基于DALSTM和联合分位数损失的海上风电功率概率预测

Probabilistic Forecasting of Offshore Wind Power Based on Dual-stage Attentional LSTM and Joint Quantile Loss Function

传统特征关联方法的预设阈值限制及分位数损失中各分位点损失的量级差异,使得海上风电功率概率预测精度受限。为了提高概率预测精度,提出了一种基于多任务联合分位数损失的双重注意力概率预测模型(MT-DALSTM)。首先,引入特征和时序双重注意力机制对特征间的关联关系和时序依赖性进行挖掘,赋予关键特征和时间点信息以注意力权重来提升功率预测的准确性;其次,在模型训练方面,采用一种基于任务不确定性的多任务联合分位数损失,通过动态调整各损失权重占比来提升最终预测结果的综合性能指标;最后,基于东海大桥海上风电场真实数据仿真验证结果表明:相比于现有的风电概率预测研究,所提方法在锐度、可靠性、综合性能指标上均具有明显提升,验证了该模型提高预测精度的有效性。

Probabilistic prediction of offshore wind power is not high in accuracy due to the predetermined threshold limitation of the traditional feature correlation method and the magnitude difference of the quantile loss in each quantile loss.To improve the probabilistic prediction accuracy,a multi-task joint quantile loss-based dual-attention probabilistic prediction model(MT-DALSTM)is proposed.Firstly,a feature and temporal dual attention mechanism is introduced to mine the correlation and temporal dependence among features,and attention weights are given to key features and time point information to improve the accuracy of power prediction.Secondly,during model training,the multi-task joint quantile loss based on task uncertainty is used to improve the final prediction results by dynamically adjusting the proportion of each loss weight.Finally,the simulation validation results based on the real data from the Donghai Bridge offshore wind farm show that the proposed method has significant improvement in sharpness,reliability and comprehensive performance indexes compared to the existing wind power probabilistic prediction studies,which verifies the effectiveness of the model in improving the prediction accuracy.