Machine Learning technologies to anticipate problems and improve the profitability of renewable plants

We know that the use of Artificial Intelligence (AI) and Machine Learning technologies is clearly effective for the early detection of problems in photovoltaic and wind plants. It is also known that innovative and well-prepared companies from the renewable industry are already increasing the profitability of their operations with such techniques. However, do we really know how?

There are many types of losses that, in addition to usually go unnoticed, are mostly detected after a period of operation, which sometimes could be months. Therefore, this results in inefficiencies. In photovoltaics, cases of these types of losses could be, for instance, string drops, tracker blockages (solar trackers) or the dirt on panels; while regarding wind power, with more complex machines, its internal regulation is normally due to overheat or misalignment with respect to the wind direction, for example.

Detecting problems could be a difficult task for several reasons: due to the variability of the resource itself, which tends to hide the reduction in production; because of the reliability and location of the sensors, for instance, in an extensive solar field, if a cloud passes over the meteorological station, it could be confusing; due to similar evidences for the same problem; or simply because of the large amount of data coming from the plants. According to the latter case, let’s think that a photovoltaic plant of about 50 MW can generate up to 80,000 measurements every day; and in order to get an idea of ??this magnitude, we can visualize it as a football stadium full of spectators, just every day!

Fig. 1: Power generated in two days, clear and cloudy measured at string box level.

This figure represents the measurement of power generated in several string boxes, in which we can observe some differences between them regarding production. This could be due to disconnected strings, blocked trackers, clipping (limitation) in the inverters, deterioration or dirt on the panels, etc.

Therefore, the challenge is to identify performance problems in this flood of variable and sometimes inconsistent data, in order to be able to carry out the proper O&M actions as soon as possible. The solution goes through advanced data analysis thanks to the application of Machine Learning techniques to these large volumes of information generated by the facilities. Continuing with the example of photovoltaics, we can develop and train an algorithm capable of detecting the moment in which the drop in production is asymmetric with respect to the forecasted measurement. In this case, it usually indicates that a tracker has been locked facing sunrise or sunset.

Fig. 2: Theoretical model of power generated at string level.
Fig. 3: Real model found by the algorithm measured at the level of string boxes (20 strings). The right curve has a reduction at the beginning.

In the previous case, an algorithm has been trained by implementing a measurement forecast and using neural networks, comparing it with the real measurement and analyzing the difference of both curves to identify any tracker blockage. Since sensorization usually adds a large number of strings and trackers, there can be an accumulation of them, which could be solved by training the algorithm for these cases. Additionally, such algorithms are also capable of calculating the lost energy associated with each of the aforementioned losses.

In a similar way, and talking now about wind power, with supervised learning techniques, the temperature signals of the critical elements of wind turbines (generator, gearbox, bearings, etc.) can be forecasted to detect situations, which —without becoming an alarm for the operator—they are quite out of the ordinary operations. In these cases, it is common for the manufacturer to implement an internal regulation that causes a performance loss or, what is more serious, it could result in a more severe problem.

Clustering is another example of a non-supervised learning technique available. In clustering, the different points of a curve are grouped into different modes of operation, allowing the signals that differentiate the operating states of a machine at every moment to be compared. The following figure shows different modes of operation in a power curve of a wind turbine: while in one of them it represents a normal performance, in another one —the yellow one— it has significant deviations in its performance.

Figure 4: Clustering of points on the wind power curve.

The main advantage of these systems is that the analyses are carried out on the data already available in the SCADA, without the need to invest in new sensors. Therefore, this results in a simpler and thus, cheaper commissioning.

In conclusion, advanced data analysis through Machine Learning is certainly a valuable help for asset managers and O&M teams to detect early problems in wind farms and photovoltaic plants. In this way, energy losses associated with unavailability cases are avoided right from the beginning.

At Isotrol, we have been performing these advanced analyses for several years through Bluence® Performance Diagnosis, which is the Big Data analytical tool in the cloud that is part of the data analysis solutions of our Bluence® platform. There it is possible to find a wide set of algorithms which performs data analysis on a daily basis, carrying out a prioritized action recommendation based on the problems identified.

Fig. 5: APD Screen – Plant situation Dashboard.
Fig. 6: APD Screen – Action suggestions.
Fig. 7: APD Screen – Out of normality analysis of critical signals.
Fig. 8: APD Screen – Classification of energy losses.

Comments are closed.