Search papers, labs, and topics across Lattice.
This paper introduces a method for probabilistic spatio-temporal forecasting using feed-forward neural networks with Gaussian weights, trained via MMAF-guided learning. The MMAF-guided learning leverages an embedding informed by the assumption of an underlying spatio-temporal Ornstein-Uhlenbeck process to preprocess the data. Experiments on synthetic and real-world datasets demonstrate calibrated forecasts across multiple time horizons, with performance comparable to or exceeding more complex architectures like convolutional or diffusion models.
Surprisingly, simple feed-forward networks, when guided by a theory-informed embedding, can outperform complex convolutional and diffusion models in probabilistic spatio-temporal forecasting.
We employ stochastic feed-forward neural networks with Gaussian-distributed weights to determine a probabilistic forecast for spatio-temporal raster datasets. The networks are trained using MMAF-guided learning, a generalized Bayesian methodology in which the observed data are preprocessed using an embedding designed to produce a low-dimensional representation that captures their dependence and causal structure. The design of the embedding is theory-guided by the assumption that a spatio-temporal Ornstein-Uhlenbeck process with finite second-order moments generates the observed data. The trained networks, in inference mode, are then used to generate ensemble forecasts by applying different initial conditions at different horizons. Experiments conducted on both synthetic and real data demonstrate that our forecasts remain calibrated across multiple time horizons. Moreover, we show that on such data, simple feed-forward architectures can achieve performance comparable to, and in some cases better than, convolutional or diffusion deep learning architectures used in probabilistic forecasting tasks.
