Dynamic path optimization and multi-objective decision-making for photovoltaic tracking systems using deep reinforcement learning

To address the core challenge of synergistically optimizing power generation efficiency, equipment wear, and environmental adaptability in photovoltaic (PV) tracking systems under dynamic conditions, this study proposes a deep reinforcement learning (DRL)-based intelligent optimization framework. A 37-dimensional state-space multi-objective decision process model is constructed, integrating spatiotemporal solar irradiance, PV panel status, environmental parameters, and historical decisions. An entropy-weighted multi-objective reward function balances conflicting objectives, while an improved Deep Deterministic Policy Gradient (CDDPG) algorithm, incorporating periodic decay learning rate, prioritized experience replay, and parallel training, resolves traditional issues of convergence oscillation, trade-off imbalance, and poor real-time performance. Validated via a dual-axis tracking platform using 2024 full-year field data from Xiaojin County, Sichuan, results show about 12.3% average efficiency gain vs. PID, Q-learning, and original DDPG, 18.7% lower equipment wear with 0.08 mm gear wear after 6 months, decision latency <8ms and 3.2% performance degradation under ± 5% noise, and 2.1% 6-month generation decay with 3.5% cross-season decay. This framework enables intelligent PV tracking upgrades with enhanced efficiency, reduced costs, and improved stability.