Machine learning techniques for monthly river flow forecasting of Hunza River, Pakistan

The forecast of river flow has high great importance in water resources and hazard management. It becomes more important in mountain areas because most of the downstream populations have high dependency for their livelihood, agriculture, and commercial activities like hydro power production. In this context, in recent times, machine learning models have got high attention due to their high accuracy in forecasting through self-learning from physical processes. In this work, we consider the potential of a data driven methods of machine learning, namely multilayer perceptron (MLP), support vector regression (SVR), and random forest (RF), are explored to forecast Hunza river flow in Pakistan using in-situ dataset for the period from 1962 to 2008. A set of five input combinations with lagged river flow values are developed based on the autocorrelation (ACF) and partial autocorrelation function (PACF) on historical river flow data. A comparative investigation is conducted to assess the performance of MLP, SVR and RF. The results of machine learning models are compared using forecasting metrics defined as correlation coefficient (R), mean absolute error (MAE), root mean square error (RMSE), and mean square error (MSE) between the observed and forecasted river flow data to assess the models’ effectiveness. The results show that RF performed the best followed by MLP and SVR. In measurable terms, superiority of RF over MPL and SVR models was demonstrated by R² = 0.993, 0.910, and 0.831, RMSE = 0.069, 0.084, and 0.104, MAE = 0.040, 0.058, and 0.062, respectively. The RF model performed 33.6% better than SVR and 17.85% to MLP. The results strengthen the argument that machine learning algorithms/models particularly RF model can be used for forecasting rivers flow with high accuracy which will further improve water and hazard management.