基于RF-TCN-SA及误差修正的风电功率超短期预测

The Ultra-Short-Term Wind Power Prediction Based on RF-TCN-SA and Error Correction

中文关键词:风电功率预测  随机森林  时间卷积神经网络  自注意力机制  误差修正

keywords:wind power prediction  random forest  time convolutional neural network  self-attention mechanism  error correction

中文摘要:为提高风电功率预测精度，提出一种结合随机森林（random forest，RF）、时间卷积神经网络（temporal convolutional network，TCN）以及自注意力机制（self-attention，SA）的预测模型。通过RF算法选择出与风电功率强相关的特征信息作为TCN的输入，采用Lookahead优化器及PRelu激活函数来提高TCN的学习、收敛性能；通过SA算法为模型不同时刻输入信息分配不同权重，以突出重要时刻信息作用，提高模型预测效果；建立误差修正模型对初步预测值进行修正，进一步提高风电功率预测精度。算例实验结果表明，所提模型相比常见循环神经网络预测模型具有更高的预测精度。

Abstract:To enhance the precision of wind power forecasting， a prediction model that integrates random forests （RF）， time convolutional neural networks （TCN）， and self-attention （SA） is proposed. Initially， the RF algorithm is used to select characteristic information that is strongly correlated with wind power as the input for the TCN. The look ahead optimizer and PReLU activation function are employed to enhance the learning and convergence capabilities of the TCN. Subsequently， SA is utilized to assign varying weights to the model's input information at different times， emphasizing the significance of information during critical periods and thereby improving the model's predictive performance. Lastly， an error correction model is established to refine the preliminary prediction values and further increase the accuracy of wind power forecasting. Simulation results indicate that this model achieves higher prediction accuracy compared to conventional recurrent neural network models.