A novel wavenets long short term memory paradigm for wind power prediction

• Published in: Applied energy 2020-07, Vol.269, p.115098, Article 115098
• Main Authors: Shahid, Farah, Zameer, Aneela, Mehmood, Ammara, Raja, Muhammad Asif Zahoor
• Format: Article
• Language: English
• Subjects: Activation functions; Long Short Term Memory (LSTM); Wavenets; Wind power prediction
• ISSN: 0306-2619; 1872-9118
• DOI: 10.1016/j.apenergy.2020.115098

•Wavelet kernels based long short term memory (WN-LSTM) for wind power prediction.•Implementation of WN-LSTM on wind farms in Europe for short term power prediction.•The efficacy, precision and robustness of WN-LSTM evaluated on performance metrics.•Comparison with established schemes supports the use of wavelet kernels in LSTM.•One-way ANOVA on 95% confidence level establishes the performance of WN-LSTM. Wind power prediction is essentially important for smooth integration of wind power into the national grid pertained to its inherent fluctuations. To facilitate the wind energy production and balance production versus market demand, a precise, efficient and robust forecast model is required to encounter highly nonlinear and intricate nature of the problem. In this study, a machine learning paradigm is presented by exploiting the strength of recurrent neural networks based Long Short Term Memory (LSTM), embedded with wavelet kernels, to encompass the dynamic behavior of temporal data. The novel idea of wavenets utilizing LSTM (WN-LSTM) with Gaussian, Morelet, Ricker and Shannon activation kernels for power prediction from various wind farms make it a unique hybrid forecast model to enrich the ultimate utilization of deep learning for vanishing gradient and wavelet transformations for nonlinear mapping. The proposed methodology, WN-LSTM, is implemented on seven wind farms based in Europe for short term wind power prediction and is evaluated in terms of various standard performance metrics, such as mean absolute error (MAE) and mean absolute percentage error (MAPE). The results are compared with the well-established existing techniques and a percentage improvement of upto 30% is observed on MAE. Multiple independent executions of the model are carried out to ensure its efficacy and robustness. Moreover, the interval prediction is complemented to quantitatively characterize the uncertainty as developing intervals, and the difference between means of predicted and actual power is 0.02 at 95% confidence level as illustrated from ANOVA based Tukey’s and Fisher’s tests.