Comparison of Two Nonparametric Models, K- nearest neighbor and M5 Decision Tree in Forecasting the River Discharge in the Karaj Catchment

Abstract

The importance of water resources planning and management, the fast growing population, and the limited surface water resources, have made the application of the new technology to forecasting of river flow. A necessity, various methods have been presented in recent years to forecast the river flow, and the data-based models are considered the most reliable for this purpose. The river flow in the Karaj Catchment has been simulated using the data based models (KNN and M5). Hydroclimatological data (discharge, precipitation, temperature and evaporation) for the 2002 to 2009 duration have been collected to carry out the simulation processes. The performance and accuracy of the models were examined and compared. The Gamma test was used to select appropriate compositions. Suitable compositions were determined as the model inputs (KNN and M5). These features were entered in to the two data-based models. Results showed that both models simulated reliable flow predictions, if the discharge had been entered as an input. The M5 model showed a better precision as compared with the KNN model. The Coefficient of determination (R2) for the KNN and M5 models were 0.97 and 0.93, respectively. The RMSE were 0.55 and 0.87, for the same two models, respectively, and the value of the KGE were 0.99, 0.96, respectively.

Keywords

• decision tree m5
• gamma test
• gupta index
• k- nearest neighbor
• river discharge