Power grid load prediction method based on multi-scale analysis

Abstract

The invention discloses a power grid load prediction method based on multi-scale analysis, and the method comprises the steps: 1, obtaining time series data of a power grid load, constructing a power grid load prediction sample set, inputting the sample set into a data preprocessing module, and constructing noise-removed multi-scale power grid time series data; 2, inputting the multi-scale power grid time sequence data into a feature extraction module to obtain local time dependence features and long-distance dependence features; 3, fusing the local time dependency feature and the long-distance dependency feature to obtain a fused feature; 4, inputting the fusion features into a prediction module to obtain a prediction value; 5, calculating an L1-L2 mixed loss function according to the predicted value and the true value, and iteratively updating the model to obtain a trained power grid load prediction model; and 6, performing prediction by using the trained power grid load prediction model. According to the method, the multi-scale characteristics of the power grid time sequence data are effectively extracted and utilized, so that the accuracy of power grid load prediction is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of time series prediction, in particular to a power grid load prediction method based on multi-scale analysis. BACKGROUND

[0002] Time series prediction is an important technical basis for improving decision-making efficiency and risk control capability in economic finance, industrial monitoring, traffic scheduling and other fields. The current technology for predicting time series data can be divided into three categories: statistical feature-based methods, traditional machine learning methods, and deep learning-based prediction methods.

[0003] Among them, the statistical feature method mainly relies on the time series, autocorrelation and periodicity characteristics of the data to construct a linear model or time series analysis model (such as ARIMA, SARIMA, etc.), and to realize the prediction of future values through parameter estimation. However, this method usually assumes that the data distribution is stable, and it is difficult to cope with the complex and variable time series structure in actual application, and the prediction error is usually large. Traditional machine learning methods, such as support vector machines and random forests, have certain flexibility in feature selection and fitting, but their performance is limited in capturing long-term dependence characteristics, trend changes and multi-scale rules of time series. In addition, traditional machine learning methods mostly rely on feature engineering, and the efficiency and accuracy of feature extraction are difficult to meet the actual prediction needs.

[0004] In recent years, deep learning-based time series prediction methods have gradually become the mainstream. Through convolutional neural networks (CNN), recurrent neural networks (RNN), long short-term memory networks (LSTM), and Transformer structures, the features in the time series can be automatically learned and extracted, reducing the dependence on artificial feature design and significantly improving the accuracy and generalization ability of prediction.

[0005] However, time series data often exhibits significantly different statistical characteristics and time series patterns at different sampling scales. For example, power load data in the power grid system may exhibit dramatic fluctuations in electricity consumption at the minute level, while at the hour or day level, it exhibits macro-trend stability. This multi-scale characteristic makes it difficult for traditional prediction methods to accurately capture the electricity consumption rules at different time scales, leading to difficulties in power grid scheduling decisions. Therefore, how to reasonably couple and integrate the feature information of power grid load time series data at multiple scales has become one of the key issues to improve the performance of prediction models. SUMMARY

[0006] The present application is to solve the above-mentioned deficiencies in the prior art, and proposes a power grid load prediction method based on multi-scale analysis, in order to reasonably couple and integrate the feature information of power grid load at multiple scales, thereby improving the accuracy of load prediction.

[0007] In order to achieve the above object, the present application realizes the technical scheme as follows: The power grid load prediction method based on multi-scale analysis has the characteristics that it comprises the following steps: Step S1, using N a plurality of sensors to collect historical power consumption data of families at fixed time intervals, thereby constructing a power grid historical load data set , wherein t represents the power grid historical load data of the N th family at the T th time step, is the total number of time steps; Step S2, constructing a power grid load prediction network, comprising a data preprocessing module, a double-branch feature extraction module, a feature enhancement module and a prediction module, and processing to obtain a predicted value ; wherein represents the predicted time step length; Step S2.1, the data preprocessing module comprises a wavelet denoising unit and a downsampling unit, and processes to obtain a denoised multi-scale power grid load time series data set , wherein represents the denoised power grid historical load data set at the th scale, and , wherein represents the power grid historical load data at the k th time step under the th scale, is the number of scales; Step S2.2, the double-branch feature extraction module comprises a Transformer branch and a CNN branch, and processes the first scales of the denoised power grid load time series data set and the last scales of the denoised power grid load time series data set respectively, to correspondingly obtain a multi-scale power grid historical load Transformer feature sequence set and a multi-scale power grid historical load CNN feature sequence set ; wherein represents the power grid historical load Transformer feature sequence at the th scale, and , wherein represents the power grid historical load fusion feature at the th time step under the th scale.Let represent the CNN feature sequence of the historical load of the power grid at the nth scale, and ,in, This represents the local characteristics of the historical load of the power grid at the t-th time step under the n-th scale. This indicates rounding down. The feature embedding dimension for the Transformer branch; Step S2.3, the feature enhancement module... Projection and interpolation upsampling processes are performed to obtain the historical load enhancement characteristic sequence of the power grid. ,in, This represents the local enhancement characteristics of the historical load of the power grid at the t-th time step under the n-th scale; simultaneously, for Interpolation and upsampling processing are performed to obtain the historical load fusion enhancement feature sequence of the power grid. ,in, This represents the historical load fusion enhancement characteristics of the power grid at the t-th time step under the n-th scale; thus, for and Weighted fusion is performed to obtain the weighted fusion feature set of historical load of the power grid. ;in, This represents the weighted and integrated characteristic sequence of historical loads of the power grid, and ,in This indicates the weighted fusion characteristics of historical load in the power grid; Step S2.4: The prediction module uses a fully connected network to... Process the data and output the predicted value. ,in, Indicates the first t After the first time step Forecast values ​​of grid load over a given time step; Step S3, based on and Constructing an L1-L2 hybrid loss function This is used to train the power grid load forecasting network to update the network parameters until the L1-L2 hybrid loss function is applied. The process continues until convergence, thus obtaining a well-trained power grid load prediction model to achieve power grid load prediction.

[0008] The power grid load forecasting method based on multi-scale analysis described in this invention is also characterized in that step 2.1 includes: Step S2.1.1, the wavelet denoising unit... The data is processed to obtain a denoised historical load dataset of the power grid. ;in, Indicates the denoised first... ta plurality of time steps N a plurality of household grid historical load data; Step S2.1.2, the down-sampling unit processes to obtain a denoised multi-scale grid historical load data set wherein, .

[0009] Further, the step 2.2 includes: Step S2.2.1, the Transformer branch performs sequence decomposition on to obtain a periodic feature set and a long-term trend feature set and performs intra-scale-inter-scale fusion on and to obtain a multi-scale grid historical load Transformer feature sequence set ; wherein, represents the long-term trend feature sequence of the nth scale grid historical load, and , represents the long-term trend feature of the nth scale at the time step; represents the periodic feature sequence of the nth scale grid historical load, , represents the periodic feature of the nth scale at the time step; Step S2.2.2, the CNN branch extracts local features of through a convolutional neural network to obtain a multi-scale grid historical load CNN feature sequence set .

[0010] Further, the sequence decomposition in step S2.2.1 is to calculate and according to formula (1) and formula (2) respectively; (1) = − (2) In formula (1) and formula (2), represents the time step index of the nth scale , represents the size of the time series aggregation window, is the stride, is the local position index within the time series aggregation window.

[0011] Furthermore, the intra-scale-inter-scale fusion in step S2.2.1 is achieved by using equations (3) and (4) to obtain the fused periodic features at the nth scale. and the long-term trend characteristics after integration Thus, by using equation (5), we can obtain : (3) (4) (5) In equations (3) and (4), Indicates a downsampling operation; Indicates an upsampling operation. This represents the periodicity feature after fusion at the (n-1)th scale. This represents the long-term trend characteristics after fusion at the (n-1)th scale. When n=1, let... = ,make = .

[0012] Furthermore, the downsampling operation in equation (3) include: Scale is Periodic characteristics Perform a linear projection to obtain the scale after the first projection. Periodic characteristics ,in, The scale after the first projection is 1. The t The periodic features at each time step, with lengths of the periodic features before and after the first projection being respectively... , ; right Perform a GELU activation function transformation to obtain the transformed scale as follows: Periodic characteristics ,in, The scale after transformation is The t Periodic characteristics of each time step; right Perform a linear projection to obtain the scale after the second projection. Periodic characteristics ,in, The scale after the second projection is 1. The t The periodic features at each time step, and the lengths of the periodic features before and after the second projection are respectively... , .

[0013] Further, the upsampling operation in formula (3) includes: linearly projecting the long-term trend feature with a scale of to obtain a long-term trend feature after the first projection with a scale of , wherein, represents the long-term trend feature of the i-th time step after the first projection with a scale of , and the lengths of the long-term trend features before and after the first projection are , respectively. t ; performing GELU activation function transformation on to obtain a long-term trend feature after the transformation with a scale of , wherein, represents the long-term trend feature of the i-th time step after the transformation with a scale of . t ; performing linear projection on to obtain a long-term trend feature after the second projection with a scale of , wherein, represents the long-term trend feature of the i-th time step after the second projection with a scale of , and the lengths of the long-term trend features before and after the second projection are , t respectively. .

[0014] Further, in step S2.3, is obtained by using formula (6) : (6) In formula (6), is a weight greater than 0 and less than 1.

[0015] Further, in step S3, the L1-L2 hybrid loss function is constructed by using formula (7) : (7) In formula (7), represents the power grid load prediction value of the i-th time step, is the power grid load true value of the i-th time step, , l ​​​​​​For the predicted time step, tanh is the hyperbolic tangent function.

[0016] The electronic device comprises a memory and a processor, wherein the memory is configured to store a program supporting the processor to execute the power grid load prediction method, and the processor is configured to execute the program stored in the memory.

[0017] The computer readable storage medium stores a computer program, wherein the computer program is configured to execute the steps of the power grid load prediction method when executed by a processor.

[0018] Compared with the prior art, the power grid load prediction method based on multi-scale analysis has the following advantages: 1. The power grid load prediction method based on multi-scale analysis can decouple complex power grid time series data, fully extract features between and within multiple scales, and improve the prediction ability of the model for power grid load.

[0019] 2. The double-branch feature extraction module can use different network architectures to extract time series features, improving the extraction ability of the model for power grid time series features.

[0020] 3. The L1-L2 hybrid loss function optimization improves the generalization and robustness of the model to adapt to different scale multi-scenario load prediction tasks. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 The power grid load prediction method based on multi-scale analysis comprises the following steps. DETAILED DESCRIPTION

[0022] In this embodiment, the power grid load prediction method based on multi-scale analysis is as follows: first, the original power grid load data is decomposed by multi-layer wavelet, and the corresponding high-frequency and low-frequency coefficients are extracted. Then, the high and low frequency coefficients obtained by decomposition are denoised by threshold strategy to weaken the interference of noise on model learning, and the coefficients are reconstructed into denoised time series data by inverse wavelet transform; secondly, the denoised power grid load is averaged down-sampled to obtain multi-scale data, which is input into the feature extraction branch module for feature extraction, and then multi-scale fusion is performed. In this way, complex time data can be easily analyzed, and the prediction ability of the model can be enhanced. Specifically, taking the power grid load prediction scene as an example, as shown in the figure, the following steps are included: Figure 1 Step S1, using N sensors to collect N ​Historical electricity consumption data of individual households, thereby constructing a historical load dataset for the power grid. ,in, Indicates the first t Each time step N Historical power grid load data for individual households T This represents the total number of time steps.

[0023] Step S2: Construct a power grid load forecasting network, including: a data preprocessing module, a dual-branch feature extraction module, a feature enhancement module, and a forecasting module, and perform... Processing is performed to obtain the predicted value. ;in, Indicates the time step of the prediction.

[0024] Step S2.1, the data preprocessing module includes: a wavelet denoising unit and a downsampling unit, and performs... The data is processed to obtain a denoised multi-scale power grid load time series dataset. ,in, Let represent the denoised historical load dataset of the power grid at the nth scale, and ,in, Represents the nth scale. Historical power grid load data at each time step k The number of scales; Step S2.1.1, Wavelet denoising unit pair The data is processed to obtain a denoised historical load dataset of the power grid. ;in, Indicates the denoised first... t Each time step N Historical load data of household power grids; among which, the wavelet denoising unit uses the Daubechies wavelet basis to decompose, threshold denoise, and reconstruct the power grid load data.

[0025] Step S2.1.2, downsampling unit The data is processed to obtain a denoised multi-scale historical load dataset of the power grid. ,in, ; The downsampling unit generates multi-scale data through one-dimensional average pooling, with a step size of 2.

[0026] Step S2.2: The dual-branch feature extraction module includes a Transformer branch and a CNN branch, and extracts features from the denoised front end respectively. Time-series dataset of power grid load at various scales and after Time-series dataset of power grid load at various scales The processing is performed, and a multi-scale power grid historical load Transformer feature sequence set and a multi-scale power grid historical load CNN feature sequence set are obtained correspondingly ; wherein, represents the n-th scale power grid historical load Transformer feature sequence, and , represents the n-th scale power grid historical load fusion feature at the t-th time step, represents the n-th scale power grid historical load CNN feature sequence, and , represents the n-th scale power grid historical load local feature at the t-th time step, represents the floor function, is the feature embedding dimension of the Transformer branch.

[0027] Step S2.2.1, the Transformer branch performs sequence decomposition on to obtain a cycle feature set and a long-term trend feature set , and performs intra-scale-inter-scale fusion on and to obtain a multi-scale power grid historical load Transformer feature sequence set ; wherein, represents the n-th scale power grid historical load long-term trend feature sequence, and , represents the n-th scale long-term trend feature at the t-th time step, represents the n-th scale power grid historical load cycle feature sequence, , represents the n-th scale cycle feature at the t-th time step, wherein, in step S2.2.1, sequence decomposition is performed according to formula (1) and formula (2) to calculate and ; (1) = − (2) In formula (1) and formula (2), represents the time step index of the n-th scale, , represents the size of the time series aggregation window,​​​​ This refers to stride length; in this implementation case, s is 24, which is one day. It is a local position index within the time series aggregation window, used to locate specific elements within the time series aggregation window.

[0028] Among them, intra-scale-inter-scale fusion is obtained by using equations (3) and (4) to obtain the periodic features after fusion at the nth scale. and the long-term trend characteristics after integration Thus, by using equation (5), we can obtain : (3) (4) (5) In equations (3) and (4), This indicates that a downsampling process is performed. This indicates that an upsampling process is performed; This represents the periodicity feature after fusion at the (n-1)th scale. This represents the long-term trend characteristics after fusion at the (n-1)th scale. When n=1, let... = ,make = .

[0029] In this embodiment, the downsampling process steps are as follows: First, the scale is Periodic characteristics Perform linear projection to obtain the periodic features after the first projection. , The scale after the first projection is 1. The t The periodic features at each time step, with lengths before and after projection being respectively... , ; Next, the periodic features after projection are analyzed. Perform a GELU activation function transformation to obtain , This indicates that the scale is after one GELU activation function transformation. The t Periodic characteristics of each time step; right Perform a linear projection to obtain the data after the second projection. , The scale after the second projection is 1. The t Seasonal data at each time step, with lengths before and after projection as follows: , .

[0030] The upsampling process steps are as follows: For scale Long-term trend characteristics Perform a linear projection to obtain the long-term trend characteristics after the first projection. , The scale after the first projection is 1. The t The long-term trend characteristics at each time step, with the lengths before and after projection being respectively... , ; For the projected data Perform a GELU activation function transformation to obtain , This indicates that the scale is after one GELU activation function transformation. The t Long-term trend characteristics at each time step; right Perform a linear projection to obtain the long-term trend characteristics after the second projection. , The scale after the second projection is 1. The t The long-term trend characteristics at each time step, with the lengths before and after projection being respectively... , .

[0031] Step S2.2.2: Extracting CNN branches through a convolutional neural network Local features are used to obtain a set of multi-scale power grid historical load CNN feature sequences. .

[0032] Step S2.3, Feature enhancement module for Projection and interpolation upsampling processes are performed to obtain the historical load enhancement characteristic sequence of the power grid. ,in, This represents the local enhancement characteristics of the historical load of the power grid at the t-th time step under the n-th scale; simultaneously, for Interpolation and upsampling processing are performed to obtain the historical load fusion enhancement feature sequence of the power grid. ,in, This represents the historical load fusion enhancement characteristics of the power grid at the t-th time step under the n-th scale; thus, for and Weighted fusion is performed to obtain the weighted fusion feature set of historical load of the power grid. ;in, This represents the weighted and integrated characteristic sequence of historical loads of the power grid, and ,in This represents the weighted fusion characteristics of historical load of the power grid; among them, the interpolation upsampling method in step S2.3 can be selected according to the characteristics of the dataset, such as nearest neighbor interpolation, Lagrange interpolation, quadratic interpolation, spline interpolation, linear interpolation, and Kriging interpolation.

[0033] In step S3, the weighted fusion is calculated according to equation (6): (6) In equation (6), It is a weight that is greater than 0 and less than 1.

[0034] Step S2.4: The prediction module uses a fully connected network to... Process the data and output the predicted value. ,in, Indicates the first t After the first time step The predicted grid load value for a given time step.

[0035] Step S3, based on and Constructing an L1-L2 hybrid loss function This is used to train the power grid load forecasting network to update the network parameters until the L1-L2 hybrid loss function is applied. The training continues until convergence, thus obtaining a well-trained power grid load prediction model to achieve power grid load prediction. The L1-L2 mixed loss function in step S5 is defined as shown in equation (7): (7) In equation (7), Indicates the first Forecast values ​​of power grid load at each time step For the first The actual value of the power grid load at each time step. , l Let tanh be the time step to be predicted, and let tanh be the hyperbolic tangent function. Indicates to The result of the operation is taken as the absolute value.

[0036] S6. Use the trained model to predict the load on the test grid.

[0037] As an embodiment, the present application proposes a multi-scale analysis method for time-series forecasting (MSAM).

[0038] In the experimental part, the MSAM method proposed by the present application is compared with four relatively influential prediction algorithms in recent years.

[0039] iTransformer is included in ICLR 2024, and its method enhances the prediction ability of the model by reversing the Token.

[0040] PatchTST is included in AAAI 2023, and the method processes the data by dividing it into patches.

[0041] FEDformer is included in ICML 2022, and its method processes the data in the frequency domain.

[0042] SCINet is included in NeurIPS 2022, and its method uses a recursive down-sampling-convolution-interaction structure to extract multiple time features from down-sampled subsequences using multiple convolution filters.

[0043] In order to verify the effectiveness of the multi-scale analysis method for time-series forecasting proposed by the present application in the field of power grid load forecasting, the prediction accuracy of MSAM, iTransformer, PatchTST, FEDformer and SCINet is compared.

[0044] The present application tests these methods on the Electricity dataset, and the average accuracy of each algorithm running ten times on the dataset is taken as the final result.

[0045] Table 1 details the main characteristics of the above-mentioned Electricity dataset.

[0046] Table 1: Datasets used in the experiment

[0047] Table 2 shows the prediction results of the above-mentioned methods on the Electricity dataset, and each row represents the prediction method that performs best on the corresponding dataset.

[0048] Table 2: Prediction comparison results of each algorithm in the experiment (The results are the average values of MSE and MAE for all prediction steps)

[0049] From the experimental results of Table 2, it can be seen that: The power grid load prediction method based on multi-scale analysis proposed in the present application has the best prediction effect on the Electricity data set. This reflects that the power grid load prediction method based on multi-scale analysis can improve the model prediction ability by denoising the data, then performing multi-branch feature extraction and multi-scale feature fusion.

Claims

1. A power grid load forecasting method based on multi-scale analysis, characterized in that, Includes the following steps: Step S1, using N Each sensor collects data at fixed time intervals. N Historical electricity consumption data of individual households, thereby constructing a historical load dataset for the power grid. ,in, Indicates the first t Each time step N Historical power grid load data for individual households T The total number of time steps; Step S2: Construct a power grid load forecasting network, including: a data preprocessing module, a dual-branch feature extraction module, a feature enhancement module, and a forecasting module, and perform... Processing is performed to obtain the predicted value. ;in, Indicates the time step of the forecast; Step S2.1: The data preprocessing module includes a wavelet denoising unit and a downsampling unit, and performs... The data is processed to obtain a denoised multi-scale power grid load time series dataset. ,in, Let represent the denoised historical load dataset of the power grid at the nth scale, and ,in, Represents the nth scale. Historical power grid load data at each time step k The number of scales; Step S2.2: The dual-branch feature extraction module includes a Transformer branch and a CNN branch, and respectively processes the denoised front... Time-series dataset of power grid load at various scales and after Time-series dataset of power grid load at various scales The data is processed to obtain a set of multi-scale power grid historical load Transformer feature sequences. and multi-scale power grid historical load CNN feature sequence set ;in, Let represent the historical load Transformer feature sequence of the power grid at the nth scale, and ,in, Represents the nth scale. Historical load fusion characteristics of the power grid at each time step Let represent the CNN feature sequence of the historical load of the power grid at the nth scale, and ,in, This represents the local characteristics of the historical load of the power grid at the t-th time step under the n-th scale. This indicates rounding down. The feature embedding dimension for the Transformer branch; Step S2.3, the feature enhancement module... Projection and interpolation upsampling processes are performed to obtain the historical load enhancement characteristic sequence of the power grid. ,in, This represents the local enhancement characteristics of the historical load of the power grid at the t-th time step under the n-th scale; simultaneously, for Interpolation and upsampling processing are performed to obtain the historical load fusion enhancement feature sequence of the power grid. ,in, This represents the historical load fusion enhancement characteristics of the power grid at the t-th time step under the n-th scale; thus, for and Weighted fusion is performed to obtain the weighted fusion feature set of historical load of the power grid. ;in, This represents the weighted and integrated characteristic sequence of historical loads of the power grid, and ,in This indicates the weighted fusion characteristics of historical load in the power grid; Step S2.4: The prediction module uses a fully connected network to... Process the data and output the predicted value. ,in, Indicates the first t After the first time step Forecast values ​​of grid load over a given time step; Step S3, based on and Constructing an L1-L2 hybrid loss function This is used to train the power grid load forecasting network to update the network parameters until the L1-L2 hybrid loss function is applied. The process continues until convergence, thus obtaining a well-trained power grid load prediction model to achieve power grid load prediction.

2. The power grid load forecasting method based on multi-scale analysis as described in claim 1, characterized in that, Step 2.1 includes: Step S2.1.1, the wavelet denoising unit... The data is processed to obtain a denoised historical load dataset of the power grid. ;in, Indicates the denoised first... t Each time step N Historical load data for individual household power grids; Step S2.1.2, the downsampling unit... The data is processed to obtain a denoised multi-scale historical load dataset of the power grid. ,in, .

3. The power grid load forecasting method based on multi-scale analysis as described in claim 1, characterized in that, Step 2.2 includes: Step S2.2.1, the Transformer branch pair Perform sequence decomposition to obtain a set of periodic features. and long-term trend characteristics set and will and Intra-scale and inter-scale fusion is performed to obtain a multi-scale power grid historical load Transformer feature sequence set. ;in, Let represent the long-term trend characteristic sequence of the historical load of the power grid at the nth scale, and , Represents the nth scale. Long-term trend characteristics at each time step; This represents the periodic characteristic sequence of the historical load of the power grid at the nth scale. , Represents the nth scale. Periodic characteristics of each time step; Step S2.2.2: The CNN branch is extracted through a convolutional neural network. Local features are used to obtain a set of multi-scale power grid historical load CNN feature sequences. .

4. The power grid load forecasting method based on multi-scale analysis as described in claim 3, characterized in that: In step S2.2.1, the sequence decomposition is calculated according to equations (1) and (2) respectively. and ; (1) = − (2) In equations (1) and (2), Indicates the time step index at the nth scale. , This indicates the size of the time series aggregation window. It's stride. It is a local location index within the time-series aggregation window.

5. The power grid load forecasting method based on multi-scale analysis as described in claim 3, characterized in that: Step S2.2.1, intra-scale-inter-scale fusion, uses equations (3) and (4) to obtain the fused periodic features at the nth scale. and the long-term trend characteristics after integration Thus, by using equation (5), we can obtain : (3) (4) (5) In equations (3) and (4), Indicates a downsampling operation; Indicates an upsampling operation. This represents the periodicity feature after fusion at the (n-1)th scale. This represents the long-term trend characteristics after fusion at the (n-1)th scale. When n=1, let... = ,make = .

6. The power grid load forecasting method based on multi-scale analysis as described in claim 5, characterized in that, downsampling operation in equation (3) include: Scale is Periodic characteristics Perform a linear projection to obtain the scale after the first projection. Periodic characteristics ,in, The scale after the first projection is 1. The t The periodic features at each time step, with lengths of the periodic features before and after the first projection being respectively... , ; right Perform a GELU activation function transformation to obtain the transformed scale as follows: Periodic characteristics ,in, The scale after transformation is The t Periodic characteristics of each time step; right Perform a linear projection to obtain the scale after the second projection. Periodic characteristics ,in, The scale after the second projection is 1. The t The periodic features at each time step, and the lengths of the periodic features before and after the second projection are respectively... , .

7. The power grid load forecasting method based on multi-scale analysis as described in claim 6, characterized in that, The upsampling operation in equation (3) include: For scale Long-term trend characteristics Perform a linear projection to obtain the scale after the first projection. Long-term trend characteristics ,in, The scale after the first projection is 1. The t The long-term trend characteristics at each time step, with lengths of the long-term trend characteristics before and after the first projection being respectively... , ; right Perform a GELU activation function transformation to obtain the transformed scale as follows: Long-term trend characteristics ,in, The scale after transformation is The t Long-term trend characteristics at each time step; right Perform a linear projection to obtain the scale after the second projection. Long-term trend characteristics ,in, The scale after the second projection is 1. The t The long-term trend characteristics at each time step, and the lengths of the long-term trend characteristics before and after the second projection are respectively... , .

8. The power grid load forecasting method based on multi-scale analysis as described in claim 1, characterized in that: In step S2.3, equation (6) is used to obtain... : (6) In equation (6), It is a weight that is greater than 0 and less than 1.

9. A power grid load forecasting method based on multi-scale analysis as described in claim 1, characterized in that: In step S3, the L1-L2 hybrid loss function is constructed using equation (7). : (7) In equation (7), Indicates the first Forecast values ​​of power grid load at each time step For the first The actual value of the power grid load at each time step. , l The time step to be predicted is given by tanh, which is the hyperbolic tangent function.

10. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store programs that support the processor in executing any of the power grid load forecasting methods of claims 1-9, and the processor is configured to execute the programs stored in the memory.

Images