Portal de Periódicos da CAPES

Autoencoder Framework for General Forecasting

2024; Springer Science+Business Media; Linguagem: Inglês

10.1007/978-3-031-61137-7_29

1611-3349

Dušan Fister, C. Peláez‐Rodríguez, L. Cornejo-Bueno, Jorge Pérez‐Aracil, Sancho Salcedo‐Sanz,

Computational Physics and Python Applications

Artificial intelligence backed forecasting systems, especially various types of autoencoders, are frequently used for short-term and medium-term weather forecasting. Sometimes, however, theirs validation is limited to specific weather occurrences, such as heatwaves or coldwaves, which limits the time period and location of forecasting significantly. We emphasise thorough model validation that validates the autoencoder's performance throughout the whole year for the whole possible area that autoencoder is trained for. Basic experimenting shows some limitations for proposed autoencoder, as at least one of the two benchmarks, i.e., the climate, overperforms performance of proposed autoencoder on average basis with regard to the utilised two stage autoencoder structure and suggests that further modifications to generalise the utilised autoencoder are needed.