Short-Term Wind Speed Prediction Based on Variational Mode Decomposition and Linear–Nonlinear Combination Optimization Model

Wei Sun, Qi GaoEnergies 2019 · DOI

TL;DR

This paper introduces a novel hybrid model, VMD-P-(ARIMA, BP)-PSOLSSVM, for short-term wind speed prediction. By combining variational mode decomposition (VMD) with phase space reconstruction (P) for data preprocessing, and then using a combination of linear (ARIMA) and nonlinear (BP) models for initial predictions, followed by a particle swarm optimized least squares support vector machine (PSOLSSVM) for secondary prediction, the model significantly improves accuracy and robustness compared to existing methods.

What Problem Does This Paper Address?

Wind power, while a crucial renewable energy source, is inherently intermittent and volatile. This unpredictability poses significant challenges for its integration into the power grid, impacting grid stability and market participation. Accurate short-term wind speed prediction is essential to mitigate these issues, enabling better grid dispatching, unit maintenance scheduling, and ultimately, improving the competitiveness of wind energy. Existing prediction models often struggle to capture the complex, nonlinear dynamics of wind speed, leading to suboptimal accuracy.

Methods

The proposed model, VMD-P-(ARIMA, BP)-PSOLSSVM, is a multi-stage approach:

  • Data Preprocessing:
    • Variational Mode Decomposition (VMD): This technique decomposes the raw wind speed signal into several stable intrinsic mode functions (IMFs) and a residual term. VMD is chosen for its ability to adaptively determine relevant frequency bands and avoid modal aliasing, offering a more rigorous mathematical foundation than methods like Empirical Mode Decomposition (EMD).
  • Phase Space Reconstruction (PSR): PSR is applied to each IMF to reconstruct a multi-dimensional time series. This step, based on chaos theory and Takens' theorem, aims to capture the underlying dynamic characteristics of the signal and define appropriate input-output matrices for the prediction models. The C-C method is used to determine the optimal delay time ($\tau$) and embedding dimension ($m$).
  • Individual Prediction Models:
    • ARIMA (Autoregressive Integrated Moving Average): A linear statistical model used to predict each IMF. ARIMA is effective at capturing linear patterns within the time series.
  • BP (Back Propagation Neural Network): A nonlinear artificial neural network model used to predict each IMF. BP excels at modeling complex nonlinear relationships.
  • Combined Prediction:
    • Linear-Nonlinear Combination: The predictions from ARIMA and BP for each IMF are combined (summed) to obtain the initial prediction for each IMF.
  • Secondary Prediction (PSOLSSVM): The prediction results from the ARIMA and BP models (for all IMFs) are then fed as input to a Particle Swarm Optimization (PSO) tuned Least Squares Support Vector Machine (LSSVM). PSOLSSVM acts as a meta-learner, optimizing the parameters of LSSVM (kernel function and regularization parameter) using PSO to achieve a more accurate and robust final prediction. This secondary prediction step is crucial for compensating for the limitations of individual models.

Datasets

The study utilizes two datasets collected from a wind farm in northern China. Both datasets represent periods with significant wind speed fluctuations.

  • Primary Dataset: 1000 data points collected at a 5-minute interval from January 1, 2017. The first 800 points were used for training, and the remaining 200 for prediction. The last 100 data points of the prediction set were used for error analysis.
  • Supplementary Dataset: 1000 data points collected at a 5-minute interval starting from March 30, 2017, also from the same district, to demonstrate the model's universality and applicability on a different, highly fluctuating wind speed period.

Key Insights and Ideas

  • Decomposition is Key: Decomposing the volatile wind speed signal into more stable IMFs using VMD significantly improves the performance of subsequent prediction models. This suggests that addressing the inherent instability of wind speed data is a critical first step.
  • Synergy of Linear and Nonlinear Models: Combining the strengths of both linear (ARIMA) and nonlinear (BP) models for initial IMF predictions, and then using these combined results as input for a powerful meta-learner (PSOLSSVM), leads to superior performance compared to using any single model or simple linear combinations.
  • PSO-Optimized LSSVM as a Powerful Meta-Learner: The PSOLSSVM effectively leverages the outputs of individual models, overcoming their limitations and capturing complex interactions to produce a highly accurate final prediction. The optimization of LSSVM parameters by PSO is crucial for its effectiveness.
  • VMD's Advantage: VMD demonstrates superior performance in data decomposition compared to EMD, leading to better overall prediction results.

Key Results

The proposed VMD-P-(ARIMA, BP)-PSOLSSVM model consistently outperformed six other benchmark models across various error metrics (MAE, MAPE, RMSE) on both the primary and supplementary datasets.

  • Primary Dataset: The proposed model achieved the lowest MAE (0.0524 m/s), MAPE (2.0055%), and RMSE (0.0715 m/s). This represents a significant improvement over separate models (e.g., VMD-ARIMA, VMD-P-BP) and other combined models. For instance, the proposed model's MAE was 53.8% lower than VMD-ARIMA and 74.3% lower than VMD-BP.
  • Supplementary Dataset: Despite higher data volatility, the proposed model still achieved the best results with MAE of 0.0591 m/s, MAPE of 1.8937%, and RMSE of 0.0729 m/s, demonstrating its robustness and applicability.
  • Improvement Indicators: The study introduced PMAE, PMAPE, and PRMSE to quantify improvements. The proposed model showed substantial improvements over other methods, particularly in reducing errors.

Limitations and Future Work

  • Data Volatility Impact: While the model performs well, the paper notes that highly volatile data can lead to slightly lower prediction accuracy and reliability compared to less volatile periods.
  • Computational Cost: The multi-stage approach, involving decomposition, multiple individual predictions, and then a final optimization, might be computationally intensive.
  • Future Work: The paper suggests that further research could explore other decomposition techniques, different combinations of linear and nonlinear models, and investigate the model's performance on datasets from diverse geographical locations and varying wind conditions. The authors also imply that exploring more advanced optimization algorithms for the secondary prediction stage could yield further improvements.

Summary updated

Abstract

Wind power, one of renewable energy resources, is a fluctuating source of energy that prevents its further participation in the power market. To improve the stability of the wind power injected into the power grid, a short-term wind speed predicting model is proposed in this work, named VMD-P-(ARIMA, BP)-PSOLSSVM. In this model, variational mode decomposition (VMD) is combined with phase space reconstruction (P) as data processing method to determine intrinsic mode function (IMF) and its input–output matrix in the prediction model. Then, the linear model autoregressive integrated moving average model (ARIMA) and typical nonlinear model back propagation neural network (BP) are adopted to forecast each IMF separately and get the prediction of short-term wind speed by adding up the IMFs. In the final stage, particle swarm optimization least squares support vector machine (PSOLSSVM) uses the prediction results of the two separate models from previous step for the secondary prediction. For the proposed method, the PSOLSSVM employs different mathematical principles from ARIMA and BP separately, which overcome the shortcoming of using just single models. The proposed combined optimization model has been applied to two datasets with large fluctuations from a northern China wind farm to evaluate the performance. A performance comparison is conducted by comparing the error from the proposed method to six other models using single prediction techniques. The comparison result indicates the proposed combined optimization model can deliver more accurate and robust prediction than the other models; meanwhile, it means the power grid dispatching work can benefit from implementing the proposed predicting model in the system.
BibTeX
@inproceedings{sun2019short,
  author = {Wei Sun and Qi Gao},
  title = {{Short-Term Wind Speed Prediction Based on Variational Mode Decomposition and Linear–Nonlinear Combination Optimization Model}},
  booktitle = {Energies},
  year = {2019},
  doi = {10.3390/en12122322},
}
Figure