Application-oriented assessment of grid-connected PV-battery system with deep reinforcement learning in buildings considering electricity price dynamics

Deep reinforcement learning (DRL) is decisive in addressing uncertainties in intelligent grid-building interactions. Using DRL algorithms, this research optimizes the operational strategy of the building's grid-connected photovoltaic-battery (PV-battery) system, and examines the economic impact of battery capacity, rooftop PV penetration, and electricity price volatility. Three algorithms are employed, each demonstrating remarkable superiority over rule-based control. Without rooftop PV, the rule-based control achieves the battery cost saving of 0.07 RMB/(d·kWh) with a capacity equal to the average building load, while the three algorithms showcase a more substantial range of 0.17–0.19 RMB/(d·kWh). The cooperation of PV introduces heightened intricacy to the DRL training process. Incorporating PV radiation information into the state space remarkably amplifies the battery's capability to consume surplus PV, thereby enhancing economic benefits within the DRL strategy. Consequently, the battery attains cost savings of approximately 0.46 RMB/(d·kWh) under 50% PV penetration. Finally, the study reveals that as electricity price volatility intensifies, the advantage of DRL becomes more conspicuous. As grid renewable penetration progresses from 24% to 50%, the superiority of DRL over rule-based control in battery's cost savings escalates from 0.11 to 0.17 RMB/(d·kWh).