A convolutional Transformer-based truncated Gaussian density network with data denoising for wind speed forecasting

• Published in: Applied energy 2023-03, Vol.333, p.120601, Article 120601
• Main Authors: Wang, Yun, Xu, Houhua, Song, Mengmeng, Zhang, Fan, Li, Yifen, Zhou, Shengchao, Zhang, Lingjun
• Format: Article
• Language: English
• Subjects: Multi-scale features; Self-attention; Transformer; Truncated Gaussian-based loss function; Wavelet soft threshold denoising; Wind speed forecasting
• ISSN: 0306-2619
• DOI: 10.1016/j.apenergy.2022.120601

Wind speed forecasting plays an important role in the stable operation of wind energy power systems. However, accurate and reliable wind speed forecasting faces four challenges: how to reduce the data noise; how to find the optimal model inputs; how to describe the complex fluctuations in wind speed; and how to design a suitable loss function to tune the forecasting model. This study proposes a novel forecasting model to address the four challenges mentioned above. First, it uses a wavelet soft threshold denoising method to reduce noise in the original wind speed time series. Second, it uses the maximal information coefficient, which measures the linear and nonlinear relationships between historical wind speed data and forecasted targets, to determine the optimal model inputs. Third, a novel convolutional Transformer-based truncated Gaussian density network is designed to characterize the complex fluctuations in wind speed. The multi-scale information from different convolutional layers is weighted using the self-attention mechanism and then fed into the Transformer network to extract temporal information. The outputs are mapped into the forecasted targets with several fully connected layers. Fourth, considering the non-negativity of wind speed, the truncated Gaussian distribution, which shows a probability of zero when the wind speed is less than zero, is employed to model the uncertainty of wind speed forecasts. This leads to designing a truncated Gaussian distribution-based loss function to train the forecasting model. The forecasting results on three real-world datasets show that the proposed model not only provides accurate deterministic wind speed forecasts but also produces reliable probabilistic wind speed forecasts. The hypothesis testing also illustrates the effectiveness of the proposed model for deterministic and probabilistic wind speed forecasting. •A convolutional Transformer-based model is proposed for wind speed forecasting.•Wavelet soft threshold denoising is used to enhance the predictability of wind data.•Maximal information coefficient is used to determine the optimal model inputs.•A truncated Gaussian distribution is built to model the forecast uncertainty.•The validity of the proposed deep model is tested through hypothesis testing.