Smart control of solar-powered oil press: Applying Reinforcement Learning for productivity and energy utilization improvement

Joévin Wiomou Bonzia*, Zhangkai Wub, Sebastian Romulia, Klaus Meissnera, Joachim Müllera
aUniversity of Hohenheim, Institute of Agricultural Engineering, Tropics and Subtropics Group, Stuttgart, Germany
bUniversity of Hohenheim, Institute of Agricultural Engineering, Artificial Intelligence in Agricultural Engineering Group, Stuttgart, Germany
*Correspondence: bonzi.wiomoujoevin@uni-hohenheim.de

The operation of the solar-powered oil press under reinforcement learning control

Components of the oil press

Oil press components

Reinforcement learning framework

Reinforcement learning framework

Simulation Performance of the RL Controller

Experiment validation

Experiment validation

BibTeX

@article{BONZI2025100772,
title = {Smart control of standalone solar-powered oil press: Applying Reinforcement Learning for productivity and energy utilization improvement},
journal = {Renewable Energy Focus},
pages = {100772},
year = {2025},
issn = {1755-0084},
doi = {https://doi.org/10.1016/j.ref.2025.100772},
url = {https://www.sciencedirect.com/science/article/pii/S1755008425000948},
author = {Wiomou Joévin Bonzi and Zhangkai Wu and Sebastian Romuli and Klaus Meissner and Joachim Müller},
keywords = {Deep reinforcement learning, Solar PV, Off-grid control, Microcontroller deployment, Rural agri-processing},
abstract = {In resources constrained rural areas, solar-powered oil extraction can be enhanced through recent advances in artificial intelligence for energy optimization. This study introduces SolPrInt, a deep reinforcement-learning (DRL) based controller for a standalone, photovoltaic-battery powered mechanical oil press. A proximal policy optimization (PPO) agent was trained in MATLAB/Simulink using 15 years of PVGIS-SARAH2 radiation data and peanut-oil extraction benchmarks. A primary training phase followed by an adversarial phase on the 5% least-sunny days reinforced robustness under low-irradiance conditions. The developed agent adapts press rotational speed to real-time PV availability, battery state of charge, and system behavior to ensure energy-efficient use of solar resources. In-silico validation achieved stable rewards and simulated throughput of 96 ± 13.5kg/d under sunny days and 90 ± 20.5kg/d under cloudy days. Compared with conventional fixed-schedule operation (08:00–18:00) under sunny and cloudy conditions, SolPrInt extends operating time, and reduces power outages, while improves oil yield by 0.7 percentage points. Experimental validation on a PV-simulator bench confirmed real-time deployment feasibility on a low-cost ESP32 microcontroller interfaced with a Kern Kraft KK20 press. These findings demonstrate the potential of PV-sensitive DRL control to improve the performance of standalone renewable energy systems, supporting reliable decentralized energy use and contributing to sustainable energy access sustainable energy access. Supplementary materials supporting this work, are available at https://bonjoe.github.io/solprint.demo/}
}