A short-term power load forecasting method, system and device
By using a multi-scale decomposable hybrid module and an adaptive multi-predictor synthesis module, the problems of insufficient multi-scale feature extraction and non-adaptive prediction models in short-term power load forecasting are solved, thus achieving efficient and accurate power load forecasting.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA RESOURCES POWER (HUBEI) SALES CO LTD
- Filing Date
- 2026-01-19
- Publication Date
- 2026-06-05
AI Technical Summary
Existing short-term power load forecasting methods suffer from problems such as insufficient multi-scale feature extraction, non-adaptive feature processing mechanisms in forecasting models, and difficulty in balancing model efficiency and accuracy. In particular, the forecasting accuracy drops significantly when faced with drastic fluctuations in power load.
A multi-scale decomposable hybrid module is used for multi-scale feature decomposition and mixing. Combined with a dual-dependency interaction module and an adaptive multi-predictor synthesis module, adaptive prediction is achieved through a time pattern selector and a parallel predictor, and the prediction strategy is dynamically adjusted.
It achieves efficient, end-to-end multi-scale feature processing, improves the prediction accuracy of power load peaks and fluctuations, overcomes the information loss and computational complexity bottlenecks of traditional models, and achieves a balance between high efficiency and high accuracy.
Smart Images
Figure CN122159176A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system technology, and specifically to a short-term power load forecasting method, system, and device. Background Technology
[0002] With the development of new energy power systems, the large-scale grid connection of renewable energy sources such as wind and solar power has brought enormous challenges to the stability of the power system. Short-Term Load Forecasting (STLF) is a core component for ensuring the safe operation of the power grid, optimizing dispatch, and facilitating electricity market transactions. However, the intermittency and volatility of renewable energy, coupled with multiple complex factors such as social activities and weather, result in modern electricity loads exhibiting highly non-stationary, multi-scale fluctuations, and strong uncertainties, posing a severe challenge to traditional forecasting models.
[0003] Currently, deep learning-based prediction methods have become mainstream, but existing technologies still have the following technical shortcomings in addressing the aforementioned challenges: First, the multi-scale features of power load are not sufficiently extracted. Real power load is a complex superposition of multiple time patterns. Existing methods are static preprocessing methods, which typically use signal decomposition to statically process the original sequence before modeling. This method not only increases the complexity of the model, but the decomposition and reconstruction process is also prone to information loss or distortion, making it difficult to effectively handle the dynamic characteristics of multi-scale entanglement.
[0004] Secondly, the feature processing mechanism of the prediction model is fixed and non-adaptive. Existing prediction models have fixed model parameters and internal feature processing logic after training. These models use a static parameter mechanism to process all types of input data. However, the dominant characteristics of power load change dynamically over different time periods. This fixed-parameter prediction mechanism cannot dynamically adjust its internal feature extraction and prediction focus according to the real-time characteristics of the input data, resulting in a significant decrease in prediction accuracy when the load fluctuates drastically or switches modes.
[0005] Third, there is a dilemma between model efficiency and accuracy. Existing models are caught in a predicament: on the one hand, while models such as Transformer attempt to capture long-term dependencies, the quadratic computational complexity of their self-attention mechanism makes their training and inference costs high, and they are prone to overfitting to local extrema in the load sequence; on the other hand, while models such as Multilayer Perceptron (MLP) are efficient, their simple linear mapping has a bottleneck of insufficient information extraction, making it difficult to capture the complex multivariate dynamic characteristics in the load data, thus limiting their prediction accuracy.
[0006] In summary, existing technologies lack a novel forecasting method that can efficiently and end-to-end process the multi-scale characteristics of power load within the model and adaptively adjust its forecasting strategy based on the real-time characteristics of the input data. Summary of the Invention
[0007] The present invention provides a short-term power load forecasting method, system and apparatus to solve at least one of the above-mentioned technical problems.
[0008] The technical solution of this invention to solve the above-mentioned technical problems is as follows: A short-term power load forecasting method, comprising: S1, acquire historical power load data and at least one covariate data as the original input sequence; S2, the original input sequence is decomposed into multiple subsequences at different time scales using a multi-scale decomposable mixing module, and the multiple subsequences are mixed to obtain multi-scale mixed features; S3, using the dual dependency interaction module to perform dual dependency interaction modeling of time dependency and channel dependency on the multi-scale hybrid features, to obtain the interactive features; S4. Based on the adaptive multi-predictor synthesis module, the time pattern selector is used to adaptively calculate the multi-scale mixed features to obtain a set of selector weights to characterize the contribution of different time patterns to future predictions at the current time point. Multiple parallel predictors of time pattern projection are used to perform parallel predictions on the interactive features to obtain a set of prediction outputs. The set of selector weights is used to perform a weighted summation of the set of prediction outputs to obtain the short-term power load prediction value.
[0009] Based on the above technical solution, the present invention can be further improved as follows.
[0010] Furthermore, S2 specifically includes: S21, Substitute the original input sequence into the multi-scale decomposable mixing module; S22, the original input sequence is decomposed by multiple downsampling to obtain multiple subsequences at different time scales; S23, the multiple subsequences at different time scales are mixed in a residual manner through a multilayer perceptron from coarse-grained to fine-grained to obtain multi-scale mixed features.
[0011] Furthermore, in step S22, the formula for decomposing the original input sequence through multiple downsampling steps is as follows: ; In the formula, Indicates the first Layer time mode; Represents the average downsampling function; when hour, This represents the first-level time pattern, and the first-level time pattern is the original input sequence; This indicates the number of downsampling times; each time pattern represents a subsequence at a different time scale. In S23, the formula for mixing the multiple subsequences at different time scales in a residual manner using a multilayer perceptron, from coarse-grained to fine-grained, is as follows: ; In the formula, Indicates the first Layered mixed information; Represents a multilayer perceptron; when hour, This represents the initial mixed information, and the initial mixed information is the first... Layered time pattern; the first layer of mixed information is the multi-scale mixed feature.
[0012] Furthermore, S3 specifically refers to: S31, Substitute the multi-scale hybrid features into the dual-dependency interaction module; S32, Perform a block-based operation on the multi-scale hybrid features to obtain block features; S33, using a multilayer perceptron shared across time steps, residual calculation is performed on the block features to aggregate dependencies in the time dimension and obtain temporal hybrid features; S34, the temporal mixing features are transposed, processed by the multilayer perceptron and inverted through a multilayer perceptron shared between channels to fuse the dependencies in the channel dimension, and the noise introduced is controlled by a scaling factor to obtain the initial interaction features. S35, perform inverse block division on the initial interaction features to obtain the interaction features.
[0013] Furthermore, in step S32, the formula for performing block segmentation on the multi-scale mixed features is as follows: ; In the formula, This represents the block operation function; This represents the multi-scale mixed feature, and , Indicates the number of channels. Indicates the sequence length; This represents the block feature, and , Indicates the number of blocks. Indicates the block length; In step S33, the formula for calculating the residual of the block results using a multilayer perceptron shared across time steps is as follows: ; In the formula, Indicates in Block features at time steps Indicates in Features following the initial interaction at the time step This represents a multilayer perceptron. Indicates in Temporal mixing features at time steps; In S34, the time mixing features are transposed, processed by the multilayer perceptron, and inverted using a multilayer perceptron shared across channels. The formula for the noise introduced, controlled by a scaling factor, is as follows: ; In the formula, Indicates in Features following the initial interaction at the time step Indicates the scaling factor. This indicates transpose.
[0014] Furthermore, in step S4, the multi-scale mixed features are adaptively calculated using a time pattern selector, specifically including: S41, the multi-scale mixed features are calculated using a decomposition layer and a gating function with learnable noise to obtain the original gated output; wherein, the calculation formula for the original gated output is: ; In the formula, This represents the original gated output. This represents the multi-scale mixed feature. Indicates a decomposition layer. Indicates standard Gaussian noise. Represents the gate function. Indicates learning noise; S42, the original gated output is processed using the Top-K function to retain the original gated output. The maximum value, and the value of the original gated output excluding The remaining values, excluding the maximum value, are exponentially scaled to obtain the Top-K function processing result; the formula for processing the original gated output using the Top-K function is as follows: ; In the formula, Represents the Top-K function. Indicates the original gated output The Middle The maximum value, This represents a constant used to adjust the selector weights. Indicates the original gated output The Middle One value; S43, the Top-K function processing result is normalized using the Softmax function to obtain a set of selector weights representing the contribution of different time patterns to future predictions at the current time point; wherein, the formula for normalizing the Top-K function processing result using the Softmax function is: ; In the formula, This represents the set of selector weights. This represents the Softmax function.
[0015] further, In S4, the formula for weighted summation of the set of predicted outputs using the set of selector weights is as follows: ; In the formula, This represents the short-term power load forecast value. Represents the weight of the first selector in the set of selector weights. Each selector weight This indicates the features after the interaction. Represents the first of the plurality of parallel predictors One predictor; Indicates the first The prediction output obtained by the first predictor predicting the post-interaction features, the first predictor... The prediction output obtained by the predictor predicting the post-interaction features is the first prediction output in the set of prediction outputs. One predicted output, This represents the total number of parallel predictors.
[0016] Furthermore, prior to S4, training of the time pattern selector and predictor is also included; The overall loss function for training the temporal pattern selector and predictor includes the prediction loss function and the selector balancing loss function; where, The prediction loss function is used to minimize the mean square error between the predicted output value and the true value. The selector balancing loss function is expressed as follows: ; In the formula, This represents the selector's balance loss function. Indicates selector weight variance Indicates selector weight The mean, This represents a very small, non-zero constant. The total loss function is expressed as: ; In the formula, Represents the total loss function, Denotes the prediction loss function, This represents the total number of trainable parameters in the model. and This is a hyperparameter.
[0017] Based on the above-mentioned short-term power load forecasting method, the present invention also provides a short-term power load forecasting system.
[0018] A short-term power load forecasting system, applied to the short-term power load forecasting method described above, includes: The data acquisition module is used to acquire historical power load data and at least one covariate data as the raw input sequence; A multi-scale decomposable mixing module is used to decompose the original input sequence into multiple sub-sequences at different time scales and mix the multiple sub-sequences to obtain multi-scale mixed features. The dual-dependency interaction module is used to perform dual-dependency interaction modeling of the time dependency and channel dependency of the multi-scale hybrid features to obtain the interactive features. An adaptive multi-predictor synthesis module is used to adaptively calculate the multi-scale mixed features using a time pattern selector to obtain a set of selector weights that characterize the contribution of different time patterns to future predictions at the current time point. Multiple parallel predictors using time pattern projection are used to perform parallel predictions on the interactive features to obtain a set of prediction outputs. The set of prediction outputs is weighted and summed using the set of selector weights to obtain the short-term power load prediction value.
[0019] Based on the above-mentioned short-term power load forecasting method, the present invention also provides a short-term power load forecasting device.
[0020] A short-term power load forecasting device includes a processor, a memory, and a computer program stored in the memory. When the computer program is executed by the processor, it implements the short-term power load forecasting method as described above.
[0021] The beneficial effects of this invention are as follows: The short-term power load forecasting method, system, and device of this invention automatically perform multi-scale decomposition and mixing within the model in an "end-to-end" manner through a multi-scale decomposition and mixing module. This overcomes the "static" and "information loss" problems caused by traditional preprocessing methods such as CEEMDAN, and can more realistically capture the multi-scale characteristics of power load. Simultaneously, this invention introduces a dynamic and adaptive forecasting mechanism through an adaptive multi-predictor synthesis module. Its time mode selector can identify the current dominant time mode in real time and dynamically allocate forecast weights to the parallel predictor best suited to handle that mode, overcoming the limitations of traditional single fixed models and greatly improving the forecasting accuracy for load peaks and fluctuations. Furthermore, the multi-scale decomposition and mixing module, the dual-dependency interaction module, and the adaptive multi-predictor synthesis module of this invention are all based on an efficient multilayer perceptron architecture, overcoming the bottleneck of high computational complexity in traditional Transformer models. In addition, through the multi-scale decomposition of the multi-scale decomposition and mixing module and the adaptive forecasting mechanism of the adaptive multi-predictor synthesis module, this invention also solves the bottleneck of insufficient information extraction in traditional MLP models, achieving a balance between high efficiency and high accuracy. Attached Figure Description
[0022] Figure 1 This is a flowchart of a short-term power load forecasting method according to the present invention. Figure 2 This is a schematic diagram of the structure of a short-term power load forecasting system according to the present invention; Figure 3 This is a schematic diagram illustrating the internal working principle of a multi-scale decomposable hybrid module. Figure 4 This is a schematic diagram illustrating the internal working principle of the adaptive multi-predictor fusion module. Figure 5 This is a schematic diagram illustrating the prediction effect under a typical daily load condition in the example. Figure 6 This is a structural block diagram of a short-term power load forecasting device according to the present invention. Detailed Implementation
[0023] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.
[0024] Example 1: like Figure 1 As shown, a short-term power load forecasting method includes: S1, acquire historical power load data and at least one covariate data as the original input sequence; S2, the original input sequence is decomposed into multiple subsequences at different time scales using a multi-scale decomposable mixing module, and the multiple subsequences are mixed to obtain multi-scale mixed features; S3, using the dual dependency interaction module to perform dual dependency interaction modeling of time dependency and channel dependency on the multi-scale hybrid features, to obtain the interactive features; S4. Based on the adaptive multi-predictor synthesis module, the time pattern selector is used to adaptively calculate the multi-scale mixed features to obtain a set of selector weights to characterize the contribution of different time patterns to future predictions at the current time point. Multiple parallel predictors of time pattern projection are used to perform parallel predictions on the interactive features to obtain a set of prediction outputs. The set of selector weights is used to perform a weighted summation of the set of prediction outputs to obtain the short-term power load prediction value.
[0025] The present invention provides a short-term power load forecasting method based on adaptive multi-scale decomposition and hybrid forecasting. This method can achieve multi-scale feature decomposition and hybridization, dynamic dependency modeling and adaptive hybrid forecasting of power load data in an efficient, end-to-end manner. This overcomes the limitations of existing technologies in processing load data, such as information loss due to static preprocessing, inability to adaptively cope with sudden changes using fixed parameter models, and difficulty in balancing computational cost and prediction accuracy.
[0026] The steps of this invention will be described in detail below.
[0027] In this invention, the covariate data in S1 can be data such as temperature or humidity.
[0028] In this invention, S2 achieves multi-scale decomposition and mixing: the original input sequence Substituting a multi-scale decomposable mixture (MDM) module, the MDM module downsamples the original input sequence multiple times. The data is decomposed into multiple subsequences at different time scales, and then these subsequences are mixed from bottom to top (i.e., from coarse-grained to fine-grained) using a multilayer perceptron (MLP) in a residual manner to obtain a multi-scale hybrid feature that fully integrates multi-scale information. .
[0029] Specifically, S2 includes S21 to S23: S21, the original input sequence Substitute the multi-scale decomposable hybrid module into the given description.
[0030] S22, the original input sequence is downsampled multiple times. The sequence is decomposed to obtain multiple subsequences at different time scales; Specifically, S22 implements multi-scale decomposition: using the original input sequence For the first-level time pattern The subsequent Layer Time Mode By analyzing the first layer of the previous layer Layer Time Mode Acquired using average downsampling, with a downsampling rate of , executed Sub-sampling operation: ; In the formula, Indicates the first Layer time mode; Represents the average downsampling function; when hour, This indicates the first-level time pattern, and the first-level time pattern The original input sequence (Right now ); This indicates the number of downsampling times; each time pattern is a subsequence at a different time scale.
[0031] S23, the multiple subsequences at different time scales are mixed in a residual manner using a multilayer perceptron, from coarse-grained to fine-grained, to obtain multi-scale mixed features. ; Specifically, S23 achieves multi-scale fusion: taking the first... Layer (i.e., the lowest layer, the coarsest granularity) time pattern As the first Layered Hybrid Information , (No. Layered Hybrid Information For the initial mixed information, i.e. From bottom to top (from) 1) The mixture is performed using a multilayer perceptron (MLP) in a residual manner, and the mixing formula is expressed as: ; In the formula, Indicates the first Layered mixed information; Represents a multilayer perceptron; when hour, This represents the initial mixed information, and the initial mixed information is the first... Layer Time Mode Layer 1 Mixed Information For the multi-scale hybrid features ,Right now .
[0032] In this invention, S3 achieves dual dependency interaction: the multi-scale hybrid features As input, the Dual Dependency Interaction (DDI) module is substituted; the DDI module models the multi-scale mixture features in parallel through a temporal mixture MLP and a channel mixture MLP. The “time dependency” (i.e., the temporal correlation of the load series) and the “channel dependency” (i.e., the mutual influence between the load and the covariate) are used to obtain an interaction post-feature.
[0033] Specifically, S3 includes S31 to S35: S31, the multi-scale hybrid features Substitute the aforementioned dual-dependency interaction module.
[0034] S32, regarding the multi-scale hybrid features Perform block segmentation to obtain block features; Specifically, the dual-dependency interaction module first processes the multi-scale hybrid features output by S2. ( , For the number of channels, The sequence length is used to perform a block division operation to obtain the block features. ( , The number of blocks, (where the block length is), i.e. ,in, This represents the block operation function; subsequently, dual dependency interaction is achieved through two parallel MLPs (i.e., a time-mixing MLP and a channel-mixing MLP) (i.e., executing S33 and S34 below).
[0035] S33, using a multilayer perceptron shared across time steps (i.e., a temporal fusion MLP), residual calculation is performed on the segmented features to aggregate dependencies along the temporal dimension, resulting in temporal fusion features; wherein, the formula for calculating the residuals of the segmented results using a multilayer perceptron shared across time steps is: ; In the formula, Indicates in Block features at time steps Indicates in Features following the initial interaction at the time step This represents a multilayer perceptron. Indicates in Temporal mixing features at time steps.
[0036] S34, the temporal mixing features are processed through a multilayer perceptron (i.e., a channel mixing MLP) shared between channels. The system performs transposition, multilayer perceptron processing, and inverse transposition to fuse dependencies along the channel dimension, and uses a scaling factor to control the introduced noise, resulting in the initial interaction-related features. The formula for transposing, processing, and inverting the temporal mixing features using a multilayer perceptron shared across channels, and controlling the introduced noise by a scaling factor, is as follows: ; In the formula, Indicates in Features following the initial interaction at the time step Indicates the scaling factor. This indicates transpose.
[0037] S35, regarding the features after the initial interaction Perform inverse block division to obtain the post-interaction features. .
[0038] In this invention, S4 achieves adaptive hybrid prediction: the multi-scale hybrid features obtained in S2 are used to perform the prediction. The interaction features obtained from S3 Substitute these values into the Adaptive Multiple Predictor Synthesis (AMS) module; the AMS module includes a "Time Pattern Selector (TP-Selector)" and a "Time Pattern Projection (TP-Projection)": (1) TP-Selector is based on the multi-scale hybrid features Adaptively calculate a set of selector weights The set of selector weights Used to characterize the contribution of different time patterns to future predictions at the current point in time; (2) Multiple parallel predictors are deployed within TP-Projection, and all predictors receive the post-interaction features. Each predictor generates a prediction output, and the prediction outputs of all predictors constitute a set of prediction outputs; (3) Finally, the AMS module uses a set of selector weights generated by TP-Selector to perform a weighted summation of a set of prediction outputs generated by TP-Projection, resulting in a final short-term power load prediction that adaptively blends the outputs of multiple predictors. .
[0039] Specifically, TP-Selector generates a set of selector weights through a mechanism that includes noise gating and a Top-K function to adaptively identify... The dominant time patterns include: S41, the multi-scale mixed features are processed through a decomposition layer and a gating function with learnable noise. Calculations are performed to obtain the original gated output; wherein, the calculation formula for the original gated output is: ; In the formula, This represents the original gated output. This represents the multi-scale mixed feature. Indicates a decomposition layer. Indicates standard Gaussian noise. Represents the gate function. Indicates learning noise; S42, the original gated output is gated using the Top-K function. Processing is performed to preserve the original gated output. In The maximum value, and the original gated output Except The remaining values, excluding the maximum value, are exponentially scaled to obtain the Top-K function processing result; wherein, the original gated output is gated by the Top-K function. The formula for processing is: ; In the formula, Represents the Top-K function. Indicates the original gated output The Middle The maximum value, This represents a constant used to adjust the selector weights. Indicates the original gated output The Middle One value; S43, the Top-K function processing result is normalized using the Softmax function to obtain a set of selector weights representing the contribution of different time patterns to future predictions at the current time point; wherein, the Top-K function processing result is normalized using the Softmax function. The formula for normalization is: ; In the formula, This represents the set of selector weights. This represents the Softmax function.
[0040] In addition, the final short-term power load forecast It is by outputting the TP-Selector's first... Selector weights , and the Parallel predictors The predicted outputs are obtained by weighted summation; the formula for weighted summation is: ; In the formula, This represents the short-term power load forecast value. Represents the weight of the first selector in the set of selector weights. Each selector weight This indicates the features after the interaction. Represents the first of the plurality of parallel predictors One predictor; Indicates the first The prediction output obtained by the first predictor predicting the post-interaction features, the first predictor... The prediction output obtained by the predictor predicting the post-interaction features is the first prediction output in the set of prediction outputs. One predicted output, This represents the total number of parallel predictors.
[0041] In this invention, prior to S4, training of the time pattern selector and the predictor is also included. The overall loss function for training the temporal pattern selector and predictor includes the prediction loss function and the selector balancing loss function; where, The prediction loss function is used to minimize the short-term power load forecast value. Compared with the actual value of short-term power load Mean squared error (MSE) between: ; In the formula, Denotes the prediction loss function, This represents the mean square error function.
[0042] The selector balancing loss function is used to optimize the weight distribution balance of the selectors. It is calculated using the coefficient of variation and aims to penalize the model's "over-reliance" on a few predictors (i.e., "load balancing"), ensuring that all predictors are adequately trained. The selector balancing loss function is expressed as follows: ; In the formula, This represents the selector's balance loss function. Indicates selector weight variance Indicates selector weight The mean, This represents a very small non-zero constant, typically 10. -8 To prevent the denominator from being zero; The total loss function is expressed as: ; In the formula, Represents the total loss function, Denotes the prediction loss function, This represents the total number of trainable parameters in the proposed model. and This is a hyperparameter.
[0043] Example 2: Based on the above-mentioned short-term power load forecasting method, the present invention also provides a short-term power load forecasting system.
[0044] like Figure 2 As shown, a short-term power load forecasting system, applied to the short-term power load forecasting method described above, includes: Data acquisition module 1 is used to acquire historical power load data and at least one covariate data as the raw input sequence; The multi-scale decomposable mixing module 2 is used to decompose the original input sequence into multiple sub-sequences at different time scales and mix the multiple sub-sequences to obtain multi-scale mixed features. Dual dependency interaction module 3 is used to perform dual dependency interaction modeling of time dependency and channel dependency on the multi-scale hybrid features to obtain the interactive features; The adaptive multi-predictor synthesis module 4 is used to adaptively calculate the multi-scale mixed features using a time pattern selector to obtain a set of selector weights that characterize the contribution of different time patterns to future predictions at the current time point. It then uses multiple parallel predictors based on time pattern projection to perform parallel predictions on the interactive features to obtain a set of prediction outputs. Finally, it uses the set of selector weights to perform a weighted summation of the set of prediction outputs to obtain the short-term power load prediction value.
[0045] The present invention discloses a short-term power load forecasting system, which provides the original input sequence during operation. (Including historical power load data and covariate data) is first fed into the multi-scale decomposable mixing module 2. The multi-scale decomposable mixing module 2 executes S2 in the method of this invention, processing the original input sequence... Decompose and blend the components, and output multi-scale blended features. Multi-scale hybrid features The data is divided into two paths: one path goes to the dual-dependency interaction module 3, and the other path goes to the time mode selector in the adaptive multi-predictor synthesis module 4. The dual-dependency interaction module 3 executes S3 in the method of this invention, processing the multi-scale mixed features. Perform time and channel dependency modeling and output post-interaction features. This is then sent to the temporal pattern projection in the adaptive multi-predictor synthesis module 4. The adaptive multi-predictor synthesis module 4 executes S4 of the method of this invention, whereby the temporal pattern selector within the adaptive multi-predictor synthesis module 4 selects the pattern based on the multi-scale blending features. Calculate a set of selector weights The temporal pattern projection within the adaptive multi-predictor synthesis module 4 is based on the post-interaction features. Generate a set of predicted outputs ( (multiple parallel prediction outputs), and finally weighted summation using a set of selector weights. Come and gather this The predicted outputs yield the final short-term power load forecast. .
[0046] Figure 3 This is a schematic diagram illustrating the internal working principle of the multi-scale decomposable hybrid module. The working principle of the multi-scale decomposable hybrid module 2 is a "U-shaped" data flow, including two processes: decomposition and hybridization. (a) Multi-scale decomposition: This process can solve the problem of insufficient multi-scale feature extraction of power load in the background technology. Using the original input sequence... For the first-level time pattern (Finest granularity). The subsequent... Layer Time Mode By analyzing the first layer of the previous layer Layer Time Mode Acquired using average downsampling (AvgPooling), with a downsampling rate of , executed The next downsampling operation yields a series of subsequences at different time scales; that is, each layer of time pattern is a subsequence at a different time scale. For example, the first... Layer Time Mode This represents a low-frequency trend in load, the second-level time pattern. This represents mid-frequency daily fluctuations, a first-level time pattern. This represents high-frequency noise.
[0047] (b) Multi-scale fusion: This process aims to fuse a series of subsequences at different time scales end-to-end within the model. Taking the first... Layer (coarsest granularity) time mode As the initial mixed information, the initial mixed information is the first... Layered Hybrid Information Then, from bottom to top, the first... Layered Hybrid Information Input a multilayer perceptron (MLP) and its output is compared with the first... Layer Time Mode Perform residual join to obtain the first... Layered Hybrid Information Finally, the multi-scale hybrid features are output to the dual-dependency interaction module 3 and the adaptive multi-predictor synthesis module 4. This is the first layer of mixed information. Layer 1 mixed information Fully integrates from the first Layer Time Mode Time pattern at level 1 All scale information.
[0048] The dual-dependency interaction module 3 can solve the dilemma in the background technology of high-precision model computational overhead and insufficient information extraction of high-efficiency model. The dual-dependency interaction module 3 first integrates multi-scale hybrid features. ( , For the number of channels, The sequence length is used to perform a block division operation to obtain the block features. ( , The number of blocks, (where the block length is specified). Subsequently, dual dependency interaction is achieved through two efficient, parallel MLPs (i.e., a time-mixing MLP and a channel-mixing MLP). (a) Temporal Mixing: This involves using an MLP shared across time steps (i.e., a temporally mixed MLP) to... Residual calculations are performed to aggregate dependencies along the time dimension, resulting in temporal hybrid features. : ; (b) Channel Mixing: Temporal mixing features are achieved through a shared MLP (i.e., a channel mixing MLP) between channels. Transpose, MLP processing, and inverse transpose are performed to fuse dependencies along the channel dimension, yielding the initial interaction features. : ; in, A learnable scaling factor is used to balance noise introduced by channel dependencies. Features after initial interaction. After the inverse block division operation, the interactive features are obtained. And it serves as the input to the adaptive multi-predictor synthesis module 4.
[0049] The adaptive multi-predictor synthesis module 4 can address the core pain point in the background technology where the feature processing mechanism of the prediction model is fixed and non-adaptive. For example... Figure 4 As shown, the adaptive multi-predictor synthesis module 4 internally employs a "two-stream" architecture: (a) Stream A (TP-Selector): This process adaptively determines which prediction mode the current input data should emphasize. It receives multi-scale mixed features. Furthermore, a set of selector weights is calculated using a mechanism that incorporates noise gating and a Top-K function. The time pattern selector generates a set of selector weights. The process is as follows: First, a decomposition layer (Decomp) and a layer with learnable noise are used. The gating function is used to calculate the original gating output. : ; in, Standard Gaussian noise is used to improve the generalization ability of the model.
[0050] Then, the original gated output is gated through a Top-K function. The Top-K function performs processing and retains the original gated output. In The maximum value (representing) The most important "dominant" time pattern), and the original gated output. Except All values other than the maximum value are exponentially scaled: ; in, Represents the Top-K function. Indicates the original gated output The Middle The maximum value, This represents a constant used to adjust the selector weights. Indicates the original gated output The Middle Values.
[0051] Finally, the selector weights are normalized using the Softmax function to obtain the final set of selector weights. A set of selector weights That is, it represents In a series of parallel predictors, each predictor has a weight at the current time point.
[0052] (b) Stream B (TP-Projection): This stream is used to perform parallel "expert predictions." It receives post-interaction features. And input it into In a parallel predictor (Predictor1 ... m) with independent structures and different parameters, each predictor independently processes the interacting features. Perform the calculations and output a prediction for each.
[0053] (c) Final aggregation (weighted summation): In Figure 4 At the end, this invention uses a set of selector weights output by the TP-Selector. In time-mode projection The weighted sum of a set of prediction outputs from each predictor yields a final short-term power load forecast that adaptively incorporates the opinions of all experts. : ; In the formula, This represents the short-term power load forecast value. Represents the weight of the first selector in the set of selector weights. Each selector weight This indicates the features after the interaction. Represents the first of the plurality of parallel predictors One predictor; Indicates the first The prediction output obtained by the first predictor predicting the post-interaction features, the first predictor... The prediction output obtained by the predictor predicting the post-interaction features is the first prediction output in the set of prediction outputs. One predicted output, This represents the total number of parallel predictors.
[0054] In this way, if the TP-Selector (flow A) determines that the current load pattern is a severe spike, it will automatically allocate more power to the predictor that is best at handling spikes (such as Predictor3) through selector weights, thereby achieving adaptive prediction.
[0055] Figure 5This is a schematic diagram illustrating the prediction effect of the present invention under a typical daily load condition. The horizontal axis represents 24 hours (in hours), and the vertical axis represents the power load (in kW). Figure 5 As shown, the actual short-term power load (solid line) exhibits complex fluctuations, including daytime peaks and nighttime troughs, with a sharp load surge around 12:00. The curve of the short-term power load prediction (dashed line) obtained through this invention highly coincides with the actual short-term power load curve. This strongly demonstrates that this invention can effectively overcome the deficiencies in the prior art. Specifically, this invention can process the multi-scale characteristics of load end-to-end, and through its adaptive hybrid prediction mechanism, it maintains extremely high prediction accuracy and robustness even when facing sharp load surges and mode switching, achieving accurate quantitative prediction of power load and verifying the beneficial effects of this invention.
[0056] Example 3: Based on the above-mentioned short-term power load forecasting method, the present invention also provides a short-term power load forecasting device.
[0057] like Figure 6 As shown, a short-term power load forecasting device includes a processor, a memory, and a computer program stored in the memory. When the computer program is executed by the processor, it implements the short-term power load forecasting method as described above.
[0058] In other words, the short-term power load forecasting device of this invention may include, but is not limited to: a processor and a memory; the memory is used to store computer programs; the processor is used to execute the short-term power load forecasting method of this invention by calling the computer programs.
[0059] In one alternative embodiment, a short-term power load forecasting device is provided, such as Figure 6 As shown. Figure 6 The illustrated short-term power load forecasting device includes a processor and a memory. The processor and memory are connected, for example, via a bus. Optionally, the short-term power load forecasting device may further include a transceiver, which can be used for data interaction between the short-term power load forecasting device and other electronic devices, such as sending and / or receiving data. It should be noted that in practical applications, the transceiver is not limited to one unit, and the structure of this short-term power load forecasting device does not constitute a limitation on the embodiments of the present invention.
[0060] The processor can be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), a PLC (Programmable Logic Controller), a FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this invention. The processor can also be a combination that implements computational functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.
[0061] A bus can include a pathway for transmitting information between the aforementioned components. The bus can be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 6 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.
[0062] The memory may be ROM (Read Only Memory) or other types of static storage devices capable of storing static information and instructions, RAM (Random Access Memory) or other types of dynamic storage devices capable of storing information and instructions, or EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited to these.
[0063] The memory stores application code (computer program) that executes the present invention, and its execution is controlled by a processor. The processor executes the application code stored in the memory to implement the content shown in the foregoing method embodiments.
[0064] The short-term power load forecasting device can also be a terminal device, which can be any device that can install applications, including at least one of smartphones, tablets, laptops, desktop computers, smart speakers, smartwatches, smart TVs, and smart in-vehicle devices.
[0065] It should be noted that, Figure 6 The short-term power load forecasting device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.
[0066] In summary, the short-term power load forecasting device of the present invention has the following technical advantages: (1) This invention uses a multi-scale decomposable hybrid module to automatically perform multi-scale decomposition and hybridization within the model in an "end-to-end" manner, overcoming the "static" and "information loss" problems caused by traditional preprocessing methods such as CEEMDAN, and can more realistically capture the multi-scale characteristics in power load.
[0067] (2) This invention introduces a dynamic and adaptive prediction mechanism through an adaptive multi-predictor synthesis module. Its TP-Selector can identify the current dominant time pattern in real time and dynamically allocate the selector weights to the parallel predictor that is best suited to handle the pattern, overcoming the limitations of the traditional single fixed model and greatly improving the prediction accuracy of load spikes and fluctuations.
[0068] (3) The multi-scale decomposable hybrid module, dual-dependency interaction module and adaptive multi-predictor synthesis module of the present invention are all based on the efficient multilayer perceptron (MLP) architecture, which overcomes the bottleneck of high computational complexity of the Transformer model. At the same time, through the multi-scale decomposition of the multi-scale decomposable hybrid module and the adaptive prediction mechanism of the adaptive multi-predictor synthesis module, the bottleneck of insufficient information extraction of the traditional MLP model is solved, and the unity of high efficiency and high accuracy is achieved.
[0069] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A short-term power load forecasting method, characterized in that, include: S1, acquire historical power load data and at least one covariate data as the original input sequence; S2, the original input sequence is decomposed into multiple subsequences at different time scales using a multi-scale decomposable mixing module, and the multiple subsequences are mixed to obtain multi-scale mixed features; S3, using the dual dependency interaction module to perform dual dependency interaction modeling of time dependency and channel dependency on the multi-scale hybrid features, to obtain the interactive features; S4. Based on the adaptive multi-predictor synthesis module, the time pattern selector is used to adaptively calculate the multi-scale mixed features to obtain a set of selector weights to characterize the contribution of different time patterns to future predictions at the current time point. Multiple parallel predictors of time pattern projection are used to perform parallel predictions on the interactive features to obtain a set of prediction outputs. The set of selector weights is used to perform a weighted summation of the set of prediction outputs to obtain the short-term power load prediction value.
2. The short-term power load forecasting method according to claim 1, characterized in that, S2 specifically includes: S21, Substitute the original input sequence into the multi-scale decomposable mixing module; S22, the original input sequence is decomposed by multiple downsampling to obtain multiple subsequences at different time scales; S23, the multiple subsequences at different time scales are mixed in a residual manner through a multilayer perceptron from coarse-grained to fine-grained to obtain multi-scale mixed features.
3. The short-term power load forecasting method according to claim 2, characterized in that, In step S22, the formula for decomposing the original input sequence through multiple downsampling steps is as follows: ； In the formula, Indicates the first Layer time mode; Represents the average downsampling function; when hour, This represents the first-level time pattern, and the first-level time pattern is the original input sequence; This indicates the number of downsampling times; each time pattern represents a subsequence at a different time scale. In S23, the formula for mixing the multiple subsequences at different time scales in a residual manner using a multilayer perceptron, from coarse-grained to fine-grained, is as follows: ； In the formula, Indicates the first Layered mixed information; Represents a multilayer perceptron; when hour, This represents the initial mixed information, and the initial mixed information is the first... Layered time pattern; the first layer of mixed information is the multi-scale mixed feature.
4. The short-term power load forecasting method according to claim 1, characterized in that, Specifically, S3 is: S31, Substitute the multi-scale hybrid features into the dual-dependency interaction module; S32, Perform a block-based operation on the multi-scale hybrid features to obtain block features; S33, using a multilayer perceptron shared across time steps, residual calculation is performed on the block features to aggregate dependencies in the time dimension and obtain temporal hybrid features; S34, the temporal mixing features are transposed, processed by the multilayer perceptron and inverted through a multilayer perceptron shared between channels to fuse the dependencies in the channel dimension, and the noise introduced is controlled by a scaling factor to obtain the initial interaction features. S35, perform inverse block division on the initial interaction features to obtain the interaction features.
5. The short-term power load forecasting method according to claim 4, characterized in that, In step S32, the formula for performing block segmentation on the multi-scale hybrid features is as follows: ； In the formula, This represents the block operation function; This represents the multi-scale mixed feature, and , Indicates the number of channels. Indicates the sequence length; This represents the block feature, and , Indicates the number of blocks. Indicates the block length; In step S33, the formula for calculating the residuals of the block results using a multilayer perceptron shared across time steps is as follows: ； In the formula, Indicates in Block features at time steps Indicates in Features following the initial interaction at the time step This represents a multilayer perceptron. Indicates in Temporal mixing features at time steps; In S34, the time mixing features are transposed, processed by the multilayer perceptron, and inverted using a multilayer perceptron shared across channels. The formula for the noise introduced, controlled by a scaling factor, is as follows: ； In the formula, Indicates in Features following the initial interaction at the time step Indicates the scaling factor. This indicates transpose.
6. The short-term power load forecasting method according to claim 1, characterized in that, In step S4, the multi-scale mixed features are adaptively calculated using a time pattern selector, specifically including: S41, the multi-scale mixed features are calculated using a decomposition layer and a gating function with learnable noise to obtain the original gated output; wherein, the calculation formula for the original gated output is: ； In the formula, This represents the original gated output. This represents the multi-scale hybrid feature. Indicates a decomposition layer. Indicates standard Gaussian noise. Represents the gate function. Indicates learning noise; S42, the original gated output is processed using the Top-K function to retain the original gated output. The maximum value, and the value of the original gated output excluding The remaining values, excluding the maximum value, are exponentially scaled to obtain the Top-K function processing result; the formula for processing the original gated output using the Top-K function is as follows: ； In the formula, Represents the Top-K function. Indicates the original gated output The Middle The maximum value, This represents a constant used to adjust the selector weights. Indicates the original gated output The Middle One value; S43, the Top-K function processing result is normalized using the Softmax function to obtain a set of selector weights representing the contribution of different time patterns to future predictions at the current time point; wherein, the formula for normalizing the Top-K function processing result using the Softmax function is: ； In the formula, This represents the set of selector weights. This represents the Softmax function.
7. The short-term power load forecasting method according to claim 1, characterized in that, In S4, the formula for weighted summation of the set of predicted outputs using the set of selector weights is as follows: ； In the formula, This represents the short-term power load forecast value. Represents the weight of the first selector in the set of selector weights. Each selector weight This indicates the features after the interaction. Represents the first of the plurality of parallel predictors One predictor; Indicates the first The prediction output obtained by the first predictor predicting the post-interaction features, the first predictor... The prediction output obtained by the predictor predicting the post-interaction features is the first prediction output in the set of prediction outputs. One predicted output, This represents the total number of parallel predictors.
8. The short-term power load forecasting method according to claim 1, characterized in that, Prior to S4, training of the time pattern selector and predictor is also included; The overall loss function for training the temporal pattern selector and predictor includes the prediction loss function and the selector balancing loss function; where, The prediction loss function is used to minimize the mean square error between the predicted output value and the true value. The selector balancing loss function is expressed as follows: ； In the formula, This represents the selector's balance loss function. Indicates selector weight variance Indicates selector weight The mean, This represents a very small, non-zero constant. The total loss function is expressed as: ； In the formula, Represents the total loss function, Denotes the prediction loss function, This represents the total number of trainable parameters in the model. and This is a hyperparameter.
9. A short-term power load forecasting system, characterized in that, The method applied to the short-term power load forecasting method as described in any one of claims 1 to 8 includes: The data acquisition module is used to acquire historical power load data and at least one covariate data as the raw input sequence; A multi-scale decomposable mixing module is used to decompose the original input sequence into multiple sub-sequences at different time scales and mix the multiple sub-sequences to obtain multi-scale mixed features. The dual-dependency interaction module is used to perform dual-dependency interaction modeling of the time dependency and channel dependency of the multi-scale hybrid features to obtain the interactive features. An adaptive multi-predictor synthesis module is used to adaptively calculate the multi-scale mixed features using a time pattern selector to obtain a set of selector weights that characterize the contribution of different time patterns to future predictions at the current time point. Multiple parallel predictors using time pattern projection are used to perform parallel predictions on the interactive features to obtain a set of prediction outputs. The set of prediction outputs is weighted and summed using the set of selector weights to obtain the short-term power load prediction value.
10. A short-term power load forecasting device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory, which, when executed by the processor, implements the short-term power load forecasting method as described in any one of claims 1 to 8.