Multidimensional data generation in water distribution systems using the Cycle-GAN

Abstract

[EN] A water distribution system (WDS) consists of thousands of components interacting with each other. To analyze the status or to manage the operation of the WDSs, two main feature values have been mainly used: nodal pressure and pipe flow rate. However, insufficient data due to the malfunction of the sensors or economical limitations interrupts the collection of abundant information in many cases. Therefore, this study proposes a WDS data generation model based on the cycle-GAN (Generative Adversarial Networks). The proposed model learns the demand time series data and its corresponding time series data of pressure or flow. All training data is two-dimensionally constructed by considering the time series data for 24 hours of each component as a single row and arranging all row data vertically. After normalizing all the data to integers from 0 to 255, they become a greyscale image. Then, the cycle-GAN model consisting of two generators and two discriminators trains those image datasets, to translate the demand image data to the WDS feature image data (i.e., pressure or flow). Firstly, based on the random seeds, the first generator in the cycle-GAN model is trained to generate the demand image, and the second generator is trained to generate the WDS feature image. After the fundamental training for the generation of their own data, the second generator trained for the feature image data starts to use the synthesized demand image results from the first generator as its seeds, not using the random noise. This process makes the second generator have the ability to translate the demand images to the WDS feature images by using the convolutional and deconvolutional functions in the neural network layers. This model was demonstrated by applying to the Mays network, which is the benchmark network consisting of 13 nodes and 21 pipes with two reservoirs.