Flood Prediction Based on Weather and Water Level Historical Data Using Recurrent Neural Networks: A Case Study of Jakarta Flood Incidents

Abstract

In spite of high sea tides in combination with subsidence cause floods in the northern part of the city, successive rainstorm along the year is considered as the decisive cause of flood incidents in the Jakarta area. Flood incidents can massively damage and inevitably disrupt most of social-economic activities of the city. Taking flood risk in many respects into account, a local flood early warning services (FEWS) as the integral part of the city flood comprehensive mitigation plan may be urgently needed.  The capability of FEWS to provide a prediction of the scale, timing, and location of the impending flood may then be used to take city-wide precautionary steps.  In this study, based on local weather and floodgate water level historical data, an attempt to develop a base model of such a FEWS using recurrent neural networks (RNN) is carried out. The local weather time series data is first concatenated with the floodgate water level data and it is then utilized to predict water level at the corresponding floodgate in 7 days ahead. The predicted water levels in turn are used to decide flood alert categories in the nearby areas surrounding the floodgates. Different types of RNN such as long short-term memory (LSTM), gated recurrent unit (GRU) and combination of LSTM-GRU are examined in order to get the best model capable of giving minimum prediction error. Our computational experiments show that all three models succeed to produce a considerable low prediction error on three different datasets with the stacked GRU model demonstrates its superiority compared to the other two models. The stacked GRU with 32 neuron each is capable of giving root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE) namely 109.73, 82.91, and 0.05 respectively on Marina Ancol floodgate dataset