Improving data-driven global weather prediction using deep convolutional neural networks on a cubed sphere (2003.11927v1)

Abstract: We present a significantly-improved data-driven global weather forecasting framework using a deep convolutional neural network (CNN) to forecast several basic atmospheric variables on a global grid. New developments in this framework include an offline volume-conservative mapping to a cubed-sphere grid, improvements to the CNN architecture, and the minimization of the loss function over multiple steps in a prediction sequence. The cubed-sphere remapping minimizes the distortion on the cube faces on which convolution operations are performed and provides natural boundary conditions for padding in the CNN. Our improved model produces weather forecasts that are indefinitely stable and produce realistic weather patterns at lead times of several weeks and longer. For short- to medium-range forecasting, our model significantly outperforms persistence, climatology, and a coarse-resolution dynamical numerical weather prediction (NWP) model. Unsurprisingly, our forecasts are worse than those from a high-resolution state-of-the-art operational NWP system. Our data-driven model is able to learn to forecast complex surface temperature patterns from few input atmospheric state variables. On annual time scales, our model produces a realistic seasonal cycle driven solely by the prescribed variation in top-of-atmosphere solar forcing. Although it is currently less accurate than operational weather forecasting models, our data-driven CNN executes much faster than those models, suggesting that machine learning could prove to be a valuable tool for large-ensemble forecasting.

• The paper improves forecast accuracy by transforming atmospheric data onto a cubed sphere grid to reduce distortion.
• It employs an enhanced U-Net CNN with skip connections and distinct weight sets for equatorial and polar regions.
• The model achieves lower RMSE and higher ACC in short-to-medium forecasts, indicating potential for efficient real-time ensemble predictions.

An Analysis of Deep Convolutional Neural Networks for Global Weather Prediction

This paper presents a sophisticated framework for enhancing global weather prediction through the application of deep convolutional neural networks (CNNs). Leveraging a cubed sphere grid for data representation, the authors improve upon previous methods by incorporating volume-conserving mapping to minimize distortion and employing a novel CNN architecture. This architecture utilizes a multi-step prediction sequence to optimize the loss function over extended prediction horizons, enhancing the stability and realism of long-term forecasts.

Key Innovations and Methodology

At the core of the methodology is the transformation of global atmospheric data from a traditional latitude-longitude grid to a cubed sphere grid. This transformation reduces grid distortion and facilitates convolution operations while maintaining natural boundary conditions. By partitioning the CNN model to utilize distinct sets of learned weights for equatorial versus polar faces of the cubed sphere, significant improvements in predictive accuracy and stability are achieved.

The CNN architecture is based on the U-Net configuration, recognized for its efficacy in image segmentation tasks, which allows for the extraction of multi-scale spatial features through a combination of convolutional layers, pooling, and up-sampling processes. This paper further enhances the U-Net with skip connections, which preserve high-resolution spatial information, thereby increasing the efficiency of long-range forecasts.

The authors incorporate a sequence prediction approach, training the CNN to minimize error across multiple predictive time steps. This anticipatory strategy improves model performance over medium-range timescales, reflecting a significant advancement beyond their prior work, which was confined to northern hemispheric forecasts.

Results and Evaluation

The proposed model demonstrates substantial improvements over traditional benchmarks, particularly for short- to medium-range weather forecasts. Against models like persistence and climatological benchmarks, the CNN-based forecasts show superior performance. Compared to a T42 spectral resolution numerical weather prediction (NWP) model, the CNN approach yields significantly lower root-mean-square errors (RMSE) and higher anomaly correlation coefficients (ACC) for the atmospheric variables of interest, namely 500 hPa geopotential height and 2-meter temperatures.

In extended forecasts, though it underperforms against state-of-the-art high-resolution operational NWP systems, the CNN model still produces realistic seasonal cycles. It successfully predicts surface temperature patterns using limited input variables, and, importantly, provides operational insights faster than traditional data-heavy models. Its computational efficiency suggests a noteworthy potential for use in real-time large-ensemble simulations where rapid generation of multiple forecasts is essential.

Implications and Future Directions

The model’s ability to produce long-term stable and realistic predictions on a cubed sphere grid indicates promising applications in operational ensemble forecasting and potential for integration with existing NWP systems. Future developments might focus on improving vertical resolution and expanding the input variable set to encompass more atmospheric dynamics, potentially increasing forecast accuracy for extreme weather patterns or seasonal anomalies.

A significant implication of this work lies in its potential to reduce computational costs associated with high-resolution ensemble predictions, offering a computationally efficient alternative for generating extensive probabilistic forecasts. Ongoing research could explore the integration of this model with hybrid data-driven and dynamical models, which might bridge the accuracy gap between purely data-driven methods and comprehensive NWP systems.

In essence, by pushing the boundaries of data-driven weather prediction through innovative use of CNNs on a cubed sphere grid, this research presents a vital step forward in leveraging machine learning for meteorological applications. The findings here underscore the transformative potential of AI in improving the efficiency and scalability of global weather forecasting.