jeffliulab/weather-forecasting-v1

Weather Forecasting

CNN-based models that forecast local weather conditions 24 hours ahead from HRRR (High-Resolution Rapid Refresh) spatial snapshots, predicting conditions at the Jumbo Statue location at Tufts University (42.408N, 71.120W).

Task

• Input: Spatial weather snapshot at time t, shape (450, 449, 42) in bfloat16
• Output: 6 weather variables at t+24h for the target grid point
• Training data: 2018-2019 HRRR hourly data
• Validation data: 2020

Prediction Targets

Variable	Unit
TMP@2m_above_ground	K
RH@2m_above_ground	%
UGRD@10m_above_ground	m/s
VGRD@10m_above_ground	m/s
GUST@surface	m/s
APCP_1hr_acc_fcst@surface	mm

Evaluation Metrics

• RMSE for TMP, RH, UGRD, VGRD, GUST
• Conditional RMSE for APCP (only when true value > 2 mm)
• AUC for binary precipitation label (APCP > 2 mm)

Models

Checkpoint	Epochs	val_loss	AUC	Notes
cnn_baseline_27ep-baseline/best.pt	27 (epoch 15)	0.6207	0.739	Baseline run

Architecture (cnn_baseline)

ResNet-style 2D CNN:

• Stem conv (stride=2) on (B, 42, 450, 449) input
• 6 Residual blocks with base_channels=64
• AdaptiveAvgPool(1) -> FC head (512->128->6)
• ~2.3M parameters

Usage

Code

jeffliulab/real_time_weather_forecasting