A similar day selection framework with heterogeneous feature integration for electric vehicle charging load forecasting

Accurate forecasting of electric vehicle (EV) charging load is essential for reliable power grid planning and operation. EV charging loads show strong temporal variability and high sensitivity to external factors. Directly using all historical samples may introduce irrelevant information and consequently reduce forecasting accuracy. Therefore, identifying representative similar-day samples is critical for reliable predition. This paper proposes a similar-day selection framework that integrates heterogeneous multi-source information through three key stages: multi-feature similarity computation, adaptive weight optimization, and pattern matching. Similarities between historical days and the target day are quantified across load, weather, and contextual feature spaces using dedicated similarity measures. A multi-objective optimization strategy is then employed to dynamically determine feature weights, enabling the construction of a candidate similar-day set. To avoid data leakage caused by the use of the target day load, an XGBoost-based pattern-matching model is developed. This model relies solely on observable contextual and weather features to discriminate among candidate days and select the final similar-day inputs based on matching probabilities. The proposed framework is validated using real-world EV charging data from Shenzhen, China. Experimental results demonstrated that the method improves both the quality of similar-day selection and the accuracy of load forecasting, achieving an average error reduction of approximately 4.5% compared with baseline models. Furthermore, this study quantifies the nonlinear influence of regional functional entropy on forecasting performance. Prediction accuracy remains stable in low-entropy regions with homogeneous functional structures, whereas it is significantly constrained in highly mixed regions with functional entropy greater than 0.94 due to increased behavioral complexity.