Data-Driven Telecommunication Outage Prediction during Hurricane Events

Abstract

Telecommunication infrastructure (TI) has become an indispensable part in modern society, and its functionality is especially vital to emergency response during hurricanes. This study bridges the gap of lack of quantitative TI outage prediction models during hurricane events. County-level TI outage and power outage time-series and demographic data across eight continental US states during 10 recent hurricane events are collected. Two types of TI outage prediction models, namely, time-independent and time-dependent models, are developed. The time-independent model is intended for rapid prehurricane preparation or posthurricane outage evaluation, and is based on the partial least-squares regression technique. Relative predictor importance is also quantified via the Shapley additive explanations for better model interpretability. Moreover, to offer temporal TI outage prediction as the hurricane unfolds in real time, the time-dependent TI outage prediction model was developed, which leverages recent advances in recurrent neural networks such as the long short-term memory (LSTM) and bidirectional LSTM networks. The time-dependent model is able to handle time-series data and offers sequential TI outage prediction as new observations become available. Comprehensive model predictive performance evaluation is carried out and the explanatory power of different predictor combinations are examined. The proposed data-driven models can offer the much-needed quantitative and rapid TI outage prediction during hurricane events.