Predictive Maintenance For Hydro Power Generators

The Challenge

The Sira-Kvina power company is one of Norway’s primary renewable energy power providers, producing approximately 5% of the country’s power from seven hydro power plants. Sira-Kvina was struggling to tie their data together from disparate and heterogeneous operational systems, and thus unable to make informed decisions. On average, one of the hydro generators going down creates a loss of approximately $1 million dollars per day, as well as negative customer sentiment and trust. Sira-Kvina needed a trusted data and analytics partner and engaged Elder Research to hand-craft a data strategy that fit their organization and their ability to absorb transformation, and to design and deploy predictive models to determine hydro generator component failure risk to enable preemptive maintenance operations and system monitoring.

The Solution

Elder Research’s dedicated data science and engineering team worked hand-in-hand with Sira-Kvina leaders, operators, and stakeholders to design and build machine learning models using various operational and maintenance data sources. Our team developed an abstract model of the generator systems that could be linked to key operational and sensor data systems. In order to aid in analytic story-telling, change management, and adoption of data-driven decisions, the generator risk and other relevant data were represented in a visual layer using our Risk Assessment Data Repository (RADR) as a prototype for the predictive maintenance tool.


Elder Research’s focused data discovery and exploration identified that the generator had moved into a damaged state eight power cycles earlier than had previously been observed by the Sira-Kvina maintenance team. We created a predictive model that was trained on the SCADA sensor data and achieved a 500% boost in predictive performance above the baseline. Thus, our models helped Sira-Kvina operators identify issues more accurately (less false positives and false negatives) and much earlier, giving operators the opportunity to strategically approach each maintenance issue in a cost-optimal fashion. Using Elder Research’s streaming data anomaly detection system called CEPtor, sensor metrics were reduced from hundreds to the ten that provided the best predictive performance, making the results easier to interpret and more actionable for the business.