Partial shading mitigation on DC link stability based on deep reinforcement learning grid-forming inverter control

Abstract

Partial shading on photovoltaic (PV) arrays induces nonlinear power–voltage characteristics that cause rapid, multi-peaked fluctuations in DC-link power injection. Conventional grid-forming (GFM) inverter controllers, which rely on fixed-gain PI regulators tuned around a nominal operating point, struggle to track these fast, nonlinear disturbances: their linear control action cannot adapt quickly enough to abrupt irradiance changes, leading to sluggish transient response and, under severe shading, DC-link voltage collapse. This paper proposes a hybrid deep reinforcement learning (DRL) based GFM inverter control strategy that replaces these fixed-gain regulators with adaptive, learning-based controllers to overcome this limitation. A Deep Deterministic Policy Gradient (DDPG) agent governs the DC–DC boost converter for maximum power point tracking (MPPT), while a Twin Delayed Deep Deterministic Policy Gradient (TD3) agent replaces the conventional inner current PI regulators of the droop-controlled GFM inverter, enabling the controller to continuously learn and adjust its response to nonlinear, time-varying shading conditions rather than relying on static gains. Performance is evaluated under four irradiance scenarios, from uniform irradiation to severe partial shading, and compared against a conventional PI–PI GFM controller. Results show that the proposed GFM (PI–TD3) controller achieves faster DC-link voltage stabilization and superior voltage regulation under all shading conditions, maintaining the DC-link at 779.1 V under severe shading compared to a collapse to 290 V with the conventional controller. These findings indicate that TD3-based DRL GFM inverter control offers a robust, adaptive solution for standalone PV systems under variable irradiance conditions.