits of one or more processors and memories. The bus can also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and will not be described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by the processor is transmitted over the wireless medium via an antenna, which further receives data and transmits it to the processor.

[0087] The processor manages the bus and general processing, and also provides various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory is used to store data used by the processor during operation.

[0088] Another embodiment of the present invention relates to a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the method embodiments described above.

[0089] That is, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0090] Those skilled in the art will understand that the above embodiments are specific embodiments for implementing the present invention, and in practical applications, various changes can be made to them in form and detail without departing from the spirit and scope of the present invention.

Claims

1. A long-term power load forecasting method, characterized in that, The method includes: Acquire historical power load, electricity price, and meteorological data within the target area to form a multivariate time series; Extract the global evolution trend, non-stationary change characteristics, and synchronous change relationship between historical power load, electricity price and meteorological data of multivariate time series as the time domain baseline prediction representation; Multivariate time series are decomposed into a low-frequency approximation component and multiple high-frequency detail components by multi-level discrete wavelet transform. The low-frequency approximation component is used to characterize the long-term trend and slow change characteristics of the multivariate time series, while the multiple high-frequency detail components are used to characterize the local fluctuations, rapid disturbances and periodic changes of the multivariate time series at different time scales. Based on the channel attention mechanism, by introducing a periodic nonlinear activation function with periodic inductive bias, the coupling relationship between historical power load, electricity price and meteorological data in the low-frequency approximation component and each high-frequency detail component, as well as the periodic fluctuations therein, are modeled to perform channel attention weighting on the low-frequency approximation component and the corresponding high-frequency detail component, so as to obtain the frequency domain baseline prediction representation. By combining time-domain baseline prediction representation and frequency-domain baseline prediction representation, the long-term power load of the target area is predicted.

2. The long-term power load forecasting method according to claim 1, characterized in that, After decomposing the multivariate time series into a low-frequency approximate component and multiple high-frequency detail components using multi-level discrete wavelet transform, the method further includes: For the low-frequency approximation component and each high-frequency detail component, multiple parallel learnable filters are used to enhance the frequency characteristics. The enhanced low-frequency approximation component and each high-frequency detail component are subjected to soft thresholding to suppress redundant disturbances with amplitudes less than a preset threshold and retain key frequency features, thereby obtaining the processed low-frequency approximation component and multiple high-frequency detail components.

3. The long-term power load forecasting method according to claim 2, characterized in that, The channel attention-based mechanism, by introducing a periodic nonlinear activation function with a periodic inductive bias, models the coupling relationship and periodic fluctuations between historical power load, electricity price, and meteorological data in the low-frequency approximation component and each high-frequency detail component, respectively. This allows for channel attention weighting of the low-frequency approximation component and its corresponding high-frequency detail component to obtain a frequency domain baseline prediction representation, including: For the processed low-frequency approximation component and each high-frequency detail component, the historical power load, electricity price and meteorological data are linearly mixed across variables through convolutional layers, and then global average pooling is performed to extract the corresponding channel global features. Channel attention weights corresponding to the global features of each channel are generated through a fully connected network containing periodic nonlinear activation functions. The attention weight of each channel is multiplied element-wise with the corresponding processed low-frequency approximation component or high-frequency detail component to obtain the channel-weighted frequency features. By combining the low-frequency approximation components and the channel-weighted frequency features corresponding to each high-frequency detail component, a frequency domain baseline prediction representation is obtained.

4. The long-term power load forecasting method according to claim 3, characterized in that, The frequency domain baseline prediction representation is obtained by combining the channel-weighted frequency features corresponding to the low-frequency approximation component and each high-frequency detail component, including: For the channel-weighted frequency features corresponding to the low-frequency approximation component and each high-frequency detail component, a learnable transformation matrix is ​​used for feature mapping. The frequency domain features obtained after feature mapping are converted back to the time domain by inverse discrete wavelet transform, and the features are fused by preset learnable fusion weights of low-frequency approximation components and each high-frequency detail component to obtain the frequency domain baseline prediction representation.

5. The long-term power load forecasting method according to claim 1, characterized in that, The method of combining time-domain baseline prediction representation and frequency-domain baseline prediction representation to predict the long-term power load of the target area includes: Based on the dynamic gating mechanism, a shared latent representation is generated by the gating network according to the characteristics of the multivariate time series and the time-domain baseline prediction representation and the frequency-domain baseline prediction representation, so as to obtain the dynamic fusion weight of the time-domain baseline prediction representation and the frequency-domain baseline prediction representation. By utilizing dynamic fusion weights, the time-domain baseline prediction representation and the frequency-domain baseline prediction representation are fused, and the long-term power load of the target area is predicted.

6. The long-term power load forecasting method according to claim 1, characterized in that, The extraction of the global evolution trend, non-stationary change characteristics, and synchronous change relationships between historical power load, electricity price, and meteorological data of multivariate time series, as a time-domain baseline prediction representation, includes: Multivariate time series are expanded into one-dimensional vectors according to channel and time dimensions; A time-domain baseline prediction representation is obtained by mapping a one-dimensional vector through a two-layer feedforward neural network.

7. The long-term power load forecasting method according to claim 1, characterized in that, The periodic nonlinear activation function that introduces a periodic inductive bias is: function.

8. A long-term power load forecasting system, characterized in that, The system includes: The sequence acquisition module is used to acquire historical power load, electricity price and meteorological data within the target area to form a multivariate time series. The time-domain feature extraction module is used to extract the global evolution trend, non-stationary change characteristics, and synchronous change relationship between historical power load, electricity price and meteorological data of multivariate time series as a time-domain baseline prediction representation; The frequency decomposition module is used to decompose a multivariate time series into a low-frequency approximation component and multiple high-frequency detail components through multi-level discrete wavelet transform. The low-frequency approximation component is used to characterize the long-term trend and slow change characteristics of the multivariate time series, while the multiple high-frequency detail components are used to characterize the local fluctuations, rapid disturbances and periodic changes of the multivariate time series at different time scales. The frequency domain feature extraction module is used to model the coupling relationship between historical power load, electricity price and meteorological data in the low-frequency approximate component and each high-frequency detail component, as well as the periodic fluctuations therein, based on the channel attention mechanism by introducing a periodic nonlinear activation function with periodic inductive bias. This allows for channel attention weighting of the low-frequency approximate component and the corresponding high-frequency detail component to obtain the frequency domain baseline prediction representation. The load forecasting module is used to forecast the long-term power load of a target area by combining time-domain baseline forecasting representation and frequency-domain baseline forecasting representation.

9. A computer device, characterized in that, include: At least one processor; And a memory communicatively connected to at least one processor; wherein the memory stores instructions executable by at least one processor, the instructions being executed by at least one processor to enable at least one processor to perform the long-term power load forecasting method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the long-term power load forecasting method as described in any one of claims 1 to 7.

Images