融合物理信息的TPA-LSTM网络应用于台风条件下海上风电场风速多步预测

Physics-informed TPA-LSTM Network Applied to Multi-step Wind Speed Prediction of Offshore Wind Farms Under Typhoon Conditions

台风条件下海上风电场风速变化大、无明显周期性,这对海上风电场的风速预测造成了极大的困难。针对此问题,提出台风条件下海上风电场风速多步预测方法。首先,针对台风预报信息与风电场风速数据时间尺度不统一的问题,提出用嵌入层网络对台风预报信息进行动态插值。其次,基于Holland气压场模型和Batts梯度风模型构建融合物理信息的神经网络,将Holland模型和Batts模型中的经验参数替换成网络可学习的参数,并针对网络训练过程中可能出现的数值问题引入适当的近似方法。最后,对含时序模式注意力机制的长短期记忆网络(temporal pattern attention long short-term memory,TPA-LSTM)进行改进,嵌入融合物理信息的神经网络,利用近40年台风期间的数据进行训练和测试。结果表明,在引入较少参数的情况下,物理信息神经网络能减少TPA-LSTM网络的训练迭代次数以及提高预测精度,所提模型相比序列到序列(sequence to sequence,Seq2Seq)模型和TPA-LSTM网络具有更高的预测精度。

Under typhoon conditions,the wind speed of offshore wind farms exhibits significant fluctuations without any discernible periodicity,making the accurate wind speed prediction a challenging task.In this paper,a multi-step wind speed prediction method is proposed for offshore wind farms under typhoon conditions.Firstly,an embedded layer network is introduced to address the non-uniformity of the time scales between the typhoon forecast information and the wind speed data of the wind farm in order to dynamically interpolate the typhoon forecast information.Secondly,a physics-informed neural network is constructed based on the Holland pressure field model and the Batts gradient wind model.The empirical parameters in the Holland and Batts models are replaced with the learnable parameters in the network,and the appropriate approximations are introduced to address the potential numerical problems during the network training.Finally,the Temporal Pattern Attention Long Short-Term Memory(TPA-LSTM)network is improved by embedding it with the physics-informed neural network,and the improved model is trained and tested by using the data collected during the typhoons over the past 40 years.Experimental results demonstrate that the proposed physical neural network is able to reduce the number of the training iterations required by the TPA-LSTM network and improve the prediction accuracy,even when introducing fewer parameters.The proposed method achieves higher prediction accuracy in comparison to the Sequence to Sequence(Seq2Seq)model and the TPA-LSTM network.