[Paper Review] A Multi-variable Stacked Long-Short Term Memory Network for Wind Speed Forecasting

Sisheng Liang, Long Nguyen+1 more|arXiv (Cornell University)|Nov 24, 2018

Energy Load and Power Forecasting4 citations

TL;DR

This paper proposes a multi-variable stacked Long Short-Term Memory (MSLSTM) network that leverages historical wind speed, wind direction, temperature, humidity, pressure, dew point, and solar radiation to improve short-term wind speed forecasting. The model uses stacked LSTM layers to capture complex temporal dependencies and multi-scale patterns, achieving state-of-the-art performance with an R² of 0.99917 on 5-minute wind speed predictions in West Texas.

ABSTRACT

Precisely forecasting wind speed is essential for wind power producers and grid operators. However, this task is challenging due to the stochasticity of wind speed. To accurately predict short-term wind speed under uncertainties, this paper proposed a multi-variable stacked LSTMs model (MSLSTM). The proposed method utilizes multiple historical meteorological variables, such as wind speed, temperature, humidity, pressure, dew point and solar radiation to accurately predict wind speeds. The prediction performance is extensively assessed using real data collected in West Texas, USA. The experimental results show that the proposed MSLSTM can preferably capture and learn uncertainties while output competitive performance.

Motivation & Objective

• Address the challenge of accurate short-term wind speed forecasting due to wind's stochastic and intermittent nature.
• Improve prediction accuracy by incorporating multiple meteorological variables beyond just wind speed.
• Develop a deep learning model capable of capturing long- and short-term temporal dependencies in wind speed time series.
• Mitigate overfitting in wind speed prediction through dropout regularization in a stacked LSTM architecture.
• Demonstrate the superiority of the proposed MSLSTM model over traditional statistical and shallow learning methods on real-world data.

Proposed method

• Utilizes a multi-variable input sequence including wind speed, wind direction, temperature, humidity, pressure, dew point, and solar radiation.
• Employs a stacked LSTM architecture with two stacked LSTM layers to learn hierarchical representations of temporal patterns.
• Applies ReLU activation functions to mitigate vanishing gradient issues and accelerate convergence.
• Incorporates a 5% dropout layer after the stacked LSTM to prevent overfitting.
• Uses dense layers for final output prediction after the stacked LSTM and dropout layers.
• Applies min-max normalization to all input features to ensure consistent scale and improve training stability.

Experimental results

Research questions

• RQ1 Can a deep learning model that integrates multiple meteorological variables improve short-term wind speed forecasting accuracy compared to single-variable models?
• RQ2 Does stacking multiple LSTM layers enhance the model's ability to capture long- and short-term dependencies in wind speed sequences?
• RQ3 How does the proposed MSLSTM model compare in performance to traditional methods like linear regression, LASSO, Ridge, and standard LSTM?
• RQ4 To what extent can the MSLSTM model reduce prediction error (MSE, RMSE, MAE) and increase R² on real-world 5-minute wind speed data?
• RQ5 Does the inclusion of atmospheric variables such as pressure and solar radiation improve prediction robustness under varying weather conditions?

Key findings

• The MSLSTM model achieved an R² of 0.99917 on the test set, significantly outperforming all baseline models.
• The MSLSTM model recorded the lowest MSE (0.00297), RMSE (0.05448), and MAE (0.04275) among all models tested.
• Compared to standard LSTM, the MSLSTM reduced RMSE by over 48% and MAE by over 58%, demonstrating the benefit of stacking LSTM layers.
• The model’s performance was robust across different weather conditions, with improved prediction stability due to multi-variable input.
• The training and test loss curves showed close alignment, indicating effective generalization and minimal overfitting.
• The model outperformed Multiple Linear Regression, LASSO, Ridge, and standard LSTM across all evaluation metrics, confirming its superiority in capturing complex wind dynamics.