INRAE's Transformer-Diffusion Model Improves Hydrological Forecasting & Imputation

A team from INRAE and ANDRA—Ferdinand Bhavsar, Lionel Benoit, Maxime Savatier, and Edith Gabriel—has introduced a transformer-based diffusion model for probabilistic imputation and forecasting of hydrological time series. Traditional statistical methods often fail to capture the high variability and sparsity of hydrometeorological data, especially when measurements are limited or missing. The proposed framework combines the sequence-processing power of transformers with the generative capacity of diffusion models, enabling realistic sampling of time series distributions even under large observation gaps.

The model was tested on joint water quantity and quality data from six sites across three adjacent headwater catchments on a limestone plateau in northeastern France. The dataset spans over 15 years and was rigorously quality-controlled for sensor drift and malfunctions by LNE metrological experts and Andra's monthly checks. Compared to several baseline approaches, the transformer-diffusion model demonstrated superior performance in both imputing missing observations and forecasting future hydrological conditions. The results highlight diffusion models' ability to handle variable missing data patterns while preserving key temporal characteristics. This work advances reliable long-term monitoring and risk assessment for floods and droughts.