Improving Hybrid Models For Precipitation Forecasting By Combining Nonlinear Machine Learning Methods

Abstract Precipitation forecast, especially on monthly and annual scales, is a key for optimal water resources management and planning, especially in semiarid climates with scarce water. The traditional hybrid models, in which two statistical models are used to separate and simulate linear and nonlinear components of precipitation time series, are still unable to provide accurate precipitation forecasts. This research aims to improve hybrid forecast models by combining one linear model and three nonlinear models with two preprocessing configurations: 1) using residuals of a linear model, representing the nonlinear component with different time steps and 2) using original time series of observations with different time steps, linear model simulations and residuals. Gene Expression Programming (GEP), Support Vector Regression (SVR) and Group Method of Data Handling (GMDH) models were used individually as in the traditional hybrid models and combinedly as in the proposed hybrid models in this study. The performance of the hybrid models was improved by different methods such as inverse variance (Iv) as an error-based method, least square regression, genetic algorithm and SVR. Two weather stations of Tabriz (annual) and Rasht (monthly) in Iran were selected to test the developed models. The results showed that Theil’s coefficient, UII, decreased in configuration one for the Tabriz station by 9% and 15% for SVR and GMDH relative to GEP, suggesting that these two models performed better than GEP in the precipitation forecast. The error criteria used in developing the proposed hybrid models with all forecast combination methods better represent observations than the hybrid model. MSE decreased by 67% and Nash Sutcliffe increased by 5% in the Rasht station in configuration two when we combined the three models using GA to obtain the improved hybrid model relative to the hybrid model combined with SVR. Generally, the hybrid models when SVR, the error based methods and GA were incorporated showed better performance than traditional hybrid models. The developed models have implications for modeling highly nonlinear systems using full advantages of machine learning methods.