Detecção de vazamentos em dados de fluxo de água com seleção e otimização automática de modelos

Abstract

Proper management of water resources is a priority issue in the modern world. An important aspect of this matter is the reduction of losses in the urban water distribution. The real-time monitoring of the distribution system followed by the application of outlier detection techniques on water flow data has been an effective alternative to reduce this index. The identification of best models and optimized parameters for detection is a challenge in this complex scenario. Therefore, the area can benefit from the recent developments in model selection and adjustment strategies, this area also known as Automated Machine Learning (AutoML). This work presents a proposal for the application of outlier detection techniques and AutoML resources on water flow data of 16 District Metering Areas (DMAs) of the water distribution system in Curitiba, state of Paraná, Brazil. An off-the-shelf AutoML tool is applied and automatic optimization of specific outlier detection algorithms is performed. The experiments conducted indicate that combining AutoML with traditional outlier detection techniques is the most effective direction.