Learning Data-Driven Control Policies for pH Regulation in a Closed-Loop Hydroponic System Using Random Forest

Precise regulation of nutrient solution parameters is critical for optimizing crop growth in hydroponic systems, with pH control being a key factor. This study presents a data-driven approach to learn control policies directly from multi-sensor measurements in a closed-loop NFT hydroponic system cultivating radishes (Raphanus sativus). A dataset of over 345,000 observations from 11 environmental and water-quality sensors was used to train a Random Forest classifier capable of determining appropriate pH control actions, including the injection of acidic or alkaline solutions, or neither.
The trained model achieved a classification accuracy of 89.3% and a balanced accuracy of 92.9%, demonstrating its ability to capture complex nonlinear relationships between sensor signals and control actions. Real-time experiments confirmed that the system maintained the pH near the target setpoint (pH ≈ 6) with rapid stabilization (∼ 35s) and minimal oscillations. Regression metrics further validated the model’s predictive capability:
R2 = 0.9524, MAE = 0.0148, RMSE = 0.0201, and MAPE = 0.255%.
These findings highlight that machine learning-based control policies can reliably reproduce and enhance human-designed control strategies without requiring explicit system modeling. This approach provides a robust and scalable framework for automated nutrient management in hydroponic cultivation, enabling efficient, precise, and sustainable operations in controlled environment agriculture.