Interval Wind-Speed Forecasting Model Based on Quantile Regression Bidirectional Minimal Gated Memory Network and Kernel Density Estimation.

To deal with the randomness, intermittence and fluctuation of wind speed, reasonable dispatch and economic prediction model are necessary. For this purpose, a new composite framework position encoding, feature extraction, quantile regression bidirectional minimal gated memory network (QRBiMGM) and kernel density estimation are proposed in this study. Firstly, numerical weather features are encoded by position encoding and aggregated with historical wind speed data. Then, the multi-head attention mechanism is applied to extract the clustering features. Subsequently, combined with the improved minimum gated memory network and quantile regression, three wind speed datasets are trained to obtain conditional quantiles at different confidence levels. Finally, the probability density function of wind speed is estimated by kernel density estimation. To verify the effectiveness of the proposed composite framework, a comparative test of four related models on three datasets is conducted. It obtains the root-mean-square error values of 2.1876 m/s for Dataset 1, respectively, using mode of conditional distribution. In terms of interval prediction, the Winkler score is 4.9130, 5.3969 and 5.6506 with 95% confidence level, which exhibits higher accuracy than the benchmark. The experimental results reveal that the combination of multi-head attention mechanism and QRBiMGM can better balance the prediction performance and timing computation, and possess strong adaptability that allows them to be widely utilized in the wind speed prediction of various wind power plants. [ABSTRACT FROM AUTHOR]