Prediction of daily drinking water demand
An accurate prediction of drinking water demand for the coming 24 hours is a truly valuable addition to the operations of any drinking water utility. Especially on days with higher demands when production capacity is critical, a constant production flow allows leveraging the production capacity maximally. Leveraging existing capacity optimally allows the postponement of investments in new -expensive- treatment works. In addition, a constant production during the day and a minimum of changes in production flow, will contribute to the best possible water quality, and will allow the limitation of chemical and energy usage in the process. A demand prediction algorithm can be extended with real-time anomaly detection to identify unexpected bigger leaks, and an algorithm to optimize distribution pump energy consumption is also included. Finally, an automated generation of optimal production setpoints will bridge the personal preferences of operators. Once they have decided upon automated demand prediction, they can trust the setpoint selection will be optimal, and everyday along the same conditions.

Spatial Insight
Dutch data science consultancy Spatial Insight is a data science consultancy focusing on the management of underground assets. We combine GIS expertise, with data science and asset management. The 10 staff team represents a century of experience in the Dutch water sector and holds a strong passion to solve the needs of utilities with data driven solutions. Spatial Insight can be considered to be a leading consultancy on the Dutch home market.

Spatial Insight has developed SI-Demand, a demand prediction algorithm based on the machine learning principle. With both the utility’s data and open-source data as an input, SI-Demand firstly predicts the total demand volume for the coming 24 hours and then projects the most likely pattern for that day on top of it. Assuming sufficiently dense input data is available, the result is a prediction of the demand for each 5 minutes, so 288 values for the coming day. The algorithm finds the most relevant patterns in historic demands, in weather forecasts, in special days (weekday-Saturday-Sunday, bank holidays, other significant events), and in other eventually available data sources (eg. known leaks). Within the current band width of available production capacity, as calculated in the SCADAsystem, SI-Demand will calculate the optimal production setpoint, and provide that setpoint to the top level of the SCADA system for implementation.

For a Dutch drinking water utility, we have trained and validated SI-Demand on historic data, and the results were satisfying, as shown in Figure 1. We have compared the existing demand prediction model of the client with SI-Demand for a relatively small supply area. Accurate predictions are harder to make for smaller supply areas, compared to bigger areas. The red line in Figure 1 shows an impressively accurate prediction of the demand on a 5 minutes granularity. The blue line is the actual demand that occurred. The purple line shows the production setpoints of SI-Demand (constant 326 m3 /h), whereas the existing algorithm produces setpoints in a band width between 300 and 350 m3 /h to deliver the predicted water flow.

SI-Anomaly and SI-Pumpenergy
We have designed SI-Anomaly, a module to detect anomalies, as an add-on to SIDemand. This module will be able to generate an alarm when the actual demand differs significantly from the predicted demand. This is a powerful early warning system for unplanned leakages or water demand. Also, a conceptual design is ready to include an optimal selection and setpoints of pumps to minimize pump energy consumption.

Essential for each machine learning algorithm is sufficiently available training data. SIDemand requires 25 up to 60 months of historic demand data for each supply area, preferably with 5 mins granularity. Also, we will need an open data source for weather forecast and special days, which are typically available on the local digital market.

Limitation of carbon footprint
Spatial Insight intends to limit the carbon footprint of its operations, and therefore we want to limit our travel movements. We can deliver the majority of our work from distance, and we prefer to work with a local supplier or consultant. We propose only to fly in to build trust, which is hard -if not impossible- to do online.

Next step
We hope and trust SI-Demand will contribute to needs for optimal treatment process control and postponement of investments. We are pleased to explore how we can define a proof-of-concept (PoC) project around SI-Demand at any site.

More information
Download this article or e-mail Ignaz Worm directly.

ignaz.worm [@]

Or if you prefer, you can contact us through our contact form.

contact us for more information

Figure 1. Demand prediction SI-Demand (red line), actual demand (blue line), production setpoint client’s
existing algorithm (orange line) and production setpoint SI-Demand (purple line).