Gaussian Process-based Spatio-Temporal Predictor
Balázs Varga · Acta Polytechnica Hungarica · 2022
This paper presents a grid-based algorithm using Gaussian Processes to predict outputs using spatially and temporally dependent data.First, independent Gaussian Processes are formulated along space and time axes.Then, these processes are coupled with a common noise in the covariance kernel.This common noise acts as a smoothing parameter, trading off accuracy at knots for extrapolation capabilities.The algorithm can predict timeseries at unmeasured locations.The efficiency of the algorithm is demonstrated in a traffic flow prediction problem.Results suggest that applying a common additive noise term capturing cross covariances improves prediction accuracy when extrapolating outside the dataset.