Clean and renewable wind energy has made an outstanding contribution to alleviating the energy crisis. However, the randomness and volatility of wind brings great risk to the integration of wind power to the grid. Therefore, it is essential to obtain reliable and efficient wind speed forecasts. Quantile-based machine learning techniques, which usually produce satisfied quantile-based prediction intervals (PIs) for wind energy, have received widespread attention. However, the obtained PIs are usually crossed and violate the monotonicity of different conditional quantiles. In addition, the completeness and quality of features directly affect the forecasting performance of the models. Therefore, mining effective and sufficient information from the limited input data helps to improve the forecasting performance. In this paper, a novel method is developed for probabilistic wind speed forecasting based on deep learning, non-crossing quantile loss, multi-scale feature (MSF) extraction, and kernel density estimation (KDE). In terms of feature extraction, sufficient MSFs with simple pattern will be extracted based on a multi-layer convolutional neural network. Attention-based long short-term memory is used to further extract and encode temporal information for features of each scale and reduce computational cost. The final feature is obtained by concatenating all the encoded feature vectors. Instead of directly outputting different conditional quantiles, this study obtains the positive difference of adjacent conditional quantiles. On this basis, a non-crossing quantile loss is designed to ensure the monotonicity of different conditional quantiles. To understand the forecasting uncertainty comprehensively, KDE is used to estimate the continuous probability distribution function for various PIs. The proposed method is verified on four wind speed datasets collected form South Dakota. The results demonstrate that the proposed method has an excellent ability of generating high-quality, high-precision, and non-crossing probabilistic wind speed forecasts.
- Attention mechanism
- Deep learning
- Multi-scale features
- Non-crossing quantile loss
- Probabilistic wind speed forecasting