Lessons Learned From Building And Operationalizing ML Weather Prediction Models

In recent years, data-driven weather prediction (DDWP) models that rely on ML have shown significant promise in providing more accurate forecasts when compared to Numerical Weather Prediction (NWP) models, and at a lower cost. The vast majority of such ML models still exist primarily in a research environment, and the use of DDWPs in operational forecasting is still in its infancy. Using ML weather models in operational forecasting opens up opportunities to validate such models in the real world, and improve the models based on user feedback. At Excarta, we have operationalized one of the world’s first ML-driven weather prediction models, providing 15 day forecasts to consumers, updated every 6 hours. In this talk, we share our experience in developing and operationalizing a medium-range ML weather prediction model, covering questions relating to the tools technologies used, setting up efficient research and development pipelines, establishing rigorous validation procedures, and incorporating customer feedback. We hope our lessons can benefit others exploring the use of AI and ML in weather-related decision making.