Bayesian Network Model for Monthly Rainfall Forecast

This paper aims at rainfall forecasting which has been one of the most challenging problems around the world. Rainfall forecasting has importance in different areas including scientific research, agriculture etc. A Bayesian network model is proposed in this paper for forecasting monthly rainfall at 21 stations in Assam, India. Bayesian network or belief network (BN) is a probabilistic graphical model which shows conditional probabilities between different variables/nodes. Rainfall at a station is taken as a variable for this model and dependencies between rainfalls at different station is shown by BN. Rainfall dependencies between different stations is found using K2 algorithm which finds BN based on a greedy search algorithm. Five local and global atmospheric parameters which include Temperature, Relative Humidity, Wind Speed, Cloud Cover and Southern Oscillation Index (SOI) are used as evidences for this model. Conditional probabilities between stations and atmospheric parameters are calculated using Maximum Likelihood Parameter Estimation (MLE). Monthly data of 20 years from 1981 to 2000 for all the parameters is used for this study which was taken from different sources. Bayesian model runs on discretized data so for this study we have taken into account three discretized values for each variable based on their distribution. Thirteen different combinations of five atmospheric parameters are studied which gives a comparison of the efficacy of different parameters in rainfall prediction. Standard data ratio 70: 30 is taken for training and testing of model. Efficiency of the model predictions is presented in the form of percentage of correct predictions for every case. Efficiency is found to be above 85 percent for most of the cases. This model can serve well for prediction of monthly rainfall. Similar model can be developed for daily data also.