Comparative analysis of statistical models in rainfall prediction

Rainfall prediction is an important part of weather prediction. Compared to conventional methods predicting rainfall rate, the approach applying historical records and data mining technology shows obviously advantage in computing cost. Many excellent works have been done attempting to build predicting model with data mining methods, however, most of them just test the predicting accuracy on data set at one specific location. In this paper, we propose two criterions to evaluate the performance of prediction ability. 11 representative subsets with different location are chosen from China Meteorological Administration (CMA)´s open dataset. Every subset is belong to one specific observing station of CMA. Three classification algorithms are tested on our prediction model. We compare varies of combination of observing station feature and classification algorithm (Naïve Bayes, Support Vector Machine and Back Propagation Neural Network). In the end, prediction accuracy of different subsets are sorted through typical features of stations (Latitude, longitude, altitude, average temperature and the prior probability of rainfall) to find their influence on prediction accuracy.