Real-time prediction of rain-impacted sewage flow for on-line control of chemical dosing in sewers

Highlights

• Rain-impacted sewage flow is predicted in real-time for chemical dosing control.

• Prediction delays are significantly reduced by considering the rainfall events.

• Two sewer networks of different characteristics are used as case studies.

• With flow prediction, lower chemical cost and better pH control are achieved.

Abstract

Chemical dosing is a commonly used strategy for mitigating sewer corrosion and odour problems caused by sulfide production. Prediction of sewage flow variation in real-time is critical for the optimization of chemical dosing to achieve cost-effective mitigation of hydrogen sulfide (H₂S). Autoregressive (AR) models have previously been used for real-time sewage prediction. However, the prediction showed significant delays in wet weather conditions. In this paper, autoregressive with exogenous inputs (ARX) models are employed to reduce the delays with rainfall data used as model inputs. The model is applied to predicting sewage flows at two real-life sewage pumping stations (SPSs) with different hydraulic characteristics and climatic conditions. The calibrated models were capable of predicting flow rates in both cases, much more accurately than previously developed AR models under wet weather conditions. Simulation of on-line chemical dosing control based on the predicted flows showed excellent sulfide mitigation performance at reduced cost.

Introduction

Sewer networks are among the most critical urban infrastructures. They collect and transport domestic and industrial wastewater to wastewater treatment plants (WWTP) for treatment and disposal. Sewer systems are vulnerable to corrosion caused by hydrogen sulfide (H₂S) generated in sewage under anaerobic conditions (Boon, 1995; Hvitved-Jacobsen et al., 2013). The emission of H₂S and other odorous compounds also induce serious odour complaints and health hazards (Jiang et al., 2015).

To minimize the generation and emission of H₂S in a whole network, a commonly used strategy is dosing of chemicals, such as oxygen (Gutierrez et al., 2008; Ochi et al., 1998), nitrate (Jiang et al., 2009; Yang et al., 2005), iron salts (Firer et al., 2008; Nielsen et al., 2005), and magnesium hydroxide (Mg(OH)₂) (Gutierrez et al., 2009). The chemical dosing rate is critical to the sulfide control performance and the operational costs. The main dosing strategies include:

• Constant dosing with which the dosing rate is maintained at a constant value, without considering hydraulic or biological dynamics and variations of the wastewater characteristics.

• Flow-paced dosing, with the dosing rate proportional to the sewage flow rate.

• Profiled dosing, where the dosing rates are pre-defined in a profile which is established from empirical knowledge and typical hydraulic conditions (Ganigue et al., 2011). The dosing strategy may have satisfactory performance, when real conditions are similar to the designed ones. However, variable weather events induce changes to sewage hydraulics and characteristics, leading to non-optimal dosing.

Recently, an on-line control strategy of chemical dosing for sulfide mitigation was developed and applied to H₂S control with Mg(OH)₂ (Ganigué et al., 2016) and FeCl₃ (Ganigué et al., 2018) dosing. One critical component of the control structures proposed was a feedforward controller that calculated the amount of Mg(OH)₂ or FeCl₃ required to balance acid or precipitate sulfide to be produced in the downstream pipes during sewage transport. As the acid/sulfide production mainly depends on the hydraulic retention time (HRT) of sewage in the pipe (Sharma et al., 2008a), the prediction of HRT is central for the control strategy (Sharma et al., 2008b). For a volume of sewage pumped into the pipe, it can be treated as a moving slug. The slug is pushed through the downstream pipe by the future sewage flow entering the sewer pipes. Due to this nature of sewage in sewers, the HRT of a sewage slug in the downstream pipe is unknown at the time when chemicals are dosed to the slug. Therefore, sewage flow has to be predicted several hours ahead in order to predict HRT of the current slug. The prediction horizon depends on the retention time of the slug in the pipe, generally in the order of hours.

An autoregressive model was developed for real-time prediction of future flow in sewers (Chen et al., 2014), which had been applied in the HRT-based on-line dosing control (Ganigué et al. 2016, 2018). The calibrated model could predict sewage flow up to six hours ahead with good accuracy during dry weather, providing critical inputs to the on-line controllers. The controller supported by the flow prediction reduced Mg(OH)₂ consumption by 15% in comparison to flow-paced dosing strategy (Ganigué et al., 2016), and FeCl₃ consumption by 31% (Ganigué et al., 2018) in real-sewer applications. In both cases, however, a delay of a few hours occurred between the real flows and predicted ones during wet weather. The delay was caused by the model not directly considering the rainfall in the catchment. Rainfall derived inflow and infiltration enter not only combined sewers, but also sanitary systems via cross-connection, defective pipes and joints, leaking manhole walls and covers or low disconnected traps (Butler and Davies, 2004). The additional flow shortens sewage HRT in the network, which affects the amount of chemicals required. Also, concentration of the chemical dosed is dramatically diluted by the increased flow. Both factors contribute to poorer control performance of sulfide in sewers (Jiang et al., 2013). Therefore, a methodology for flow prediction directly considering the impact of rainfalls is of great importance for the optimization of sulfide mitigation.

To develop a model for predicting the rain-impacted sewage flow, in a review of hydrological forecasting methods (O'Connell and Clarke, 1981), it was highlighted that a predicting model should be able to reflect the recently observed state, to provide immediate response to the time-varying series, and to compute efficiently. Bidwell (1971) predicted the stream flow out of a catchment, by considering the catchment as a discrete non-linear system with rainfall as the input. A certain memory period of rainfall values was determined by stepwise multiple regression. However, this method had low ability in accounting for evaporation, and caused over-estimation during long dry periods. In a study of hydrological runoff, the surface runoff and rainfall derived infiltration were separated based on unit hydrograph through ARMA (autoregressive moving average) transfer function modelling (Novotny and Zheng, 1989), but not as an exogenous input. In Novotny and Zheng (1989) and several other studies (Ding and Chen, 2005; Ding et al., 2006; Gelfan et al., 1999), several algorithms for model parameter estimation were introduced, and the impact of model structure, order and parameter values on the accuracy of prediction was analysed. These studies provided useful guidance on ARMA modelling. They showed that the accuracy of on-line flow prediction highly relied on the order of the model used. They also revealed that an increase of the model parameters can effectively improve the results for off-line prediction (Ding et al., 2006). However, the model cannot handle an entire non-stationary process and thus requires further study before being applied to predict sewage flow during wet weather.

This study aims to develop a methodology that is able to achieve real-time flow prediction of the sewage under variable weather conditions. This was achieved through fitting flows and rainfalls to an autoregressive moving average with exogenous inputs (ARMAX) model. The rainfall data was taken into consideration as an external input to improve the prediction accuracy by decreasing the delays. The calibration and validation of the model was conducted with data from two sewage pumping stations (SPSs) with different hydraulic characteristics and climatic conditions, collecting sewage in a small and a large catchment, respectively. The value of the models in supporting real-time chemical dosing control was then demonstrated through simulation studies.