An autoencoder neural network for time series prediction: application on the evolution of cesium-137 concentrations in French rivers
Kathleen Pele est post-doctorante à l’IRSN.
Nowadays, monitoring the impact of human activities on the environment is a major challenge as many pollutants can be found in the different ecosystems. It is therefore important to understand their past and present evolution to predict future consequences. We are interested here in the impacts of the nuclear industry on rivers that started in the middle of the 20th century. This study focuses on the cesium-137, a radionuclide that has a lifetime of about thirty years and is found significantly in the releases from the main nuclear industries and has been heterogeneously deposited on the French watersheds after accidental events such as Tchernobyl.
Understanding the fate of this contaminant is therefore essential for the protection of these hydrosystems and thus to better understand their resilience’s. In this original approach, we propose a deep learning tool to predict the concentration of this radionuclide in the suspended solids and sedimentary archives of rivers. The objective is to predict the concentration from a set of variables providing information on water discharge, soil washout flux and industrial radioactive releases. These multivariate time series data sets present very different and sometimes complex temporal dynamics at several levels and non-linear relationships.
We propose the use of an autoencoder neural network, “Hierarchical Attention-Based Recurrent Highway Networks”, characterized by its ability to extract the most relevant temporal and spatial information from the databases. This multivariate modelling is also interesting for its ability to understand the relationships between variables that can play the role of both explanatory and target variables. Once optimized, the model generates first results in agreement with the real concentration curves by correctly following the main trends, even if inaccuracies remain, notably due to the quantity of data available. This error rate is tolerated because the objective was mainly to know the evolution of the concentration over long periods and not to have a precision to the day.
The originality of this model is its capacity to make predictions on different catchment areas. With these first interesting results, further investigations should concern different radionuclides than cesium-137 to better understand the robustness of the model.
Sorbonne Université – Campus Pierre et Marie Curie
METIS-IPSL, salle Darcy