Source-linked AI summary
Predicting Wind Turbine Power Using Machine Learning and Weather Forecasts
Khivishta Boodhoo, Isaac Triguero, Josh Plumbly, Bruce Nicolson, Nicholas Watson
TL;DR
The study examines data preprocessing, data volume, and transferability challenges in wind-turbine power prediction. It designs a machine-learning pipeline and reports improved prediction performance, with low-power forecasts supporting maintenance scheduling.
Problem
Data preprocessing effects and model transferability across wind turbines require further study.
Method
An ML pipeline was designed to forecast wind-turbine power production and identify future low-power periods for maintenance scheduling.
Results
The ANN improved from R2 score of 0.94 to 0.98, while the model had a lower NRMSE (11% compared to 22%).
Takeaways & Limitations
Low-power predictions can identify future periods more suitable for scheduling maintenance.
Takeaways & Limitations
Limited similarity between wind turbines restricts full appraisal of methods for real-world deployment, and relevant data are often unavailable when deploying models on another dataset.
Abstract
from arXiv · showhide
Offshore wind turbines are widely used to generate renewable energy, but their maintenance can result in decreased efficiency due to forced shutdowns. Accurate wind turbine power predictions can identify periods of low power that would be ideal for scheduling maintenance. However, the effects of data volume, feature selection, and data preprocessing on the performance of such power prediction models have not been thoroughly studied. Besides, current models have limited transferability between different wind turbines. Therefore, this study developed a baseline Linear Regression for performance comparison with a more complex Artificial Neural Network model to predict the power output of a wind turbine, using weather conditions only to enhance applicability. A range of data preprocessing techniques were studied, and models were trained on one month and one year of data to determine the effects of data preprocessing and volume on model performance. Feature selection was explored using a Random Forest Regressor. The best results from the different models showed that the Artificial Neural Network models provided the highest accuracy, with an R2 score of 0.98 and a low Mean Absolute Error of 194, when compared with the baseline model (R2 score of 0.94 and Mean Absolute Error of 441). The model performance is comparable to the range of results in past studies, with the advantage that the proposed method leverages a separate weather dataset from a nearby weather station, enabling future applications for similar wind turbines in different locations. The Artificial Neural Network model was then used to identify 4-h periods of low power predictions over 2 months (simulating application for future periods), providing power output savings of approximately 2000 kW for each maintenance event.
1 Introduction
Offshore wind turbine maintenance can reduce output, motivating accurate power forecasts to identify low-generation periods. The study examines preprocessing, data volume, feature selection, model complexity, and transferability for machine-learning-based forecasting.
- Motivation: Offshore maintenance can disrupt power output because turbines require shutdowns, inspections, repairs, and lengthy access procedures.Routine maintenance is reported as approximately 40 h per year, with offshore operations adding logistical constraints.
- Motivation: Accurate forecasts can identify 1–6-h low-power windows suitable for maintenance scheduling and support operational decisions.Forecasts may also flag disparities between predicted and actual output and help manage supply-demand variability.
- Research challenge: Wind-power forecasting is difficult because wind speed is uncertain and volatile, while increasing wind generation requires accurate and reliable predictions.The difficulty is linked to the variable relationship between meteorological conditions and turbine performance.
- Research gap: Many models are evaluated only on training-related data rather than independent future or other-turbine datasets, limiting appraisal for real-world deployment.The paper specifically highlights limited transferability when input features are not readily available for another turbine.
- Research gap: Data preprocessing, feature selection, data volume, and their effects on model performance have not been studied thoroughly.The study also identifies unsynchronised data streams and irrelevant or redundant features as practical data challenges.
- Study aims: The study compares an ANN with an LR baseline while testing preprocessing, data volume, feature selection, and maintenance scheduling using weather data.It also demonstrates use of a separate weather dataset to broaden the method’s potential application.
2 Methodology
The methodology combines two datasets, preprocessing and feature-selection steps, and LR and ANN models. An independent nearby weather-forecast dataset is then used to assess deployment and maintenance scheduling.
- Preprocessing: Dataset 1 undergoes filtering, invalid-state removal, feature selection, four-minute resampling, outlier removal, and normalisation.The preprocessing addresses startup and shutdown effects, brake states, negative power values, abnormal observations, and scale differences.
- Models: The pipeline develops an LR baseline as a benchmark before evaluating a more complex ANN model.The methodology explicitly positions LR as a comparison point for ANN performance.
- Maintenance scheduling: Predicted energy outputs are used to identify future low-power periods for maintenance scheduling.The pipeline connects model deployment with operational timing decisions.
- Deployment data: The weather forecasts come from a MET station approximately 14.5 miles from the studied turbine and are used as independent deployment data.This design tests prediction in a scenario separate from the training and testing dataset.
- Deployment data: The weather dataset contains forecasts at hourly intervals extending up to 72 h into the future.The forecasts were assembled from structured tables originally available at four-hour intervals.
- Datasets: The study uses a turbine dataset for model training and testing and a separate weather-forecast dataset for deployment assessment.Dataset 2 evaluates performance independently from the data used to train and test the models.
2.3 Exploratory Data Analysis
The exploratory analysis visualized the turbine power curve with high-density scatter plots and classified four types of outliers. These patterns guided the study’s outlier-identification and removal procedure.
- High-density scatter plots show a main power-curve region at approximately 500 data points per pixel and denser regions above 2500 data points per pixel.
- The denser regions occur toward the bottom and top of the power curve.
- Type 1 outliers are bottom-stacked points associated with failures, outages, unplanned maintenance, or measurement and communication errors.
- Type 2 outliers reflect wind curtailment, which limits turbine power output below normal levels even when wind speed exceeds rated speed.
- Type 3 outliers are sparse points linked to extreme weather or random noise, while Type 4 outliers are top-curve stacks caused by wind-speed sensor failures.
- The study used power-curve visualization to identify and remove the described outliers.
2.4 Preprocessing of Datasets 1 and 2
The preprocessing pipeline filtered operational anomalies, selected weather-related features, reduced data resolution, removed multivariate outliers, and normalized model inputs and outputs. These steps were applied to prepare SCADA, met-mast, and weather-forecast datasets for modeling.
- The pipeline tracked data-point counts through startup filtering, brake-state removal, feature selection, resampling, missing-value removal, outlier removal, and repetitive-value filtering.Dataset 1 decreased from 28,604,042 initial SCADA data points per feature to 48,956 after the listed filtering stages.
- The preprocessing framework also aligned met-mast feature heights with the weather-forecast dataset and applied feature scaling to both datasets.
- Removing all brake-condition features reduced the number of power values less than or equal to zero.
- 99% of negative power-output data points were attributed to brake conditions, while the remaining 1% may have resulted from sensor errors.
- Random Forest feature selection reduced the initial 900 SCADA and met-mast features to weather-related inputs including wind speed, wind direction, and air pressure.
- Data were aggregated from 1-second measurements to 4-minute intervals to balance computational efficiency and model performance.The same 4-minute resampling was used before Mahalanobis-distance outlier detection to reduce computational load while retaining feature variability over time.
- Mahalanobis distance was selected over z-score and Euclidean distance because it accounts for covariance structure and unequal variable variances.It was used to remove outliers from the power curve.
- Min-max normalization placed model input features on a common scale using each feature’s minimum and maximum values.
2.5 Model Development
The study develops baseline Linear Regression and feedforward Artificial Neural Network models to predict wind-turbine power from weather data. It compares data volumes, preprocessing, feature inputs, and model configurations using standard predictive-performance metrics.
- Baseline Linear Regression: A baseline Linear Regression model predicts turbine power from weather data, using wind speed as its single input feature after Random Forest feature selection.
- Model selection: The ANN was selected as suitable for approximately 50,000 data points and multiple preprocessed features because it can perform on large datasets and generalise to unseen data.
- Artificial Neural Network: The feedforward Artificial Neural Network uses back-propagation to update weights and activation functions to transform hidden-node inputs into output estimates.
- Data preparation: Models used 60:20:20 training, validation, and testing splits, while a 40:30:30 split was also investigated without significantly affecting performance.
- Hyperparameter tuning: Grid search with three- and five-fold cross-validation varied batch size, epochs, learning rate, neuron count, and hidden-layer count to identify network configurations.
- Training design: Models were trained on March 2019 data and on January–December 2019 data to investigate how data volume affects predictive performance.
- Evaluation: Performance was assessed with R2, Mean Absolute Error, Root Mean Squared Error, and normalised Root Mean Squared Error.
2.6 Model Deployment- Maintenance Scheduling
The trained model uses 60-day METEO forecasts as deployment inputs to predict future turbine power. Four-hour prediction windows are then screened for low generation to guide maintenance timing.
- Forecast deployment: METEO forecasts from 01-01-2022 to 01-03-2022 were used to simulate future power predictions when weather forecasts become available.
- Deployment evaluation: Predicted power from the METEO dataset was compared with actual turbine output, while maintenance-presence data established an average maintenance duration of 3–4 hours.
- Maintenance scheduling: The predictions were aggregated into 4-hour periods to identify low-generation windows suitable for engineers to perform maintenance.
3 Results and Discussion
Results compare Linear Regression and ANN models across preprocessing choices, input features, and one-month versus one-year training datasets. The ANN generally achieved lower errors and higher fit, while longer training data and preprocessing improved performance and low-power scheduling supported maintenance decisions.
- Model comparison: The ANN model performed better than the baseline because its R2 was closer to 1 and its prediction error was lower.
- Data volume: 11% NRMSE was obtained for the preprocessed ANN trained on one year of data, compared with 22% for Linear Regression.
- Feature selection: Additional input features improved performance, although wind speed remained the most important and influential feature.
- Error performance: 194 MAE was reported for the ANN, lower than 441 for the regression baseline under the same preprocessing techniques.
- Preprocessing: Preprocessing produced higher accuracies and lower errors for ANN models when comparing actual power with predictions from different datasets.
- Independent forecast deployment: Predicted and actual power trends were similar for the 2022 comparison period, with an average power difference of approximately 600 kW.
- Maintenance scheduling: The maximum and minimum average energy outputs for 4-hour blocks were over 24000 KWh and approximately 500 kWh, with about eight periods below 1000 kWh.
4 Conclusion
An ML pipeline combined weather data, preprocessing, feature selection, Linear Regression benchmarking, and ANN modelling to forecast wind-turbine power. ANN performance improved with preprocessing and more training data, supporting low-power maintenance scheduling.
- Pipeline: The study designed an ML pipeline using SCADA and nearby weather-forecast data, with outlier removal, feature selection, normalisation, Linear Regression benchmarking, and ANN modelling.The ANN was trained with hyperparameter tuning using TensorFlow and Keras.
- Preprocessing and modelling: Fine-tuning considerably improved ANN accuracy, while preprocessing also produced a large effect for the baseline regression model.The conclusion reports improvements using MAE and R2, and separately highlights preprocessing effects on the baseline.
- Preprocessing and modelling: R2 increased from 0.31 to 0.94, demonstrating the importance of data preprocessing even for the simplest benchmarked model.
- Training-data volume: With one year instead of one month of training data, the baseline improved from R2 0.81 to 0.94 and NRMSE 33.60 % to 22 %.
- Training-data volume: With one year instead of one month of training data, ANN performance improved from R2 0.94 to 0.98 and NRMSE 21.09 % to 11 %.
- Model comparison: On one year of data with the same preprocessing, ANN achieved NRMSE 11% versus 22% for baseline regression and had deployment error 610.83 lower.
- Maintenance scheduling: Approximately 2000 kW could be saved in a 4-h period by scheduling maintenance during low predicted power output.ANN predictions were deployed on weather forecasts to calculate average energy outputs in 4-h blocks and identify low-production periods.
5 Data Availability
Datasets related to the article are available through the ORE Catapult data-collection site.
- The article’s datasets can be found at the ORE Catapult data-collection URL.
6 Declaration of Competing Interest
The authors declare no known competing financial interests or personal relationships that could have influenced the reported work.
- The authors declare no known competing financial interests or personal relationships that could have influenced the reported work.
7 Data Availability
Datasets related to the article are available through the ORE Catapult data-collection site.
- The article’s datasets can be found at the ORE Catapult data-collection URL.
8 Declaration of Competing Interest
The authors declare no known competing financial interests or personal relationships that could have influenced the reported work.
- No known competing financial interests are declared.
- No personal relationships are declared that could have influenced the reported work.
- The declaration covers interests or relationships that could have appeared to influence the work.