Kim & Kim's TSCoNet Forecasts Climate with Accurate Uncertainty

Modern deep learning models excel at forecasting correlated spatio-temporal variables like regional precipitation and temperature, but they typically report no uncertainty. Forcing uncertainty via maximum likelihood often degrades accuracy, especially when variables are strongly correlated. TSCoNet solves this by decoupling the mean and variance estimation. In the first stage, a convolutional-recurrent neural network (CNN-LSTM) learns accurate mean forecasts. In the second stage, the model holds the mean fixed and refines a shared representation to estimate predictive variance, which is then recalibrated to produce well-calibrated prediction intervals. This Gaussian copula framework allows TSCoNet to handle complex dependencies across both space and multiple output variables.

The model was evaluated on simulated non-stationary spatial fields on the sphere and on a real dataset of monthly precipitation and temperature for 50 cities from 2000 to 2020. Results show that TSCoNet matches the accuracy of a strong deterministic forecaster while also supplying calibrated prediction intervals—something deterministic models simply cannot provide. This gives practitioners a single tool that delivers both accurate point forecasts and reliable uncertainty quantification for multivariate spatio-temporal data. The work has immediate implications for climate risk assessment, agriculture planning, and any domain where trust in forecast reliability is as important as the forecast itself.