In 2017, Google-owned research company DeepMind released a chess engine called AlphaZero. This complex, Neural Network-based program was given the rules of chess and told to play “itself”. After a few hours of training AlphaZero not only became capable of beating any human player, but was also better than any existing chess engine. Truly an inspiring example of machine learning!
In many ways, chess is similar to some of the real-world problems we see in the offshore wind industry. Pieces moving on the board in carefully pre-planned sequences are like technicians on vessels, moving between turbines requiring maintenance. Planning chess moves and offshore wind farm work are both immensely complicated challenges; the possible combinations of moves exceeding the number of stars in the observable universe.
So, we pose a question – could brutally efficient programs like AlphaZero be used to solve complex real-world problems, such as planning offshore wind farm maintenance? We believe so, yes. Will this happen in near future? Probably not, but we explain below why.
Chess vs. offshore wind maintenance planning
One of the key differences between chess and planning offshore work is the number and variability of model inputs. All the information that went into the AlphaZero Neural Network was the starting positions of all pieces, legal moves, constraints, and winning condition. When planning offshore maintenance, you need to take into account many more factors that change on a daily – and sometimes hourly – basis such as:
– Vessels available (including their spec)
– People available (including their qualifications)
– Weather forecasts
– Tasks to be done and their priority
– Locations of turbines and O&M base(s)
Chess would become much more complex if some squares became inaccessible to certain pieces at random points in the game (think turbine access restrictions), or if a piece attempted to capture another one but did not manage to (e.g. failing to complete a maintenance task).
Although most wind farms collate a vast array of O&M data, it is often stored in disparate systems (even in Excel spreadsheets), making it tricky to use as input data for optimisation tools. Since every wind farm uses a different set of tools, even if there was a brilliant work planning maintenance AI, it wouldn’t be easy to apply it across different sites.
Furthermore, in a system as complex as an offshore wind farm, human error could easily squander an optimal work plan devised by a state-of-the-art model. Wind farms need reliable systems to monitor and enforce (where possible) adherence to the plan.
Getting experienced planners to change their ways, just because a machine says it’s better to do so will be a challenge. It’s very likely that optimal work plans may depart from industry-wide practices (such as having teams on the same vessel all day), and end-users of such a model are likely to lack trust in its outputs.
Let’s say we have met all of the above prerequisites (each of which Sennen has been focusing on in order to drive operational efficiency and improve asset management), the next step is to ask the following question:
Are the costs associated with developing and applying an optimisation tool worth the potential gains to be achieved from optimised work plans?
While at this stage it’s almost impossible to calculate, we believe that in the long term, they are.
However, there is lower hanging fruit to be picked before day-to-day offshore wind maintenance is optimised by an AlphaZero-like model. And it’s solving these trivialities that happen frequently and require focus across many offshore sites:
– Ensuring that out of date qualifications do not cause last minute plan changes and delays
– Having systems and processes that enable real-time, flexible plan changes and ensure clear communication of those changes between planners/operations/marine control to avoid confusion and minimise human error
– Improved gathering, analysis and visualisation of operational data, to highlight any inefficiencies in planning and carrying out O&M (when and why things don’t go according to the plan)
By doing this, we’re building up the architecture and valuable datasets for training machine learning models.
In tackling the above issues, we’re getting wind farm operators to move their processes away from Excel spreadsheets and whiteboards. This is a necessity if we are serious about using AI for work planning optimisation. Not to mention the cost savings and health and safety benefits that stem from these (which you can read about here in our London Array case study here).
It’s a pleasure watching machine learning technologies develop at the pace they have been in the last 10 years. Undoubtedly, there is huge potential to apply those methods to helping us create better offshore work plans, ultimately leading to lower cost of offshore energy.
This revolution will not happen overnight. To begin with, smart algorithms could suggest groups of tasks to be done by maintenance teams to maximise the amount of work done while minimising technician time spent on vessel. Over time, machine learning algorithms could be introduced to learn from the plans humans have made in the past and improve on them. If that happens, the planner’s role could be reduced to checking and adjusting the work plans made by AI.
For now, we must focus on building a solid foundation for future models, so that they can be integrated seamlessly with the relevant data and procedures used for planning and monitoring offshore wind farm work.
Whether you’re playing chess, or planning work on an offshore wind farm, thinking only one move ahead won’t get you very far. Let’s not make any blunders in this game against climate change.
Sennen’s technology is deployed on almost 4GW of renewable energy projects and growing. Request our product data sheet here for information on our bespoke functionality and dashboards.