Implementasi Deep Reinforcement Learning untuk Optimasi Energi Sistem Panel Surya pada Penghitung Ikan Otomatis

Abstract

Standalone battery-based photovoltaic systems can experience suboptimal energy utilization when energy-flow decisions rely on fixed control rules that do not adapt to changing operating conditions. This study implemented Deep Reinforcement Learning (DRL) based on a Deep Q-Network (DQN) on an ESP32 as a system-level energy manager and evaluated its performance against conventional Maximum Power Point Tracking (MPPT). The experimental evaluation was conducted for six days, 21–26 June 2026, from 10:00 to 16:00 WIB, using the same prototype and a counterbalanced ABBA switching design. Data were sampled every 10 minutes, producing 444 valid observations consisting of 222 observations for MPPT and 222 for DRL. The evaluated parameters were panel output power, cumulative energy, and power stability. The DRL mode produced an average panel power of 46.75 W, compared with 45.08 W for MPPT, corresponding to a 3.70% increase. The final report summary also records cumulative energy of 1,693.93 Wh for DRL and 1,633.82 Wh for MPPT, while the coefficient of variation was 16.07% and 16.57%, respectively. DQN inference latency on ESP32 was 35–85 ms per decision. Daily-block inferential analysis reported a paired t-test of p < 0.001 and Wilcoxon signed-rank p = 0.031 (n = 6 days). The findings indicate that system-level DRL energy-flow management can improve the utilization of photovoltaic power while maintaining comparable or slightly better relative power stability under the tested conditions.