摘要
With the advancement of energy transition,residential photovoltaic(PV)systems face intermittency challenges that impact grid stability.While battery integration enhances resilience,existing approaches exhibit critical gaps:(1)underdeveloped hybrid modeling frameworks balancing physical interpretability and data-driven accuracy;(2)reinforcement learning(RL)strategies prioritizing economic gains over grid stability,risking localized fluctuations;and(3)performance evaluations lacking systematic assessment across varying PV-battery capacities.To bridge these gaps,this study proposes a hybrid framework combining physical energy flow constraints with XGBoost-based machine learning for robust forecasting.Two optimization strategies,proximal policy optimization(PPO)and rule-based control(RBC),are developed for charge-discharge scheduling,explicitly incorporating grid stability metrics.Multi-scenario analysis evaluates performance under varying capacities and initial states of charge(SOC).Results demonstrate the hybrid model’s superiority over physics-based benchmarks,significantly improving prediction accuracy,with R2 increasing from 0.70 to 0.95 for SOC and from 0.83 to 0.98 for grid power.Both PPO and RBC enhance efficiency and stability versus baseline:the energy self-sufficiency rate rises from 10.6%to 79.3%(PPO)and 82.4%(RBC),while grid power fluctuations decrease from 2.6 kWh to 1.66 kWh(PPO)and 1.38 kWh(RBC).Crucially,RBC achieves higher stability and interpretability near boundaries,whereas PPO excels in long-term optimization but exhibits boundary-condition sensitivity.Results further reveal that PV-battery capacity and initial SOC influence strategy performance.This study establishes a structured technical pathway encompassing hybrid forecasting model development,stability-oriented optimization design,and scenario-based performance evaluation,providing an integrated solution to enhance grid resilience and energy autonomy in residential PV-battery systems.
