**Authors:** Paweł Knes, Phong B. Dao
**Published:** _Energies_, 2024
**Link:** [MDPI Energies](https://doi.org/10.3390/en17205055)
---
### Summary
This study presents a comparative analysis of different Machine Learning ([[ML]]) models and a Cointegration-based method for monitoring and detecting faults in wind turbines. It proposes a cascading monitoring framework that uses [[ML]] for long-term predictions and #Cointegration for short-term confirmation of anomalies, aiming to improve reliability and reduce false alarms.
---
### Key Points
- **Machine Learning Models:**
- Uses Artificial Neural Networks #ANN, Linear Regression #LR, Random Forest #RF, and Gradient Boosting Tree #GBT.
- These models are effective for early fault detection but require large datasets and high computational costs.
- RF and GBT were found to be the best performers for long-term predictions.
- **Cointegration Method:**
- Originating from econometrics, it identifies **long-term relationships** between turbine parameters (like wind speed, temperature, power generation).
- Simpler and requires less data compared to [[ML]] but less effective for **early detection**.
- Ideal for confirming **short-term anomalies** detected by [[ML]] models.
- **Cascading Monitoring Framework:**
- [[ML]] models provide **early fault predictions** (outer monitoring).
- Cointegration **validates anomalies** flagged by [[ML]] (inner monitoring).
- This combination reduces false alarms and pinpoints immediate risks.
---
### How It Works
1. **Data Collection:** SCADA data (Supervisory Control and Data Acquisition) from a 5-year wind turbine dataset.
2. **Training & Validation:**
- [[ML]] models are trained on long-term data to predict faults.
- Cointegration is used to detect anomalies in real-time.
3. **Monitoring:**
- When an [[ML]] model flags a fault, the Cointegration model verifies the event before triggering an alarm.
---
### Performance Results
- **[[ML]] Models:** Detected faults **38 days** before actual failure.
- **Cointegration Model:** Detected faults **5.5 days** before failure.
- **Combined Approach:** Reduced **false alarms** and improved **fault verification**.
---
### Strengths
- Early detection of faults with [[ML]].
- Reliable short-term validation with #Cointegration.
- Minimizes **false positives** and **unnecessary shutdowns**.
- Lower **computational cost** for short-term monitoring.
---
### Limitations
- [[ML]] models require **large datasets** and **high computation**.
- Cointegration alone is not sufficient for very early detection.
- Needs well-structured SCADA data for optimal performance.