Abstract
Isolated island communities relying on diesel- or LPG-fuelled generation for electricity, heating, cooling, and freshwater face compounded fuel- and water-security risks that integrated, renewable-based multi-generation systems can help mitigate; capacity sizing and real-time dispatch are, however, rarely integrated within a single AI-managed architecture in the existing literature. This study investigates whether a two-stage framework — metaheuristic capacity sizing followed by reinforcement-learning-based real-time dispatch — can improve the techno-economic and environmental performance of such a system beyond conventional rule-based operation. A novel solar tetra-generation system is accordingly proposed for off-grid isolated islands, simultaneously supplying electricity, space heating, cooling, and freshwater. The architecture integrates photovoltaic arrays, gas engine units with heat and power recovery, evacuated tube solar collectors, battery banks, hot water storage tanks, an electric chiller (selected over an absorption chiller owing to the high ambient temperature and humidity of the Persian Gulf coastal environment), and a reverse osmosis desalination unit; all thermal demand is satisfied exclusively through evacuated tube collector output and gas engine waste heat recovery, without any auxiliary fossil-fueled heater, and surplus thermal energy is redirected to preheat RO feedwater, reducing its annual electricity consumption by 746kWh/year (5%). The principal methodological novelty lies in the sequential coupling of ANN-PSO capacity sizing with DQN-based real-time dispatch: a seven-design-variable Artificial Neural Network (ANN) surrogate accelerates Particle Swarm Optimization (PSO)-based capacity sizing by 10–15%, and the resulting configuration is subsequently governed by Deep Q-Network (DQN)-based real-time dispatch of five subsystems (battery, gas engines, RO unit, feedwater preheater, and chiller) via a 72-action policy conditioned on an 11-parameter system state. Through reward-maximizing adaptive dispatch, DQN-based energy management increases freshwater production by 30% (from 2.3 to 3.0 m³/day) by reallocating previously curtailed photovoltaic surplus, without requiring additional membrane capacity. Under local LPG pricing (0.69 $/kg), the optimized configuration for Sadaf Island (Persian Gulf) achieves renewable fractions of 43.3% for electricity and 43.6% for heating, reducing LPG consumption and CO₂ emissions by 67.3% relative to a conventional LPG-only baseline, with a comparative annual benefit of 10,580 $/year and a payback period of 3.8years. Under international market-rate pricing of LPG (1.71 $/kg), the renewable electricity fraction increases to 74.0%, the payback period shortens to 1.7 years, and the comparative annual benefit rises to 41,800 $/year. These results demonstrate that coupling AI-based capacity sizing with adaptive real-time dispatch offers a practical, transferable route to reducing fossil-fuel dependence and strengthening water security in isolated, renewable-rich coastal communities.