Attention-Augmented Recurrent Deep Q-Network with Adaptive CLLC Resonant Converter for Intelligent Multi-Source EV Fast Charging and Hybrid Thermoelectric-PCM Thermal Management: A Comprehensive Simulation Study

Abstract: This paper presents a simulation-based investigation of an Attention-Augmented Recurrent Deep Q-Network (A-RDQN) for energy management, executed on an Adaptive Capacitive-LLC (ACLLC) bidirectional resonant converter within an isolated renewable DC microgrid. This microgrid can power three fast-charging stations for electric vehicles (EVs) at the same time. The A-RDQN uses a four-head self-attention mechanism in a bidirectional LSTM (BiLSTM) encoder to model long-term temporal patterns in the variable profiles of photovoltaic (PV) and wind energy generation. This lets it look ahead 12 seconds, which purely reactive controllers can't do. A cascaded Hybrid Thermoelectric Cooler- Phase Change Material (HTC-PCM) system takes care of battery thermal safety. It uses a shared reward function and a two-layer thermal prediction sub-network to optimize the current and coolant flow rate of the Peltier module at the same time. In four different dynamic scenarios, MATLAB R2024a/Simulink-PLECS co-simulations show that the A-RDQN can restore the battery's state of charge (SoC) from 20% to 92% in just 710 seconds. It also keeps the peak cell temperature at 41 °C, limits total harmonic distortion (THD) to 0.68%, and achieves a bus voltage settling time of 0.019 seconds. It does better than FCS-MPC and ANN-based controllers in every metric that was tested. The ACLLC's secondary switched-capacitor bank also makes sure that Zero Voltage Switching (ZVS) happens across the 60-115% rated-load range, keeping the peak conversion efficiency above 96.8%.

Keywords: Adaptive CLLC resonant converter, attention mechanism, battery thermal management, deep reinforcement learning, dueling DQN, EV fast charging, hybrid thermoelectric-PCM, LSTM, model predictive control, renewable DC microgrid, and V2G.