Volume 64

Explaining Deep Reinforcement Learning based Control Strategy Optimization of an Industrial Cooling System with SHAP Arthur Stobert, Tobias Lademann, Richard Stobert, Matthias Weigold

Abstract

Industrial energy supply systems are key enablers for energy-efficient and cost-effective production, yet their operation is characterized by complex dynamics, competing objectives, and increasing uncertainty. While deep reinforcement learning (DRL) has demonstrated significant potential for optimizing control strategies in such systems, its practical adoption remains limited due to the lack of interpretability and transparency of learned policies.
This paper presents an explainability-driven analysis of a DRL-based operational strategy applied to a real-world industrial central cooling system in the metal processing industry. A combination of qualitative behavioral analysis and SHapley Additive exPlanations (SHAP) is used to interpret the learned control policy and to reveal the decision logic underlying the agent’s actions. The DRL-based strategy achieves energy cost reductions of 29 % during training and 15.7 % on an unseen validation year compared to conventional rule-based control.
The explainability analysis shows that the agent autonomously prioritizes free cooling whenever favorable thermal conditions are present and activates mechanical chilling only when internal temperature gradients indicate insufficient heat rejection capability. Furthermore, the results reveal that internal thermodynamic states dominate the agent’s decision-making, while external signals such as electricity price and weather conditions have only minor influence in the investigated system.
By demonstrating how DRL-based control strategies can be interpreted and validated using XAI methods, this work contributes to improving transparency, trust, and acceptance of learning-based control in industrial energy supply systems.

Keywords explainable artificial intelligence, interpretable machine learning, SHAP, industrial energy supply systems, control strategy optimization, deep reinforcement learning