Fault Prediction and Early-Detection in Large PV Power Plants Based on Self-Organizing Maps

Alessandro Betti, Mauro Tucci, Emanuele Crisostomi, Antonio Piazzi, Sami Barmada, Dimitri Thomopulos2021 · DOI

TL;DR

This paper introduces a novel data-driven approach using Self-Organizing Maps (SOMs) to predict and detect faults in large photovoltaic (PV) power plants up to 7 days in advance. By developing a unique Key Performance Indicator (KPI) based on cell occupancy, the method effectively identifies anomalous behavior, enabling proactive maintenance and potentially significant energy gain.

What Problem Does This Paper Address?

Maintaining the optimal performance of large-scale photovoltaic (PV) power plants is crucial for their economic viability and reliable energy production. Traditional maintenance strategies often rely on reactive measures or scheduled interventions, which can be costly and inefficient. While analytical monitoring systems exist, accurately detecting and predicting malfunctions in complex PV systems, especially those with a vast number of interconnected components like inverters, remains a challenge. The sheer volume of data generated by these plants, coupled with the intricate interactions between electrical and environmental factors, makes it difficult to model and predict failures using conventional methods. This paper aims to bridge this gap by developing an automated, unsupervised, and accurate methodology for early fault prediction in large PV plants, moving beyond simple anomaly detection to proactive forecasting.

Methods

The core of the proposed solution lies in a data-driven approach utilizing a Self-Organizing Map (SOM), a type of unsupervised artificial neural network. The SOM is trained on historical SCADA data representing nominal (healthy) operating conditions of the PV plant. This training process creates a low-dimensional (typically 2D) map that preserves the topological relationships of the high-dimensional input data, effectively modeling the "normal" behavior of the system.

A novel Key Performance Indicator (KPI), termed "Monitoring of Cell Occupancy," is introduced. This KPI quantifies how the observed operational data deviates from the learned nominal behavior. Specifically, it measures the change in the probability of cells within the trained SOM being activated by new data points compared to the training data. A significant deviation in this probability indicates an anomaly.

The methodology involves:

  1. Data Pre-processing: Cleaning and preparing SCADA data, including handling missing values (using k-Nearest Neighbors imputation), removing statistical outliers, and detrending seasonal trends.
  2. SOM Training: Training the SOM on a dataset of normal PV plant operations.
  3. KPI Calculation: For new, incoming data, calculating the KPI daily based on the deviation of cell occupancy from the trained nominal state.
  4. Warning Generation: Establishing thresholds and rules based on the KPI's value, its derivative, and its persistence over time to generate warning levels of varying severity. This helps distinguish between minor fluctuations and genuine impending faults.

Datasets

The study utilizes historical SCADA data and operation and maintenance (O&M) logs from three different PV plants:

  • Plant A: Located in Romania, with an installed capacity of 9.8 MW, equipped with 35 inverter modules of manufacturer type 1.
  • Plant B: Located in Greece, with an installed capacity of 2.8 MW, equipped with 7 inverter modules of manufacturer type 1.
  • Plant C: Located in Greece, with an installed capacity of 4.9 MW, equipped with 25 inverter modules of manufacturer type 2.

The data spans testing periods of at least six months, with one plant having a full year of data. The SCADA data includes various electrical (DC/AC current, voltage, power) and environmental (irradiance, temperatures) signals. The O&M logs provide timestamps and descriptions of actual fault events. The paper does not introduce new benchmark datasets but rather uses real-world data from these operational PV plants.

Key Insights and Ideas

  • Unsupervised Learning for Unbalanced Data: The paper effectively leverages unsupervised learning (SOMs) to handle the inherent imbalance between nominal and faulty data in PV systems. This is a significant advantage over supervised methods that struggle with rare fault occurrences.
  • Predictive Power Beyond Detection: The core innovation is the shift from simply detecting anomalies to predicting incipient faults. The KPI's ability to signal issues days in advance is a critical step towards proactive maintenance.
  • Generic Fault Prediction: The method focuses on predicting generic faulty conditions rather than specific fault types. This makes it broadly applicable across different inverter technologies and plant configurations, requiring only historical SCADA data and a fault taxonomy.
  • KPI as a Health Indicator: The developed KPI, based on cell occupancy deviation, serves as a robust indicator of the plant's "health," with a value of 1 representing normal operation and decreasing values signaling degradation.
  • Multi-level Warning System: The tiered warning system, incorporating KPI thresholds, derivative, and persistence, allows operators to prioritize responses based on the severity and likelihood of an impending fault.

Key Results

The proposed SOM-based monitoring system demonstrated strong performance across the three case studies:

  • Early Prediction: The system was able to predict incipient faults an average of 6 to 7 days in advance of their occurrence.
  • High True Positive Rate (TPR): The TPR consistently exceeded 92% across all test cases, indicating a high ability to correctly identify actual faults.
  • Low False Negative Rate (FNR): The FNR remained low, with values ranging from 2% to 8%, meaning very few actual faults were missed.
  • Manageable False Positive Rate (FPR): The FPR varied, with the lowest being 1% for Plant C and the highest 18% for Plant B. While some false alarms are generated, the system's logic (especially using the KPI derivative) helps minimize them.
  • Potential Energy Gain: For Plant A, the analysis suggested that timely interventions based on the SOM predictions could have led to an energy gain of up to 20%.

Limitations and Future Work

  • Generic Fault Classification: The current model predicts generic faulty conditions but does not classify the specific type of fault. The authors acknowledge this limitation and are developing a supervised fault-classification tool to integrate into the system.
  • Data Quality Dependence: The performance relies heavily on the quality and availability of historical SCADA data and O&M logs for training and validation.
  • Online Deployment: While the system is being installed for online monitoring, a full evaluation of its real-time performance and feedback from plant operators is pending.
  • Threshold Tuning: The thresholds for warning levels were tuned based on the specific case studies. Further validation and potential adaptive tuning mechanisms might be necessary for broader deployment.

The authors express optimism about the system's potential and are actively working on enhancing its capabilities, particularly in predicting specific fault classes.

Summary updated

BibTeX
@misc{betti2021fault,
  author = {Alessandro Betti and Mauro Tucci and Emanuele Crisostomi and Antonio Piazzi and Sami Barmada and Dimitri Thomopulos},
  title = {{Fault Prediction and Early-Detection in Large PV Power Plants Based on Self-Organizing Maps}},
  year = {2021},
  doi = {10.3390/s21051687},
}
Figure