Application of neural networks for flood forecasting in the Itajaí-Açu River in Blumenau, SC, Brazil

Abstract

Artificial neural networks (ANN) have been used successfully in predicting variables based on previous events, however, there are few studies on the application of this solution to river levels forecasting in flood events. This study aimed to evaluate the application of ANN to the levels forecasting of the Itajaí-Açu river in the short term in the municipality of Blumenau, Santa Catarina, Brazil. Blumenau was chosen as a study area for its extensive history of flooding. Rain and river level data from the telemetric stations located in the hydrographic basin of the Itajaí-Açu river of the Brazilian National Water Resources Information System (SNIRH) – National Water Agency (ANA) – were used to train the neural networks. Both data have a frequency of 15 min. 7 hydrological alert events registered by the limnimetric station installed in the municipality of Blumenau were selected. The data were collected and gathered according to their location and type, and were scaled to the same unit of measurement: cm for levels and mm for precipitation. Two types of neural networks were used: Long Short-Term Memory (LSTM) and Multi Layer Perceptron (MLP). To evaluate the performance of the neural networks, the following parameters were used: coefficient of determination (R2), Nash-Sutcliffe Efficiency Coefficient (NSE), Root-mean Square Error (RMSE), Mean Square Error (MSE), the Mean Absolute Error (MAE) and the Mean Absolute Percentage Error (MAPE). The model with the best performance – LSTM model with 6 h forecast horizon – presented: R2 = 0.996594; NSE = 0.9995548; RMSE = 7.72 cm; MSE = 59.65 cm; MAE = 4.82 cm; MAPE = 1.89%; and MSE val. = 0.000035. The study showed that the LSTM model, with simple pre-processing, is capable to predicting the water level of the Itajaí-Açu river during extreme flood events, with high precision, presenting better results compared to the MLP model. This study presented a proposal for a viable level forecasting model solution, which can be applied as a real-time forecasting tool for the study area. This study contributes to the development of support systems for the management of water resources and to mitigate the impacts caused by disasters, covering social, economic and environmental spheres.