Electric vehicle charging load prediction method based on optimized fuzzy neural network

Abstract

The invention discloses an electric vehicle charging load prediction method based on an optimized fuzzy neural network, and the method comprises the steps: obtaining a plurality of historical load data of an electric vehicle, and processing all historical load data to generate a multi-dimensional feature data set; dividing an input feature space for the multi-dimensional feature data set based on an improved fuzzy clustering algorithm to obtain a clustering result and an activation intensity-contribution degree two-dimensional evaluation index; iteratively optimizing FNN parameters based on an improved adaptive differential evolution algorithm; constructing a rule contribution degree analysis model based on an orthogonal experimental design, setting a quantitative truncation threshold with the cumulative interpretation degree greater than or equal to 85%, and realizing self-adaptive compression of the scale of the rule base; an activation intensity threshold value dynamic calculation method is used, low-efficiency rules are automatically identified through sliding window statistics, and rule pruning is achieved; and outputting a load prediction result, and applying non-negativity and peak constraint to a prediction value.

Description

Technical Field

[0001] This invention relates to the field of load forecasting technology, and in particular to a method for forecasting electric vehicle charging load based on an optimized fuzzy neural network. Background Technology

[0002] With the global energy structure transformation, electric vehicles (EVs), as an important carrier of clean energy transportation, have seen their market penetration rate continue to rise. Along with the increase in the number of EVs, the demand for charging infrastructure has experienced explosive growth. However, the randomness and spatiotemporal uncertainty of large-scale EV charging loads pose significant challenges to grid dispatching, charging station planning, and operation.

[0003] Traditional prediction methods (such as Monte Carlo simulation) struggle to effectively integrate multi-dimensional features such as charging behavior data (charging duration, power curve), environmental variables (holidays), and user profiles (travel patterns), resulting in limited prediction accuracy. Statistical models (such as ARIMA) cannot capture abrupt changes in charging load; machine learning models (such as SVM and random forest) rely on manual feature engineering and lack generalization ability; while deep learning methods (such as LSTM) can extract temporal dependencies, they suffer from poor interpretability and high training costs.

[0004] Traditional optimization algorithms (such as standard PSO and genetic algorithms) are prone to getting trapped in local optima and have slow convergence speeds when optimizing fuzzy neural network parameters. For example, the fixed inertia weights of the PSO algorithm struggle to balance global exploration with local exploitation; the fixed scaling factor (F) and crossover probability (CR) of the Differential Evolutionary Algorithm (DE) cannot adapt to the needs of different evolutionary stages. In fuzzy neural networks (FNNs), the rule library generated by traditional grid partitioning methods suffers from combinatorial explosion (the number of rules grows exponentially with the feature dimension), leading to low computational efficiency. Existing pruning methods rely on manually set thresholds and lack adaptive mechanisms. Summary of the Invention

[0005] Technical objective: To address the shortcomings of existing technologies, this invention discloses a method for predicting electric vehicle charging load based on an optimized fuzzy neural network.

[0006] Technical solution: To achieve the above technical objectives, the present invention adopts the following technical solution.

[0007] A method for predicting electric vehicle charging load based on an optimized fuzzy neural network includes the following steps:

[0008] Step A: Obtain several historical load data of electric vehicles, and generate a multidimensional feature data set after processing all historical load data;

[0009] Step B: Divide the input feature space of the multidimensional feature data set based on the improved fuzzy clustering algorithm to obtain the clustering results and the dual-dimensional evaluation index of activation intensity and contribution; the Gaussian membership function parameters include the center and width of the Gaussian membership function; the improved fuzzy clustering algorithm introduces a multimodal distance dynamic trade-off between the statistical correlation and distribution differences of the features by fusing Mahalanobis distance and Wasserstein distance;

[0010] Step C: Use the multidimensional feature dataset as input samples, and iteratively optimize the parameters of the fuzzy neural network based on the improved adaptive differential evolution algorithm; the parameters of the fuzzy neural network include Gaussian membership function parameters and regular weights;

[0011] Step D: Further optimize the parameters of the optimized fuzzy neural network obtained in Step C through rule pruning and dynamic fusion of external factors to obtain the optimized fuzzy neural network; calculate the activation intensity and contribution of the rules based on orthogonal experimental design, and then realize rule pruning, that is, perform rule pruning on the initial rule base generated in Step B; use the multidimensional feature data set obtained in Step A to construct the temperature-holiday attention weight matrix to realize the dynamic fusion of external factors.

[0012] Step E: Use the optimized fuzzy neural network to predict the historical load data of the electric vehicle to be predicted, output the load prediction result, and apply non-negativity and peak constraints to the predicted value.

[0013] Beneficial effects:

[0014] 1. This invention discloses an electric vehicle charging load prediction method based on an optimized fuzzy neural network, which can efficiently handle the nonlinearity, time-varying nature, and uncertainty of charging load. By dynamically optimizing the membership function parameters and rule weights of the fuzzy neural network (FNN), the prediction accuracy is significantly improved, effectively capturing the complex correlation between charging behavior and external factors (such as holidays), and more accurately reflecting the distribution characteristics of actual charging demand.

[0015] 2. This invention addresses the limitations of traditional optimization algorithms by proposing a dynamic parameter adaptive mechanism to solve the problem of PSO, DE, and other algorithms easily getting trapped in local optima. By introducing an elite archiving strategy and hybrid mutation operations, an efficient balance is achieved between global exploration and local exploitation, enabling the prediction model to maintain stable convergence even in complex scenarios.

[0016] 3. This invention innovatively designs a data-driven rule pruning mechanism, which selects high-contribution rules (cumulative contribution ≥ 85%) through orthogonal experiments and automatically removes redundant rules by combining activation intensity thresholds, thereby compressing the rule base size, significantly reducing computational complexity, and ensuring model interpretability.

[0017] 4. This invention is applicable to modeling of various charging scenarios, especially constant current-constant voltage (CC-CV) charging curve fitting in fast / slow charging modes, which can more effectively predict the actual situation of charging stations in the area. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method of the present invention;

[0019] Figure 2 Flowchart of the improved JADE algorithm;

[0020] Figure 3 This is a schematic diagram of fuzzy rule generation and pruning;

[0021] Figure 4 A performance comparison chart of charging quantity prediction results with traditional algorithms on the training and test sets;

[0022] Figure 5 A scatter plot comparing the charging demand prediction results with those of traditional algorithms on the test set;

[0023] Figure 6 A graph comparing the seven-day charging demand predicted by the FNN optimized for JADE with the actual value.

[0024] Figure 7 A bar chart comparing the charging demand at different stations on the same day with the actual values, to improve JADE's optimized FNN predictions. Detailed Implementation

[0025] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0026] Example

[0027] As attached Figure 1 To be continued Figure 3 As shown in the figure, an electric vehicle charging load prediction method based on an optimized fuzzy neural network in this embodiment includes the following steps:

[0028] Step A: Obtain several historical load data of electric vehicles (hereinafter referred to as EVs), and generate a multi-dimensional feature data set after processing all historical load data;

[0029] Step A specifically includes the following steps:

[0030] Step A1: Obtain several historical load data of EVs, and use dual timestamp verification and anomaly cleaning to remove outliers from the historical load data. After outlier removal, use STL (Seasonal-Trend decomposition procedure based on Loess) to decouple multi-scale time series features from the historical load data to obtain a time series with multiple features. Among them, dual timestamp verification refers to logical verification of charging start and end time and duration threshold constraints. The time series decomposition includes the extraction of daily, weekly and monthly periodic components, and the fusion of external features such as temperature and holidays.

[0031] This embodiment uses data from several charging stations within a certain area as samples, namely historical load data. Each historical load data point includes eleven features: charging start time, charging end time, charging pile type, voltage, current, charging amount, temperature, whether it is a holiday, whether it is a weekend, location, and charging pile number. A dual timestamp verification mechanism is used to preprocess the historical load data, removing abnormal records such as abnormal times, voltages, currents, and charging amounts.

[0032] The calculation formula for the dual timestamp verification mechanism is as follows:

[0033]

[0034] In the formula, T min and T max These represent the minimum and maximum allowed time boundaries in historical load data, t. start t is the charging start time. end The charging end time is specified. If any condition in the calculation formula of the dual timestamp verification mechanism is not met, the load data is determined to be an abnormal record and removed. After removing abnormal records, the charging duration is calculated using the charging start time and charging end time; the STL decomposition algorithm is used to decouple multi-scale time-series features.

[0035] Step A2: Encode the external features in the time series to generate time series data with composite features as samples;

[0036] The model dynamically embeds and encodes categorical variables for external factors in time series. It constructs a dynamic encoding model for three external variables: temperature, whether it is a holiday, and whether it is a weekend. Using the TemporalFusionTransformers (TFT) feature cross-module, it calculates the temperature-holiday combination effect through attention weights, generating physically interpretable composite features (such as "high temperature + weekday" and "low temperature + holiday") to enhance the model's ability to characterize complex scenarios. Furthermore, it performs numerical conversion on charging pile types (AC slow charging, DC fast charging, and super-fast charging) (corresponding to 1, 2, and 3 respectively) to generate charging type features.

[0037] In this example, the characteristics such as charging pile type, temperature, whether it is a holiday, whether it is a weekend, etc. are numerically converted to generate relevant charging type characteristics, where the charging pile type includes: AC slow charging, DC fast charging, and super fast charging.

[0038] Step A3: Perform dynamic normalization on the samples and construct a multidimensional feature matrix to generate a feature matrix with expandable dimensions. Each value in the matrix is ​​in the range of [0, 1], which serves as a multidimensional feature data set.

[0039] The data in the sample are normalized using a min-max function according to the parameters of the training set, and the output variables retain the inverse normalization parameters; the time series decomposition components, external factor encoding, and original statistical features are vertically concatenated to generate a feature matrix with expandable dimensions.

[0040] The normalization process in step A3 is as follows:

[0041] For the input sample data, i.e., X∈R N×D The feature transformation is performed column-wise, i.e., normalization is performed. The calculation formula is as follows:

[0042]

[0043] In the formula, N is the number of samples, and D is the feature dimension of the samples. and Let x be the minimum and maximum values ​​of the j-th feature in the sample, corresponding to the upper and lower bounds U and L in subsequent steps. ij Let j be the value of the j-th dimension feature of the i-th data in the sample. The sample is normalized; normalization result

[0044] The inverse normalization process refers to: normalizing the result X... norm The original dimensions can be recovered through inverse transformation when predicting subsequent outputs. The formula for inverse normalization is:

[0045]

[0046] In the formula, and It needs to be obtained from the normalized parameters saved during the training phase. This is the result of inverse normalization.

[0047] Step B: Divide the input feature space of the multidimensional feature data set using an improved fuzzy clustering algorithm to obtain the clustering results and a dual-dimensional evaluation index of activation strength and contribution. The Gaussian membership function parameters include the center and width of the Gaussian membership function. The improved fuzzy clustering algorithm introduces a multimodal distance that fuses Mahalanobis distance and Wasserstein distance to dynamically weigh the statistical correlation and distribution differences between features. The contribution corresponds to the multimodal distance fusion weight λ.

[0048] Specifically, the input feature space is partitioned based on improved fuzzy clustering (FCM). This improved fuzzy clustering incorporates a multimodal distance metric fusion (combining Mahalanobis distance and Wasserstein distance, i.e., Mahalanobis-Wasserstein distance) to replace traditional Euclidean distance, dynamically balancing statistical correlation and distributional differences between features to improve the robustness of clustering high-dimensional nonlinear data. The covariance matrix dynamically corrects for scale differences and correlation interference between features. When generating Gaussian membership functions based on cluster centers, the inter-class spacing adaptively adjusts the standard deviation parameter, effectively improving the fitting accuracy of fuzzy intervals to the data distribution. After generating an initial rule base through the Cartesian product of the feature space, a dual-dimensional evaluation index of activation strength and contribution is constructed to provide a quantitative basis for subsequent rule pruning.

[0049] Step B, the clustering process based on the improved fuzzy clustering algorithm, specifically includes:

[0050] Step 1: Perform feature correlation analysis on the multidimensional feature dataset and determine the multimodal distance fusion weights; calculate the linear correlation coefficient between features using the covariance matrix, where the Mahalanobis distance takes into account the covariance matrix and is based on the mean of the Pearson correlation coefficient. Compared with the historical mean μ ρ The multimodal distance fusion weight λ is dynamically determined using the following formula:

[0051]

[0052] In the formula, β λ λ is the weight sensitivity coefficient, and the multimodal distance fusion weight is the contribution. In other words, the contribution of Mahalanobis distance and Wasserstein distance is dynamically weighed through λ.

[0053] Step B2: Initialize cluster centers for the multidimensional feature data set using K-means++; generate initial cluster centers for the multidimensional feature data set using the K-means++ algorithm, based on the multimodal distance D. multi (x i C prev Calculate the sample selection probability P(x) i The formula is:

[0054]

[0055] In the formula, D multi (x i c k )=λD mahalanobis (x i c k )+(1-λ)D sserstein (x i c k ), for sample x i To cluster center c k Multimodal distance, fusing Mahalanobis distance and Wasserstein distance. C prev ={c1, c2, ..., c k} represents the set of selected cluster centers, initially empty and gradually expanded during iteration. D mahalanobis (x i c k ) represents the Mahalanobis distance, D wasserstein (x i c k ) represents the Wasserstein distance.

[0056] During the initialization process, the first step is to randomly select the first center c. j The second step is to calculate the multimodal distance D from all samples to the nearest center. multi (x i C prev The third step is to calculate the probability P(x). i Select the next center, and repeat until K centers have been selected.

[0057] Step B3: Dynamically update the cluster centers and distance metric. During each cluster center update, calculate the feature covariance matrix S in real time. k Furthermore, it eliminates feature scale differences and multicollinearity interference through incremental updates, and dynamically adjusts the metric formula for multimodal distance.

[0058] The Mahalanobis distance formula is as follows:

[0059]

[0060] In the formula Sk -1 S is the inverse of the covariance matrix of the k-th cluster, and S is the covariance matrix of the t-th iteration. k The formula for calculating (t) is:

[0061]

[0062] In the formula, t is the number of iterations.

[0063] The Wasserstein distance is calculated using a Gaussian kernel approximation, as shown in the following formula:

[0064]

[0065] In the formula, σ w The Gaussian kernel bandwidth is adaptively set to the median of the sample distance, denoted as _median{||x_{\pi} ... i -c k ||2|x i ∈C k}, p k (y), p i (y) represents the cluster center c. k With sample x i The probability density function is generated through kernel density estimation.

[0066] Specifically, first calculate the fuzzy membership degree, i.e., the sample x i Membership degree u of cluster k ik The calculation formula is:

[0067]

[0068] In the formula, m is the fuzziness index, which controls the degree of fuzziness in the membership distribution.

[0069] Then iteratively update the cluster centers. After each update, recalculate the covariance matrix S. k With Gaussian kernel bandwidth σ w Dynamically adjust the multimodal distance metric. Update cluster centers c using weighted updates. k The formula is:

[0070]

[0071] In the formula, For sample x i The membership degree of cluster k is affected by the fuzzy exponent m, where 1 ≤ i ≤ n; n is the total number of samples.

[0072] Step B4: Dynamically generate Gaussian membership function parameters; calculate the multimodal distance density ρ of intra-cluster samples based on clustering results. k Dynamically adjust the standard deviation parameter σ of the Gaussian function ikFor high-density regions, i.e., ρ k <0.5ρ max Using a smaller standard deviation η is the scaling factor, ρ max This represents the maximum density across all clusters. For sparse regions, the standard deviation range is widened to avoid overfitting. This widening is achieved by adding noise, specifically a small noise fluctuation to the samples. Finally, the Gaussian membership function is calculated.

[0073] Wherein, the modal distance density ρ of cluster k k The formula is:

[0074]

[0075] In the formula, n k c is the number of samples in cluster k. k ρ is the cluster center. k It reflects the degree of clustering of samples under multimodal distance.

[0076] The formula for calculating the Gaussian membership function of the i-th feature of the k-th cluster is:

[0077]

[0078] In the formula, c ik Let c be the center value of cluster k in the i-th dimension, which can be directly inherited from the cluster center c. k The center and width of the Gaussian membership function are obtained based on the Gaussian membership function.

[0079] Step B5: Use the Cartesian product of the feature space to generate a rule combination matrix as the initial rule base.

[0080] If the input features contain d dimensions, and each dimension generates m fuzzy sets, then m fuzzy sets are generated through Cartesian product. d The rule has the following form:

[0081] If x1 is A 1k and x2 is A 2k and…then y=ω 0k +ω 1k x1+…+ω dk x d

[0082] In the formula, A ik Let k be the fuzzy set of the i-th feature, defined by the Gaussian membership function μ. ik (x) is defined. ω 0k ω lk …ω dkThe linear parameters are regular consequents, which are solved by minimizing the error between the model's predicted output and the true value.

[0083] Step B6: Calculate the activation intensity according to the Gaussian membership function product rule, and complete the dual-dimensional evaluation index of activation intensity and contribution.

[0084] The activation intensity α of input sample x for the j-th rule j The product of the membership degrees of each dimension:

[0085]

[0086] In the formula, μ ij (x i Let be the membership degree of the i-th feature to the j-th rule fuzzy set, and let its parameter c be... ik σ ik The value is dynamically determined in step B4. In this example, there are 11 input features, and each feature generates 3 Gaussian membership functions. The contribution value corresponds to the multimodal distance fusion weight λ; the activation intensity α j The larger the value of λ, the more the current input "triggers" the rule, and the stronger the rule's "involvement" in the current prediction; it can filter out rules that respond highly to input data and identify which rules are active in the current scenario. The larger the contribution value λ, the greater the "actual contribution" of the rule to the final prediction result; it can distinguish between rules that are "highly activated but have little impact on the output" and rules that "play a key role in the prediction result".

[0087] In step B, the input feature space is divided using improved fuzzy clustering (FCM) on the multidimensional feature data set. The statistical correlation and distribution differences between features are analyzed by multimodal distance metric (Massaurus-Wasserstein distance), providing original feature support for generating Gaussian membership functions and initial rule base.

[0088] Step C: Using the multidimensional feature dataset as input samples, iteratively optimize the parameters of the fuzzy neural network (FNN) based on the improved adaptive differential evolution algorithm (JADE). The parameters of the fuzzy neural network include Gaussian membership function parameters and rule weights. The rule weights are the rule weights of the initial rule base and are also used as optimization parameters in step C. Using the test set RMSE as the fitness function, iteratively optimize the membership function center, width, and rule weights in the FNN parameters. In step C, the Gaussian membership function parameters obtained in step B are used as the initial parameters of the FNN, and iterative optimization is performed using the improved adaptive differential evolution algorithm to improve the fitting accuracy of the fuzzy interval to the data distribution.

[0089] Step C, the process of iteratively optimizing FNN parameters based on improved JADE, specifically includes:

[0090] Step C1: Using the multidimensional feature dataset as input samples, perform a Logistic-Chebyshev-Sobol hybrid initialization operation on the parameters of the fuzzy neural network. A Logistic-Chebyshev-Sobol hybrid initialization mechanism is designed, where 50% of individuals are generated through Logistic-Chebyshev hybrid chaos, 30% through Sobol sequences, and 20% through Gaussian perturbation fine-tuning of existing individuals. This combination of Logistic-Chebyshev hybrid chaos generation and Sobol sequence generation enhances population diversity. The initial population individuals generated by the hybrid initialization contain the FNN parameters.

[0091] 50% of the individuals are generated using Logistic-Chebyshev hybrid chaos, calculated as follows:

[0092]

[0093] In the formula, the fusion weight λ1∈[0,1], here taken as 0.5, to balance the contributions of the two chaotic mappings. Logistic mapping: x n+1 =4x n ·(1-x n ), x n ∈(0,1),x n Let x0 be the chaotic value in the nth iteration, with the initial value randomly generated within the range (0.1, 0.9). Chebyshev mapping: y n+1 =cos(k·arccos(y) n )), y n ∈(-1, 1), k = 3, y n Let y0 be the chaotic value in the nth iteration, and let y0 be the initial value randomly generated within the range (-0.9, 0.9). The order parameter is fixed at 3, which determines the degree of chaos.

[0094] 30% of individuals are generated using Sobol sequences, and low-discrepancy sequences are generated using Gray code transformation to ensure a uniform distribution of samples in the parameter space. The formula is as follows:

[0095] S i =[s i，1 s i，2 , ..., s i，q ],i=1,2,…,N·30%

[0096] In the formula, q represents the number of optimization parameters, i.e., the problem dimension, and N represents the population size.

[0097] By fine-tuning the existing individuals using Gaussian perturbation, 20% of the individuals can be obtained. The formula is as follows:

[0098]

[0099] In the formula, x i Randomly selected from the first 80% of generated individuals (chaotic or Sobol). σ1 is the perturbation intensity (controls the fine-tuning amplitude, default is 0.05 times the parameter range), U and L are the upper and lower bounds of the parameters, i.e., inherited from the boundary definitions in step A3;

[0100] The multimodal distance metric in step B, with its dynamic adjustment logic, provides prior information on the feature space distribution of the diversity (Logistic-Chebyshev-Sobol mixture mechanism) and mutation strategy (ABL-OCL-JADE) of the population initialization in step C based on covariance matrix and kernel density estimation, thereby indirectly improving the search efficiency of the optimization algorithm.

[0101] Step C2: Construction of Dynamic Archive Management Mechanism. An empty archive set is created, with a maximum capacity of 20% of the population size. A First-In-First-Out (FIFO) elimination strategy is used to keep the archive dynamically updated. The dynamic archive management mechanism incorporates historical high-quality solutions into the mutation operation; the specific process is as follows:

[0102] Create an empty archive collection A t Each generation retains the top 10% of elite solutions in the archive, employing a first-in, first-out (FIFO) elimination strategy to keep the archive dynamically updated. First, implement the archive update rules, saving historical high-quality solutions to guide mutation. The archive update rule formula is:

[0103] A t =A t-1 ∪{x i t-1 |f(x i t-1 ) <f(x i t )},|A t |≤0.2N

[0104] In the formula, N is the archive capacity, i.e., the population size, which is taken as 20% of the population size as the archive capacity for high-quality solutions. f(x) is the fitness function, i.e., the RMSE of the test set. A t-1 For the previous archive set, x i t-1 For individuals in the previous archive set;

[0105] Then implement the archive sampling strategy. Archives participate in mutation, and during the mutation operation, they are sampled probabilistically. Select a basis vector from the archive; the calculation formula is:

[0106]

[0107] In the formula, To save from archive A t The basis vectors selected from the data to participate in the mutation.

[0108] Step C3: Dynamically adjust the scaling factor F and the crossover probability CR. The scaling factor F adopts Lehmer mean adaptation, and the crossover probability CR adopts Beta distribution adaptation. A Lehmer mean update strategy for the scaling factor F and a Beta distribution adaptation mechanism for the crossover probability CR are proposed to achieve global-local balance. The dynamic adjustment of the scaling factor F and the crossover probability CR directly affects the search accuracy of the FNN parameters.

[0109] The scaling factor F is updated using the Lehmer mean, with the initial scaling factor F. 0 =0.5, Lehmer's mean formula:

[0110]

[0111] In the formula, N s To determine the number of successful solutions (solutions with better fitness than their parents), F i t-1 The scaling factor F used in the previous generation of successful solutions t The scaling factor used for the current generation of successful solutions; The RMSE of the test set for the i-th solution is consistent with the error index of the activation strength calculation in step B6; RMSE t-1 This represents the average RMSE of the previous generation. The formula for RMSE is:

[0112]

[0113] In the formula, N test y is the number of samples in the test set. i Let y be the actual value of the i-th sample. i Let be the predicted value for the i-th sample.

[0114] The crossover probability CR adopts an adaptive Beta distribution, and the Beta distribution parameter update formula is as follows:

[0115]

[0116] In the formula, Δ i = 1 (if solving for i is successful) or 0 (otherwise), The CR value is taken from the previous generation individual. t b t Here are the shape parameters of the Beta distribution, all initially set to 2. The CR sampling formula is:

[0117] Step C4: A hybrid mutation strategy based on boundary learning, oppositional collaborative learning, and adaptive differential evolution (ABL-OCL-JADE) is employed to optimize the FNN parameters through collaborative optimization and elite preservation, improving the rule's adaptability to load prediction. The ABL-OCL-JADE hybrid mutation strategy performs mutation operations on individuals containing FNN parameters, exploring better values ​​in the parameter space. The mutation operations include ABL mutation (adaptive boundary learning), AOCL mutation (adaptive oppositional center learning), and hybrid ADE mutation, with the mutated individual v... i The allocation formula is:

[0118]

[0119] In the formula, P ABL P represents the proportion of individuals with ABL variants. AOCL The proportion of AOCL variant individuals; The ADE mutant individuals (offspring) generated for the i-th individual correspond to the global exploration current-to-pbest / 1 and the local development rand-to-best / 2 in the hybrid mutation strategy in the following steps; the ADE hybrid mutation strategy includes two parts: global search and local development, and its individuals correspond to the individuals generated by these two parts.

[0120] ABL Adaptive Boundary Learning Mutation Formula:

[0121]

[0122] In the formula, For the ABL variant individuals (offspring) generated from the i-th individual, α ABL =α0(1-t / T) β For the adaptive learning rate, we take α0 = 0.8, β = 1.2, and x i For the i-th individual (parent generation), x pbest pbest represents the top 10% of high-quality solutions in the archive. The initial elite ratio is 0.2. The iteration progress ratio is given by T, where T is the total number of iterations, and pbest is randomly selected. The optimal individual. r1 is a random number in [0, 1], used to randomly select the direction of boundary learning (upper or lower boundary). U and L are the upper and lower bounds of the parameters, inherited from U and L in step A, and f() is the fitness function.

[0123] The AOCL adaptive contrast center learning mutation formula is:

[0124]

[0125] In the formula, For the AOCL variant generated for the i-th individual, the opposing learning weight λ O =0.3 + 0.7exp(-5t / T), decreasing with iteration. This is an adaptive balance factor, based on the distance between the current solution and the optimal / worst solution. (Opposite solution) Random perturbation standard deviation σ ε =0.01 (UL). Central solution M is set to 20, meaning the top 20% of high-quality solutions are used for calculation.

[0126] In the hybrid mutation strategy, the global exploration current-to-pbest / 1 is calculated using the following formula:

[0127]

[0128] In the formula, For the globally explored mutant individual generated for the i-th individual, F i x is the scaling factor for the i-th individual. i Let x be the i-th individual in the current population. pbest For elite individuals, x r1 x r2 Two different individuals were randomly selected. CR Given the mean CR of the current generation, the tendency for high CR is: when μ CR →1, The Beta distribution generates large CR values ​​(close to 1), promoting the generation of more dimensions from variant individuals. Enhance global search capabilities.

[0129] The formula for local exploitation rand-to-best / 2 in the hybrid mutation strategy is:

[0130]

[0131] In the formula, The locally exploratory mutant individual generated by the i-th individual, x best The current globally optimal individual, x r1 x r2 x r3 Three distinct individuals were randomly selected. Low CR propensity: when μ CR →0, the Beta distribution generates small CR values ​​(close to 0), which leads to more dimensions being retained from the parent individual x. i Then, we can refine and develop the current optimal region.

[0132] Step C5: Binomial crossover operation to enhance population diversity and preserve superior genes from the parent generation. The binomial crossover operation fuses the parent generation with the variant individuals, preserving superior genes.

[0133] The binomial cross formula is:

[0134]

[0135] In the formula, u i，d For the new offspring after crossover, v i，d For the mutant individuals generated in the above steps, x i，d For the parent individuals, CR is the crossover probability, which is the adaptive Beta distribution of step C, d rand The forced cross dimension is randomly selected to ensure that at least one dimension comes from the mutation vector. i represents the individual, and d represents the dimension.

[0136] Step C6: Selection operation, using greedy selection. The greedy selection operation uses the RMSE of the test set as the fitness function, selecting the better individuals to enter the next generation.

[0137] The formula for greedy selection is:

[0138]

[0139] In the formula, f(x) i The fitness function is calculated using the RMSE formula from the test set, and the data is archived and updated. The winning individual is stored in dynamic archive A t .

[0140] Step C7: Update the global optimal solution based on the RMSE iteration results of the fitness function, set an early stopping mechanism, terminate the evolution when the fitness improvement is less than 1e-4 for 15 consecutive generations, avoid invalid iterations, and output the FNN parameters after iterative optimization.

[0141] In the mutation phase, Adaptive Boundary Learning (ABL) and Adaptive Opposites Center Learning (AOCL) are introduced, combined with a hybrid mutation strategy of current-to-pbest / 1 and rand-to-best / 2, forming the ABL-OCL-JADE algorithm. After the crossover operation, a greedy selection is performed, and the FNN parameters are iteratively optimized using the RMSE of the test set as the fitness function, thereby improving the convergence speed and global robustness.

[0142] Step D: Further optimize the parameters of the optimized fuzzy neural network obtained in Step C through rule pruning and dynamic fusion of external factors to obtain the optimized fuzzy neural network; calculate the activation intensity and contribution of the rules based on orthogonal experimental design, and then realize rule pruning, that is, perform rule pruning on the initial rule base generated in Step B; use the multidimensional feature data set obtained in Step A to construct the temperature-holiday attention weight matrix to realize the dynamic fusion of external factors.

[0143] In this invention, a rule contribution analysis model is constructed based on orthogonal experimental design. Redundant rules are eliminated based on a "dual-dimensional evaluation index of activation intensity and contribution," meaning that rules with a cumulative contribution of ≥85% and an activation intensity of ≥1e-6 are retained, achieving adaptive compression of the rule base size and forming a streamlined rule base. Furthermore, a dynamic calculation method for activation intensity thresholds is used, and inefficient rules are automatically identified through sliding window statistics, simultaneously implementing rule pruning and overcoming the shortcomings of traditional pruning methods that rely on manual experience. A learnable temperature-holiday attention weight matrix is ​​introduced into the external factor fusion layer, and the influence of features is dynamically adjusted through a gating mechanism, enhancing the model's responsiveness to sudden events and its adaptability to complex scenarios.

[0144] Step D, based on the FNN parameters output in Step C, further optimizes the structure and performance of the FNN through rule pruning and dynamic fusion of external factors, reducing computational complexity and improving model robustness. Step C iteratively optimizes the FNN parameters using an improved adaptive differential evolution algorithm (ABL-OCL-JADE), outputting parameters including Gaussian membership function centers, widths, and rule weights. Step D then further simplifies and optimizes the FNN's rule base based on these optimized parameters.

[0145] Step D's rule pruning uses a two-step screening method. The process of step D includes:

[0146] Step D1: Orthogonal experiment-driven rule contribution modeling. A three-factor, three-level orthogonal experimental matrix is ​​constructed, using rule activation strength, consequent weight, and input feature importance as experimental factors. Analysis of variance is used to quantify the independent contribution of each rule to the output variable. Redundant rules are removed from the initial rule base based on a dual-dimensional evaluation index of activation strength and contribution. Rules with a cumulative contribution ≥ 85% and an activation strength ≥ 1e-6 are retained, achieving adaptive compression of the rule base size and forming a streamlined rule base, thus realizing rule pruning.

[0147] The importance of input features is indirectly reflected by activation intensity; higher feature importance results in higher activation intensity. The design rule energy index (REI) is calculated using the following formula:

[0148]

[0149] In the formula, REI r Let α be the rule energy index of the r-th rule, where the range of rule h in the denominator is 1 ≤ h ≤ R. R is the size of the subset of rules involved in the calculation; here, R = K, where K is the total number of rules. r (x k Let y be the activation intensity of the r-th rule in the k-th sample. r (x k) is the consequent output of the r-th rule in the k-th sample, which is generated according to the rule combination matrix in step B5, ω r This represents the weight of the r-th rule, which is also the consequent weight. The subscript h corresponds to the same content as r. When the cumulative REI value reaches 85%, a rule truncation mechanism is triggered, automatically retaining the key rule set. The calculated rule contributions are sorted in descending order, while rules with activation strength less than 1e-6 are removed.

[0150] The calculation of output rule weights includes: the weighted contribution of each rule to the final output, calculated using the following formula:

[0151]

[0152] In the formula, α r Let θ be the activation strength of the r-th rule, which is also the product of Gaussian membership functions. k The parameter is the consequent of the rule, which belongs to the linear weights, and is the linear parameter ω of the consequent of the rule in step B5. 0k ω 1k …ω dk The activation strength of a rule provides a basis for rule pruning, i.e., activation strength threshold filtering.

[0153] The contribution of a rule is comprehensively evaluated and defined as the normalized activation-weight product, used to quantify the relative impact of the rule on the model output. The formula is as follows:

[0154]

[0155] In the formula, y r Let |ω be the output value of the r-th rule, i.e., the linear combination of consequents. r ·y r | represents the absolute value of the weighted output. The sum of contributions to all rules.

[0156] In this example, there are three approaches to weight allocation. The first is implicit weighting, which uses membership function parameters for allocation. The center and width of the Gaussian function determine the influence of features on the activation strength of rules. The second is explicit weighting, which uses consequent parameters for allocation. By training linear coefficients optimized by least squares, different output weights are directly assigned to features. The third is dynamic weighting, implemented through normalized weights, which dynamically adjusts the contribution ratio of each rule based on the input during inference.

[0157] Rule truncation settings: Cumulative contribution truncation, for those already truncated by ρ k The formula for sorting the rule set in descending order, retaining the first M rules, is:

[0158]

[0159] In the formula, ρ k To determine the contribution of the rule, it needs to be normalized in advance, so that ∑ρ k =1, and the threshold of 0.85 is an adjustable parameter.

[0160] Low activation strength rules are eliminated by further eliminating rules from the truncated rule set according to the following formula:

[0161] max(α k <10 -6 (Maximum activation strength across all training samples).

[0162] In the formula, α k Let k be the activation intensity vector of rule k across all samples, with a threshold of 10. -6 This is an engineering criterion for near-zero activation.

[0163] Step D2: Using the multi-dimensional feature data set obtained in Step A, construct a temperature-holiday attention weight matrix to achieve dynamic fusion of external factors; among which, the construction of the multi-factor attention weight matrix and the design of the temperature-holiday joint attention mechanism are carried out, and the formula is as follows:

[0164]

[0165] In the formula, σ s The Sigmoid activation function maps the input to the (0, 1) interval and is used to generate attention weights, determining the degree of attention given to different features. T is the temperature encoding vector, H is the holiday state matrix, and C is the regular features, i.e., the features other than T and H in the multidimensional feature dataset of step A. These are the transposes of the temperature encoding vector weight matrix, the holiday state matrix weight matrix, and the regular feature weight matrix, respectively, which perform linear transformations on the temperature feature T, holiday feature H, and regular feature C. ⊙ represents the element-wise multiplication mathematical expression. b is the bias term, used to adjust the output of the linear transformation and increase the model's fitting ability. tanh is the hyperbolic tangent function. Step D2, based on the multidimensional feature data set from step A, amplifies the strong influencing factors such as temperature and holidays in the original features through classification, dynamic weighting, and error-driven updates, while weakening the interference of regular features, thus improving the model's adaptability to complex scenarios.

[0166] Step D3: Introduce knowledge-guided supplementary factors into the fuzzy neural network to compensate for temperature and holiday features, as well as the output of the fuzzy neural network. For temperature and holiday features, the knowledge-guided supplementary factors address potential issues such as insufficient information mining and poor model adaptability in the fuzzy neural network's processing of these features, making the model's utilization of temperature and holiday features more accurate. For the output of the fuzzy neural network, domain knowledge is used for compensation, combining domain knowledge (such as knowledge of different scenarios and patterns in load forecasting) to correct and improve the model output.

[0167] A knowledge-guided compensation factor is injected into the fuzzy neural network, and a domain knowledge compensation module is embedded in the output layer. The formula is as follows:

[0168]

[0169] In the formula, y final This is the final output value, which is also the final value after knowledge compensation and correction, y pred Let f be the original predicted value of the model (output before compensation), f be the domain knowledge constraint function (a function that encodes prior knowledge of "historical peak patterns, holiday / extreme temperature effects"), and y be the original predicted value of the model. pred The function is an abbreviation, where β1 is the adaptive learning rate parameter. φ(t) is a compensation function based on historical peak patterns, which automatically activates the compensation channel when holidays / extreme temperatures are detected. The compensation function is designed as follows:

[0170] The first step is to define the compensation scenario. Scenario 1: Holidays (t holiday The encoding interval is [0, 1], t holiday =1, indicating a holiday); Scenario 2: Extreme temperature (t temp Set t temp <-10℃ or t temp >35°C is the extreme temperature, and the encoding range is [0, 1]. temp =1 represents the extreme temperature). The input to the compensation function φ(t) is the scene encoding vector: t = [t holiday , t temp , t hour ], t hour This represents the number of hours in a day, used to capture intraday peak patterns.

[0171] The second step is the dynamic compensation function formula (Gaussian + piecewise). A compensation function using double Gaussian superposition and piecewise activation is employed to simultaneously cover the effects of intraday peak temperatures during holidays and the ongoing impact of extreme temperatures.

[0172]

[0173] Among them, the holiday compensation sub-function φholiday (t) Use a Gaussian function to fit the daily load peak during holidays (10-12 am and 6-8 pm are the charging peaks).

[0174]

[0175] In the formula, A holiday The compensation range for holidays (average additional load based on historical statistics, in kW); t 01 t 02 Peak times during holidays; σ 01 σ 02 The peak diffusion coefficient (controls the attenuation range of the compensation amount over time; for example, σ = 2 indicates that there is significant compensation within 2 hours before and after the peak).

[0176] Extreme temperature compensation sub-function φ temp (t), which uses a linear-exponential mixture function to fit the persistent effect of extreme temperatures (the higher / lower the temperature, the greater the compensation, and it persists over time).

[0177] φ temp (t)=A temp ·(1+exp(-k·t hour ))·|T d |·t temp

[0178] In the formula, A temp The base amplitude for extreme temperature compensation (e.g., the basic compensation amount per °C temperature difference); k is the time decay coefficient (controlling the decay of the compensation amount over time (e.g., k = 0.1 indicates that the compensation amount gradually decreases after dusk)); T d This is the difference between the actual temperature and the comfortable temperature (taken as 26℃).

[0179] The multidimensional feature data set obtained in step A is used as the basic data for rule contribution analysis in step D. The activation intensity and contribution of rules are calculated through orthogonal experimental design to provide a quantitative basis for rule pruning. At the same time, external variables such as temperature and holidays in the features are used to construct the attention weight matrix and dynamically integrate external factors.

[0180] Step E: Utilize the optimized fuzzy neural network to predict the historical load data of the electric vehicles to be predicted, output the load prediction results, and apply non-negativity and peak value constraints to the predicted values. Median filtering is used to correct abnormal fluctuations, ensuring that the results conform to actual power grid operating conditions.

[0181] The optimized fuzzy neural network processing procedure includes: using the multidimensional feature data set obtained in step A, the parameters of the fuzzy neural network are optimized after processing in steps B, C, and D respectively, resulting in the optimized fuzzy neural network. The historical load data of the electric vehicle to be predicted is used to generate a multidimensional feature data set through step A, which is then input into the optimized FNN for prediction.

[0182] The optimized FNN parameters are applied to the FNN for prediction, outputting load prediction results. Non-negativity and peak constraints are applied to the predicted values. The non-negativity constraint ensures that all load prediction values ​​conform to physical meaning, avoiding unreasonable results such as negative values. The peak constraint limits the predicted value to no more than 1.2 times the maximum value in the training set, effectively preventing overshoot in the prediction results. Median filtering is used to correct abnormal fluctuations, and outliers exceeding the threshold are replaced with the median value in a sliding window to ensure that the results conform to actual power grid operating conditions. The load prediction results can be presented as an Excel file comparing the predicted charging load and the actual load, as well as several visualizations, performance graphs, etc.

[0183] Step E includes the following steps:

[0184] Step E1: Train the optimized fuzzy neural network. Based on the optimized parameters obtained in the previous steps, first decode the optimal parameters, and then generate fuzzy rules based on the decoded optimal parameters. After obtaining all fuzzy rules, perform prediction training on the training and test sets and calculate performance metrics. For example, the input historical charging dataset can be divided into training and test sets in an 8:2 ratio for model training and testing. In this example, the performance metrics are Root Mean Square Error (RMSE), Normalized Root Mean Square Error (NRMSE), and Mean Absolute Percentage Error (MAPE). Their calculation formulas are as follows:

[0185]

[0186] In the formula, n is the sample size, y i For the i-th actual value, Let be the i-th predicted value.

[0187]

[0188] In the formula, Max(y) i () represents the maximum actual value. This represents the maximum predicted value.

[0189]

[0190] In the formula, y i For the i-th actual value, Let be the i-th predicted value.

[0191] Step E2: Predicted value correction. This includes nonnegativity constraints, peak constraint, and outlier median replacement. Specific steps include:

[0192] Step E21: Nonnegativity constraint, the formula for the nonnegativity constraint in step E1 is:

[0193]

[0194] In the formula y pred Let be the original predicted value vector, and max(·, 0) is the element-wise non-negative operation.

[0195] Step E22: Peak constraint, the peak constraint formula in step E2 is as follows:

[0196]

[0197] In the formula That is, the maximum output value of the training set, 1.2 is the preset peak tolerance coefficient (adjustable).

[0198] Step E23: Median outlier replacement. The outlier replacement formula in step E3 is:

[0199]

[0200] In the formula y valid ={y pred |y pred For finite real numbers, the median is calculated as median(y) = Q. 0.5 (y), which is the sample median.

[0201] Step E3: Visualization Output. Based on the prediction data and performance index data, output relevant visualization content. In this embodiment, the prediction results include: Figure 4 Performance comparison chart of charging quantity prediction results with traditional algorithms on training and test sets. Figure 5 A scatter plot comparing the charging demand prediction results of traditional algorithms on the test set. Figure 6 Improved JADE optimizes the comparison curve between the seven-day charging demand predicted by FNN and the actual value. Figure 7 A bar chart comparing the charging demand at different stations on the same day with the actual value predicted by the improved JADE-optimized FNN.

[0202] It should be noted that the prediction results output in step E have also undergone an inverse normalization operation, which corresponds to the normalization operation in step A. The normalized prediction values ​​are restored to their original dimensions (actual charging load values) to ensure the physical meaning of the results.

[0203] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting electric vehicle charging load based on an optimized fuzzy neural network, characterized in that, Includes the following steps: Step A: Obtain several historical load data of electric vehicles, and generate a multidimensional feature data set after processing all historical load data; Step B: Divide the input feature space of the multidimensional feature data set based on the improved fuzzy clustering algorithm to obtain the clustering results and the dual-dimensional evaluation index of activation strength and contribution; the Gaussian membership function parameters include the center and width of the Gaussian membership function; The improved fuzzy clustering algorithm incorporates statistical correlations and distributional differences among multimodal distance dynamic trade-off features that fuse Mahalanobis distance and Wasserstein distance; Step C: Use the multidimensional feature dataset as input samples, and iteratively optimize the parameters of the fuzzy neural network based on the improved adaptive differential evolution algorithm; the parameters of the fuzzy neural network include Gaussian membership function parameters and regular weights; Step D: Further optimize the parameters of the optimized fuzzy neural network obtained in Step C through rule pruning and dynamic fusion of external factors to obtain the optimized fuzzy neural network; calculate the activation intensity and contribution of the rules based on orthogonal experimental design, and then realize rule pruning, that is, perform rule pruning on the initial rule base generated in Step B; use the multidimensional feature data set obtained in Step A to construct the temperature-holiday attention weight matrix to realize the dynamic fusion of external factors. Step E: Use the optimized fuzzy neural network to predict the historical load data of the electric vehicle to be predicted, output the load prediction result, and apply non-negativity and peak constraints to the predicted value.

2. The electric vehicle charging load prediction method based on an optimized fuzzy neural network according to claim 1, characterized in that, Step A specifically includes the following steps: Step A1: Obtain several historical load data of Ev, and use dual timestamp verification and anomaly cleaning to remove outliers from the historical load data. After outlier removal, use the STL time series decomposition method to decouple multi-scale time series features to obtain a time series with multiple features. Among them, dual timestamp verification refers to logical verification of charging start and end time and duration threshold constraints. The time series decomposition includes the extraction of daily, weekly and monthly periodic components, and integrates external features such as temperature and holidays. Step A2: Encode the external features in the time series to generate time series data with composite features as samples; Step A3: Perform dynamic normalization on the samples and construct a multidimensional feature matrix to generate a feature matrix with expandable dimensions. Each value in the matrix is ​​in the range of [0, 1], which serves as a multidimensional feature data set.

3. The electric vehicle charging load prediction method based on an optimized fuzzy neural network according to claim 1, characterized in that: Step B, the clustering process based on the improved fuzzy clustering algorithm, specifically includes: Step 1: Perform feature correlation analysis on the multidimensional feature data set and determine the multimodal distance fusion weights; calculate the linear correlation coefficient between features using the covariance matrix for the multidimensional feature data set, where Mahalanobis distance takes into account the covariance matrix, and dynamically determine the multimodal distance fusion weights based on the mean of Pearson correlation coefficients and historical mean, with the multimodal distance fusion weights serving as the contribution. Step B2: Initialize cluster centers for the multidimensional feature data set using K-means++; generate initial cluster centers for the multidimensional feature data set using the K-means++ algorithm, and calculate sample selection probability based on multimodal distance; Step B3: Dynamically update cluster centers and distance metrics. Each time cluster centers are updated, the feature covariance matrix is ​​calculated in real time, and incremental updates are used to eliminate feature scale differences and multicollinearity interference, dynamically adjusting the multimodal distance metric formula. Step B4: Dynamically generate Gaussian membership function parameters; Gaussian membership function parameters include the center and width of the Gaussian membership function; Step B5: Generate a rule combination matrix using the Cartesian product of the feature space as the initial rule base; Step B6: Calculate the activation intensity according to the Gaussian membership function product rule, and complete the dual-dimensional evaluation index of activation intensity and contribution.

4. The electric vehicle charging load prediction method based on an optimized fuzzy neural network according to claim 1, characterized in that: Step C, the process of iteratively optimizing the parameters of the fuzzy neural network based on the improved adaptive differential evolution algorithm, specifically includes: Step C1: Using the multidimensional feature dataset as input samples, perform a Logistic-Chebyshev-Sobol hybrid initialization operation on the parameters of the fuzzy neural network: Design a Logistic-Chebyshev-Sobol hybrid initialization mechanism, where 50% of individuals are generated through Logistic-Chebyshev hybrid chaos, 30% of individuals are generated through Sobol sequences, and 20% of individuals are fine-tuned from existing individuals through Gaussian perturbation; the existing individuals are those generated by both Logistic-Chebyshev hybrid chaos and Sobol sequences. Step C2: Construction of dynamic archive management mechanism: Create an empty archive set, set the maximum capacity to 20% of the population size, and use a first-in-first-out elimination strategy to keep the archive dynamically updated; Step C3: Dynamically adjust the scaling factor F and the crossover probability CR. The scaling factor F adopts Lehmer mean adaptive adjustment, and the crossover probability CR adopts Beta distribution adaptive adjustment. Step C4: Optimize the parameters of the fuzzy neural network using a hybrid mutation strategy of ABL-OCL-JADE for collaborative optimization and an elite preservation strategy; Step C5: Perform a binomial crossover operation; Step C6: Perform the selection operation using greedy selection; Step C7: Update the global optimal solution based on the fitness function iteration results, set an early stopping mechanism, terminate the evolution when the fitness improvement is less than 1e-4 for 15 consecutive generations, avoid invalid iterations, and output the fuzzy neural network parameters after iterative optimization.

5. The electric vehicle charging load prediction method based on an optimized fuzzy neural network according to claim 1, characterized in that: Step D, the rule pruning process, includes: removing redundant rules from the initial rule base based on a dual-dimensional evaluation index of activation strength and contribution, thus achieving rule pruning; orthogonal experiment-driven rule contribution modeling, constructing a three-factor, three-level orthogonal experiment matrix, using rule activation strength, consequent weight, and input feature importance as experimental factors, and quantifying the independent contribution of each rule to the output variable through analysis of variance.

Images