Predicting Disruptions in advance

Predicting Disruptions

in advance

Accurate predictions between seven days and one month in advance.

Worker in a modern factory identifing predictive disruptions in advance

Success Story

reading time: 3 min

Energy & Utilities

Predictive Quality

Predictive Maintenance

Anomaly Detection

100%

Automatic alarming

Accurate

Predictions between 7 days and 1 month in advance

Improved

Resource planning and efficiency

Context

Our client serves as the primary heating infrastructure provider for more than 2 million residents across one of Europe's largest capitals. A limited number of staff have to do inspections and maintenance work. There was an automatic warning system in place, but it didn’t not inform on the severity of the issue or it could be a false alarm. However, each warning has to be followed up, which meant that the employees had to abort their ongoing maintenance/inspection work to investigate the new warning. This was stressful and cumbersome, especially as the inspections and maintenance work team has limited resources.

Challenge

In a Proof of Concept (POC), it was essential to address the following three inquiries:

Can disruptions and the type of disruption be predicted in advance?
If yes, then how far ahead?
When does an ongoing maintenance need to be paused to investigate a more serious disruption/warning?

An additional challenge was to convert the annotations of workers into standardized information that can be used for modeling.

Assignment

Our task was to implement a model (A), which predicts if there would be a disruption within a time window. This time window was set to values between 1 day up to 1 month. An additional model (B), was implemented to predict the disruption type, in case model A predicted a high probability of a disruption.

Solution

Model A was a recurrent neural network (LSTM), which takes in a time series (i.e the last X hours of data) to predict whether a disruption will occur in the next Y hours/days. Model B was a Gradient Boosting Tree Ensemble to predict the disruption type (related to priority of warning).

We used Tensorflow, Gradient Boosting trees, pandas and Scikit-Learn for a solution that predicted and automatically reported disruption between 7 days and 1 month in advance.

With craftworks, it was simple and uncomplicated. After an initial conversation, we could start a project within a short amount of time. While the data scientists at craftworks supported us in drawing valuable conclusions from our data, the close collaboration together with their domain experts made it possible to have fast results within a month.

Stefan Kermer

Wien Energie, Head of Innovation and Strategic Projects