<p>Accurate precipitation forecasting is essential for mitigating floods, optimizing agricultural planning, managing water resources, and ensuring transportation safety. However, it remains challenging due to complex nonlinear drivers and pronounced spatiotemporal variability. The primary prediction target in the main experiments of this study is the hourly precipitation (tp) from the ERA5 reanalysis dataset. The research framework is entirely constructed based on ERA5 data, and station-observed precipitation is not used as training or validation labels, thereby avoiding potential misunderstandings of “self-fitting” ERA5 variables. It should be emphasized that this design is not intended to simply reproduce ERA5 precipitation using its other variables; rather, it aims to learn a data-driven surrogate relationship between multivariate atmospheric states and point-specific precipitation responses. This framework can support applications such as rapid precipitation estimation, temporal downscaling, and serving as a computationally efficient alternative to direct reanalysis outputs. On this basis, to address this, we propose MSLEFT-Transformer, a framework that comprehensively considers the spatial characteristics, global trends, local details, and temporal dependencies of the precipitation processes, enabling precise modeling of precipitation dynamics. Experiments on 82 stations across Guangdong Province, China, show that MSLEFT-Transformer achieves coefficient of determination (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">$\varvec{R}^{\varvec{2}}$</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mrow> <mi mathvariant="bold-italic">R</mi> </mrow> <mrow> <mn mathvariant="bold">2</mn> </mrow> </msup> </math></EquationSource> </InlineEquation>) values ranging from 0.725 to 0.995, outperforming baseline algorithms at more than half of the stations. A paired one-tailed t-test confirms that the improvements in <InlineEquation ID="IEq2"> <EquationSource Format="TEX">$\varvec{R}^{\varvec{2}}$</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mrow> <mi mathvariant="bold-italic">R</mi> </mrow> <mrow> <mn mathvariant="bold">2</mn> </mrow> </msup> </math></EquationSource> </InlineEquation> over four baseline models are statistically significant (p < 0.01). Moreover, the model attains a mean absolute error (MAE) of 0.134 mm/h and a root mean square error (RMSE) of 0.350 mm/h, both substantially lower than all comparison algorithms. Further adaptive analysis across plains, mountainous, and hilly terrains verifies its robustness. Overall, MSLEFT-Transformer demonstrates significant advantages in accuracy, stability, and terrain adaptability, providing an effective solution for fine-grained precipitation prediction in complex meteorological scenarios.</p>

A MSLEFT-Transformer precipitation forecasting method fusing multi-scale spatial and temporal features

期刊论文
发表日期：

2026年2月14日

查看全文

Cong Li,
Tao Huang,
Pengping Lv,
Xupeng Ren

摘要

Accurate precipitation forecasting is essential for mitigating floods, optimizing agricultural planning, managing water resources, and ensuring transportation safety. However, it remains challenging due to complex nonlinear drivers and pronounced spatiotemporal variability. The primary prediction target in the main experiments of this study is the hourly precipitation (tp) from the ERA5 reanalysis dataset. The research framework is entirely constructed based on ERA5 data, and station-observed precipitation is not used as training or validation labels, thereby avoiding potential misunderstandings of “self-fitting” ERA5 variables. It should be emphasized that this design is not intended to simply reproduce ERA5 precipitation using its other variables; rather, it aims to learn a data-driven surrogate relationship between multivariate atmospheric states and point-specific precipitation responses. This framework can support applications such as rapid precipitation estimation, temporal downscaling, and serving as a computationally efficient alternative to direct reanalysis outputs. On this basis, to address this, we propose MSLEFT-Transformer, a framework that comprehensively considers the spatial characteristics, global trends, local details, and temporal dependencies of the precipitation processes, enabling precise modeling of precipitation dynamics. Experiments on 82 stations across Guangdong Province, China, show that MSLEFT-Transformer achieves coefficient of determination ( $\varvec{R}^{\varvec{2}}$ $R^{2}$ ) values ranging from 0.725 to 0.995, outperforming baseline algorithms at more than half of the stations. A paired one-tailed t-test confirms that the improvements in $\varvec{R}^{\varvec{2}}$ $R^{2}$ over four baseline models are statistically significant (p < 0.01). Moreover, the model attains a mean absolute error (MAE) of 0.134 mm/h and a root mean square error (RMSE) of 0.350 mm/h, both substantially lower than all comparison algorithms. Further adaptive analysis across plains, mountainous, and hilly terrains verifies its robustness. Overall, MSLEFT-Transformer demonstrates significant advantages in accuracy, stability, and terrain adaptability, providing an effective solution for fine-grained precipitation prediction in complex meteorological scenarios.