Work

Probabilistic seasonal forecasts with conditional variational autoencoders

Year: 2024
Type: dissertation
Abstract: While traditional numerical models are the primary source of seasonal predictions, they are computationally expensive to run and induce drifts and biases in their outputs. Machine learning offers an alternative by learning complex patterns from data, potentially overcoming these limitations. This project investigates the use of generative neural networks to create computationally efficient seasonal forecast models. These models can incorporate randomness (stochasticity) and generate ensemble forecasts, mimicking the capabilities of established methods at a lower cost. We implement a Conditional Variational AutoEncoder (cVAE) approach, previously successful in six-month predictions. Here, we focus on predicting summer’s average temperature (three months) based on May’s climate data. The network learns from data generated by the sixth Coupled Model Intercomparison Project (CMIP) and is evaluated against real-world observations from the fifth generation of ECMWF reanalysis for the global climate and weather (ERA5) reanalysis dataset. To isolate the method’s performance, we employ detrending, e.g. removing long-term trends from the data using a reference period independent of the testing data. Finally, we compare our model’s predictions to established baselines commonly used in seasonal forecasting for a comprehensive evaluation. Our approach managed to deliver seasonal forecasts that matched the performance of SEAS5, the leading model, across various regions. Although the general proficiency of our model falls short of SEAS5, it surpasses standard benchmarks like climatology and persistence. Consequently, our predictions provide a rapid and computation-light estimate of the seasonal climate, retaining the skillfulness from numerical methods to a certain degree.
Author: Peraza González, Alejandro
Language: English
Open Access status: green
OpenAlex ID: W7005289610
Cites: 49
Cited by:
Related to: 10