CLEVER: A Novel Approach for Improving EV Charging Duration and Load Predictions Using Curriculum Learning Approach

Abstract

Despite advancements in electric vehicle (EV) charging prediction models, existing approaches are suffering from complex charging patterns. The curriculum learning (CL) is a training approach which resembles the natural human learning progression by introducing training samples through different patterns, hence efficiently structuring the learning process. While the CL has been successfully used in other domains, its application in EV charging prediction remains unexploited. In this work, the CL is to be leveraged for the first time to improve the EV charging behavior predictions in both EV charging duration and charging load prediction. A feature-based curriculum learning approach, named CLEVER (Curriculum Learning EV chargER), is proposed for predicting charging session load and duration. CLEVER employs an advanced data stratification mechanism that introduces training samples progressively according to complexity metrics computed from temperature variations, state of charge variations, and temporal patterns. The CLEVER method integrates a CL strategy with a staged schedule mechanism over four neural network architectures: ANN, DNN, LSTM, and GRU. The performances obtained exhibit notable gains, where CL scores 20.9% reduction of Mean Absolute Error for GRU-based forecasting of EV charging duration and 2.2% improvement for DNN-based charging load forecasting. The CLEVER methodology shows considerable improvements in predicting the duration of EV charging, with, as many as 23.0% reductions in Mean Absolute Error with GRU models on Level 1 chargers and a near 20.7% improvement with DNN models on DC Fast Chargers. For EV charging load forecasts, curriculum learning produces consistent, but modest, gains, with improved up to 2.4% with ANN models on DC Fast Chargers and 1.6-2.1% improvements across different neural network architectures. This comprehensive analysis across different charger types, user groups, vehicle models, temperatures, and temporal patterns makes CL a superior approach to enhancing EV charging infrastructure management and grid stability.