AI-based management and dispatch for a photovoltaic-thermal-electric-hydrogen integrated energy system

• “Edge-deployed LSTM” achieves 4.7% MAPE for ultra-short-term PV forecasting. • DRL scheduling increased the self-consumption rate of PV to 95.8%. • 1D-CNN fault diagnosis achieves 98.8% accuracy for system reliability. • Synergistic “electric-thermal-hydrogen storage” ensures reliable multi-energy supply. • Lightweight AI framework enables real-time control on low-cost STM32 platform. The intermittency of photovoltaic (PV) power generation leads to significant PV curtailment issues, which limits its large-scale application. To overcome the limitations of existing research—such as a narrow focus on single energy forms or reliance on simulations—this study designs and implements a Photovoltaic-Thermal-Electric-Hydrogen Integrated Energy System (IES) that combines multi-energy storage with artificial intelligence (AI) technology. The novel contributions of this work encompass an integrated “prediction–optimization–diagnosis” AI framework deployed on edge hardware (STM32) for real-time control, along with synergistic electric–thermal–hydrogen storage coordinated by AI. Additionally, long short-term memory (LSTM), deep reinforcement learning (DRL), and 1D convolutional neural network (1D-CNN) models are deployed end-to-edge on a low-cost microcontroller. Specifically, the framework employs LSTM for PV prediction (achieving a MAPE of 4.7%), DRL for power dispatch (resulting in 95.8% self-consumption and a 5.2 percentage-point reduction in PV curtailment rate), and a 1D-CNN for fault diagnosis (with 98.8% accuracy). Lightweight deployment on the STM32 platform further enhances operational efficiency. Experimental results demonstrate the superiority of the proposed approach over model predictive control (MPC), improving energy efficiency, economy (with a 13.9% cost reduction), and reliability, while also contributing to sustainability outcomes such as reduced carbon emissions and higher renewable energy penetration.