基于时空高斯混合注意力的时空风速预测模型

patiotemporal Wind Speed Prediction Model Based on Spatiotemporal Gaussian Mixture Attention

中文关键词:风速预测  时空依赖性  时空高斯混合注意力  动态时间规整

keywords:wind speed forecasting  spatiotemporal dependency  spatiotemporal gaussian mixture attention  dynamic time warping

中文摘要:风能由于其可持续性和无污染性，受到越来越多的关注。为解决现有预测方法难以捕捉复杂时空依赖性的问题，提出时空高斯混合注意力的时空风速预测模型。通过季节-趋势分解，将风速数据分解为趋势部分和季节部分，分别使用多层感知机预测趋势，并通过注意力机制预测季节性波动。设计时空高斯混合注意力层，有效融合风速的周期性时间信息与长期空间特征。通过动态时间规整算法提取全球空间上下文特征，进一步增强模型的时空依赖性捕捉能力。实验结果显示，所提模型在4个真实数据集上表现优异，在平均绝对误差和均方根误差指标上分别提高8.42%和9.16%，显著优于现有最先进的预测方法。

Abstract:Wind energy is gaining attention due to its sustainability and zero pollution.We propose a spatiotemporal wind speed prediction model based on spatiotemporal Gaussian mixture attention to address the challenge of capturing complex spatiotemporal dependencies in existing forecasting methods.First，seasonal-trend decomposition splits wind speed data into trend and seasonal components.A multilayer perceptron predicts the trend，while an attention mechanism captures seasonal variations.We design a spatiotemporal Gaussian mixture attention layer to integrate periodic temporal patterns and long-term spatial features.Then a dynamic time-warping algorithm to extract global spatial context，enhancing spatiotemporal dependency modeling.Experimental results demonstrate that the proposed model performs exceptionally well on four real-world datasets，with improvements of 8.42% and 9.16% in mean absolute error and root mean squared error metrics，respectively，compared to state-of-the-art methods.