Novel Ensemble Method for Long Term Rainfall Prediction

Abstract

In the field of weather forecasting especially in rainfall prediction many researchers employed different data mining techniques to deal with that problem by using different predictors. This paper proposes a novel method to develop long-term weather forecasting model for rainfall prediction by using ensemble technique. Monthly meteorological data that obtained from Central Bureau of Statistics Sudan from 2000 to 2012, for 24 meteorological stations distributed among the country has been used. The dataset contained date, minimum temperature relative humidity, wind direction and rainfall as the predictors. In the experiments we built 10 base algorithm models (Gaussian Processes, Linear Regression, Multilayer Perceptron, IBk, KStar, Decision Table, M5Rules, M5P, REP Tree and User Classifier.), 7 Meta algorithms(Additive Regression, Bagging, Multi Scheme, Random Subset, Regressionby Discretization, Stacking, and Vote).The new novel ensemble method has been constructed based of Meta classifier Vote combining with three base classifiers IBK, K-star and M5P.The models have been evaluated by using correlation coefficient; mean absolute error and root mean-squared error as performance metrics. Also we use the both time taken to build the model and time taken to test model on supplied test set to compare and differentiate among the models results show that the new novel ensemble method has the best performance comparing to both basic and Meta algorithms.