Open AccessUsing the mutual information technique to select explanatory variables in artificial neural networks for rainfall forecasting
Author(s) -
Babel Mukand S.,
Badgujar Girish B.,
Shinde Victor R.
Publication year - 2015
Publication title -
meteorological applications
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.672
H-Index - 59
eISSN - 1469-8080
pISSN - 1350-4827
DOI - 10.1002/met.1495
Subject(s) - mean squared error , artificial neural network , computer science , mutual information , index (typography) , explanatory model , statistics , data mining , econometrics , machine learning , artificial intelligence , mathematics , world wide web
ABSTRACT The artificial neural network ( ANN ), a data‐driven approach, is a powerful tool for forecasting rainfall. However, selecting the appropriate explanatory variables in order to develop ANN models for this purpose is a major challenge. Recent studies in various fields have highlighted the usefulness of the mutual information ( MI ) technique in identifying explanatory variables for application in non‐linear problems, which, however, has largely been unexplored in forecasting rainfall. The present study was carried out to fill this knowledge gap. Three ANN models were developed, with different explanatory variables, to forecast the rainfall in Mumbai, India. Model A used temporal data of past rainfall events, Model B used selected meteorological data apart from rainfall and Model C used those variables identified by the MI technique. When the results of Model C were compared with those of Models A and B, a reduction of 5.79 and 4.11% in normalized mean square error, respectively, 16.66 and 12.90% improvement in efficiency index, respectively, and 3.22 and 4.24% reduction in the root mean square error, respectively, were observed. Thus, this study highlights the superiority of the MI technique in selecting explanatory variables for ANN modelling, not only because of the enhanced performance of the model with respect to various indicators but also because this performance has been achieved with a simple ANN architecture.