A fine-grained collaborative forecasting method for power load

Abstract

The present invention discloses a fine-grained collaborative prediction method for electric loads, the method comprising: obtaining multivariate time series input data; slicing the input data to obtain multiple data patches; processing the data patches through dynamic weight variables and a short-term time series feature extraction module to obtain enhanced feature representations; inputting the enhanced feature representations into an encoder to generate coding features; inputting the data patches into a linear prediction layer to obtain a first prediction component; generating a second prediction component based on the enhanced feature representation and the coding features; and obtaining a collaborative prediction result for electric loads based on the fusion of the first prediction component and the second prediction component. The present invention can effectively collaboratively model the complex dependencies of time and variable dimensions through AVSTFE, channel-independent Transformer, and dual-path prediction heads, overcoming the limitation of existing methods that are difficult to balance the two, thereby significantly improving the prediction accuracy of multivariate time series, especially fine-grained electric loads.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a fine-grained power load collaborative prediction method. Background Art

[0002] The huge population base creates a massive demand for electricity. Electricity generation, transportation, and consumption are all simultaneous processes, and electricity demand and population distribution vary widely from region to region. This creates significant challenges in meeting people's electricity needs. Accurately forecasting electricity data at all levels is key to solving the problem of power energy allocation. Fine-grained load forecasting, in particular, can provide control and decision-making support for power allocation and management, avoiding unplanned power production and irrational allocation, and significantly alleviating regional electricity demand pressures.

[0003] Fine-grained forecasting of power load faces significant challenges. Research on coarse-grained forecasting of power load data has yielded significant results. Coarse-grained forecasting of the same data requires significantly less data than fine-grained forecasting, and the data also contains less information and detail. Coarse-grained forecasting requires lower accuracy and detail than fine-grained forecasting. Coarse-grained test sets cannot evaluate the quality of forecast results at a more microscopic and detailed level, resulting in a coarse-grained forecast that often easily achieves good results. However, real-world electricity consumption fluctuates rapidly, and coarse-grained forecasting, due to its loss of data detail and greater granularity, cannot effectively meet the needs of power load forecasting. Providing real-time, detailed decision-making for power production and dispatch requires fine-grained forecasting with a finer granularity. Compared to coarse-grained forecasting, fine-grained forecasting uses finer data granularity, which translates to more detailed information and more anomalies. Furthermore, capturing power trends at a fine granularity makes it more difficult to capture power data. These factors pose significant challenges to fine-grained forecasting.

[0004] For such fine-grained multivariable power load forecasting tasks (usually including the load itself as well as multiple related variables such as voltage, current, and meteorology), existing forecasting methods still have limitations in processing data. Although deep learning-based methods, especially the Transformer model, have shown potential in processing time series, they still have shortcomings in co-processing the temporal characteristics within the sequence (such as long-term and short-term dependencies, periodicity) and the interactions between different power-related variables. Some models sacrifice the processing power or computational efficiency of long-term time series patterns in order to capture the relationship between variables (channel dependence), while other models adopt a channel-independent strategy to focus on temporal characteristics, but may ignore the key interactions between variables, such as the possible lag in the impact of voltage fluctuations on load. This inability to effectively balance and co-model time and variable dimension information limits the accuracy and reliability of existing methods in fine-grained power forecasting scenarios, especially when longer-term forecasts are required. Summary of the Invention

[0005] An embodiment of the present invention provides a fine-grained power load collaborative forecasting method to solve the problem in the existing technology that it is impossible to effectively balance and collaboratively model time and variable dimension information, which limits the accuracy and reliability of the existing method in fine-grained power forecasting scenarios when longer-term forecasts are required.

[0006] In one aspect, an embodiment of the present invention provides a fine-grained power load collaborative forecasting method, comprising:

[0007] Get multivariate time series input data;

[0008] Slicing the input data to obtain a plurality of data patches;

[0009] Processing the data patch by a dynamic weight variable and a short-term temporal feature extraction module to obtain an enhanced feature representation;

[0010] Inputting the enhanced feature representation into an encoder to generate encoded features;

[0011] Inputting the data patch into a linear layer to obtain a first prediction component;

[0012] generating a second prediction component based on the enhanced feature representation and the encoded feature;

[0013] The power load collaborative prediction result is obtained by fusing the first prediction component and the second prediction component.

[0014] In a possible implementation, slicing the input data to obtain a plurality of data patches includes:

[0015] Preprocessing the input data to obtain processed data;

[0016] The processed data is separated in time sequence to obtain a plurality of data patches.

[0017] In a possible implementation, preprocessing the input data to obtain processed data includes:

[0018] The input data is subjected to resampling processing and missing value mean supplement processing to obtain the processed data.

[0019] In a possible implementation, the processing of the data patch by using a dynamic weight variable and a short-term temporal feature extraction module to obtain an enhanced feature representation further includes:

[0020] Separating the data patches according to the parity of time steps;

[0021] Performing a value embedding operation in the data patch to restore the dimension of the data patch;

[0022] At least two two-dimensional convolutional layers with different convolution kernel sizes are used to capture the local temporal patterns of the data patch and the interaction characteristics between variables from different receptive field scales;

[0023] The interactive features include features that are correlated across time steps;

[0024] Obtaining global features of the data patch by combining average pooling and maximum pooling operations;

[0025] A one-dimensional convolutional layer is used to process the global features to calculate the dynamic weights of each variable.

[0026] In a possible implementation, inputting the enhanced feature representation into an encoder to generate encoding features includes:

[0027] The enhanced feature representation is input into a channel-independent Transformer encoder for time series processing to obtain encoded features.

[0028] In a possible implementation, inputting the data patch into a linear prediction layer to obtain a first prediction component includes:

[0029] The sliced ​​data patch is input into a linear projection layer to extract and predict the long-term trend component as the first prediction component.

[0030] In a possible implementation, generating a second prediction component according to the enhanced feature representation and the encoding feature includes:

[0031] The coding feature is subjected to nonlinear transformation extraction by a multi-layer perceptron to obtain a nonlinear dynamic component as the second prediction component.

[0032] In a possible implementation, after obtaining the power load collaborative prediction result by fusing the first prediction component and the second prediction component, the method further includes:

[0033] The power load collaborative forecasting result is inversely normalized to restore the original data range of the power load collaborative forecasting result.

[0034] The fine-grained power load collaborative forecasting method of the present invention has the following advantages:

[0035] Through AVSTFE, channel-independent Transformer, and dual-path prediction head, we can effectively and collaboratively model the complex dependencies between time and variable dimensions, overcoming the limitations of existing methods that find it difficult to balance the two, thereby significantly improving the prediction accuracy of multivariate time series, especially fine-grained power loads. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0037] Figure 1 A flowchart of a fine-grained power load collaborative forecasting method provided in an embodiment of the present application;

[0038] Figure 2 A schematic diagram of the overall framework of the ChannelMixer method for a fine-grained power load collaborative forecasting method provided in an embodiment of the present application;

[0039] Figure 3 A schematic diagram of the structure of a dynamic weight variable and short-term temporal feature extraction (AVSTFE) module of a fine-grained power load collaborative forecasting method provided in an embodiment of the present application;

[0040] Figure 4 A schematic diagram of the workflow of the odd-even slice transformation (CoST) unit in the AVSTFE module of a fine-grained power load collaborative forecasting method provided in an embodiment of the present application;

[0041] Figure 5 A schematic diagram of the structure of a dynamic weight extraction (AVIE) module in an AVSTFE module of a fine-grained power load collaborative forecasting method provided in an embodiment of the present application;

[0042] Figure 6A schematic diagram of the attention dimension of a Transformer encoder for a fine-grained power load collaborative prediction method provided in an embodiment of the present application;

[0043] Figure 7 A comparison chart of the prediction results of a fine-grained power load collaborative forecasting method provided in an embodiment of the present application with a prediction length of 96 on an ETTm1 dataset, the true value, and other comparison methods;

[0044] Figure 8 A comparison chart of the prediction results of a fine-grained power load collaborative forecasting method provided in an embodiment of the present application with a prediction length of 720 on a Weather dataset, the true value, and other comparison methods. DETAILED DESCRIPTION

[0045] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0046] Figure 1 A flowchart of a fine-grained power load collaborative forecasting method provided by an embodiment of the present invention is provided. The embodiment of the present invention provides a fine-grained power load collaborative forecasting method, including:

[0047] Get multivariate time series input data;

[0048] Slicing the input data to obtain a plurality of data patches;

[0049] Processing the data patch by a dynamic weight variable and a short-term temporal feature extraction module to obtain an enhanced feature representation;

[0050] Inputting the enhanced feature representation into an encoder to generate encoded features;

[0051] Inputting the data patch into a linear layer to obtain a first prediction component;

[0052] generating a second prediction component based on the enhanced feature representation and the encoded feature;

[0053] The power load collaborative prediction result is obtained by fusing the first prediction component and the second prediction component.

[0054] The slicing of the input data to obtain a plurality of data patches comprises:

[0055] Preprocessing the input data to obtain processed data;

[0056] The processed data is separated in time sequence to obtain a plurality of data patches.

[0057] The preprocessing of the input data to obtain processed data comprises:

[0058] The input data is subjected to resampling processing and missing value mean supplement processing to obtain the processed data.

[0059] The step of processing the data patch by using the dynamic weight variable and the short-term time series feature extraction module to obtain the enhanced feature representation further comprises:

[0060] Separating the data patches according to the parity of time steps;

[0061] Performing a value embedding operation in the data patch to restore the dimension of the data patch;

[0062] At least two two-dimensional convolutional layers with different convolution kernel sizes are used to capture the local temporal pattern of the data patch and the interaction characteristics between variables from different receptive field scales;

[0063] The interactive features include features that are correlated across time steps;

[0064] Obtaining global features of the data patch by combining average pooling and maximum pooling operations;

[0065] A one-dimensional convolutional layer is used to process the global features to calculate the dynamic weights of each variable.

[0066] Inputting the enhanced feature representation into an encoder to generate encoding features comprises:

[0067] The enhanced feature representation is input into a channel-independent Transformer encoder for time series processing to obtain encoded features.

[0068] Inputting the data patch into a linear prediction layer to obtain a first prediction component comprises:

[0069] The sliced ​​data patch is input into a linear projection layer to extract and predict the long-term trend component as the first prediction component.

[0070] Generating a second prediction component according to the enhanced feature representation and the encoded feature comprises:

[0071] The coding feature is subjected to nonlinear transformation and extracted by a multi-layer perceptron to obtain a nonlinear dynamic component as the second prediction component.

[0072] After obtaining the power load collaborative prediction result by fusing the first prediction component and the second prediction component, the method further includes:

[0073] The power load collaborative forecasting result is inversely normalized to restore the original data range of the power load collaborative forecasting result.

[0074] Exemplarily, this application proposes a ChannelMixer method, which aims to effectively and collaboratively model the temporal features and variable relationships in multivariate time series to improve prediction accuracy.

[0075] Reference Figure 2 ,The overall framework of the ChannelMixer model is mainly composed of four core functional modules:

[0076] Input and Slicing Module: Receive Multivariate Time Series Data (Where L is the sequence length and V is the number of variables). First, the input data Preprocessing is performed using resampling and mean filling.

[0077] Specifically, we obtain multivariate time series input data and slice it to obtain multiple data patches. The preprocessing step may include resampling at a specified granularity and filling missing values ​​(such as mean filling). Then, we divide the preprocessed input time series data into a series of discrete data patches. , where N is the number of patches, P is the time step of each patch (i.e., patch length), and V is the number of variables.

[0078] In the figure, X L Is the original data patch The output after processing by the linear prediction head (LinearHead) represents the long-term trend component of the prediction.

[0079] X T is the encoding feature The output after processing by the nonlinear prediction head (Non-libearHwad) represents the predicted nonlinear dynamic component.

[0080] yes Z The data after dimension transposition (Permute) is used to adapt the subsequent nonlinear prediction head.

[0081] Z It is the output of the Transformer encoder and is the encoding feature that represents the long-term temporal dependency information.

[0082] yes The data after dimension transposition (Permute) is used to adapt to the input of the Transformer encoder.

[0083] yes The data after positional encoding and value embedding is used as the final input of the Transformer encoder.

[0084] is the final forecast result, which is composed of the long-term trend component X L and nonlinear dynamic components X T The results are obtained by adding and fusion, and then inverse normalization.

[0085] L is the length (number of time steps) of the original input time series.

[0086] V is the number of variables (or channels).

[0087] N is the number of patches generated after slicing.

[0088] P is the length of each patch (i.e. the number of time steps within the patch).

[0089] D It is the embedding dimension (feature dimension) inside the Transformer model.

[0090] T is the length of the final prediction sequence (how many time steps into the future to predict).

[0091] Dynamic weight variable and short-term temporal feature extraction (AVSTFE) module: Figure 3 As shown, this module receives data patches , extracts local time-dependent features through the internal convolutional neural network structure, and combines the channel attention mechanism to dynamically adjust the weight distribution of variables related to the prediction target, and finally generates an enhanced representation .

[0092] in, It is the data consisting of all even-numbered time steps in the original patch after the even-odd slice transformation.

[0093] It is the data consisting of all odd-numbered time steps in the original patch after the even-odd slice transformation.

[0094] is an even number slice The representation after the value embedding operation is performed to restore the original patch dimension.

[0095] It is an odd-numbered slice The representation after the value embedding operation is performed to restore the original patch dimension.

[0096] yes Output after feature extraction by large-core CNN (convolutional neural network).

[0097] yes Output after feature extraction by large-core CNN (convolutional neural network).

[0098] yes Output after processing by the first layer dynamic weight extraction module.

[0099] yes Output after processing by the first layer dynamic weight extraction module.

[0100] yes Output after feature extraction by small-kernel CNN.

[0101] yes Output after feature extraction by small-kernel CNN.

[0102] yes The output after processing by the second-layer dynamic weight extraction module represents the final features of the odd branches.

[0103] yes The output after processing by the second-layer dynamic weight extraction module represents the final features of the odd branches.

[0104] It is the final output of the AVSTFE module, which consists of the output of the odd and even branches and The added fusion is an enhanced feature representation.

[0105] Odd-Even Slice Transform (CoST) unit: Figure 4 As shown, the parity slice transformation unit is used to separate the data patches according to the parity of the time step and perform a value embedding operation to restore the dimension. The input patch Interval acquisition is performed on the time step dimension P, which is divided into odd patches and even patches Afterwards, these sliced ​​patches are restored to their original patch length P using value embedding (e.g., linear layer) to obtain the processed odd-even patches and This processing helps improve the model’s robustness to noise and effectively increases the original time range that each patch can cover.

[0106] Multi-scale convolutional feature extraction unit: The multi-scale convolutional feature extraction unit includes at least two two-dimensional convolutional layers with different convolution kernel sizes to capture local temporal patterns and variable interaction features at different scales. The unit performs multi-scale feature extraction on the odd and even patches after CoST processing. It usually includes at least two two-dimensional convolutional layers with different convolution kernel sizes. For example, a large-core CNN (such as a convolution kernel size of 4×8) and a small-core CNN (such as a convolution kernel size of 4×4) can be used. The large-core CNN focuses on capturing time series features and relationships between variables within a larger time window, and then the small-core CNN further refines and extracts local temporal patterns and variable interaction details. Batch normalization (BatchNorm) and activation functions (such as GeLU) are usually connected after the convolution operation. The general representation is:

[0107]

[0108] in represent or .

[0109] Dynamic weight extraction unit: such as Figure 5 As shown in the figure, in order to distinguish the contribution of different input variables to the prediction results, after each convolution operation, the extracted features are passed to the AVIE module. This module first performs a feature map Average and max pooling are applied simultaneously across both the time step and patch dimensions to capture global features, preserving trend information and highlighting key time point features. The pooled results are then concatenated and dimensionally permuted before being fed into a one-dimensional convolutional layer (Conv1D). The sigmoid activation function then generates the dynamic weight ω for each variable. This weight ω is then multiplied back onto the original feature map to dynamically adjust the variable's features. The weight calculation formula is:

[0110]

[0111] in, ω It is the final calculated weight, which is used to achieve dynamic adjustment of variable characteristics. Subsequently, the relevant features can be influenced based on this weight to adjust the degree of effect of the features in the model and other processes.

[0112] Permute: This is a "permutation" operation that adjusts the dimensional order of a tensor (a multidimensional array commonly used in scenarios such as deep learning). For example, a tensor with the original dimensions [batch, channel, height, width] can be adjusted to [batch, height, width, channel], etc., to adapt to the dimensional requirements of subsequent calculations.

[0113] σ : refers to the Sigmoid activation function, which can map the input value to the (0,1) interval, play a role in nonlinear transformation of the convolution result and limit the output range, so that the output can be used as a meaningful value such as weight.

[0114] Conv: is the abbreviation of "Convolution", which stands for one-dimensional convolution layer (combined with the description of "one-dimensional convolution layer" in the article). It extracts local features by sliding the convolution kernel on the input data. Here it is used to further process the spliced ​​features and explore the associations between features.

[0115] Concat: It is the abbreviation of "Concatenate", which is a splicing operation. It splices the two features gavg and gmax together according to a specific dimension (such as the channel dimension) to integrate different feature information.

[0116] It is the average pooling calculated on the input features, It is the result of Global Max Pooling (GMP, abbreviated here) calculated on the input features. It is a feature representation obtained by taking the maximum value in a certain dimension, which can extract the most prominent feature information.

[0117] The convolution kernel size k of the one-dimensional convolution layer is adaptively determined according to the number of input variables V. .

[0118] Channel-independent Transformer encoder module: Figure 6 As shown, the enhanced feature representation Xp∈R output by the AVSTFE module V×N×P First, the dimension is adjusted (for example, transposing and possibly merging the N and V dimensions or treating N as a batch dimension and focusing on the P-length sequence of each variable V), then position encoding is applied and embedded into the D-dimensional feature space, finally transforming it into the input X required by the encoder d ∈R V×P×D (or N×V×P×D and then process each variable one by one). The Transformer encoder processes the enhanced feature representation of each variable separately. That is, the same Transformer encoder structure is shared, and the time series (Xdi ∈R 1×P×D ) for learning. Its internal self-attention mechanism acts on the time step dimension (the number of tokens is P), focusing on mining the long-term temporal patterns and quasi-periodic information of the variable itself without being directly interfered with by the current state of other variables, thereby generating the encoding feature Z.

[0119] Dual-path long-term global time series feature extraction and prediction module: This module uses a dual-path prediction head to predict the trend component and the dynamic component respectively.

[0120] Linear prediction head (trend component prediction): the sliced ​​raw data patch It is processed through a linear projection layer (without an activation function) to extract the long-term trend component of the time series. The output is .

[0121]

[0122] The Flatten operation adjusts the patch of each variable to a one-dimensional sequence that adapts to the linear layer input, and Linear performs a linear transformation on the input data.

[0123] Nonlinear prediction head (dynamic component prediction): The enhanced feature representation output by the AVSTFE module and the encoded features output by the Transformer encoder are input into a nonlinear prediction head based on a multi-layer perceptron (MLP). The MLP is used to integrate the temporal features and obtain the nonlinear dynamic component of the required prediction length. The output is:

[0124]

[0125] Fusion and output: The long-term trend component XL extracted by the linear prediction head (as the first prediction component) and the nonlinear dynamic component extracted by the nonlinear prediction head are combined (as the second prediction component) is fused, usually by direct addition. The fused prediction result is normalized. Finally, through the inverse operation of instance normalization, derevin,

[0126]

[0127] The prediction results are restored to the dimensional range of the original data to obtain the final prediction results.

[0128] In one possible embodiment, to verify the effectiveness of the ChannelMixer method of this application, experiments can be conducted on a series of publicly available multivariate time series datasets, such as ETT (ETTh1, ETTh2, ETTm1, ETTm2 - power transformer temperature data), Weather (meteorological data), Electricity (power consumption data), ILI (influenza-like illness data), and ExchangeRate (exchange rate data). These datasets have different numbers of time steps, numbers of variables, and sampling frequencies. The model training settings may include:

[0129] Data partitioning: Usually divided into training set, validation set and test set in chronological order (70% to 10% to 20% ratio).

[0130] Normalization: Instance normalization is performed on the input data, that is, each time series is normalized separately, which helps model training.

[0131] Learning rate strategy: A 1-cycle learning rate strategy can be used. For example, the maximum learning rate is set to 0.0001. The learning rate gradually increases from 1 / 25 of the maximum learning rate to the maximum in the early stage of training, and then gradually decays in the later stage of training.

[0132] Loss function: A combination of mean square error (MSE) and mean absolute error (MAE) can be used. , where the weight a can be set to 0.5.

[0133] L : Represents the total loss function after combination, which is a quantitative representation of the error between the model prediction result and the true value. It is used to measure the quality of the model prediction and guide the optimization of model parameters.

[0134] a : is the weight coefficient, used to control the mean square error ( MSE ) and mean absolute error ( MAE ) in the combined loss, is set to 0.5 in this paper, which means that the contribution of the two to the total loss is half each. It can also be adjusted according to actual needs. MSE and MAE influence.

[0135] L MSE : Refers to the mean square error (MeanSquaredError), which is calculated as the mean of the square of the difference between the predicted value and the true value. The formula is , n is the number of samples, is the true value, is the predicted value and is sensitive to outliers because squaring will amplify the deviation.

[0136] L MAE : Mean Absolute Error, which is calculated as the mean of the absolute values ​​of the differences between the predicted value and the true value. The formula is ,compared to MSE , it is more robust to outliers and will not excessively affect the loss calculation due to individual extreme deviations.

[0137] Batch size: for example, 128 or 256.

[0138] Number of training epochs: For example, 100 to 200 epochs with early stopping.

[0139] Hyperparameters: The convolution kernel size in the AVSTFE module can be set to (4×8, 4×4), the patch size can be set to 16, etc. The Transformer encoder can contain multiple layers (e.g., 2 layers), and each layer has multiple attention heads (e.g., 8).

[0140] Experimental results (such as Figure 7 and Figure 8 Generally, the results show that the ChannelMixer method of this application outperforms or is on par with a variety of existing advanced baseline models (such as PatchTST, DLinear, MICN, TimesNet, FEDformer, Autoformer, Informer, etc.) in prediction tasks with multiple datasets and different prediction lengths (such as 96, 192, 336, and 720 steps). For example, when making a 96-step prediction on the ETTm1 dataset (such as Figure 7 As shown in Figure 2, the ChannelMixer prediction curve can more closely follow the peak and valley changes of the true value. When making a 720-step prediction on the Weather dataset (as shown in Figure 2), the ChannelMixer prediction curve can more closely follow the peak and valley changes of the true value. Figure 8 As shown in Figure 3 ), ChannelMixer also maintains accurate predictions of the overall trend and major cycle profiles. These results demonstrate that our approach can effectively improve forecast accuracy and robustness by collaboratively modeling information in the time and variable dimensions.

[0141] This method is particularly suitable for fine-grained load forecasting of electricity, providing more accurate control and decision-making support information for power production planning, optimized dispatching, and transmission and distribution management. Furthermore, the method of this application can also be applied to other fields with similar data characteristics, such as weather forecasting, financial market analysis, industrial process monitoring, and other multivariate time series forecasting scenarios.

[0142] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including all changes and modifications that fall within the scope of the present invention and the preferred embodiments.

[0143] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A fine-grained power load collaborative forecasting method, characterized in that: include: Get multivariate time series input data; Slicing the input data to obtain multiple data patches; Preprocessing the input data to obtain processed data; Separating the processed data in a time sequence to obtain a plurality of data patches; Processing the data patch by a dynamic weight variable and a short-term temporal feature extraction module to obtain an enhanced feature representation; Separating the data patches by the parity of their time steps; Performing a value embedding operation in the data patch to restore the dimension of the data patch; At least two two-dimensional convolutional layers with different convolution kernel sizes are used to capture the local temporal pattern of the data patch and the interaction characteristics between variables from different receptive field scales; The interactive features include features that are correlated across time steps; Obtaining global features of the data patch by combining average pooling and maximum pooling operations; Using a one-dimensional convolutional layer to process the global features to calculate the dynamic weights of each variable; Inputting the enhanced feature representation into an encoder to generate encoded features; Inputting the data patch into a linear layer to obtain a first prediction component; Inputting the sliced ​​data patch into a linear projection layer to extract and predict a long-term trend component as the first prediction component; generating a second prediction component based on the enhanced feature representation and the encoded feature; Performing nonlinear transformation on the coding feature through a multi-layer perceptron to extract a nonlinear dynamic component as the second prediction component; The power load collaborative prediction result is obtained by fusing the first prediction component and the second prediction component.

2. A fine-grained power load collaborative forecasting method according to claim 1, characterized in that: The preprocessing of the input data to obtain processed data comprises: The input data is subjected to resampling processing and missing value mean supplement processing to obtain the processed data.

3. A fine-grained power load collaborative forecasting method according to claim 1, characterized in that: Inputting the enhanced feature representation into an encoder to generate encoding features comprises: The enhanced feature representation is input into a channel-independent Transformer encoder for time series processing to obtain encoded features.

4. A fine-grained power load collaborative forecasting method according to claim 1, characterized in that: After obtaining the power load collaborative prediction result by fusing the first prediction component and the second prediction component, the method further includes: The power load collaborative forecasting result is inversely normalized to restore the original data range of the power load collaborative forecasting result.

Images