A data-driven surrogate model for forecasting functionality states of transportation networks during extreme rainfall events based on the ConvLSTM method

Abstract

A quick foresight (or a short-term forecast) of highly probable functionality states of transportation networks under extreme rainfall events can provide excellent reference for proactive risk mitigation interventions, such as rescue, evacuation, and emergency material scheduling. Physics-based flood simulation and flow-based transportation network analysis at community scales are usually time-consuming, therefore can hardly meet the computational efficiency demand of functionality forecast of roadway networks, let alone that of real-time interactive mitigation decisions. Furthermore, unlike traffic forecasting under ordinary (non-hazardous) conditions, a functionality forecast under a heavy rainfall event has to reflect the dynamic non-local spatial correlations and the non-stationary temporal dependences among functionality states of network components, which are due to the meteorological characteristics of the rainfall process as well as the geographical and hydrological conditions of the community.

To address the above challenges, this study proposes a data-driven surrogate model based on the deep learning method to forecast the dynamic functionality state of a roadway network to a non-stationary rainfall-induced flooding evolution. The surrogate model employs the convolutional long short-term memory (ConvLSTM) network to capture the spatiotemporal features of dynamic network functionality states. The major contribution of this ConvLSTM-based surrogate model is two-fold: i) it implements dual-state input of extreme rainfall and traffic observation sequence through the Encoder-Decoder framework to capture the spatiotemporal impact of rainfall events on roadway network; ii) it accomplishes stable multi-step-ahead traffic response predictions under non-stationary rainfall conditions. A case study is conducted with a flood-prone community in Zhejiang Province, China, during a torrential rainfall to illustrate the potential application of the ConvLSTM-based model.