An evaluation of various data pre-processing techniques with machine learning models for water level prediction

Floods are the most frequent type of natural disaster. It destroys wildlife habitat, damages bridges, railways, roads, properties, and puts millions of people at risk. As such, flood detection systems have been developed to monitor the changes of water level and raise an alarm should there be imminent danger. River water level prediction is a significant task in flood mitigation planning and floodplains management. Usually, using raw data of rainfall series directly with machine learning (ML) regression methods, does not result in sufficiently good prediction accuracy. The raw data should be pre-processed using specific techniques to enhance their quality a priori to being applied to the prediction methods. This paper serves to address the stated problem by utilizing various data pre-processing techniques such as the Variational Mode Decomposition (VMD), Bagging, Boosting, Bagging-VMD, and Boosting-VMD to enhance the quality of input data and thus culminating in improved model accuracy. The five proposed pre-processing techniques were applied to the observed daily rainfall series of the Dungun river basin, Malaysia, for the period starting from November to February (Northeast Monsoon) from 1996 to 2016. Two machine learning models, the base models (Ori), that is the artificial neural network (ANN) and the support vector regression (SVR), were used in conjunction with the data pre-processing methods. The comparison between the ML methods with and without data pre-processing was done. It was found that prediction of water levels with the two ML methods of SVR and ANN together with the Boosting-VMD was superior to those results derived with just the base original model (Ori). The advantage of the enhanced models (respectively, founded on SVR and ANN) over the original models (SVR and ANN) is best reflected in the performance statistics. Numerical results in terms of root mean square error (RMSE) of (0.42, 0.20 vs 1.85,1.82), mean absolute percentage error (MAPE) of (4.36, 2.82 vs 18.89, 22.56), mean absolute error (MAE) of (0.28,0.16 vs 1.25, 1.41), and Nash–Sutcliffe efficiency coefficient (NSE) (0.96, 0.99 vs 0.25, 0.27) were obtained for the respective models. Additionally, various data visualization graphs such as hydrographs, residual hydrographs, peak-estimates, and box and whisker plots were illustrated to compare between various data pre-processing techniques. The experimental results showed that both the Boosting and the Boosting-VMD methods showed better performance over the other techniques. The Boosting-ANN model was found to be the better model to predict river water levels with the lowest RMSE (0.19), MAPE (2.72), and MAE (0.15) and the highest NSE (0.99).