A comparative study of severe thunderstorm among statistical and ANN methodologies

Area of Science:

• Meteorology and Atmospheric Science
• Artificial Intelligence in Weather Forecasting

Background:

• Severe thunderstorms are extreme weather events causing significant local damage.
• Accurate prediction with adequate lead time is vital for mitigating storm-related calamities.
• Existing prediction methods can be enhanced with advanced computational techniques.

Purpose of the Study:

• To introduce and evaluate novel machine learning approaches, Naïve Bayes and Radial Basis Function Network (RBFN), for severe thunderstorm prediction.
• To compare the performance of RBFN and Naïve Bayes against traditional methods like Multilayer Perceptron (MLP) and K-nearest neighbor (KNN).
• To assess prediction accuracy and lead time for severe storm events using specific weather parameters.

Main Methods:

• Application of machine learning algorithms including Naïve Bayes, MLP, KNN, and RBFN to meteorological data.
• Utilizing specific weather parameters not previously emphasized for these predictive models.
• Comparative analysis of the predictive performance of the selected algorithms.

Main Results:

• Radial Basis Function Network (RBFN) demonstrated superior performance compared to Naïve Bayes, MLP, and KNN.
• RBFN achieved a 95% prediction accuracy for severe squall-storms and 94% for no-storm scenarios.
• The developed models provided a significant lead time of 10-12 hours for predictions.

Conclusions:

• Radial Basis Function Network (RBFN) is a highly effective method for accurate severe thunderstorm prediction.
• The study highlights the potential of advanced machine learning techniques in improving weather forecasting lead times and accuracy.
• Findings from Kolkata, India, suggest broader applicability in similar meteorological contexts.