A method and system for modeling variable relationships of cross-scale cross-band power consumption loads

Abstract

This invention discloses a method and system for modeling cross-scale and cross-frequency band electricity load variable relationships, belonging to the field of intelligent energy prediction technology. The method includes: acquiring multivariate time series of electricity load and converting it into channel-frequency and channel-scale views; generating and sorting ranking scores based on global indicators; permuting the channel-frequency and channel-scale views; using a pass-through proxy to propagate the gradient of the ranking scores to the permuted views to maintain gradient propagation; arranging potentially correlated segments into adjacent positions; performing cross-deformable convolution on the rearranged views to dynamically capture cross-variable, cross-frequency band, and cross-scale cross-correlation; and performing bi-branch prediction based on the captured cross-correlation to obtain the final electricity load prediction result. This invention alleviates the problem of non-local distribution of key relationships in the views, reduces noise interference, and improves the accuracy and robustness of electricity load prediction.

Description

Technical Field

[0001] This invention relates to the field of intelligent energy forecasting technology, and more specifically to a method and system for modeling the relationship between cross-scale and cross-frequency band electricity load variables. Background Technology

[0002] With the increasing proportion of renewable energy generation and the increasingly complex operating environment of the power system, electricity load is influenced by multiple factors, including meteorological conditions, renewable energy output, calendar factors, historical electricity prices, user behavior, and dispatch strategies, exhibiting strong volatility, non-stationarity, multi-scale, and multi-frequency coupling characteristics. The accuracy of electricity load forecasting largely depends on modeling the complex relationships between different influencing factors. Accurately characterizing the cross-influence relationships of variables such as historical electricity load, meteorological factors, renewable energy output, historical electricity prices, and dispatch information at different time scales and frequency bands is key to improving the accuracy of electricity load forecasting.

[0003] In recent years, time series research has primarily focused on intra-frequency and intra-scale correlations, i.e., the relationships between different channels within each frequency or scale after frequency or scale decomposition of the time series. However, these methods neglect potentially more important cross-correlations, i.e., relationships across different frequencies and scales, as well as between different channels. In real-world datasets, cross-correlations are sometimes even more significant than their corresponding intra-frequency correlations. Capturing such relationships is a challenging two-dimensional modeling problem. Inspired by the significant success of computer vision in capturing two-dimensional dependencies, time series are transformed into channel-frequency and channel-scale views with image structures. Compared to the original time series format, these views provide richer representations of temporal features and offer a more intuitive layout for two-dimensional cross-correlations. In this context, applying convolutional networks to channel-frequency and channel-scale views is an efficient and effective strategy. However, this image-like structure with embedded temporal information has the following characteristics that significantly affect model performance:

[0004] The key variable relationships are not concentrated. Electricity load is influenced by multiple factors, including historical load, weather conditions, renewable energy output, historical electricity prices, calendar factors, and dispatch strategies. The effective correlations between these factors often span different variables, frequency bands, and scales, and are not limited to a local neighborhood. Ordinary convolution cannot directly capture these non-locally distributed key relationships, and expanding the receptive field of convolution can easily introduce noise and redundant information.

[0005] Load variable views lack regular spatial attributes. Channel-frequency and channel-scale views differ from natural images, where row and column dimensions represent variables, frequencies, or scales respectively, lacking a clear geometric spatial structure. Directly using convolution, deformable convolution, or attention mechanisms for modeling can easily lead to problems such as difficulty in learning sampling locations, interference from noisy features, and spurious overfitting, thereby reducing the stability and accuracy of electricity load forecasting.

[0006] Therefore, providing a method and system that can flexibly handle channel-frequency views and channel-scale views, and effectively capture cross-variable, cross-frequency band, and cross-scale cross-correlation relationships of electricity load variables, is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] To address the problems of existing electricity load forecasting methods that struggle to capture cross-variable, cross-frequency band, and cross-scale cross-correlation, and the low prediction accuracy caused by the non-concentrated distribution of key relationships in the view and the lack of regular spatial attributes, this invention proposes a cross-scale and cross-frequency band electricity load variable relationship modeling method and system. Through view generation, adaptive two-dimensional rearrangement, cross-deformable convolution, and bi-branch prediction, high-precision electricity load forecasting is achieved.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, embodiments of the present invention provide a method for modeling the relationship between cross-scale and cross-frequency band electricity load variables, comprising the following steps: S1. Obtain the multivariate time series of electricity load and convert it into a channel-frequency view and a channel-scale view; S2. Generate a ranking score based on the global indicator and rank accordingly. Permutate the channel-frequency view and channel-scale view. Pass the gradient of the ranking score to the permuted view through a pass-through proxy to maintain gradient propagation. Arrange potential related segments related to power load to adjacent positions. S3. On the rearranged view, the sampling points are initialized in a cross shape and cross-deformable convolution is performed through a dynamic mask modulation mechanism to dynamically capture cross-variable, cross-frequency band and cross-scale cross-correlation. S4. Based on the captured cross-correlation, perform bi-branch prediction to obtain the final electricity load prediction result.

[0009] Furthermore, in step S1, the generation of the channel-scale view includes: performing multi-scale sequence feature extraction operations on the multivariate time series using multiple sets of parallel one-dimensional convolution operators, embedding the feature at the sequence level via a linear layer, thereby generating the channel-scale view. The generation of the channel-frequency view includes: performing a fast Fourier transform on the multivariate time series data and dividing it into multiple non-overlapping segments, embedding the segments using a linear layer, and generating the channel-frequency view.

[0010] Furthermore, in step S2, a sorting score is generated based on the global indicator and sorted accordingly. The channel-frequency view and channel-scale view are then permuted, specifically including: Global indicators for the channel dimension and global indicators for the frequency or scale dimension are generated respectively. The dimensions of the global indicators are summed to obtain the channel sorting score and the frequency or scale sorting score. The argmax function is used to calculate the target arrangement order corresponding to each sorting score. The rows and columns of the channel-frequency view or the channel-scale view are permuted according to the target arrangement order.

[0011] Furthermore, in step S2, the gradient of the sorting score is passed to the permuted view through a pass-through proxy to maintain gradient propagation, and potential related segments associated with electricity load are arranged to adjacent positions, specifically including: The rows and columns of the channel-frequency view or the channel-scale view are permuted according to the target arrangement order to obtain the permuted view; Simultaneously, the sorting score is replaced according to the same target sorting order to obtain the replaced sorting score; During backpropagation, the ratio of the permuted sort score to the permuted sort score after gradient truncation is calculated, and a gradient-containing all-one matrix is ​​constructed as the gradient injection matrix. The gradient injection matrix is ​​multiplied element-wise with the permuted view to obtain the final rearranged view, so that the gradient of the sorting score is backpropagated to the parameter update process of the global indicator.

[0012] Furthermore, in step S3, on the rearranged view, a cross-shaped initialization of sampling points is adopted and cross-deformable convolution is performed through a dynamic mask modulation mechanism. Specifically, this includes: asymmetrically expanding the receptive field along both the frequency and channel dimensions, with the initial set of sampling points being:

[0013] in, represents the receptive field size in the frequency dimension and the channel dimension, respectively, and x and y represent the offset of the sampling point relative to the center point in the frequency dimension and the channel dimension, respectively. Represents a two-dimensional integer coordinate space; An offset is generated using two-dimensional depthwise separable convolution, and an adaptive sampling position is obtained by combining the initial sampling points. Features are extracted at the adaptive sampling position through bilinear interpolation, which are used to interact with the center point features to strengthen the correlation between variables that have a significant impact on power load fluctuations.

[0014] Furthermore, the adaptive sampling position formula is as follows:

[0015] in, This represents the offset learned to adapt to the correlation distribution among electricity load variables. This indicates the adaptive sampling location used to capture electrical load-related features. Let M represent the initial set of sampling locations, N represent the number of variable channels related to electrical load, and M represent the number of channels. f This represents the number of frequency bands after dividing the electricity load data into frequency bands, num represents the total number of sampling points corresponding to each center point, Φ1 represents the two-dimensional depthwise separable convolution operator, and CF shuf This represents the rearranged channel-frequency view.

[0016] Furthermore, in step S3, dynamically capturing cross-variable, cross-frequency, and scale cross-correlation specifically includes: a modulation mechanism based on dynamic masking to learn modulation coefficients and masking thresholds for interaction between variables in the electricity load data; and using a differentiable soft mask to suppress the values ​​of the modulation coefficients below the masking threshold to near zero, thereby obtaining the masked modulation coefficients. The extracted features are modulated using the masked modulation coefficients, and the modulated features are fused with the center point features to obtain an enhanced power load view that captures cross-correlation. Furthermore, the specific calculation method for the differentiable soft mask is as follows: Using convolution to learn modulation coefficients and mask thresholds through interactions between variables in electricity load data Calculate the intermediate variables and the masked modulation coefficients:

[0017] in, 'g' represents the modulation coefficient, and 'g' represents the temperature coefficient. Indicates an intermediate variable.

[0018] Furthermore, in step S4, a two-branch prediction is performed based on the captured cross-correlation to obtain the final electricity load prediction result, specifically including: The original order of the processed channel-scale view and channel-frequency view is restored to obtain the channel-scale enhancement features and channel-frequency enhancement features with the original order restored. The channel-scale enhancement features that restore the original order are flattened and input into a multilayer perceptron to obtain the temporal scale prediction result. The channel-frequency enhancement features that restore the original order are input into two multilayer perceptrons respectively to predict the real and imaginary parts of the frequency domain result; The real and imaginary parts are combined into complex-valued data, and an inverse fast Fourier transform is performed to obtain the initial prediction result in the time domain. The initial time-domain prediction results are refined using a linear layer (Linearf) to obtain frequency branch prediction results. The time-domain scale prediction results and the frequency-domain scale prediction results are added together and fused to obtain the final electricity load prediction result.

[0019] In a second aspect, embodiments of the present invention provide a cross-scale, cross-frequency band electricity load variable relationship modeling system, which applies the method described in the first aspect, and the system includes: The view generation module is used to acquire multivariate time series of electricity load and convert them into channel-frequency view and channel-scale view; An adaptive two-dimensional rearrangement module is used to generate a sorting score based on a global indicator and sort accordingly, permuting the channel-frequency view and the channel-scale view, passing the gradient of the sorting score to the permuted view through a pass-through proxy to maintain gradient propagation, and arranging potentially related segments related to power load to adjacent positions. The cross-deformable convolution module is used to perform cross-deformable convolution on the rearranged view by initializing sampling points in a cross shape and using a dynamic mask modulation mechanism to dynamically capture cross-variable, cross-frequency band and cross-scale cross-correlation. The dual-branch prediction module is used to perform dual-branch prediction based on the captured cross-correlation to obtain the final electricity load prediction result.

[0020] The beneficial effects of the above-described technical solutions provided in the embodiments of the present invention include at least the following: This invention transforms one-dimensional time series into an image-like structure with rich temporal features by constructing channel-frequency views and channel-scale views. This provides an intuitive layout basis for capturing two-dimensional cross-correlation across variables, frequency bands, and scales, overcoming the limitation of existing methods that only focus on correlations within frequency and scale.

[0021] This invention uses adaptive two-dimensional rearrangement to generate robust sorting scores based on global indicators, arranging segments with potential dependencies in the view to adjacent positions, effectively alleviating the problem of non-local distribution of key variable relationships in the view; at the same time, it solves the gradient discontinuity problem of two-dimensional discrete sorting operations through a pass-through surrogate mechanism.

[0022] This invention employs a cross-deformable convolution method and a non-uniform cross initialization strategy to independently and asymmetrically expand the receptive field along the frequency and channel dimensions, with the number of sampling points increasing only linearly. This significantly reduces computational overhead while capturing long-distance cross-correlation. Simultaneously, it enhances robustness against spurious correlations and noise interference in complex power data by adaptively shielding noise and irrelevant features through a dynamic masking modulation mechanism.

[0023] This invention employs a dual-branch prediction method, performing independent predictions in the scale domain and frequency domain and then fusing the results for output. This fully leverages the trend characteristics of electricity load data at different scales and the periodic characteristics at different frequency bands, thereby further improving prediction accuracy. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0025] Figure 1 This is a flowchart of the cross-scale and cross-frequency band power load variable relationship modeling method provided in the embodiments of the present invention.

[0026] Figure 2 This is a framework diagram of a cross-scale, cross-frequency band power load variable relationship modeling system provided in an embodiment of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Example 1: like Figure 1 As shown in the figure, this invention discloses a method for modeling the relationship between cross-scale and cross-frequency band electricity load variables, including the following steps: S1. Obtain the multivariate time series of electricity load and convert it into a channel-frequency view and a channel-scale view to reconstruct the one-dimensional time series signal into a two-dimensional image structure with rich frequency and time domain feature representations, providing a structured input basis for subsequent two-dimensional convolution modeling.

[0029] S2. Generate ranking scores based on global indicators and rank accordingly. Permutate the channel-frequency view and channel-scale view. Pass the gradient of the ranking scores to the permuted view through a pass-through proxy to maintain gradient propagation. Arrange potential related segments related to electricity load to adjacent positions to alleviate the problem of non-local distribution of key variable relationships in the view, so that the convolutional network can effectively capture long-distance dependencies in the local neighborhood.

[0030] S3. On the rearranged view, the sampling points are initialized in a cross shape and cross-deformable convolution is performed through a dynamic mask modulation mechanism to dynamically capture cross-variable, cross-frequency band and cross-scale cross-correlation, so as to break through the limitation of the fixed receptive field of traditional convolution, realize non-uniform and adaptive correlation modeling, and suppress the interference of noise and irrelevant features.

[0031] S4. Based on the captured cross-correlation, perform bi-branch prediction to obtain the final electricity load prediction result, so as to make full use of the trend characteristics of electricity load data at different scales and the periodic characteristics at different frequency bands to obtain the final electricity load prediction result.

[0032] This invention reconstructs one-dimensional time-series signals into a two-dimensional image-like structure by constructing channel-frequency and channel-scale views. This provides an intuitive layout basis for capturing cross-variable, cross-frequency, and cross-scale cross-correlation, overcoming the limitations of existing methods that only focus on intra-frequency or intra-scale correlations. Through adaptive two-dimensional rearrangement, segments with potential dependencies in the view are arranged in adjacent positions, effectively alleviating the problem of non-local distribution of key relationships. A pass-through surrogate mechanism is used to solve the gradient discontinuity problem of discrete sorting operations. Through cross-deformable convolution, a non-uniform cross initialization strategy is adopted to independently and asymmetrically expand the receptive field along the frequency and channel dimensions, so that the number of sampling points increases only linearly. This reduces computational overhead while capturing long-distance cross-correlation. Combined with a dynamic masking modulation mechanism, noise and irrelevant features are adaptively shielded, enhancing robustness against spurious correlations and noise interference. Through bi-branch prediction, predictions are made independently in the scale and frequency domains and then fused for output, making full use of the trend and periodic features of electricity load data, further improving prediction accuracy.

[0033] The following is a detailed description of the cross-scale and cross-frequency band power load variable relationship modeling method of the present invention: Step S1: Obtain the multivariate time series of electricity load and convert it into a channel-frequency view and a channel-scale view. Specifically: Preliminary extraction of time series features is performed. These features include historical electricity load data, meteorological data, historical electricity price data, calendar data (including date type, holiday identifiers, seasonal information, and dispatch strategy-related data). These multivariate data collectively constitute a set of influencing factors for electricity load forecasting. After preliminary extraction, they are reorganized into channel-scale and channel-frequency views.

[0034] (1) Generation of channel-scale view: To avoid using average pooling operations that may lead to information loss, a series of parallel one-dimensional convolution operations are employed, with a kernel size of [missing value]. , Indicate the number of scales to extract multi-scale sequences:

[0035] in, Indicates the first The first channel in the Sequences at various scales The scale is represented by N, the number of variable channels is represented by T, and the length of the input sequence is represented by d. j Indicates the first The size of the one-dimensional convolution kernel corresponding to each scale.

[0036] Subsequently, Each linear layer is embedded at the sequence level to obtain a channel-scale view. ,in, Indicates the size of the hidden dimension.

[0037] (2) Generation of channel-frequency view: To decouple different frequencies, the data is first divided into... after the Fast Fourier Transform. There are several non-overlapping segments. The data following each segment can be represented as... This represents the hidden layer dimension. Subsequently, a linear layer is used for embedding at the fragment level to obtain the channel-frequency view. .

[0038] Step S2: Generate a ranking score based on the global indicator and rank accordingly. Permutate the channel-frequency view and channel-scale view. Pass the gradient of the ranking score to the permuted view through a pass-through proxy to maintain gradient propagation. Arrange potentially related segments associated with electrical load to adjacent positions. Specifically: First, an adaptive two-dimensional reordering is performed to address the issue that related components in the channel-frequency and channel-scale views are often non-locally distributed. Guided by a global indicator, a robust ranking score is adaptively generated to determine the ranking in the scale or frequency dimension and the channel dimension.

[0039] After the rearrangement operation, adjacent segments are more likely to have potential dependencies, enabling the convolutional network to directly capture two-dimensional dependencies. (This will be discussed later.) and Perform the same operation on both views, to The view is used as an example for illustration.

[0040] To mitigate the impact of short-term noise on channel and frequency ranking learning, a global indicator is introduced. By taking a long-term perspective, the indicator can learn robust and generalizable sorting.

[0041] The channel ranking score is obtained by summing the dimensions of the indicators. and frequency sorting score Subsequently, argmax is used to calculate the target permutation order for rearrangement. and :

[0042] in, Indicates the order in which the channel targets are arranged. Indicates the order of targets on the frequency scale. This represents the index for finding the maximum value. This indicates that the result is a vector containing N real numbers. This means that the result is a sequence containing M. f A vector of real numbers, Indicates the number of segments.

[0043] The objective of this invention is based on and Sorting is performed, but the discreteness and non-differentiability of standard sorting operations prevent it from being directly used for gradient-based optimization. As an alternative, soft sorting methods provide a differentiable approximation by calculating the probability distribution of the relative order of elements; however, this method may introduce noise and compromise the accuracy and intuitiveness of the sorting results.

[0044] To address this problem, this invention proposes a method that maintains both gradient propagation and accuracy, achieving results similar to traditional sorting methods. (Applied sequentially) and right Perform a permutation to obtain a view with a learnable order. At the same time, apply the same substitution. and ,get and This is to ensure the correct propagation of the gradient subsequently. Because... The gradient will be lost due to non-differentiable permutation operations, which can be addressed by injection. and The gradient, explicitly defined during backpropagation, actually employs a pass-through surrogate to ensure adaptive ranking learning:

[0045] in, and This represents a matrix of all ones with a gradient. Methods for truncating gradients This represents the final rearranged view. Recording is done during the rearrangement process. and This is so that the reverse rearrangement can be performed later.

[0046] Step S3: On the rearranged view, sample points are initialized in a cross shape and cross-deformable convolution is performed using a dynamic mask modulation mechanism to dynamically capture cross-variable, cross-frequency band, and cross-scale cross-correlation. Specifically: Cross-deformable convolution is performed, which dynamically adapts to the shape of the convolution kernel, thus effectively capturing geometric transformations. Specifically, for the initial set of sampling positions on the regular sampling grid of the convolution window... It predicts a data-related offset from the input features. By incorporating this offset, the sampling point is moved to a new adaptive position. .

[0047] To better capture cross-correlation in views lacking well-defined spatial properties, a cross-deformable convolution method is proposed. This method enhances the effectiveness of standard deformable convolution on channel-frequency and channel-scale views by introducing non-uniform cross-expansion initialization to adapt to specially designed image-like views and achieving robust interaction through dynamic mask-based modulation.

[0048] For ease of description, a two-dimensional grid coordinate system is established on the view, and it is assumed that the given convolution window is... Centered on this coordinate system, it is used to capture segments related to the central segment. In this coordinate system, shaft and The axes correspond to the frequency dimension and the channel dimension, respectively.

[0049] Initial sampling is performed; the standard deformable convolution initialization strategy uses a square shape and uniform sampling. When the receptive field size is... At that time, the initial shape Defined as:

[0050] This square initialization causes the number of sampling points to increase quadratically relative to the receptive field, resulting in significant computational overhead and noise. To efficiently expand the receptive field to capture complex cross-correlation, this invention designs a cross-shaped initialization method that expands linearly along both the frequency and channel dimensions. Its layout is formally defined as follows:

[0051] in, represents the receptive field size in the frequency dimension and the channel dimension, respectively, and x and y represent the offset of the sampling point relative to the center point in the frequency dimension and the channel dimension, respectively. This represents a two-dimensional integer coordinate space. This method enables the model to expand its receptive field independently and asymmetrically along both dimensions, with the number of sampling points increasing only linearly. More importantly, it allows the model to capture internal correlations along the frequency and channel dimensions by default, and further explore cross-dimensional dependencies based on this.

[0052] In addition, the standard strategy will be applied to each location Initializing a sampling point presents a discrepancy with the non-uniform distribution of cross-correlation. To address this issue, this invention designs a non-uniform sampling strategy for different locations. Different numbers of sampling points are assigned. Specifically, the location... Number of samples at the location Distance from it to the center point The distribution is determined and follows a normal distribution:

[0053] in, , and It is a hyperparameter used to adjust the normal distribution.

[0054] The initial set of all sampling points in all convolutional windows is formally defined as:

[0055] in, This represents the total number of sampling points corresponding to each center point. This strategy encourages adaptive 2D rearrangement, moving more relevant segments closer to the grid center.

[0056] After initialization, use two-dimensional depthwise separable convolution. Generate offset This is then used to calculate the adaptive sampling position. :

[0057] in, This represents the offset learned to adapt to the correlation distribution among electricity load variables. This indicates the adaptive sampling location used to capture electrical load-related features. Let M represent the initial set of sampling locations, N represent the number of variable channels related to electrical load, and M represent the number of channels. f This represents the number of frequency bands after dividing the electricity load data into frequency bands, num represents the total number of sampling points corresponding to each center point, Φ1 represents the two-dimensional depthwise separable convolution operator, and CF shuf This represents the rearranged channel-frequency view.

[0058] Due to adaptive position Typically, fractional coordinates cannot be perfectly aligned with discrete grid points on the feature map; therefore, bilinear interpolation is used, denoted as . Extracting features at these precise locations Used to interact with the center point: .

[0059] Adaptive Interaction: Considering that the number of effective dependencies varies across different convolutional windows, a modulation mechanism based on dynamic masks is proposed for cross-deformable convolution. This mechanism is based on adaptive thresholding. In modulation coefficient The model adaptively masks irrelevant interactions between the center point and the sampling points, enabling it to dynamically capture sparse dependencies.

[0060] Additional convolutions are used to learn the modulation coefficients for the interactions between variables in the electricity load data. and mask threshold :

[0061] in, and and The same type of convolution is used. To ensure gradient propagation, a differentiable soft mask is used. Values ​​below the mask threshold are suppressed to near zero, thus obtaining the masked modulation coefficients:

[0062] in, The modulation coefficient is represented by g, and the temperature coefficient is preferably 0. , This represents an intermediate auxiliary variable.

[0063] Finally, feature interactions are performed between sampling points and their corresponding center points to capture cross-correlation, and the coefficients are used to determine the correlation. Modulation:

[0064] in, This represents the modulated characterization. This represents the characterization after aggregation. This represents the output representation with residuals after the interaction.

[0065] Step S4: Perform bi-branch prediction based on the captured cross-correlation to obtain the final electricity load prediction result, specifically: Two-branch prediction is used: After applying cross-deformable convolution, respectively... and Perform an inverse rearrangement operation to restore its original order, resulting in channel-scale augmented features with the original order restored. and channel-frequency enhancement features that restore the original order Then, predictions are made independently on each branch.

[0066] (1) Channel-scale branch prediction: Unlike most existing methods, this approach uses a more expressive MLP-based prediction head to replace the linear prediction head, thereby enhancing prediction capabilities.

[0067] in, Used to flatten Convert to prediction results Indicates the predicted length.

[0068] (2) Channel-frequency branch prediction: First, predict the data in the frequency domain, then transform the results back to the time domain and refine the predictions:

[0069] in, Represented as the real part, Represented as the imaginary part, Subsequently, and Combined into complex-valued data, and then subjected to an inverse fast Fourier transform to obtain... and :

[0070] in, Used to refine the prediction results in the time domain. The final prediction is:

[0071] Existing electricity load forecasting methods that incorporate external influencing factors have shortcomings in modeling relationships across variables, scales, and frequency bands. This invention improves the ability to capture complex variable relationships and enhances prediction robustness through adaptive two-dimensional rearrangement, cross-deformable convolution, non-uniform cross initialization, and dynamic mask modulation mechanisms.

[0072] The following example illustrates the cross-scale and cross-frequency band power load variable relationship modeling method of the present invention: This invention underwent performance testing on the Nordic Electricity Market Dataset (NP). This dataset records relevant data from the NordPool electricity market. This experiment selected electricity load as the prediction target and used wind power and historical electricity prices as external influencing factors to verify the effectiveness of this invention in electricity load prediction tasks involving external influencing factors. Performance was compared with mainstream baseline models, including TimeKAN, TimeXer, TiDE, DLinear, PatchTST, DUET, Amplifier, xPatch, and TFT, using MSE, MAE, and other evaluation metrics to verify performance.

[0073] With an input step size of 168 and a prediction step size of 24, the future electricity load is predicted using historical electricity load, wind power, and historical electricity prices. The specific results are shown in Table 1.

[0074] Table 1: Electricity load forecast results from the Nordic electricity market dataset

[0075] Experimental results show that the method of this invention achieves optimal performance in the electricity load forecasting task on the Nordic electricity market dataset. Compared with representative forecasting models such as TimeKAN, TimeXer, TiDE, DLINear, PatchTST, DUET, Amplifier, xPatch, and TFT, the method of this invention reduces the MSE to 0.011 and the MAE to 0.070, achieving optimal results in both evaluation metrics, indicating that this invention can more accurately predict future electricity load changes.

[0076] Furthermore, compared to the suboptimal TFT method, the MSE of the method in this invention is reduced from 0.013 to 0.011, a relative reduction of approximately 15.4%; the MAE is reduced from 0.081 to 0.070, a relative reduction of approximately 13.6%. This indicates that the present invention, through channel-frequency view and channel-scale view modeling, can more fully capture the cross-variable, cross-scale, and cross-frequency cross-relationships between historical electricity load, wind power, and historical electricity prices; at the same time, adaptive two-dimensional rearrangement and cross-deformable convolution can reduce irrelevant features and noise interference, thereby improving the accuracy and robustness of electricity load prediction.

[0077] Example 2: like Figure 2 As shown, based on the same inventive concept, this embodiment of the invention also provides a cross-scale, cross-frequency band electricity load variable relationship modeling system, which includes: The view generation module is used to acquire multivariate time series of electricity load and convert them into channel-frequency view and channel-scale view; An adaptive two-dimensional rearrangement module is used to generate a sorting score based on a global indicator and sort accordingly, permuting the channel-frequency view and the channel-scale view, passing the gradient of the sorting score to the permuted view through a pass-through proxy to maintain gradient propagation, and arranging potentially related segments related to power load to adjacent positions. The cross-deformable convolution module is used to perform cross-deformable convolution on the rearranged view by initializing sampling points in a cross shape and using a dynamic mask modulation mechanism to dynamically capture cross-variable, cross-frequency band and cross-scale cross-correlation. The dual-branch prediction module is used to perform dual-branch prediction based on the captured cross-correlation to obtain the final electricity load prediction result.

[0078] The process involves using a view generation module to convert the input time series into a graph-like structure. By modeling this graph-like structure, complex cross-correlation is captured simultaneously in both channel-frequency and channel-scale views. The channel-scale view characterizes the impact of different driving variables on electricity load at scales such as short-term fluctuations, medium-term trends, and long-term cycles, while the channel-frequency view characterizes the daily, weekly, peak load fluctuations, periodic variations, and high-frequency disturbance components within the electricity load and its driving variables. The same operation is then performed in parallel on both views.

[0079] An adaptive 2D reordering module adaptively generates robust rankings based on information from the frequency or scale dimension and the channel dimension. Simultaneously, an efficient method is proposed to ensure correct gradient propagation during the 2D reordering process. In this way, discrete channel-frequency and channel-scale views are reordered, adaptively arranging potentially dependent components in adjacent positions.

[0080] Subsequently, on the rearranged view, cross-shaped initialization of sampling points is employed, and a cross-deformable convolution is performed using a dynamic mask modulation mechanism to extract frequency or scale information and capture complex cross-correlation. Two key enhancement steps are designed for standard deformable convolution. The first enhancement is non-uniform cross initialization, enabling the model to expand its receptive field independently and asymmetrically along two dimensions while performing non-uniform sampling, thus better adapting to the rearranged view. The second enhancement is a dynamic mask-based modulation mechanism that selectively masks noise and irrelevant features by learning to generate dynamic masks.

[0081] Finally, the original order of the views is restored, and the final prediction result is generated based on the multilayer perceptron in the two-branch prediction module.

[0082] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0083] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for modeling the relationship between electricity load variables across scales and frequency bands, characterized in that, Includes the following steps: S1. Obtain the multivariate time series of electricity load and convert it into a channel-frequency view and a channel-scale view; S2. Generate a ranking score based on the global indicator and rank accordingly. Permutate the channel-frequency view and channel-scale view. Pass the gradient of the ranking score to the permuted view through a pass-through proxy to maintain gradient propagation. Arrange potential related segments related to power load to adjacent positions. S3. On the rearranged view, the sampling points are initialized in a cross shape and cross-deformable convolution is performed through a dynamic mask modulation mechanism to dynamically capture cross-variable, cross-frequency band and cross-scale cross-correlation. S4. Based on the captured cross-correlation, perform bi-branch prediction to obtain the final electricity load prediction result.

2. The method as described in claim 1, characterized in that, In step S1, the generation of the channel-scale view includes: performing multi-scale sequence feature extraction on the multivariate time series using multiple sets of parallel one-dimensional convolution operators, embedding the feature at the sequence level through a linear layer, thereby generating the channel-scale view. The generation of the channel-frequency view includes: performing a fast Fourier transform on the multivariate time series data and dividing it into multiple non-overlapping segments, embedding the segments using a linear layer, and generating the channel-frequency view.

3. The method as described in claim 1, characterized in that, In step S2, a sorting score is generated based on a global indicator and sorted accordingly. The channel-frequency view and channel-scale view are then permuted, specifically including: Global indicators for the channel dimension and global indicators for the frequency or scale dimension are generated respectively. The dimensions of the global indicators are summed to obtain the channel sorting score and the frequency or scale sorting score. The argmax function is used to calculate the target arrangement order corresponding to each sorting score. The rows and columns of the channel-frequency view or the channel-scale view are permuted according to the target arrangement order.

4. The method as described in claim 3, characterized in that, In step S2, the gradient of the sorting score is passed to the permuted view through a pass-through proxy to maintain gradient propagation, and potential related segments associated with electricity load are arranged to adjacent positions. Specifically, this includes: The rows and columns of the channel-frequency view or the channel-scale view are permuted according to the target arrangement order to obtain the permuted view; Simultaneously, the sorting score is replaced according to the same target sorting order to obtain the replaced sorting score; During backpropagation, the ratio of the permuted sort score to the permuted sort score after gradient truncation is calculated, and a gradient-containing all-one matrix is ​​constructed as the gradient injection matrix. The gradient injection matrix is ​​multiplied element-wise with the permuted view to obtain the final rearranged view, so that the gradient of the sorting score is backpropagated to the parameter update process of the global indicator.

5. The method as described in claim 1, characterized in that, In step S3, on the rearranged view, sampling points are initialized in a cross shape and cross-deformable convolution is performed using a dynamic mask modulation mechanism. Specifically, this includes: asymmetrically expanding the receptive field along both the frequency and channel dimensions, with the initial set of sampling points as follows: in, represents the receptive field size in the frequency dimension and the channel dimension, respectively, and x and y represent the offset of the sampling point relative to the center point in the frequency dimension and the channel dimension, respectively. Represents a two-dimensional integer coordinate space; An offset is generated using two-dimensional depthwise separable convolution, and an adaptive sampling position is obtained by combining the initial sampling points. Features are extracted at the adaptive sampling position through bilinear interpolation, which are used to interact with the center point features to strengthen the correlation between variables that have a significant impact on power load fluctuations.

6. The method as described in claim 5, characterized in that, The formula for the adaptive sampling position is: in, This represents the offset learned to adapt to the correlation distribution among electricity load variables. This indicates the adaptive sampling location used to capture electrical load-related features. Let M represent the initial set of sampling locations, N represent the number of variable channels related to electrical load, and M represent the number of channels. f This represents the number of frequency bands after dividing the electricity load data into frequency bands, num represents the total number of sampling points corresponding to each center point, Φ1 represents the two-dimensional depthwise separable convolution operator, and CF shuf This represents the rearranged channel-frequency view.

7. The method as described in claim 1, characterized in that, In step S3, the dynamic capture of cross-variable, cross-frequency, and scale cross-correlation specifically includes: a modulation mechanism based on dynamic masking to learn modulation coefficients and masking thresholds for interaction between variables in the electricity load data; and using a differentiable soft mask to suppress the values ​​of the modulation coefficients below the masking threshold to near zero, thereby obtaining the masked modulation coefficients. The extracted features are modulated using the masked modulation coefficients, and the modulated features are fused with the center point features to obtain an enhanced view of power load after capturing cross-correlation.

8. The method as described in claim 7, characterized in that, The specific calculation method for the differentiable soft mask is as follows: Using convolution to learn modulation coefficients and mask thresholds through interactions between variables in electricity load data Calculate the intermediate variables and the masked modulation coefficients: in, 'g' represents the modulation coefficient, and 'g' represents the temperature coefficient. Indicates an intermediate variable.

9. The method as described in claim 1, characterized in that, In step S4, a two-branch prediction is performed based on the captured cross-correlation to obtain the final electricity load prediction result, specifically including: The original order of the processed channel-scale view and channel-frequency view is restored to obtain the channel-scale enhancement features and channel-frequency enhancement features with the original order restored. The channel-scale enhancement features that restore the original order are flattened and input into a multilayer perceptron to obtain the temporal scale prediction result. The channel-frequency enhancement features that restore the original order are input into two multilayer perceptrons respectively to predict the real and imaginary parts of the frequency domain result; The real and imaginary parts are combined into complex-valued data, and an inverse fast Fourier transform is performed to obtain the initial prediction result in the time domain. The initial time-domain prediction results are refined using a linear layer (Linearf) to obtain frequency branch prediction results. The time-domain scale prediction results and the frequency-domain scale prediction results are added together and fused to obtain the final electricity load prediction result.

10. A system for modeling the relationship between cross-scale and cross-frequency band electricity load variables, characterized in that, The system comprises: The view generation module is used to acquire multivariate time series of electricity load and convert them into channel-frequency view and channel-scale view; An adaptive two-dimensional rearrangement module is used to generate a sorting score based on a global indicator and sort accordingly, permuting the channel-frequency view and the channel-scale view, passing the gradient of the sorting score to the permuted view through a pass-through proxy to maintain gradient propagation, and arranging potentially related segments related to power load to adjacent positions. The cross-deformable convolution module is used to perform cross-deformable convolution on the rearranged view by initializing sampling points in a cross shape and using a dynamic mask modulation mechanism to dynamically capture cross-variable, cross-frequency band and cross-scale cross-correlation. The dual-branch prediction module is used to perform dual-branch prediction based on the captured cross-correlation to obtain the final electricity load prediction result.

Images