Optimization of a Low-Power Inverter for AI-Based Solar Power Systems Using Reinforcement Learning Sukarman Purba (a), Bakti Dwi Waluyo (a*), Wanapri Pangaribuan (a), Selly Annisa Binti Zulkarnain (a)
a) Department of Electrical Engineering Education, Universitas Negeri Medan, Medan, Indonesia.
*bakti_dw[at]unimed.ac.id
Abstract
The rapid growth of rooftop photovoltaic (PV) installations has intensified the demand for compact, high-efficiency inverters capable of maintaining power quality under variable irradiance. This study proposes a reinforcement-learning-driven control strategy for a low-power single-phase inverter that interfaces residential PV arrays with local AC loads. A Deep Q-Network (DQN) agent is trained to minimize total harmonic distortion (THD) and switching losses while tracking a sinusoidal reference under wide-ranging solar conditions. The research follows a two-stage methodology. First, a detailed MATLAB/Simulink model of the inverter-incorporating real PV I-V data and an L-filter-is used to train the DQN offline with experience replay and prioritized sampling. Second, the learned policy is ported to a 32-bit microcontroller and validated on a 500 W hardware prototype fed by a 1 kW PV array and programmable DC source. Experimental results demonstrate that the proposed controller achieves a THD of 1.8 % at rated power, outperforming conventional PI (3.9 %) and model-predictive control (2.4 %) baselines. Efficiency improves by 2.7 % owing to optimized switching sequences, and dynamic response time to irradiance steps (800 →- 300 W/m2) is reduced by 35 %. Reliability tests over a 72-hour duty cycle confirm stable operation without thermal derating. These findings indicate that reinforcement learning can deliver high power quality and energy savings in low-power solar inverters, supporting the broader adoption of AI-based control in distributed renewable energy systems.
Keywords: Low-power inverter- Reinforcement learning- Solar PV- Total harmonic distortion- Energy efficiency
