Comparative evaluation of intelligent algorithms to improve adaptive neuro-fuzzy inference system performance in precipitation modelling

Abstract

Existent algorithms to train adaptive neuro-fuzzy inference system (ANFIS) such as gradient descent and least-square methods are useful, but they have problems such as trapping in local optimum as well as high-volume of computations. This paper examines the application of genetic algorithm (GA), ant colony optimization for the continuous domain (ACOR), particle swarm optimization (PSO) and differential evolution (DE) for improving the performance of ANFIS models in simulating monthly rainfall magnitudes. Monthly rainfall records of nine foothills to arid weather stations of Isfahan province, central Iran were used for evaluating the proposed methodology. The correlation coefficient and non-linear sensitivity analysis were performed on the available rainfall records. Monthly rainfall modelling was carried out based on three input groups and used separately in five applied models (e.g. ANFIS, ANFIS-PSO, ANFIS-ACOR, ANFIS-DE, and ANFIS-GA). The obtained results showed that proposed hybrid models had better accuracy than the simple ANFIS model in escaping local optima. The best model was ANFIS-ACOR (with R2 = 0.92, RMSE = 2.73 mm and SI = 0.26), while sole ANFIS model (with R2 = 0.33, RMSE = 9.98 mm and SI = 0.90) presented the worst output. Moreover, the obtained results showed that: (1) Coupling the optimization algorithms with ANFIS can improve its performance in modeling the monthly rainfall time series; (2) the correlation coefficient and sensitivity analysis of input data provided a valid input selection that could improve the model accuracy with neglecting the redundant input parameters.