Local Off-Grid Weather Forecasting with Multi-Modal Earth Observation Data

Abstract: Urgent applications like wildfire management and renewable energy generation require precise, localized weather forecasts near the Earth's surface. However, forecasts produced by machine learning models or numerical weather prediction systems are typically generated on large-scale regular grids, where direct downscaling fails to capture fine-grained, near-surface weather patterns. In this work, we propose a multi-modal transformer model trained end-to-end to downscale gridded forecasts to off-grid locations of interest. Our model directly combines local historical weather observations (e.g., wind, temperature, dewpoint) with gridded forecasts to produce locally accurate predictions at various lead times. Multiple data modalities are collected and concatenated at station-level locations, treated as a token at each station. Using self-attention, the token corresponding to the target location aggregates information from its neighboring tokens. Experiments using weather stations across the Northeastern United States show that our model outperforms a range of data-driven and non-data-driven off-grid forecasting methods. They also reveal that direct input of station data provides a phase shift in local weather forecasting accuracy, reducing the prediction error by up to 80% compared to pure gridded data based models. This approach demonstrates how to bridge the gap between large-scale weather models and locally accurate forecasts to support high-stakes, location-sensitive decision-making.

• The paper introduces a multi-modal GNN framework that integrates global ERA5 data with local MADIS observations to bridge global and local weather patterns.
• The paper uses a message-passing neural network to iteratively refine node features, achieving up to an 82% reduction in mean squared error compared to traditional models.
• The paper demonstrates practical improvements in off-grid forecasting for applications like wildfire management and renewable energy through rigorous testing over a 2019-2023 dataset.

An Analysis of Multi-modal Graph Neural Networks for Localized Off-grid Weather Forecasting

The paper presents a sophisticated approach to improving localized weather forecasting using multi-modal graph neural networks (GNNs). The primary focus is on enhancing prediction accuracy at off-grid locations, crucial for applications like wildfire management and renewable energy generation.

Key Methodological Insights

The authors propose an innovative methodology by integrating heterogeneous data through a GNN framework. This approach effectively bridges the gap between global and local weather models, traditionally based on gridded datasets like ERA5. Key to this model is the downscaling of global forecasts to localized predictions, leveraging local historical weather observations.

1. Heterogeneous Graph Construction: The model constructs a graph with nodes representing both gridded global data (ERA5) and off-grid weather stations (MADIS). This structure facilitates message passing, where each node aggregates information from its neighbors, thus integrating global atmospheric dynamics with local patterns.
2. Message Passing Neural Network (MPNN): By applying a message-passing mechanism, the model updates the node features iteratively, using both local and interpolated global data. This process enhances the prediction accuracy at specific off-grid locations.
3. Dataset and Evaluation: Spanning 2019-2023, the dataset comprises both ERA5 and MADIS data from across the Northeastern United States. The study reports substantial biases in ERA5, particularly in near-surface wind estimates. The proposed GNN model addresses these biases, outperforming other forecasting methods.

Numerical Results and Model Performance

The model demonstrates a significant decrease in mean squared error (MSE) compared to baseline methods. Notably, the GNN model achieved a 55.22% reduction in MSE compared to the best-performing MLP model and an 82.55% reduction compared to interpolated ERA5. These results underscore the model's capability in leveraging global and local data to produce accurately localized weather forecasts.

Implications and Future Directions

The implications of this research are multi-faceted. Practically, the ability to provide accurate localized forecasts can greatly enhance decision-making in critical fields such as agriculture and disaster management. Theoretically, the success of integrating multi-modal data into GNN frameworks could inspire further research into other domains requiring localized predictions.

Future developments could explore integrating additional data modalities, such as satellite imagery, which could potentially enhance the model's accuracy and applicability. Additionally, expanding the research scope to cover other regions and climates would provide a comprehensive evaluation of the model's scalability and robustness.

In conclusion, this paper offers a formidable contribution to the domain of weather forecasting, presenting a method that not only addresses current biases in reanalysis products like ERA5 but also sets a foundation for future advancements in AI-driven weather prediction.