Machine learning to reduce leakage at the Örby field
The Örby field and the Örby drinking water treatment plant, together constitute a technically complex system that contributes to securing the distribution of drinking water.
The overarching goal of this project was to investigate whether machine learning can be used to reduce water losses at Örbyfältet. Reducing these losses is important from both an environmental and sustainability perspective, as well as to decrease the costs associated with water losses. Additionally, the project aims to increase the level of knowledge about machine learning within the water sector and to demonstrate how to work with machine learning from both a development and operational perspective.
From a technical standpoint, the project comprised two main parts: consumption forecasts for drinking water (30 days ahead and high-resolution for 12 hours ahead) and investigating whether machine learning can be used to optimize operational strategies regarding pumping and levels in the field. Creating and evaluating an accurate forecast model for drinking water consumption has many other applications and can be of great benefit to several water utilities. For example, a consumption forecast can be used to detect anomalies and identify leaks in the drinking water distribution network (by comparing expected consumption with actual consumption), as well as contribute to more secure drinking water production and distribution.
The results show that we can successfully predict water consumption for the next 28 days on a daily level and for 12 hours on an hourly level within an error margin of approximately 2-3%. We have evaluated several variables to investigate which of them can improve a machine learning forecast. We found that different algorithms respond differently to various variables. Although some improvement can be seen when certain variables are included, a sufficiently good result could be achieved by using historical data on water consumption. This allows for easier implementation of the developed results. For the optimization of operational strategies regarding pumping and levels, it was assessed that there was not enough high-resolution data available to justify further work in this area. Additionally, the project report includes a comprehensive description of machine learning and data science, as well as considerations when applying these techniques. Our hope is that the final results can be used as a reference for more applications within the Swedish water sector.
