Development of machine learning models for short-term water level forecasting

Abstract

The impact of precise river flood forecasting and warnings in preventing potential victims along with promoting awareness and easing evacuation is realized in the reduction of flood damage and avoidance of loss of life. Machine learning models have been used widely in flood forecasting through discharge. However the usage of discharge can be inconvenient in terms of issuing a warning since discharge is not the direct measure for the early warning system. This paper focuses on water level prediction on the Storå River, Denmark utilizing several machine learning models.

The study revealed that the transformation of features to follow a Gaussian-like distribution did not improve the prediction accuracy further. Additional data through different feature sets resulted in increased prediction performance of the machine learning models. Using a hybrid method for the feature selection improved the prediction performance as well. The Feed-Forward Neural Network gave the lowest mean absolute error and highest coefficient of determination value. The results indicated the difference in prediction performance in terms of mean absolute error term between the Feed-Forward Neural Network and the Multiple Linear Regression model was 0.003 cm. It was concluded that the Multiple Linear Regression model would be a good alternative when time, resources, or expert knowledge is limited.