Edge-based real-time tracking of carbon emission flow in power systems with dynamic network pruning

Carbon emission flow is a virtual network flow attached to the physical power flow. Accurate tracking of carbon emission flow distribution within power systems constitutes a critical prerequisite for enhancing carbon emission accounting frameworks and refining responsibility allocation mechanisms. However, existing carbon emission flow models lack dynamic tracking capability for renewable-induced carbon emission flow variations. In this context, this paper proposes a lightweight edge computing method based on dynamic network pruning for real-time tracking of carbon emission flows in power systems. First, a coupled spatiotemporal graph convolutional network-transformer model is developed to accurately capture dynamic variations in grid carbon emission flow distribution induced by PV output and load fluctuations. Second, an adaptive mechanism is established, correlating core model structural redundancy with real-time measured PV output volatility to determine model simplification feasibility. This mechanism specifically identifies suitable periods for simplification when substantial carbon emission flow shifts are unlikely. Then, a dynamic pruning method is developed on edge devices to adaptively lighten the tracking model guided by this feasibility determination at resource-constrained edge devices. Finally, the effectiveness of proposed algorithms is verified in power systems with a complex network.