Prediction of heat waves in Pakistan using quantile regression forests

Abstract

The rising temperature due to global warming has caused an increase in frequency and severity of heat waves across the world. A statistical model known as Quantile Regression Forests (QRF) has been proposed in this study for the prediction of heat waves in Pakistan for different time-lags using synoptic climate variables. The gridded daily temperature data of Princeton's Global Meteorological Forcing (PGF) was used for the reconstruction of historical heat waves and the National Centers for Environmental Prediction (NCEP) reanalysis data was used to select the appropriate set of predictors to forecast the heat waves using QRF. The performance of QRF in prediction of heat waves was compared with classical random forest (RF). The results showed superior performance of QRF in detecting heat waves compared to RF. The QRF model was able to predict the triggering and departure dates of heat waves with 1 to 10 days lead times at various levels of accuracy. The model was able to predict the triggering dates of 2 to 3 out of 3 heat waves in the month of May, 8 to 12 out of 13 heat waves in June and 2 out of 2 in July and the departure dates of 3 out of 3 in May, 10 out of 13 in June and 2 out of 2 in July with an accuracy of up to ±5 days. The evaluation of different atmospheric variables revealed that wind and relative humidity are the major factors that define the heat waves in Pakistan. The research proved the advantage of QRF model to predict the conditional quantiles that help to explain some extreme behaviors of temperature.