A Case Study on Targeting a Practically Simple Pre-Training Method to Enhance Machine Learning-Based Climate Prediction Models

This study introduced a pre-training method for machine learning-based climate prediction models. The method leverages the advantage of climate events (some theoretical knowledge) to address their limitation (small sample size). It consists of the following steps: generating artificial samples via composite analysis of high and low anomaly events, pre-training predictive models with these samples, and selecting an optimal pre-trained model that most closely matches the observational training set from numerous repeated experiments with only the model’s random number seeds being varied. Sensitivity experiments demonstrate that this pre-training method not only substantially improves predictive skills but also significantly reduces prediction instability. This simple and practical pre-training method is applicable not only to the climate prediction events in this study but also to all climate events for which composite analysis is applicable.