Power load forecasting method, device, and storage medium based on signal decomposition and multidimensional information

Abstract

This invention relates to a method, apparatus, and storage medium for power load forecasting based on signal decomposition and multidimensional information. The method acquires and preprocesses a raw dataset, which includes historical time-series power load data and weather data; it obtains correlations and selects dimensions with correlations greater than a preset threshold to obtain a filtered dataset; it extracts features using a hybrid decomposition strategy of variational mode decomposition-whale optimization algorithm, and applies Kalman filtering to the extracted features to obtain a smoothed multidimensional dataset; based on the smoothed multidimensional dataset, it uses a pre-constructed model based on a time block-double CONET structure to perform prediction, obtaining the final power load forecast result. Compared with existing technologies, this invention has advantages such as effectively improving the accuracy of power load forecasting based on complex data and adaptability to complex load change patterns.

Description

Technical Field

[0001] This invention relates to the field of power load forecasting technology, and in particular to a power load forecasting method, apparatus and storage medium based on signal decomposition and multidimensional information. Background Technology

[0002] Electricity load forecasting methods can be categorized into historical data-based forecasting methods and real-time data-based forecasting systems. If a power load forecasting system can accurately match historical load patterns with real-time load data, it can provide timely forecast results. High forecast accuracy is an advantage, but a disadvantage is that the system may over-rely on historical data, making the prediction of future load changes challenging. Real-time data-based forecasting systems analyze the differences between current load conditions and historical load patterns; when there is a significant deviation between real-time load and expectations, the forecasting system will issue a warning. Therefore, better predicting load changes in power systems is currently a hot topic in power system research.

[0003] From a model learning perspective, traditional statistical learning, machine learning, and deep learning are all commonly used methods in power load forecasting systems. Statistical learning algorithms typically include autoregressive models and moving average models. Machine learning methods include support vector machines, decision trees, and random forests. Deep learning methods utilize neural networks and convolutional neural networks for load forecasting. However, single forecasting models have many limitations when dealing with complex data, including insufficient generalization ability, susceptibility to overfitting, high sensitivity to specific data types, weak robustness, and insufficient adaptability. This leads to difficulties in capturing complex load change patterns when facing nonlinear and high-dimensional features due to model limitations. The increasing volume of data and the diversity of load patterns restrict the development of traditional methods. Furthermore, previous methods have failed to effectively extract the latent features contained in nonlinear sequences, resulting in shortcomings in forecast accuracy. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art by providing a power load forecasting method, device, and storage medium based on signal decomposition and multidimensional information, which can realize medium- and long-term power load forecasting, effectively improve the accuracy of power load forecasting based on complex data, and cope with complex load change patterns.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] According to a first aspect of the present invention, a power load forecasting method based on signal decomposition and multidimensional information is provided, comprising the following steps: S1, acquiring and preprocessing an original dataset, wherein the original dataset includes historical time series data of power load and weather data; S2, based on the preprocessed dataset, acquiring correlations and selecting dimensions with correlations greater than a preset threshold to obtain a filtered dataset; S3, based on the filtered dataset, performing feature extraction using a hybrid decomposition strategy of variational mode decomposition-whale optimization algorithm, and performing Kalman filtering on the extracted features to obtain a smoothed multidimensional dataset; S4, based on the smoothed multidimensional dataset, performing prediction using a pre-constructed model based on a time block-double CONET structure to obtain the final power load forecasting result.

[0007] As a preferred technical solution, the preprocessing process specifically includes missing value processing and feature normalization.

[0008] As a preferred technical solution, S2 specifically includes: normalizing the preprocessed power load data and weather data, wherein the normalized power load data has corresponding weather dimension information; performing correlation analysis on the normalized power load data and the corresponding weather dimension information using the random forest method, and calculating the correlation coefficient between each weather dimension information and the corresponding power load data; and selecting the weather dimension that is positively correlated with the power load data based on the correlation coefficient.

[0009] As a preferred technical solution, the variational mode decomposition-whale optimization algorithm hybrid decomposition strategy specifically includes: performing variational mode decomposition on the screened dataset using signal decomposition methods to obtain a finite number of intrinsic mode function components and residuals, and extracting corresponding features; and using the whale optimization algorithm to optimize the key parameters in the variational mode decomposition process, wherein the key parameters include the center frequency and the number of intrinsic mode function components.

[0010] As a preferred technical solution, Kalman filtering is performed on each intrinsic mode function component, and the processed intrinsic mode function components are reconstructed to obtain a smoothed multidimensional dataset.

[0011] As a preferred technical solution, the variational mode decomposition process is expressed as follows:

[0012]

[0013] In the formula, u k Let ω represent the k-th eigenmode function component. k Let K represent the center frequency of the k-th intrinsic mode function component, where K is the number of intrinsic mode function components. Here, δ(·) is the convolution operator, and δ(·) represents the Dirac distribution. For partial derivative functions, v k Let f represent the eigenvalue of the k-th intrinsic mode function component, t represent time, and f represent the function value.

[0014] As a preferred technical solution, the prediction model based on the time block-double CONET structure includes a feature embedding layer. Specifically, S4 includes: taking the smoothed multidimensional dataset as the input sequence, and obtaining the corresponding deep features after passing through the feature embedding layer; extracting the corresponding temporal features using time blocks based on the deep features; performing normalization and denormalization processing using the double CONET structure based on the temporal features, and making medium- and long-term predictions of the power load, thereby outputting the corresponding power load prediction results.

[0015] As a preferred technical solution, normalization and denormalization processing is performed using a double CONET structure, the specific process of which is as follows:

[0016]

[0017] In the formula, These are all regression coefficients for the window. Given a single variable sequence, the horizontal coefficient at time step number t, where L represents the time span and N represents the number of layers.

[0018] According to a second aspect of the present invention, a power load forecasting device based on signal decomposition and multidimensional information is provided, comprising a memory, a processor, and a program stored in the memory, wherein the processor executes the program to implement the method described therein.

[0019] According to a third aspect of the present invention, a storage medium is provided having a program stored thereon, which, when executed, implements the method described thereon.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] 1. The power load prediction method provided by this invention is based on signal decomposition and multidimensional information. It combines time block and double CONET structure to construct a prediction model. After extracting features from the preprocessed and screened multidimensional dataset using a hybrid decomposition strategy of variational mode decomposition-whale optimization algorithm, it uses Kalman filter for further noise reduction. Then, it uses the constructed model based on time block-double CONET structure to make predictions and obtain the final power load prediction results. It fully integrates the advantages of variational mode decomposition, whale optimization algorithm and Kalman filter in signal processing, which can avoid the limitations of single model prediction of charge change and effectively improve the accuracy of power load prediction.

[0022] 2. This invention proposes an efficient normalization and denormalization method to optimize input and output, reduce feature differences, accelerate convergence speed, and improve prediction accuracy and efficiency. It can perform medium- and long-term forecasts of power load, improve the accuracy of power load forecasting based on complex data, and improve adaptability to complex load change patterns.

[0023] 3. The power load forecasting method provided by this invention utilizes time block processing to upgrade one-dimensional tensors into two-dimensional information, obtains deep features from the embedding layer, effectively represents long sequence dependencies, and avoids the influence of various periodic factors (such as weather, holidays, etc.) in time series data. It can effectively solve the problem of linear models extracting high-dimensional features and reducing the limitations of external variable features when modeling nonlinear sequence relationships. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the method flow provided in an embodiment of the present invention. Detailed Implementation

[0025] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0026] Example

[0027] like Figure 1 As shown, this embodiment provides a power load forecasting method based on signal decomposition and multidimensional information. The method includes the following steps:

[0028] Step S1: Obtain the raw dataset and preprocess it. The raw dataset includes historical time series data of electricity load and weather data.

[0029] The data preprocessing process includes missing value handling and feature normalization. Specifically, linear interpolation can be used to fill in missing values, which in turn include missing values ​​for power load and multi-dimensional weather information for the area collected on that day. At the same time, Z-score scaling is used to reduce the data to a specific range and then normalization is performed.

[0030] Step S2: Based on the preprocessed dataset, obtain the correlation and select the dimensions with a correlation greater than a preset threshold to obtain the filtered multidimensional dataset.

[0031] Specifically, the degree of correlation is determined by correlation analysis, that is, by correlation analysis, the degree of influence of different factors on power load is obtained and presented as correlation coefficient.

[0032] This embodiment employs the random forest method for correlation analysis to reduce the dimensionality of multidimensional data. After Z-score normalization of the collected power load and weather data, the correlation between the weather dimension information and the power load data is calculated. After obtaining the correlation coefficient, the dimensions that are positively correlated with the power load data are selected, that is, the dimensions with a weight greater than 0 are filtered out.

[0033] Step S3: Based on the filtered multidimensional dataset, feature extraction is performed using a hybrid decomposition strategy of Variational Mode Decomposition (VMD) and Whale Optimization (WOA), and the extracted features are then processed by Kalman filtering to obtain a smoothed multidimensional dataset. Specifically, this includes:

[0034] Step S31: Make a preliminary prediction on the historical power load data. When the data for N periods is predicted, use the data before the prediction time point N-1 as the point to update the training.

[0035] Step S32: Use signal decomposition methods to perform variational mode decomposition on the filtered time series data to obtain a limited number of intrinsic mode function components (IMFs) and residuals, so as to capture the key features and hidden information of the signal.

[0036] Step S33: Optimize key parameters in the VMD process using the whale optimization algorithm. Key parameters include center frequency and the number of IMFs to enhance the accuracy and stability of the decomposition.

[0037] Step S34: Perform Kalman filtering on each IMF component to further reduce noise and smooth the signal.

[0038] Step S35: Reconstruct the processed IMF components to obtain smoothed multidimensional time series data.

[0039] The center frequency is obtained by continuously updating the center frequency and gradually adjusting the frequency band of each mode. The variational problem is expressed as:

[0040]

[0041] In the formula, u k Let ω represent the k-th eigenmode function component. k Let K represent the center frequency of the k-th intrinsic mode function component, where K is the number of intrinsic mode function components. Here, δ(·) is the convolution operator, and δ(·) represents the Dirac distribution. For partial derivative functions, v k Let f represent the eigenvalue of the k-th intrinsic mode function component, t represent time, and f represent the function value.

[0042] The Whale Optimization Algorithm (WOA) cleverly uses the minimum envelope entropy value as the fitness function to optimize VMD parameters: First, the position vector [K, α] of the whale pod is initialized; then, the envelope entropy is used as the fitness function to calculate the fitness of each whale; finally, the position vector is iteratively updated using an iterative formula until the optimal VMD parameters are obtained. The specific execution flow of the Whale Optimization Algorithm (WOA) is as follows:

[0043] (1) Initialize the whale population: Randomly initialize a group of individual whales, each representing a set of VMD parameters (such as the number of modes K, penalty factor α, etc.);

[0044] (2) Calculate fitness: For each individual whale, use its parameters to perform VMD decomposition, and calculate fitness based on the quality of the decomposition results (such as reconstruction error, modality smoothness, etc.);

[0045] (3) Update position: Update the position of individual whales according to the rules of the whale optimization algorithm, including simulating whales’ prey encirclement, bubble attack and random search behavior;

[0046] (4) Select the optimal solution: In each iteration, record the currently found optimal parameters and the corresponding fitness.

[0047] Termination condition judgment: Check whether the algorithm meets the termination condition (such as the number of iterations, fitness threshold, etc.). If it is not met, return to continue iterating; otherwise, output the optimal VMD parameters.

[0048] The Kalman filter is a method that uses the state equations of a linear system to estimate the optimal state of input and output observation data. In data processing, it is an effective method for noise reduction and data recovery. The Kalman filter consists of two stages: prediction and correction. The specific execution flow for Kalman filtering each IMF component is as follows:

[0049] (1) Initialize the state, set the initial state and covariance matrix of the Kalman filter;

[0050] (2) In the prediction phase, the state and covariance matrix of the next moment are predicted based on the system dynamic model;

[0051] (3) In the update step, the predicted state and covariance matrix are updated using the new observation data to obtain the optimal estimate;

[0052] (4) Smooth the output of the Kalman filter to reduce random fluctuations in the estimation.

[0053] (5) Repeat the prediction and update steps until all observation data has been processed. Finally, output the smoothed signal as the final result.

[0054] Steps S1 to S3 preprocess the original input dataset and perform noise reduction based on signal decomposition, aiming to provide clean and normalized input data for models based on the TimeBlock-Double CONET structure.

[0055] Step S4: Based on the smoothed multidimensional time series data, normalization and denormalization are performed using a pre-built model based on the TimeBlock-Double CONET structure to make medium- and long-term forecasts of the power load and obtain the final power load forecast results.

[0056] This step specifically includes five sub-steps: feature embedding sub-step, time block processing sub-step, normalization and denormalization processing sub-step, dual CONET fusion sub-step, and prediction output sub-step.

[0057] (1) Feature embedding sub-step

[0058] The smoothed multidimensional time series data is used as the input sequence, which is then transformed through a feature embedding layer to obtain a deep feature representation, i.e., deep features:

[0059]

[0060] In the formula, T represents the sequence length, and d model This step represents the dimension values. It enhances the representation of the raw data and provides a foundation for capturing complex dependencies.

[0061] For the time span L:

[0062]

[0063] (2) Time block processing sub-step

[0064] The embedded data is input into a time block structure. Each time block processes the data through transformation, feature extraction, and dimensionality reduction to extract key temporal features and transform these features back into a one-dimensional representation. After this, convolution is used to extract temporal variations.

[0065]

[0066] In the formula, l represents the sequence of time blocks.

[0067] The input sequence is processed through an embedding layer to obtain deep features, which are then used for prediction in four steps:

[0068] 1) Transformation: First, transform the one-dimensional time-series features of the input. Periodicity extraction is performed, and the periodicity is transformed into a two-dimensional tensor to represent the two-dimensional temporal variation, where Represents each frequency component X 1D The intensity, the frequency k with the highest intensity corresponds to the most significant period length {f1,…,f k}k{p1,…,p k}

[0069] 2) Extraction: Two-dimensional tensor Because of its two-dimensional locality, information is extracted using convolution. Here, the classic Inception model is used, namely:

[0070] 3) Dimensionality reduction: For the extracted temporal features, convert them back to one-dimensional space for information aggregation.

[0071] 4) Adaptive fusion: The next step will yield a one-dimensional representation. It involves weighted summation of the response frequencies to obtain the final output sequence information required in the normalized and denormalized architecture.

[0072] (3) Normalization and denormalization sub-step

[0073] The time-block processed sequence data undergoes normalization and denormalization using the CONET model. This process evaluates and adjusts the data's distribution characteristics to adapt to the model's learning of long-term dependencies and periodic changes. The normalization and denormalization process is represented as follows:

[0074]

[0075] In the formula, These are all regression coefficients for the window. Given a single variable sequence, the horizontal line coefficient at time step number t, where N represents the number of layers.

[0076] (4) Double CONET fusion sub-step

[0077] The internal spatial offset and interval spatial offset of time series data are processed and fused independently using the BACKCONET and HORI-CONET structures. This is mainly used to reduce the internal spatial shift and interval spatial shift in time series, and to enhance the model's adaptability and accuracy in series prediction.

[0078] BACKCONET is specifically designed for understanding input. The output of the spatial distribution, and the spatial distribution HORICONET.

[0079] (5) Predicting the output sub-step

[0080] Based on the above processing steps, the model outputs the final power load forecast result, which takes into account the long-term dependence, periodic changes and distribution characteristics of the data, aiming to provide high-accuracy load forecast.

[0081] The prediction model based on the TimeBlock-Dual CONET structure provided in this embodiment is compared with other combined prediction models. Compared with other models combining models through VMD mode decomposition, the comparison of MAE and MSE also shows that the model has higher prediction accuracy, indicating that the improved scheme is preferred. Table 1 shows the comparison results of the statistical parameters of each model. The calculation formulas for MAE and MSE are as follows:

[0082]

[0083] In the formula, m is the number of samples; y i The actual value; These are predicted values.

[0084] Table 1 Comparison of statistical parameters for each model

[0085]

[0086] Furthermore, this embodiment also provides a power load forecasting device based on signal decomposition and multidimensional information, including a memory, a processor, and a program stored in the memory. When the processor executes the program, it implements the aforementioned method. The specific implementation process is the same as the aforementioned method steps, and will not be repeated here.

[0087] Furthermore, this embodiment also provides a storage medium storing a program, which, when executed, implements the aforementioned method. The specific implementation process is the same as the aforementioned method steps, and will not be repeated here.

[0088] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A power load forecasting method based on signal decomposition and multidimensional information, characterized in that, Includes the following steps: S1. Obtain the raw dataset and preprocess it. The raw dataset includes historical time series data of power load and weather data. S2, based on the preprocessed dataset, obtain the correlation and select the dimensions with a correlation greater than a preset threshold to obtain the filtered dataset; S3. Based on the filtered dataset, feature extraction is performed using a hybrid decomposition strategy of variational mode decomposition-whale optimization algorithm, and the extracted features are processed by Kalman filtering to obtain a smoothed multidimensional dataset. S4. Based on the smoothed multidimensional dataset, a pre-built model based on the time block-double CONET structure is used to make predictions, and the final power load prediction results are obtained. The prediction model based on the time block-dual CONET structure includes a feature embedding layer, and S4 specifically includes: The smoothed multidimensional dataset is used as an input sequence, and after passing through the feature embedding layer, the corresponding deep features are obtained. Based on the aforementioned deep features, corresponding temporal features are extracted using time blocks; Based on the aforementioned time-series characteristics, normalization and denormalization processing are performed using a dual CONET structure, and medium- to long-term power load forecasts are conducted, thereby outputting the corresponding power load forecast results.

2. The power load forecasting method based on signal decomposition and multidimensional information according to claim 1, characterized in that, The preprocessing process specifically includes missing value handling and feature normalization.

3. The power load forecasting method based on signal decomposition and multidimensional information according to claim 1, characterized in that, S2 specifically includes: The preprocessed power load data and weather data are normalized, and the normalized power load data has corresponding weather dimension information. The random forest method was used to perform correlation analysis on the normalized power load data and the corresponding weather dimension information, and the correlation coefficient between each weather dimension information and the corresponding power load data was calculated. Based on the correlation coefficient, the weather dimension that is positively correlated with the power load data is selected.

4. The power load forecasting method based on signal decomposition and multidimensional information according to claim 1, characterized in that, The hybrid decomposition strategy of variational mode decomposition-whale optimization algorithm specifically includes: The filtered dataset is subjected to variational mode decomposition using signal decomposition methods to obtain a finite number of intrinsic mode function components and residuals, and the corresponding features are extracted. The whale optimization algorithm is used to optimize key parameters in the variational mode decomposition process, including the center frequency and the number of intrinsic mode function components.

5. The power load forecasting method based on signal decomposition and multidimensional information according to claim 4, characterized in that, Each intrinsic mode function component is subjected to Kalman filtering, and the processed intrinsic mode function components are reconstructed to obtain a smoothed multidimensional dataset.

6. The power load forecasting method based on signal decomposition and multidimensional information according to claim 4, characterized in that, The process of variational mode decomposition is represented as follows: In the formula, u k Let ω represent the k-th eigenmode function component. k Let K represent the center frequency of the k-th intrinsic mode function component, where K is the number of intrinsic mode function components. Here, δ(·) is the convolution operator, and δ(·) represents the Dirac distribution. For partial derivative functions, v k Let f represent the eigenvalue of the k-th intrinsic mode function component, t represent time, and f represent the function value.

7. The power load forecasting method based on signal decomposition and multidimensional information according to claim 1, characterized in that, Normalization and denormalization are performed using a double CONET structure. The specific process is as follows: In the formula, These are all regression coefficients for the window. Given a single variable sequence, the horizontal coefficient at time step number t, where L represents the time span and N represents the number of layers.

8. A power load forecasting device based on signal decomposition and multidimensional information, comprising a memory, a processor, and a program stored in the memory, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1-7.

9. A storage medium having a program stored thereon, characterized in that, When the program is executed, it implements the method as described in any one of claims 1-7.

Images