Multi-element improved prediction method based on charging load and charging facility planning method

Abstract

The invention discloses a multivariate improvement prediction method based on a charging load and a charging facility planning method. The prediction method comprises the following steps: obtaining historical charging data and an association sequence of an electric vehicle charging facility; generating an adversarial network through a time sequence, constructing historical charging data into an enhanced data sequence, and calculating an MLE index of the enhanced data sequence; screening out K strong correlation sequences related to the charging load of the electric vehicle from the correlation sequence; an LSTM network is constructed, hyper-parameters of the LSTM network are adjusted according to the MLE index, and a charging load prediction model of the electric vehicle charging facility is obtained; and inputting the K strong correlation sequences into the electric vehicle charging facility charging load prediction model to obtain a charging load prediction result of the electric vehicle charging facility.

Description

Technical Field

[0001] This invention relates to the field of charging planning technology, and in particular to a multivariate improved prediction method based on charging load and a charging facility planning method. Background Technology

[0002] The charging load generated by the massive number of electric vehicles connected to the power distribution network has strong spatiotemporal distribution uncertainty. In order to eliminate the problems of local overload and power quality deterioration caused by its randomness and volatility to the power supply and distribution network and to meet the charging needs, it is necessary to plan the charging facilities in a reasonable manner.

[0003] However, in practice, electric vehicle charging infrastructure in most cities is still in the early stages of construction. Existing charging infrastructure planning relies heavily on traditional experience or simple statistical models, failing to fully capture the dynamic evolution and nonlinear correlation characteristics of charging load. Furthermore, it fails to effectively address the complex scenarios involving chaotic characteristics and the coupling of multiple influencing factors, resulting in a lack of scientific rigor and foresight in planning schemes. This often leads to problems such as supply-demand mismatch and polarized utilization rates. Moreover, most designs are static, making it difficult to adapt to future growth and evolution of charging demand. They are prone to failure when the external environment changes, requiring frequent adjustments and modifications, significantly increasing construction and operating costs. In summary, existing technologies face multiple bottlenecks in charging infrastructure planning, including insufficient data support, limited accuracy of prediction models, and poor adaptability of planning schemes. These limitations make it difficult to meet the dual demands of safe operation of the power distribution network and large-scale charging of electric vehicles. Therefore, there is an urgent need for a charging infrastructure planning method that can overcome data limitations, accurately predict charging load, and achieve dynamic optimization. Summary of the Invention

[0004] In view of the above-mentioned prior art, the present invention provides a multi-dimensional improved prediction method based on charging load and a charging facility planning method, which mainly solves the technical problems existing in the above-mentioned background art.

[0005] To achieve the above objectives, the technical solution of this invention is implemented as follows: The first aspect of this invention discloses a multivariate improved prediction method based on charging load, the prediction method comprising the following steps: Obtain historical charging data and associated sequences for electric vehicle charging facilities; By using a time-series generative adversarial network, historical charging data is constructed into an augmented data sequence, and the MLE exponent of the augmented data sequence is calculated. K strongly correlated sequences related to electric vehicle charging load were selected from the associated sequences; An LSTM network is constructed, and the hyperparameters of the LSTM network are adjusted according to the MLE exponent to obtain a charging load prediction model for electric vehicle charging facilities. K strongly correlated sequences are input into the electric vehicle charging facility charging load prediction model to obtain the electric vehicle charging facility charging load prediction results.

[0006] Optionally, the associated sequence includes a temperature sequence. ,humidity Rainfall Wind speed Solar irradiance Date sequence data of historical charging load data sampling points for charging facilities at corresponding times. And population sequence data within a 10km radius of the i-th charging facility. Serial data of per capita GDP of the area to be planned .

[0007] Optionally, historical charging data can be constructed into an enhanced data sequence using a time-series generative adversarial network, specifically including: The historical charging data is dimensionality reduced by using the embedding function in the time series generative adversarial network to obtain several first feature values, and the static and temporal features of the historical charging data are extracted simultaneously. The reconstruction function in the time series generative adversarial network is used to reconstruct the dimensionality-reduced historical charging data, obtain the reconstruction sequence, calculate several second feature values ​​corresponding to the reconstruction sequence, calculate the reconstruction loss of the reconstruction sequence, and if the reconstruction loss exceeds a set threshold, the embedding and reconstruction process is repeated. New sequence data is generated based on the static and temporal features. The generated sequence data is then encoded together with the features extracted by the embedding function and input into the discriminator. The generated sequence data is then judged to be true or false in conjunction with the historical charging data. The generated data that is judged to be true is retained. The generation loss is calculated based on the distribution consistency between the generated data and the historical charging data. If the generation loss exceeds a set threshold, the embedding and generation process is repeated. The reconstruction loss and generation loss are iterated repeatedly until both meet the set threshold, and finally an enhanced data sequence consistent with the feature patterns and distribution characteristics of the historical charging data is obtained.

[0008] Optionally, before calculating the MLE exponent of the augmented data sequence, phase space reconstruction is performed on the augmented data sequence, which specifically includes: The maximum time delay search range is set based on the length of the augmented data sequence; By iterating through each time delay within the search range, the enhanced data sequence is reconstructed into a first sequence and a second sequence; Calculate the marginal probabilities of the first sequence and the second sequence respectively, and calculate the joint probability of the first sequence and the second sequence; Based on marginal probabilities and joint probabilities, the average mutual information function for calculating the enhanced data sequence is obtained. Based on the calculation results of the average mutual information function, the time delay corresponding to the first occurrence of a local minimum is taken as the final time delay; Using the determined final time delay, construct an m-dimensional spatial trajectory and an m+1-dimensional spatial trajectory; Calculate the nearest neighbor of each m-dimensional trajectory point, and calculate the distance mutation ratio of the nearest neighbor in the m+1 dimension. Determine the optimal embedding dimension based on the threshold of the distance mutation ratio. Based on the optimal embedding dimension and the final time delay, the enhanced data sequence is reconstructed into a high-dimensional trajectory vector.

[0009] Optionally, the Wolf method can be used to calculate the MLE exponent of the high-dimensional trajectory vector, and its chaotic characteristics can be determined based on the MLE exponent.

[0010] Optionally, the chaotic characteristics can be determined based on the MLE index, specifically including: when the MLE index > 0, it indicates that the load evolution has chaotic characteristics, is sensitive to future disturbances, and is difficult to predict in the long term; when the MLE index < 0 or the MLE index = 0, the load change tends to be stable or periodic, and the prediction uncertainty is low.

[0011] Optionally, the number of hidden layer units, the Dropout rate, and the learning rate of the LSTM network can be determined based on the MLE index.

[0012] Optionally, K strongly correlated sequences related to electric vehicle charging load can be selected from the associated sequences, specifically including: calculating the augmented data sequence and the temperature sequence respectively. ,humidity Rainfall Wind speed Solar irradiance Date sequence data of historical charging load data sampling points for charging facilities at corresponding times. And population sequence data within a 10km radius of the i-th charging facility. Serial data of per capita GDP of the area to be planned The Pearson correlation coefficient between them was used to select k strongly correlated sequences that are strongly correlated with electric vehicle charging load. These k strongly correlated sequences were then included in the association sequence.

[0013] A second aspect of this invention discloses a charging facility planning method based on the prediction method described in any one of the foregoing claims, the planning method comprising: The coordinates of multiple charging facilities to be constructed are generated in advance, and the strong correlation sequence between each charging facility and the electric vehicle charging load is obtained. The strongly correlated sequence of each charging facility is input into the electric vehicle charging facility charging load prediction model to obtain the predicted charging load data. Based on load forecast data, indicators including average load, maximum load, minimum load, and charging facility utilization rate are calculated. If the calculated results of the corresponding indicators are lower than the preset threshold, the m-th charging facility to be constructed is removed; otherwise, it is retained.

[0014] The beneficial effects of this invention are as follows: By acquiring historical charging data of the planned area for electric vehicle charging facilities, location parameters of the charging facilities, and weather factor sequence data, date sequence data, population sequence data, and per capita GDP sequence data corresponding to the charging load of these charging facilities; then, based on a time-series generative adversarial network, an enhanced data sequence for generating the historical charging load data of the original charging facilities is constructed, and K sequences strongly correlated with the electric vehicle charging load are selected; on this basis, chaos theory analysis is introduced, and the chaotic characteristics of the charging load sequence are extracted by calculating the maximum Lyapunov exponent, and a continuous mapping function is designed based on this exponent to dynamically configure the key hyperparameters of the Long Short-Term Memory Recurrent Neural Network (LSTM), thereby constructing a multivariate improved prediction model that integrates TimeGAN and chaotic features. Using this model, the charging load of the proposed charging facilities under different coordinates is obtained, and the maximum charging load, utilization rate, and other indicators of the charging facilities are quantitatively analyzed, ultimately forming a reasonable charging facility planning scheme for the planned area. The proposed method transforms the traditional static planning problem, which relies on limited historical data, into a dynamic optimization problem that is oriented towards multi-scenario evolution by introducing data augmentation and chaotic optimization mechanisms. This not only significantly improves the scientific nature and foresight of the planning process, but also enhances the robustness and feasibility of the planning results in the face of the uncertainties of the large-scale development of electric vehicles. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the multivariate improved prediction method based on charging load in the embodiments of this application; Figure 2 This is a flowchart illustrating the multivariate improved prediction method and charging facility planning method based on charging load in the embodiments of this application. Detailed Implementation

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. In the following description, the expression "some embodiments" refers to a subset of all possible embodiments; however, it should be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments and can be combined with each other without conflict.

[0017] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention can be practiced without one or more of these details. In other instances, certain technical features well-known in the art have not been described in order to avoid obscuring the invention.

[0018] It should be understood that the present invention can be embodied in various forms and should not be construed as being limited to the embodiments set forth herein. Rather, providing these embodiments will make the disclosure thorough and complete, and will fully convey the scope of the invention to those skilled in the art. Furthermore, the terminology used herein is intended only to describe particular embodiments and is not intended to limit the invention. When used herein, the singular forms “a,” “an,” and “the” are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “compose” and / or “comprising,” when used in this specification, identify the presence of the stated features, integers, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups. When used herein, the term “and / or” includes any and all combinations of the associated listed items.

[0019] It should also be noted that when an element is referred to as being "fixed to" another element, it can be directly attached to the other element or there may be an intervening element. When an element is referred to as being "connected to" another element, it can be directly connected to the other element or there may be an intervening element. The terms "vertical," "horizontal," "inner," "outer," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible implementation.

[0020] To fully understand this invention, a detailed structure will be presented in the following description to illustrate the technical solution proposed by this invention. Optional embodiments of the invention are described in detail below; however, in addition to these detailed descriptions, the invention may have other embodiments.

[0021] Please refer to the attached document. Figures 1 to 2 The first aspect of this application provides a multivariate improved prediction method based on charging load, the prediction method comprising the following steps: S1. Obtain historical charging data and associated sequences of electric vehicle charging facilities; Specifically, acquiring all of the areas to be planned N The charging load history curve for each of the charging facilities is denoted as . , , , , Representing the first i The first, t, and e-th data points of the historical charging load data for each charging facility. The coordinates of the i-th charging facility are: .

[0022] Based on the coordinates of the i-th charging facility, obtain the corresponding association sequence, which includes the temperature sequence of the i-th charging facility. ,humidity Rainfall Wind speed Solar irradiance Date sequence data of historical charging load data sampling points for charging facilities at corresponding times. And population sequence data within a 10km radius of the i-th charging facility. Serial data of per capita GDP of the area to be planned

[0023] S2. Using a time-series generative adversarial network, historical charging data is constructed into an enhanced data sequence, and the MLE exponent of the enhanced data sequence is calculated.

[0024] In some alternative implementations, the historical charging load data of the i-th charging facility is processed through an embedding function in a time-series generative adversarial network and a time-series feature extraction mechanism. Dimensionality reduction involves preserving the static characteristics of the data, such as statistical attributes like average and peak loads, while also capturing temporal characteristics, such as the alternation of load peaks and troughs and time-series dependencies. The final output consists of S core feature values, denoted as follows: , ... ; Furthermore, by using a reconstruction function in a time-series generative adversarial network, the dimensionality-reduced historical charging data is back-mapped to the original charging load data space, thus reconstructing historical charging load data. To verify the consistency of features between the reconstructed data and the original data, calculations were performed. The corresponding S eigenvalues , ... The feature differences between the two are calculated using a reconstruction loss function, and the formula for calculating the loss value L1 is as follows:

[0025] In the formula, To reconstruct the sequence feature values, These are the original embedded feature values.

[0026] If the loss value L1 exceeds the preset threshold, it means that the reconstructed data has not fully preserved the original features. The embedding and reconstruction process needs to be repeated until L1 meets the accuracy requirements to ensure the effectiveness of the reconstruction mechanism.

[0027] Furthermore, based on the static and temporal features extracted by the embedding function, a new load sequence is generated using a Time Series Generative Adversarial Network (TimeGAN). This generated result is then co-encoded with the original features output by the embedding function and input into the discriminator. The discriminator uses the original historical charging data... As real samples, the generated data is judged to be true or false, and the generated data that the discriminator judges as "true" is retained. Simultaneously calculate the original data. The pointwise probability distribution F1 (i.e., the probability density of the load value at each time point) and the generated data The pointwise probability distribution F2 is used to quantify the difference between the two distributions using the loss function L2. The formula for calculating the loss value L2 is as follows:

[0028] If L2 exceeds the set threshold, it indicates that the generated data deviates significantly from the distribution of the real data, and the embedding and generation process needs to be repeated until L2 meets the requirements.

[0029] The iteration terminates when L1 ≤ and L2 ≤ the preset threshold. At this point, the generator has the ability to stably generate high-fidelity data. The generator generates a large number of payload sequences in batches. All generated data that are judged as "true" by the discriminator and meet the loss requirements together constitute the augmented data sequence. .

[0030] By iteratively optimizing the reconstruction loss L1 and the generation loss L2, and continuously refining the parameters of the embedding, reconstruction, and generation processes, an augmented data sequence that is highly consistent with the original historical charging data in terms of feature patterns and distribution characteristics is finally obtained. This sequence not only expands the data scale but also covers the charging load changes in multiple scenarios (such as different peak and valley periods and different load intensities), completely breaking through the limitations of the original historical charging data being sparse and covering limited scenarios.

[0031] S3. Select K strongly correlated sequences related to electric vehicle charging load from the associated sequences; Specifically, the augmented data sequences are calculated separately. With temperature sequence ,humidity Rainfall Wind speed Solar irradiance And population sequence data within a 10km radius of the i-th charging facility. Serial data of per capita GDP of the area to be planned The Pearson correlation coefficient between them was used to select k strongly correlated sequences with electric vehicle charging load, and these sequences were then reordered as follows: , ,……, The k strongly correlated sequences selected are included in the associated sequences, for example, temperature sequences. ,humidity Rainfall Wind speed These are four sequences that are strongly correlated with the charging infrastructure and the electric vehicle charging load, and they are reordered as follows: , ,……, .

[0032] In some implementations, phase space reconstruction of the augmented data sequence is performed before calculating the MLE exponent, specifically including: Based on enhanced data sequences The length M is used to set the maximum time delay search range. ; Then based on each time delay Iterate through each time delay within the search range and reconstruct the enhanced data sequence into the first sequence. and the second sequence Simultaneously, X and Y will be divided into... A number of equal-width intervals.

[0033] Calculate the marginal probabilities of the first sequence respectively Marginal probability of the second sequence And calculate the joint probability of the first sequence and the second sequence. ; Based on marginal probabilities and joint probabilities, the average mutual information function for calculating the enhanced data sequence is obtained. The expression for the average mutual information function is as follows:

[0034] Based on the expression for the time delay τ, the delay τ corresponding to the first occurrence of a local minimum is determined as the final time delay, ensuring minimal redundancy and maximum independence among the components of the reconstructed vector. Its expression is:

[0035] Starting with m=2, increment the time trial up to m=max, for each m. Utilize the determined final time delay. Constructing m-dimensional spatial trajectories and (m+1) dimensional spatial trajectory ,right Find its nearest neighbor and calculate the distance abrupt change ratio of that nearest neighbor in dimension m+1. Determine the optimal embedding dimension based on the threshold of the distance abrupt change ratio, where the expression for calculating the distance abrupt change ratio is:

[0036] like If its decrease is less than 0.01, then take the current value. m For optimal embedding dimension .

[0037] Based on the optimal embedding dimension and the final time delay, the augmented data sequence is reconstructed into a high-dimensional trajectory vector, the expression of which is:

[0038] In some alternative implementations, a simplified Wolf algorithm is applied to calculate the MLE exponent (maximum Lyapunov exponent) of the high-dimensional trajectory vector, specifically as follows: Set the reference point as Then the initial distance between the two points is In time t After step, it evolved into d ( t If the system exhibits chaotic characteristics, then the neighbor trajectories will diverge exponentially, satisfying the dynamic relationship shown in the following equation.

[0039]

[0040] Taking the natural logarithm of both sides of the above equation and simplifying further, we get: ,in, b =ln d (0). Calculate its regression slope. This refers to the enhanced data sequence of historical charging load data. The estimated value of the Lyapunov index.

[0041] Further transformation of the simplified result yields the specific function for calculating the maximum Lyapunov exponent, as shown in the following equation:

[0042] Based on the calculated maximum Lyapunov index Determine the dynamic properties of the charging load sequence. When When the value is greater than 0, it indicates that the load evolution has chaotic characteristics, the system is sensitive to future disturbances, and long-term prediction is difficult; when <0 or =0, indicating that load changes tend to be stable or periodic, and the prediction uncertainty is low. This criterion will serve as the core basis for the dynamic configuration of hyperparameters in subsequent LSTM prediction models.

[0043] This phase space reconstruction method enables the optimal time delay and embedding dimension to reconstruct the charging load enhancement data sequence to the greatest extent possible. The inherent dynamic characteristics avoid redundancy or information loss in the reconstructed vectors caused by improper selection of the optimal time delay. Furthermore, the distance mutation ratio of the (m+1)-dimensional trajectory ensures that the embedding dimension m is minimal and sufficient; it avoids trajectory compression distortion due to excessively small m, and also prevents increased computational complexity due to excessively large m introducing redundant dimensions. The reconstructed high-dimensional trajectory vector fully preserves the spatiotemporal correlation and evolution of the original sequence, providing a precise and reliable phase space foundation for subsequent calculation of the maximum Lyapunov exponent (MLE) using the Wolf method. This makes the extraction of chaotic characteristics more consistent with the true dynamic nature of the charging load sequence, thus providing a scientific basis for the dynamic configuration of hyperparameters in the subsequent LSTM model and significantly improving the accuracy and stability of charging load prediction.

[0044] S4. Construct an LSTM network and adjust the hyperparameters of the LSTM network according to the MLE exponent to obtain a charging load prediction model for electric vehicle charging facilities. Specifically, the basic structure of a Long Short-Term Memory Recurrent Neural Network (LSTM) is first built, and the input and output dimensions of the model are clarified. The input parameters are set as K sequences that are strongly correlated with the charging load after being screened by Pearson correlation coefficient, and the output parameters are the charging load data, ensuring that the model focuses on the mapping relationship between the core influencing factors and the charging load.

[0045] Based on the above The key hyperparameters of the LSTM model are dynamically determined through a continuous mapping function, achieving dynamic adaptation of the hyperparameters. Among these, the number of hidden layer units in the LSTM prediction model... The calculation is performed using the following formula:

[0046] In the formula, This is the minimum threshold for the number of hidden layer units. This represents the maximum threshold for the number of hidden layer units. , where is the adjustment coefficient used to control the influence of the MLE index on the number of hidden layer units. Dropout rate of LSTM prediction models Calculated using the following formula:

[0047] In the formula, This represents the minimum boundary value for the Dropout rate. This represents the maximum boundary value of the Dropout rate. The adaptation coefficients are used to ensure the model's generalization ability under complex sequences.

[0048] Learning rate of LSTM prediction model Calculated using the following formula:

[0049] In the formula, The initial learning rate, The parameters are adjusted to balance the convergence speed and stability of model training.

[0050] Through a dynamic mapping mechanism between the MLE exponent and hyperparameters, the LSTM model can accurately adapt to the dynamic characteristics of the charging load sequence: when When the load sequence is greater than 0, it exhibits strong chaotic characteristics, and the system is sensitive to disturbances. The model automatically increases the number of hidden layer units to capture complex temporal dependencies, adjusts the Dropout rate to suppress overfitting, and optimizes the learning rate to ensure training convergence. When the value is ≤0, the sequence tends to be stable or periodic, and the model simplifies hyperparameter configuration accordingly, avoiding resource redundancy and overfitting risks. Simultaneously, the model input focuses on strongly correlated influencing factors, and training is conducted using massive amounts of augmented data generated by Time Series Generative Adversarial Networks (TimeGAN). This effectively overcomes the bottleneck of traditional models relying on limited historical data and struggling to cope with uncertainty, significantly improving the accuracy and stability of charging load prediction. The constructed prediction model can provide reliable support for predicting the charging load of proposed facilities, making subsequent charging facility site selection and capacity determination more aligned with actual needs. This enhances the scientific rigor, foresight, and feasibility of planning schemes, effectively balancing user charging demand, equipment utilization, and grid safety operation.

[0051] The input and output sequences are divided into training, testing, and validation sets. The model parameters are optimized using the training set, the prediction accuracy is verified using the testing set, and the model performance is fine-tuned using the validation set. Finally, the charging load prediction model for electric vehicle charging facilities is completed.

[0052] A second aspect of the present invention discloses a charging facility planning method based on the prediction method, characterized in that the planning method includes: Based on experience or at fixed distances, the coordinates of multiple charging facilities to be built are generated in advance, and a strong correlation sequence related to the electric vehicle charging load is obtained for each charging facility. The strongly correlated sequence of each charging facility is input into the electric vehicle charging facility charging load prediction model to obtain the predicted charging load data. Based on load forecast data, indicators including average load, maximum load, minimum load, and charging facility utilization rate are calculated. If the calculated results of the corresponding indicators are lower than the preset threshold, the m-th charging facility to be constructed is removed; otherwise, it is retained.

[0053] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multivariate improved prediction method based on charging load, characterized in that, The prediction method includes the following steps: Obtain historical charging data and associated sequences for electric vehicle charging facilities; By using a time-series generative adversarial network, historical charging data is constructed into an augmented data sequence, and the MLE exponent of the augmented data sequence is calculated. K strongly correlated sequences related to electric vehicle charging load were selected from the associated sequences; An LSTM network is constructed, and the hyperparameters of the LSTM network are adjusted according to the MLE exponent to obtain a charging load prediction model for electric vehicle charging facilities. K strongly correlated sequences are input into the electric vehicle charging facility charging load prediction model to obtain the electric vehicle charging facility charging load prediction results.

2. The multivariate improved prediction method based on charging load according to claim 1, characterized in that, The associated sequence includes a temperature sequence. ,humidity Rainfall Wind speed Solar irradiance Date sequence data of historical charging load data sampling points for charging facilities at corresponding times. And population sequence data within a 10km radius of the i-th charging facility. Serial data of per capita GDP of the area to be planned .

3. The multivariate improved prediction method based on charging load according to claim 1, characterized in that, Using a time-series generative adversarial network, historical charging data is constructed into an enhanced data sequence, specifically including: The historical charging data is dimensionality reduced by using the embedding function in the time series generative adversarial network to obtain several first feature values, and the static and temporal features of the historical charging data are extracted simultaneously. The reconstruction function in the time series generative adversarial network is used to reconstruct the dimensionality-reduced historical charging data, obtain the reconstruction sequence, calculate several second feature values ​​corresponding to the reconstruction sequence, calculate the reconstruction loss of the reconstruction sequence, and if the reconstruction loss exceeds a set threshold, the embedding and reconstruction process is repeated. New sequence data is generated based on the static and temporal features. The generated sequence data is then encoded together with the features extracted by the embedding function and input into the discriminator. The generated sequence data is then judged to be true or false in conjunction with the historical charging data. The generated data that is judged to be true is retained. The generation loss is calculated based on the distribution consistency between the generated data and the historical charging data. If the generation loss exceeds a set threshold, the embedding and generation process is repeated. The reconstruction loss and generation loss are iterated repeatedly until both meet the set threshold, and finally an enhanced data sequence consistent with the feature patterns and distribution characteristics of the historical charging data is obtained.

4. The multivariate improved prediction method based on charging load according to claim 3, characterized in that, Before calculating the MLE exponent of the augmented data sequence, phase space reconstruction is performed on the augmented data sequence, which specifically includes: The maximum time delay search range is set based on the length of the augmented data sequence; By iterating through each time delay within the search range, the enhanced data sequence is reconstructed into a first sequence and a second sequence; Calculate the marginal probabilities of the first sequence and the second sequence respectively, and calculate the joint probability of the first sequence and the second sequence; Based on marginal probabilities and joint probabilities, the average mutual information function for calculating the enhanced data sequence is obtained. Based on the calculation results of the average mutual information function, the time delay corresponding to the first occurrence of a local minimum is taken as the final time delay; Using the determined final time delay, construct an m-dimensional spatial trajectory and an m+1-dimensional spatial trajectory; Calculate the nearest neighbor of each m-dimensional trajectory point, and calculate the distance mutation ratio of the nearest neighbor in the m+1 dimension. Determine the optimal embedding dimension based on the threshold of the distance mutation ratio. Based on the optimal embedding dimension and the final time delay, the enhanced data sequence is reconstructed into a high-dimensional trajectory vector.

5. The multivariate improved prediction method based on charging load according to claim 14, characterized in that, The Wolf method is used to calculate the MLE exponent of high-dimensional trajectory vectors, and the chaotic characteristics are determined based on the MLE exponent.

6. The multivariate improved prediction method based on charging load according to claim 5, characterized in that, The chaotic characteristics of load evolution are determined based on the MLE index, specifically: when the MLE index > 0, it indicates that the load evolution has chaotic characteristics, is sensitive to future disturbances, and is difficult to predict in the long term; when the MLE index < 0 or the MLE index = 0, the load change tends to be stable or periodic, and the prediction uncertainty is low.

7. The multivariate improved prediction method based on charging load according to claim 6, characterized in that, The number of hidden layer units, Dropout rate, and learning rate of the LSTM network are determined based on the MLE index.

8. The multivariate improved prediction method based on charging load according to claim 4, characterized in that, From the association sequences, K strongly correlated sequences related to electric vehicle charging load were selected, specifically including: calculating the augmented data sequence and the temperature sequence respectively. ,humidity Rainfall Wind speed Solar irradiance Date sequence data of historical charging load data sampling points for charging facilities at corresponding times. And population sequence data within a 10km radius of the i-th charging facility. Serial data of per capita GDP of the area to be planned The Pearson correlation coefficient between them was used to select k strongly correlated sequences that are strongly correlated with electric vehicle charging load. These k strongly correlated sequences were then included in the association sequence.

9. A method for planning charging facilities based on the prediction method described in any one of claims 1-8, characterized in that, The planning method includes: The coordinates of multiple charging facilities to be constructed are generated in advance, and the strong correlation sequence between each charging facility and the electric vehicle charging load is obtained. The strongly correlated sequence of each charging facility is input into the electric vehicle charging facility charging load prediction model to obtain the predicted charging load data. Based on load forecast data, indicators including average load, maximum load, minimum load, and charging facility utilization rate are calculated. If the calculated results of the corresponding indicators are lower than the preset threshold, the m-th charging facility to be constructed is removed; otherwise, it is retained.

Images