Title:Online Learning of Interconnected Neural Networks for Optimal Control of an HVAC System

Abstract:Optimizing the operation of heating, ventilation, and air-conditioning (HVAC) systems is a challenging task, requiring the modeling of complex nonlinear relationships among HVAC load, indoor temperatures, and outdoor environments. This paper proposes a new strategy for optimal operation of an HVAC system in a commercial building. The system for indoor temperature control is divided into three sub-systems, each of which is modeled using an artificial neural network (ANN). The ANNs are then interconnected and integrated into an optimization problem for temperature set-point scheduling. The problem is reformulated to determine the optimal set-points using a deterministic search algorithm. After the optimal scheduling is initiated, the ANNs undergo online learning repeatedly, mitigating the overfitting. Case studies are performed to analyze the performance of the proposed strategy, compared to the strategies with a pre-determined temperature set-point, an ideal physics-based building model, and conventional ANN-based building models. The case study results confirm that the proposed strategy is effective in terms of the HVAC energy cost, practical applicability, and training data requirement.