Machine learning to optimize the success rate of offshore maintenance voyages
Deputy Head of AI Center of Excellence
Elia is responsible for managing the MOG OSY platform, which transmits power from 4 offshore windfarms totaling 1 GW of capacity. This task requires regular maintenance of the platform, for which Elia crew must travel to the platform by air (via helicopters) or sea (via CTVs Crew Transfer Vessels). These journeys are not without risk, and some trips to the platform were unsuccessful, meaning the crew ultimately did not reach the platform.
The AI Center of Excellence, together with the Offshore team in Belgium, set out to build a tool to recommend the optimal time to make a trip to the platform. To do this, we use past data of the local weather conditions and historical daily progress reports (DPRs) recorded by the team. This past data is used to train a machine learning model that advises the team whether the weather conditions will lead to a successful trip out to the platform. An experienced member of the crew still makes the final decision as to whether or not to make the final voyage. This additional advice from the model helps save both time of the crew not having to go out on voyages that are ultimately not successful, as well as the money lost from unsuccessful voyages.
Data Preparation and Analysis
First, the data had to be extracted from the daily progress reports, which were stored in Microsoft Word files. This step involved working closely with the colleagues from the Offshore team to understand the reports. The weather data also needed to be prepared.
Once the data was ready, the team spent time analyzing the data to see if there were recognizable patterns that lead to platform approach failures, visualizing the main factors influencing voyage success, such as wave height.
Model Training and Evaluation
Multiple machine learning models were then trained using the prepared and analyzed data. This step proved difficult given the limited number of past failures on which to train the model. Finally, the best-preforming model was put to the test against a real expert from the offshore team. The model showed fewer “false positive” errors than the Elia crew member, which in this case means an error where the model said the trip would be a success, when in fact it was not successful.
Next Steps and Implementation
Given the number of failed voyages to the platform that occurred at the start of the project, and the performance of the model developed, the tool would have saved Elia 33.5% in costs incurred from unsuccessful voyages. However, over the course of project, the distribution of the underlying data changed, which changed the amount of savings the tool could deliver: How did this happen? As time went on, the offshore team suffered far fewer incomplete voyages than they previously had. Furthermore, given that the platform is now in operational phase and no longer in the initial installation phase, the team foresees fewer trips in general, and in particular fewer trips that must be completed in a very short time frame. For this reason, the Offshore team decided against implementing the tool. However, they found a lot value in the insights gained in the exploratory analysis phase of the project.