Chapter 11 - A new heuristic model for monthly streamflow forecasting: outlier-robust extreme learning machine

Abstract

Streamflow plays an important role in several hydraulic and hydrologic applications. Therefore, forecasting streamflow accurately is critical for better understanding of streamflow characteristics and variability over the time. This chapter proposed and discussed four heuristic extreme learning machine (ELM) models used for forecasting streamflows. Subsequently, the chapter includes a case study where application of the four heuristic ELM models is demonstrated in forecasting monthly streamflows using data from two stations in Turkey, namely, Topluca and Tozkoy stations. Furthermore, the case study involves investigating the comparative feasibility and robustness of four heuristic methods. The proposed heuristic models are (i) the outlier-robust extreme learning machine (ORELM), (ii) the regularized extreme learning machine (RELM), (iii) the weighted regularized extreme learning machine (WRELM), and (iv) the original ELM. Results obtained using the proposed models were compared with those obtained using the multiple linear regression models. Using monthly streamflow dataset measured between 1994 and 2007, the models were developed using 70% of the dataset as training data, and 30% as validation data, according to six scenarios. A total of four statistical metrics were used to evaluate performance of the models: correlation coefficient (R2), Nash–Sutcliffe efficiency (NSE), root mean square error (RMSE), and mean absolute error. At Topluca Station, the ORELM models provided the R2 values ranging from 0.604 to 0.858, NSE values from 0.313 to 0.735, and RMSE values from 11.121 to 17.827m3/s, respectively. At Tozkoy Station, the ORELM models provided the R2 values ranging from 0.615 to 0.911, NSE values from 0.291 to 0.829, and RMSE values from 3.164 to 6.439m3/s, respectively. Overall, the results suggested that the proposed ELM models have capability to forecast monthly streamflow with reliable accuracy.