Short-term wind power prediction based on hybrid optimization clustering

Single monolithic models for short-term wind power prediction often struggle to achieve high precision when applied to large, complex datasets, primarily due to their inability to account for significant inter-sample heterogeneity. To overcome this limitation, this paper introduces a short-term wind power prediction method based on hybrid optimisation clustering. The methodology begins by employing the Randomised Dependence Coefficient to analyse nonlinear dependencies in high-dimensional data and select the most salient meteorological features. Subsequently, a hybrid clustering method partitions the data into more homogeneous subsets. This involves first reducing data dimensionality with Uniform Manifold Approximation and Projection to preserve both global and local structures, followed by the Ant Lion Optimiser that refines the k-means cluster centres to enhance clustering accuracy. Within each resulting cluster, the training data quality is further improved by selecting the most representative weather data samples using the entropy weight method combined with a random forest model. Finally, an improved Bidirectional Gated Recurrent Unit is trained on each subset for high-precision, short-term wind power prediction. Experimental results demonstrate the superiority of the proposed method over other comparison models, achieving significant improvements in R2, MAE, and RMSE metrics, thereby validating its effectiveness in short-term wind power prediction.