From total nightmare to wild dream
A wind turbine is a wonderful piece of mechanics that uses wind to generate electricity. The force of the blowing wind operates three blades that, connected to a rotor, spin an electrical generator. Under the hood of those gigantic turbines we sometimes see offshore, or on the top of a hill, a number of electronic components work orchestrated by a controller. In addition to the controller, in a turbine there are components like anemometer, blades, brake, gear box, low- and high-speed shaft, pitch, generator. A wind turbine has an average lifetime of about 20 years and during this time any component it is made of is subject to regular maintenance and, in some cases, replacement.
A solar panel is made of a number of solar cells. Each cell consists of four layers: an anti-reflective coating captures and retains sunlight thus ensuring that two underneath silicon layers can convert solar energy into electricity. The silicon layers and, more importantly, their actual chemical configuration is crucial for the process to work. At the very top of the cell, strips of conductor material move electricity out to its final destination. Electricity is generated as direct current that an inverter turns into alternate current.
Generally, a solar panel comes with a 20-year guarantee and usually it keeps working after that time at least at 80% of the original capacity. Solar panels may need direct as well as indirect maintenance. Direct maintenance is fixing or replacing parts. Indirect maintenance includes ancillary tasks like cleaning surface from dust, weeding and cutting grass.
How would you perform any due maintenance on these types of generation units?
Scheduled maintenance is the most common option. In this case, components are checked periodically according to a predetermined schedule that doesn’t necessarily take into account the actual state of the component, weather conditions or exceptional situations. It’s a good-enough approach with a number of drawbacks. For example, maintenance may be scheduled when it is not necessary given the workload of the component. Imagine a broken turbine that stays out of order for a long period because it is inaccessible due to severe winter conditions. In this case, the consumption of its components may be significantly lower than other turbines in the farm. When scheduled maintenance occurs on that particular turbine it could be a bit ahead of real need. Another drawback of scheduled maintenance is that, well, it may also happen too late when the component is already broken or seriously damaged.
Condition-based maintenance uses fixed rules to calculate the ideal time for maintenance based on the actual wear of components as reported by embedded sensors. Typically, a control room software would allow to trigger alarms based on KPIs calculated out of the raw numbers returned by sensors. The good news is that no unnecessary maintenance is carried out and by fine-tuning alarms one can reserve a margin to postpone or anticipate maintenance according to volatile conditions such as bad or good weather. The weakest point of condition-based maintenance is that alarms parameters must be set by the rule of thumb and are subject to the highly variable human ability to learn from numbers and mistakes.
It seems then that both scheduled and condition-based maintenance may work but none is perfect. The next step is predictive maintenance.
Predictive maintenance aims at hunching what would be a good time to intervene. It requires constant monitoring of the generation units and abundance of data from internal components. This is already the case with most commercial solutions out there that collect raw device data and aggregate into sensible indicators. Numbers alone aren’t sufficient though.
Analytical algorithms are required along with an effective data model that figures out the hidden patterns of when failures on a given type of component is more likely to happen. In addition to sensors that communicate state and performance numbers via industrial protocols such as OPC and its variations additional equipment and logic is needed. For example, sensors to measure noise, heat and humidity, ultrasonic microphones, and cameras. Arranging a data model that result in an alarm in presence of certain measurable conditions is not a trivial task. It requires machine learning techniques and empirical algorithms. It’s a good challenge for data scientists.
Predictive maintenance is not here yet but it is being talked about a lot meaning that it really addresses a core concern for the energy industry. The good part is that predictive maintenance is not sci-fi but it’s getting real and more production-ready every day thanks to the growing availability of cognitive services in most cloud and software platforms.
