CritiCoder: An End-to-End Uncertain Regression Network for Robust Macroscopic Pressure Models in Water Distribution Systems

Abstract

Due to the massive uncertain disturbances in water distribution systems (WDSs), it is rather challenging to construct robust macroscopic pressure models. In this article, we propose the CritiCoder, an end-to-end uncertain regression neural network, to build macroscopic pressure models in WDSs and estimate the distribution of uncertain disturbances. By separating the normal regression process into two parts, the CritiCoder decomposes the output into two parts brought by observable and unobservable variates separately. Two subnetworks, Coder and Critic, make up of the CritiCoder. Through the reconstruction of data flow, the Coder is expected to approximate ideal outputs from observable variates. Meanwhile, the loss function of the Critic is redesigned to consider the effectiveness of uncertain disturbances in output brought by unobservable variates. Experiments on a practical application of WDSs in a Chinese mega-city show superior performance of CritiCoder. Especially for pressure-monitoring nodes more severely impacted by disturbances, the performance decline of the CritiCoder is less compared with other baseline methods.