Comment on “Deep reinforcement learning approach for MPPT control of partially shaded PV systems in Smart Grids”

Abstract

In a recent work, (Avila et al., 2020), an environment was developed and reported for partial shading conditions (PSC) in the open-source OpenAI Gym platform. This work presented deep reinforcement learning (DRL) techniques to address the maximum power point tracking (MPPT) problem of a photovoltaic (PV) array under PSC. Two DRL algorithms, namely, Deep Deterministic Policy Gradient (DDPG) and Twin Delayed DDPG (TD3) were investigated. A deviation of less than 1%, compared to the theoretical maximum power, was claimed for the DDPG algorithm. Based on the presented fact-based investigations, this comment highlights the issues in the reported approach. The main issues are found to be the approximate PV array modeling, erroneous performance metric and erroneous choice of DRL algorithms. Of the presented PSCs, the proposed technique in Avila et al. (2020) always attained the very first peak of the PV characteristic of the PV system. It may be noted that the first peak may not always be the global maximum power point. Overall, based on the presented investigations this comment demonstrates that the DDPG technique used in Avila et al. (2020) is not able to effectively address the MPPT problem in PV array under PSC.