Development of a Digital Twin to estimate flow in a water supply network

Abstract

The digital twins have been emerging as an important solution for industrial processes for providing simulation models capable of imitating the physical system. However, there is not yet an unanimous approach for creating digital twins, considering the difficulties of using an accurate technique that have a rapid update to the new outputs caused by changes in the system as equipment efficiency, for example. In this sense, this work compares 3 machine learning techniques to modeling a water supply network (specifically, predict the flow in a point of the network), specifically a classic Artificial Neural Network and CatBoost, against the proposed algorithm, the KNN (K-Nearest Neighbors) which have incremental learning. Due to the incremental learning approach, it can be shown that KNN have superior performance than CatBoost and ANN (these techniques have transfer learning and need retraining, respectively), having better performance in both the first phase (training) and the last (the incremental learning), showing its potential for application in digital twins.