Temperatura negli edifici: il Reinforcement Learning

Potenzialità e limiti del Reinforcement Learning per il controllo della temperatura negli edifici

Luca Ferrarini, Alberto Valentini - Politecnico di Milano - DEIB

(in lingua inglese) - Introduction: HVAC Control; Objective and Workflow: Design and implementation of a DRL agent for temperature control in a medium-sized building - Case Study - Reinforcement Learning Agent - RL Control Strategies - Robustness Analysis - Conclusions and Future Developments

- The model-based approach reduces the RL training time (months/years - days/hours)
- Cascade control exhibits superior performance with respect to centralized control
- RL-based controller with cascade configuration provides performance comparable to MPC
- RL agent robustness analysis regarding uncertainties in the model parameters (image)
- Reducing the mismatch between reference and simplified model introducing a variable fan coil thermal transmission coefficient (?) and COP
- Hyperparameter optimization algorithm (Bayesian optimization)
- Combining Model Predictive Control and Reinforcement Learning

Attached, you can download the full pdf of the presentation.