Aiming at the problems of weak data timing coupling and lagging anomaly detection in dynamic scenarios of energy metering system, we propose a collaborative architecture of timing fusion Transformer and neural ODE to realize real-time governance and anomaly detection of the whole-link data. The multi-head self-attention mechanism is designed to dynamically aggregate voltage, current and other multi-dimensional timing correlations, and ensure the causality of timing based on causal mask; the continuous time modeling feature of neural ODE is used to replace the discrete computation and reduce the time consuming redundant inference. Experiments show that the architecture in dynamic energy metering scenarios, data governance end-to-end delay ≤ 14ms, voltage/current mutation anomaly detection accuracy of 98.25%, edge computing node memory occupancy of only 0.83GB, and real-time data reporting compliance rate of 98%, which has a high timeliness and strong scenario adaptability in the energy metering all-link governance.
