Photovoltaic fluctuation-adapted dynamic network pruning for low-voltage distribution network edge computing

The inherent volatility of photovoltaic (PV) output necessitates the use of high-complexity deep learning (DL) models for accurate predictions. However, such models operate at full capacity even during stable PV output periods, consuming redundant computational resources and overloading resource-constrained edge devices in low-voltage distribution network (LVDN). To address the above issue, this paper proposes a dynamic network pruning framework that adaptively adjusts DL model complexity based on PV fluctuations. Firstly, a PV fluctuation-sensitive channel importance assessment method is proposed to identify the redundant structures in DL models. Subsequently, a lightweight optimization framework with PV operational constraints is developed to adjusts pruning thresholds based on PV output uncertainty and edge resource availability. Finally, a dynamic network pruning technique is proposed to adaptively balance model accuracy and computational complexity in response to real-time LVDN operation status and PV output volatility, ensuring pruned sub-networks align with the evolving PV data characteristics. The empirical results show that the proposed method can provide a practical solution for deploying lightweight DL models on edge devices. Specifically, our method effectively compresses 72 % FLOPs of the DL model in PV fluctuation challenging environments with slight accuracy degradation.