Research on Energy Efficiency-Accuracy Co-Optimization and Adaptive Inference for Big Data of Population Mobility Under Edge Computing

Population mobility big data underpins urban governance, traffic planning, and public safety, yet traditional cloud processing suffers from high latency, privacy risks, and resource waste. Edge computing offers a lightweight, low-latency, privacy-preserving alternative, but existing edge inference schemes fail to balance energy efficiency and accuracy amid multi-source heterogeneous, spatio-temporally sparse, and dynamically changing mobility data. This paper proposes an energy efficiency-accuracy co-optimization framework and adaptive layered inference mechanism. First, a lightweight representation and low-bit shared quantization model reduces edge-side storage and computational overhead. Second, a deep reinforcement learning-based strategy dynamically selects edge nodes and partitions models to adapt to heterogeneous terminal computing power, bandwidth, and data distributions. Finally, a multi-objective optimization function balances accuracy and energy efficiency for joint optimization of inference accuracy, energy consumption, and latency. Experiments on mobile phone signaling, census, and traffic datasets show that compared with cloud and static edge inference, the proposed method reduces edge energy consumption by 42.7%, cuts latency by 58.3%, and improves memory access efficiency by 2.56×, with accuracy loss below 1.5%, enabling efficient, reliable large-scale population mobility monitoring and decision-making.