Comparación de modelos físicos y de inteligencia artificial para predicción de niveles de inundación

Abstract

Hydrology has used traditional methods for flood level forecasting. However, this type of forecast can lead to accuracy issues, caused by the nonlinear behavior of floods and limitations by not including all variables, such as water flow, level and precipitation. Consequently, some scientists began to use unconventional methods based on artificial intelligence models, to forecast floods more precisely and rigorously. This paper compares the HEC-RAS one-dimensional flow transit model with an artificial intelligence model based on Artificial Neural Networks, developed in MatLab to predict floods. The results were analyzed using six statistical indicators: mean absolute error (MAE), mean squared error (MSE), mean absolute percentage error (MAPE), square root of the MSE, Pearson correlation coefficient (CC), and concordance correlation coefficient (ρc). In addition, the efficiency coefficient was calculated, and used in a virtual tool called Hydrotest. The analysis shows that forecast models that use neural networks have accurate results, given their closeness to the real data: MAPE between 11.95 and 12.51, CC between 0.90 and 0.92, ρc between 0.84 and 0.87, and a coefficient of efficiency larger than 0.8. The study was conducted on a section of the upper Bogotá River, in Colombia, between the Florence Bridge and Tocancipá hydrological stations. Flow data was taken from the Regional Autonomous Corporation of Cundinamarca (CAR), from September 2009 to October 2013.