Blog & Daily weather news | Infoplaza

AI and weather forecasting: regional higher-resolution weather models

Written by Christian Versloot | Oct 31, 2024 7:29:44 AM

In the last few years, there have been significant advances in AI-based weather forecasting. These advances, with new models like GraphCast, Pangu-Weather and FourCastNet being used worldwide, have mostly focused on global weather forecasting. However, global weather modelling – whether AI-based or NWP-based – comes with various drawbacks. That’s why regional NWP-based weather models, such as Harmonie, ICON-D2 and AROME, have been integral to weather forecasting. Recently, efforts to create AI-based regional weather models have shown initial results and yielded valuable insights. In this article, we will explore these developments.

Firstly, we’ll share the differences between global and regional weather modelling. What can you expect from a global weather model, and what are its limitations? Why are regional models particularly useful for specific cases? Knowing the answers to these questions leads us to the concept of boundary conditions – an important part of regional weather modelling. Then, we’ll illustrate recent advances in AI-based weather forecasting and provide some future directions.

Global vs regional weather modelling

In numerical weather prediction (NWP), weather models like the GFS and ECMWF models are well-known. These models are global models, which means – as the name suggests – that they produce outputs for the entire world. While the benefits of such models are directly reflected in the global weather forecasts issued by Infoplaza, there are also drawbacks to global weather models. Computational limitations are one of them.

Recall that NWP is computationally intensive, because model output is derived from physics equations computed from an initial state. Consequently, the resolution of global models must be limited in order to produce forecasts in acceptable time frames. For example, the GFS model displayed in the image below, has an approximate resolution of 12 by 12 kilometers. In other words, each cell in the forecast contains a single value for temperature, wind speed, precipitation, and so forth, for an area of 144 square kilometers. This means that while larger-scale weather patterns such as large storm systems can be captured, finer-scale weather events like thunderstorms are represented less adequately. Furthermore, global weather models typically produce forecasts in multi-hour time steps, such as 3 to 6 hours, resulting in a very coarse-grained weather forecast.

To mitigate these drawbacks, a wide range of regional (or limited-area) models are available next to global models. When issuing weather forecasts, Infoplaza makes use of these models, and generates the best possible mix of these models for each location to the most accurate weather forecast. A regional weather model, like a global model, produces weather forecasts, but for a much smaller domain. For example, the Harmonie model predicts weather for large parts of northwest Europe; the AROME model focuses on France and surrounding countries, and the ICON-D2 weather model focuses on Germany. Typically, these models are produced by meteorological institutions and consortia with a regional interest. They trade-off global forecast availability for a higher resolution in both space and time. For example, the Harmonie model displayed below has an approximate resolution of 2 by 2-kilometer and is available with hourly time steps. However, as mentioned, the domain is limited, as is the forecast horizon: while GFS extends many days into the future, the predictions of Harmonie don’t go beyond a few days only.

Figure 1: Left: global output of the GFS weather model at 12-kilometer resolution. Right: regional output of the KNMI Harmonie weather model at 2-kilometer resolution. Source: I’m Weather

Boundary conditions

An important consideration for regional weather modelling is the concept of boundary conditions. At the boundaries of model domains, the models are blind to the outside the boundary, as indicated by the gray area in the image above. For example, suppose a storm system is moving across the Atlantic towards Europe. Its low pressure, high winds and significant precipitation amounts will cause severe weather in Ireland, the United Kingdom and eventually the Netherlands. If we would just let the regional weather model run without this information, it might not model this weather system at all, leading to a very poor forecast. By forcing the finer grid to take this information from a coarse, global model as boundary conditions, we may make the model aware of these incoming weather conditions. Applying boundary conditions via boundary forcing involves using the outputs from another weather model and incorporating them into the parts outside the regional model’s domain. Our wave modellers, for example, who are heavy users of small-scale weather models, use these coarser-grained models to steer the boundaries of their model domains to incoming weather conditions.

Figure 2: multiple grids nested into each other. The green grid would use the blue grid as its boundary conditions; the blue grid would use the black grid as its boundary conditions. In this case, the arrows suggest unidirectional boundary forcing: the coarser grid is used as boundary conditions for the finer grid. Bidirectional forcing is also possible, where computations in the finer grid are reflected in the coarser grid.

From global to regional AI-based weather models

AI-based weather models are typically global. The first and second generation of models were trained on ERA5 data, which is a global reanalysis dataset. In other words, even though their performance is competitive to global NWP models and sometimes even surpasses those models, they are global. They thus suffer from the same limitations as NWP-based global models: limited resolution in space and time, limited resolving of smaller-scale weather phenomena and thus limited representation of actual, surface-level weather.

After these generations of AI models were released, researchers started thinking – would it be possible to use the learnings from these global models and develop a regional AI-based weather model instead? The answer is yes. It is possible to develop regional AI-based weather models, and the first ones have emerged with surprisingly good performance.

One of the first AI-based limited-area weather models is Neural-LAM, created by Oskarsson et al. (2023). Inspired by GraphCast, which uses graph neural networks for weather forecasting, they first adapted GraphCast to a limited-area setting. After finding issues and changing the approach, a functional model was created by training it with 2021-2023 forecasts from the MetCoOp Ensemble Prediction System (MEPS), of which the operational area covers the Nordics. Boundary forcing was performed using data from the MEPS dataset, too. Once trained, generating a 57-hour forecast only takes 1.5 seconds on a single machine equipped with a powerful graphics card. While no extensive comparison with were made with non-AI weather models, the unrealistic smoothing commonly observed in AI-based weather forecasts is observed here, too – and listed as a future development.

Figure 3: a forecast of net longwave solar radiation flux generated by the Neural-LAM weather model. In the image, (a) represents the prediction target, while (b) represents the forecast generated by the best candidate AI weather model tested in the Neural-LAM project.

In a collaboration with ECMWF, Met Norway since adapted the model architecture to be interoperable with its AIFS model. Now starting with a global model, a regional grid with a higher resolution was nested into it, in a so-called stretched-grid approach. According to Nipen et al. (2024), “[t]he goal of this stretched-grid approach is for weather systems to move seamlessly from the global domain into the regional domain and out again without any explicit treatment of the boundary between the domains”. In other words, it is a model that is trained simultaneously from both global and high-resolution analysis data without explicit boundary forcing. Compared to the resolution of AIFS – approximately 30 by 30 kilometers – the resolution of the Nordics domain is 2.5 by 2.5 kilometers in this new model.

Whenever such models are created, it is important to verify them. The model was compared with the MEPS system that is used operationally by Met Norway, as well as the IFS model (i.e. the ECMWF model) at 0.1-degree resolution. Additionally, point observations from Met Norway’s SYNOP stations were used to compare the model to, after some treatment to make them ready for comparison. Overall, for many surface-level variables, significantly lower Root Mean Squared Error (RMSE) values are achieved when compared to the NWP models. However, the model is not competitive in all cases. Still, these results demonstrate that an AI-based regional weather model is possible, possibly warranting its use in operational weather forecasting if challenges are overcome.

Figure 4: model output from the model created by Nipen et al. (2024). Lower-resolution weather beyond the model domain is resolved with a higher resolution within the domain, leading to more detailed weather forecasts. Source: ECMWF.

Future directions

Many interesting developments are currently being undertaken in the realm of AI-based weather forecasting, both in the context of regional models and other kinds of challenges. Recently, ECMWF has commenced coordinating an EU-level project called WeatherGenerator which aims to use machine learning in novel ways for weather forecasting. According to ECMWF, “[t]he fundamental idea of the WeatherGenerator is to build one machine learning tool that can be used and adapted for a large number of specific tasks”. In other words, it will be a so-called foundation model, like the ChatGPT-like models that are able to improve many language-related tasks, including programming.

What’s next: a recap and outlook

For now, this is the final main article of our series on AI and weather forecasting. The next article will provide a recap, summarizing the advancements discussed so far and offering an outlook on future directions. Thank you for reading this series and we hope that you have been able to improve your understanding of AI-based weather modelling! If there are any questions, feel free to get in touch via our contact page.

References

ECMWF. (2024). Data-driven regional modelling. https://www.ecmwf.int/en/about/media-centre/aifs-blog/2024/data-driven-regional-modelling

Met Norway. (2024). MetCoOp. https://www.met.no/en/projects/metcoop

Nipen, T. N., Haugen, H. H., Ingstad, M. S., Nordhagen, E. M., Salihi, A. F. S., Tedesco, P., ... & Chantry, M. (2024). Regional data-driven weather modeling with a global stretched-grid. arXiv preprint arXiv:2409.02891.

Oskarsson, J., Landelius, T., & Lindsten, F. (2023). Graph-based Neural Weather Prediction for Limited Area Modeling. arXiv preprint arXiv:2309.17370.