Seasonal Forecasting Using Machine Learning Algorithms for the Continental Europe

The ability to accurately predict seasonal and sub-seasonal weather patterns is of great significance in climate modeling, as extreme weather events and climate change become more prevalent. Furthermore, current climate models require extensive computational resources to generate monthly forecasts. Recent advancements in machine learning, accessibility of the vast amount of data, and efficiency of the ML models motivate the use of machine learning based approaches for seasonal forecasting.In this study, we present a sub-seasonal weather forecast model using a UNet based deep learning architecture that enables the learning of long-range spatiotemporal information. Our model is trained with surface air temperature data obtained from 10 different state-of-the-art CMIP6 model output, and finetuned using ERA5 atmospheric reanalysis dataset to increase the generalization in forecasting with real weather modeling data. We focus on the monthly temperature forecast of continental Europe and compare our results with physics-based climate models and ML methods. We evaluate our models using Rooted Mean Square Error (RMSE). Moreover, we analyze the effect of incorporating ancillary data such as topography maps into our model. Our method outperforms linear regression techniques by a high margin. We have found that using the temperature information from the preceding 3 to 4 years of data can improve the performance.We design different experimental settings to arrange monthly historical temperature information given to the deep learning model. In addition to the use of historical data consequently, we investigate the performance of periodical arrangements.