One of the main barriers being faced to implement final applications based in machine learning is the complexity of the results triggered from the algorithms to be understandable at business level.
The trend in the sector consists of implementing algorithms easy-to-understand to entry into the market. This is, of course, a good approach to begin, but it means that only a small potential of machine learning is putting into value and the most powerful techniques are still to be discovered.
In the following post we will explain the different techniques and new results that machine learning may offer to users at business level. Specifically, we will dive into clustering technique to illustrate Isotrol’s use cases and support opportunities identification, boosting machine learning technology implementation and interpretation.
Machine learning, and artificial intelligence in general,is becoming widely popular within the renewable energy sector. As example, anomaly and low efficiency detectionfor O&M optimization are two of the most extended fields of action, considering how they affect costs.
Machine learning techniques are divided into two main groups:
- Supervised learning: popularly based in regression techniques.
- Unsupervised learning: mainly supported by clustering techniques.
Most of available prediction and assets monitoring tools are based in the first group.
We know that the results obtained from a regression technique are easily interpretable by final users, as it commonly represents the predicted value for a predefined signal. Tools supported by a regression approach are typically based in the continuous comparison of the real value of a signal versus a predicted value by the regression algorithm, so a warning or alarm is reported to the user when a significant difference between values occurs.
Image 1 Regression approach
Image 2. Example of the results from a prediction based in regression techniques.
In the Image 1, v3 represents the signal to be predicted. Image 2 illustrate a specific example where v3 is “Generator bearing temperature”, compared with its predicted value. In image 2, a red frame appears when the real value of “Generator bearingtemperature” differs from its predicted value considering a level of confidence.
The real value of clustering techniques
As mentioned, regression techniques are more extended than clustering ones, not such popular nowadays.
But, why they aren’t? From the technical point of view, this make no sense, as the unsupervised learning techniques are indeed powerful. According to our researches, the main barrier for this technique to entry the market is the need of bringing togethermachine learning and business knowledge to successfully select the business case to be faced through clustering and interpret the results coming from a clustering algorithm.
These two points are not that critical in regression just because its fundamentals are easier to understand for the public than these of clustering, so no deep knowledge in machine learning is needed to identify or evaluate results.
On the other hand, although clustering approach requires higher analytical skills by the users, the additional benefits of implementing this techniquemay deserve it.
Beyond the identification of signals with deviated behavior (as in regression), the result from a clustering analysis includes the following advantages:
- Multi-behavior identification and characterization in a single step (clusters).
- Identification of new behavior(new clusters) appearing, never seen before in the historical data.
- Characterization of each behavior appearing in the data (provided by the centroids).
Then, clustering techniques not only provide the value of one specific signal of the system but also all possible interactions between every available signal which add information about the problem to be solved or process to be monitored.
The next image represents the schema for a clustering approach that may be compared with image 1, which represents the schema for a regression approach.
Image 3. Clustering approach.
Practical applications of clustering in wind farms and PV stations
We in Isotrol have been recently checking the benefits of applying clustering techniques in renewable assets’ operations and maintenance. Taking our own experience into consideration, the main conclusion is that these techniques provide new and interesting possibilities for renewable plants’ owners and managers. The following use cases support this outcome:
Image 4 Clustering in the wind power curve
Image 5. Clustering in Active Power Vs Time and DC Current Vs Time from PV systems.
As we can see, by incrementing a little the effort of explaining and plotting results, interpretation of clustering results become easier, especially when they are supported by business cases.
Reliable implementation of these techniques is only possible when they are supported by a robust technical structure and team whose pillars relies in high specialization in both, machine learning and renewable energy generation, as Isotrol´s team.
We in Isotrol believe in the value of a continuous innovation and technology development to enhance our products. Our main aim is giving our best to our customers and the energy sector. That is why we keep working specifically on identifying new use cases to bring to the sector, supported by the last technologies and, more important, triggering the results in the business language to boost the real potential of our tools and services.