| Popis: |
Forecasting water level from natural resources like rivers is very challenging due to the complexity of water sections and various environmental effects. In the Ile-de-France region around Paris, the Syndicat des Eaux d’Ile de France (SEDIF) provides drinking water for more than 4 million inhabitants. Most of this water is produced by three large interconnected plants located on rivers Seine, Marne and Oise. When a flood happens, the water utility may change its operating strategy. If necessary, a plant is stopped another one takes over to maintain the Water Distribution System (WDS) while consumers are always supplied with drinking water. In this paper, a statistical approach is adopted to forecast river water levels when flood event occurs. Two ensemble models are designed to estimate values of the water level up to 48 h. The first modeling is a combination of a single-layer neural network tuned by Gradient Boosting (GB). The second method is a sequential version of the first one where a gradient boosting model is trained at each time step and the training includes the prediction of previous step models. These two models are evaluated to forecast water level located at the three production plants with horizons of 6, 12, 24 and 48 h. Real data from plants are used as well as upstream monitoring stations i.e. six stations on Marne, eighteen stations on Seine and ten stations on Oise. The extensive experiments show that the sequential model outperforms the other one and allow us to highlight the effectiveness of the proposed approach. |
