Bayesian network for monthly rainfall forecast: a comparison of K2 and MCMC algorithm

Abstract

Rainfall prediction has been one of the most challenging problems around the globe and has significance in many fields of science and technology. A Bayesian approach is used for monthly mean rainfall prediction at 21 stations in Assam, India. The inter-station rainfall dependencies/independencies are represented using Bayesian network (BN) structure and five atmospheric variables which include, Temperature, Relative Humidity, Wind Speed, Cloud Cover, and Southern Oscillation Index are used as predictors. Two different BN structure learning algorithms, K2 and Markov Chain Monte Carlo (MCMC) algorithm, are used. Thirteen different models are developed using different combinations of five predictors. K2 algorithm outperformed MCMC algorithm for all combinations. A combination of Temperature, Cloud cover, and Wind speed performed best for K2 algorithm giving 91.27% correct predictions, whereas a combination containing all the atmospheric variables performed best for MCMC algorithm giving 88.56% correct predictions. Thirteen stations out of 21 stations have accuracy above 90% for K2 algorithm, whereas only eight stations have accuracy above 90% for MCMC algorithm. Stations in the western part of the state performed better than stations in the other parts of the state.