A machine learning-based surrogate model for the identification of risk zones due to off-stream reservoir failure

Abstract

With the modification of the Regulations of the Hydraulic Public Domain of Spain in 2008, approximately 70.000 owners of off-stream reservoirs are obligated to present a classification assessment on the potential risk due to failure, which requires complex procedures. This work proposes a simplified methodology based on Machine Learning, which allows identifying risk zones at any point at the affected area based on the physical characteristics of the reservoir and the surrounding terrain. Random Forest algorithm is applied to two datasets generated with synthetic cases designed and modelled in Iber. Two methods were tested for balancing the datasets: synthetic minority over-sampling and random under-sampling. Results show high accuracy on both models, although the Random Forest model adjusted with random under-sampling presented better results for the estimation of risk zones. In conclusion, this work found that the simplified method based on Machine Learning can be a useful tool to owners and government administrations, having an equally reliable estimation than current methods and reducing the computational time and resources.