A Short-Term Power Load Forecasting Method Combining CDC-VMD and Echo State Networks

Abstract

This invention claims protection for a short-term power load forecasting method combining CDC-VMD and echo state networks. First, the Local Directional Centrality (CDC) clustering algorithm is used to accurately cluster the raw load data. The Davidson-Bolding Index (DBI) is then used to compare with other clustering algorithms, revealing that the CDC algorithm has the best clustering effect. Next, Variational Mode Decomposition (VMD) is used to process the load data, decomposing the load sequence into multiple Intrinsic Mode Functions (IMFs) to reduce the complexity of the input data for the neural network prediction model. The sample entropy (SE) of each IMF component is calculated, and components with similar sample entropy values ​​are merged into a new sequence to reduce the number of models required for training. Finally, a Deep Echo State Network (DESN) is used to fit the relationship between historical data and predicted data for each sequence, and the prediction results of each model are superimposed to obtain the final predicted value. This invention offers higher accuracy compared to traditional prediction methods.

Description

Technical Field

[0001] This invention belongs to the technical field of short-term power load forecasting methods, and in particular, it is a short-term power load forecasting method that combines clustering algorithms, variational mode decomposition, and echo state networks. Background Technology

[0002] In today's society, electricity, as an indispensable element of modern life, profoundly impacts everyone's life in every aspect. Large-scale power systems are crucial to a nation's economic and industrial lifeline. With the surge in electricity demand from various industries, the power system must rationally and scientifically allocate power resources to meet the needs of relevant industries without wasting excess power. Unlike other industries (such as chemical and coal industries), the power industry has its own unique characteristics because electricity cannot be stored on a large scale for long periods; therefore, its production and consumption must occur simultaneously. Due to the inherent characteristics of electricity, the power system must promptly and quickly understand and grasp changes in load trends and adjust power generation in real time to achieve a dynamic balance. In recent years, with the large-scale establishment of China's power system and the increasingly sophisticated development of related markets in the power sector, power load forecasting is gradually becoming an important research topic in the field of power systems. Short-term power load forecasting is a new method developed under this environment, which can make power system management more intelligent and economical.

[0003] Currently, there are two main methods for short-term load forecasting: classical forecasting methods and modern forecasting methods. Classical forecasting methods refer to traditional forecasting methods that apply statistical theory, with time series analysis being the most common. Time series analysis reveals the pattern of load changes over time based on historical load variations. This pattern can be used as a basis for future load changes. Its advantage lies in its relatively good forecasting performance when the power grid is operating normally. However, if the load is affected by uncertainties such as power outages, the forecasting effect of this method cannot meet the target requirements. In recent years, due to the rapid development of deep learning and artificial intelligence...

[0004] Machine learning methods are increasingly being used in short-term load forecasting. A typical approach is to train artificial neural networks using large amounts of electricity load data to build forecasting models. Examples include recurrent neural networks (RNNs) and long short-term memory neural networks (LSTMs). Although neural network forecasting models can effectively fit the nonlinear relationship between historical data and external factors, they still have many shortcomings, such as getting trapped in local optima, relying on manual parameter adjustments, and being prone to overfitting.

[0005] Therefore, a combined prediction method integrating data signal processing and neural networks is needed. This method overcomes the shortcomings of traditional methods, which insufficiently process the original power load sequence and directly use it to build prediction models. Instead, it allows for the decomposition and analysis of the original power load sequence to extract effective features. Furthermore, considering the emergence of sequences with similar modal complexity during decomposition, using all of them for model training would require building multiple prediction models, significantly increasing training time and difficulty. Therefore, this method utilizes sample entropy to merge highly similar sequences into new sequences, reducing the number of models needed for training and accelerating the training process. The decomposed sequences can effectively leverage the strong nonlinear fitting ability of echo state networks, improving the accuracy of load prediction.

[0006] CN114529049A discloses a short-term power load forecasting method based on VMD-CNN-LSTM, comprising: acquiring power load data for a certain region, retaining the date and daily power consumption data to form an original dataset; preprocessing the dataset, including completing missing data in abnormal datasets, correcting biased data in abnormal datasets, decomposing it into multiple components using the VMD algorithm, normalizing the components, and dividing them into training, validation, and test sets according to a certain ratio; constructing a short-term power load forecasting model based on VMD-CNN-LSTM, which includes an input layer, a convolutional neural network layer, an activation layer, a long short-term memory network layer, a fully connected layer, and an output layer; using the obtained training set to train the VMD-CNN-LSTM short-term power load forecasting model to obtain a short-term power load forecasting model with a suitable fit; and using the obtained power load forecasting model to predict short-term power load. This invention also includes a power load forecasting system.

[0007] (1) In the data preprocessing stage, in addition to completing missing data and correcting biased data in the dataset, this invention also uses the CDC clustering algorithm to cluster the original dataset. This clustering algorithm divides each data into a specific group, so that the data in the same group have highly similar feature patterns. This makes it easier for the prediction model to learn the feature patterns of the data more quickly when making predictions, resulting in more accurate prediction results and smaller errors. (2) In the VMD modal decomposition stage, this invention adds the step of "calculating the sample entropy of each decomposed load sequence and merging modal components with similar sample entropy values ​​to form new components". This allows the model to quickly learn the regularity of new components during model training, greatly reducing the complexity of training and significantly improving the training speed and prediction accuracy. (3) This invention improves the structure of the ESN model by constructing a deep architecture to capture the multi-scale dynamics of time series. An encoder layer is constructed on the basis of the ESN model structure, and feature connections are added between each encoder layer and the last output layer. High-dimensional projection is fully utilized to obtain a rich multi-scale dynamic representation of the input, and finally the DESN model is obtained. Compared to the CNN-LSTM model, the DESN model can reduce the complexity of data training to simple linear regression, which greatly improves the shortcomings of the high training complexity and long training time of the CNN-LSTM model. Moreover, the DESN model still maintains efficient computation without relying on the gradient propagation algorithm. (3) In order to overcome the shortcomings of unclear and ambiguous grasp of data feature patterns in the data preprocessing stage, this invention uses the CDC clustering algorithm to classify data with similar feature patterns, so that the data in the same group have highly similar feature patterns. In order to overcome the shortcomings of high training complexity and long training time of traditional models such as CNN-LSTM, this invention uses the DESN model for training, which can effectively reduce the complexity and training time of model training. Summary of the Invention

[0008] This invention aims to solve the problems of the prior art. It proposes a short-term power load forecasting method combining CDC-VMD and echo state networks. The technical solution of this invention is as follows:

[0009] A short-term power load forecasting method combining CDC-VMD and echo state networks includes the following steps:

[0010] Step 1: Clean the raw power load data and use a clustering algorithm to cluster the cleaned load data. Analyze the clustering results using the Davidson-Bolding Index (DBI). Based on the DBI index, the CDC clustering algorithm is found to have the best clustering performance.

[0011] Step 2: The clustered load sequences are decomposed using variational mode VMD to obtain the individual decomposed load sequences;

[0012] Step 3: Calculate the sample entropy of each decomposed load sequence, and merge modal components with similar sample entropy values ​​to form new components.

[0013] Step 4: Normalize the merged components and map them to the range [0,1].

[0014] Step 5: Construct time series input-output label pairs using a sliding window; these input-output label pairs are used to train the prediction model.

[0015] Step 6: Construct a short-term power load forecasting model that combines CDC-VMD and echo state network, and use the Adam optimization algorithm to adjust the model weight parameters and find the optimal value of the network;

[0016] Step 7: Feed the normalized components into the Deep Echo State Network (DESN) for training;

[0017] Step 8: Reconstruct the load sequence of each trained component to obtain the predicted load data and output the prediction result.

[0018] Furthermore, step 1 involves clustering the original load data, specifically including the following steps:

[0019] (1) Calculate the angular variance, which is a measure of local orientation centrality. The angular variance is used to characterize the uniformity of the K nearest neighbor distribution of each point.

[0020]

[0021] α i Let i represent the angle value, k represent the number of angles, and k represent the current α. i For which angle value?

[0022] (2) Based on the local orientation centrality measure, a threshold is set to divide the boundary points and interior points, where the DCM value of the boundary points is greater than the threshold and the DCM value of the interior points is less than the threshold.

[0023] (3) Generate connections between constrained interior points, i.e., divide each cluster. The specific method is: calculate the connection between the boundary point and each interior point p. i The closest distance r i Calculate the distance d between interior points. ij If r i +r j ≥d ij Then the interior point p i With p j Those in the same cluster, otherwise in different clusters;

[0024]

[0025] Internal point cluster partitioning criteria:

[0026] d(p i ,p j )≤r i +r j (3)

[0027] The Davidson-Bourdin Index (DBI) was used to analyze the clustering results. The calculation formula is as follows:

[0028]

[0029] In the formula: w is the average distance from the data within a cluster to the cluster centroid, representing the dispersion of time series within cluster i. i The DBI represents the cluster centers of each cluster; for each cluster, the DBI first finds the average internal class distance. and The sum is divided by the maximum distance between the two cluster centers; then these maximum values ​​are added together to get the average. The CDC clustering algorithm has the lowest DBI index and the best clustering effect.

[0030] Furthermore, step 2 involves using variational mode decomposition (VMD) to decompose the clustered load sequences into individual decomposed load sequences, specifically including the following steps:

[0031] (1) Initialization Let the k-th modal component and the center frequency be represented respectively. For the Lagrange operator, the number 1 in the upper left corner indicates the first iteration.

[0032] (2) For each subsequence, continuously update according to (5) and (6) to obtain and

[0033]

[0034]

[0035] In the formula: Wiener filtering for the current remaining components, The frequency center of the corresponding modal component is ω, where ω is the frequency value. What do they represent respectively? What does α represent? What do they represent respectively? What does α represent? Let f(t) and f(t) represent the original sequence, respectively. (The Fourier transform of α, where α is the quadratic penalty factor.)

[0036] (3) For all ω≥0, update

[0037]

[0038] τ represents the noise tolerance, and K represents the total number of modes.

[0039] (4) Determine whether the iteration termination condition is satisfied:

[0040]

[0041] If the termination condition is not satisfied, repeat steps (2) and (3). If the condition is satisfied, the iteration terminates, and K decomposed subsequences are obtained; ε represents the similarity coefficient;

[0042] Further, in step 3: Calculate the sample entropy of each decomposed sequence, and merge the modal components with similar sample entropy values to form new components. Specifically: For a time series {x(n)} composed of N points = x(1), x(2), … x(N), the calculation method of sample entropy is as follows:

[0043] (1) Form a vector sequence of dimension m by serial numbers, X m (1), …, X m (N - m + 1), where X m (i) = {x(i), x(i + 1), …, x(i + m - 1)}, (1 ≤ i ≤ N – m + 1); [[ID=2,8]]

[0044] (,2) Define the distance d[X m (i), X m (j)] between X m (i) and X m (j) as the absolute value of the maximum difference between their corresponding elements;

[0045] <^{,}

[0046] (3) Given a threshold r, record the number of j for which d[X m (i), X m (j)] < r, denoted as B i , and for 1 ≤ i ≤ N - m, record the ratio of B i to N - m + 1 as:

[0047]

[0048] (4) Calculate the average value for all to obtain:

[0049] <^{,}

[0050] (5) Increase the dimension to m + 1, and calculate X m+1 (i) and Xm+1 (j) The number of items whose distance is less than or equal to r, denoted as A. i , Defined as:

[0051]

[0052] (6) Define A m (r) is:

[0053]

[0054] B m (r) and A m (r) represents the matching probability of sequence pair m and m+1, respectively. The sample entropy is defined as follows:

[0055]

[0056] When N takes finite values, the estimated value of the sample entropy is:

[0057]

[0058] Furthermore, step 4: normalize the merged new components, mapping them to the range [0,1]. Specifically, normalization involves scaling the original data proportionally within a specific interval, converting it into dimensionless pure numerical values, and then using the min-max normalization method to map the load to the range [0,1]. The calculation formula is as follows:

[0059]

[0060] Where, x min and x max ...

[0061] Furthermore, step 6, constructing a deep echo state network (DESN) short-term power load prediction model, specifically involves using a DESN to model and predict the time series of each modality of power load. The DESN is a hierarchical ESN model, whose main idea is to capture the multi-scale dynamics of the time series by constructing a deep architecture. Instead of directly stacking multiple ESNs, DESN constructs encoder layers and adds feature connections between each encoder layer and the last output layer, fully utilizing high-dimensional projection to obtain a rich multi-scale dynamic representation of the input. Moreover, DESN training is performed layer by layer on the entire model, with the only trainable layer being the last output layer. It maintains high computational efficiency without relying on gradient propagation algorithms. During network training, the Adam optimizer is used to adjust the network weight parameters and find the optimal values ​​for the network.

[0062] The advantages and beneficial effects of this invention are as follows:

[0063] This invention combines Local Oriented Centrality (LOC) clustering, Variational Mode Decomposition (VMD), and Deep Echo State Networks (DESN). LOC clustering is used to accurately cluster the original load data. The Davies-Bouldin Index (DBI) is used to compare with other clustering algorithms, revealing that the CDC algorithm has the best clustering effect. Then, Variational Mode Decomposition (VMD) is employed to decompose the original data into subsequences with limited bandwidth, extracting effective features from the load sequences. Simultaneously, sample entropy is used to measure the complexity of each modal component; modal components with similar sample entropy values ​​are merged into new components, effectively improving the model's training speed and predictive ability. Finally, Deep Echo State Networks (DESN) are used as the prediction model. Traditional Echo State Networks (ESNs) have a very simple training process and possess the characteristics of nonlinear echo states and short-term memory, which are very useful for modeling dynamic systems. However, a single ESN network cannot handle input signals that require complex hierarchical processing. To address the high volatility of short-term power load sequences, Deep Echo State Networks (DESNs) are used to model and predict time series of various power load modes. DESN is a hierarchical ESN model whose main idea is to capture the multi-scale dynamics of time series by constructing a deep architecture. Instead of directly stacking multiple ESNs, DESN constructs encoder layers and adds connections (called feature connections) between each encoder layer and the last output layer. This fully utilizes high-dimensional projection to obtain rich multi-scale dynamic representations of the input. Furthermore, DESN training is performed layer by layer on the entire model, with the only trainable layer being the last output layer. This maintains high computational efficiency without relying on gradient propagation algorithms. Therefore, combining local orientation centrality clustering algorithm, variational mode decomposition and temporal convolutional network for power load forecasting can not only perform efficient and accurate clustering of raw load data, but also decompose the original load sequence to extract effective features of the load sequence, and utilize the structural advantages of deep echo state network to improve the training speed and prediction accuracy of the model.

[0064] The innovative steps of this invention are as follows: (1) Step 1: Use a clustering algorithm to perform CDC clustering on the cleaned load data, and use the Davidson-Bolding index (DBI) to analyze the clustering results. According to the DBI index, the clustering performance of the CDC clustering algorithm is optimal; (2) Step 3: Calculate the sample entropy of each decomposed load sequence, and merge modal components with similar sample entropy values ​​to form new components; (3) Step 7: Put the normalized components into the Deep Echo State Network (DESN) for training.

[0065] The "prediction method combining CDC-VMD and echo state network" proposed in this invention can quickly learn the regular characteristics of load data when predicting power load, and greatly reduce the complexity of data prediction, thus significantly improving the prediction speed and prediction accuracy. This is mainly due to (1) In the data preprocessing stage, in addition to completing the missing data and correcting the biased data in the dataset, this invention also uses the CDC clustering algorithm to cluster the original dataset. This clustering algorithm divides each data into a specific group, so that the data in the same group have highly similar feature patterns. This makes it possible for the prediction model to learn the feature patterns of the data more quickly when predicting the data, so that the prediction results are more accurate and the error is smaller. (2) In the VMD modal decomposition stage, the step of "calculating the sample entropy of each decomposed load sequence and merging modal components with similar sample entropy values ​​to form new components" is added. This allows the model to quickly learn the regularity of the new components during model training, greatly reducing the training complexity and significantly improving the training speed and prediction accuracy. (3) The structure of the ESN model is improved by constructing a deep architecture to capture the multi-scale dynamics of the time series. An encoder layer is constructed based on the ESN model structure, and feature connections are added between each encoder layer and the last output layer. This fully utilizes high-dimensional projection to obtain a rich, multi-scale dynamic representation of the input, ultimately resulting in the DESN model. The DESN model reduces the complexity of data training to that of simple linear regression, while maintaining high computational efficiency without relying on gradient propagation algorithms. Attached Figure Description

[0066] Figure 1 This invention provides a preferred embodiment of a short-term power load forecasting flowchart that combines CDC-VMD and echo state networks.

[0067] Figure 2 This is a schematic diagram of a deep echo state network structure. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0069] The technical solution of the present invention to solve the above-mentioned technical problems is:

[0070] like Figure 1 As shown, this implementation provides a short-term power load forecasting method that combines local orientation centrality clustering algorithm, variational mode decomposition, and deep echo state network, including the following steps:

[0071] Step 1: Cluster the raw load data using three algorithms: Bi-Kmeans, FCM, and CDC. The Davies-Bouldin Index (DBI) is then used to analyze the clustering results. Based on the DBI index, CDC clustering is deemed to have the best clustering performance. The CDC clustering algorithm flowchart is as follows:

[0072] (1) Calculate the angular variance (a measure of local orientation centrality). The angular variance is used to characterize the uniformity of the distribution of K nearest neighbors (KNN) of each point. That is, the smaller the variance, the more uniform the distribution, meaning that the K nearest neighbors (KNN) of the target point are evenly distributed around it, surrounding the target point; the larger the variance, the more uneven the distribution, meaning that the K nearest neighbors (KNN) of the target point are biased towards a single direction and cannot be surrounded by K nearest neighbors (KNN).

[0073]

[0074] (2) Based on the local orientation centrality measure, a threshold is set to divide the boundary points and interior points, where the DCM value of the boundary points is greater than the threshold and the DCM value of the interior points is less than the threshold.

[0075] (3) Generate connections between constrained interior points, i.e., divide each cluster. Specific method: Calculate the connections between boundary points and each interior point p. i The closest distance r i Calculate the distance d between interior points. ij If r i +r j ≥d ij Then the interior point p i With p j Those in the same cluster, otherwise in different clusters;

[0076]

[0077] Internal point cluster partitioning criteria:

[0078] d(p i ,p j )≤r i +r j (3)

[0079] To verify the effectiveness of the proposed load clustering method, the Davies-Bouldin Index (DBI) was used to analyze the clustering results. The calculation formula is as follows:

[0080]

[0081] In the formula: w is the average distance from the data within a cluster to the cluster centroid, representing the dispersion of time series within cluster i. i This represents the cluster center of each cluster. For each cluster, DBI first finds the average intra-class distance (...). and The sum of the distances within each cluster is divided by the maximum distance between the two cluster centers. These maximum distances are then summed to obtain the average. A smaller DBI indicates smaller intra-cluster distances and larger inter-cluster distances. Comparing the clustering results of the CDC algorithm with Bi-Kmeans and FCM, it was found that the CDC algorithm has the lowest DBI and the best clustering effect.

[0082] Step 2: Apply Variational Mode Decomposition (VMD) to the load sequence to obtain the decomposed sequence. The complete process of Variational Mode Decomposition is as follows:

[0083] (1) Initialization

[0084] (2) For each subsequence, continuously update according to (5) and (6) to obtain and

[0085]

[0086]

[0087] In the formula: Wiener filtering for the current remaining components, The frequency center of the corresponding modal component is ω, where ω is the frequency value.

[0088] (3) For all ω≥0, update

[0089]

[0090] (4) Determine whether the iteration termination condition is met:

[0091]

[0092] If the termination condition is not met, repeat steps (2) and (3). If the condition is met, the iteration terminates, and the K subsequences after decomposition are obtained.

[0093] Step 3: Calculate the sample entropy of each decomposed sequence, and merge the modal components with similar sample entropy values to form new components. Specifically: Use sample entropy to evaluate the complexity of each component. The lower the value of the sample entropy, the higher the similarity of the sequence and the lower the complexity. For a time series {x(n)} = x(1), x(2), … x(N) consisting of N points, the calculation method of sample entropy is as follows:

[0094] (1) Form a vector sequence of dimension m by serial numbers, X m (1), …, X m (N - m + 1), where X m (i) = {x(i), x(i + 1), …, x(i + m - 1)}, (1 ≤ i ≤ N - m + 1);

[0095] (2) Define the distance d[X m (i), X m (j)] between vector X m (i) and X m (j) as the absolute value of the maximum difference between their corresponding elements;

[0096]

[0097] (3) Given a threshold r, record the number of j for which d[X m (i), X m (j)] < r, denoted as B i , and for 1 ≤ i ≤ N - m, record the ratio of B i to N - m + 1 as:

[0098]

[0099] (4) Calculate the average value for all to obtain:

[0100]

[0101] (5) Increase the dimension to m + 1, and calculate the number of X m+1 (i) and X m+1 (j) whose distance is less than or equal to r, denoted as A i , defined as:

[0102]

[0103] (6) Define A m (r) as:

[0104]

[0105] B m (r) and Am (r) represents the matching probability of sequence pair m and m+1, respectively. The sample entropy is defined as follows:

[0106]

[0107] When N takes finite values, the estimated value of the sample entropy is:

[0108]

[0109] Step 4: Normalize the merged new load components, mapping the load data to the range [0,1]. Scale the original data proportionally within a specific interval, converting it into dimensionless pure numerical values. Use the min-max normalization method to map the load to the range [0,1]. The calculation formula is as follows:

[0110]

[0111] Where, x min and x max ...

[0112] Step 5: Construct time series input-output label pairs using a sliding window; these input-output label pairs are used to train the model.

[0113] Step 6: Construct a short-term power load prediction model combining CDC-VMD and Echo State Networks. Traditional Echo State Networks (ESNs) have a very simple training process and possess nonlinear echo states and short-term memory characteristics, which are very useful for dynamic system modeling. However, a single ESN network cannot handle input signals requiring complex hierarchical processing. To address the high volatility of short-term power load sequences, Deep Echo State Networks (DESNs) are used to model and predict the time series of each power load modality. DESN is a hierarchical ESN model. Its main idea is to capture the multi-scale dynamics of the time series by constructing a deep architecture. DESN does not directly stack multiple ESNs; instead, it constructs encoder layers and adds connections (called feature connections) between each encoder layer and the last output layer, fully utilizing high-dimensional projection to obtain rich multi-scale dynamic representations of the input. Furthermore, DESN training is performed layer by layer on the entire model, with the only trainable layer being the last output layer. It maintains high computational efficiency without relying on gradient propagation algorithms. During network training, the Adam optimizer is used to adjust the network weight parameters and find the optimal values ​​for the network.

[0114] Step 7: Feed the normalized components into Deep Echo State Networks (DESN) for training;

[0115] Step 8: Reconstruct the load sequence of each trained component (i.e., algebraic summation) to obtain the predicted load data and output the prediction results.

[0116] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0117] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0118] The above embodiments should be understood as illustrative only and not as limiting the scope of protection of the present invention. After reading the description of the present invention, those skilled in the art can make various alterations or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A short-term power load forecasting method combining CDC-VMD and echo state networks, characterized in that, Includes the following steps: Step 1: Clean the raw power load data and use a clustering algorithm to cluster the cleaned load data. Analyze the clustering results using the Davidson-Bolding Index (DBI). Based on the DBI index, the CDC clustering algorithm is found to have the best clustering performance. Step 2: The clustered load sequences are decomposed using variational mode VMD to obtain the individual decomposed load sequences; Step 3: Calculate the sample entropy of each decomposed load sequence, and merge modal components with similar sample entropy values ​​to form new components. Step 4: Normalize the merged components and map them to the range [0,1]. Step 5: Construct time series input-output label pairs using a sliding window; these input-output label pairs are used to train the prediction model. Step 6: Construct a short-term power load forecasting model that combines CDC-VMD and echo state network, and use the Adam optimization algorithm to adjust the model weight parameters and find the optimal value of the network; Step 7: Feed the normalized components into the Deep Echo State Network (DESN) for training; Step 8: Reconstruct the load sequence of each trained component to obtain the predicted load data and output the prediction results; Step 1 involves clustering the raw load data, specifically including the following steps: (1) Calculate the angular variance, which is a measure of local orientation centrality. The angular variance is used to characterize the uniformity of the K nearest neighbor distribution of each point. （1） Represents the angle value. The number of angles represented. Representing the present For which angle value? (2) Based on the local orientation centrality measure, a threshold is set to divide the boundary points and interior points, where the DCM value of the boundary points is greater than the threshold and the DCM value of the interior points is less than the threshold; (3) Generate connections between constrained interior points, i.e., divide each cluster. The specific method is to calculate the connection between the boundary points and each interior point. closest distance Calculate the distance between interior points ,if Then the internal point and Those in the same cluster, otherwise in different clusters; （2） Internal point cluster partitioning criteria: （3） The Davidson-Bourdin Index (DBI) was used to analyze the clustering results. The calculation formula is as follows: （4） In the formula: It is the average distance from the data within a class to the cluster centroid, representing the cluster class. The degree of dispersion of each time series in the data. The DBI represents the cluster centers of each cluster; for each cluster, the DBI first finds the average internal class distance. and The sum is divided by the maximum distance between the two cluster centers; then these maximum values ​​are added together to get the average. The CDC clustering algorithm has the lowest DBI index and the best clustering effect.

2. The short-term power load forecasting method combining CDC-VMD and echo state networks according to claim 1, characterized in that, Step 2 involves using variational mode decomposition (VMD) to decompose the clustered load sequences into individual decomposed load sequences, specifically including the following steps: (1) Initialization , , ; ( Let the k-th modal component and the center frequency be represented respectively. The number 1 in the top left corner indicates the first iteration; (2) For each subsequence, continuously update according to (5) and (6) to obtain and ; （5） （6） In the formula: Wiener filtering for the current remaining components, The frequency center of the corresponding modal component, This is the frequency value; Representing the original sequence ,as well as Fourier transform, It is a secondary penalty factor; (3) For all ,renew : （7） Indicates noise tolerance. Indicates the total number of modes; (4) Determine whether the iteration termination condition is met: （8） If the termination condition is not met, repeat steps (2) and (3). If the condition is met, the iteration terminates, and the K subsequences after decomposition are obtained. This represents the similarity coefficient.

3. The short-term power load forecasting method combining CDC-VMD and echo state networks according to claim 1, characterized in that, Step 3: Calculate the sample entropy of each decomposed sequence, and merge modal components with similar sample entropy values ​​to form new components. Specifically, for a time series consisting of N points... The method for calculating sample entropy is as follows: (1) Arrange the vectors into a sequence of dimensions m according to their serial numbers. ,in , ; (2) Define vectors and Distance between It is the absolute value of the maximum difference among the corresponding elements of the two; （9） (3) Given a threshold r, record of The number, denoted as ,for ,Bundle and The ratio is denoted as: (10) (4) For all The average value is: （11） (5) Increase the dimension to ,calculate and Distance less than or equal to The number of them is denoted as , Defined as: （12） (6) Definition for: （13） and Sequence pairs and The matching probability of a point, where the sample entropy is defined as: （14） when When taking finite values, the estimated value of the sample entropy is: （15） 4. The short-term power load forecasting method combining CDC-VMD and echo state networks according to claim 3, characterized in that, Step 4: Normalize the merged new components, mapping them to the range [0,1]. Specifically, normalization involves scaling the original data proportionally within a specific interval, converting it into dimensionless pure numerical values, and then using the min-max normalization method to map the load to the range [0,1]. The calculation formula is as follows: （16） in, and These are the minimum and maximum values ​​of the sample data, respectively. This represents the value after mapping.

5. The short-term power load forecasting method combining CDC-VMD and echo state networks according to claim 4, characterized in that, Step 6: Constructing a Deep Echo State Network (DESN) short-term power load prediction model. Specifically, the DESN is used to model and predict the time series of each modality of power load. The DESN is a hierarchical ESN model. Its main idea is to capture the multi-scale dynamics of the time series by constructing a deep architecture. Instead of directly stacking multiple ESNs, the DESN constructs an encoder layer and adds feature connections between each encoder layer and the last output layer. It makes full use of high-dimensional projection to obtain a rich multi-scale dynamic representation of the input. Moreover, the DESN is trained layer by layer, with the only trainable layer being the last output layer. It maintains high computational efficiency without relying on gradient propagation algorithms. During network training, the Adam optimizer is used to adjust the network weight parameters and find the optimal values ​​of the network.

Images