A Hybrid LSTM--Vision Transformer Architecture for Predicting HRRR Forecast Errors

Technical Summary: A Hybrid LSTM–Vision Transformer Architecture for Predicting HRRR Forecast Errors

Problem Statement
High-resolution Numerical Weather Prediction (NWP) systems, such as the High-Resolution Rapid Refresh (HRRR) model, exhibit systematic forecast errors often linked to unresolved planetary boundary layer (PBL) processes, convection, and terrain-induced circulations. Previous research (Evans et al., 2025) demonstrated that Long Short-Term Memory (LSTM) networks could successfully predict these errors using surface mesonet observations. However, that work identified a critical limitation: performance degraded significantly during periods of complex vertical atmospheric evolution, suggesting that surface-only data is insufficient to capture the atmospheric structures driving these errors.

Methodology
To address the limitation of surface-only inputs, the authors developed a hybrid LSTM–Vision Transformer (LSTM–ViT) framework. This architecture integrates temporal sequence learning from surface data with attention-based processing of vertically resolved atmospheric profiles.

• Data Sources: The model utilizes data from the New York State Mesonet (NYSM), combining:
  • Standard Network: 127 surface stations providing 16 meteorological variables (e.g., temperature, wind, humidity, precipitation).
  • Profiler Network: 17 stations equipped with microwave radiometers (MWR) providing vertical profiles (surface to 5 km) of temperature, humidity, liquid water content, and other variables at 10-minute intervals.
  • HRRR Forecasts: Three versions of the HRRR model (v2, v3, v4) serving as the baseline forecasts to be corrected.
• Architecture:
  • LSTM Encoder: Processes the temporal sequence of surface observations and HRRR forecast fields, compressing the recent atmospheric history into a latent state.
  • Vision Transformer (ViT) Encoder: Treats the MWR data as image-like arrays (Time × Height × Variable). It flattens the time-height dimensions into tokens and uses multi-headed self-attention to learn relationships between observations at different altitudes and times. Positional and temporal embeddings are added to preserve the structure of the vertical profile.
  • Fusion Mechanism: The latent representations from the LSTM and ViT encoders are concatenated and passed through a "fusion" Multi-Layer Perceptron (MLP). This MLP maps the combined features into the decoder space, applying LeakyReLU activation and dropout to prevent overfitting.
  • LSTM Decoder: Uses the fused representation to generate sequential forecast-error predictions for specific lead times (1–18 hours).
  • Post-Processing: A linear post-processing function is applied to tailor predictions to individual stations and variables.
• Training: Models were trained on data from 2018–2022, validated on 2023 data, and tested on 2024–2025 data. An outlier-focused loss function was employed to improve the representation of rare, high-impact events.

Key Contributions

1. Novel Hybrid Architecture: The paper introduces a unified framework that combines recurrent sequence modeling (LSTM) for temporal surface dynamics with attention mechanisms (ViT) for vertical atmospheric structure, specifically tailored for NWP error correction.
2. Integration of Vertical Profiling: It demonstrates the utility of incorporating microwave radiometer profiles into machine learning models for forecast error prediction, moving beyond the limitations of surface-only inputs.
3. Targeted Variable Prediction: The framework is explicitly trained to predict forecast errors for three critical variables: accumulated precipitation, 10-m wind speed, and 2-m temperature.

Results
The hybrid LSTM–ViT model was evaluated against the baseline LSTM and the raw HRRR forecasts across 17 profiler stations.

• Precipitation Errors: The LSTM–ViT framework achieved approximately a twofold increase in predictive skill relative to the baseline LSTM. It significantly reduced the over-prediction of small positive errors and improved the detection of negative errors (dry bias), although some asymmetry remained. The largest gains occurred at shorter lead times (1–12 hours) and during convective periods.
• Wind Errors: The hybrid model showed consistent improvements over the baseline LSTM, particularly at shorter lead times (1–10 hours). It reduced the systematic positive bias observed in the baseline LSTM and produced a more temporally balanced error representation. However, wind error prediction remained highly station-dependent due to complex local terrain and PBL interactions.
• Temperature Errors: The LSTM–ViT model improved upon the baseline LSTM at early to intermediate lead times (5–14 hours), particularly in capturing localized and temporally evolving error signatures. It performed better than the baseline during peak PBL overturning (solar noon) but showed reduced skill during nocturnal conditions compared to the baseline.
• Lead Time Dependency: Across all variables, the relative advantage of the LSTM–ViT over the baseline LSTM was most pronounced at shorter forecast lead times. As lead time increased, the predictive value of the observed vertical profile diminished, and the baseline LSTM sometimes demonstrated more stable skill in correcting persistent, large-scale biases.

Significance and Claims
The authors claim that this research provides a physically meaningful pathway for improving forecast-error prediction in operational NWP systems. By combining temporal sequence learning with vertically informed attention mechanisms, the model captures forecast-error characteristics associated with PBL processes that surface-based approaches miss.

The study concludes that while the LSTM–ViT framework offers enhanced guidance for forecasters regarding model bias and confidence—particularly in rapidly evolving atmospheric regimes—the two architectures (LSTM and LSTM–ViT) exhibit complementary strengths. The hybrid model excels at resolving localized, dynamically evolving processes at short lead times, whereas the standalone LSTM remains effective at capturing broader-scale, temporally persistent biases at extended lead times. The authors suggest that the greatest operational value may lie in using these as complementary components of a unified framework rather than competing methodologies.

The paper acknowledges limitations, including temporal displacement in predictions (shifting events by 1–2 hours), reduced skill in nocturnal regimes, and the need for improved data quality and observational coverage to ensure training stability across all stations. Future work is proposed to focus on Bayesian Neural Networks for uncertainty calibration and reliability.