Data-Driven Adaptive Electric Vehicle Load Forecasting Using Cluster-Weighted Reasoning and Local Regularized Regression

• Using dynamic cumulative vectorization to display the real-time status of the system with the ability to maintain the principle of causality and prevent future information leakage into the training process. • Designing an adaptive dimensionality reduction mechanism for automatic extraction of hidden features, with the aim of accurately modeling structural load changes over hourly intervals. • Providing a method for intelligent neighborhood optimization that determines the optimal number of neighbors for each scenario through knee-point detection based on the Kneedle difference function, establishing a balance between accuracy (bias) and stability (variance). • Developing an adaptive forecasting framework based on the integration of intelligent clustering and regularized regression for identifying various load behavior pattern. In this paper, a Clustered Weighted Case-Based Reasoning (CWCBR) framework is proposed for the simultaneous forecasting of energy demand and the number of incoming Electric Vehicles (EVs) in parking facilities. First, by mapping raw transactions to vector space, the charging session data is structured in the form of number and energy matrices to obtain a dynamic and cumulative feature vector. This process represents the real state of the system at each hour of representation by intelligently merging submatrices up to the current hour and without leaking future information. To manage the data complexity, the dimensionality reduction process was implemented hourly and independently through time-aware principal component analysis (TA-PCA). In the next step, by mapping the test sample to the historical neighborhood space, the optimal learning boundary of the manifold is determined. In this process, by using the Savitzky-Golay filter and slope threshold control to remove outliers, a coherent set of the most similar scenarios is extracted to forecast the future time step. Finally, the selected neighbors are modeled as local experts via ridge regression. Then, the weight of each is adjusted based on the structural similarity to the test sample. This recursive cycle is evaluated on ACN-Data database. The results show that the proposed model, with an average error of 4.99%, has a more accurate performance than reference methods: GRU (5.19%), LSTM (5.74%), and persistence (6.45%). Also, the error stability in the range of 1.08 to 1.24% in different time horizons indicates the robustness of this framework against error propagation