Predicting Flood Events with Streaming Data: A Preliminary Approach with GRU and ARIMA

The most frequent flooding situations in the Azores are caused by torrential rainfall, which is difficult to predict due to its characteristics. Using data collected by sensors, rainfall and stream flow values, from natural occurring flood events, we describe results from learning RNNs, ARIMA time series forecast model, and a warning system that can empower civil protection decision-makers to safeguard property and lives. For dealing with all the difficulties resulting from forecasting rare events undersampling are used to get a richer sample of positive events, combined with simulation of new events. GRU and ARIMA models performed better than LSTM, using the hit hate measure and 30% of the positive events as test sample and 70% for learning sample. Even though the alert messages are sent by SMS to relevant deciders, two apps were developed to deploy the forecast models. An WWW application to manage the alerts and the sensors spread by all the islands, and a mobile app for operational staff working in all Azores archipelago.