How Hafslund Kraft improved inflow forecasting with an Azure platform for Machine Learning

A robust Machine Learning platform for inflow forecasting

Hafslund Kraft recognized the need to build a platform to productionize its hybrid models for water inflow. These models were developed to more accurately reflect the milder and wetter weather trend Hafslund Kraft has experienced in its watercourses. For instance, higher rainfall during summer increased pressure on the company’s hydroelectric facilities.

“The physically-based models in use were developed in the 1970s, the model itself has gone through several iterations to become the industry standard it is today” said Martin Heltberg, Hydrologist at Hafslund Kraft. “These physically-based models are calibrated on all the historical data Hafslund Kraft has, going back to the 1980s. Due to warmer and wetter climate, the weather is no longer as representative of what we experience out in the field, so we needed to improve this model. This is where Machine Learning had a role to play and where SoftwareOne’s expertise was called upon in order to build a robust Machine Learning platform for inflow forecasting.”

“We need to predict the water inflow into all the hydroelectric facilities Hafslund Kraft has” added Heltberg. “This is an important factor: it helps Hafslund Kraft optimize the production planning of these plants. A deviation in inflow predictions can result in insufficient water to meet production targets or, in the worst case, water spilling over the reservoir's edge.”

A learning process

Hafslund Kraft partnered with SoftwareOne to develop a platform on Azure to facilitate the deployment of hybrid models, composed of physically-based and Machine Learning models, for predicting water inflow at various hydroelectric facilities. The initial roll-out targeted three hydroelectric facilities, with the possibility to expand the solution across Hafslund Kraft’s full network of 80 facilities.

For Hafslund Kraft, the project served as a valuable opportunity to deepen the company’s understanding of forecasting and the potential of Machine Learning to enhance predictions. “The project followed an iterative development process: as more data was added to the platform, the models could be retrained with the aim of improving forecasting accuracy.,” explained Anne Gunn Kraabøl, Leder Ressursgrunnlag (Head Hydropower resource management), at Hafslund Kraft. “One of the key aspects of implementing Machine Learning is to enhance Hafslund Kraft’s knowledge of how to implement Machine Learning techniques to solve complex problems.”

From SoftwareOne’s perspective, the human and process-driven approach was equally critical. SoftwareOne’s methodology emphasizes that people and processes are key to successfully implementing AI and delivering great results with the technology. When faced with a complex and changing environment such as inflow prediction, learning on the go is essential.

Practical implementation

Hafslund Kraft has currently twelve Machine-Learning models in production, each corresponding to a catchment area where snowmelt and rainfall accumulate. Sensors in the catchment area collect a lot of different data, from temperature, precipitation, unregulated waterflow to snow. These sensors connect to Hafslund Kraft’s central database where it is being used in the inflow predictions. These sensors collect important data which is an important input into the Machine Learning algorithms.

“With help from SoftwareOne, we built a Machine Learning platform in Azure, which integrates with our other systems,” said Heltberg. “We continuously feed it with new data, and it feeds us with new predictions, which go back into our database. This ends up in the model used to optimize production planning.”

Increased reliability in Machine Learning for inflow predictions

Once Hafslund Kraft implemented the new platform, the results quickly justified the investment. “We were cautious at ,” explained Kraabøl. “We didn’t know if the outscaling of the Machine Learning models would provide good predictions for the new catchments. It is also dependent on the quality of our existing data, such as inflow, temperature, and precipitation.  It was extremely rewarding to see how technology could be used on our vast historical data to improve inflow forecasting.”

Heltberg elaborated on the challenge of predicting flooding events: “The most difficult forecasts are for big floods – identifying precisely when the flood peak will arrive. By combining Machine Learning with our existing physically-based model, we hope to predict these more accurately than before. According to our own benchmarking so far, we see that the hybrid models are better at predicting the floods. However, we see that the hybrid models do not perform as good during the winter with low inflow.”

For Hafslund Kraft, this project represented a ‘hydrological’ step forward. The company was able to marry its wealth of scientific data and domain knowledge with SoftwareOne’s contribution: the technical infrastructure and architecture of Azure, along with the coding expertise to create Machine Learning pipelines, so that Hafslund Kraft could enhance its inflow planning.

“This project took us much further than we had been before,” said Heltberg. “Together with SoftwareOne, we have built a platform that has made our work easier and more accurate. Since it was designed as a modular system, we can easily edit individual components without breaking anything.”

“And for us at SoftwareOne, we have also gained valuable insights into the unique challenges faced by hydro production companies through our work with Hafslund Kraft. This experience has been enlightening, and we are eager to leverage this knowledge to enhance our partnership in future projects.”  adds Valsdottir.

The solution provided Hafslund Kraft greater insight into when to adjust hydroelectric production. Importantly, it has also increased reliability in inflow predictions, which is particularly critical during flood situations, especially ahead of anticipated heavy rainfall in the summer months.

Looking ahead, Hafslund Kraft is exploring opportunities to extend the Azure platform to other areas of its business, such as wind power, to further enhance forecasting and resource management.

Next steps

The scalable way that SoftwareOne devised the platform on Azure means that it is easily replicable in other fields, for example, in wind power production planning. This is attractive to Hafslund Kraft as a company with diverse operations in multiple fields.

“We worked very well together with SoftwareOne,” said Kraabøl. “The way we have developed this platform makes it easy to scale and adopt new Machine Learning techniques or use it on other datasets.”

Why SoftwareOne

Hafslund Kraft identified that it needed a technology partner to help build and manage a platform in Azure, so that it could harness the benefits of Machine Learning. SoftwareOne was able to provide a collaborative team, marrying its technological understanding and expertise with Hafslund Kraft’s domain and scientific knowledge.

SoftwareOne’s team came with in-depth experience of building, using and managing projects on the Azure platform, so were able to transfer these skills to Hafslund Kraft, equipping the company to manage the project themselves in future. The prospect of expanding this knowledge transfer to other parts of Hafslund Kraft’s operations, such as wind power, was an additional attraction.