Exploring the Potential of Data-Driven AI Models for Operational Weather Forecasting in Finland

Recent years have witnessed substantial advancements in the development of data-driven artificial intelligence (AI) models aimed at weather forecasting. These models have demonstrated remarkable capabilities, showcasing promising results when compared against conventional Numerical Weather Prediction (NWP) models. This study aims to explore the potential of AI-based forecasts in daily operational weather forecasting, despite the inherent limitations in parameter scope and spatial resolution when compared with prevalent Limited Area Models employed by National Meteorological Services (NMSs).

Finnish Meteorological Institute (FMI) has been running the Pangu-Weather AI model, developed by Huawei, in an operational-like configuration since the summer of 2023. The Github repository provided by ECMWF (ecmwf-lab) was used to build the configuration to run the Pangu-Weather model (the trained parameters of Pangu-Weather are made available under the terms of the BY-NC-SA 4.0 license). The Pangu-Weather forecasts are initialized using ECMWF HRES model analysis fields, and the forecasts are generated with a spatial resolution of 0.25 degrees and temporal resolution identical to ECMWF HRES.

For verification and validation purposes the forecasted fields from Pangu-Weather are made available to FMI forecasters through meteorological Smartmet and Geoweb workstations and point forecasts are stored for verification against observations from Finnish Synop stations.

The primary objective is to assess the physical realism of Pangu-Weather's forecasted weather scenarios for Finland. We consider the surface parameters temperature, mean sea level pressure, wind speed and direction and investigate how well these fields compare with ECMWF HRES forecasts. Point forecasts are also compared with the post-processing methods used by FMI in operational weather forecasting, including Model Output Statistics (MOS) based on ECMWF HRES and FMI Model Blend, which is a consensus model that incorporates inputs from various NWP models.

The preliminary results indicate that Pangu-Weather forecast fields are in general physically realistic, displaying a considerable degree of consistency in pressure and wind fields. Some deviation in the forecasted wind and pressure fields is noticeable in the positioning of low-pressure system centers. Overall, Pangu-Weather maintains a consistent weather development over consecutive lead times, although it employs a hierarchical temporal aggregation approach by integrating multiple AI models for distinct lead times. Verification results against Synop observations in Finland show that for temperature forecasts, Pangu-Weather produces competitive results compared to ECMWF HRES and post-processing methods. However, for wind speed forecasts, Pangu-Weather tends to slightly underforecast the observed wind speeds, possibly due to the coarser resolution and orographic complexities.

This research underscores the potential of data-driven AI models in operational weather forecasting while also identifying opportunities for refinement and optimization. Data-driven AI is capable of generating realistic scenarios for the forecasted parameters, offering insights into weather trends. The outcomes of this study contribute to the ongoing dialogue surrounding AI's role in meteorological forecasting, offering valuable information for both meteorological researchers and practitioners.