【摘要】 为提高非平稳风速短期预测精度，采用滑动窗口构建滚动分解（RD）与预测方法，通过对比多种神经网络模型的预测结果并结合变分模态分解（VMD）方法，提出采用RD-CNN-BiLSTM神经网络组合模型对非平稳风速序列进行预测。进一步地，针对神经网络组合模型的预测精度不足，提出一种融合误差修正和RD-CNN-BiLSTM神经网络组合模型的非平稳风速短期预测方法。结果表明，对实测风速数据设置滑动窗口并进行一次预测延伸，对延伸后的数据进行分解与修剪重构可有效弱化边界效应的影响并可避免信息泄漏问题。CNN-BiLSTM组合模型相比传统单一的神经网络模型，如自回归积分滑动平均（ARIMA）、反向传播（BP）、长短期记忆网络（LSTM）和双向长短期记忆网络（BiLSTM）模型具有更高的预测精度。采用二次抛物线的误差修正方法相比线性误差修正方法更优，尤其在高风速区间，前者对预测精度的提升效果尤为显著。通过其他时段的实测风速数据预测结果，验证了所提出的融合误差修正和RD-CNN-BiLSTM神经网络组合模型在非平稳风速短期预测方面具有较高的预测精度与泛化性。更多还原

【Abstract】 To improve the accuracy of short-term prediction for non-stationary wind speeds, a rolling decomposition(RD) and prediction method was built by using a sliding window. By comparing the predictive results of various neural network models and incorporating the variational mode decomposition(VMD) algorithm, an RD-CNN-BiLSTM neural network combined model was proposed for predicting non-stationary wind speed sequences. Furthermore, addressing the predictive accuracy limitations of the neural network combined model, a short-term predictive approach for non-stationary wind speeds that integrates the error correction and RD-CNNBiLSTM neural network combined model was proposed. The research results indicate that applying sliding windows to the actual wind speed data and carrying out a single prediction and extension, followed by decomposition, trimming, and reconstruction of the extended data, the impact of boundary effects can be effectively weakened and the information leakage can be prevented. The CNN-BiLSTM combined model outperforms traditional individual neural network models, such as autoregressive integrated moving average(ARIMA), back propagation(BP), long short-term memory(LSTM), and bidirectional long short-term memory(BiLSTM) models, in terms of the predictive accuracy. It was found that the quadratic parabola error correction method is superior to the linear error correction method, particularly in high wind speed ranges, where the predictive accuracy improvement by the former method is especially significant. The predictive results from actual wind speed data at different time intervals further validate that the proposed approach of integrating error correction and RD-CNN-BiLSTM neural network combined model has high predictive accuracy and generalization in short-term prediction of non-stationary wind speeds.更多还原