One of the most common forms of natural disasters is flood, which is an overflow of water. Time series modeling plays a key role in many real-world problems. Forecasting time series helps to track stock prices, or even to anticipate electricity demand. Predicting floods is one of the main challenges for disaster management. We often want to predict over a long-term period, for instance several days. However, the majority of existing time series models are only capable of predicting one time point ahead. Multi-step forecasting, which considers multiple time points, is often a challenging task because of the lack of information and accumulation of errors.
Due to climate change and global warming, floods are more likely to increase both in size and frequency. In Antwerp, water levels may be influenced by other factors, such as the water levels in Vlissingen, wind direction, wind speed, water temperature, water discharge, air pressure and air temperature. In this thesis, we study multivariate and multi-step deep learning models using Long short-term memory (LSTM) for half-hourly water levels prediction. We investigate 6 architectures of LSTM networks and combine them with standard feedforward neural networks (convolutional neural networks). We use previous week meteorological data to predict water levels up to 24 hours ahead. The presented models are multi-site, and thus, able to make water level predictions for different sites, simultaneously.
Two statistical evaluation parameters, namely, the mean absolute error and the root mean square error are used to assess how the models perform. We compared our models to a baseline univariate statistical model (ARIMA). The results show that the ARIMA model performs as well as the deep learning models. Then, we decided to enlarge our sample of data and compared the ARIMA model with our best-performing deep learning model. The deep learning model significantly provided better results.