Power load prediction method and system
By performing wavelet transform and Fourier transform decomposition on the power load signal, and combining a parallel deep learning architecture of one-dimensional convolutional neural network and long short-term memory network, the shortcomings of existing power load prediction methods in feature extraction and noise processing are solved, and higher accuracy and robustness of power load prediction are achieved.
Patent Information
- Application Number
- CN202511321384.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-16
- Publication Date
- 2026-02-03
AI Technical Summary
Existing power load forecasting methods are insufficient in handling the complexity, nonlinearity, nonstationarity, and spatiotemporal coupling characteristics of modern power system load data, especially in terms of feature extraction depth, the length and accuracy of time-dependent capture, and robustness to noise and abrupt changes.
A power load forecasting system and method are adopted, including data acquisition, preprocessing, feature extraction and forecasting modules. The signal is decomposed by wavelet transform and fast Fourier transform, and deep learning is combined with one-dimensional convolutional neural network and long short-term memory network to decompose the power load signal into trend, seasonal and residual components. Deep information interaction and fusion are carried out, and a parallel deep learning architecture is constructed to capture spatiotemporal features.
It significantly improves the accuracy and robustness of power load forecasting, better adapts to the complex load characteristics of modern power systems, reduces forecasting errors, and improves the reliability of forecasting results.
Smart Images
Figure CN121457794A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power technology, and specifically to a power load forecasting method and system. Background Technology
[0002] Accurate load forecasting is crucial for the planning, dispatching, stable operation, and economic efficiency of modern power systems. Especially with the increasing penetration of renewable energy and the growing complexity of electricity consumption patterns, load sequences exhibit strong nonlinearity, non-stationarity, multi-scale fluctuations, and noise interference, posing significant challenges to forecasting technologies. Current mainstream load forecasting methods and their limitations include: Traditional statistical methods (such as Autoregressive Differential Moving Average (ARIMA) and multiple linear regression): These methods are usually based on linear assumptions, making it difficult to capture complex nonlinear dynamics. They have high requirements for data stationarity, limited effectiveness when processing high-dimensional and large-scale data, and poor adaptability to rapid changes in load patterns.
[0003] Traditional machine learning methods (such as Support Vector Regression (SVR), Random Forest, Artificial Neural Networks (ANN)) are superior to statistical methods in handling nonlinearity, but they often fail to fully explore the deep spatiotemporal features of data, especially long-distance time dependencies and complex relationships between features. They are also sensitive to noise and data quality and are prone to overfitting.
[0004] Existing deep learning methods (such as Recurrent Neural Networks (RNNs), Long Short-Term Memory Networks (LSTMs), and Gated Recurrent Units (GRUs)) excel at capturing time dependencies. However, they still face several challenges in practical applications: insufficient noise handling, incomplete feature representations, simplistic feature fusion mechanisms, and limited adaptability to non-stationarity, among other issues. Limitations of existing technologies: Existing power load forecasting methods still have many shortcomings in handling the complexity, nonlinearity, non-stationarity, and spatiotemporal coupling characteristics of modern power system load data, especially in terms of feature extraction depth, the length and accuracy of time-dependent capture, and robustness to noise and abrupt changes.
[0005] Based on this, the present invention designs a power load forecasting method and system to solve the above problems. Summary of the Invention
[0006] In view of the above-mentioned shortcomings of the existing technology, the present invention provides a power load forecasting method and system.
[0007] To achieve the above objectives, the present invention provides the following technical solution: A power load forecasting system includes a data acquisition module; The data acquisition module is used for acquiring raw load data; The data acquisition module is connected to a data preprocessing module, which is used for raw load data preprocessing and signal decomposition. The data preprocessing module is connected to the feature extraction module, which extracts deep features from the time domain and frequency domain and fuses them to obtain a comprehensive feature vector. The data preprocessing module is connected to a fully connected layer, which is used to synthesize feature vectors for deep information interaction; The fully connected layer is connected to the prediction module, which is used for prediction.
[0008] A forecasting method for an electricity load forecasting system includes the following steps: Step 1: The data acquisition module collects raw load data; Step 2: The data preprocessing module preprocesses the raw load data to obtain a clean load signal, and decomposes the clean load signal to obtain trend components, seasonal components and residual components. Step 3: The feature extraction module marks the clean load signal as time-domain data and adds the seasonal components and residual components to convert it into frequency-domain data; Step 4: Extract time-domain deep feature vectors and frequency-domain deep feature vectors from time-domain data and frequency-domain data respectively, and then fuse the time-domain deep feature vectors and frequency-domain deep feature vectors to obtain a comprehensive feature vector that contains both time-domain and frequency-domain information; Step 5: Input the comprehensive feature vector into the fully connected layer. The fully connected layer performs deep information interaction and finally outputs the predicted power load values for the next n time points.
[0009] Furthermore, the specific steps of step 2 are as follows: Step 21: Periodically impute missing values and correct outliers in the original load data, and then normalize the data to obtain the normalized values. Step 22: Perform multi-scale decomposition on the normalized value using wavelet transform, and then perform threshold processing on the high-frequency detail components to filter out random noise and obtain a clean load signal. Step 23: Decompose the pure load signal into trend components, seasonal components, and residual components using the STL method.
[0010] Furthermore, step 21 is detailed below: Step 211: Arrange the historical power load data in chronological order to construct a dataset containing missing values; Step 212: Find missing values Y(T,t) The value at the same time on the day before and after Y(T - 1,t), Y(T + 1,t) and the value at the moment before and after the same dayY(T,t - 1), Y(T,t + 1) Then, a weighted average is performed to obtain the missing values. Y(T,t) The missing values are added to the dataset containing the missing values to obtain the complete dataset. Y(T,t) represents the missing values at time t on day T. Step 213: In the time dimension, determine whether the difference between the data in the dataset and its neighboring data is greater than a preset time dimension threshold. If the determination is no, proceed to step 214. If the determination is yes, use smoothing correction to perform correction, and then proceed to step 214. In the period dimension, determine whether the difference between the data in the dataset and its neighboring data is greater than a preset period dimension threshold. Points exceeding the preset threshold will be smoothed. If the determination is no, proceed to step 214. If the determination is yes, use smoothing correction to perform correction, and then proceed to step 214. Step 214: Perform max-min normalization on the data obtained in step 213, then map the data to the [0,1] interval to eliminate the influence of dimensions and obtain the normalized value after processing.
[0011] Furthermore, the maximum-minimum normalization calculation method is as follows: in, Normalized value The minimum value among the data obtained in step 213. The maximum value among the data obtained in step 213. The data obtained in step 213.
[0012] Furthermore, step 22 is detailed below: Step 221: Use wavelet basis functions to perform multi-scale decomposition on the normalized value, decomposing it into a low-frequency approximate component and n high-frequency detail components at different scales; Step 222: Calculate high-frequency detail components ; Step 223: Calculate the high-frequency detail components By merging with the low-frequency approximate component, and reconstructing the denoised pure signal through inverse wavelet transform, the signal is denoised.
[0013] Furthermore, the wavelet basis functions are as follows: in, It is a normalized value. For wavelet basis functions, For scaling function, For high-frequency detail components, It is a low-frequency approximation component.
[0014] Furthermore, step 222 is performed as follows: Step 2221: Calculate the noise intensity using the median absolute deviation method to obtain the standard deviation estimate sigma; Step 2222: Calculate the dynamic threshold λ according to the general threshold rule; Step 2223: For high-frequency detail components whose absolute value is less than λ, set them directly to zero (consider them as noise). For high-frequency detail components whose absolute value is greater than λ, shrink them towards zero by a distance of λ. The formula is as follows: .
[0015] Furthermore, the specific steps for adding the seasonal components and residual components to convert them into frequency domain data are as follows: Step A: The seasonal component and the residual component are added point by point to obtain the high-frequency fluctuation signal; Step B: Use the Fast Fourier Transform algorithm to convert the high-frequency fluctuation signal into frequency domain data.
[0016] Furthermore, the specific operation of step B is as follows: the high-frequency fluctuation signal is sliced using a sliding window, the amplitude spectrum is calculated, and the first half of the amplitude spectrum is taken as the frequency domain data.
[0017] Beneficial effects: This invention improves data quality and noise resistance: It uses wavelet transform signal processing method for refined denoising preprocessing, effectively separates and suppresses non-stationary noise, and protects the key fluctuation characteristics of the signal. Compared with traditional fixed threshold or filtering methods, it can more intelligently handle non-stationary noise of power load, retain important transient characteristics, and significantly improve the robustness of the model in real noisy environments.
[0018] This invention is based on the decoupling and reconstruction of data characteristics: it introduces a seasonal-trend decomposition technique based on local weighted regression to further decompose the denoised signal into trend components, seasonal components and residual components. By selectively recombining the trend components, seasonal components and residual components, two data streams with different physical meanings are constructed.
[0019] This invention enables in-depth mining of frequency domain features: it employs a fast Fourier transform algorithm to extract the frequency domain features of a signal and utilizes a one-dimensional convolutional neural network to efficiently learn the periodic patterns within it.
[0020] This invention presents a dual-stream parallel deep learning architecture: a parallel processing structure comprising a one-dimensional CNN network and a two-layer LSTM network. The one-dimensional CNN network focuses on analyzing the frequency domain characteristics of high-frequency signals composed of seasonal and residual components, efficiently learning their periodic patterns. The two-layer LSTM network focuses on processing data streams representing overall load trends and major changes, accurately capturing long-term time dependencies. Employing a "divide and conquer" strategy, it avoids mutual interference between different data components, far superior to the traditional approach of directly inputting mixed signals into a single model. Furthermore, deep fusion through subsequent fully connected layers achieves complementary advantages, resulting in a comprehensive performance superior to simple model concatenation or haphazardly inputting all features into a complex network.
[0021] This invention provides a short-term and ultra-short-term power load forecasting solution with higher prediction accuracy, stronger robustness, and better adaptability to the complex load characteristics of modern power systems, thereby significantly reducing prediction errors and improving the reliability of prediction results under various operating conditions. Attached Figure Description
[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.
[0023] Figure 1 This is a block diagram of a power load forecasting system according to the present invention. Detailed Implementation
[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present 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 the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0025] The present invention will be further described below with reference to embodiments.
[0026] Example 1: Please refer to Figure 1 A power load forecasting system, including a data acquisition module; The data acquisition module is used for acquiring raw load data; The data acquisition module is connected to a data preprocessing module, which is used for raw load data preprocessing and signal decomposition. The data preprocessing module is connected to the feature extraction module, which extracts deep features from the time domain and frequency domain and fuses them to obtain a comprehensive feature vector. The data preprocessing module is connected to a fully connected layer, which is used to synthesize feature vectors for deep information interaction; The fully connected layer is connected to the prediction module, which is used for prediction.
[0027] Example 2, based on Example 1: A prediction method for a power load forecasting system, comprising the following steps: Step 1: The data acquisition module collects raw load data; Step 2: The data preprocessing module preprocesses the raw load data to obtain a clean load signal, and decomposes the clean load signal to obtain trend components, seasonal components and residual components. Step 3: The feature extraction module marks the clean load signal as time-domain data and adds the seasonal components and residual components to convert it into frequency-domain data; Step 4: Extract time-domain deep feature vectors and frequency-domain deep feature vectors from time-domain data and frequency-domain data respectively, and then fuse the time-domain deep feature vectors and frequency-domain deep feature vectors to obtain a comprehensive feature vector that contains both time-domain and frequency-domain information; Step 5: Input the comprehensive feature vector into the fully connected layer. The fully connected layer performs deep information interaction and finally outputs the predicted power load values for the next n time points.
[0028] This invention provides a short-term and ultra-short-term power load forecasting solution with higher prediction accuracy, stronger robustness, and better adaptability to the complex load characteristics of modern power systems, thereby significantly reducing prediction errors and improving the reliability of prediction results under various operating conditions.
[0029] Step 2 is detailed below: Step 21: Periodically impute missing values and correct outliers in the original load data, and then normalize the data to obtain the normalized values. Step 22: Perform multi-scale decomposition on the normalized value using wavelet transform, and then perform threshold processing on the high-frequency detail components to filter out random noise and obtain a clean load signal. Step 23: Decompose the pure load signal into trend components, seasonal components, and residual components using the STL method.
[0030] Step 21 is as follows: Step 211: Arrange the historical power load data in chronological order to construct a dataset containing missing values; Step 212: Find missing values Y(T,t) The value at the same time on the day before and afterY(T - 1,t), Y(T + 1,t) and the value at the moment before and after the same day Y(T,t - 1), Y(T,t + 1) Then, a weighted average is performed to obtain the missing values. Y(T,t) The missing values are added to the dataset containing the missing values to obtain the complete dataset. Y(T,t) represents the missing values at time t on day T. Step 213: In the time dimension, determine whether the difference between the data in the dataset and its neighboring data is greater than a preset time dimension threshold. If the determination is no, proceed to step 214. If the determination is yes, use smoothing correction to perform correction, and then proceed to step 214. In the period dimension, determine whether the difference between the data in the dataset and its neighboring data is greater than a preset period dimension threshold. Points exceeding the preset threshold will be smoothed. If the determination is no, proceed to step 214. If the determination is yes, use smoothing correction to perform correction, and then proceed to step 214. Step 214: Perform max-min normalization on the data obtained in step 213, then map the data to the [0,1] interval to eliminate the influence of dimensions and obtain the normalized value after processing.
[0031] The calculation method for the maximum-minimum normalization process is as follows: in, Normalized value The minimum value among the data obtained in step 213. The maximum value among the data obtained in step 213. The data obtained in step 213.
[0032] Step 22 is detailed below: Step 221: Use wavelet basis functions to perform multi-scale decomposition on the normalized value, decomposing it into a low-frequency approximate component and n high-frequency detail components at different scales; Step 222: Calculate high-frequency detail components ; Step 223: Calculate the high-frequency detail components By merging with the low-frequency approximate component, and reconstructing the denoised pure signal through inverse wavelet transform, the signal is denoised.
[0033] The wavelet basis functions are as follows: in, It is a normalized value. For wavelet basis functions, For scaling function, For high-frequency detail components, It is a low-frequency approximation component.
[0034] Step 222 is performed as follows: Step 2221: Calculate the noise intensity using the median absolute deviation method to obtain the standard deviation estimate sigma; The standard deviation estimate sigma is an indicator used to describe the intensity of noise. The larger the value, the stronger the noise. The larger the calculated dynamic threshold λ, the more strictly strong noise can be filtered out. Step 2222: Calculate the dynamic threshold λ according to the general threshold rule; Step 2223: For high-frequency detail components whose absolute value is less than λ, set them directly to zero (consider them as noise). For high-frequency detail components whose absolute value is greater than λ, shrink them towards zero by a distance of λ. The formula is as follows: .
[0035] Step 23 is as follows: Step 231: Input the clean load signal and a specified period into a seasonal-trend decomposition method based on local weighted regression, and iteratively use local multinomial regression to smooth the data, decomposing it into trend components, seasonal components and residual components.
[0036] For example, if there are 96 data points sampled every 15 minutes, then a specified period is set to 96.
[0037] The specific steps for adding the seasonal components and residual components to convert the data into frequency domain data are as follows: Step A: The seasonal component and the residual component are added point by point to obtain the high-frequency fluctuation signal; Step B: Use the Fast Fourier Transform algorithm to convert the high-frequency fluctuation signal into frequency domain data.
[0038] The specific operation of step B is as follows: the high-frequency fluctuation signal is sliced using a sliding window, the amplitude spectrum is calculated, and the first half of the amplitude spectrum is taken as the frequency domain data.
[0039] Because the spectrum is symmetrical, only the first half needs to be taken to form a complete frequency domain data. The specific operation of slicing is to convert the high-frequency fluctuating signal from the time dimension to the frequency dimension using the Fast Fourier Transform algorithm.
[0040] The conversion calculation is as follows: Where X[k] is the transformed frequency domain representation, k represents the frequency index from 0 to N-1; N represents the length of the window; This represents the input load value, where n is the index of the time point, from 0 to N-1; The core transformation kernel is represented by a frequency detector.
[0041] The window length is one week's worth of data; Calculate the amplitude spectrum: Take the absolute value of X[k] and normalize it to obtain the amplitude spectrum.
[0042] Amplitude represents the "strength" or "importance" of each frequency component in the signal.
[0043] Step 4: Extract time-domain deep feature vectors and frequency-domain deep feature vectors from the time-domain data and frequency-domain data respectively, and then fuse the time-domain deep feature vectors and frequency-domain deep feature vectors to obtain a comprehensive feature vector that contains both time-domain and frequency-domain information.
[0044] The specific steps for step 4 are as follows: Step 41: Input the time-domain data into the LSTM network to learn the long-term temporal dependencies of the load sequence and obtain the time-domain deep feature vector; An LSTM network is a two-layer LSTM network. An LSTM network receives time-domain data and learns and remembers long-term dependencies in the time-series data through internal "forget gate," "input gate," and "output gate," ultimately outputting a time-domain deep feature vector representing the time-domain information. Step 42: Input the frequency domain data into a one-dimensional CNN network to learn the implicit periodicity and frequency combination patterns, and obtain the frequency domain deep feature vector; CNN (Convolutional Neural Network): A feedforward network consisting of multiple one-dimensional convolutional layers, batch normalization layers, ReLU activation functions, and pooling layers. CNNs receive frequency domain data, automatically learn and extract the most representative combinations of frequency patterns through convolutional kernels, and finally output a frequency domain deep feature vector representing frequency domain information. Step 43: Concatenate the feature vectors learned from the two branches to form a comprehensive feature vector that combines information from both the time and frequency domains; The torch.cat algorithm is used to concatenate the frequency domain depth feature vector and the time domain depth feature vector to form a comprehensive feature vector that contains information from both the time and frequency domains.
[0045] The fully connected layer is a fully connected network consisting of two layers.
[0046] This invention improves data quality and noise resistance: it uses wavelet transform signal processing method for refined denoising preprocessing, effectively separates and suppresses non-stationary noise, and protects the key fluctuation characteristics of the signal. Compared with traditional fixed threshold or filtering methods, it can more intelligently handle non-stationary noise of power load, retain important transient characteristics, and significantly improve the robustness of the model in real noisy environments.
[0047] This invention is based on the decoupling and reconstruction of data characteristics: it introduces a seasonal-trend decomposition technique based on local weighted regression to further decompose the denoised signal into trend components, seasonal components and residual components. By selectively recombining the trend components, seasonal components and residual components, two data streams with different physical meanings are constructed.
[0048] This invention enables in-depth mining of frequency domain features: it employs a fast Fourier transform algorithm to extract the frequency domain features of a signal and utilizes a one-dimensional convolutional neural network to efficiently learn the periodic patterns within it.
[0049] This invention presents a dual-stream parallel deep learning architecture: a parallel processing structure comprising a one-dimensional CNN network and a two-layer LSTM network. The one-dimensional CNN network focuses on analyzing the frequency domain characteristics of high-frequency signals composed of seasonal and residual components, efficiently learning their periodic patterns. The two-layer LSTM network focuses on processing data streams representing overall load trends and major changes, accurately capturing long-term time dependencies. Employing a "divide and conquer" strategy, it avoids mutual interference between different data components, far superior to the traditional approach of directly inputting mixed signals into a single model. Furthermore, deep fusion through subsequent fully connected layers achieves complementary advantages, resulting in a comprehensive performance superior to simple model concatenation or haphazardly inputting all features into a complex network.
[0050] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A power load forecasting system, comprising a data acquisition module, characterized in that: The data acquisition module is used for acquiring raw load data; The data acquisition module is connected to a data preprocessing module, which is used for raw load data preprocessing and signal decomposition. The data preprocessing module is connected to the feature extraction module, which extracts deep features from the time domain and frequency domain and fuses them to obtain a comprehensive feature vector. The data preprocessing module is connected to a fully connected layer, which is used to synthesize feature vectors for deep information interaction; The fully connected layer is connected to the prediction module, which is used for prediction.
2. A prediction method for the power load prediction system as described in claim 1, characterized in that, Includes the following steps: Step 1: The data acquisition module collects raw load data; Step 2: The data preprocessing module preprocesses the raw load data to obtain a clean load signal, and decomposes the clean load signal to obtain trend components, seasonal components and residual components. Step 3: The feature extraction module marks the clean load signal as time-domain data and adds the seasonal components and residual components to convert it into frequency-domain data; Step 4: Extract time-domain deep feature vectors and frequency-domain deep feature vectors from time-domain data and frequency-domain data respectively, and then fuse the time-domain deep feature vectors and frequency-domain deep feature vectors to obtain a comprehensive feature vector that contains both time-domain and frequency-domain information; Step 5: Input the comprehensive feature vector into the fully connected layer. The fully connected layer performs deep information interaction and finally outputs the predicted power load values for the next n time points.
3. The prediction method according to claim 2, characterized in that, Step 2 is detailed below: Step 21: Periodically impute missing values and correct outliers in the original load data, and then normalize the data to obtain the normalized values. Step 22: Perform multi-scale decomposition on the normalized value using wavelet transform, and then perform threshold processing on the high-frequency detail components to filter out random noise and obtain a clean load signal. Step 23: Decompose the pure load signal into trend components, seasonal components, and residual components using the STL method.
4. The prediction method according to claim 3, characterized in that, Step 21 is as follows: Step 211: Arrange the historical power load data in chronological order to construct a dataset containing missing values; Step 212: Find missing values Y(T,t) The value at the same time on the day before and after Y(T-1,t), Y(T+1,t) and the value at the moment before and after the same day Y(T,t-1),Y(T,t+1) Then, a weighted average is performed to obtain the missing values. Y(T,t) The missing values are added to the dataset containing the missing values to obtain the complete dataset. Y(T,t) represents the missing values at time t on day T. Step 213: In the time dimension, determine whether the difference between the data in the dataset and its neighboring data is greater than a preset time dimension threshold. If the determination is no, proceed to step 214. If the determination is yes, use smoothing correction to perform correction, and then proceed to step 214. In the period dimension, determine whether the difference between the data in the dataset and its neighboring data is greater than a preset period dimension threshold. Points exceeding the preset threshold will be smoothed. If the determination is no, proceed to step 214. If the determination is yes, use smoothing correction to perform correction, and then proceed to step 214. Step 214: Perform max-min normalization on the data obtained in step 213, then map the data to the interval [0, 1] to eliminate the influence of dimensions and obtain the normalized value after processing.
5. The prediction method according to claim 4, characterized in that, The calculation method for the maximum-minimum normalization process is as follows: in, Normalized value The minimum value among the data obtained in step 213. The maximum value among the data obtained in step 213. The data obtained in step 213.
6. The prediction method according to claim 5, characterized in that, Step 22 is detailed below: Step 221: Use wavelet basis functions to perform multi-scale decomposition on the normalized value, decomposing it into a low-frequency approximate component and n high-frequency detail components at different scales; Step 222: Calculate high-frequency detail components ; Step 223: Calculate the high-frequency detail components By merging with the low-frequency approximate component, and reconstructing the denoised pure signal through inverse wavelet transform, the signal is denoised.
7. The prediction method according to claim 6, characterized in that, The wavelet basis functions are as follows: in, It is a normalized value. For wavelet basis functions, For scaling function, For high-frequency detail components, It is a low-frequency approximation component.
8. The prediction method according to claim 2, characterized in that, Step 222 is performed as follows: Step 2221: Calculate the noise intensity using the median absolute deviation method to obtain the standard deviation estimate sigma; Step 2222: Calculate the dynamic threshold λ according to the general threshold rule; Step 2223: For high-frequency detail components whose absolute value is less than λ, set them directly to zero; for high-frequency detail components whose absolute value is greater than λ, shrink them towards zero by a distance of λ. The formula is as follows: 。 9. The prediction method according to claim 2, characterized in that, The specific steps for adding the seasonal components and residual components to convert the data into frequency domain data are as follows: Step A: The seasonal component and the residual component are added point by point to obtain the high-frequency fluctuation signal; Step B: Use the Fast Fourier Transform algorithm to convert the high-frequency fluctuation signal into frequency domain data.
10. The prediction method according to claim 9, characterized in that, The specific operation of step B is as follows: the high-frequency fluctuation signal is sliced using a sliding window, the amplitude spectrum is calculated, and the first half of the amplitude spectrum is taken as the frequency domain data.