An innovative multi-head attention model with BiMGRU for real-time electric vehicle charging management through deep reinforcement learning.

Electric vehicle (EV) charging management combines forecasting, pricing, and scheduling, with pricing and forecasts significantly influencing scheduling models. AI-based forecasting frequently uses LSTM and GRU to handle complex dependencies, but attention-based mechanisms excel at capturing long-term dependencies. Managing EV charging is difficult due to limited battery capacity and unpredictable variables such as traffic, user behaviour, and electricity prices. Researchers prefer model-free approaches incorporating deep reinforcement learning (DRL) to address these challenges. This paper describes a DRL-based solution for optimizing in-home EV charging, presented as a Markov decision process (MDP). We present the novel modified GRU (MGRU) model, which builds on GRU, and an advanced model, the multi-head attention-based bi-directional MGRU ("MHA-BiMGRU"). This innovative model optimizes EV charging by utilizing historical energy prices and automatically scheduling actions based on real-time electricity pricing to meet user needs and reduce charging costs. The extensive simulations with several other variants of the proposed model and related RNN-based models conclusively confirm the effectiveness of the proposed model "MHA-BiMGRU"compared to conventional RNN-based models (LSTM and JANET) in increasing user satisfaction and significantly lowering EV owners' charging expenses. Additionally, the proposed model reduces charging costs by 13% and 26% when using DQN and 15% and 36% when using DDPG, compared to the related LSTM and JANET-based models, respectively. [ABSTRACT FROM AUTHOR]