On the application of generative modelling for seasonal climate predictions

Reliable probabilistic predictions at the seasonal time scale are critical for key societal sectors such as agriculture, energy, and water management. Current operational approaches face significant challenges: General Circulation Models (GCMs) are computationally expensive and often limited by low spatial resolution and model deficiencies. At the same time, traditional statistical methods struggle due to significant modelling assumptions, such as linearity or homoscedasticity. Generative models emerge as a cost-effective, promising alternative, offering the potential to model complex nonlinear climate dynamics inherently probabilistically and at a reduced computational cost. Yet, training these algorithms on the short span of current reanalysis datasets results in almost certain overfitting due to the imbalance between trainable parameters and available training samples.

In this context, the present study compares the effectiveness of different generative methodologies in predicting gridded fields of temperature and rainfall seasonal anomalies. The predictions cover all four seasons and are initialised one month before the start of the season, aligning with most climate services providers. We employ climate model output from CMIP6  and CEMS-lens2 during training and ERA5 reanalysis data during testing to circumvent the short span of current reanalysis and observational datasets. We analyse the method's performance in predicting interannual anomalies beyond the climate change-induced trend. We show that the model's ensemble generation capabilities allow it to provide diverse ensemble members, allowing the derivation of relevant probabilistic information and potentially reliable predictions. While climate change trends dominate the skill of temperature predictions, additional skill over the climatological forecast in regions influenced by known teleconnections is found. We reach similar conclusions based on the validation of precipitation predictions.

This work further demonstrates the effectiveness of training generative models on climate model output for seasonal predictions, providing skilful seasonal climate predictions beyond the induced climate change trend at time scales and lead times relevant for user applications, motivating further research.