Neural network-based model predictive control for residential four-asset energy co-optimization with vehicle-to-home integration and carbon-aware dispatch

HVAC dispatch alongside rooftop photovoltaic (PV) generation, battery storage, and a vehicle-to-home (V2H)-capable electric vehicle creates a high-dimensional, coupled control problem that conventional rule-based approaches cannot solve efficiently. This paper presents a physics-informed neural network-based model predictive control (NN-MPC) framework that keeps building thermal comfort central while co-optimizing a full four-asset residential energy system: PV, battery, heat-pump HVAC, and EV, under time-of-use pricing and real-time carbon intensity. A feedforward neural network, trained offline with a cost-augmented physics loss, replaces the online optimizer in a receding-horizon loop, cutting inference time from seconds to sub-milliseconds. Evaluated via a full-year Monte Carlo simulation (30 runs) using Algiers climate data and benchmarked against rule-based and mixed-integer linear programming controllers, NN-MPC achieves a 50.8% reduction in thermal comfort deviation, 7.5% lower electricity cost, 21.4% lower grid consumption, and 26.3% lower carbon emissions, recovering 87% of the MILP optimality gap at over 2,400 times lower cost.