Combined Sewer Overflow (CSO) is a major problem to be addressed by many cities. Understanding the behavior of sewer system through proper urban hydrological models is an effective method of enhancing sewer system management. Conventional deterministic methods, which heavily rely on physical principles, is inappropriate for real-time purpose due to their expensive computation. On the other hand, data-driven methods have gained huge interests, but most studies only focus on modeling a single component of the sewer system and supply information at a very abstract level. In this paper, we proposed the DeepCSO model, which aims at forecasting CSO events from multiple CSO structures simultaneously in near real time at a citywide level. The proposed model provided an intermediate methodology that combines the flexibility of data-driven methods and the rich information contained in deterministic methods while avoiding the drawbacks of these two methods. A comparison of the results demonstrated that the deep learning based multi-task model is superior to the traditional methods.
In recent years, increased impermeable surface, extreme rainfall event and urbanization have resulted in more frequent Combined Sewer Overflow (CSO). Owing to the demand for on-time information, a lot of cities have developed the surveillance system to offer insights into the performance of CSO structures (Montserrat et al. 2015;Power 2016;Ayyeka 2017). Intelligent urban infrastructures such as smart sewer system will become the backbone of future cities (Jaokar 2015). To give sewer surveillance system 'intelligence', both data acquisition and extract useful information from the collected data are indispensable.
In this context, developing a versatile urban hydrological model is imperative to capture useful information from a large amount of collected data and to enhance various tasks. Indeed, how to effectively leverage the data collected by ubiquitous infrastructure sensors through proper modeling techniques has become a sticking point for future intelligent sewer system management (Wu & Rahman 2017).
In general, methods involved in urban hydrological modeling can be classified into two major categories: deterministic and data-driven methods (Nourani et al. 2014). Admittedly, deterministic methods can provide fully detailed information for sewer systems. However, deterministic methods require sophisticated foreknowledge about the sewer system, incorporate a huge number of parameters and the simulation are based on numerical methods. These characteristics make the model construction, calibration and computation of deterministic methods extremely complex. Therefore, deterministic methods are inappropriate for application in real time purpose (El-Din & Smith 2002). Another disadvantage of deterministic methods is that the computation of deterministic methods is based on given rainfall, it cannot provide future hydrological information (Chiang et al. 2010). Accurate hydrological time series forecasting could support engineers’ decision-making, pinpoint the vulnerable part of the sewer system in advance, warm up sewer control facilities or early warning peak events. Hence, hydrological time series forecasting is often a prerequisite for successful sewer system control.
By contrast, data-driven methods are flexible in model development, it avoids complicated hydraulic/hydrological theories by learning from data without human intervention. Moreover, Data-driven methods can produce future hydrological data by being fed with current and previous data. Many research efforts have been done to enrich data-driven approaches for hydrological time series forecasting. Due to capable of handling non-linear and non-stationary problems, the Artificial Intelligence (AI) methods have shown promise among numerous data-driven approaches. A particularly popular sub-set of AI used for hydrological time series forecasting is the machine learning. Typical machine learning algorithms include Support Vector Regression (SVR) and various artificial neural network (ANN) structures. Unlike shallow ANN structures, deep learning models extract high-level abstractions in data through processing data by the internal layers, thus, deep learning is able to provide efficient high-dimensional interpolators that cope with multiple scales and heterogeneous information (Marçais & de Dreuzy 2017). Deep learning has made revolutionary strides in recent years, typical examples of deep learning include AlphaGo (Silver et al. 2016) and the latest Google translation system (Google 2016). Deep learning method has also shown its superior performance compare to traditional methods on traffic time series forecasting (Hsu 2017;Ma et al. 2015;Kanestrøm 2017), and hence employed by Uber (Laptev et al. 2017) for their ride request forecasting system.
Although as the most promising data-driven methods, machine learning/deep learning has presented its power in many studies, we could still find two major deficiencies by summarizing previous researches.
First, the success of deep learning in both academia and industry suggests a natural prospective interest for the use of deep learning for hydrological time series forecasting, but there are very few reports studied the performance of deep learning on hydrological data. Second, in order to forecast urban hydrological time series in near real time, data-driven methods seem a good alternative to deterministic methods, although the latter method could provide fully detailed information. However, in most urban hydrological studies, researchers only focus on predicting hydrological time series for a single component of the sewer system. This kind of model can only provide information at a very abstract level. One may develop models separately for individual parts of the sewer system, but this approach neglecting the existed physical correlation of sewer components. Moreover, a system with many independent models is less efficient due to redundant information contained in these models (Bezuglov et al. 2016), maintain
This content is AI-processed based on open access ArXiv data.
