Reduction of Multicollinearity and Nonlinear Problem in Rainfall Prediction Using Data-Driven Environmental Models Based on MIR
Year: 2025
Type: article
Abstract: It is well acknowledged that uncertainty is a basic aspect of forecasting for weather, seasonal climate, and hydrology.Random Forest and the Stochastic Response Surface Method (SRSM) offer the advantage of high accuracy for extremely nonlinear situations and is a helpful strategy for handling the non-convergence of computing results in the reliability study of weather forecasting.Multicollinearity, characterized by high correlation among explanatory variables, is not considered in machine learning model and SRSM.Mutual Information Ratio (MIR), which maximizes the joint dependency with rainfall to determine the most pertinent and least redundant set of parameters, has been used in this study.This study investigates the association between rainfall and seventeen explanatory variables, including maximum temperature, minimum temperature, dry bulb temperature at 9:30 am, dry bulb temperature at 18:30 pm, wet bulb temperature at 9:30 am, wet bulb temperature at 18:30 pm, dew point temperature at 9:30 am, dew point temperature at 18:30pm, humidity at 9:30am, humidity at 18:30 pm, pressure at 9:30 am, pressure at 18:30 pm, wind speed at 9:30 am, wind speed at 18:30 pm, maximum wind speed, sunshine and evaporation.As categorical variables, the data were obtained from the Meiktila weather station's Myanmar Meteorological Data.By lowering the multicollinearity issue, the outcome provided the amount of components needed to explain rainfall prediction.The suggested reliability analysis system uses random forest, the Least Absolute Shrinkage and Selection Operator (LASSO) regression-based SRSM, the Ridge regression-based SRSM, the Elastic net regression-based SRSM in combination with the Mutual Information Ratio (MIR) based conditional entropy approach to address and lessen the effect of multicollinearity present in weather forecasting of reliability analysis.Comparative analysis results of are displayed in this system.According to the adjusted R-squared, Standard Error (SE) and P-value, this system compares the performance between original Random Forest and regression methods, Random Forest and the regression methods following the use of MIR.The accuracy results of Random Forest and the Stochastic Response Surface Method (SRSM) after applying Mutual Information Ratio (MIR) are greater than the other existing models.Hence, Mutual Information Ratio (MIR) can be recommended as the appropriate method for complex data which is applicable in the domain of rainfall prediction for weather data.
Language: English
Subfield: Environmental Engineering
Field: Environmental Science
Domain: Physical Sciences
SDG: Climate action
Open Access status: hybrid
OpenAlex ID: W4409406836