Advanced Rule-Based System for Rainfall Occurrence Forecasting by Integrating Machine Learning Techniques

Though the magnitude of future rainfall is important in most water resources applications, many applications require its occurrence/nonoccurrence rather than its magnitude such as in agricultural systems management, drought management systems, regulated deficit irrigation for various crops, short-term municipal water demand modeling and management, and reservoir operation. The occurrence of rainfall is a classification problem that also affects day-to-day human activities and management. However, most of the work on rainfall forecasting is for rainfall magnitude, and very few studies on rainfall occurrence forecasting have been carried out in the past. Also, few artificial intelligence and machine learning techniques have been utilized in rainfall magnitude forecasting but not any work registered so far for forecasting rainfall occurrence using these methods. The proposed novel approach in this paper integrates two machine learning methods, artificial neural network (ANN) and decision tree (DT), which are capable of making rainfall occurrence forecasting comprehensible and accurate. For this purpose, the rules have been extracted by generating a DT using the input-output data obtained from an ANN rainfall occurrence forecasting model. Daily climatic data are employed to illustrate the methodology developed in this study. The obtained results show that during training, ANN models learned a fixed set of rules for rainfall occurrence forecasting. The obtained rules are simple and can be used as a tool for rainfall occurrence forecasting.