A deep learning approach for hydrological time-series prediction: A case study of Gilgit river basin

Abstract

Streamflow prediction is a significant undertaking for water resources planning and management. Accurate forecasting of streamflow always being a challenging task for the hydrologist due to its high stochasticity and dynamic patterns. Several traditional and the deep learning models have been applied to simulate the complex nature of the hydrological system. However, to develop and explore a better expert system for prediction is a continuous exertion for hydrological studies. In this study, a deep neural network, namely a one-dimensional convolutional neural network (1D-CNN) and extreme learning machine (ELM) are explored for one-step-ahead streamflow forecasting for three-time horizons (daily, weekly and monthly) in Gilgit River, Pakistan. The 1D-CNN model gained incredible popularity due to its state-of-the-art performance and nominal computational complexity; while ELM model performed superfast as compared to traditional/deep learning architecture, gives comparable performance with fast execution rate. A comparative analysis is presented to assess the performance of the 1D-CNN related to the ELM model. The performance measurement matrices defined as the correlation coefficient (R2), mean absolute error (MAE) and root mean square error (RMSE) computed between the observed and predicted streamflow to evaluate the 1D-CNN and ELM model efficacy. The results indicated that the ELM model performed relatively better than the 1D-CNN model based on predefined statistical measures in three-time scale. In numerical terms, the superiority of ELM over 1D-CNN model was demonstrated by R2 = 0.99, MAE = 18.8, RMSE = 50.14, and R2 = 0.97, MAE = 136.59, RMSE = 230.9, for daily streamflow (testing phase) respectively. Based on our findings, it can be concluded that the ELM model would be an alternative to the 1D-CNN model for highly accurate streamflow forecasting in mountainous regions of the world.