Temporal Evaluation of Adaptive Neuro-Fuzzy Inference System for Rainfall Time Series Modeling

Abstract

Accurate prediction of future rainfall based on current conditions and historical events is important for both weather forecasting and water resource management domains. Adaptive Neuro-Fuzzy Inference System (ANFIS) is state-of-the- art soft computing technique extensively applied by many meteorologists and civil engineers to forecast rainfall and runoff. ANFIS has been frequently reported the superior performance over conventional statistical and mathematical modeling methods. The adaptive and learning abilities through artificial neural network architecture in addition to the uncertainty handling capability with the fuzzy inference system are the key ingredients of the ANFIS's success. In this work, we present the temporal evaluation of ANFIS on modeling rainfall time series. The main purpose of this empirical study is to observe the performance of ANFIS on predicting future rainfall based on historical data with varying time frames. For the temporal evaluation, we perform monthly rainfall data lagging from 1, 3, 6, 12 up to 18 months. Predictive performance of ANFIS with different time-frame data has been evaluated and compared against other efficient modeling techniques including linear regression, chi-squared automatic interaction detection, support vector regression, and artificial neural network. The experimental results reveal that ANFIS is the best model to predict short and medium-term rainfall in temporal dimension of 1 to 3 month lagging periods. Nonetheless, the conventional linear regression technique shows the best performance on predicting long-term rainfall with lagging periods from 6 to 18 months.

Keywords

• rainfall prediction
• adaptive neuro-fuzzy inference system
• anfis
• linear regression