考虑空间相关性的MSCNN LSTM Attention能见度预测模型  

MSCNN-LSTM-Attention model for visibility prediction with spatial correlation consideration

作　　者：王小建[1] 苏彤 马飞[1] 林智婕 白元旦 郭庆元 魏俊涛 黄凯 徐玉凤 WANG Xiaojian;SU Tong;MA Fei;LIN Zhijie;BAI Yuandan;GUO Qingyuan;WEI Juntao;HUANG Kai;XU Yufeng(School of Economics and Management,Chang an University,Xi an 710064,China;Xi'an Meteoroogy,Xi'an 710016,China;Xi'an Traffic Information Center,Xi'an 710018,China)

机构地区：[1]长安大学经济与管理学院,西安710064 [2]西安市气象局,西安710016 [3]西安市交通信息中心,西安710018

出　　处：《安全与环境学报》2025年第4期1622-1632,共11页Journal of Safety and Environment

基　　金：国家自然科学基金项目(72104034,72104037);陕西省社会科学基金项目(2024R009);陕西省教育厅2021年度重点科学研究计划项目(21JP007);陕西省交通运输厅交通科技项目(23-12K)。

摘　　要：准确预测能见度对保障交通运输安全具有重要意义。针对现有方法在能见度预测时对影响因素空间相关性考虑不足导致预测精度较低的问题,研究构建了一种考虑空间相关性的能见度预测模型。利用一维多尺度卷积神经网络(Multi-Scale Convolutional Neural Network, MSCNN)提取能见度以预测各影响因素下不同精细度的空间特征,并将其进行线性融合得到多因素空间特征,实现对能见度预测影响因素的空间特征提取;利用Attention机制加强对关键信息关注的优势以对长短期记忆神经网络(Long-Short Term Memory Neural Network, LSTM)方法进行改进,进而增强模型对重要时序信息关注的能力和模型预测的准确性,实现在考虑影响因素空间相关性下对能见度的预测。以2021—2023年西安市逐时气象数据和污染物数据为试验数据,采用均方根误差(RMSE)、平均绝对误差(MAE)和R2指标对模型进行评价。试验结果显示,研究模型MAE下降26.3%~39.1%,RMSE下降25%~40%,R2提升3.7%~16.4%,能见度预测精度较高。Visibility prediction is crucial for ensuring transportation safety.This paper addresses the significant prediction errors observed in existing research,which often overlook the spatial correlation among various influencing factors.We propose a visibility prediction model that effectively incorporates spatial correlation to improve accuracy.By utilizing the robust spatial feature extraction capabilities of Multi-Scale Convolutional Neural Networks(MSCNN),we extract spatial features at various levels of detail for each visibility prediction influencing factor through multi-scale convolutional kernel operations.The features obtained from each convolutional kernel are then linearly fused to generate the multi-site single-factor spatial features for each influencing factor.The multi-site single-factor spatial features for each influencing factor are linearly fused to create multi-site multi-factor spatial features.Subsequently,the multi-site multi-factor spatial features from the two types of influencing factors are combined through linear fusion,enabling the comprehensive extraction of spatial features relevant to visibility prediction.The Long Short-Term Memory(LSTM)method is enhanced through the incorporation of an Attention mechanism,which assigns varying weights to the hidden layer states at different time steps.This approach allows for weighted processing of the hidden layer states based on their corresponding weights,enabling the model to capture more critical temporal information within the time series.This enhancement addresses the limitations of LSTM in distinguishing the importance of temporal features and retaining key information.Hourly meteorological and pollutant data from Xi an,collected between 2021 and 2023,are utilized as experimental data.The model s performance is evaluated using metrics such as Root Mean Square Error(RMSE),Mean Absolute Error(MAE),and R-squared(R 2).Experimental results indicate that the proposed model achieves a reduction in MAE ranging from 26.3%to 39.1%,a decrease in RMSE bet

关 键 词：环境科学技术基础学科 能见度预测 空间相关性 一维多尺度卷积神经网络 长短期记忆神经网络 注意力机制 

分 类 号：X16[环境科学与工程—环境科学]