TL;DR
This paper introduces a novel condition monitoring system for wind farms using probabilistic multi-layer perceptrons (PMLPs) with transfer learning. The system accurately predicts wind turbine power output under normal conditions, enabling early detection of anomalies and outperforming existing methods in key performance metrics.
What Problem Does This Paper Address?
Wind farms, while crucial for renewable energy, face high operational and maintenance costs due to frequent failures and challenging environmental conditions. Effectively monitoring these turbines is essential to reduce downtime and costs. Existing methods often rely on SCADA data but struggle with the inherent complexities of wind farm operations, such as:
- Data Volume and Variability: SCADA systems generate vast amounts of data over long periods, requiring models that can scale and adapt to weather fluctuations.
- Data Sparsity: Individual turbines may have significant periods of missing data due to maintenance or installation, making it difficult to build reliable models for them alone.
- Need for Probabilistic Predictions: Simple point estimates of power output are insufficient; probabilistic predictions are needed to quantify uncertainty and enable robust anomaly detection.
- Inter-Turbine Dependencies: A wind turbine's performance is influenced by its neighbors. Effective monitoring should leverage this collective information.
Methods
The core of the proposed system is a Probabilistic Multi-Layer Perceptron (PMLP). This neural network is designed to predict not just the expected power output but also its variance, assuming a normal distribution with input-dependent mean and variance (heteroscedasticity). Key methodological aspects include:
- Branching Architecture: The PMLP features a "common branch" for initial layers, followed by two separate "branches" for predicting the mean and the standard deviation of the output power. This allows for flexible modeling of both aspects.
- Transfer Learning (Fine-tuning): To address data sparsity and leverage information across turbines, a Large PMLP (LPMLP) is introduced. This involves pre-training a model on data from all turbines in the farm and then fine-tuning it for individual turbines. This "borrowing strength" approach significantly improves performance, especially for turbines with limited data.
- Condition Monitoring with CUSUM Charts: The probabilistic predictions from the PMLP are integrated into a Cumulative Sum (CUSUM) control chart. This statistical tool monitors deviations from the predicted normal behavior, flagging potential anomalies. The CUSUM chart is specifically adapted for wind farm data.
Datasets
The study utilizes SCADA data from six Senvion MM92 wind turbines at Kelmarsh wind farm in the UK. The dataset spans from January 2016 to July 2021, containing over 1.7 million data points with 110 variables, including 10-minute averages, standard deviations, minimum, and maximum values for various operational parameters. Data filtering was performed to exclude periods of standby, warnings, and operational stops, ensuring the training data represented normal behavior. The final filtered dataset comprised 846,968 data points.
Key Insights and Ideas
- Probabilistic Predictions are Crucial: Simply predicting the mean power output is insufficient for effective condition monitoring. Modeling the variance (heteroscedasticity) is vital for accurately assessing deviations and understanding uncertainty.
- Transfer Learning is a Game-Changer for Data Sparsity: The LPMLP approach, leveraging data from all turbines to pre-train a model, significantly boosts performance for individual turbines, especially those with incomplete historical data. This is a practical solution for real-world wind farm scenarios.
- CUSUM Charts Enhance Anomaly Detection: Integrating probabilistic predictions into CUSUM charts provides a statistically sound and interpretable method for flagging potential faults, offering a clear visual representation of system status.
- Architecture Matters: The paper explores different PMLP architectures (A1 and A2), demonstrating that while both perform well, architectural choices and the transfer learning strategy contribute to improved performance.
Key Results
The proposed PMLP and LPMLP models, particularly with architecture A1, demonstrated superior performance compared to a sparse Gaussian process and a Bayesian neural network across several metrics:
- Reduced Prediction Errors: LPMLP models achieved the lowest Root Mean Square Error (RMSE) and Mean Absolute Error (MAE). For instance, LPMLP-A1 achieved an RMSE of 27.38 kW and MAE of 14.81 kW.
- Improved Coverage Probabilities: The models exhibited good coverage probabilities, meaning their predicted confidence intervals were more likely to contain the actual power output. LPMLP-A1 achieved 94.56% 95% coverage and 98.40% 99% coverage.
- Faster Training Times: The LPMLP models, due to transfer learning, offered significantly reduced training times compared to traditional methods like sparse Gaussian processes. LPMLP-A2 fine-tuned in just 2 minutes after pre-training.
- Early Fault Detection: The CUSUM control charts successfully identified anomalies, with one example showing an early warning 3 days before a recorded forced outage. The system achieved a recall of 0.77 and precision of 0.68 with a decision interval of 15, providing an average lead notice time of 15.49 hours.
Limitations and Future Work
- Component-Level Diagnosis: While the system detects anomalies, it doesn't pinpoint the specific component causing the issue. Future work could extend the methodology to monitor other SCADA variables (e.g., temperatures, pressures) to achieve component-level diagnosis.
- Transfer Learning Refinements: The paper uses parameter initialization for transfer learning. Future research could explore using parameter estimates to construct more informative prior distributions for Bayesian neural networks, potentially leading to even better performance.
- Automated Decision Interval Selection: While a method for selecting the CUSUM decision interval (I) is discussed, further work could focus on more automated and robust methods for this selection, especially in scenarios with limited fault data.
Summary updated