Dynamic PMV/PPD prediction using Lagrangian-driven deep clustering and reinforcement learning for HVAC energy optimization

Forecasting indoor thermal comfort significantly influences the energy optimisation approach in buildings. The empirical Fanger's formula for a predicted mean vote (PMV) is the most accepted to be the closest to measuring human sensation of thermal comfort, which is characterised by their complex and nonlinear dynamical properties. The weights belonging to a wide independent range of the Lagrange formula cannot be used to fit data points of PMV. The clustering lagrangian model using reinforcement learning for dynamic thermal comfort (CLIMURLDTC) allows greater flexibility in tackling such problems. Each cluster easily achieves a high-level degree of fitting by the Lagrangian function. The challenging task is cluster matching between their boundaries' weights, which is effectively tuned to be matched using the reinforcement learning (RL) algorithm. The corresponding results showed that the proposed PMV mode effectively maintained thermal comfort while reducing heating, ventilation, and air conditioning (HVAC) systems energy usage. The study also demonstrated significant energy savings using the proposed PMV as a guideline for energy management. Further, it reduced total energy consumption by 22.17%, showcasing the potential for energy efficiency improvements in HVAC systems. Various statistical indices were used to evaluate the CLIMURLDTC model's performance.