Work

CTDE Multi-Agent Deep Q-Network for Grid-Interactive Energy Management of a Saharan Microgrid

Year: 2026
Type: article
Abstract: We propose a grid-interactive, decentralized multi-agent Deep Q-Network (MA-DQN) with Centralized Training, Decentralized Execution for energy management in a Saharan hybrid microgrid (Boujdour, Morocco). Independent agents supervise PV, wind, diesel and batteries; a low-overhead PCC coordinator enforces import/export exclusivity, zero-crossing deadband and dwell time. A multi-objective reward balances cost, CO? and grid exchange, with penalties on ±10% voltage and SOC violations. On a 24-h case (demand 991.1 kWh), the learned policy delivers renewables-first dispatch (PV 56.4%, diesel 34.4%, grid 9.2%), battery throughput ±7.6 kWh, and 272.5 kg CO? (~275 g/kWh). Versus a PSO baseline, cost falls 1250?1120 MAD (?10.4%), CO? 318.9?272.5 kg (?14.6%), and inference latency 12.8?2.1 s (?83.6%). Sensitivity on the CO? weight ? yields tunable cost–carbon trade-offs (? 0.1?2.0: 310?240 kg CO? at +15.5% cost). Stress tests (PV depression/outage, +20% evening peak) maintain supply and power-quality; ablation confirms reduced PCC chattering versus single-agent DQN.
Language: English
Open Access status: diamond
OpenAlex ID: W7163386355
FWCI: 1.299
Cites:
Cited by: 1
Related to: 10