CN-AEBench is a comprehensive multi-source atmospheric and environmental dataset integrating ground meteorological observations, environmental monitoring, and ECMWF IFS data. This visualization platform demonstrates the dataset's reliability and consistency through three representative extreme weather events across different climate zones in China, validating the data quality and applicability for atmospheric and environmental research and machine learning applications.
Severe Fog-Haze Event
Analysis of severe winter fog-haze pollution event in North China Plain, demonstrating the dataset's capability to capture complex atmospheric processes and aerosol-meteorology interactions.
View Visualization βExtreme Rainstorm & Typhoon
Investigation of extreme precipitation & typhoon event in South China coastal region, showcasing the dataset's ability to capture severe convective weather systems and their environmental impacts.
View Visualization βDust Storm Event
Examination of severe dust storm at Taklamakan Desert edge, validating the dataset's performance in extreme arid conditions and dust aerosol monitoring capabilities.
View Visualization βWe selected four representative regions across China for comprehensive evaluation:
Characterized by arid climate, complex terrain, and frequent dust events influenced by the Gobi Desert.
Located in the Yunnan-Guizhou Plateau with mild climate year-round and unique monsoon patterns.
Features extreme continental climate with harsh winters and significant seasonal variations.
China's economic hub with subtropical monsoon climate, urban heat island effects, and typhoon influences.
Each region is trained and tested independently to ensure robust regional performance.
Validation Set: Days 1-10 of Oct 2023, Jan/Apr/Jul/Sep/Dec 2024, Mar/Jun 2025
Test Set: Days 14-end of the same months
Training Set: All remaining data
Overall ratio - Training:Validation:Test = 7:1:2
MAE (Mean Absolute Error) and Corr (Correlation Coefficient) for basic accuracy assessment
Soft-DTW (Soft Dynamic Time Warping) measures shape similarity while allowing temporal shifts, crucial for capturing evolutionary patterns
Enhanced RQE (Relative Quantile Error) variants specifically designed for extreme weather prediction. See paper for details.
Configuration: 24 steps β 1\4\12\24 steps
Evaluation: Lead times at 1h (nowcasting), 4h, 12h, 24h (short-term), and overall performance
Configuration: 48 steps β 1\24\48\72\96\120 steps
Evaluation: Lead times at 1h (nowcasting), 24h, 48h, 72h, 96h, 120h (extended range), and overall performance
Experiments are conducted for both TSF (Time Series Forecasting) and ST-GNN (Spatio-Temporal Graph Neural Networks), with additional ablation studies on covariate analysis to assess multivariate prediction capabilities.