MDA Renewables specialise in finding suitable locations and obtaining planning permissions for on-shore wind and solar installations. They also operate and manage some of their own turbines in South Wales.

Large wind farms generally run SCADA monitoring systems which is often price prohibitive for small and single installations. We were asked to come up with a solution for excessive downtime on a particularly remote turbine in one of the Welsh Valleys during its initial 'teething' period, which would often stop functioning and lose significant periods of available energy generation.

After exploring various options, a 2 part system was chosen. We decided to program a small embedded device to sit at the turbine, and a set of APIs to receive and process data in the cloud. The turbine had a small LAN in the control housing for various monitoring and remote access systems so we worked with the suppliers of the turbine to expose some read-only items related to the health and generating state of the machine on the LAN.

We used one of the team's favourite 'Made in Wales' products, a Raspberry Pi, for the embedded device. This was chosen for ease of programming, low cost and ability to replace without too much technical knowledge in case of a hardware failure. The device gathers data from the turbine at set intervals, collates, and then transmits it to the APIs at larger intervals.

The APIs store the data and aggregate daily figures. If a turbine indicates there is windspeed to allow generation to occur, but no power is being generated, stakeholders are contacted automatically. If a turbine misses more than 1 data submission window, stakeholders are again contacted, providing a dead man's switch style system to a total system failure in the control housing.

This system has transformed the generating abilities of the turbine and allowed tricky teething issues to be diagnosed as soon as they occur, and fixed with software or hardware upgrades.

With the turbine now streaming a continuous record of wind conditions and power output to the cloud, we had exactly the data needed to go a step further than monitoring. We trained a machine learning model that predicts how much electricity the turbine should be generating from the prevailing conditions - learning the relationship between wind speed and direction, air density, time of day and the turbine's own performance curve.

Fed with short-range wind forecasts for the site, the model produces a rolling forecast of expected generation, giving MDA Renewables a view of the energy they can expect to bring online in the hours and days ahead. The same predictions sharpen our fault detection: instead of only alerting when wind is present but no power is generated, the system now compares live output against the model's expected figure and flags any meaningful shortfall, catching underperformance long before it would show up as a total stoppage.

• PHP
• Python
• scikit-learn
• Google AppEngine
• Windows IoT
• C#