Forecasting of compound dry-hot events based on preceding sea surface temperature patterns and machine learning: A case study of Pearl River Basin

Abstract

[Objective] In the context of climate change, compound dry-hot events(CDHEs) have shown an increasing trend in both frequency and intensity, posing severe threats to social and ecological systems. To enhance the early warning and prevention capability for CDHEs, it is imperative to establish a high-precision forecasting model. [Methods] Preceding sea surface temperature fluctuation patterns(PSFP) were constructed as predictors based on SST variations during different periods. The data were processed using the Loess seasonal-trend decomposition method. Three machine learning models—support vector machine(SVM), random forest( RF), and extreme learning machine( ELM)—were employed to develop deterministic, ensemble, and probabilistic forecasting models(PSFP-SVM, PSFP-RF, and PSFP-ELM) for CDHEs. Taking the Pearl River Basin as the study area, four representative summer CDHEs in 1989, 1991, 1999, and 2011 were selected for model validation. [Results] The result showed that the models could effectively predict the spatiotemporal distribution characteristics of summer CDHEs in the Pearl River Basin with 1-and 3-month lead time. Among them, the PSFP-ELM model performed best, achieving an R² higher than 0. 5 and an accuracy exceeding 0. 8. Its forecast result showed strong agreement with observations and superior performance at the 3-month lead time. The forecasting accuracy of the PSFP-based models was significantly higher than that of traditional circulation factors-based models, with performance improvements exceeding tenfold. [Conclusion] The findings provide a scientific basis for risk early warning and decision-making of disaster prevention and mitigation regarding CDHEs.

Keywords

compound dry-hot events / forecasting / machine learning / preceding sea surface temperature fluctuation / STL decomposition / Pearl River Basin / probabilistic forecasting / influencing factors