In order to support the monitoring of a deep geological repository for radioactive waste (i.e. Andra’s Cigéo project) during its operating phase, real data acquired by a network of sensors and virtual data from predictive models (i.e. numerical simulation) will be compared. Considering secular time scale monitoring, the implementation of spatio-temporal (4D) interpolation techniques may prove useful to ensure the data consistency overtime, especially under dynamic conditions. Over extended periods, the evolution of the sensor network geometry may evolve due to possible losses of aging sensors, their replacements or the incorporation of new sensors when possible. For such tasks, graph neural networks are efficient, given that graphs can be used to accurately depict physical phenomena (such as thermal conduction) and to model geometrical and local dependencies. In this work, we leverage the availability of the experimental data acquired in Andra’s Underground Research Laboratory (URL) to train two sets of graph neural networks: the first set of graph neural networks are designed to assess the data’s integrity and the second to interpolate missing data for effective monitoring. The selected experiment treated in this work emulates the thermal loading of a high-level waste demonstrator cell. In particular, we use the data along a distributed sensor (i.e. optic fiber) and work with a unidimensional dataset. The first type of graph neural networks inputs the field of temperature from the sensors at current and past steps and return the state of each individual sensor, faulty or not. These networks are derived from the GraphSAGE model designed by Hamilton & al. 2017 [1]. The trained networks are ultimately compared against a thresholding method, proving their effectiveness. The second type of graph neural networks inputs the field of temperature of healthy sensors at current and past steps and return an approximation of the temperature for the faulty sensors. These networks are derived from message passing neural networks models designed by Gilmer & al. 2017 [2]. The trained networks are compared against widespread interpolation methods such as kriging, establishing their effectiveness. Both models include elements from the graph networks framework developped by Battaglia & al. 2018 [3].