AI in the water sector

Real-time monitoring and control

When it comes to sensor analytics there are two main areas our clients in the water industry focus on: real-time monitoring and control.

Monitoring is about keeping an eye on processes and assets and ensuring they function as they should. This includes the detection of abnormal events to provide early warnings.

Control is about optimising processes based on data and reducing human effort. For example, control of a pumping station can be based on a predicted water level, taking into account historical and expected precipitation, evaporation and groundwater levels, as well as data about other pumping stations and dams in the area.

Even if the application is different, the questions asked are always similar. Part of our work as data scientists is to understand a complex problem and then find the right techniques to process data based on these questions. We apply three main techniques to achieve this goal.

Predicting future value

Sometimes we want to know more than whether the sensor will exceed a limit value. We might want to know how long an incident will last, for instance. Or how much the sensor value will exceed the limit, and when the peak will be reached.

This can be done by analysing historical data over a certain spread of time, and developing a model that predicts current or future values. For example, a model could learn how a given water level correlates with historical and actual measurements of precipitation, other water levels and evaporation.

This enables the model to predict the impact of different intensities of rain showers at different times. After all, a heavy shower after a period of drought has a different impact than it would after a wet period.

With machine learning and the right quality of data, we can predict the full course of future sensor values. Sometimes it’s even possible to show the impact of different interventions, like what would happen to the water level if you were to change the valve position of a weir. This allows experts to make more informed decisions about which measure will lead to the best results or minimal disruption.

Anomaly detection with quantile regression

With anomaly detection, we look at present and past data to detect when a system is behaving differently than we’d expect.

To do this, we use machine learning to paint a picture of what a normally functioning system looks like under different conditions. If we have enough data about how a system should function in different circumstances, we can compare this with the actual situation using both measured and predicted values.

When there’s a discrepancy between what a system should look like and what it does look like, we have an anomaly. And this can help draw attention to problems that need addressing.

For instance, the water level at a sewerage pumping station can be elevated, but if this is caused by rain drainage this is normal behaviour. 

However, if an anomaly is indicated, then there may be a genuine problem. Perhaps the pumping station is processing a rain shower much more slowly than expected due to a blockage in the system.

Or perhaps there is simply a 'sensor drift'; a natural tendency for sensors to deteriorate over time due to aging components. Either way, the problem should be investigated.

It is important to mention that we’re not looking for a specific known problem with this approach, but rather for unknown problems that cause the sensors to deviate from the standard.

These are just three types of analyses we could carry out on sensor data, but there are many more. If you’re looking to get more insight from your data, we’d be happy to help you get started. Get in touch to talk to one of our experts.