A Bayesian Generative Adversarial Network (GAN) to Generate Synthetic Time-Series Data, Application in Combined Sewer Flow Prediction

TL;DR

This paper presents a Bayesian GAN approach to generate synthetic time-series data for urban sewer flow prediction, addressing data scarcity and enhancing model accuracy through data augmentation, with a focus on combined sewer systems.

Contribution

The study introduces a GAN-based method for generating synthetic time-series data to improve sewer flow prediction models, demonstrating its effectiveness and limitations in different weather conditions.

Findings

GAN successfully generates synthetic data matching real data distribution.

Data augmentation improves peak flow prediction accuracy.

Ensemble models combining real and synthetic data perform better in various scenarios.

Abstract

Despite various breakthroughs in machine learning and data analysis techniques for improving smart operation and management of urban water infrastructures, some key limitations obstruct this progress. Among these shortcomings, the absence of freely available data due to data privacy or high costs of data gathering and the nonexistence of adequate rare or extreme events in the available data plays a crucial role. Here, Generative Adversarial Networks (GANs) can help overcome these challenges. In machine learning, generative models are a class of methods capable of learning data distribution to generate artificial data. In this study, we developed a GAN model to generate synthetic time series to balance our limited recorded time series data and improve the accuracy of a data-driven model for combined sewer flow prediction. We considered the sewer system of a small town in Germany as the test case. Precipitation and inflow to the storage tanks are used for the Data-Driven model development. The aim is to predict the flow using precipitation data and examine the impact of data augmentation using synthetic data in model performance. Results show that GAN can successfully generate synthetic time series from real data distribution, which helps more accurate peak flow prediction. However, the model without data augmentation works better for dry weather prediction. Therefore, an ensemble model is suggested to combine the advantages of both models.