Reproducing human decisions in reservoir management: the case of lake Lugano

Abstract

The objective of this study is to identify a model able to represent the behavior of the historical decision maker (DM) in the management of lake Lugano, during the period 1982–2002. The DM decides every day how much water to release from the lake. We combine hydrological knowledge and machine learning techniques to properly develop the model. As a predictive tool we use lazy learning, namely local linear regression. We setup a daily predictor, which achieves good accuracy, with a mean absolute percentage error around 8.5%. Yet, the behavior of the model is not fully satisfactory during the floods. In fact, from an interview with a domain expert, it appears that the DM can even update the release decision every 6 hours during emergencies. We have therefore developed a refined version of the model, which works with a variable time step: it updates the release decision once a day in normal conditions, and every 6 hours during emergencies. This turns out to be a sensible choice, as the error during emergencies (which represent about 5% of the data set) decreases from 9 to 3 m³/sec.