SYSTEM MODELING AND DATA-DRIVEN APPROACH FOR DYNAMIC RISK EARLY WARNING AND CONTROL THROUGHOUT THE POWER EXTENSION PROCESS

Abstract

Risk factors are highly coupled and continuously evolve throughout the entire power extension operation process. Existing risk management methods, primarily based on static experience, are insufficient to support real-time early warning and effective control. To address this issue, this paper uses multi-source time-series data from the entire power extension process as input. First, a Dynamic Bayesian Network (DBN) is constructed to describe the causal relationships and stage evolution characteristics of risks, enabling online updates of risk states within a probabilistic inference framework. Subsequently, a Long Short-Term Memory (LSTM) network is used to deeply model the risk-related time-series characteristics and predict potential risk evolution trends. By fusing the risk state probabilities output by the DBN with the LSTM prediction results, a unified dynamic risk early warning index is formed, and targeted control decision support is achieved based on the reverse inference of key risk nodes. Experimental results show that the average early warning lead time of the DBN+LSTM fusion model reaches 18.7 minutes; the risk state identification accuracy during the power extension stage is 86.8%; the overall false alarm rate and missed alarm rate are 5.21% and 1.82%, respectively. The research outcomes demonstrate that this method provides an effective technical approach for dynamic risk early warning and control throughout the entire process of power grid expansion, combining mechanistic explanation and data-driven capabilities.