STIFF: A forecasting framework for spatio-temporal data

Nowadays spatiotemporal forecasting has been drawing more and more attention from academic researchers and industrial practitioners for its promising applicability to complex data containing both spatial and temporal characteristics. To meet this increasing demand we propose STIFF (SpatioTemporal Integrated Forecasting Framework) in this paper. Following a divide-and-conquer methodology, it 1) first constructs a stochastic time series model to capture the temporal characteristic of each spatially separated location, 2) then builds an artificial neural network to discover the hidden spatial correlation among all locations, 3) finally combines the previous individual temporal and spatial predictions based upon statistical regression to obtain the overall integrated forecasting. After the framework description a real-world case study in a river catchment, which bears abrupt water flow rate fluctuation, obtained from a British catchment with complicated hydrological situations, is presented for illustration purpose. The effectiveness of the framework is shown by an enhanced forecasting accuracy and more balanced behaviors.