基于VMD-Transformer-LSTM-XGBoost的短期风电机组出力混合预测模型

A Short-Term Wind Power Forecasting Method Based on VMD-Transformer-LSTM-XGBoost

摘要: 风电功率时间序列具有明显的非平稳性和多尺度波动特征，使高精度短期预测面临较大挑战。针对传统模型难以同时刻画趋势、周期及高频扰动等不同时间尺度结构的问题，本文提出一种融合自适应变分模态分解(VMD)、模糊熵复杂度分析、Transformer-LSTM深度特征提取与XGBoost回归的两阶段短期风电功率预测方法。首先，以理论功率序列为分解对象，通过贝叶斯优化在训练集上自适应确定VMD的模态数与惩罚参数，并采用严格的零数据泄露策略。随后，利用模糊熵度量各IMF的复杂度特征，将其重构为低频趋势、中频周期与高频扰动三类协同模态(Co-IMFs)，以增强输入特征的物理可解释性与稳定性。在特征提取阶段，构建融合Transformer全局依赖建模能力与LSTM局部时序记忆能力的DeepBlock网络，并通过贝叶斯优化确定其最优结构与训练参数，最终由XGBoost完成非线性回归预测。基于甘肃瓜州某风电场2023~2025年15分钟分辨率数据的实验结果表明，所提出方法在MAE、RMSE与R²等指标上均优于多种基准模型及消融模型，验证了该两阶段多尺度混合框架在复杂风电功率预测任务中的有效性。

Abstract: Wind power time series exhibit pronounced non-stationarity and multiscale fluctuations, posing significant challenges for high-accuracy short-term forecasting. To address the difficulty of conventional models in simultaneously capturing trend, periodic, and high-frequency components, this study proposes a two-stage short-term wind power forecasting framework integrating adaptive Variational Mode Decomposition (VMD), fuzzy-entropy-based complexity analysis, Transformer-LSTM deep feature extraction, and XGBoost regression. Theoretical power is first selected as the decomposition target, and the VMD mode number and penalty parameter are adaptively determined on the training set via Bayesian optimization under a strict zero-data-leakage strategy. The intrinsic mode functions (IMFs) are then reconstructed into three collaborative components—low-frequency trend, mid-frequency periodicity, and high-frequency disturbance—based on fuzzy entropy, enhancing the interpretability and stability of the input features. A DeepBlock network combining Transformer-based global dependency modeling and LSTM-based local temporal learning is employed for feature extraction, followed by XGBoost to perform nonlinear regression. Experiments conducted on 15-minute resolution data from a wind farm in Guazhou, Gansu Province (2023~2025) demonstrate that the proposed method consistently outperforms multiple benchmark and ablation models in terms of MAE, RMSE, and R², validating the effectiveness of the proposed multiscale two-stage forecasting framework.