Machine Learning Method in Prediction Streamflow Considering Periodicity Component

Abstract

Accurately prediction of streamflow is very important issue for sustainable management of water resources. In this chapter, the applicability of three intelligent data analytic techniques based on the long short-term memory (LSTM) network, extreme learning machines (ELM), and random forest (RF) algorithms is examined in prediction of monthly streamflow. Data from two stations, Kohala and Garhihabibullah, Pakistan, are employed in the study and periodicity component was also considered as model input. The models' outcomes were evaluated using three statistics, root mean square error, mean absolute error, and determination coefficient. The results from the applied models indicated that the LSTM provided superior accuracy to the ELM and RF methods. The relative RMSE and MAE differences between the LSTM and ELM/RF are 6.4/31.9 m3/s and 9.1/24.3 m3/s for the Kohala and 5.9/11.9 m3/s and 8.9/20.8 m3/s for the Garhihabibullah, respectively. Including periodicity input (month number of the output) considerably improved the models' efficiency in prediction monthly streamflow in both stations: improvements in the RMSEs of the optimal LSTM, ELM, and RF models were 20.8{\%}, 20.5{\%}, and 3.7 for the Kohala and 11.9{\%}, 6.9{\%}, and 1{\%} for the Garhihabibullah, respectively. The ELM model was ranked as the second best. The chapter recommends the LSTM as a good alternative for streamflow prediction in the studied area.