Brought to you by
Sewer overflow prediction
Data Scientist, Advanced Analytics and AI solutions, GHD Digital; Bachelor of Engineering (Civil), Queensland University of Technology; Bachelor of Engineering (Electrical), University of Queensland
Extreme wet weather events can cause significant disruption to wastewater infrastructure, including causing sewage overflows.
Overﬂow prediction currently uses operational experience and hydraulic models, which are difﬁcult and costly to calibrate. A cost-effective, near real-time solution has been needed that can reduce environmental impact and improve operational responses.
In conjunction with Queensland water utility Unitywater, GHD Digital has co-created a groundbreaking sewer overﬂow prediction model that uses a novel neural network framework to predict localised rainfall and estimate the probability of sewer pump station overﬂ ow up to six hours in advance.
In place of traditional physics-based models, the solution draws inspiration from the computer vision ﬁeld, harnessing the spatiotemporal relationship between radar imagery and sewer overﬂow SCADA data.
The solution draws inspiration from the computer vision ﬁeld, harnessing the spatiotemporal relationship between radar imagery and sewer overﬂow SCADA data.
This neural network model was developed and implemented by GHD Digital Data Scientist Jeffrey Fisher.
GHD Digital and Unitywater’s tool consists of a recurrent, convolutional neural network that was developed using Google’s Tensorﬂow v 2.0 machine-learning libraries.
The topology of the model is based on a neural network developed by researchers from the Hong Kong University of Science and Technology.
The model consists of two submodels: a “nowcast” model that forecasts the next six-hour sequence of rainfall radar images, and a “prediction” model that takes this forecast sequence and provides a probability of sewer overﬂ ow for each pump station.
A key innovation is that the only input into the model is a sequence of image ﬁles. No explicitly measurable data such as rainfall intensity, wind speed or time of day is used, nor is there any reference to a computationally intensive network hydraulic model.
The non-deterministic nature of this model allows for rapid calculation to be made, in contrast to the time-consuming, deterministic hydraulic modelling techniques that have traditionally been used.
“A useful tool with strong community beneﬁt that demonstrates the strength of partnerships between groups with complementary areas of expertise. This is an excellent example of an old-world problem being solved in new ways.
“This kind of predictive modelling using new sources of data should become the bread and butter of engineering in the future and it is good to see these kinds of applications being implemented. If successful, it seems the type of process that could be implemented in any city, meaning its potential is far-reaching.”