Improving Hourly Heavy Rainfall Forecasts With Intensity‐Aware Deep Learning

Abstract Heavy rainfall is a key trigger of urban pluvial flooding, and accurate rainfall forecasting is essential for early warnings and emergency response. However, hourly station rainfall series are typically characterized by severe zero inflation and heavy‐tailed extremes. Conventional regression losses that minimize average error tend to overlearn the dominant dry and light‐rainy cases and produce overly conservative forecasts. To address these issues, we propose a Zero‐Inflated Level‐Weighted Asymmetric (ZILA) loss and incorporate it into a long short‐term memory (LSTM) model to account for zero inflation and asymmetric costs under heavy rainfall. Applied to hourly rainfall forecasting in Hong Kong, ZILA‐trained LSTM delivers the best performance in process consistency, peak‐intensity estimation and threshold‐based event detection. At the 1‐hr lead time, compared with LSTMs trained using MAE, MSE, and Tweedie losses, it achieves the highest Kling–Gupta Efficiency, reduces peak RMSE for heavy‐rainfall events with intensities greater than 10, 30 and 50 mm/hr by 28%–47%, and improves heavy rainfall event‐detection skill by 3.2–8.3 fold. ZILA maintains a forecasting advantage as rainfall severity and forecast lead time increase, with predictive skill degrading more slowly than baseline losses despite widening disparities across station‐information tiers. The results also reveal that richer station information becomes more important as rainfall severity increases, while less‐information stations lose warning skill faster with lead time. Cross‐validation confirms robustness across data splits and evaluation settings. These improvements highlight the practical value of intensity‐aware loss design for heavy rainfall early warning in Hong Kong and its broader methodological significance for extreme‐rainfall prediction.

Authors

Institutions

Publication Details

Journal
Journal of Geophysical Research Atmospheres
Published
2026-11-14
DOI
https://doi.org/10.1029/2026jd047775
Primary Topic
Hydrological Forecasting Using AI
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Improving Hourly Heavy Rainfall Forecasts With Intensity‐Aware Deep Learning

Jiachuan Yang, Y Q Xia
Journal of Geophysical Research Atmospheres
Hydrological Forecasting Using AI
article

Improving Hourly Heavy Rainfall Forecasts With Intensity‐Aware Deep Learning

Jiachuan Yang, Y Q Xia
article en

Abstract

Abstract Heavy rainfall is a key trigger of urban pluvial flooding, and accurate rainfall forecasting is essential for early warnings and emergency response. However, hourly station rainfall series are typically characterized by severe zero inflation and heavy‐tailed extremes. Conventional regression losses that minimize average error tend to overlearn the dominant dry and light‐rainy cases and produce overly conservative forecasts. To address these issues, we propose a Zero‐Inflated Level‐Weighted Asymmetric (ZILA) loss and incorporate it into a long short‐term memory (LSTM) model to account for zero inflation and asymmetric costs under heavy rainfall. Applied to hourly rainfall forecasting in Hong Kong, ZILA‐trained LSTM delivers the best performance in process consistency, peak‐intensity estimation and threshold‐based event detection. At the 1‐hr lead time, compared with LSTMs trained using MAE, MSE, and Tweedie losses, it achieves the highest Kling–Gupta Efficiency, reduces peak RMSE for heavy‐rainfall events with intensities greater than 10, 30 and 50 mm/hr by 28%–47%, and improves heavy rainfall event‐detection skill by 3.2–8.3 fold. ZILA maintains a forecasting advantage as rainfall severity and forecast lead time increase, with predictive skill degrading more slowly than baseline losses despite widening disparities across station‐information tiers. The results also reveal that richer station information becomes more important as rainfall severity increases, while less‐information stations lose warning skill faster with lead time. Cross‐validation confirms robustness across data splits and evaluation settings. These improvements highlight the practical value of intensity‐aware loss design for heavy rainfall early warning in Hong Kong and its broader methodological significance for extreme‐rainfall prediction.

Journal of Geophysical Research AtmospheresVol. 131(22)
Hong Kong University of Science and Technology (HK)
Hydrological Forecasting Using AI
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon