A total load prediction method for an electric vehicle charging community
By combining CEEMDAN and VMD decomposition techniques with a BiGRU network and attention mechanism in a parallel prediction model, the problems of insufficient model analytical capability and high computational cost in total load prediction of electric vehicle charging communities are solved, achieving high-precision and fast load prediction results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- WUHAN UNIV OF SCI & TECH
- Filing Date
- 2026-06-16
- Publication Date
- 2026-07-14
AI Technical Summary
Existing methods for predicting total load in electric vehicle charging communities suffer from problems such as insufficient model analytical capabilities, high computational costs, an explosion of feature dimensions, and a surge in computing power when dealing with load data characterized by high volatility and high noise, making it difficult to achieve accurate load prediction.
The CEEMDAN decomposition technique is used to extract the IMF components. The feature vector is constructed by combining the sample entropy and zero-crossing rate. The parallel prediction model is constructed by using probe optimization and VMD adaptive secondary decoupling. The model includes a BiGRU network, an attention mechanism and a fully connected network layer, which independently model and predict low-frequency, mid-frequency and high-frequency subsequences.
It significantly improves the accuracy and speed of total load forecasting for electric vehicle charging communities, reduces high-frequency noise interference, enhances the model's generalization ability, and ensures the reliability and consistency of the forecast results.
Smart Images
Figure CN122393922A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of electric vehicle load forecasting technology, and more specifically to a method for forecasting the total load of an electric vehicle charging community. Background Technology
[0002] Currently, in recent years, electric vehicles, as important electrical devices, have seen a rapid increase in penetration rates in communities. Impacted by the large-scale, unregulated charging of electric vehicles, the total load time series in communities exhibits typical characteristics of high volatility and high noise, with frequent and dramatic transient high-frequency abrupt changes on the load curve, known as "spurts." Accurate forecasting of the total community load is therefore a fundamental guarantee for effectively preventing distribution transformer overload and ensuring the safe and stable operation of the power system.
[0003] However, for the complex load data mentioned above, conventional short-term power load forecasting models typically employ signal decomposition techniques such as Empirical Mode Decomposition (EMD), Ensemble Empirical Mode Decomposition (EEMD), or Fully Adaptive Noise Ensemble Empirical Mode Decomposition (CEEMDAN) to attempt to decompose the non-stationary original sequence into multiple relatively stationary subsequences to reduce noise interference. However, existing decomposition prediction methods have significant limitations and defects in practical applications. First, conventional single-step decomposition algorithms often fail to completely separate true features from irregular noise, resulting in insufficient analytical capabilities for key information. Second, traditional decomposition techniques inevitably generate a large number of intrinsic mode functions (IMFs). If independent prediction models are directly constructed for all components, it can easily lead to an explosion in feature dimensions and a surge in computational overhead. Furthermore, some studies have introduced Variational Mode Decomposition (VMD) for secondary decoupling, but the selection of VMD core parameters often relies on extremely time-consuming heuristic swarm intelligence optimization algorithms, resulting in excessively high computational costs and making it difficult to implement in practical engineering.
[0004] Therefore, how to provide a method for predicting the total load of electric vehicle charging communities that can solve the above problems is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] In view of this, the present invention provides a total load prediction method for communities with electric vehicle charging facilities, which is suitable for communities with a high proportion of electric vehicles and drastic load fluctuations; it has higher accuracy: especially at sudden change points such as charging peaks and late-night concentrated charging, the prediction error is significantly reduced.
[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for predicting the total load of a community including electric vehicle charging stations includes the following steps: S1: Obtain the basic power consumption sequence and electric vehicle charging pile power sequence for the same time period, and preprocess them to generate the corresponding community total load dataset; S2: Process the total load dataset of the community to obtain the corresponding low-frequency trend component, mid-frequency periodic component and high-frequency impact component; S3: Process the high-frequency impact components to obtain the corresponding high-frequency sub-sequence; S4: Construct a parallel prediction model, taking the low-frequency trend component, mid-frequency periodic component, and high-frequency subsequence as a sequence set, and inputting the sequence set into the parallel prediction model for prediction to obtain independent prediction results corresponding to each subsequence; S5: Linearly sum the independent prediction results to output the final community total load prediction result.
[0007] Preferably, S1 includes: S11: Obtain the basic power consumption sequence and electric vehicle charging pile power sequence for the same time period; S12: Clean up the missing values in the basic electricity consumption power sequence and the electric vehicle charging pile power sequence, and perform power aggregation at the same time sequence node to obtain the corresponding community household total basic electricity consumption power sequence and electric vehicle total charging power sequence. S13: Linearly superimpose the total basic electricity consumption sequence of community households and the total charging power sequence of electric vehicles at the same time node to form the total load dataset of the community.
[0008] Preferably, S2 includes: S21: Perform CEEMDAN decomposition on the total community load dataset to obtain the corresponding IMF components, construct the corresponding feature vectors based on the IMF components, and simultaneously construct the corresponding two-dimensional physical feature matrix using the feature vectors. S22: The feature vector is reduced in dimensionality using the K-Mean clustering algorithm to obtain the corresponding low-frequency trend component, mid-frequency periodic component and high-frequency impact component.
[0009] Preferably, S21 includes: S211: Add the total community load dataset to a preset white noise set to obtain the corresponding set signal; S212: Perform EMD decomposition on the set signal to obtain the corresponding first-order intrinsic mode function, and calculate the corresponding first-order residual sequence based on the first-order intrinsic mode function; S213: Add the (k-1)th order white noise mode component extracted by EMD decomposition to the (k-1)th order residual sequence to generate the corresponding sequence to be decomposed, and update the corresponding residual sequence at the same time. S214: Iterate through S213 until the number of extreme points of the latest updated residual signal is less than or equal to two, so that the envelope cannot be effectively divided, and then terminate the iteration to obtain the corresponding IMF component. S215: Calculate the sample entropy and zero-crossing rate of each IMF component, and construct the corresponding feature vector based on the calculation results. At the same time, construct the corresponding two-dimensional physical feature matrix using the feature vector.
[0010] Preferably, S3 includes: S31: Extract a continuous sequence of a preset front end length from the high-frequency impact component as an optimization probe segment; S32: Initialize the number of mode decompositions in VMD, perform VMD pre-decomposition on the optimization probe fragment, and extract the center frequency corresponding to the highest frequency mode generated by the decomposition; S33: Step by step, increase the number of modal decompositions, perform VMD pre-decomposition on the optimization probe segment again, and calculate the first-order absolute difference of the center frequency of the highest frequency mode under adjacent modal decomposition numbers; S34: Compare the first-order absolute difference with the preset convergence threshold. If it is greater than or equal to the threshold, return to execute S33. Otherwise, terminate the iteration and use the current modal decomposition number as the global optimal modal number. S35: The complete high-frequency impact component is decoupled by VMD using the global optimal mode number to obtain the corresponding high-frequency subsequence.
[0011] Preferably, S4 includes: S41: The low-frequency trend component, the mid-frequency periodic component, and the high-frequency subsequence are taken as the sequence set to be predicted, and the sequence set is preprocessed to obtain the corresponding input feature vector and the corresponding true prediction label. S42: Construct a parallel prediction model, wherein the parallel prediction model is implemented based on a BiGRU network, an attention mechanism layer, and a fully connected network layer; S43: Construct an independent BiGRU network for each sequence in the sequence set, and inject the input feature vector into the BiGRU network to output the corresponding bidirectional hidden state vector; S44: Input the bidirectional hidden state vector into the attention mechanism layer and output the corresponding attention weights; S45: The bidirectional hidden state vectors are weighted and summed according to the attention weights to obtain the corresponding global context feature vectors. These vectors are then input into the fully connected network layer for nonlinear mapping, and the independent prediction results under normalized dimensions are output.
[0012] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method for predicting the total load of electric vehicle charging communities, which has the following beneficial effects: 1. This invention splits the community load into two sequences: basic electricity consumption and electric vehicle charging power. After cleaning and time-series alignment, the sequences are aggregated. Missing values and outliers are preprocessed uniformly. CEEMDAN is used to adaptively decompose the total load, extracting the IMF component. A feature vector is constructed by combining sample entropy and zero-crossing rate. The complex load is automatically decomposed into low-frequency trend components, mid-frequency periodic components, and high-frequency impact components, so that each component with different characteristics performs its own function. This reduces the difficulty of predicting complex loads with a single model, ensures the time-series consistency and amplitude reliability of the total load dataset, and significantly improves the accuracy of subsequent decomposition and prediction. 2. This invention utilizes probe optimization and VMD adaptive secondary decoupling, using the first-order difference of the center frequency as the convergence criterion to achieve VMD parameter adaptation, and further finely decouples the high-frequency components to obtain a stable high-frequency subsequence, effectively suppressing high-frequency noise and improving prediction accuracy. 3. This invention models each component independently, and independently models and predicts low-frequency, mid-frequency, and high-frequency subsequences. BiGRU captures bidirectional temporal dependencies, adapts to the correlation characteristics before and after the load, the attention mechanism automatically focuses on key periods, strengthens the weight of important information, and the fully connected layer completes accurate fitting and dimensional normalization, resulting in faster overall prediction speed and stronger generalization ability. Attached Figure Description
[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0014] Figure 1 The present invention provides an overall flowchart of a total load prediction method for electric vehicle charging communities. Detailed Implementation
[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0016] See Figure 1 As shown in the figure, an embodiment of the present invention discloses a method for predicting the total load of a community including electric vehicle charging stations, comprising the following steps: S1: Obtain the basic power consumption sequence and electric vehicle charging pile power sequence for the same time period, perform preprocessing, and generate the corresponding community total load dataset X(t); S2: Process the total load dataset of the community to obtain the corresponding low-frequency trend component, mid-frequency periodic component and high-frequency impact component; S3: Process the high-frequency impact components to obtain the corresponding high-frequency sub-sequence; S4: Construct a parallel prediction model, taking the low-frequency trend component, mid-frequency periodic component, and high-frequency subsequence as a sequence set, and inputting the sequence set into the parallel prediction model for prediction to obtain independent prediction results corresponding to each subsequence; S5: Linearly sum the independent prediction results to output the final community total load prediction result.
[0017] In one specific embodiment, S1 includes: S11: Obtain the basic electricity consumption sequence of M community households within the same time period. Electric vehicle charging station power sequence for N households owning electric vehicles Where i = 1, 2, ..., M, j = 1, 2, ..., N, and M <N; S12: Linear interpolation is used to clean up missing values in the basic electricity consumption power sequence and the electric vehicle charging pile power sequence, and power aggregation is performed at the same time-series nodes to obtain the corresponding community household total basic electricity consumption power sequence. Total charging power sequence of electric vehicles Where t = 1, 2, ..., T, and T represents the total number of data points; S13: Linearly superimpose the total basic electricity consumption sequence of community households and the total charging power sequence of electric vehicles at the same time node to form the community total load dataset, i.e. .
[0018] In one specific embodiment, S2 includes: S21: Perform CEEMDAN decomposition on the total community load dataset to obtain the corresponding IMF components, and construct the corresponding feature vectors based on the IMF components. Simultaneously, the corresponding two-dimensional physical feature matrix is constructed using the feature vectors; S22: The feature vector is reduced in dimensionality using the K-Mean clustering algorithm to obtain the corresponding low-frequency trend component, mid-frequency periodic component and high-frequency impact component.
[0019] In one specific embodiment, S21 includes: S211: Preset the size of the white noise set to L, and define the operator. Given the k-th mode component of a signal generated by the EMD algorithm, add L noise amplitude adjustment parameters to the community total load dataset X(t). Adaptive white Gaussian noise sequence Construct a set of signals The noise set index is p = 1, 2, ..., L; S212: Perform EMD decomposition on the collected signals to obtain the first-order intrinsic mode functions of preliminary decoupling. And calculate the first-order residual sequence after stripping the mode based on the first-order intrinsic mode function. ; S213: To the (k-1)th order residual sequence By incorporating the (k-1)th order white noise mode component extracted through EMD decomposition, a new sequence to be decomposed is constructed. The first-order modal components are extracted and the ensemble average is calculated to obtain the k-th order modal components. At the same time, update the corresponding residual sequence. ,in This represents the noise amplitude adjustment parameter in the (k-1)th order residual decomposition process. It is used to weight the (k-1)th order white noise mode components obtained after EMD decomposition so that the amplitude of the auxiliary noise added to the residual sequence matches the fluctuation scale of the current residual sequence. S214: Iterate through S213 until the number of extreme points in the latest updated residual signal is less than or equal to two, causing the envelope to be unable to be effectively divided. At this point, the iteration terminates. The highly volatile original community total load sequence is completely decoupled. The sum of the intrinsic mode functions with different degrees of stability and the final residual yields the corresponding IMF components, i.e. ; S215: Calculate the sample entropy and zero-crossing rate of each IMF component, and construct the corresponding feature vector based on the calculation results. At the same time, construct the corresponding two-dimensional physical feature matrix using the feature vector.
[0020] Specifically, the implementation process of S215 includes: S2151: Eigenmode functions of the length T of the decoupled output Set the embedding dimension m and the tolerance threshold r, where , This represents the standard deviation of the corresponding modal components; Calculate the sample entropy of each component separately. ,in , These represent the number of sequence template matching pairs that satisfy Chebyshev distance less than or equal to r under embedding dimensions m+1 and m, respectively. S2152: Calculate the zero-crossing rate of each eigenmode function from its own mean. The frequency domain center frequency characteristic of the signal is characterized by the following expression:
[0021] In the formula, Let represent the time series mean of the k-th IMF. Indicates an indicator function, Represents a symbolic function; S2153: Combine the sample entropy and zero-crossing rate of each eigenmode function to construct an feature vector. The maximum-minimum normalization strategy is used to map the features of all components to the same dimension interval [0,1]. Subsequently, a two-dimensional physical feature matrix covering time-domain complexity and frequency-domain centrality can be constructed.
[0022] Specifically, the implementation process of S22 includes: K-Mean clustering algorithm is used to analyze feature vectors Dimensionality reduction is performed, reconstructing redundant modes into low-frequency trend components, mid-frequency periodic components, and high-frequency impulse components. The specific process includes: S221: Set the initial number of clusters for the K-Means clustering algorithm to 3, and include K... ce Using the two-dimensional normalized physical feature matrix of the intrinsic mode functions as input, unsupervised clustering is performed, and the cluster centers are iteratively updated until the objective function converges, obtaining the coordinates of the three cluster centers. Cluster index ; S222: Calculate separately Euclidean distance to the origin of the two-dimensional physical feature space The calculation formula is: ; S223: According to The three clusters are sorted in ascending order by their size, and then physically mapped into low-frequency trend clusters, mid-frequency periodic clusters, and high-frequency impulse clusters according to the signal complexity corresponding to the distance from smallest to largest. S224: Integrate members belonging to the same cluster Linear superposition at the same time points completes the dimensionality reduction and reconstruction of modal features, and sequentially outputs low-frequency trend components representing the basic electricity consumption behavior of the community. The mid-frequency periodic component representing the regular circadian rhythm And the high-frequency impact component that accurately represents the abrupt changes in the disordered charging characteristics of electric vehicles. .
[0023] In one specific embodiment, S3 includes: S31: Extract the high-frequency impact component As the complex target signal to be analyzed, a preset front-end length is used to reduce the computational overhead of parameter optimization. The continuous sequence is used as the optimization probe fragment, in which ; S32: Initialize the mode decomposition number of VMD The optimization probe fragment is pre-decomposed using VMD to extract the center frequency corresponding to the highest frequency mode generated by the decomposition. ; S33: Stepwise increase the modal decomposition number, i.e., update The VMD pre-decomposition of the optimization probe fragment is performed again, and the first-order absolute difference of the center frequency of the highest frequency mode under the adjacent modal decomposition number is calculated. ; S34: The first-order absolute difference... With preset convergence threshold The comparison is performed. If the result is greater than or equal to the given value, the process returns to step S33. If this occurs for the first time, the iteration terminates. This indicates that the high-frequency impact mode has fully converged and no longer generates new high-frequency features with independent physical meaning. At this point, the optimization iteration is terminated, and the current mode decomposition number is set. As the globally optimal mode number ; S35: Employs the globally optimal number of modes. For the complete high-frequency impact component of length T Perform secondary decoupling of VMD, and completely decompose it into A number of independent high-frequency subsequences with limited bandwidth and strong regularity. Subsequence index .
[0024] In one specific embodiment, S4 includes: S41: The low-frequency trend component Intermediate frequency periodic components High-frequency subsequences As a set of sequences to be predicted , set of sequences Each sequence in the dataset is divided into a training set, a validation set, and a test set in a 70:15:15 ratio; the specific process of preprocessing the sequence set may include: For each sequence It calculates the maximum and minimum values and fits a scaling parameter using only its training set data, and then applies this parameter to the global sequence to obtain a dimensionless normalized sequence. Then, a sliding time window of length S is used in the sequence. Stepwise truncation is performed to construct autoregressive sample pairs: for a given current prediction time t, continuous data within its historical window is extracted as the input feature vector. and the corresponding true prediction labels ; S42: Construct a parallel prediction model, wherein the parallel prediction model is implemented based on a BiGRU network, an attention mechanism layer, and a fully connected network layer; S43: Construct an independent BiGRU network for each sequence in the sequence set, and process the input feature vector... The input is fed into the BiGRU network at the internal time step. BiGRU calculates the forward hidden state respectively With backward hidden state The two are then concatenated and merged into a complete bidirectional hidden state vector. To synchronously capture the historical and future evolution dependencies of time-series data; S44: Transfer the bidirectional hidden state vector The input attention mechanism layer uses a non-linear activation function to compute... Attention score Attention weights are obtained by normalization using the Softmax function. This weight dynamically assesses the importance of the abrupt change characteristics at each historical moment within the sliding window to the current prediction target; S45: Perform a weighted summation of the bidirectional hidden state vectors based on the attention weights to obtain the global context feature vector. The data is then fed into a fully connected network layer for nonlinear mapping, and the output is an independent prediction result under normalized dimensions. .
[0025] Specifically, the implementation process of S5 includes: S51: Extract independent prediction results At the same time, call S41 to get By using the inverse mapping function to reverse the process, we can obtain the predicted values of each independent component that are restored to the true physical power dimension. ; S52: Based on the sequence set The original composition mapping relationship will be denormalized. The predicted value strictly corresponds to , as well as ; S53: At the same time node t+1, the predicted values of all the above subsequences are linearly superimposed, and the mathematical aggregation model is as follows:
[0026] Final output This serves as an accurate short-term total load forecast for a microgrid in a community with electric vehicle charging facilities at time t+1.
[0027] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.
[0028] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for predicting the total load of a community including electric vehicle charging stations, characterized in that, Includes the following steps: S1: Obtain the basic power consumption sequence and electric vehicle charging pile power sequence for the same time period, and preprocess them to generate the corresponding community total load dataset; S2: Process the total load dataset of the community to obtain the corresponding low-frequency trend component, mid-frequency periodic component and high-frequency impact component; S3: Process the high-frequency impact components to obtain the corresponding high-frequency sub-sequence; S4: Construct a parallel prediction model, taking the low-frequency trend component, mid-frequency periodic component, and high-frequency subsequence as a sequence set, and inputting the sequence set into the parallel prediction model for prediction to obtain independent prediction results corresponding to each subsequence; S5: Linearly sum the independent prediction results to output the final community total load prediction result.
2. The total load forecasting method for a community including electric vehicle charging stations according to claim 1, characterized in that, S1 includes: S11: Obtain the basic power consumption sequence and electric vehicle charging pile power sequence for the same time period; S12: Clean up the missing values in the basic electricity consumption power sequence and the electric vehicle charging pile power sequence, and perform power aggregation at the same time sequence node to obtain the corresponding community household total basic electricity consumption power sequence and electric vehicle total charging power sequence. S13: Linearly superimpose the total basic electricity consumption sequence of community households and the total charging power sequence of electric vehicles at the same time node to form the total load dataset of the community.
3. The total load forecasting method for a community including electric vehicle charging stations according to claim 1, characterized in that, S2 includes: S21: Perform CEEMDAN decomposition on the total community load dataset to obtain the corresponding IMF components, construct the corresponding feature vectors based on the IMF components, and simultaneously construct the corresponding two-dimensional physical feature matrix using the feature vectors. S22: The feature vector is reduced in dimensionality using the K-Mean clustering algorithm to obtain the corresponding low-frequency trend component, mid-frequency periodic component and high-frequency impact component.
4. The total load forecasting method for a community including electric vehicle charging stations according to claim 3, characterized in that, S21 includes: S211: Add the total community load dataset to a preset white noise set to obtain the corresponding set signal; S212: Perform EMD decomposition on the set signal to obtain the corresponding first-order intrinsic mode function, and calculate the corresponding first-order residual sequence based on the first-order intrinsic mode function; S213: Add the (k-1)th order white noise mode component extracted by EMD decomposition to the (k-1)th order residual sequence to generate the corresponding sequence to be decomposed, and update the corresponding residual sequence at the same time. S214: Iterate through S213 until the number of extreme points of the latest updated residual signal is less than or equal to two, so that the envelope cannot be effectively divided, and then terminate the iteration to obtain the corresponding IMF component. S215: Calculate the sample entropy and zero-crossing rate of each IMF component, and construct the corresponding feature vector based on the calculation results. At the same time, construct the corresponding two-dimensional physical feature matrix using the feature vector.
5. The total load forecasting method for a community including electric vehicle charging stations according to claim 1, characterized in that, S3 includes: S31: Extract a continuous sequence of a preset front end length from the high-frequency impact component as an optimization probe segment; S32: Initialize the number of mode decompositions in VMD, perform VMD pre-decomposition on the optimization probe fragment, and extract the center frequency corresponding to the highest frequency mode generated by the decomposition; S33: Step by step, increase the number of modal decompositions, perform VMD pre-decomposition on the optimization probe segment again, and calculate the first-order absolute difference of the center frequency of the highest frequency mode under adjacent modal decomposition numbers; S34: Compare the first-order absolute difference with the preset convergence threshold. If it is greater than or equal to the threshold, return to execute S33. Otherwise, terminate the iteration and use the current modal decomposition number as the global optimal modal number. S35: The complete high-frequency impact component is decoupled by VMD using the global optimal mode number to obtain the corresponding high-frequency subsequence.
6. The total load forecasting method for a community including electric vehicle charging stations according to claim 1, characterized in that, S4 includes: S41: The low-frequency trend component, the mid-frequency periodic component, and the high-frequency subsequence are taken as the sequence set to be predicted, and the sequence set is preprocessed to obtain the corresponding input feature vector and the corresponding true prediction label. S42: Construct a parallel prediction model, wherein the parallel prediction model is implemented based on a BiGRU network, an attention mechanism layer, and a fully connected network layer; S43: Construct an independent BiGRU network for each sequence in the sequence set, and inject the input feature vector into the BiGRU network to output the corresponding bidirectional hidden state vector; S44: Input the bidirectional hidden state vector into the attention mechanism layer and output the corresponding attention weights; S45: The bidirectional hidden state vectors are weighted and summed according to the attention weights to obtain the corresponding global context feature vectors. These vectors are then input into the fully connected network layer for nonlinear mapping, and the independent prediction results under normalized dimensions are output.