A similarity-based predictive scheduling method for dynamic electric vehicle charging load management.

Real-time energy management at public electric vehicle (EV) parking lots is a complex challenge that involves the distribution grid stability considering user demand uncertainties. This paper develops a multi-stage, data-driven control framework that focuses on both load profile smoothing and ensuring user constraints such as desired charge level at departure. In this regard, the proposed algorithm tries to overcome the limitations of classical methods by integrating three layers: historical similarity-based prediction, dynamic predictive optimization with a genetic algorithm, and a final repair stage. Comprehensive evaluations of different stochastic scenarios ensure the operational robustness of this algorithm. Peak-to-average ratio (PAR) of the entire network is reduced by shifting load from nighttime peak hours to off-peak hours. This performance, which has a high convergence with the global optimal state, demonstrates the system's effectiveness in peak shaving and valley filling strategies. Additionally, the system has prevented destructive stresses on the transformers by limiting the maximum load ramp rate. Final repair mechanism ensures final state of charge (SoC) error for all EVs to below 0.1% on the user side. Ultimately, with an average processing time of a few seconds and a 35% reduction in charging pile occupancy, this method acts as a real-time solution that enhances network resilience while postponing the need for heavy investments in physical infrastructure development.