Machine learning helps wastewater management

By Oliver Harrison, Sales Engineer

Introduction

Water authorities aim to transport waste water as cost-effectively as possible. Analysing the performance of the sewer networks is key for identifying potential issues developing in the sewer network and detecting where improvements can be made. Pump station efficiency can be improved, wastewater treatment optimised and response time to potential issues reduced. In this article I’ll explain how machine learning in waste water network catchments can improve wastewater management performance using predictive monitoring.

OFWAT governs the UK water authorities, which are subjected to compliance requirements and are scrutinised on key performance metrics. Huge monetary fines (and rewards) are distributed on performance of these key objectives, known as outcome delivery incentives (ODIs).

Two common performance commitments in AMP 7 (April 2020- April 2025) for all UK Water Authorities are to improve on:
– Internal sewer flooding
– Pollution incidents

Exceptional performance on these commitments will yield great rewards for water authorities and predictive network performance will enhance the chances of excelling in these two key metrics.

Machine Learning and Data Mining

Machine Learned (ML) intelligence is the scientific study of algorithms and statistical models that computer systems use to effectively perform a task. e.g. notifying water authorities of potential blockages and sewer flooding to prevent or mitigate issues developing in the sewer network. ML relies on pattern recognition and will automatically notify users when patterns exceed the expected bandwidth of the predicted pattern. ML builds a mathematical model or ‘digital twin’ based on previous data in order to make predictions.

Data mining focuses on discovery of the unknown or irregularities in the data set. Due to the unsteady nature of waste water networks, abnormalities can vary vastly. Human intelligence and feedback from past issues will be inputted into the model to improve accuracy. Future predictions and issues can therefore be categorised accordingly to notify the end user what the potential issue could be and how the issue can be resolved. ML intelligence will be used not only for the site where feedback was gathered, but for all monitoring locations inputted into the model.

Instrumentation and Data Inputs

Buying new flow meters and water level sensors  is costly to water authorities. Data gathered from instrumentation is a valuable, intangible asset which is keyed into the model to create predictions. It is essential that data gathered is accurate and instrumentation is installed at optimal locations. Confidence in product accuracy and installation technique is vital in ensuring the data collected is true. If the data gathered false predictions it will be useless and investment wasted.

Utilising existing instrumentation is the most cost-effective method for predictive analysis for water authorities. The majority ‘hotspot’ areas and CSOs in the waste water network have incumbent instrumentation installed. The nature of Cloud-based ML is that unlimited data can be inputted. For machine learning to be used in predictive analysis, the data must be verified and the instrumentation calibrated. Finally, potential telemetry protocols must enable data from varying instrumentation types and manufacturers’ data to be utilised in the model build.

During dry-weather flow (DWF) monitoring systems can accurately predict what the hourly inflow is for every hour of the week. This is helpful to detect deviation from the normal situation not only for sewer blockages, but can also identify illegal discharges, infiltration, misconnections or how new developments are affecting the network. During DWF typical diurnal patterns are expected, this makes predictions easier and more accurate than during rainfall events.

Live rainfall and rainfall forecasts are applied for accurate predictions. To increase accuracy antecedent rainfall needs to be taken into consideration. Using antecedent rainfall ground saturation levels can be inferred which is crucial in identifying the effects of inflow and infiltration on the sewer system; this will reduce ‘false alerts’.

The nature of rainfall events is that they vary in intensity and duration and are subject to ever changing atmospheric conditions which can make accurate predictions challenging. Accuracy bandwidths are therefore included in the predictions and alerts are only generated when the bandwidth is exceeded. As is the case with ML the more data inputted and over time the prediction becomes increasingly accurate, the bandwidth of the expected pattern will therefore be reduced over time.

The individual site characteristics of each site will be assessed. If any environmental or human influences affect the prediction then data for these factors will be included in the model build. Potential external influences on the sewer system include (but are not limited to):

• Tide times
• River levels
• Ground saturation levels
• Industrial Influence
• Geology of the land
• Sewer condition
• Silt/rubble movement
• Seasonal variation (including holiday parks etc.)

Identifying monitoring locations

When a catchment is currently unmonitored then understanding the catchment and past issues is crucial in identifying potential ‘hotspots’ for monitoring. Sharing knowledge and understanding what the water authority wants to achieve by monitoring a certain catchment is important information when making recommendations regarding where to measure and what type of instrument is most suitable.

The potential monitoring locations should be assessed for suitability and alternative locations identified where necessary.

Deliverables and Benefits

Successful integration of data to build a prediction model will deliver an early warning system for blockages and sewer flooding 24 hours and day, 7 days a week. Machine learning in a waste water network will enable a confident alarm protocol to be developed, which will increase confidence in real alerts/ alarms and reduce false alarms. Water authorities will be able to view what is occurring in their network at all times through a software platform and alarms/alerts will be sent directly to their emergency call-out operations for immediate response to issues.

Management of assets, equipment failures and maintenance schedules will be identified and planning made in a timely manner.

Additional benefits to catchment monitoring are as follows:

• Verification of hydraulic models
• Inform of capacity assessment
• Integration of sewer, pluvial and fluvial models
• Reduction in processing costs of alarms

Conclusion

Predictive analytics analyses current and historical facts to make predictions about the future or otherwise unknown events. Machine learning in waste water enables a huge amount of data to be processed to generate accurate predictions. The predictive model becomes increasingly accurate and dependable the more data that is inputted. Unlimited data streams can be uploaded depending on the individual site characteristics. Machine learning will not only learn how one site operates, but if other sites react in similar ways it will recognise the issue and issue an alert/alarm for full network coverage every hour of every day.