Short-term wind power prediction based on adaptive Stacking integrated learning model

Abstract Aiming at the problem that when a single combined model uses multiple features for short-term wind power prediction with a small number of training samples, it is difficult to capture the characteristics of time series data, resulting in a large generalization error of the prediction model. This paper proposes a short-term wind power prediction method based on an adaptive Stacking integrated learning model. This ensemble prediction method selects four basic models with different working principles, namely Random Forest (RF), Transformer, Bidirectional Long Short-Term Memory Network (BiLSTM), and Gated Temporal Convolutional Network (GTCN). Since the meta-learner of Stacking ensemble learning cannot fuse each basic model differentially, this limits the advantage of the fusion of model prediction results to a certain extent, leading to the accumulation of errors. In this paper, by introducing an adaptive dynamic attention mechanism, weights are assigned to the preliminary prediction results of each basic model to form weighted input features. Finally, the weighted input features of each basic model are sent to the meta-learner for ensemble training, and the prediction results are mapped and fused to obtain the final wind power prediction result. According to the actual power generation data of a certain offshore wind farm in Fujian Province, taking the data in December as an example, the Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and correlation coefficient ( <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" overflow="scroll"> <mml:msup> <mml:mrow> <mml:mi>R</mml:mi> </mml:mrow> <mml:mn>2</mml:mn> </mml:msup> </mml:math> ) of the proposed model are 97.13, 65.77, and 0.91 respectively. The comparison results with multiple models show that the proposed method has higher prediction accuracy.