Meta-learning-based reinforcement learning with deterministic voltage clamping for global MPPT in grid-connected solar PV systems

Rapid solar photovoltaic (PV) grid integration requires robust global maximum power point tracking (GMPPT) under partial shading conditions (PSCs). Traditional continuous reinforcement learning (RL) controllers rely on static hyperparameters, rendering them susceptible to local peak trapping and power oscillations during rapid environmental transients. To resolve these computational and dynamic limitations, this paper presents a meta-learning-based reinforcement learning framework featuring an adaptive meta-deep deterministic policy gradient (AM-DDPG) controller for simultaneous GMPPT and automated voltage regulation in grid-connected PV systems. The core algorithmic novelty is an outer meta-learning layer governed by meta-plasticity principles that dynamically tunes the inner DDPG agent’s hyperparameters through a power–duty-cycle sensitivity gradient. To bridge advanced artificial intelligence with physical hardware safety standards, a deterministic override layer strictly clamps intermediate DC-link voltage fluctuations within safe operating limits (500 V–600 V). A grid-forming (GFM) inverter stage completes the system, providing autonomous synchronization through decentralized droop mechanisms. Evaluated under complex, time-varying, real-world shading profiles, the AM-DDPG framework consistently achieves 99.4%-99.8% tracking efficiency, converges within 12 ms, and suppresses the steady-state DC-link voltage ripple by 62% compared with conventional DDPG. These findings highlight a safe, computationally efficient AI control solution for resilient energy extraction in modern power grids.