Pipes

PREDICTING SEWER CHOKES THROUGH MACHINE LEARNING  
Advanced analytics to predict the number, location and type of chokes in sewer assets
B Cameron, M McGowan, C Mitchell, J Winder, R Kerr, M Zhang
Publication Date (Web): 1 November 2017
DOI: https://doi.org/10.21139/wej.2017.035


In 2015, sewer chokes cost Sydney Water and its customers over $10 million. Sewer chokes are blockages typically caused by external factors such as tree roots, fats and grease, and foreign items in the pipes such as wet wipes. Chokes may lead to sewage overflows through designated sewer system overflow points, or uncontrolled overflows onto public or private property, or rarely inside houses. 

The likelihood of any particular main choking may be influenced by many environmental, social and structural factors. Traditional forecasting and prediction techniques are currently not capable of accurately predicting sewer chokes in the Sydney Water network. This limits the extent of the network where preventive maintenance strategies can be economically employed. 

Sydney Water, through working with CSIRO’s Australian Digital and Data Innovations Group, Data61, has used machine learning techniques to analyse factors that may contribute to sewer chokes and have developed a pilot model to predict the likelihood of future chokes in every sewer main asset. This may enable Sydney Water to shift from the reactive approach towards a more proactive approach to sewer choke management.  

The pilot predictive model was developed in four stages: 

1.  Factor analysis – identifying the relative importance of each feature (potential factor present in dataset) to contribute to accurate predictions. 
2.  Model development and training – identifying and ‘learning’ the patterns in the data which lead to chokes. These latent blockage patterns were then incorporated into four different algorithms, one for each main choke type. 
3.  Validation – The model was then ‘back tested’ to predict chokes in 2015. 
4.  Prediction – Potential chokes, their location and likelihood, were then predicted for 2016.  

The project showed that machine learning and factor analysis techniques can be applied to improve the prediction of sewer chokes. The model provided a 33% improvement in the ability to predict sewer chokes when considering pipes with the highest probability of chokes. However, the project highlighted the potential need for improved or new data that is currently not readily available.

Using data analytics can assist Sydney Water to improve our ability to predict future events, such as sewer chokes. This will support Sydney Water’s corporate strategy by enabling the business to deliver world class service to our customers, the community and the city. 

The project showed that a machine learning model could be built to predict the occurrence of sewer chokes if a significant amount of relevant data is available. The ability to predict the number, location and type of chokes likely to occur may be able to inform future resourcing, business and operational planning. 

 

Click here to read the full paper


Please note: You need to login to access this member-only content.
Not a member? You can purchase the paper in our Online Document Library.