A1870
Title: An innovative hybrid approach for forecasting spatio-temporal variables
Authors: Antonella Congedi - University of Salento (Italy) [presenting]
Sandra De Iaco - University of Salento (Italy)
Abstract: Climate change represents one of the main ecosystem problems, and accurate predictions of meteorological factors are essential to address extreme events, rising sea levels, loss of biodiversity, and to integrate energy-saving strategies with renewables. Forecasting methodologies include statistical and machine learning (ML) models, which capture the complex spatio-temporal dynamics of climate variables. An innovative hybrid approach combining spatio-temporal geostatistical techniques and artificial neural networks is proposed to analyze two meteorological datasets. The combination of kriging with ConvLSTM2D network is applied in the first case study to model solar radiation daily recorded in 2020 over a regular grid covering the Lombardy Region. In the second application, kriging combined with SparseConvLSTM network is used to model temperature weekly observed at meteorological stations of the Veneto Region from 2019 to 2022. A comparison with alternative methods is presented, and results highlight the advantages of using hybrid approaches instead of pure traditional or ML approaches. Future studies could extend the application of the proposed methodology to further fields or combine spatio-temporal kriging with other ML approaches.
