Optimization of an Automated Substrate Irrigation System Using the SAC Reinforcement Learning Agent

This study presents the optimization of an automated mushroom substrate irrigation system by integrating a Soft Actor-Critic (SAC) reinforcement learning agent with a recursive LSTM prediction model. The system, based on a Siemens S7-1200 PLC, CS650 dielectric sensors, and an Ethernet-based data architecture, provides real-time control of humidity, temperature, and electrical conductivity. Experimental data analysis shows that the SAC agent increases the episodic reward from 20–32 to 90–100 units over 200 episodes, stably maintaining the substrate moisture in the range of 61–65%. The LSTM model achieved a Validation Loss of 0.016–0.022, accurately predicting the hydro-physical parameters. Compared to traditional PID controllers, the SAC-based system reduces humidity deviations by 35–40%, reduces the risk of overwatering and drying out, and increases mycelium colonization. The results confirm that the developed cyber-bioprocess platform increases the stability of the mushroom cultivation process, water use efficiency, and product quality and shows potential for industrial application, which must be validated in larger-scale trials.