摘要
基于深度学习的温度预测是利用历史气象数据训练深度学习模型、学习气象数据的模式和关联性,以预测未来温度。该技术在天气预报、农业生产、能源管理和环境管理等领域具有重要的应用价值。现有方法在采用对比学习进行温度时间序列预测时,难以保持多个气象站点间数据的相关性,同时在多尺度特征提取时难以权衡多个尺度特征对于预测结果影响的差异。为了解决上述问题,本文提出基于对比学习的多站点温度预测模型,主要包括多站点联合模块(multi-site association module,MSAM)和多尺度提取模块(multi-scale extraction module,MSEM)。MSAM利用通道注意力机制和空洞卷积,捕获不同气象站点之间的相关性;MSEM通过分支权重调控机制和因果卷积,权衡各尺度特征对预测结果影响的差异。所提模型在采用对比学习的温度预测时表现出色。与CoST模型相比,在河北省气象数据集和湖南省气象数据集上,预测未来24 h内的均方误差分别降低了4.31%和4.85%。
The deep learning-based temperature prediction utilizes historical weather data to train a deep learning mod⁃el,which learns the patterns and correlations in meteorological data to forecast future temperatures.This technology holds significant applications in various fields such as weather forecasting,agricultural production,energy management,and environmental management.Existing methods face challenges when using contrastive learning for temperature time series prediction,as they struggle to maintain the data correlations between multiple meteorological stations and find it difficult to balance the impact of different-scale features on prediction results.To address these issues,this paper propos⁃es a contrastive learning-based multi-site temperature prediction model,consisting of a multi-site association module(MSAM)and a multi-scale extraction module(MSEM).MSAM utilizes channel attention mechanisms and dilated convolu⁃tions to capture the inter-site correlations among different meteorological stations,while MSEM balances the influence of various scale features on prediction results through branch weight control and causal convolutions.The proposed model excels in temperature prediction using contrastive learning and reduces the mean squared error(MSE)for forecasting the next 24 hours by 4.31%and 4.85%in the meteorological datasets of Hebei Province and Hunan Province,respectively,when compared to the CoST model.
作者
曹慧博
张军
CAO Huibo;ZHANG Jun(School of Artificial Intelligence,Hebei University of Technology,Tianjin 300401,China)
出处
《河北工业大学学报》
2026年第2期29-37,70,共10页
Journal of Hebei University of Technology
基金
河北省自然科学基金(F2023202001)。
关键词
对比学习
多站点温度预测
多尺度
通道注意力
分支权重调控机制
contrastive learning
multi-site temperature prediction
multiscale
channel attention
branch weight modu⁃lation mechanis
