Rosedale wastewater treatment plant, New Zealand

Cities need sustainable wastewater solutions

As city populations grow, wastewater plants must increase their treatment capacity – and the need is rising rapidly. Asset owners and operators are under pressure to extend the life of existing infrastructure efficiently, safely and without service interruption. On top of that, they must build in climate resilience and meet carbon emissions reduction targets.

These challenges call for smarter, better ways to manage infrastructure, and that’s where data-based solutions like digital twins come in. A digital twin is an accurate virtual model of an asset that can be updated with performance data from the physical asset via live sensor input, as well as using external sources such as weather radar.

We identified that a digital twin could help manage cost and greenhouse gas emissions at the Rosedale wastewater treatment plant in Auckland, helping owner Watercare to improve control of operational costs and work towards its strategic objective of halving operational carbon by 2030.

An integrated team from Mott MacDonald and Watercare developed a digital twin featuring real-time dashboards to monitor operational costs and emissions. The system relies on soft sensors that detect instrumentation drift and failure, for improved quality and accuracy of operational data.

The plant at Rosedale will support population growth in Auckland

Insights into plant performance

After creating a digital twin for the commissioning of a new bioreactor at Rosedale, we saw it could also help cut operational cost and carbon.

New Zealand’s most populous city is growing rapidly, so its wastewater infrastructure needs to keep pace – and in a cost-effective, sustainable way. More than 1.5M people already live in Auckland but the population is forecast to reach 1.9M by 2035.

Opened in 1962, the Rosedale wastewater treatment plant was built with capacity to process wastewater from 240,000 people. Presciently, it was designed so this could later be ramped up by a third to serve a population of 320,000.

In 2018, Mott MacDonald was appointed as principal engineer on the upgrade, from concept design through to construction monitoring and commissioning. Our project team overcame the challenges of managing complex connections, live tie-ins and shutdowns, ensuring the plant kept running throughout.

This is truly innovative and a practical example of how data and analytics can generate new insights and improve decision making, reducing commissioning time and the risk of pollution incidents. It’s an industry first and a step change in how to commission treatment plants.

Anna Whitmore

Digital solutions consultant

Two years later we developed a digital twin to generate insights into plant performance and scenario testing to support the commissioning of a fourth modified Ludzack-Ettinger (MLE) bioreactor and associated plant at Rosedale. It combines biological modelling and machine learning with predictive rainfall and real-time data, supported by our Moata digital solutions platform.

The digital twin can, for instance, predict how an extreme weather event will affect plant performance, enabling operators to simulate a range of scenarios and trial different operating procedures virtually before implementing them on the plant.

“This is truly innovative and a practical example of how data and analytics can generate new insights and improve decision making, reducing commissioning time and risk,” says Anna Whitmore, senior consultant with Mott MacDonald.

But developing the digital twin for commissioning the new bioreactor turned out to be only a first step. Anna explains: “We worked with our client Watercare to ask, what if we developed a more comprehensive, intelligent digital twin to improve sensor reliability and enhance existing data?

“This would give us new insights into how to reduce operational cost and carbon and improve the plant’s whole-life performance.” And that’s exactly what our team went on to do.

Combining biological modelling and machine learning with predictive rainfall and real-time data

How soft sensors improve instrument reliability

Analysis using data science helped us determine the thresholds that will trigger an alert to check for probe failure.

Accurate data is critical for optimising performance, yet low-quality operational data due to unreliable instrumentation is common at wastewater treatment plants. A major cause is sensor drift, where instrument readings falsely increase or decrease over time in a way that can be difficult for operators to spot visually.

Using a combination of biological, machine learning and statistical modelling, we developed ‘soft’ or ‘virtual’ sensors of key plant performance parameters to identify drift in physical sensors’ data. Operators are alerted to drift as it occurs so they can calibrate, clean or replace the corresponding physical sensor, improving the quality of data captured.

Making use of historic data

To develop this system, we first reviewed historic sensor data alongside maintenance records to identify past drift events. After learning about wastewater quality characteristics from Watercare’s instrument technicians and plant operators, we worked out how to cleanse the data to remove anomalies and spikes.

The next step was to develop a model. After determining the most appropriate kind of model for the specific data type, we identified the inputs required, trained the model on historic maintenance and operations data to simulate plant performance, and deployed soft sensors connected to SCADA (supervisory control and data acquisition) instrument data. For each parameter we monitored the deviation between physical and soft sensor averages and determined whether sensor drift had occurred.

Finally, we configured email and/or text alarms to alert operators when an instrument’s drift threshold is exceeded, reviewed each soft sensor’s feedback to determine if it had correctly identified drift or raised a false alarm, and adjusted the configuration as required.

Plotting the measuring points at Rosedale

Three ways to detect sensor drift

We used a variety of modelling techniques to develop the soft sensors, depending on the type of data monitored.

Biological models

A core component of the original digital twin developed for the plant’s upgrade programme is a live BioWin model, a treatment process simulator that determines values based on biological and chemical reactions in the plant. We generated a soft sensor for ammonia (NH3) from this model, using influent data to calculate expected ammonia concentrations.

During heavy rain, the model combines data on influent flow and rainfall with a fixed ammonia diurnal load profile to determine the approximate influent concentration. The digital twin assumes increased flow caused by rain is more dilute.

Machine learning models

To create the soft sensors to monitor suspended solids and nitrite (NO2-) and nitrate (NO3-) levels, we used a machine learning model that automatically identifies complex relationships in input and output data to increase predictive performance.

By training the model on previous probe readings, it was possible to get the best of both worlds: the short-term patterns modelled from the probe and the trend of the lab data.

Anna Whitmore

Digital solutions consultant

“Watercare’s process engineers would identify typical parameters that impact total suspended solids, such as flow rate and load, and our data scientists would train the model based on the features that best predicted the data,” explains Anna Whitmore.

“By training the model on previous probe readings, it was possible to get the best of both worlds: the short-term patterns modelled from the probe and the trend of the lab data.”

Statistical models

An MLE bioreactor’s performance is strongly influenced by dissolved oxygen levels, because dissolved oxygen is vital for the removal of ammonia. The plant adjusts airflow into the bioreactor cells based on probe values to maintain a target level of dissolved oxygen. If the probe is incorrect, airflow could be higher than required, unnecessarily driving up energy costs.

We used a statistical model to determine probe sensor drift by monitoring the relationship between the air control valve position and average probe values for dissolved oxygen. When the valve opening fails to raise the dissolved oxygen value, a sensor drift alert is triggered.

Physical sensors for dissolved oxygen tend to be unreliable and difficult to maintain. They are also critical for process control and major contributors to total plant energy consumption, therefore identifying sensor drift is important in improving overall process performance and energy optimisation.

Opportunities for expansion

As data accumulates from the 20 soft sensors now in operation at Rosedale, this builds a basis on which to roll out monitoring further.

Watercare’s operational team are continuously recording drift alerts and outcomes and feeding this information back to our data science team. Long-term performance monitoring will build confidence in the sensors’ accuracy and open up opportunities to roll soft sensors out for other critical parameters, and at other sites.

When scaling to other plants or parameters a choice must be made between modelling methods, and the various methods have different strengths.

A biological model takes considerable initial effort to calibrate, including comprehensive plant-wide sampling data and expert process knowledge. But once calibrated, it can be used to virtually trial control strategies and interventions.

A biological model will be the best solution when a well-calibrated model already exists, but a machine learning model will be a faster and cheaper solution when quality training data is available.

Anna Whitmore

Digital solutions consultant

A machine learning model is quicker and cheaper to deploy and, based on the results from Rosedale, can generate more accurate soft sensors. However, because it learns relationships from historic data, it will not immediately reflect an operational change in the plant. This type of model is therefore less effective in trialling of asset management scenarios.

“Different situations will require different techniques,” says Anna Whitmore. “A biological model will be the best solution when a well-calibrated model already exists, but a machine learning model will be a faster and cheaper solution when quality training data is available.”

Senior consultant Anna Whitmore and principal process engineer David Hume helping to install nitrous oxide monitors at Rosedale

Visual dashboards on expenditure and emissions

We used our Moata digital platform to generate visual readouts of daily carbon and cost outputs.

Typically, operational expenditure (opex) and carbon emissions generated by wastewater treatment processes are calculated and reported monthly, with many values based on theoretical factors. Calculations for emissions from bioreactors, for example, are based on global climate change guidelines and the assumption of a static population.

“It is crucial to increase knowledge and understanding of opex and emissions sources to enable improved control and reduction,” says Anna Whitmore. “By combining over 50 data streams from sources such as telemetry, laboratory sampling databases and maintenance systems, we developed real-time operational dashboards to enable Watercare to identify hotspots of opex and greenhouse gas production across the plant.”

It is crucial to increase knowledge and understanding of opex and emissions sources to enable improved control and reduction.

Anna Whitmore

Digital solutions consultant

To develop the dashboards, we first identified sources of operational expenditure and emissions. These were broadly split into several categories: energy consumption, chemical consumption, maintenance, labour, process emissions, effluent discharge and biosolids disposal.

We reviewed each source to identify what data was available for calculating cost and carbon outputs, then worked out baseline calculations for each source. This was used to develop data pipelines for automatic ingestion of data into the digital twin in our Moata Smart Water app. The data was cleaned – for example, by excluding negative values – to improve the accuracy and completeness of input streams.

We then codified these calculations in Moata to automatically calculate daily operational expenditure and carbon emission values. The final step was to develop standardised operational dashboards that visually model the results for real, virtual and predictive scenarios, through an easy-to-use browser interface.

Harness the power of data with Moata

Moata, our digital platform, hosts a range of solutions that use the power of data to solve today’s most pressing infrastructure challenges around the globe. It is open, secure, scalable and adaptable, delivering predictive power in a geospatial context, through advanced analytics and machine learning.

Revealed: the cost versus carbon correlation

Results from the dashboards reveal that process emissions at Rosedale make up 94% of total on-site greenhouse gas emissions.

They show about 100t of CO2e are produced from MLE process emissions each week, equivalent to NZD$442,000 per year using the cost of a typical carbon credit price of NZD$85 per tonne of CO2e (New Zealand emission unit, NZU). Whilst wastewater is not currently in the NZ Emissions Trading Scheme, there is the potential for it to be in future.

These results reinforce the need to consider the connection between opex and emissions. By creating dashboards which include both values, we can identify interrelationships and monitor trends when trialling strategies to optimise wastewater treatment for better cost and climate performance.

This is especially relevant to the operation of Rosedale’s MLE bioreactors where aeration has a significant impact on both opex, making up 10% of total opex, and nitrous oxide production, which represents 57% of total emissions produced by the plant.

Innovation unlocked

The hidden potential of plant data revealed at Rosedale helps point the way to new strategies for wastewater management.

“Plant operators rely on data to control processes and for strategic analysis,” says principal process engineer David Hume, of Mott MacDonald’s water team in New Zealand. “So it is critical for any plant optimisation that this data is accurate and reliable.”

Soft sensors have the potential to significantly improve the quality of a plant’s operating data.

David Hume

Principal process engineer, water

The analysis of plant data at Rosedale has shown how collaboration between process engineers and data scientists can unlock innovations. David explains: “Soft sensors have the potential to significantly improve the quality of a plant’s operating data, either through improved sensor maintenance routines or reducing reliance on unreliable field instrumentation.”

The dashboard is an exciting new development, allowing plant operators to easily access the information they need to understand existing and predicted plant performance and how process emissions are generated.

David adds: “Digital twins, by combining modelling and machine learning with real-time data and powerful visualisation, will generate new insights into plant performance and how to develop control strategies to reduce operational costs and greenhouse gas production. They will therefore play an important role in the future of wastewater treatment.”

Digital twins, by combining modelling and machine learning with real-time data and powerful visualisation, will generate new insights into plant performance.

David Hume

Principal process engineer, water