Spatio-temporal forecasting models typically generate predictions only at locations where sensors are physically installed, limiting their applicability in scenarios requiring full spatial coverage. Many real-world applications, however, demand forecasts at arbitrary geolocations where no measurement data are available.

This paper introduces a graph convolutional recurrent neural network (GCRNN)–based method that enables forecasting at such locations. By leveraging spatial relationships and contextual information from nearby monitored areas, the proposed approach generates accurate predictions even in the absence of direct observations.

The method was evaluated on real-world meteorological, traffic, and air pollution datasets, achieving significant improvements in MAPE (up to 66.97%) compared to the state-of-the-art GConvGRU model.

View ARIEN’s Community on Zenodo to read the complete publication.