Rainfall Forecasting with Hybrid and Machine Learning Models Based on Hyperparameter Optimization

Abstract

Time-series analysis in hydrology plays an important role in the efficient use of water resources, prediction of flood risks, and crop production. However, considering many parameters in the solution of hydrological problems complicates the time-series analysis. For this reason, the use of artificial intelligence methods in hydrology studies has become interesting. The support vector regression (SVR) model and random forest (RF) model are frequently used in hydrological time-series forecasting. The SVR model reduces the experimental measurement errors as well as giving the statistical properties of the data used. The RF model is often used in the solution of regression. However, in both models, realistic values can be reached with different parameters in predicting total monthly rainfall (TMR). The reason for this is that the parameters used in the SVR model and RF model in time-series analysis should be used in different values for each period. In this study, the SVR model is used to forecast monthly rainfall amounts. The artificial bee colony (ABC) algorithm is used to obtain optimum values of SVR model parameters. In the proposed approach, TMR values in the years between January 1938 and December 2016 of Elazig, Erzincan, Erzurum, and Van stations in the eastern part of Turkey are used. The performance of the proposed model is compared with search-based SVR and RF models and hybrid SVR-genetic algorithm and RF-genetic algorithm. For the four stations, the results showed that the hybrid model performed significantly better. The proposed SVR-ABC model provided more realistic values than the search-based SVR and RF models.