RINA - Démonstrateur d’utilisation de l’intelligence artificielle pour la gestion opérationnelle des risques naturels gravitaires géologiques | FEREC 2020

of giving more or less importance to the precision or to the recall. The greater the the greater β β the importance of the recall. β = (1 + β 2 ) · · β 2 · + 5. Results of models on RN1 data This section will present the performance of the different models. Firstly the performance of the operational code, following the results of the different models produced by artificial intelligence described in the previous section 4.1. Training and validation general settings As the period before the construction of protective elements (2000-2007) had more events it will be studied more in depth. It was decided that the period between the 01/01/2000 and the 25/02/2004 would be used to train the model while the data between 02/01/2005 and 27/12/2007 would be used to test the model. This permits us to avoid the year of 2004, that is distinct from the others, maintain temporal continuity of training and test data and also similar to reality as the operational code, that will be compared to our results, was also based on the events previous to 2004. The predicted value, the output, was whether there would or would not be a rock reaching the road (0 when no rock fell, 1 if one or more rocks fell). One DNN model experimented with an additional class (0 representing no rock falls, 1 representing 1 rock fall, 2 representing more than 1 rock fall). The input data for those predictions was the daily rain of the day (P0) and of the four previous days (Pi for i between 1 and 4, where “i” is the number of days before the studied date). 5.1. Operation code The operational code takes into account the rain of previous days and uses it to define if the risk of blocks reaching the road is too high. The code studied is the R7 that dictates the road will be closed when: ● The daily rain in the last 24 h is superior to 15 mm ● The daily rain in the last 72 h is superior to 30 mm This means that, in a certain instant, if the rain in the last 24h was superior to 15 mm the road will be closed, and if in the last 72h there was a period of 24h when the rain was superior to 30 mm the road will also be closed. As we are working with daily data a small adaptation must be made. It will be considered the model predicts an event if: 32