Load prediction method based on power load multi-scale feature extraction

Abstract

The invention relates to the field of power load prediction, and provides a load prediction method based on power load multi-scale feature extraction, and the method comprises the steps: obtaining a power load curve, carrying out the processing, obtaining a new data sequence, obtaining the maximum information coefficient of a load feature X and a load time sequence decomposition component Y, carrying out the correlation analysis according to the maximum information coefficient, and carrying out the prediction of the load feature X and the load time sequence decomposition component Y. Sorting according to the correlation to obtain a load feature subset; obtaining component final load feature subsets according to the new data sequence, and superposing the subsets to obtain a load feature set; selecting a plurality of data subsets with different lengths from the load feature set, inputting the data subsets into the model in sequence to obtain a plurality of single prediction models corresponding to the data subsets with different output lengths, and integrating the single prediction models to obtain a prediction integrated model; according to the method, the power load time sequence is decomposed by using empirical mode decomposition, and the power load characteristics are screened by using the load characteristics and the maximum information coefficient of the load component, so that the prediction precision is improved, and the calculation amount is reduced.

Description

Technical Field

[0001] The present invention relates to the field of electric load forecasting, and more specifically, to a load forecasting method based on multi-scale feature extraction of electric load. Background Art

[0002] Electric load is an important part of the power system. As the consumer of electric energy, it has an important impact on the analysis, design and control of the power system. The extraction of electric load characteristics is the key prerequisite for load forecasting. Since electric load is easily affected by various factors, and at the same time, with the large-scale access of energy storage, electric vehicles, and renewable distributed power sources, it brings difficulties to the extraction of electric load characteristics. In order to cope with the challenges brought by the complex environment, in-depth research has been carried out at home and abroad on the decomposition of the original load time series and the correlation analysis of load influencing factors.

[0003] Scholars at home and abroad have also proposed many methods to accurately extract load characteristics and achieve accurate classification of load characteristics. However, these methods all have different degrees of defects. For example, the method of using the time and season of load characteristic sample collection as feature vectors cannot reflect the essential characteristics of load dynamic characteristics and has great subjectivity;

[0004] Taking the parameters of the induction motor comprehensive model as feature vectors and the model response under standard voltage excitation as feature vectors are two commonly used feature extraction methods at present. However, both of these methods must select the model structure and perform parameter identification to obtain the load dynamic characteristic feature vectors. Model structure errors and parameter identification errors will be inevitable, and the calculation amount is large.

[0005] Accurate load forecasting can enable energy suppliers and grid operators to predict the fluctuations of energy use, thereby promoting more efficient grid management and the integration of renewable energy. Load forecasting is not only a technical need but also the cornerstone of the pursuit of energy sustainability, because it reduces the carbon footprint by optimizing the balance between energy demand and renewable energy supply. However, the complexity brought by renewable energy may be beyond the reach of traditional methods.

[0006] In order to forecast electric load, the forecasting method has changed from traditional statistical methods to complex machine learning techniques. The initial electric load forecasting mainly relied on traditional regression models and time series analysis. These methods lack flexibility and cannot capture complex patterns in the data. Secondly, models such as convolutional neural networks and multi-layer perceptrons in neural networks are widely used and can detect non-linear relationships and intricate dependencies in large datasets. Furthermore, it has turned to variants such as long short-term memory (LSTM) and gated recurrent units in recurrent neural networks. These models are good at processing sequential data, but they also have their own challenges and limitations. Summary of the Invention

[0007] To solve the above technical problems, the present invention provides a load forecasting method based on multi-scale feature extraction of electric power load. By decomposing the original electric power load time series using empirical mode decomposition and then screening the electric power load features using the maximum information coefficient between the load features and the load components, the forecasting accuracy is effectively improved and the computational amount is reduced.

[0008] A load forecasting method based on multi-scale feature extraction of electric power load includes the following steps:

[0009] Step 1: Obtain the electric power load curve and process the electric power load curve through empirical mode decomposition to obtain a new data sequence h 1 (c);

[0010] Step 2: Obtain the maximum information coefficient between the load feature X and each empirical mode decomposition component Y of the load time series;

[0011] Step 3: Conduct a correlation analysis on the load feature X based on the target imf component according to the maximum information coefficient, and sort the load feature X according to the correlation to obtain a load feature subset T;

[0012] Step 4: Obtain the final load feature subsets of different imf components according to the new data sequence Superimpose different final load feature subsets to obtain a load feature set T BEST ;

[0013] Step 5: Select data subsets with multiple different output lengths from the load feature set T BEST and input them into the Autoformer model in sequence to obtain multiple single forecasting models corresponding to the data subsets with different output lengths;

[0014] Step 6: Integrate multiple single forecasting models and output a forecasting integrated model for electric power load forecasting.

[0015] Preferably, the Step 1 includes:

[0016] Step 1.1: Obtain the upper envelope u 1 (c) and the lower envelope v 1 (c) of the original load time series f(c), and obtain the average value m 1 (c) of the upper envelope u 1 (c) and the lower envelope v 1 (c):

[0017]

[0018] Step 1.2: Obtain a new data sequence h by subtracting the average envelope from the original load data 1 (c):

[0019] h 1 (c) = f(c) - m 1 (c)

[0020] Preferably, after the said Step 1.2, it further includes:

[0021] Step 1.3: Determine whether h 1 (c) meets the intrinsic mode function. If not, regard h 1 (c) as the new f(c), and repeat the processing procedures of Steps 1.1 and 1.2 wherein, is the difference sequence after m iterations, and β 1 (c) is the first qualified imf component;

[0022] Step 1.4: Obtain the remaining component r 1 (c):

[0023] r 1 (c) = f(c) - β 1 (c)

[0024] Step 1.5: Regard r 1 (c) as the new load time series, and repeat Steps 1.1 to 1.4 to obtain each order of imf components;

[0025] Step 1.6: Obtain the original data sequence:

[0026]

[0027] wherein, β i (c) is the i-th imf component, and r n (c) is the residual component.

[0028] Preferably, the determination of whether h 1 (c) meets the intrinsic mode function specifically is:

[0029] By determining that the difference between the number of extreme points and the number of zero-crossing points of the difference h 1 (c) is 0 or 1;

[0030] The upper and lower envelope lines of the difference h 1 (c) are obtained by The calculated mean value is 0.

[0031] Preferably, the said Step 2 includes:

[0032] Step 2.1: Construct preliminary features based on time type data, weather characteristic data, and historical load data, and initialize the original feature set S;

[0033] Step 2.2: Calculate the information entropy O of the load feature components

[0034]

[0035]

[0036] where X j is the feature affecting each component, A is the value space of this feature variable, S is the original feature set, which is the set of all features X j , p(X j ) is the probability density of X j , p(Y k ) is the probability density of Y k , Y k are the imf components after the EMD decomposition of the load, B is the value space of this Y k component, and Q is the set of all imf components;

[0037] Step 2.3: Obtain the mutual information D(X j , Y k ) between each load feature variable and each imf component:

[0038] D(X j , Y k ) = O(X j ) - O(X j │Y k ) = O(Y k ) - O(Y k │X j )

[0039] where O(X j │Y k ) and O(Y k │X j ) are conditional entropies;

[0040] Step 2.4: Obtain the mutual information D(X j , Y k ) that satisfies the following relational expression:

[0041] 0 ≤ D(X j , Y k ) ≤ min{O(X j ), O(Y k )}

[0042] where D(X j , Y k ) is the variable Xj and Y k mutual information;

[0043] Step 2.5: Obtain the maximum correlation information coefficient for feature X j and each IMF component Y k :

[0044]

[0045] where D max (X j , Y k ) is the maximum correlation information coefficient of variables X j and Y k .

[0046] Preferably, the said step 3 includes: Step 3.1: Obtain the symmetric uncertainty SU j of variable X k and Y max :

[0047]

[0048] Step 3.2: Sort each SU max (X j , Y k ) obtained in step 3.1 from large to small. The larger the value of SU max (X j , Y k ) indicates that the load feature X j has a greater impact on the load component Y k , and the feature importance is higher. Store the sorted features in set T.

[0049] Preferably, after the said step 3.2, it further includes:

[0050] Step 3.3, perform redundancy analysis on the load features. D(X j , Y k ) satisfies the following conditions:

[0051]

[0052] where X i ∈T and X i ∈T, i≠j; delete redundant features to obtain the load feature subset T.

[0053] Preferably, when selecting multiple data subsets with different output lengths from the load feature set and inputting them into the Autoformer model in step 5, it is necessary to perform standardized preprocessing on the data subsets, using the Z-score standardization method to ensure that the data input into the model is within the same dimension range;

[0054] Meanwhile, during the process of inputting into the Autoformer model, according to the input requirements of the model, the time series data is reasonably windowed, and the sliding window step size is set so that the model can learn the temporal relationship of the data.

[0055] Preferably, when integrating multiple single prediction models in step 6, the weighted average method is adopted, and weights are assigned according to the prediction accuracy of each single prediction model on the validation set.

[0056] Specifically, the load feature set is first divided into a training set, a validation set, and a test set in chronological order, with a ratio of 7:2:1. During the validation phase, accuracy metrics of each single prediction model are recorded using the root mean square error or the mean absolute error.

[0057] Finally, when making predictions, the output results of each model are weighted and summed according to the corresponding weights to obtain the output of the prediction integrated model for the final power load prediction.

[0058] Preferably, the time type data includes months, weeks, and holidays, and the weather feature data includes temperature and humidity.

[0059] Preferably, each of the multiple data subsets contains a same data set to be measured.

[0060] Preferably, integrating multiple said single prediction models to output a prediction integrated model for power load prediction is specifically as follows:

[0061] Among multiple said single prediction models, retaining the prediction models for the same data set to be measured;

[0062] Calculating and outputting the average value of the prediction models for multiple said same data sets to be measured, and the average value of the prediction models is the prediction integrated model for power load prediction.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] 1. The present invention decomposes the original power load time series by introducing empirical mode decomposition (EMD) to extract the intrinsic mode function (IMF) components, decomposes the complex load data into multi-scale components, and retains the essential characteristics of the load signal. Further, the maximum correlation information coefficient is used to screen the features to ensure that only the key features that make significant contributions to the prediction are retained, thereby greatly improving the accuracy of load feature extraction and laying a high-quality data foundation for the subsequent prediction link.

[0065] 2. The present invention adopts a prediction model based on the Autoformer architecture. This model captures long-term dependencies and periodic patterns in time series through an autocorrelation mechanism, solving the problem of insufficient modeling ability for complex load dynamic changes in the prior art. In addition, through a multi-model integration strategy, that is, averaging the model results with different prediction lengths, the adaptability to load patterns at different time scales is further enhanced, the prediction accuracy is improved, and it can adapt to both short-term and long-term load changes simultaneously to meet the requirements of different application scenarios.

[0066] 3. The present invention decomposes the load sequence and only screens key features for prediction, reducing the burden of redundant data on model calculation. The autocorrelation mechanism of the Autoformer model significantly reduces the time complexity and avoids the exponentially increasing computational requirements in the traditional attention mechanism as the sequence length grows. At the same time, the computational complexity is significantly reduced on the premise of ensuring prediction accuracy, improving the practicality and deployment efficiency of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is a schematic diagram of the step flow of the method of the present invention.

[0068] Figure 2 It is a schematic diagram of the process of feature extraction and screening in step 2 of the present invention.

[0069] Figure 3 It is a schematic diagram of the prediction model architecture of the present invention.

[0070] Figure 4 It is a schematic diagram of the Autoformer model architecture with feature extraction added in the embodiment of the present invention.

[0071] Figure 5 It is a schematic diagram of the output of part of the data in the embodiment of the present invention under the Autoformer model and the comparison with the true value. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0072] The following further describes in detail the embodiments of the present invention in conjunction with the drawings. The following embodiments are used to illustrate the present invention, but cannot be used to limit the scope of the present invention.

[0073] The present invention provides a load prediction method based on multi-scale feature extraction of electric power load, including the following steps:

[0074] Step 1: Obtain the electric power load curve, and process the electric power load curve through empirical mode decomposition to obtain a new data sequence h 1 (c);

[0075] Step 2: Obtain the maximum information coefficient of the load feature X and each empirical mode decomposition component Y of the load time series.

[0076] Step 3: Perform a correlation analysis on the load characteristic X based on the target imf component according to the maximum information coefficient, and sort the load characteristic X according to the correlation to obtain a load characteristic subset T;

[0077] Step 4: Obtain the final load characteristic subsets of different imf components according to the new data sequence Superimpose different final load characteristic subsets to obtain a load characteristic set T BEST ;

[0078] Step 5: Select multiple data subsets with different output lengths from the load characteristic set T BEST and input them into the Autoformer model in sequence to obtain multiple single prediction models corresponding to the data subsets with different output lengths;

[0079] Step 6: Integrate multiple single prediction models and output a prediction integration model for power load forecasting.

[0080] Example 1:

[0081] As Figures 1 - 3 shown, this example provides a load forecasting method based on multi-scale feature extraction of power loads. This method decomposes the original power load time series using empirical mode decomposition, and then uses the maximum information coefficient between the load characteristics and the load components to perform molecular screening on the power load characteristics.

[0082] Step 1.1: Obtain the upper envelope u 1 (c) and the lower envelope v 1 (c) of the original load time series f(c), and obtain the average value m 1 (c) of the upper envelope u 1 (c) and the lower envelope v 1 (c):

[0083]

[0084] Step 1.2: Obtain a new data sequence h 1 (c) after subtracting the average envelope from the original load data:

[0085] h 1 (c) = f(c) - m 1 (c)

[0086] Step 1.3: Determine whether h 1 (c) satisfies the two conditions of the intrinsic mode function (IMF):

[0087] By judging that the difference between the number of extreme points and the number of zero-crossing points of h 1 (c) is 0 or 1;

[0088] Difference h 1 The upper and lower envelopes of (c) are determined by The calculated average value is 0.

[0089] If the IMF condition is not satisfied, then h 1 (c) is regarded as the new f(c), and the processing procedures of steps 1.1 and 1.2 are repeated, wherein, is the difference sequence after m iterations, and β 1 (c) is the first qualified imf component;

[0090] Step 1.4: Obtain the remaining component r 1 for separating the first imf component (c):

[0091] r 1 (c) = f(c) - β 1 (c)

[0092] Step 1.5: Regard r 1 (c) as the new load time series, and repeat steps 1.1 to 1.4, and the imf components of each order can be obtained in turn until the screening ends when the given termination condition is satisfied to obtain all imf components and the remaining components;

[0093] Step 1.6: The original data sequence can be represented by these imf components and a mean or trend term r n (c), and obtain the original data sequence:

[0094]

[0095] wherein, β i (c) is the i-th imf component, and r n (c) is the residual component.

[0096] Step 2.1, construct preliminary features according to time type data (month, week, holiday), weather characteristic data (temperature, humidity) and historical load data. The preliminary features are shown in Table 1, and the original feature set S is initialized;

[0097] Table 1 Preliminary features

[0098]

[0099]

[0100] Step 2.2, calculate the information entropy O of the load characteristic component,

[0101]

[0102] Among them, X j is a feature affecting each component, A is the value space of the feature variable, S is the original feature set, which is the set of all features X j . p(X j ) is the probability density of X j , p(Y k ) is the probability density of Y k . Y k is each imf component after load EMD decomposition, B is the value space of this Y k component, and Q is the set of all imf components;

[0103] Step 2.3: Obtain the mutual information D(X j , Y k ) between each load feature variable and each imf component:

[0104] D(X j , Y k ) = O(X j ) - O(X j │Y k ) = O(Y k ) - O(Y k │X j )

[0105] Among them, O(X j │Y k ) and O(Y k │X j ) are conditional entropies;

[0106] Step 2.4: Obtain that the mutual information D(X j , Y k ) satisfies the following relational expression:

[0107] 0 ≤ D(X j , Y k ) ≤ min{O(X j ), O(Y k )}

[0108] Among them, D(X j , Y k ) is the mutual information between variables X j and Y k ;

[0109] Step 2.5: Obtain the maximum correlation information coefficient between the feature X j and each imf component Y k :

[0110]

[0111] Among them, Dmax (X j , Y k ) is the maximum correlation information coefficient of variables X j and Y k .

[0112] Step 3.1: Obtain the symmetric uncertainty SU j of variable X k and Y max :

[0113]

[0114] Step 3.2: Sort each SU max (X j , Y k ) obtained in Step 3.1 from large to small. The larger the value of SU max (X j , Y k ), the greater the influence of the load feature X j on the load component Y k , and the higher the feature importance. Store the sorted features in set T

[0115] Step 3.3: Conduct redundancy analysis on the load features. D(X j , Y k ) satisfies the following conditions:

[0116]

[0117] Among them, X i ∈ T and X i ∈ T, i ≠ j; Delete redundant features to obtain the load feature subset T

[0118] Example 2:

[0119] As Figures 1 - 3 shown, this embodiment provides an integrated prediction model for power load prediction by integrating multiple single prediction models, specifically:

[0120] Among multiple single prediction models, retain the prediction models for the same dataset to be measured; When selecting multiple data subsets with different output lengths from the load feature set and inputting them into the Autoformer model in Step 5, it is necessary to perform standardized preprocessing on the data subsets. The Z-score standardization method is used to ensure that the data input into the model is within the same dimension range;

[0121] At the same time, during the process of inputting into the Autoformer model, according to the input requirements of the model, perform reasonable window partitioning on the time series data and set the sliding window step size so that the model can learn the time series relationship of the data

[0122] Calculate and output the average value of the prediction models for multiple identical datasets to be measured. The average value of the prediction models is the prediction ensemble model for power load prediction. When integrating multiple single prediction models, the weighted average method is used, and weights are assigned according to the prediction accuracy of each single prediction model on the validation set.

[0123] Specifically, first divide the load feature set into a training set, a validation set, and a test set in chronological order, with a ratio of 7:2:1. During the validation phase, use the root mean square error or mean absolute error to record the accuracy metrics of each single prediction model.

[0124] Finally, when making predictions, sum the output results of each model weighted by the corresponding weights to obtain the output of the final prediction ensemble model for power load prediction.

[0125] Specifically, one of the main advantages of Autoformer is its ability to adjust the prediction length. Utilizing the function of Autoformer, a set of models with different prediction lengths is adopted to improve the overall prediction accuracy. This set consists of multiple Autoformer models, each trained with a different output length, including 96, 144, and 192 points, to capture various temporal patterns in the data, and only the first 96 points are used to output the final ensemble model. This method can provide a more comprehensive understanding of the power load behavior within different time ranges.

[0126] Table 2 Output results for each prediction length

[0127] Forecast Length MAE for 96 Points MSE For Points 96 points 0.3099 0.1983 144 0.3112 0.2088 192 0.3375 0.2284 Ensemble 0.3058 0.1964

[0128] By combining the prediction results of models with different output lengths of 96, 144, and 192 points, different temporal insights are effectively aggregated, thereby obtaining more robust and accurate prediction results, which can provide a more detailed understanding and prediction of power load patterns, meet the needs of different prediction horizons and scenarios, and fully utilize the advantages of Autoformer in processing long sequences and prediction length adaptability, making it a powerful tool in the power load prediction framework.

[0129] By evaluating the performance of the prediction model at 96 points, 144 points, and 192 points, especially analyzing the MAE and MSE of the first 96 points in each case, and combining the prediction results of different models, the overall prediction accuracy was optimized. A notable trend was highlighted by the results: as the prediction length increased, the accuracy of the initial 96 points decreased slightly, and the MAE and MSE values of the 144-point and 192-point predictions increased compared to the 96-point prediction. This result indicates that although the model can handle longer sequences, it is optimized for shorter and more immediate predictions.

[0130] However, the most significant improvement was seen in the ensemble method, which combines predictions of different lengths. The prediction ensemble model achieved the best levels of MAE (0.3058) and MSE (0.1964) for the first 96 points, outperforming the performance of any single-length model. This improvement highlights the efficacy of the ensemble strategy, which leverages the strengths of each individual prediction length and mitigates its weaknesses.

[0131] The embodiments of the present invention are given for purposes of illustration and description. Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. A load forecasting method based on multi-scale feature extraction of power load, characterized in that: The steps include: Step 1: Obtain the power load curve, and process the power load curve through empirical mode decomposition to obtain a new data sequence h1(c); in the IMF component; Step 2: Obtain the maximum information coefficient of each empirical mode decomposition component Y of the load characteristic X and the load time series; Step 3: performing a correlation analysis on the load feature X based on the target imf component according to the maximum information coefficient, and sorting the load feature X according to the correlation to obtain a load feature subset T; Step 4: Obtain the final load characteristic subsets of different IMF components according to the new data sequence Superimpose different final load characteristic subsets to obtain the load characteristic set T BEST ; Step 5: From the load characteristic set T BEST Select multiple data subsets with different output lengths and input them into the Autoformer model in sequence to obtain multiple single prediction models corresponding to the data subsets with different output lengths; Step 6: Integrate multiple single prediction models and output a prediction integration model for power load prediction.

2. A load forecasting method based on multi-scale feature extraction of power load according to claim 1, characterized in that: The step 1 comprises: Step 1.1: Obtain the upper envelope u1(c) and the lower envelope v1(c) of the original load time series f(c), and obtain the average value m1(c) of the upper envelope u1(c) and the lower envelope v1(c): Step 1.2: Obtain the new data sequence h1(c) after subtracting the average envelope from the original load data: h1(c)=f(c)-m1(c) step 1 . 3 :Determine whether h1(c) satisfies the eigenmode function. If not, treat h1(c) as the new f(c) and repeat the process of steps 1.1 and 1.

2. in, is the difference sequence iterated m times, β1(c) is the first qualified imf component; Step 1.4: Get the residual component r1(c) of separating the first imf component: r1(c)=f(c)-β1(c) Step 1.5: Consider r1(c) as a new load time series, repeat steps 1.1 to 1.4 to obtain the IMF components of each order; Step 1.6: Get the original data sequence: Among them, β i (c) is the i-th imf component, r n (c) is the residual component.

3. A load forecasting method based on multi-scale feature extraction of power load according to claim 2, characterized in that: The determination of whether h1(c) satisfies the eigenmode function is specifically as follows: By judging whether the difference between the number of extreme value points and the number of zero-crossing points of the difference h1(c) is 0 or 1; The upper and lower envelopes of the difference h1(c) are given by The calculated mean is 0.

4. A load forecasting method based on multi-scale feature extraction of power load according to claim 3, characterized in that: The step 2 comprises: Step 2.1, construct preliminary features based on time type data, weather characteristic data and historical load data, and initialize the original feature set S; Step 2.2, calculate the information entropy O of the load characteristic component, Among them, X j is the feature that affects each component, A is the value space of the feature variable, S is the original feature set, which is all the features X j The set of p(X j ) is X j The probability density, p(Y k ) is Y k The probability density of k is the imf component after load EMD decomposition, B is the Y k The value space of the component, Q is the set of all imf components; Step 2.3: Obtain the mutual information D(X j , Y k ): D(X j ,AND k )=O(X j )-O(X j |And k )=OR(AND) k )-OR(AND k |X j ) Among them, O(X j |Y k ) and O(Y k |X j ) is the conditional entropy; Step 2.4: Obtain mutual information D(X j , Y k ) satisfies the following relationship: 0≤D(X j ,AND k )≤min{O(X j ),OR(AND) k )} Among them, D(X j , Y k ) is the variable X j and Y k The mutual information of Step 2.5: Get the feature X j and each imf component Y k The maximum correlation coefficient between: Among them, D max (X j , Y k ) is the variable X j and Y k The maximum correlation coefficient.

5. A load forecasting method based on multi-scale feature extraction of power load according to claim 4, characterized in that: The step 3 includes: Step 3.1: Obtain variable X j With Y k Symmetric uncertainty SU max : Step 3.2: Each SU obtained in step 3.1 max (X j ,Y k ) are sorted from large to small, SU max (X j , Y k ) value is larger, indicating that the load characteristic X j For load component Y k The greater the influence, the higher the feature importance, and the sorted features are stored in set T. Step 3.3, perform redundancy analysis on load characteristics, D(X j ,Y k ) meets the following conditions: Among them, X i ∈T and X i ∈T,i≠j; delete redundant features and obtain the load feature subset T.

6. A load forecasting method based on multi-scale feature extraction of power load according to claim 5, characterized in that: The time type data includes month, week, and holiday, and the weather characteristic data includes temperature and humidity.

7. A load forecasting method based on multi-scale feature extraction of power load according to claim 1, characterized in that: When multiple data subsets with different output lengths are selected from the load feature set and sequentially input into the Autoformer model in step 5, the data subsets need to be standardized and preprocessed by using the Z-score standardization method to ensure that the data input into the model are in the same dimensional range; At the same time, when inputting the Autoformer model, the time series data is reasonably divided into windows according to the input requirements of the model, and the sliding window step size is set so that the model can learn the temporal relationship of the data.

8. The load forecasting method based on multi-scale feature extraction of power load according to claim 1 is characterized in that: When integrating multiple single prediction models in step 6, a weighted average method is used to assign weights according to the prediction accuracy of each single prediction model on the validation set; Specifically, the load feature set is first divided into a training set, a validation set, and a test set in chronological order, with a ratio of 7:2:

1. In the validation stage, the accuracy indicators of each single prediction model are recorded using root mean square error or mean absolute error; Finally, when making predictions, the output results of each model are weighted and summed according to the corresponding weights to obtain the final prediction integrated model output for power load prediction.

9. The load forecasting method based on multi-scale feature extraction of power load according to claim 1 is characterized in that: The plurality of data subsets all contain a same segment of the data set to be tested.

10. A load forecasting method based on multi-scale feature extraction of power load according to claim 9, characterized in that: Integrate the multiple single prediction models to output a prediction integration model for power load prediction, specifically: Among the multiple single prediction models, retain the prediction model of the same data set to be tested; The average value of the prediction models of the plurality of identical data sets to be tested is calculated and output, and the average value of the prediction models is a prediction integration model for power load prediction.

Images