Temporally Correlated Deep Learning-Based Horizontal Wind-Speed Prediction

Lintong Li, Jose Escribano Macias, Mingwei Zhang, Shenghao Fu, Mingyang Huang, Xiangmin Yang, Tianyu Zhao, Yuxiang Feng, Mireille Elhajj, Arnab Majumdar, Panagiotis Angeloudis, Washington Ochieng2024 · DOI

TL;DR

This paper introduces a novel approach to predicting horizontal wind speed by leveraging a comprehensive set of Quality Indicators (QIs) and employing temporally correlated deep learning models. The study demonstrates that Bidirectional Long Short-Term Memory (BiLSTM) networks, when trained with these enriched datasets, significantly outperform traditional methods, offering more accurate and reliable wind speed forecasts crucial for aviation safety and renewable energy optimization.

What Problem Does This Paper Address?

Accurate wind speed prediction is paramount for a multitude of applications, ranging from optimizing wind energy production to ensuring flight safety in aviation. Traditional methods often struggle with the inherent complexity, non-linearity, and temporal dependencies of wind patterns. Existing machine learning models, particularly those that treat data as independent points in time, fail to capture the crucial sequential nature of wind speed, leading to significant prediction errors, especially during sudden wind changes. This can result in operational delays, increased accident risks, and reduced efficiency in energy generation. The paper aims to bridge this gap by developing a more robust prediction model that accounts for temporal correlations and a wider array of influential meteorological factors.

Methods

The researchers explored both traditional point-wise machine learning models and advanced temporally correlated deep learning models.

Point-wise Models: * Polynomial Regression: A simple model fitting a polynomial function to the data. * Random Forest: An ensemble method that builds multiple decision trees. * Multilayer Perceptron (MLP): A foundational neural network architecture. * ARIMA (AutoRegressive Integrated Moving Average): A statistical model for time series forecasting.

Temporally Correlated Deep Learning Models: * LSTM (Long Short-Term Memory): A type of recurrent neural network (RNN) designed to handle sequential data and long-term dependencies. * GRU (Gated Recurrent Unit): A simplified version of LSTM, also designed for sequential data. * CNN-LSTM: Combines Convolutional Neural Networks (CNNs) for feature extraction with LSTMs for time-series prediction. * Bi-LSTM (Bidirectional Long Short-Term Memory): An extension of LSTM that processes the input sequence in both forward and backward directions, allowing it to capture context from both past and future data.

The study also introduced a novel set of Quality Indicators (QIs) beyond the standard atmospheric pressure, relative humidity, and temperature, including factors like divergence, cloud cover, geopotential, and ozone mass mixing ratio. Feature importance analysis was conducted using the Permutation Feature Importance (PFI) method to identify the most influential QIs. Data normalization to a [0, 1] range was applied to all QI values before feeding them into the models.

Datasets

The primary dataset used in this study is from the ECMWF Reanalysis v5 (ERA5). This dataset provides hourly estimates of atmospheric, ocean-wave, and land-surface quantities from 1950 to 2022. For the experiments, data from 31 May and 1 June 2015 were selected to capture changeable weather conditions. The analysis focused on four capital cities (Beijing, London, Nairobi, and Washington DC) representing diverse climates, with data collected within a 5-degree latitude and longitude radius around each city. Additionally, data at specific pressure levels (875, 900, 925, 950, 975, and 1000 hPa) relevant to aircraft flight trajectories were utilized.

Key Insights and Ideas

  • Temporal Correlation is Key: The study strongly reinforces that wind speed is a temporally dependent phenomenon. Point-wise models, which ignore this, inherently perform worse than models that consider past data.
  • BiLSTM Dominance: The Bidirectional LSTM (BiLSTM) architecture emerged as the superior model, consistently achieving the best performance across various metrics. This highlights the advantage of processing sequential data in both forward and backward directions for capturing complex wind patterns.
  • Beyond Traditional QIs: Expanding the set of Quality Indicators beyond the commonly used atmospheric pressure, relative humidity, and temperature significantly improved prediction accuracy. This suggests that a more holistic view of meteorological factors is crucial for accurate wind speed forecasting.
  • Feature Importance Matters: Not all QIs are equally important. Geopotential, ozone mass mixing ratio, potential vorticity, and temperature were identified as the most influential factors, while relative humidity, surprisingly, showed less importance in this specific dataset. This implies that careful feature selection and analysis can lead to more efficient and accurate models.
  • CNN-LSTM's Feature Extraction Power: While BiLSTM performed best overall, the CNN-LSTM's ability to extract salient features through convolutional layers was noted, particularly in reducing maximum prediction errors in certain scenarios.

Key Results

  • BiLSTM Outperforms: BiLSTM achieved the best results for MAE and RMSE across all tested scenarios, demonstrating its superior predictive capability for horizontal wind speed.
  • Enriched QIs Improve Accuracy: Using all fourteen QIs as input resulted in significantly better MAE and RMSE metrics compared to using only atmospheric pressure, relative humidity, and temperature. In most cases, the maximum prediction errors were also smaller with the full set of QIs.
  • Deep Learning Superiority: Deep learning models (LSTM, GRU, CNN-LSTM, Bi-LSTM) consistently outperformed traditional point-wise machine learning models (PR, RF, GBDT, MLP) and ARIMA in terms of prediction accuracy (MAE and RMSE).
  • GRU's Speed Advantage: GRU offered a faster training time compared to LSTM due to fewer parameters, making it a viable option when rapid predictions are needed, though generally with slightly lower accuracy.
  • MAX Metric Insights: While BiLSTM generally led in MAE and RMSE, CNN-LSTM sometimes showed better performance for the MAX (maximum prediction error) metric, indicating its strength in handling extreme values.

Limitations and Future Work

  • Interdependence of QIs: The paper acknowledges that the QIs are not independent. Future work should focus on analyzing these correlations and potentially reducing the dimensionality of the input features to improve efficiency and reduce storage and training time.
  • Seasonality and Extreme Weather: The authors suggest that future research should consider seasonality and the impact of extremely severe weather events on wind-speed prediction.
  • Real-time Aircraft Operations: The ultimate goal is to integrate these accurate wind-speed predictions into real-time aircraft operations to mitigate deviations, reduce delays, and prevent accidents.

This research provides a strong foundation for more accurate and reliable wind speed forecasting, with significant implications for safety-critical applications and the advancement of renewable energy technologies.

Summary updated

BibTeX
@misc{li2024temporally,
  author = {Lintong Li and Jose Escribano Macias and Mingwei Zhang and Shenghao Fu and Mingyang Huang and Xiangmin Yang and Tianyu Zhao and Yuxiang Feng and Mireille Elhajj and Arnab Majumdar and Panagiotis Angeloudis and Washington Ochieng},
  title = {{Temporally Correlated Deep Learning-Based Horizontal Wind-Speed Prediction}},
  year = {2024},
  doi = {10.3390/s24196254},
}
Figure