Machine Learning-Based Temperature Prediction for European Climate Change Adaptation: A ConvLSTM Approach Using ERA5 Reanalysis Data

The constant increase of extreme events in Europe is affecting the general well-being. Historical climate model predictions have been used previously, but they do not demonstrate the capability to forecast long-term and spatiotemporal patterns of extreme heat events, particularly at the continental scale. This study evaluates a machine learning-driven temperature prediction system with Convolutional Long Short-Term Memory (ConvLSTM) networks trained on ERA5 reanalysis data to generate multi-year-in-advance temperature predictions for European summer months (June, July, and August). 30 years of monthly mean ERA5 2-meter temperature was used in the model over Europe with three stacked ConvLSTM layers and then batch normalization as well as dropout regularization. The model was found to have a good predictive capacity for all the months of summer, validated with MAE from 1.5-2.5°C and RMSE from 2.0-3.5°C. Orographic influences in mountainous areas, latitudinal temperature gradients, pressure gradients, gradients, and Mediterranean heat dome patterns were the key climate features the model successfully emulated. Sustained warming trends for every month of summer were revealed with temporal analysis, with the most significant increase in July showing temperature anomalies of more than 2-3°C above the long-term average. The ConvLSTM approach suitably bridges the time gap between short-term numerical weather predictions and long-term climate projections and provides valuable capability for climate adaptation planning and early warning systems. The ability of the model to learn both interannual variability and long-term climate trends at the same time is a significant development over the traditional statistical approach.