Comparative performance of recently introduced Deep Learning models for Rainfall-Runoff Modelling

Machine learning and specifically deep learning has been applied in solving numerous hydrology related problems in the past. Furthermore, extensive research has been done on the evaluation and comparison of performances of different Machine learning techniques applied in solving hydrology related problems. In this research, the possible reasons behind these performance variations are being assessed. The performance of recently introduced deep learning techniques for rainfall-runoff modelling are being evaluated by looking in to the possible modelling set-up and training procedures. Therefore, model set-up and training procedures such as: normalization techniques, input variable selection (feature selection), sampling techniques, model complexity, optimization techniques and random initialization of weights are being examined closely in order to improve the performances of different deep learning techniques for rainfall-runoff modelling. As a result, this study is trying to answer whether these factors have significant effect on the model accuracy.

The experiments are being conducted on different deep learning models such as: LSTMs, GRUs and MLPs as well as non-deep learning models such as: XGBoost, Random Forest, Linear Regression and Naïve models. Deep learning frameworks including TensorFlow and Keras are being implemented on Python. For better generalization, study areas from three different climatic zones namely: Bagmati catchment in Nepal, Yuna catchment in Dominican Republic and Magdalena catchment in Colombia are chosen to implement this experimental research. Additionally, in situ meteorological and stream flow data are being used for the rainfall-runoff modelling.

The preliminary model results show that model performances in case of Bagmati catchment are higher as compared to the other catchments. The LSTMs and MLPs are performing good with NSE values of 0.71 and 0.72 respectively. Most importantly, the linear regression model was outperforming the other models with NSE up to 0.75 in case of considering 6 days lagged rainfall input. This implies the relationship between daily rainfall and runoff data from Bagmati catchment may not be as complex. On the contrary, the 3-hourly data from Yuna catchment shows results with lower values for the performance metrics. This may be an indication of more complex relationships within the Yuna catchment.

This research provides key elements of the modelling process, especially in setting up and training deep learning models for rainfall-runoff modelling. The comparative analysis performed here, provides a basis of performance variations on different basins. This work contributes to the experiences in understanding machine learning requirements for different types of river basins.