Deep learning assessment of rural rooftop solar potential and storage optimization

Rapid deployment of distributed rooftop photovoltaic (PV) systems can substantially reshape distribution-level net load and increase the likelihood of reverse power flow, making spatially explicit resource assessment and demand-side flexibility increasingly important. This study develops a bottom-up framework integrating multi-source remote sensing, deep-learning-based rooftop segmentation, geometric correction, spatiotemporal PV-potential estimation, net-load screening, and household PV-battery energy storage system (PV-BESS) scheduling for rural Qingdao, China. On the internal validation set, SegFormer outperformed U-Net and supported regional mapping of 282.04 km 2 of projected rooftop area, of which 127.34 km 2 was estimated to be technically usable. Spatial analysis showed an overall decline in building and PV-potential density with increasing distance from urban centers. City-scale and 443-household cohort analyses showed progressively deeper midday net-load troughs and reverse-power-flow conditions at higher PV penetration levels; these results are interpreted as planning-level screening rather than feeder operating-limit verification. A 24-h receding-horizon MILP-based model predictive control framework was then applied over a 153-day heating-season household dataset to evaluate coordinated PV-BESS operation across low-, medium-, and high-load groups. The 5-, 8-, and 10-kWh configurations served as reference storage cases for the three groups under the tested assumptions, while additional sensitivity analyses showed that rooftop estimates and PV-BESS performance depend on geometric, tariff, technology-cost, and irradiance assumptions. The proposed framework provides an evidence-bounded basis for spatially differentiated rooftop-PV planning and demand-side flexibility assessment in rural energy transitions.