Load forecasting method based on error correction and fourier decomposition

Abstract

This invention discloses a load forecasting method based on error correction and Fourier decomposition. The forecasting model used includes a first forecasting branch and a second forecasting branch. The first forecasting branch obtains a load baseline forecasting error sequence based on historical continuous characteristics and historical weekday discrete characteristics. This sequence is de-standardized and then added point-by-point to the direct-regulation load baseline forecasting sequence to obtain a direct-regulation load forecasting sequence based on error correction. The second forecasting branch extracts low-frequency and high-frequency components from the historical load sequence through Fourier transform and forecasts them separately to obtain low-frequency and high-frequency load forecasting sequences. These two sequences are then added point-by-point to obtain a direct-regulation load forecasting sequence based on Fourier decomposition. The direct-regulation load forecasting sequence based on error correction and the direct-regulation load forecasting sequence based on Fourier decomposition are then weighted point-by-point to obtain the final direct-regulation load forecasting sequence. This method fully utilizes the baseline forecasting curve to ensure accurate forecasting direction while precisely capturing load fluctuations.

Description

Technical Field

[0001] This invention belongs to the field of power system load forecasting technology, specifically relating to a load forecasting method based on error correction and Fourier decomposition. Background Technology

[0002] In the context of market-based electricity operation and high-proportion integration of renewable energy, direct load forecasting affects not only grid dispatching plans but also market participants' judgments on supply and demand, electricity price trends, and trading positions. In actual operation, a benchmark load forecast curve published by the grid on the previous day usually exists, but this benchmark forecast is subject to systematic deviations due to factors such as changes in photovoltaic output, sudden weather changes, and cyclical changes in electricity consumption behavior.

[0003] Existing load forecasting models often directly predict load without fully utilizing existing benchmark forecast curves. They also lack directional constraints for correcting for loads "above or below the benchmark." When the model's prediction error direction is opposite to the actual deviation direction, even if the absolute error is small, it can lead to reverse judgments in market decisions. On the other hand, load sequences simultaneously contain information such as slowly changing components, intraday cycles, weekly cycles, sudden ramp-ups, and peak fluctuations. Existing models fail to effectively utilize this information, resulting in reduced forecast accuracy. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the technical problem this invention aims to solve is to provide a load forecasting method based on error correction and Fourier decomposition.

[0005] The present invention solves the aforementioned technical problem by adopting the following technical solution: A load forecasting method based on error correction and Fourier decomposition, characterized by the following steps: Step 1: Obtain the actual direct load sequence, the direct load benchmark prediction sequence, and the photovoltaic output prediction sequence for the target area, and further construct multi-source characteristic sequences, including the load benchmark prediction error sequence, the prediction error change sequence, the benchmark prediction change sequence, the photovoltaic ramp-up sequence, the clear sky index sequence, the hourly sine coding sequence, the hourly cosine coding sequence, and the weekday discrete characteristic sequence. Step 2: The multi-source feature sequence is sliced ​​using the sliding window method to generate multiple samples, resulting in a dataset. Each sample includes historical continuous features, future continuous features, historical weekday discrete features, and future weekday discrete features. Historical continuous features include the load baseline prediction error, prediction error change, baseline prediction change, photovoltaic ramp-up amount, clear sky index, hourly sine code, hourly cosine code, and actual direct-load load value for the historical time step. Future continuous features include the baseline prediction change, photovoltaic ramp-up amount, clear sky index, hourly sine code, and hourly cosine code for the future time step. Historical weekday discrete features and future weekday discrete features include the weekday discrete features for the historical time step and the future time step, respectively. Step 3: Construct a two-branch load forecasting model, including a first forecasting branch and a second forecasting branch; The first prediction branch includes a historical encoding side and a future encoding side. On the historical encoding side, the discrete features of historical days are embedded into dense vectors to obtain historical day-of-week embedding features. The historical day-of-week embedding features are concatenated with historical continuous features to obtain a historical fusion feature sequence. The historical fusion feature sequence is transposed and then passed through an dilated causal convolutional network for historical state encoding to obtain a historical state encoding sequence. On the future encoding side, the discrete features of future days are embedded into dense vectors to obtain future day-of-week embedding features. The future day-of-week embedding features are concatenated with future continuous features to obtain a future fusion feature sequence. The future fusion feature sequence is flattened and then encoded to obtain a future state encoding. The future state encoding is concatenated with the historical state encoding of the last historical time step to obtain a global state fusion feature. The global state fusion feature is mapped through two fully connected layers to obtain a load baseline prediction error sequence. The load baseline prediction error sequence is de-standardized and then added point-by-point to the direct-adjustment load baseline prediction sequence to obtain a direct-adjustment load prediction sequence based on error correction. In the second prediction branch, the column containing the actual value of the direct-regulation load is extracted from the historical continuous features to obtain the historical load sequence. A Fast Fourier Transform (FFT) is performed on the historical load sequence to obtain multiple frequency components. The three lowest frequency components are used to construct a low-frequency spectrum. An Inverse Fast Fourier Transform (IFFT) is performed on the low-frequency spectrum to obtain the low-frequency load sequence. The point-by-point difference between the historical load sequence and the low-frequency load sequence is calculated to obtain the high-frequency load sequence. The low-frequency load sequence is flattened to obtain a low-frequency feature vector. The future continuous features are flattened to obtain a future feature vector. The low-frequency feature vector and the future feature vector are concatenated and then subjected to a linear transformation to obtain the low-frequency load prediction sequence. The high-frequency load sequence is flattened to obtain a high-frequency feature vector. The high-frequency feature vector and the future feature vector are concatenated and then subjected to a linear transformation to obtain the high-frequency load prediction sequence. The low-frequency load prediction sequence and the high-frequency load prediction sequence are added point-by-point to obtain the direct-regulation load prediction sequence based on Fourier decomposition. Finally, the direct load forecast sequence based on error correction and the direct load forecast sequence based on Fourier decomposition are summed point by point to obtain the final direct load forecast sequence. Step 4: Use the dataset to train the dual-branch load forecasting model, and then use the trained model for direct load forecasting.

[0006] Furthermore, during the training process, the first prediction branch calculates the error correction loss according to equation (1); (1) In the formula, To correct for error loss, , They are respectively The predicted and actual load baseline forecast errors at each time step. To predict the total number of time steps; The direction penalty coefficient; The second forecasting branch calculates the load forecasting loss according to equations (2)-(6): (2) (3) (4) (5) (6) In the formula, For load forecasting losses, Based on the mean square error loss, For peak losses during high load periods, The loss is due to the consistency of the distribution between the load forecast sequence and the actual load sequence. For the gradient loss of load change, , They are respectively The predicted load value at each time step versus the actual load value. , They are respectively The predicted load value at each time step versus the actual load value. , After sorting the load forecast sequence and the actual load sequence in ascending order respectively The predicted load value at each time step versus the actual load value. To take the absolute value; for Peak mask at time step, when When the actual load at a time step is higher than the peak threshold, ,otherwise ; , , All are weighting coefficients.

[0007] Furthermore, the extended causal convolutional network includes multiple residual convolutional blocks. Each residual convolutional block includes a first one-dimensional convolutional layer, a first length pruning layer, a first activation layer, a first random dropout layer, a second one-dimensional convolutional layer, a second length pruning layer, a second activation layer, and a second random dropout layer. The output features of the second random dropout layer are residually concatenated with the input features of the residual convolutional block to obtain the output features of the residual convolutional block.

[0008] Compared with the prior art, the present invention has at least the following beneficial effects: 1. The first forecasting branch uses the direct-regulation load benchmark forecast as the anchor point. By learning the benchmark forecast error, the load benchmark forecast is corrected to obtain the load forecast sequence. It makes full use of the load benchmark forecast curve disclosed by the power grid before the day to improve the forecast accuracy. It also reduces the difficulty of the model fitting the load from scratch. At the same time, the error loss with directional penalty can ensure the directional consistency of the benchmark forecast error and avoid reverse judgment in market decision-making caused by reverse forecast.

[0009] The second prediction branch explicitly extracts low-frequency and high-frequency components from historical load sequences using Fourier frequency domain filtering, resulting in low-frequency and high-frequency load sequences. These sequences represent the slower-changing parts of the overall load and the intraday fluctuations, peak-valley changes, and faster-changing parts, respectively. By distinguishing between slow load changes and periodic fluctuations on a frequency scale, the model can better capture load fluctuations.

[0010] 2. The first prediction branch models historical and future states separately. It fully utilizes features from historical time steps, such as actual direct-load values, error predictions, and baseline predictions, to encode historical states. It also utilizes features from future time steps, such as baseline prediction changes, photovoltaic ramp-up amounts, and clear-sky indexes, to encode future states, representing constraints such as baseline load trends, photovoltaic output changes, and electricity consumption cycles within the prediction period. By fusing historical and future state encodings, the model can simultaneously consider the evolution of historical baseline prediction errors and changes in future operating conditions when predicting load baseline prediction errors. This improves the model's adaptability to load ramp-up, photovoltaic fluctuations, intraday cycles, and weekday differences, thereby enhancing the accuracy and stability of load error correction. Furthermore, this modeling approach clearly distinguishes between historically observable information and future pre-obtainable information, avoiding the use of future actual loads and future true errors during the prediction process, thus reducing the risk of future information leakage. Attached Figure Description

[0011] Figure 1 This is the overall flowchart; Figure 2The flowchart for the first prediction branch; Figure 3 This is the flowchart for the second prediction branch. Detailed Implementation

[0012] Specific embodiments are given below with reference to the accompanying drawings. These specific embodiments are only used to describe the technical solution of the present invention in detail and are not intended to limit the scope of protection of this application.

[0013] like Figure 1-3 As shown, this invention provides a load forecasting method based on error correction and Fourier decomposition, comprising the following steps: Step 1: Using hourly sampling intervals, obtain the actual direct-load sequence, the direct-load baseline prediction sequence, and the photovoltaic output prediction sequence for the target area; based on these sequences, further construct multi-source feature sequences, including the load baseline prediction error sequence, the prediction error change sequence, the baseline prediction change sequence, the photovoltaic ramp-up sequence, the clear sky index sequence, the hourly sine coding sequence, the hourly cosine coding sequence, and the weekday discrete feature sequence. Specifically, the load baseline prediction error is obtained by subtracting the actual value of the direct-controlled load from the baseline prediction value at the same time step, which represents the deviation between the existing baseline prediction and the actual operating result; the prediction error change is obtained by subtracting the load baseline prediction error of the current time step from the load baseline prediction error of the previous time step, which represents the direction and magnitude of the change in the baseline prediction deviation; the baseline prediction change is obtained by subtracting the direct-controlled load baseline prediction value of the current time step from the direct-controlled load baseline prediction value of the previous time step, which represents the rise or fall of the baseline prediction curve; and the photovoltaic ramp-up is obtained by subtracting the photovoltaic output baseline prediction value of the current time step from the photovoltaic output baseline prediction value of the previous time step. The value represents the increase or decrease in photovoltaic power output forecast; the maximum photovoltaic power output forecast value within the previous 30 days is used as the surrogate value of installed capacity; for data with less than 30 days at the beginning of the sequence, the first subsequent surrogate value of installed capacity is used to supplement it; the cosine value of the solar zenith angle is calculated based on the latitude and longitude of the target region and the time corresponding to each time step, while the negative values ​​of the solar zenith angle cosine value generated when the sun is below the horizon are adjusted to zero; the solar zenith angle cosine value is multiplied by the surrogate value of installed capacity to obtain the theoretical value of photovoltaic power output; the baseline forecast value of photovoltaic power output is divided by the theoretical value of photovoltaic power output to obtain the clear sky index, while the clear sky index is limited to the range (0, 1.5] to reduce the impact of outlier values.

[0014] The week number of the time step is determined based on the timestamp, and the week discrete feature is obtained to characterize the periodic differences such as weekdays and weekends. At the same time, sine and cosine encodings are performed according to the position of the time step in the 24 hours of a day to continuously represent the periodic relationship between adjacent time steps and the first and last time steps across days.

[0015] Step 2: Construct samples using the sliding window method to obtain the dataset; Using a historical window of 168 hours, a future window of 48 hours, and a sliding step of 24 hours, the multi-source feature sequence is sliced ​​to generate multiple samples, each of which includes continuous historical features. Future continuous characteristics Historical weekday discrete characteristics Discrete characteristics of future weeks ; Historical continuous features include eight features for the historical time step: load baseline forecast error, forecast error change, baseline forecast change, photovoltaic ramp-up, clear sky index, hourly sine code, hourly cosine code, and actual value of direct load. Future continuous features include five features for the future time step: baseline forecast change, photovoltaic ramp-up, clear sky index, hourly sine code, and hourly cosine code. Historical weekday discrete features and future weekday discrete features include weekday discrete features for the historical time step and future time step, respectively.

[0016] The third step is to construct a dual-branch load forecasting model, which mainly includes a first forecasting branch based on error correction and a second forecasting branch based on Fourier decomposition. The two forecasting branches share historical continuous features and future continuous features. The first forecasting branch corrects the load baseline forecast by learning the deviation between the actual load value and the forecast value, thus obtaining a direct-load forecasting sequence based on error correction. The second forecasting branch is used to directly forecast the direct-load, thus obtaining a direct-load forecasting sequence based on Fourier decomposition.

[0017] The first prediction branch includes a historical encoding side and a future encoding side. The input to the historical encoding side includes historical continuous features and historical weekday discrete features. The historical weekday discrete features are mapped into a 3D dense vector through a weekday embedding layer, resulting in 168×3 historical weekday embedding features. The historical weekday embedding features are concatenated with historical continuous features to obtain a 168×11 historical fusion feature sequence. To adapt to the input format of a one-dimensional temporal convolutional network, the 168×11 historical fusion feature sequence is transposed to 11×168. The transposed historical fusion feature sequence is then encoded into historical states through an dilated causal convolutional network, resulting in a 64×168 historical state encoding sequence. The historical state encoding at the last time step represents the historical comprehensive state within 168 hours before the prediction starting point. The input to the future encoding side includes future continuous features and future weekday discrete features. The future weekday discrete features are mapped into a 3D dense vector through a weekday embedding layer, resulting in 48×3 future weekday embedding features. The future weekday embedding features are concatenated with the future continuous features to obtain a 48×8 future fusion feature sequence. After flattening the future fusion feature sequence, it is passed through a future feature encoding layer that includes linear transformation, activation function, and random dropout operation to obtain a 128×1 future state code. The 64×1 historical state code is concatenated with the 128×1 future state code to form a 192×1 global state fusion feature. This global state fusion feature is passed through two fully connected layers, first mapped to 64×1 latent features, and then mapped to a 48×1 load baseline prediction error sequence. After inverse standardization, the load baseline prediction error sequence is added point-by-point to the direct-adjustment load baseline prediction sequence to correct the load baseline prediction, resulting in an error-corrected direct-adjustment load prediction sequence.

[0018] The dilated causal convolutional network is composed of multiple stacked residual convolutional blocks. This embodiment includes 7 residual convolutional blocks with output channels of 32, 32, 32, 32, 32, 64, and 64 respectively. The kernel length is 3, and the dilation factor increases with the number of network layers. Each residual convolutional block includes a first one-dimensional convolutional layer, a first length-pruned layer, a first activation layer, a first random dropout layer, a second one-dimensional convolutional layer, a second length-pruned layer, a second activation layer, and a second random dropout layer. The output features of the second random dropout layer are residually concatenated with the input features of the residual convolutional block to obtain the output features of the residual convolutional block. The length-pruned layer is used to remove redundant time positions generated by convolution padding, ensuring that each layer's output still maintains 168 time steps and preventing future information leakage.

[0019] The second prediction branch directly predicts the direct-load using only historical and future continuous features. It extracts the columns containing the actual direct-load values ​​from the historical continuous features, obtaining a historical load sequence with a dimension of 168×1. A fast Fourier transform is performed on the historical load sequence to obtain the original spectrum, which includes multiple frequency components arranged from low to high. The first frequency component represents the overall average load level over the past 168 hours, while the remaining low-frequency components represent the slower-changing parts of the load, and the high-frequency components represent the faster-changing parts. The first three frequency components are used to construct a low-frequency spectrum with the same dimension as the original spectrum, and all other frequency positions in the low-frequency spectrum are set to zero. An inverse fast Fourier transform is performed on the low-frequency spectrum to obtain a low-frequency load sequence with a dimension of 168×1. The historical load sequence and the low-frequency load sequence are subtracted point-by-point to obtain a high-frequency load sequence with a dimension of 168×1. The low-frequency load sequence represents the slower-changing parts of the direct-load, while the high-frequency load sequence represents the intraday fluctuations, peak-valley variations, and faster-changing parts of the direct-load. Furthermore, the point-by-point addition of the low-frequency and high-frequency load sequences can reconstruct the historical load sequence before decomposition.

[0020] On the one hand, the 168×1 low-frequency load sequence is flattened into a 168×1 low-frequency feature vector, and the 48×5 future continuous feature vector is flattened into a 240×1 future feature vector. The low-frequency feature vector and the future feature vector are concatenated and then subjected to a linear transformation to obtain the low-frequency load prediction sequence. On the other hand, the 168×1 high-frequency load sequence is flattened into a 168-dimensional high-frequency feature vector. The high-frequency feature vector and the future feature vector are concatenated and then subjected to a linear transformation to obtain the high-frequency load prediction sequence. The low-frequency load prediction sequence and the high-frequency load prediction sequence are added point by point to obtain the direct-adjustment load prediction sequence based on Fourier decomposition.

[0021] The direct-load forecast sequence, based on error correction and output from the first forecast branch, and the direct-load forecast sequence based on Fourier decomposition and output from the second forecast branch, are summed point-by-point with weights to obtain the final direct-load forecast sequence. By adjusting the weights of the two branches, a balance can be struck between emphasizing baseline forecast error correction and emphasizing low-frequency and high-frequency load patterns.

[0022] Step 4: Train the dual-branch load forecasting model using the dataset, and then use the trained model for direct load forecasting. During the training process, the first prediction branch uses the true value of the load baseline prediction error as the target and calculates the loss using the error correction loss function of equation (1); (1) In the formula, To correct for error loss, , They are respectively The predicted and actual load baseline forecast errors at each time step. To predict the total number of time steps; The direction penalty coefficient is set to 2 in this embodiment; When the prediction error is in the same direction as the actual error and When the signs are the same, the direction penalty term is 0; when the directions are opposite, the direction penalty term is 0. and If the sign is opposite, an additional penalty is incurred, thereby improving the consistency of the sign between the load baseline forecast error and the actual value in the standardized error space.

[0023] The second forecasting branch targets the actual direct load value and employs constraints including baseline error, peak error, sequence distribution differences, and consistency of changes between adjacent time steps. Therefore, the load forecasting loss function is: (2) (3) (4) (5) (6) In the formula, For load forecasting losses, Based on the mean square error loss, For peak losses during high load periods, The loss is due to the consistency of the distribution between the load forecast sequence and the actual load sequence. For the gradient loss of load change, , They are respectively The predicted load value at each time step versus the actual load value. , They are respectively The predicted load value at each time step versus the actual load value. , After sorting the load forecast sequence and the actual load sequence in ascending order respectively The predicted load value at each time step versus the actual load value. To take the absolute value; for The peak mask for each time step uses the 80th percentile of all actual load values ​​in the current training batch as the peak threshold. When the actual load value at a certain time step is higher than this peak threshold, it indicates that the time step is in the top 20% of the load range, and this is considered a peak load period, with the peak mask applied accordingly. Otherwise ; , , All of these are weighting coefficients, which are set to 0.03, 0.03 and 0.30 respectively in this embodiment.

[0024] The first prediction branch uses the Adam optimizer with a learning rate of 0.0003 and is trained for 25 epochs; the second prediction branch uses the Adam optimizer with a learning rate of 0.0005 and is trained for 30 epochs; the batch size is set to 32; each branch trains 3 sets of models to form an ensemble, and the corresponding model parameters are saved when the validation loss is minimized.

[0025] Example This embodiment uses a rolling forecasting method to test the load forecasting effect in different test intervals. Each test interval covers all hourly data within the corresponding natural month. The GRU model is selected as the comparison model, and the directional accuracy is used as the evaluation index to measure the consistency between the model's predicted load deviation direction and the actual load deviation direction.

[0026] Table 1 Rolling Forecast Results

[0027] As shown in Table 1, the orientation accuracy of the method of the present invention varies in different test intervals. In the test intervals of September 2025, October 2025, November 2025, December 2025, February 2026, and March 2026, the orientation accuracy of the method of the present invention is higher than that of the GRU model. In September 2025 and October 2025, the orientation accuracy of the method of the present invention reached 64.86% and 64.78%, respectively, which are 8.05 percentage points and 8.46 percentage points higher than that of the GRU model. In December 2025 and February 2026, the orientation accuracy of the method of the present invention reached 68.55% and 68.60%, respectively. Overall, the method of this invention achieved a higher directional accuracy than the GRU model in 6 out of 7 test intervals, indicating that by combining baseline prediction error correction and Fourier decomposition, the method of this invention can accurately capture the direction of baseline prediction deviation, as well as the slower-than-usual parts of overall load change, intraday fluctuations, peak-valley changes, and faster-than-usual changes, thereby improving the accuracy of load fluctuation fitting and thus enhancing the model's ability to predict the direction of load deviation.

[0028] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A load forecasting method based on error correction and Fourier decomposition, characterized in that, Includes the following steps: Step 1: Obtain the actual direct load sequence, the direct load benchmark prediction sequence, and the photovoltaic output prediction sequence for the target area, and further construct multi-source characteristic sequences, including the load benchmark prediction error sequence, the prediction error change sequence, the benchmark prediction change sequence, the photovoltaic ramp-up sequence, the clear sky index sequence, the hourly sine coding sequence, the hourly cosine coding sequence, and the weekday discrete characteristic sequence. Step 2: The multi-source feature sequence is sliced ​​using the sliding window method to generate multiple samples, resulting in a dataset. Each sample includes historical continuous features, future continuous features, historical weekday discrete features, and future weekday discrete features. Historical continuous features include the load baseline prediction error, prediction error change, baseline prediction change, photovoltaic ramp-up amount, clear sky index, hourly sine code, hourly cosine code, and actual direct-load load value for the historical time step. Future continuous features include the baseline prediction change, photovoltaic ramp-up amount, clear sky index, hourly sine code, and hourly cosine code for the future time step. Historical weekday discrete features and future weekday discrete features include the weekday discrete features for the historical time step and the future time step, respectively. Step 3: Construct a two-branch load forecasting model, including a first forecasting branch and a second forecasting branch; The first prediction branch includes a historical encoding side and a future encoding side. On the historical encoding side, the discrete features of historical days are embedded into a dense vector to obtain the historical day embedding features. The historical weekday embedding features are concatenated with historical continuous features to obtain a historical fused feature sequence. This historical fused feature sequence is then transposed and passed through an dilated causal convolutional network for historical state encoding, resulting in a historical state encoding sequence. On the future encoding side, the future weekday discrete features are embedded into a dense vector to obtain future weekday embedding features. These future weekday embedding features are concatenated with future continuous features to obtain a future fused feature sequence. This future fused feature sequence is flattened and then encoded to obtain the future state encoding. Finally, the future state encoding is concatenated with the historical state encoding of the last historical time step to obtain the global state fused feature. The global state fusion features are mapped through two fully connected layers to obtain the load baseline prediction error sequence. After inverse standardization, the load baseline prediction error sequence is added point by point to the direct-adjustment load baseline prediction sequence to obtain the direct-adjustment load prediction sequence based on error correction. In the second prediction branch, the column containing the actual value of the direct-regulation load is extracted from the historical continuous features to obtain the historical load sequence. A Fast Fourier Transform (FFT) is performed on the historical load sequence to obtain multiple frequency components. The three lowest frequency components are used to construct a low-frequency spectrum. An Inverse Fast Fourier Transform (IFFT) is performed on the low-frequency spectrum to obtain the low-frequency load sequence. The point-by-point difference between the historical load sequence and the low-frequency load sequence is calculated to obtain the high-frequency load sequence. The low-frequency load sequence is flattened to obtain a low-frequency feature vector. The future continuous features are flattened to obtain a future feature vector. The low-frequency feature vector and the future feature vector are concatenated and then subjected to a linear transformation to obtain the low-frequency load prediction sequence. The high-frequency load sequence is flattened to obtain a high-frequency feature vector. The high-frequency feature vector and the future feature vector are concatenated and then subjected to a linear transformation to obtain the high-frequency load prediction sequence. The low-frequency load prediction sequence and the high-frequency load prediction sequence are added point-by-point to obtain the direct-regulation load prediction sequence based on Fourier decomposition. Finally, the direct load forecast sequence based on error correction and the direct load forecast sequence based on Fourier decomposition are summed point by point to obtain the final direct load forecast sequence. Step 4: Use the dataset to train the dual-branch load forecasting model, and then use the trained model for direct load forecasting.

2. The load forecasting method based on error correction and Fourier decomposition according to claim 1, characterized in that, During training, the first prediction branch calculates the error correction loss according to equation (1); （1） In the formula, To correct for error loss, , They are respectively The predicted and actual load baseline forecast errors at each time step. To predict the total number of time steps; The direction penalty coefficient; The second forecasting branch calculates the load forecasting loss according to equations (2)-(6): （2） （3） （4） （5） （6） In the formula, For load forecasting losses, Based on the mean square error loss, For peak losses during high load periods, The loss is the distribution consistency loss between the load forecast sequence and the actual load sequence. For the gradient loss of load change, , They are respectively The predicted load value at each time step versus the actual load value. , They are respectively The predicted load at each time step versus the actual load. , After sorting the load forecast sequence and the actual load sequence in ascending order respectively The predicted load value at each time step versus the actual load value. To take the absolute value; for Peak mask at time step, when When the actual load at a time step is higher than the peak threshold, ,otherwise ; , , All are weighting coefficients.

3. The load forecasting method based on error correction and Fourier decomposition according to claim 1 or 2, characterized in that, The extended causal convolutional network includes multiple residual convolutional blocks. Each residual convolutional block includes a first one-dimensional convolutional layer, a first length pruning layer, a first activation layer, a first random dropout layer, a second one-dimensional convolutional layer, a second length pruning layer, a second activation layer, and a second random dropout layer. The output features of the second random dropout layer are residually concatenated with the input features of the residual convolutional block to obtain the output features of the residual convolutional block.

Images