Power system load prediction method and device based on multi-scale time-frequency domain analysis

Abstract

The invention provides a power system load prediction method and device based on multi-scale time-frequency domain analysis, and relates to the technical field of power system data processing. The method comprises the steps of obtaining a historical time sequence of a power system load; performing normalization processing on the historical time sequence to obtain normalized time sequence data; extracting periodic features and non-periodic features of a multi-time-scale time-frequency domain in the normalized time series data; for each time scale in the multiple time scales, predicting a prediction sequence of the time scale according to the periodic feature and the non-periodic feature of the time-frequency domain of the time scale; and carrying out adaptive integration on the multi-time scale prediction sequence, and carrying out inverse normalization processing to obtain a power system load prediction result. According to the embodiment of the invention, high-precision prediction is realized by cooperatively extracting periodic features and non-periodic features of different time scales (such as a day period and a week period) in the time sequence data.

Description

Technical Field

[0001] This application relates to the field of power system data processing technology, and in particular to a power system load forecasting method and apparatus based on multi-scale time-frequency domain analysis. Background Technology

[0002] The power system is inextricably linked to people's lives and production; ensuring its stable operation is crucial from both economic and safety perspectives. Load forecasting, based on historical power system data, provides a forecast of future power load levels to assist power authorities in developing appropriate dispatching plans. In recent years, with strategic reforms of the power system, load forecasting has faced even greater challenges.

[0003] Load forecasting is categorized based on the time span between the forecast date and the current time, including long-term, medium-term, short-term, and ultra-short-term load forecasts. Different time spans correspond to different functions. Long-term load forecasts typically span years and are used to guide future planning for a specific region; medium-term load forecasts span months and are used for market assessments over a shorter timeframe; short-term forecasts typically span days and primarily provide a reference for power grid dispatching; ultra-short-term forecasts span hours and are mainly used for data monitoring and emergency warnings.

[0004] In the task of short-term or ultra-short-term load forecasting, a large number of studies have been constantly proposing new methods to reduce the error of load forecasting. These methods can be divided into two categories: 1) traditional forecasting methods and 2) machine learning forecasting methods.

[0005] Traditional forecasting methods were primarily used in early load forecasting tasks. For example, the multiplicative AR (Auto-Regressive) model uses least squares to estimate the AR parameters and ultimately makes predictions based on the optimal AR parameters. The ARMA (Auto-Regression and Moving Average) model was also a widely used method in the early days. However, due to performance limitations, traditional methods have been gradually replaced by machine learning-related methods.

[0006] In recent years, the rapid development of deep learning has spurred the emergence of a variety of deep prediction models, including those based on recurrent neural networks (such as DeepAR and LSTNet), convolutional neural networks (such as TCN and SCINet), Transformers (such as Informer and Autoformer), and graph neural networks (such as MTGNN, StemGNN, and AGCRN). DeepAR stands for DeepAutoRegressive, a probabilistic prediction model based on autoregressive recurrent neural networks; LSTNet stands for Long and Short-term Time-series Network, a time-series prediction model combining long short-term memory and convolutional layers; TCN stands for Temporal Convolutional Network, a time-series modeling architecture using dilated causal convolutions; SCINet stands for Sample Convolution and Interaction Network, a model that improves time-series prediction through sample convolution and interaction decomposition; Transformer is a transformer; Informer stands for Information-efficient Transformer, a Transformer optimized for long-series prediction, proposing a sparse attention mechanism; MTGNN stands for Multivariate Time Series Graph Neural Network, a multivariate time-series graph neural network that automatically learns dependencies between variables; StemGNN stands for Spatio-Temporal Enhanced Graph Neural Network, a graph neural network that enhances spatio-temporal modeling, combining spectral convolution and temporal encoding; AGCRN stands for Adaptive Graph Convolutional Recurrent. Network is an adaptive graph convolutional recurrent network that can dynamically learn graph structures.

[0007] While these deep learning models achieve good predictive performance in specific scenarios, their complex network architecture often leads to high computational overhead during training or inference. Furthermore, when available training data is limited, a large number of parameters can easily affect the model's robustness. Therefore, methods based on Multi-Layer Perceptrons (MLPs), such as DLinear, have attracted attention due to their simple structure, low complexity, and superior performance. DLinear stands for Decomposition Linear Network. However, these methods rely on pointwise mapping to capture temporal relationships, making it difficult to handle global dependencies in time series, failing to capture periodic patterns effectively, and being susceptible to local fluctuations and redundant information, leading to performance bottlenecks. Therefore, it is urgent to solve this technical problem. Summary of the Invention

[0008] In view of the above problems, this application is made to provide a power system load forecasting method, apparatus, and electronic device based on multi-scale time-frequency domain analysis to overcome or at least partially solve the above problems. The technical solution is as follows: Firstly, a power system load forecasting method based on multi-scale time-frequency domain analysis is provided, the method comprising: Obtain historical time series of power system load; The historical time series is normalized to obtain normalized time series data; Extracting periodic and non-periodic features in the time-frequency domain across multiple time scales from normalized time-series data; For each time scale in the multi-time scale, the prediction sequence for that time scale is predicted based on the periodic and non-periodic characteristics of the time-frequency domain. The prediction sequences at multiple time scales are adaptively integrated and then inversely normalized to obtain the power system load prediction results.

[0009] Secondly, a power system load forecasting device based on multi-scale time-frequency domain analysis is provided, the device comprising: The input module is used to obtain the historical time series of power system load; The normalization module is used to normalize historical time series data to obtain normalized time series data. The multi-timescale learning and frequency domain enhancement module is used to extract the periodic and non-periodic features of the time-frequency domain at multiple time scales in the normalized time series data; for each time scale in the multi-time scale, the prediction sequence of that time scale is predicted based on the periodic and non-periodic features of the time-frequency domain at that time scale. The prediction fusion layer is used to adaptively integrate prediction sequences across multiple time scales. The inverse normalization module is used to perform inverse normalization on the results of adaptive integration of prediction sequences at multiple time scales to obtain the power system load prediction results.

[0010] Thirdly, an electronic device is provided, comprising a processor and a memory, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the power system load forecasting method based on multi-scale time-frequency domain analysis as described in any of the preceding claims.

[0011] By means of the above technical solutions, the power system load forecasting method, device and electronic equipment based on multi-scale time-frequency domain analysis provided in the embodiments of this application propose a multi-scale periodic parallel modeling framework, which achieves high-precision forecasting by collaboratively extracting periodic and non-periodic features of different time scales (such as daily cycle and weekly cycle) in time series data. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the description of the embodiments of this application will be briefly introduced below.

[0013] Figure 1 A flowchart of the power system load forecasting method based on multi-scale time-frequency domain analysis provided in an embodiment of this application is shown; Figure 2 This paper illustrates the overall framework diagram for parallel modeling of daily and weekly cycles provided in an embodiment of this application. Figure 3 This paper illustrates a framework diagram of the daily cycle learning and frequency domain enhancement module provided in an embodiment of this application. Figure 4 This diagram illustrates the framework of the periodic learning and frequency domain enhancement module provided in an embodiment of this application. Figure 5 This illustration shows a comparison between the daily cycle processing module and the CycleNet method in the overall model framework provided in this application embodiment; Figure 6 A schematic diagram illustrating the prediction effect of the verification cycle module provided in an embodiment of this application is shown; Figure 7 The diagram shows the structure of the power system load forecasting device based on multi-scale time-frequency domain analysis provided in this application embodiment; Figure 8 A structural diagram of a power system load forecasting device based on multi-scale time-frequency domain analysis provided in another embodiment of this application is shown; Figure 9A structural diagram of an electronic device provided in an embodiment of this application is shown. Detailed Implementation

[0014] Exemplary embodiments of the present application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application will be thorough and complete, and will fully convey the scope of the present application to those skilled in the art.

[0015] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such use can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the term "comprising" and its variations should be interpreted as open-ended terms meaning "including but not limited to."

[0016] To address the aforementioned technical problems, embodiments of this application provide a power system load forecasting method based on multi-scale time-frequency domain analysis, such as... Figure 1 As shown, the power system load forecasting method based on multi-scale time-frequency domain analysis may include the following steps S101 to S105: Step S101: Obtain the historical time series of power system load; Step S102: Normalize the historical time series to obtain normalized time series data; Step S103: Extract the periodic and non-periodic features of the time-frequency domain at multiple time scales from the normalized time-series data. Step S104: For each time scale in the multi-time scale, predict the prediction sequence of that time scale based on the periodic and non-periodic characteristics of the time-frequency domain of that time scale. Step S105: Adaptively integrate the prediction sequences of multiple time scales and perform inverse normalization to obtain the power system load prediction results.

[0017] This embodiment proposes a multi-scale periodic parallel modeling framework, which achieves high-precision prediction by collaboratively extracting periodic and non-periodic features of different time scales (such as daily and weekly cycles) from time series data.

[0018] This application provides a possible implementation method in its embodiments, where the normalized time-series data has... D The time series X of a variable or channel, the objective is based on L Historical observations predict the futureH The numerical value of the step, expressed mathematically as follows: ; For elements in time series X, t For time, R For the set of real numbers, This represents the load forecast results for the power system.

[0019] The statistical properties of time series data (such as the mean) often change over time; this phenomenon is called distribution drift. When a model trained on historical training data is applied to future data, this drift can lead to a decline in model performance. This embodiment introduces the RevIN (Reversible Instance Normalization) strategy. Specifically, after data input and before output, the data needs to undergo a process of deleting and adding back statistical attributes.

[0020] Normalization module: ; Inverse normalization module: ; In the formula, μ and σ These represent the mean and standard deviation of the input time window, respectively. ε This is a small constant introduced to ensure numerical stability. This method employs invertible instance normalization that does not include learnable affine parameters.

[0021] This application provides a possible implementation method where multiple time scales include daily and weekly cycles. This allows for the extraction of periodic and non-periodic features in the daily time-frequency domain of the normalized time-series data, as well as the extraction of periodic and non-periodic features in the weekly time-frequency domain of the normalized time-series data.

[0022] Figure 2 This paper illustrates the overall framework diagram for parallel modeling of daily and weekly cycles provided in an embodiment of this application. Figure 2 The process includes inputting historical time series data of power system load, normalization, daily cycle learning and frequency domain enhancement modules, weekly cycle learning and frequency domain enhancement modules, a prediction fusion layer, inverse normalization, and outputting power system load prediction results. Here, the daily cycle learning and frequency domain enhancement module focuses on capturing the smallest significant daily cycle and its associated aperiodic features; the weekly cycle learning and frequency domain enhancement module models macroscopic weekly cycle patterns; then, the prediction fusion layer adaptively integrates the multi-scale prediction results and performs inverse normalization to obtain the power system load prediction results.

[0023] This application provides a possible implementation method where, in steps S103 and S104 above, when multiple time scales include the daily cycle, the periodic and non-periodic features of the daily cycle in the time-frequency domain are extracted from the normalized time-series data; based on the periodic and non-periodic features of the daily cycle in the time-frequency domain, the predicted sequence of the daily cycle is predicted, which may specifically include the following steps A1 to A5: Step A1: The normalized time-series data after extracting the periodic and non-periodic features in the daily time-frequency domain is denoted as... , , L 0 less than or equal to L ; Step A2, given a priori day cycle length W 0 D When there are multiple variables, a learnable cycle is generated. Initialized as a matrix of all zeros, with a cycle period of [missing information]. Q Global sharing among variables, and obtaining periodic components through periodic copying operations. C , length is W 0 cycle period Q It will work together with the set backbone prediction module to perform gradient backpropagation training, and finally learn a representation that is different from the initial zero value and can reveal the internal periodicity of the sequence. Step A3, from Elimination cycle Q The generated periodic components C Historical cycle section To obtain the residual components ; Step A4, Input the segmented frequency domain pattern learning module used to extract mid-to-high frequency features of the sequence, and obtain the prediction results of the residual components. ; Step A5, With cycle period Q The generated periodic components C The prediction cycle part By superimposing the data, a daily cycle prediction sequence is obtained. .

[0024] This application provides a possible implementation method, which involves the following steps to control the cycle period. Q Perform alignment and repetition operations to obtain periodic components C Historical cycle section and prediction cycle part : Cycle Q Shift left t modW 0 bits obtained ,in t mod W 0 represents the current sequence sample in Q The relative position index in; Will repeat Next, splice together get: ; Will repeat Next, splice together get: .

[0025] This application provides a possible implementation method, which employs the following processing procedure to... And perform adaptive filtering: ; ; In the formula, This represents the Fast Fourier Transform along the time dimension. This represents the inverse fast Fourier transform along the time dimension; Indicates to Learnable parameters for filtering; ⊙ indicates element-wise multiplication.

[0026] Figure 3 This illustration shows a framework diagram of the diurnal cycle learning and frequency domain enhancement module provided in an embodiment of this application. Figure 3 This includes the time-series data after input normalization (i.e., input...). ), daily periodic feature learning module, alignment and repetition to obtain periodic components C Historical cycle section Elimination (i.e., from input) Elimination cycle Q The generated periodic components C Historical cycle section To obtain the residual components The segmented frequency domain pattern learning module (which can be denoted as the Time_seg module) is used to extract mid-to-high frequency features from the sequence, and can extract residual components. Input the segmented frequency domain pattern learning module used to extract mid-to-high frequency features of the sequence, and obtain the prediction results of the residual components. Alignment and repetition to obtain periodic components C The prediction cycle part Adaptive filtering and superposition (soon) Compared with the adaptive filtering process (Overlay), outputting the predicted sequence for the daily cycle. .

[0027] This application embodiment provides a possible implementation method, in which step A4 above will... Input the segmented frequency domain pattern learning module used to extract mid-to-high frequency features of the sequence, and obtain the prediction results of the residual components. Specifically, this may include the following steps A4-1 to A4-3: Step A4-1, for Simplified representation Based on the preset hyperparameter seg_num, x The average score is 2 seg_num Each sub-fragment is zero-padded at the beginning and end, and the sub-fragments are stacked to obtain the desired result. ; Step A4-2, for Fast Fourier Transform along the time dimension to obtain , and the learnable parameters processed by the softmax activation function Obtained by point-by-point multiplication , and then along The first dimension is summed, completing the adaptive weighted summation operation along the first dimension, and the frequency domain extraction operation along the time and variable dimensions, resulting in frequency domain data with extracted features. ;

[0028] ; ; In the formula, It is a set of learning parameters; It is the element in the set of learning parameters; Step A4-3, for Perform an inverse transform, and obtain the output after passing through a multilayer perceptron. , yes A simplified representation of .

[0029] In this embodiment, the segmented frequency domain pattern learning module, namely the Time_seg module, innovatively integrates time-series segmentation, Fourier transform, and learnable frequency domain weight matrix, adaptive feature weighting, and other methods to achieve targeted extraction of high-frequency features (such as sudden fluctuations and short-term trends) in aperiodic components. This effectively solves the problem of insufficient modeling of mid-to-high frequency information in traditional methods and significantly improves the prediction accuracy of aperiodic components.

[0030] This application embodiment provides a possible implementation method. When multiple time scales include a periodic cycle in steps S103 and S104 above, the periodic and non-periodic features of the periodic cycle in the time-frequency domain are extracted from the normalized time-series data. Based on the periodic and non-periodic features of the periodic cycle in the time-frequency domain, the predicted sequence of the periodic cycle is predicted, which may specifically include the following steps B1 to B7: Step B1, the normalized time-series data after extracting periodic and non-periodic features in the time-frequency domain is denoted as... , , L 1 is less than or equal to L ; Step B2, given a priori periodic length W 1 of D When there are multiple variables, a learnable cycle is generated. Initialized as a matrix of all zeros, with a cycle period of [missing information]. Global sharing among variables, and obtaining periodic components through periodic copying operations. , length is W 1 cycle period It will work together with the set backbone prediction module to perform gradient backpropagation training, and finally learn a representation that is different from the initial zero value and can reveal the internal periodicity of the sequence. Step B3, will After average pooling, key sequences containing long-period features are retained. ; Step B4, will Elimination cycle The generated historical periodic components yield residual components. ; Step B5, will Input the segmented frequency domain pattern learning module used to extract mid-to-high frequency features of the sequence, and obtain the prediction results of the residual components. ; Step B6, will With cycle period The generated predicted periodic components are superimposed to obtain a preliminary predicted sequence of the periodic cycle. ; Step B7, will Inputting a linear layer yields a periodic prediction sequence. .

[0031] In this embodiment, capturing periodic features requires inputting at least one longer period of time-series data into the system, significantly increasing computational complexity. However, for short-term predictions, the model only needs to extract key information about trend changes in the weekly data. Based on considerations of reducing computational load and not compromising model performance, the periodic learning and frequency domain enhancement module comprises three sub-modules: a pooling module, a periodic feature learning module, and a linear layer. By fusing pooling and multilayer perceptron (MLP) collaborative extraction and restoration architecture, key information from long sequences is extracted, improving computational efficiency while maintaining prediction accuracy.

[0032] Figure 4 This illustration shows a framework diagram of the periodic learning and frequency domain enhancement module provided in an embodiment of this application. Figure 4 This includes the time-series data after input normalization (i.e., input...). Pooling (soon) After average pooling, key sequences containing long-period features are retained. ), Periodic feature learning module, alignment and repetition to obtain periodic components Historical cyclical components in the data, removal (soon) Elimination cycle The generated historical periodic components yield residual components. The segmented frequency domain pattern learning module (which can be denoted as the Time_seg module) is used to extract mid-to-high frequency features from sequences. Input the segmented frequency domain pattern learning module used to extract mid-to-high frequency features of the sequence, and obtain the prediction results of the residual components. Alignment and repetition to obtain periodic components The predicted periodic components, adaptive filtering, and superposition (i.e., the...) By superimposing the predicted periodic component after adaptive filtering, a preliminary predicted sequence of the periodicity is obtained. ), linear layer (soon) Input linear layer), output periodic prediction sequence .

[0033] This application embodiment provides a possible implementation method. Step S105 above adaptively integrates the prediction sequences of multiple time scales and performs inverse normalization processing to obtain the power system load prediction result. Specifically, it may include the following step C1: Step C1, Prediction sequence of daily cycles and the predicted sequence of the cycle Adaptive integration and inverse normalization are then performed to obtain the power system load forecast results. ; ; In the formula, ⊙ represents element-wise multiplication. and These are learnable parameters processed by the softmax activation function.

[0034] The above introduces Figure 1 The embodiments shown have various implementation methods for each step. The following will further illustrate the power system load forecasting method based on multi-scale time-frequency domain analysis of this application through specific embodiments.

[0035] Electricity load data exhibits a clear periodicity in its overall trend, and identifying these periodic patterns is crucial for model prediction. Previous methods have typically employed different approaches to capture these patterns. These include the long-range modeling capabilities of Transformers, the use of large convolutional kernels in ModernTCN (Modern Temporal Convolutional Network) to enhance temporal dependencies, and the optimization of long-sequence modeling capabilities through segmented iterative processing in SegRNN (Segment Recurrent Neural Network).

[0036] CycleNet (Cyclical Neural Network) uses globally shared periodic segments to represent power load patterns and generates multi-period predictions by repeating these segments. This explicit modeling of periodicity is more efficient than previous methods that relied on complex architectures. However, CycleNet has the following drawbacks: (1) CycleNet effectively captures the periodic patterns of time series data through an explicit periodic modeling mechanism. From the perspective of frequency domain analysis, this method performs well in learning low-frequency components (corresponding to periodic features), but due to the limitations of its basic network architecture built with linear layers or multilayer perceptrons, it has insufficient ability to represent mid-to-high frequency components (reflecting short-term fluctuations and non-periodic features), thus restricting the optimization space of model prediction accuracy.

[0037] (2) The existing CycleNet architecture has a significant effect on feature extraction at a single cycle scale (such as daily cycle). However, in the face of the multi-scale nested cycle characteristics (coupling effect of daily cycle and weekly cycle) in power load data, the current simplified processing strategy of using a fixed maximum cycle length W will introduce spectral confusion, which will lead to model performance degradation.

[0038] (3) In the task of power system load forecasting, in order to effectively extract long-period features such as the cycle period, a longer input time window is required. However, long-sequence input will significantly increase the computational complexity, resulting in a non-linear increase in the time and space overhead of model training and inference.

[0039] This embodiment constructs a complete frequency-domain adaptive feature extraction and fusion framework. The core innovations of this scheme are mainly reflected in three aspects: First, a learnable frequency-domain filtering mechanism is proposed, which intelligently filters different frequency components through dynamic spectrum weighting technology, particularly enhancing the extraction capability of low-frequency periodic features. Second, a multi-scale time-frequency joint analysis method is developed, combining time-series segmentation, short-time Fourier transform, and adaptive frequency-domain weighting, effectively solving the problem of insufficient modeling of mid-to-high frequency aperiodic features in traditional methods. Finally, a hierarchical long-period extraction architecture is designed, employing pooling and cross-scale feature fusion techniques to achieve explicit modeling of long-period patterns for the first time. The core of this scheme lies in the aforementioned special methodological system for time-domain and frequency-domain feature processing, especially the innovative technical solutions such as dynamic spectrum weighting, time-frequency joint analysis, and multi-scale feature fusion.

[0040] This embodiment specifically demonstrates... Figure 2 , Figure 3 , Figure 4 The framework shown is used to implement power system load forecasting, and the case study is as follows.

[0041] 1) Introduction to the dataset The ETT (Electricity Transformer Temperature) open-source dataset includes ETTh2 with a time interval of 1 hour and ETTm2 with a time interval of 15 minutes, containing multivariate data such as load, oil temperature, and ambient temperature of power transformers from a certain site.

[0042] Load dataset: Electricity load values ​​for a city at 15-minute intervals.

[0043] 2) Calculation example 1 To verify that the Time_seg module has better prediction performance for medium and high frequency data compared to a simple multilayer perceptron, only the daily cycle processing module in the overall model framework is compared with the CycleNet method.

[0044] Training and testing were performed using a GPU (Graphics Processing Unit). The training epoch was 30, and the batch size was set to 128. The final test results are as follows: Figure 5 As shown.

[0045] existFigure 5 In Chinese, MSE stands for Mean Squared Error, which is the average of the squared differences between the predicted and actual values. It reflects the accuracy of the prediction, and the smaller the value, the better.

[0046] MAE stands for Mean Absolute Error, which is the average of the absolute differences between the predicted and actual values, and measures the degree of deviation in the prediction.

[0047] As shown in the results of Example 1, the overall framework plus the daily cycle learning and frequency domain enhancement module have steadily improved the prediction performance of various datasets and prediction lengths. Thus, the Time_seg framework proposed in this method breaks through the prediction bottleneck of medium and high frequency data.

[0048] 3) Calculation example 2 To verify the impact of the weekly cycle module on prediction performance, a weekly cycle learning and frequency domain enhancement module and a daily cycle learning and frequency domain enhancement module were added to the overall framework to process weekly and daily cycle data. This was compared to having only a daily cycle learning and frequency domain enhancement module. Figure 6 As shown. The input time window length for processing daily periodic data is... L 0=96, the length of the input time window for processing periodic data is L 1 = 168 or 672.

[0049] exist Figure 6 In Chinese, MSE stands for Mean Squared Error, which is the average of the squared differences between the predicted and actual values. It reflects the accuracy of the prediction, and the smaller the value, the better.

[0050] MAE stands for Mean Absolute Error, which is the average of the absolute differences between the predicted and actual values, and measures the degree of deviation in the prediction.

[0051] As shown in Example 2, adding a periodic module to extract periodic characteristics significantly improves the prediction performance for each dataset.

[0052] This embodiment can achieve the following technical effects: (1) By designing a learnable frequency domain filtering mechanism, spectral domain features are selected for the input subsequence, significantly enhancing the ability of the periodic feature learning module to capture low-frequency components (such as daily and weekly cycles). This module optimizes the modeling effect of periodic patterns by dynamically suppressing high-frequency noise and redundant information.

[0053] (2) The Time_seg module innovatively integrates time-series segmentation, Fourier transform and learnable frequency domain weight matrix, adaptive feature weighting and other methods to achieve targeted extraction of high-frequency features (such as sudden fluctuations and short-term trends) in non-periodic components. This effectively solves the problem of insufficient modeling of medium and high frequency information in traditional methods and significantly improves the prediction accuracy of non-periodic components.

[0054] (3) For long-period patterns such as the weekly cycle, a key feature extraction architecture based on pooling and linear projection is proposed. Through cross-scale feature fusion, explicit decoupling and joint modeling of nested long-period features are realized for the first time.

[0055] It should be noted that the sequence numbers of the steps in the above embodiments do not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. In practical applications, all the above possible implementation methods can be arbitrarily combined in a combined manner to form possible embodiments of this application, which will not be described in detail here.

[0056] Based on the power system load forecasting method based on multi-scale time-frequency domain analysis provided in the above embodiments, and based on the same inventive concept, this application also provides a power system load forecasting device based on multi-scale time-frequency domain analysis.

[0057] Figure 7 This is a structural diagram of a power system load forecasting device based on multi-scale time-frequency domain analysis provided in an embodiment of this application. Figure 7 As shown, the power system load forecasting device based on multi-scale time-frequency domain analysis may specifically include an input module 710, a normalization module 720, a multi-time-scale learning and frequency domain enhancement module 730, a prediction fusion layer 740, and an inverse normalization module 750.

[0058] Input module 710 is used to obtain the historical time series of power system load; The normalization module 720 is used to normalize historical time series data to obtain normalized time series data. The multi-timescale learning and frequency domain enhancement module 730 is used to extract the periodic and non-periodic features of the time-frequency domain of the multi-timescale time series data after normalization; for each timescale in the multi-timescale, the prediction sequence of the timescale is predicted based on the periodic and non-periodic features of the time-frequency domain of that timescale. Prediction fusion layer 740 is used for adaptive integration of prediction sequences at multiple time scales; The inverse normalization module 750 is used to perform inverse normalization processing on the results of adaptive integration of prediction sequences at multiple time scales to obtain the power system load prediction results.

[0059] This application provides a possible implementation method in its embodiments, where the normalized time-series data has... D The time series X of 1 variable, the objective is based on L Historical observations predict the future H The numerical value of the step, expressed mathematically as follows: ; For elements in time series X, t For time, R For the set of real numbers, This represents the load forecast results for the power system.

[0060] This application provides a possible implementation method, where multiple time scales include daily and weekly cycles; for example... Figure 8 As shown, the multi-timescale learning and frequency domain enhancement module 730 may include a daily cycle learning and frequency domain enhancement module 731 and a weekly cycle learning and frequency domain enhancement module 732. The daily cycle learning and frequency domain enhancement module 731 is used for: The normalized time-series data after extracting the periodic and non-periodic features in the daily time-frequency domain is denoted as... , , L 0 less than or equal to L ; Given a priori day cycle length W 0 D When there are multiple variables, a learnable cycle is generated. Initialized as a matrix of all zeros, with a cycle period of [missing information]. Q Global sharing among variables, and obtaining periodic components through periodic copying operations. C , length is W 0 cycle period Q It will work together with the set backbone prediction module to perform gradient backpropagation training, and finally learn a representation that is different from the initial zero value and can reveal the internal periodicity of the sequence. from Elimination cycle Q The generated periodic components C Historical cycle section To obtain the residual components ; Will Input the segmented frequency domain pattern learning module used to extract mid-to-high frequency features of the sequence, and obtain the prediction results of the residual components. ; Will With cycle period Q The generated periodic components C The prediction cycle part By superimposing the data, a daily cycle prediction sequence is obtained. .

[0061] This application embodiment provides a possible implementation, wherein the daily period learning and frequency domain enhancement module 731 is further used for: The following steps are used to determine the cycle period. Q Perform alignment and repetition operations to obtain periodic components C Historical cycle section and prediction cycle part : Cycle Q Shift left t mod W 0 bits obtained ,in t mod W 0 represents the current sequence sample in Q The relative position index in; Will repeat Next, splice together get: ; Will repeat Next, splice together get: .

[0062] This application embodiment provides a possible implementation, wherein the daily period learning and frequency domain enhancement module 731 is further used for: The following processing procedure is used for And perform adaptive filtering: ; ; In the formula, This represents the Fast Fourier Transform along the time dimension. This represents the inverse fast Fourier transform along the time dimension; Indicates to Learnable parameters for filtering; ⊙ indicates element-wise multiplication.

[0063] This application embodiment provides a possible implementation, wherein the daily period learning and frequency domain enhancement module 731 is further used for: right Simplified representation Based on the preset hyperparameter seg_num, x The average score is 2 seg _numEach sub-fragment is zero-padded at the beginning and end, and the sub-fragments are stacked to obtain the desired result. ; right Fast Fourier Transform along the time dimension to obtain , and the learnable parameters processed by the softmax activation function Obtained by point-by-point multiplication , and then along The first dimension is summed, completing the adaptive weighted summation operation along the first dimension, and the frequency domain extraction operation along the time and variable dimensions, resulting in frequency domain data with extracted features. ;

[0064] ; ; In the formula, It is a set of learning parameters; It is the element in the set of learning parameters; right Perform an inverse transform, and obtain the output after passing through a multilayer perceptron. , yes A simplified representation of .

[0065] This application embodiment provides a possible implementation, wherein the periodic learning and frequency domain enhancement module 732 is further used for: The normalized time-series data after extracting periodic and non-periodic features in the time-frequency domain is denoted as... , , L 1 is less than or equal to L ; Given a priori period length W 1 of D When there are multiple variables, a learnable cycle is generated. Initialized as a matrix of all zeros, with a cycle period of [missing information]. Global sharing among variables, and obtaining periodic components through periodic copying operations. , length is W 1 cycle period It will work together with the set backbone prediction module to perform gradient backpropagation training, and finally learn a representation that is different from the initial zero value and can reveal the internal periodicity of the sequence. Will After average pooling, key sequences containing long-period features are retained. ; Will Elimination cycle The generated historical periodic components yield residual components. ; Will Input the segmented frequency domain pattern learning module used to extract mid-to-high frequency features of the sequence, and obtain the prediction results of the residual components. ; Will With cycle period The generated predicted periodic components are superimposed to obtain a preliminary predicted sequence of the periodic cycle. ; Will Inputting a linear layer yields a periodic prediction sequence. .

[0066] This application embodiment provides a possible implementation, wherein the inverse normalization module 750 is further used for: Prediction sequence of daily cycles and the predicted sequence of the cycle Adaptive integration and inverse normalization are then performed to obtain the power system load forecast results. ; ; In the formula, ⊙ represents element-wise multiplication. and These are learnable parameters processed by the softmax activation function.

[0067] Based on the same inventive concept, embodiments of this application also provide an electronic device, including a processor and a memory, wherein a computer program is stored in the memory, and the processor is configured to run the computer program to execute the power system load forecasting method based on multi-scale time-frequency domain analysis of any of the above embodiments.

[0068] In an exemplary embodiment, an electronic device is provided, such as Figure 9 As shown, Figure 9 The illustrated electronic device 900 includes a processor 901 and a memory 903. The processor 901 and the memory 903 are connected, for example, via a bus 902. Optionally, the electronic device 900 may also include a transceiver 904. It should be noted that in practical applications, the transceiver 904 is not limited to one type, and the structure of this electronic device 900 does not constitute a limitation on the embodiments of this application.

[0069] Processor 901 may be a CPU (Central Processing Unit), GPU, DSP (Digital Signal Processor), ASIC (Application Specific Integrated Circuit), FPGA, or other programmable logic device, transistor logic device, hardware component, 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 application. Processor 901 may also be a combination that implements computing functions, such as including one or more microprocessor combinations, a combination of DSP and microprocessor, etc.

[0070] Bus 902 may include a pathway for transmitting information between the aforementioned components. Bus 902 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 902 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 9 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.

[0071] The memory 903 may be a ROM (Read Only Memory) or other type of static storage device capable of storing static information and instructions, RAM (Random Access Memory) or other type of dynamic storage device capable of storing information and instructions, or an 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 thereto.

[0072] The memory 903 stores computer program code that executes the scheme of this application, and its execution is controlled by the processor 901. The processor 901 executes the computer program code stored in the memory 903 to implement the content shown in the foregoing method embodiments.

[0073] Among them, electronic devices include, but are not limited to: mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), and in-vehicle terminals (such as in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 9 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0074] Those skilled in the art will clearly understand that the specific working process of the systems, devices, and modules described above can be referred to the corresponding process in the foregoing method embodiments. For the sake of brevity, it will not be repeated here.

[0075] Those skilled in the art will understand that the technical solution of this application, or all or part of it, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several program instructions to cause an electronic device (e.g., a personal computer, server, or network device) to execute all or part of the steps of the methods described in the embodiments of this application when running the program instructions. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0076] Alternatively, all or part of the steps of the foregoing method embodiments can be implemented by hardware (such as electronic devices like personal computers, servers, or network devices) associated with program instructions. The program instructions can be stored in a computer-readable storage medium. When the program instructions are executed by the processor of the electronic device, the electronic device executes all or part of the steps of the methods described in the embodiments of this application.

[0077] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that within the spirit and principles of this application, modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the corresponding technical solutions to leave the protection scope of this application.

Claims

1. A power system load forecasting method based on multi-scale time-frequency domain analysis, characterized in that, The method includes: Obtain historical time series of power system load; The historical time series is normalized to obtain normalized time series data; Extracting periodic and non-periodic features in the time-frequency domain across multiple time scales from normalized time-series data; For each time scale in the multi-time scale, the prediction sequence for that time scale is predicted based on the periodic and non-periodic characteristics of the time-frequency domain. The prediction sequences at multiple time scales are adaptively integrated and then inversely normalized to obtain the power system load prediction results.

2. The method according to claim 1, characterized in that, If the normalized time series data has D The time series X of 1 variable, the objective is based on L Historical observations predict the future H The numerical value of the step, expressed mathematically as follows: ; For elements in time series X, t For time, R For the set of real numbers, This represents the load forecast results for the power system.

3. The method according to claim 2, characterized in that, Multiple time scales include daily and weekly cycles; Extract the periodic and non-periodic features of the daily time-frequency domain from the normalized time-series data; Based on the periodic and non-periodic characteristics of the daily cycle in the time-frequency domain, a predicted sequence for the daily cycle is generated, including: The normalized time-series data after extracting the periodic and non-periodic features in the daily time-frequency domain is denoted as... , , L 0 less than or equal to L ; Given a priori day cycle length W 0 D When there are multiple variables, a learnable cycle is generated. Initialized as a matrix of all zeros, with a cycle period of [missing information]. Q Global sharing among variables, and obtaining periodic components through periodic copying operations. C , length is W 0 cycle period Q It will work together with the set backbone prediction module to perform gradient backpropagation training, and finally learn a representation that is different from the initial zero value and can reveal the internal periodicity of the sequence. from Elimination cycle Q The generated periodic components C Historical cycle section To obtain the residual components ; Will Input the segmented frequency domain pattern learning module used to extract mid-to-high frequency features of the sequence, and obtain the prediction results of the residual components. ; Will With cycle period Q The generated periodic components C The prediction cycle part By superimposing the data, a daily cycle prediction sequence is obtained. .

4. The method according to claim 3, characterized in that, The following steps are used to determine the cycle period. Q Perform alignment and repetition operations to obtain periodic components C Historical cycle section and prediction cycle part : Cycle Q Shift left t mod W 0 bits obtained ,in t mod W 0 represents the current sequence sample in Q The relative position index in; Will repeat Next, splice together get: ； Will repeat Next, splice together get: 。 5. The method according to claim 4, characterized in that, The method further includes: The following processing procedure is used for And perform adaptive filtering: ； ； In the formula, This represents the Fast Fourier Transform along the time dimension. This represents the inverse fast Fourier transform along the time dimension; Indicates to Learnable parameters for filtering; ⊙ indicates element-wise multiplication.

6. The method according to claim 3, characterized in that, Will Input the segmented frequency domain pattern learning module used to extract mid-to-high frequency features of the sequence, and obtain the prediction results of the residual components. ,include: right Simplified representation Based on the preset hyperparameter seg_num, x The average score is 2 seg_num Each sub-fragment is zero-padded at the beginning and end, and the sub-fragments are stacked to obtain... ; right Fast Fourier Transform along the time dimension to obtain , and the learnable parameters processed by the softmax activation function Obtained by point-by-point multiplication , and then along The first dimension is summed, completing the adaptive weighted summation operation along the first dimension, and the frequency domain extraction operation along the time and variable dimensions, resulting in frequency domain data with extracted features. ; ； ； In the formula, It is a set of learning parameters; It is the element in the set of learning parameters; right Perform an inverse transform, and obtain the output after passing through a multilayer perceptron. , yes A simplified representation of .

7. The method according to claim 3, characterized in that, Extract the periodic and non-periodic features of the time-frequency domain of the normalized time-series data; Based on the periodic and non-periodic characteristics of the periodic time-frequency domain, a predicted sequence of the periodicity is generated, including: The normalized time-series data after extracting periodic and non-periodic features in the time-frequency domain is denoted as... , , L 1 is less than or equal to L ; Given a priori period length W 1 of D When there are multiple variables, a learnable cycle is generated. Initialized as a matrix of all zeros, with a cycle period of [missing information]. Global sharing among variables, and obtaining periodic components through periodic copying operations. , length is W 1 cycle period It will work together with the set backbone prediction module to perform gradient backpropagation training, and finally learn a representation that is different from the initial zero value and can reveal the internal periodicity of the sequence. Will After average pooling, key sequences containing long-period features are retained. ; Will Elimination cycle The generated historical periodic components yield residual components. ; Will Input the segmented frequency domain pattern learning module used to extract mid-to-high frequency features of the sequence, and obtain the prediction results of the residual components. ; Will With cycle period The generated predicted periodic components are superimposed to obtain a preliminary predicted sequence of the periodic cycle. ; Will Inputting a linear layer yields a periodic prediction sequence. .

8. The method according to claim 7, characterized in that, The forecast sequences at multiple time scales are adaptively integrated and then inversely normalized to obtain the power system load forecast results, including: Prediction sequence of daily cycles and the predicted sequence of the cycle Adaptive integration and inverse normalization are then performed to obtain the power system load forecast results. ; ； In the formula, ⊙ represents element-wise multiplication. and These are learnable parameters processed by the softmax activation function.

9. A power system load forecasting device based on multi-scale time-frequency domain analysis, characterized in that, The device includes: The input module is used to obtain the historical time series of power system load; The normalization module is used to normalize historical time series data to obtain normalized time series data. The multi-timescale learning and frequency domain enhancement module is used to extract the periodic and non-periodic features of the time-frequency domain at multiple time scales in the normalized time series data; for each time scale in the multi-time scale, the prediction sequence of that time scale is predicted based on the periodic and non-periodic features of the time-frequency domain at that time scale. The prediction fusion layer is used to adaptively integrate prediction sequences across multiple time scales. The inverse normalization module is used to perform inverse normalization on the results of adaptive integration of prediction sequences at multiple time scales to obtain the power system load prediction results.

10. An electronic device, characterized in that, The system includes a processor and a memory, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the power system load forecasting method based on multi-scale time-frequency domain analysis as described in any one of claims 1 to 8.

Images