Bias Correction of Maximum Temperature Forecasts for Ensemble-Based model using Various Machine Learning Techniques for Hyderabad Station

The increasing frequencies of extreme weather events like heavy precipitation, drought, heatwaves, etc, have been associated with climate change in recent years. The reliability of air temperature forecasts at 2 meters above the surface is vital when trying to prepare for potential weather-related disasters, such as heat waves. In recent years, there has been a lot of emphasis placed on the prediction of heatwave conditions over India by using deterministic Numerical Weather Prediction (NWP) models. Despite improvements in model physics and resolution, deterministic NWP models have difficulties predicting extreme events at longer lead times. As the model integrates over time, errors grow due to the uncertainty associated with the initial conditions. This uncertainty is taken into account using ensemble prediction systems (EPSs). Heatwaves are now being predicted in India using EPSs due to their better performance in predicting events with longer lead times. The intensity of extreme events is typically underestimated by these models because EPSs typically have a low resolution and are also affected by the systematic biases present in the parent deterministic models. So to make the forecast more reliable, bias correction of the maximum temperature forecasts from EPSs is required.

This study focuses on the comparative assessment of various machine learning techniques for bias correction of maximum temperature from ensemble forecasts of maximum temperature for Hyderabad station. Three machine learning techniques were used in this study, namely Random  Forest,  Gradient Boost, and Support Vector Machine. The temperature forecasts used in this study were obtained from the National Centre for Medium Range Weather Forecasting (NCMRWF) global ensemble prediction system (EPS—called the NEPS). The NEPS configuration is based on the UK Met Office Global and Regional Ensemble Prediction System (MOGREPS).  The climatology used in this study is obtained from the MOGREPS data available on TIGGE and the observations for the maximum temperature are from the Indian Meteorological Department (IMD) station data (1985-2021), this data is used as the training set. The objective of this research study is to improve the accuracy of temperature forecasts by utilizing machine learning techniques for the bias correction of maximum temperature in order to improve model performance, primarily based on metrics such as Root Mean Square Error (RMSE). The initial raw RMSE values for Day 3, Day 5, and Day 7 are recorded as 2.1461, 2.4741, and 2.811, respectively. By examining the refined RMSE values for these specific forecast days, model corrections are revealed using Support Vector Machines (SVM), Gradient Boosting (GB), and Random Forest (RF) . After correction, the SVM model achieves improvements of 18.42%, 26.29%, and 27.42% in RMSE, demonstrating its increased predictive accuracy for Days 3, 5, and 7. Similarly, the RMSE reductions for GB on Day 3, Day 5, and Day 7 are observed at 18.77%, 26.23%, and 28.16%, while RF exhibits reductions of 39.21%, 28.24%, and 22.5% for the corresponding forecast days.  The percentage reductions indicate the improved accuracy attained by bias correction employing various machine learning methods.

Keywords: Heat Waves, Ensemble Prediction Systems, Support Vector Machine, Random Forest, Gradient boost.