Аңдатпа
While Deep Reinforcement Learning (DRL) offers near-optimal control for Home Energy Management Systems (HEMS), its deployment on IoT edge devices is severely hindered by the high computational and memory footprints of standard neural architectures.This paper proposes a computationally lightweight framework that integrates a Time-series Dense Encoder (TiDE) with Proximal Policy Optimization (PPO) for intelligent electric vehicle charging on resource-constrained edge hardware.The frozen TiDE encoder generates compact 64-dimensional latent representations of future load trajectories, augmenting the RL agent's observation space with predictive context at linear computational complexity O(L).Experimental validation on the CityLearn 2022 benchmark demonstrates 5.9% and 27.1% cost reductions under time-of-use and dynamic pricing, respectively, with zero safety constraint violations across all evaluated scenarios.Benchmarking against Model Predictive Control confirms near-optimal performance (2.8% optimality gap) while achieving 255× inference speedup.A systematic comparison against offpolicy DRL alternatives (SAC, TD3) on the identical observation space confirms that TiDE-PPO achieves statistically comparable reward (p > 0.09) while maintaining a 22-46× smaller model footprint, making it the only edge-deployable DRL solution.Hardware-in-the-Loop testing on an ESP32 microcontroller validates edge feasibility: 34 ms execution per control step, 3.0% SRAM utilization, and 16.9 mJ energy per inference.Post-Training Quantization from Float32 to INT8 achieves 3.8× model compression (60.25 → 15.84 KB) with zero accuracy degradation (p = 1.0).Duty-cycle analysis confirms autonomous battery-powered operation exceeding 5 years under conservative derating assumptions.Ablation studies confirm that latent state augmentation outperforms both raw forecast and no-forecast baselines.