Day-ahead electricity price prediction framework, device and method based on time-frequency domain feature learning
By employing a multi-encoder architecture based on time-frequency domain feature learning and a customized loss function, the problems of noise suppression and feature redundancy in electricity price data in the electricity market are solved, achieving high-precision day-ahead electricity price forecasting and improving forecast accuracy and robustness.
Patent Information
- Application Number
- CN202511479850.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-16
- Publication Date
- 2026-01-30
AI Technical Summary
Existing technologies lack effective noise suppression mechanisms in the electricity market, electricity price data has redundant multi-dimensional features, long-sequence prediction models rely on large-scale datasets and ignore local features, and traditional loss functions are unable to capture the peak and valley features of electricity prices, resulting in insufficient accuracy in day-ahead electricity price prediction.
A time-frequency domain feature learning-based approach is adopted, including a multi-encoder and single-decoder architecture. It combines variational modulus decomposition (VMD) technology and the ReliefF algorithm for data preprocessing, utilizes patching technology and BiLSTM, TCN, and Transformer encoders for feature extraction and fusion, and optimizes prediction accuracy through a customized loss function.
It improved the accuracy of electricity price forecasting. Experimental verification showed that the accuracy rate increased by more than 12.46%, and the MSE decreased by 52.10% to 79.76%, enhancing the ability to capture electricity price volatility and the robustness of forecasting.
Smart Images
Figure CN121440591A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power system and artificial intelligence technology, specifically relating to a day-ahead electricity price forecasting framework, device and method based on time-frequency domain feature learning, which is applicable to scenarios such as strategic decision-making, power generation dispatch optimization and risk management for power market participants. Background Technology
[0002] Against the backdrop of electricity market liberalization reforms and the increasing complexity of energy systems, accurate day-ahead price forecasting is crucial for power generation companies to optimize dispatch, for grid operators to provide risk warnings, and for market participants to formulate trading strategies. However, the increasing penetration of renewable energy in power generation has led to stronger nonlinearity and volatility in electricity price sequences, posing significant challenges to day-ahead forecasting.
[0003] The existing technology has the following main shortcomings:
[0004] 1. Insufficient data preprocessing: There is a lack of effective noise suppression mechanisms, making it difficult to remove high-frequency noise from electricity price data, and there is redundant information in multi-dimensional features, which affects the model learning efficiency.
[0005] 2. Limitations of model architecture: Existing long sequence prediction models rely on large-scale datasets to learn long-term dependencies, but the training samples of electricity price data are limited; at the same time, they neglect the extraction of local features, and long input windows are prone to information loss.
[0006] 3. Limitations of the loss function: The traditional MSE loss function is difficult to capture the peak and trough characteristics of electricity prices, and the direct prediction paradigm ignores the autocorrelation of the label sequence, which limits the improvement of prediction accuracy. Summary of the Invention
[0007] To overcome the shortcomings of existing technologies, this invention proposes a day-ahead electricity price prediction framework, device, and method based on time-frequency domain feature learning, enabling accurate prediction of highly volatile electricity prices.
[0008] The technical solution adopted by this invention to solve its technical problem is: a day-ahead electricity price prediction method based on time-frequency domain feature learning, comprising the following steps:
[0009] The data collected includes electricity price data, total power, the proportion of renewable energy output, and real-time data on the bidding space for thermal power plants.
[0010] The real-time data is input into a pre-built electricity price prediction model to obtain the electricity price prediction result;
[0011] The electricity price prediction model is the EPformer model, which includes multiple encoders and a single decoder.
[0012] Preferably, before inputting the real-time data into a pre-built electricity price prediction model, the method includes:
[0013] The real-time data is preprocessed, including: using variational modular decomposition (VMD) technology to denoise the electricity price data, and using the ReliefF algorithm combined with expert domain knowledge to select electricity price features from the power grid operation data;
[0014] Accordingly, inputting the real-time data into the pre-built electricity price prediction model includes:
[0015] The preprocessed real-time data is input into the electricity price prediction model.
[0016] Preferably, the multi-encoder includes an independently encoded time encoder and a feature encoder, and the single decoder includes a predictive decoder for fusing the outputs of the time encoder and the feature encoder to predict electricity prices.
[0017] Preferably, the time encoder includes patching technology and a bidirectional long short-term memory neural network (BiLSTM) model, and the steps for using the time encoder are as follows:
[0018] S1. Input the electricity price data or the denoised electricity price data into the time encoder, and use the patch processing technology to divide the electricity price data into multiple review windows, and each review window into multiple patches;
[0019] S2. Input the patch into the shared fully connected layer, obtain the characteristics of each patch, and then perform feature fusion on all patch links.
[0020] Preferably, the core components of the feature encoder are a temporal convolutional network (TCN) and a Transformer encoder, and the steps for using the feature encoder are as follows:
[0021] S1. Input the electricity price data or the denoised electricity price data into the feature encoder. The TCN extracts features of various patterns from the electricity price sequence and integrates the time series features.
[0022] S2. Input the high-dimensional feature information extracted by TCN into the Transformer encoder, obtain the initial weight matrix through the multi-head self-attention mechanism, calculate the attention score matrix, normalize the attention score into attention weights, and extract features by weighted summation of the value matrix.
[0023] Preferably, the prediction decoder includes a feature fusion block, a BiLSTM prediction decoding model, and a Transformer decoder for decoding output. The steps for using the prediction decoder are as follows:
[0024] S1. Perform feature fusion on the electricity price features encoded by the time encoder and the multiple electricity price-related features obtained through screening;
[0025] S2. Input the fused features into the decoder. The BiLSTM model in the decoder uses the units and hidden states of the temporal encoder BiLSTM as the initial state and converts the fused feature input into an output sequence.
[0026] S3. Then, the output sequence of BiLSTM and the feature encoding output are jointly input into the self-attention mechanism layer of the Transformer decoder, so that the decoder can pay attention to the semantic information of the entire electricity price feature sequence output by the encoder when generating the prediction sequence.
[0027] S4, the self-attention mechanism layer outputs a feature matrix H, where the output W at each position is independently passed through two fully connected feedforward networks, enhancing the model's ability to capture complex patterns and relationships in the data and completing the decoding process of the Transformer decoder;
[0028] The output sequence of the S5 Transformer decoder is passed through a fully connected layer to obtain the predicted electricity price.
[0029] Preferably, the method further includes: a customized loss function based on a time-frequency domain feature learning strategy to improve the accuracy of the obtained predicted electricity price. The customized loss function is a time-domain loss function constructed by adding a loss penalty factor to the traditional time-domain loss function.
[0030] Preferably, improving the accuracy of the obtained predicted electricity price includes:
[0031] The data collected includes electricity price data, total power, the proportion of renewable energy output, and real-time data on the bidding space for thermal power plants.
[0032] The real-time data is preprocessed, including: using variational modular decomposition (VMD) technology to denoise the electricity price data, and using the ReliefF algorithm combined with expert domain knowledge to select electricity price features from the power grid operation data;
[0033] The preprocessed real-time data is input into a pre-built electricity price prediction model to obtain the electricity price prediction result;
[0034] A customized loss function based on a time-frequency domain feature learning strategy is used to improve the accuracy of the predicted electricity price. The customized loss function is constructed by adding a loss penalty factor to the traditional time-domain loss function to build an improved time-domain loss function. At the same time, a frequency-domain loss function is constructed to project the obtained day-ahead predicted electricity price sequence onto an orthogonal basis and convert it into frequency components, thereby reducing the interdependence caused by time autocorrelation. Finally, the improved time-domain loss function and the frequency-domain loss function are integrated to obtain the customized loss function based on the time-frequency domain feature learning strategy.
[0035] Preferably, a day-ahead electricity price forecasting device based on time-frequency domain feature learning includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-described day-ahead electricity price forecasting method based on time-frequency domain feature learning.
[0036] The denoised electricity price data is input into the time encoder. The electricity price data is divided into multiple review windows using patch processing technology, and each review window is divided into multiple patches. The patches are input into a shared fully connected layer. After obtaining the features of each patch, all patches are linked together to fuse the features. The fused features are input into the BiLSTM model, which goes through storage units, gating mechanisms, and state update processes.
[0037] The denoised electricity price data is input into the feature encoder. The TCN extracts features of various patterns from the electricity price sequence and integrates time series features. The high-dimensional feature information extracted by the TCN is input into the Transformer encoder. The initial weight matrix is obtained through the multi-head self-attention mechanism. The attention score matrix is calculated and the attention score is normalized into attention weights. The weighted summation of the value matrix is performed to extract features.
[0038] First, the electricity price features encoded by the time encoder and multiple electricity price-related features obtained through screening are fused to enrich the feature information;
[0039] The fused features are input into the decoder. The BiLSTM model in the decoder uses the units and hidden states of the temporal encoder BiLSTM as the initial state to convert the fused feature input into an output sequence.
[0040] The output sequence of BiLSTM and the feature encoding output are then jointly input into the self-attention mechanism layer of the Transformer decoder, enabling the decoder to focus on the semantic information of the entire electricity price feature sequence output by the encoder when generating the prediction sequence. The self-attention mechanism layer outputs a feature matrix H, where the output W at each position is independently passed through two fully connected feedforward networks, further enhancing the model's ability to capture complex patterns and relationships in the data, and completing the decoding process of the Transformer decoder.
[0041] Where FFnn(W) is the output of the feedforward network; max(0,) is the ReLU activation function; W1 and W2 are the learnable weight matrices; b1 and b2 are the learnable bias terms; the output sequence of the Transformer decoder is finally passed through a fully connected layer to obtain the predicted electricity price.
[0042] The advantages of this invention are:
[0043] 1. Multi-encoder-single-decoder architecture
[0044] Temporal encoder: Employs patching techniques to segment long input sequences and combines BiLSTM to capture long-distance temporal dependencies; Feature encoder: Utilizes TCN to extract local features and performs global feature encoding through a Transformer encoder; Prediction decoder: Fuses the outputs of the two encoders to generate the final prediction result.
[0045] 2. Joint loss function in the time and frequency domains
[0046] Time domain loss: A penalty factor is introduced on the basis of traditional MSE to focus on peak electricity prices; Frequency domain loss: FFT transformation is used to suppress autocorrelation of the label sequence and improve prediction accuracy; Joint optimization: Time and frequency domain features are learned collaboratively to enhance the ability to capture the periodicity and volatility of electricity prices.
[0047] 3. Technological advantages
[0048] Patch processing technology reduces the risk of gradient explosion in long sequences and preserves local feature details; feature fusion mechanism supplements market supply and demand information and improves prediction robustness; experimental verification shows that the accuracy is improved by more than 12.46% and the MSE is reduced by 52.10% to 79.76% compared with the baseline model. Attached Figure Description
[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0050] Figure 1 This is a schematic diagram of the day-ahead electricity price prediction framework based on time-frequency domain feature learning of the present invention. Detailed Implementation
[0051] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0052] Example 1
[0053] Please see Figure 1 As shown, a day-ahead electricity price forecasting method based on time-frequency domain feature learning includes the following steps:
[0054] The data collected includes electricity price data, total power, the proportion of renewable energy output, and real-time data on the bidding space for thermal power plants.
[0055] The real-time data is input into a pre-built electricity price prediction model to obtain the electricity price prediction result;
[0056] The electricity price prediction model is the EPformer model, which includes multiple encoders and a single decoder.
[0057] Furthermore, before inputting the real-time data into the pre-built electricity price prediction model, the method includes:
[0058] The real-time data is preprocessed, including: using variational modular decomposition (VMD) technology to denoise the electricity price data, and using the ReliefF algorithm combined with expert domain knowledge to select electricity price features from the power grid operation data;
[0059] Accordingly, inputting the real-time data into the pre-built electricity price prediction model includes:
[0060] The preprocessed real-time data is input into the electricity price prediction model.
[0061] Furthermore, the multi-encoder includes an independently encoded time encoder and a feature encoder, and the single decoder includes a predictive decoder for fusing the outputs of the time encoder and the feature encoder to predict electricity prices.
[0062] Furthermore, the time encoder includes patching technology and a bidirectional long short-term memory neural network (BiLSTM) model. The steps for using the time encoder are as follows:
[0063] S1. Input the electricity price data or the denoised electricity price data into the time encoder, and use the patch processing technology to divide the electricity price data into multiple review windows, and each review window into multiple patches;
[0064] S2. Input the patch into the shared fully connected layer, obtain the characteristics of each patch, and then perform feature fusion on all patch links.
[0065] Furthermore, the core components of the feature encoder are the Temporal Convolutional Network (TCN) and the Transformer encoder. The steps for using the feature encoder are as follows:
[0066] S1. Input the electricity price data or the denoised electricity price data into the feature encoder. The TCN extracts features of various patterns from the electricity price sequence and integrates the time series features.
[0067] S2. Input the high-dimensional feature information extracted by TCN into the Transformer encoder, obtain the initial weight matrix through the multi-head self-attention mechanism, calculate the attention score matrix, normalize the attention score into attention weights, and extract features by weighted summation of the value matrix.
[0068] Furthermore, the prediction decoder includes a feature fusion block, a BiLSTM prediction decoding model, and a Transformer decoder for decoding the output. The steps for using the prediction decoder are as follows:
[0069] S1. Perform feature fusion on the electricity price features encoded by the time encoder and the multiple electricity price-related features obtained through screening;
[0070] S2. Input the fused features into the decoder. The BiLSTM model in the decoder uses the units and hidden states of the temporal encoder BiLSTM as the initial state and converts the fused feature input into an output sequence.
[0071] S3. Then, the output sequence of BiLSTM and the feature encoding output are jointly input into the self-attention mechanism layer of the Transformer decoder, so that the decoder can pay attention to the semantic information of the entire electricity price feature sequence output by the encoder when generating the prediction sequence.
[0072] S4, the self-attention mechanism layer outputs a feature matrix H, where the output W at each position is independently passed through two fully connected feedforward networks, enhancing the model's ability to capture complex patterns and relationships in the data and completing the decoding process of the Transformer decoder;
[0073] The output sequence of the S5 Transformer decoder is passed through a fully connected layer to obtain the predicted electricity price.
[0074] Furthermore, the method also includes: a customized loss function based on a time-frequency domain feature learning strategy to improve the accuracy of the obtained predicted electricity price. The customized loss function is a time-domain loss function constructed by adding a loss penalty factor to the traditional time-domain loss function.
[0075] Furthermore, the improvement of the accuracy of the obtained predicted electricity price includes:
[0076] The data collected includes electricity price data, total power, the proportion of renewable energy output, and real-time data on the bidding space for thermal power plants.
[0077] The real-time data is preprocessed, including: using variational modular decomposition (VMD) technology to denoise the electricity price data, and using the ReliefF algorithm combined with expert domain knowledge to select electricity price features from the power grid operation data;
[0078] The preprocessed real-time data is input into a pre-built electricity price prediction model to obtain the electricity price prediction result;
[0079] A customized loss function based on a time-frequency domain feature learning strategy is used to improve the accuracy of the predicted electricity price. The customized loss function is constructed by adding a loss penalty factor to the traditional time-domain loss function to build an improved time-domain loss function. At the same time, a frequency-domain loss function is constructed to project the obtained day-ahead predicted electricity price sequence onto an orthogonal basis and convert it into frequency components, thereby reducing the interdependence caused by time autocorrelation. Finally, the improved time-domain loss function and the frequency-domain loss function are integrated to obtain the customized loss function based on the time-frequency domain feature learning strategy.
[0080] Furthermore, it includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described day-ahead electricity price forecasting method based on time-frequency domain feature learning.
[0081] The denoised electricity price data is input into the time encoder. The electricity price data is divided into multiple review windows using patch processing technology, and each review window is divided into multiple patches. The patches are input into a shared fully connected layer. After obtaining the features of each patch, all patches are linked together and the features are fused. The fused features are input into the BiLSTM model, which goes through storage units, gating mechanisms, and state update processes.
[0082] The denoised electricity price data is input into the feature encoder. The TCN extracts features of various patterns from the electricity price sequence and integrates time series features. The high-dimensional feature information extracted by the TCN is input into the Transformer encoder. The initial weight matrix is obtained through the multi-head self-attention mechanism. The attention score matrix is calculated and the attention score is normalized into attention weights. The weighted summation of the value matrix is performed to extract features.
[0083] First, the electricity price features encoded by the time encoder and multiple electricity price-related features obtained through screening are fused to enrich the feature information;
[0084] The fused features are input into the decoder. The BiLSTM model in the decoder uses the units and hidden states of the temporal encoder BiLSTM as the initial state to convert the fused feature input into an output sequence.
[0085] The output sequence of BiLSTM and the feature encoding output are then jointly input into the self-attention mechanism layer of the Transformer decoder, enabling the decoder to focus on the semantic information of the entire electricity price feature sequence output by the encoder when generating the prediction sequence. The self-attention mechanism layer outputs a feature matrix H, where the output W at each position is independently passed through two fully connected feedforward networks, further enhancing the model's ability to capture complex patterns and relationships in the data, and completing the decoding process of the Transformer decoder.
[0086] Where FFnn(W) is the output of the feedforward network; max(0,) is the ReLU activation function; W1 and W2 are the learnable weight matrices; b1 and b2 are the learnable bias terms; the output sequence of the Transformer decoder is finally passed through a fully connected layer to obtain the predicted electricity price.
[0087] Technology application scenarios and scalability
[0088] This invention is applicable to the formulation of daily trading strategies by electricity market participants, the optimization of dispatching by power generation companies, and the risk management of grid operators. The framework supports adapting to different market data characteristics by adjusting parameters such as patch length and encoder layers, and can integrate a real-time data update mechanism to provide decision support for the dynamic optimization of the power system.
[0089] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.
Claims
1. A day-ahead electricity price forecasting method based on time-frequency domain feature learning, characterized in that: The method comprises the following steps: Collecting real-time data including electricity price data, total power, new energy output proportion and thermal power bidding space; Inputting the real-time data into a pre-constructed electricity price prediction model to obtain an electricity price prediction result; The electricity price prediction model is an EPformer model, which comprises a plurality of encoders and a single decoder. 2.The day-ahead electricity price forecasting method based on time-frequency domain feature learning according to claim 1, characterized in that: Before inputting the real-time data into the pre-constructed electricity price prediction model, the method comprises: Preprocessing the real-time data, including denoising the electricity price data by using a variational mode decomposition (VMD) technique, and selecting electricity price features from power grid operation data by using a ReliefF algorithm combined with expert field knowledge; Correspondingly, the step of inputting the real-time data into the pre-constructed electricity price prediction model comprises: Inputting the preprocessed real-time data into the electricity price prediction model. 3.The day-ahead electricity price forecasting method based on time-frequency domain feature learning according to claim 1, characterized in that: The plurality of encoders comprises an independently coded time encoder and a feature encoder, and the single decoder comprises a prediction decoder for fusing the outputs of the time encoder and the feature encoder to perform electricity price prediction. 4.The day-ahead electricity price forecasting method based on time-frequency domain feature learning according to claim 3, characterized in that: The time encoder comprises a patch processing technique and a bidirectional long short-term memory (BiLSTM) neural network model, and the use steps of the time encoder are as follows: S1, inputting the electricity price data or the denoised electricity price data into the time encoder, and dividing the electricity price data into a plurality of review windows by using the patch processing technique, each review window being divided into a plurality of patches; S2, inputting the patches into a shared fully connected layer, obtaining features of each patch, and fusing the features of all the patches. 5.The day-ahead electricity price forecasting method based on time-frequency domain feature learning according to claim 1, characterized in that: The core component of the feature encoder is a time convolution network (TCN) and a Transformer encoder, and the use steps of the feature encoder are as follows: S1, inputting the electricity price data or the denoised electricity price data into the feature encoder, and extracting various mode features from the electricity price sequence by the TCN to integrate time sequence features; S2, inputting the high-dimensional feature information extracted by the TCN into the Transformer encoder, obtaining an initialized weight matrix through a multi-head self-attention mechanism, calculating an attention score matrix, normalizing the attention score, weighting and summing the value matrix, and extracting features. 6.The day-ahead electricity price forecasting method based on time-frequency domain feature learning according to claim 3, characterized in that: The prediction decoder comprises a feature fusion block, a BiLSTM prediction decoding model and a Transformer decoder for decoding output, and the use steps of the prediction decoder are as follows: S1, fusing the electricity price features coded by the time encoder and the plurality of electricity price related features selected; S2, inputting the fused features into the decoder, and taking the units and hidden states of the time encoder BiLSTM as initial states, and inputting the fused features into the decoder to convert the fused features into an output sequence; S3, inputting the output sequence of the BiLSTM and the feature coding output into the self-attention mechanism layer of the Transformer decoder, so that the decoder can focus on the semantic information of the entire electricity price feature sequence output by the encoder when generating a prediction sequence. S4, the self-attention mechanism layer outputs a feature matrix H, wherein the output W of each position is independently enhanced through two layers of fully connected feedforward networks, improving the ability of the model to capture complex patterns and relationships in the data, and completing the decoding process of the Transformer decoder; S5, the output sequence of the Transformer decoder is input into a fully connected layer to obtain a predicted electricity price. 7.The day-ahead electricity price forecasting method based on time-frequency domain feature learning according to claim 1, characterized in that: The method further comprises: improving the accuracy of the obtained predicted electricity price based on a customized loss function of a time-frequency domain feature learning strategy, wherein the customized loss function is a time domain loss function constructed by adding a loss penalty factor to a traditional time domain loss function. 8.The day-ahead electricity price forecasting method based on time-frequency domain feature learning according to claim 7, characterized in that: The accuracy improvement of the obtained predicted electricity price comprises: collecting real-time data including electricity price data, total power, new energy output ratio and thermal power bidding space; preprocessing the real-time data, including denoising the electricity price data using VMD technology and selecting electricity price features from power grid operation data using the ReliefF algorithm combined with expert domain knowledge; inputting the preprocessed real-time data into a pre-constructed electricity price prediction model to obtain an electricity price prediction result; improving the accuracy of the obtained predicted electricity price based on a customized loss function of a time-frequency domain feature learning strategy, wherein the customized loss function is an improved time domain loss function constructed by adding a loss penalty factor to a traditional time domain loss function, and a frequency domain loss function is also constructed, the obtained day-ahead predicted electricity price sequence is projected onto an orthogonal basis to convert it into frequency components, reducing the mutual dependence caused by time autocorrelation, and finally integrating the improved time domain loss function and the frequency domain loss function to obtain the customized loss function based on the time-frequency domain feature learning strategy.
9. A day-ahead electricity price forecasting framework based on time-frequency domain feature learning, characterized in that: A computer program product comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the day-ahead electricity price prediction method based on the time-frequency domain feature learning. 10.A day-ahead electricity price prediction device based on time-frequency domain feature learning, characterized in that: It comprises: a collection module for collecting real-time data including electricity price data, total power, new energy output ratio and thermal power bidding space; an input module for inputting the real-time data into a pre-constructed electricity price prediction model to obtain an electricity price prediction result; wherein the electricity price prediction model is an EPformer model comprising multiple encoders and a single decoder.