HAPPY

• Machine learning-based weather prediction.

The model was trained using data from 2000 to 2024. First, the original ERA5 daily and monthly datasets were interpolated to a 1.5° grid. Weekly data were then generated by averaging the daily data. Based on the prediction date (the Monday three weeks ahead), the input and output datasets were organized into pairs. The input data include: Monthly forecast data, Weekly data of predictive variables for the previous 20 weeks, Weekly upper-level data for the previous 10 weeks, and Elevation data. The output data consist of daily predictions for seven specific days: Monday, Wednesday, Friday, and Sunday of the third week ahead, and Tuesday, Thursday, and Saturday of the fourth week ahead. The monthly forecast model was trained using monthly data from 1940 to 2024. It takes the previous 15 months of historical data and other surface variables as input and predicts the monthly mean two months ahead. The monthly forecast data used as input correspond to the predicted month of the target date.

TAS: The input includes 20 consecutive weeks of historical daily 2-meter temperature (tas) and 10 consecutive weeks of historical daily geopotential height at 200 hPa, 300 hPa, and 500 hPa. The monthly forecast model (same as for MSLP) requires 15 weeks of historical data for U10, V10, SST, TAS, and MSLP. MSLP: The input includes 20 consecutive weeks of historical daily mean sea level pressure (mslp) and 10 consecutive weeks of historical daily geopotential height at 500 hPa and 850 hPa, specific humidity at 700 hPa, divergence and potential vorticity at 900 hPa. The monthly forecast model (same as for TAS) requires 15 weeks of historical data for U10, V10, SST, TAS, and MSLP. TP: The input includes 20 consecutive weeks of historical daily total precipitation (tp) and 10 consecutive weeks of historical daily geopotential height at 200 hPa, 300 hPa, and 500 hPa, specific humidity at 700 hPa, and cloud cover at 800 hPa. The monthly forecast model requires 15 weeks of historical data for U10, V10, SST, T2M, MSLP, and TP.

Only ERA5 reanalysis data.

Model Framework The overall prediction framework consists of two stages. First, the monthly values of each variable are predicted using more than one year of surface variable data. Then, using over three months of upper-level variable data, the monthly forecast results, and elevation data, the model predicts daily values for seven specific days: Monday, Wednesday, Friday, and Sunday of the third week ahead, and Tuesday, Thursday, and Saturday of the fourth week ahead. The weekly mean is then calculated from the predicted data for the third and fourth weeks, and the quintile category is determined for submission. Each target variable (TAS, MSLP, and TP) is trained with its own independent model. Monthly Forecast Model Input data: Surface variables from the previous 15 months. Model architecture: A hybrid structure combining a Convolutional Neural Network (CNN) and a Temporal Convolutional Network (TCN). Output data: Single-variable (TAS, MSLP, or TP) forecasts two months ahead. Main Prediction Model Input data: Weekly data of the target variable from the previous 20 weeks, Weekly upper-level data from the previous 10 weeks, Monthly forecast data, and Elevation data. Model architecture: For near-surface air temperature (TAS), six deep learning architectures were tested. All demonstrated stable convergence during training (with overall ACC gradually increasing and reaching above 0.75). The architectures include: Convolutional Neural Network (CNN) CNN + GRU Hybrid Model Transformer–CNN Hybrid Model CBAM-based model combining spatial attention mechanisms and residual blocks Residual Squeeze-and-Excitation + CNN Model Residual U-Net Model In comparison, for mean sea level pressure (MSLP) and total precipitation (TP), only the Residual U-Net architecture achieved stable and effective performance during training. Output prediction: Daily forecasts for seven specific days—Monday, Wednesday, Friday, and Sunday of the third week ahead, and Tuesday, Thursday, and Saturday of the fourth week ahead. Post-processing For each model, the predicted data for the third and fourth weeks are averaged, and the resulting mean is compared with the historical climatological quartiles to determine the quintile category. For TAS, six models were used to generate historical climatological quartiles and corresponding quintile predictions. The final TAS result is obtained by averaging the quintile outputs from all six models. For MSLP and TP, the quintile prediction is derived directly from the single Residual U-Net model.

No, we plan to start writing the related content after gaining better control over the model’s performance.

No.

The model was trained entirely on a gaming laptop. Although the computational power was limited, which constrained both data input volume and model complexity, the model was successfully completed and submitted. The reason for firmly choosing to train on a gaming laptop was twofold: first, it ensured stable computational access throughout the entire competition period; second, I did not have access to large or fast computing resources, so I aimed to develop models optimized for low computational capacity. When training became slow or unresponsive, I simplified the model architecture, reduced the number of input variables, and repeated the process several times until it could run smoothly.

The model is still in its early development stage, and many aspects are being adjusted and optimized. For this submission cycle, it became clear that the conversion from the predicted true values of sea level pressure and precipitation to their quintile categories could be improved by using predicted historical values to calculate climatological means—similar to how it was done for near-surface air temperature. In future work, I plan to refine this conversion process to achieve better evaluation scores, and to carefully analyze the prediction performance to identify issues and further optimize the model.

The main motivation and foundation for developing this model came from testing monthly-scale predictions using ERA5 single-level data. By predicting tas one month ahead based on the previous 15 months of historical data, the model achieved excellent RMSE and ACC performance, which gave me strong confidence to proceed with subsequent developments.

This team has chosen to keep its participants anonymous.