Study: overcoming fault detection and classification in large-scale solar PV farms

 

  • Scientists from Malaysia and Thailand have developed a novel machine-learning model for predicting the maintenance needs of large-scale solar PV plants.
  • According to a recently published scientific paper touting the efforts, the model utilizes two machine learning techniques – K-Means and long-short-term memory (LSTM) – and has a root mean square error (RMSE) of 0.7766.
  • The aim of the tool is to overcome “fault detection and classification” often found in traditional operating systems, the paper claims.

“The conventional operation and maintenance (O&M) of solar PV systems do not utilize machine learning for fault detection and classification,” the paper states. “This poses challenges for plant operators, especially those managing large-scale solar (LSS) PV plants, who typically rely on manual approaches to screen large amounts of electrical data and inspect numerous string panels. Consequently, the cost of O&M is high.”

K-Means is an algorithm for data segmentation and groups similar data points into clusters. The researchers use the algorithm to cluster the electrical current of string modules with environmental factors such as global irradiance and module temperature. A center or mean point is then created for the cluster to represent typical behavior.

Then the LSTM technique, trained on historical data, kicks in. The technique aims to detect anomalies in the predicted electrical current of string modules, which would then alert operators of maintenance needs.

“LSTMs can handle sequential data with variable length input by using gate mechanism to decide which information is important to persist and which to discard in each time step, thus making predictions based on the past trends and patterns in the input sequence,” the paper states. “LSTMs can do this by using a special type of memory cell that can store information for long periods of time, as well as gates that control the flow of information into and out of the cell.”

The data for the training, as well as the method’s precision analysis, is based on information sourced from a large-scale solar PV plant located in the center of Malaysia. A turnkey and sub-inverter were used as test cases, monitoring 420 string modules and a total of 8,400 PV modules. Compared to the collected data, the model has a root mean squared error (RMSE) of 0.7766.

The relative error is then compared with the number set by the benchmark model based on Artificial Neural Network (ANN). “LSTM and ANNs are often compared because they are both belong to neural networks and commonly used in various natural language processing, computer vision and speech recognition tasks,” the paper states. They found the LSTM is more precise with an average LSTM relative error of 4.316% and an ANN relative error of 4.363%.

The algorithm was described in the study “Anomaly detection using K-Means and long-short term memory for predictive maintenance of large-scale solar (LSS) photovoltaic plant,” which was recently published in Energy Reports. The research group was formed by scientists from the University of Technology Malaysia and the Chiang Mai Rajabhat University.

Author: Lior Kahana

This article was originally published in pv magazine and is republished with permission.

Leave A Reply

About Author

Green Building Africa promotes the need for net carbon zero buildings and cities in Africa. We are fiercely independent and encourage outlying thinkers to contribute to the #netcarbonzero movement. Climate change is upon us and now is the time to react in a more diverse and broader approach to sustainability in the built environment. We challenge architects, property developers, urban planners, renewable energy professionals and green building specialists. We also challenge the funding houses and regulators and the role they play in facilitating investment into green projects. Lastly, we explore and investigate new technology and real-time data to speed up the journey in realising a net carbon zero environment for our children.

Receive the week’s most popular stories in your inbox every Saturday morning