This work introduces a novel multi-agent reinforcement learning framework with periodic adversarial training to enhance coordinated control and power quality in multi-microgrid systems facing false data injection attacks. The architecture combines global state sharing for improved coordination with adversarial training for resilience against cyber threats. Evaluations on a modified IEEE 33-bus system with three microgrids demonstrate the enhanced multi-agent reinforcement learning (EMARL) method's superiority, achieving a median voltage deviation of 0.003 per unit during normal operations—91.9% better than uncontrolled systems, 81.3% better than multi-agent soft actor-critic (MASAC), and 50.0% better than multi-agent deep deterministic policy gradient (MADDPG). Under random attacks, the median deviation of EMARL remains within 0.01 per unit, outperforming MASAC (0.018 per unit) and MADDPG (0.02 per unit). The framework consistently maintained voltages within the 0.95-1.05 per unit range.

This article introduces a way to optimize the power output of a photovoltaic (PV) system by implementing a maximum power point tracking (MPPT) technique. The algorithm is based on adaptive particle swarm optimization (APSO) under partial shading conditions (PSC). APSO aims to solve the issue that conventional MPPT algorithms cannot track the optimal global solution when the photovoltaic array's power-voltage (P-V) curve shows multiple peaks under PSC. In APSO, the algorithm adaptively adjusts the learning factor and inertia weight to optimize convergence speed and precision. The simulation results prove that APSO outperforms the conventional particle swarm optimization (PSO) algorithm by rapidly and precisely tracking the maximum power point (MPP) under uniform illumination and static or dynamic PSC. Moreover, the APSO exhibits fewer power fluctuations during the tracking process.

This paper uses the neural network to realize a photovoltaic (PV) system's maximum power point tracking. After Bayesian optimization of hyperparameters, the model can accurately determine the maximum power point voltage according to solar irradiance and temperature. The optimized PV system not only adapts to different operating conditions but also achieves the optimal power output under uniform irradiance, static irradiance shading, and dynamic shading, ensuring the maximum efficiency of the PV system. Moreover, the optimized model is superior to the unoptimized model and the traditional perturbation and observation method in terms of accuracy. The improved model enhances the practicality and reliability of the PV system, helping to improve PV efficiency and reduce energy loss.