Modeling uncertainty with engression: a deep generative time-series approach

Deep learning enables precise environmental predictions and simulations across spatial and temporal scales. However, reliable uncertainty estimation with generative capabilities remains crucial for actionable forecasting and simulation, yet robust and simple quantification methods remain challenging. Recently, engression, a generative approach for model-agnostic training and uncertainty quantification, has been proposed. We evaluate its feasibility for environmental time-series modeling, specifically applying it to rainfall-runoff prediction using long short-term memory (LSTM) networks. As a benchmark, we use quantile regression, a generalization of the mean absolute error (MAE). Our results show that engression-LSTM is an effective and easy-to-use method for generative modeling, outperforming quantile regression in uncertainty quantification. Furthermore, qualitative analysis of the generated runoff time series indicates high dynamic fidelity. These findings highlight engression-LSTM as a promising approach for incorporating uncertainty into environmental time-series modeling.

Journal article

Submitted to Geophysical Research Letters