Dandelion Optimizer-Based Reinforcement Learning Techniques for MPPT of Grid- Connected Photovoltaic Systems

Abstract

The integration of photovoltaic (PV) into electric power systems has been widely explored and adopted to address the problems associated with the depletion of fossil fuels and the release of greenhouse gases. PV panels convert sunlight into electricity, minimizing the reliance on fossil fuels and mitigating environmental pollution. It is crucial to optimally utilize the PV power in the system; hence maximum power point tracking (MPPT) algorithms have been developed to ensure optimal performance of grid-connected PV systems at the maximum power point (MPP) despite changes in weather conditions. Moreover, deep reinforcement learning (DRL) developments provide a promising approach for optimizing grid-connected PV systems, replacing the conventional proportional-integral-derivative (PID) controllers. However, there is limited research evaluating the efficiency of these systems using DRL techniques. This paper proposes a new dandelion optimizer (DO)-based DRL for MPPT of grid-connected photovoltaic systems and evaluates the proposed method for a 100-MW PV plant connected to a 33-kV distribution system. The proposed DRL technique uses proximal policy optimization (PPO) and deep deterministic policy gradient (DDPG) algorithms for continuous states and discrete or continuous action spaces to adjust the PV-measured voltage based on a reference one produced via DO-PPO and DO-DDPG methods. To test the effectiveness and practicality of the introduced methods, simulations were conducted using actual input data of a 100 MW PV plant connected to a 33-kV distribution system for typical days in summer and winter seasons using MATLAB/Simulink software. The proposed implemented methods were evaluated by comparing their simulation results with other techniques: DO-PID, particle swarm optimization (PSO), and incremental conductance (InC-PI). The findings revealed that the efficiencies of the DC-DC boost and the voltage source converters using the introduced methods were 84.25%- 85.90%, and 78.33%- 81.10% on a summer day while they were 92.77%- 95% and 86.70%- 89.50% on a winter day, respectively, which proves that these methods were efficient and effective, indicating their promising potential for future applications.

Keywords

• Dandelion optimizer
• deep deterministic policy gradient
• deep reinforcement learning
• maximum power point tracking
• PV systems
• proximal policy optimization