基于注意力机制的多尺度时序预测模型及其在空气质量预报中的应用

孙佳婕

doi:10.70693/202609178002

文献检索 | 登录 | 注册

开放获取　·　同行评审　·　学术期刊

科学与工程研究维普检索国际期刊

《科学与工程研究》（Science and Engineering Research）是由香港 Wisvora Publishing 出版、面向全球发行的国际中文学术期刊。本刊聚焦自然科学与工程技术领域的理论创新、技术发展与实践应用，关注基础研究、应用研究、工程实践及多学科交叉融合中的前沿问题。倡导跨学科、跨领域的多元研究视角，打通科学研究与工程实践的沟通壁垒，推动科研成果的传播、共享与转化，助力科学技术创新与工程领域高质量发展。

在线投稿模板下载最新一期

发行周期：双月刊发文量：1 篇

公告 《科学与工程研究》创刊征稿覆盖理工科各学科方向 新刊《科学与工程研究》为双月刊，论文将被维普数据库收录，面费优惠价 880 元 / 篇，欢迎广大科研工作者踊跃投稿。详情

在线投稿下载模版过刊浏览

切换期刊

期刊栏目

刊期目次

2026年1卷1期

查看全部 >

加入编委

期刊订阅

添加您的邮件地址以接收即将发行期刊数据：

Open Access Article

PDF

基于注意力机制的多尺度时序预测模型及其在空气质量预报中的应用

A Multi-Scale Time-Series Forecasting Model Based on Attention Mechanism and Its Application in Air Quality Prediction

作者：孙佳婕

单位：河池学院

卷期：2026年1卷1期

DOI：10.70693/202609178002

页码：1-7

发布时间：2026-09-17

总浏览量：190

摘要

城市空气质量预报是大气污染精准防控与公众健康风险管理的重要基础。受排放源、气象条件与大气化学过程的多重耦合影响，污染物浓度时间序列普遍具有强非线性、非平稳性与多尺度特征：小时尺度上的短期波动、日尺度上的交通与生活排放节律、周尺度上的生产活动周期以及季节尺度上的气象演变相互叠加，使精确预报成为一项具有挑战性的任务。传统统计模型难以刻画上述非线性关系，单一结构的深度学习模型则难以同时捕捉不同时间尺度的时序特征。针对这一问题，本文提出一种基于注意力机制的多尺度时序预测模型（multi-scale attention LSTM，MSA-LSTM）。该模型首先通过多尺度滑动窗口将输入序列分解为小时、日、周三个时间尺度的子序列，分别由双向长短时记忆网络（BiLSTM）提取时序特征；随后引入多头自注意力机制对各尺度特征进行自适应加权融合，使模型能够依据预测任务动态调整不同尺度的贡献；最后由全连接输出层生成多步预测结果。以某市空气质量监测站点2023年1月至2024年12月逐小时监测数据（示意数据集）为研究对象，与ARIMA、随机森林、LSTM、Transformer、Informer等基线模型进行系统对比。结果表明（示意结果），MSA-LSTM在PM2.5浓度1 h、6 h、12 h、24 h等多步预测任务上均取得最优性能，其中24 h预测平均绝对误差为6.2 μg/m³，均方根误差为9.8 μg/m³，决定系数为0.93；消融实验进一步验证了多尺度分解与注意力融合模块的独立贡献。研究结果表明，多尺度建模与注意力机制的有机结合能够显著提升污染物浓度预测精度，可为多学科交叉背景下的数据驱动环境建模提供方法参考。

关键词

深度学习；时序预测；注意力机制；多尺度特征；空气质量；PM2.5

Abstract

Urban air quality forecasting is a fundamental basis for precise air pollution control and public health risk management. Affected by the multiple coupling of emission sources, meteorological conditions, and atmospheric chemical processes, pollutant concentration time series generally exhibit strong nonlinearity, non-stationarity, and multi-scale characteristics: short-term fluctuations on the hourly scale, traffic and living emission rhythms on the daily scale, production activity cycles on the weekly scale, and seasonal meteorological evolution are superimposed, making accurate forecasting a challenging task. To address this issue, this paper proposes a multi-scale attention LSTM model (MSA-LSTM). The model first divides the input sequence into sub-sequences of hourly, daily, and weekly scales through multi-scale sliding windows, and extracts temporal features at each scale with bidirectional long short-term memory networks (BiLSTM). A multi-head self-attention module is then introduced to adaptively fuse the multi-scale features so that the model can dynamically adjust the contributions of different scales according to the forecasting task. Finally, a fully connected layer generates multi-step forecasting results. Using hourly monitoring data of an urban air quality station from January 2023 to December 2024 (illustrative dataset), systematic comparative experiments are conducted against ARIMA, random forest, LSTM, Transformer, and Informer. The results (illustrative) show that MSA-LSTM achieves the best performance in multi-step forecasting tasks at 1 h, 6 h, 12 h, and 24 h horizons, with a mean absolute error of 6.2 μg/m³, a root mean square error of 9.8 μg/m³, and a coefficient of determination of 0.93 for 24 h forecasting. Ablation studies further verify the independent contributions of the multi-scale decomposition and attention fusion modules. The results indicate that the combination of multi-scale modeling and attention mechanisms can significantly improve pollutant concentration forecasting accuracy, providing methodological references for data-driven environmental modeling in multidisciplinary contexts.

Keywords

deep learning; time-series forecasting; attention mechanism; multi-scale features; air quality; PM2.5

下载PDF

引用本文

孙佳婕. 基于注意力机制的多尺度时序预测模型及其在空气质量预报中的应用[J]. 科学与工程研究. 2026, 1 (1): 1-7. DOI: 10.70693/202609178002.

APA引用

孙佳婕. (2026). 基于注意力机制的多尺度时序预测模型及其在空气质量预报中的应用. 科学与工程研究, 1 (1), 1-7. https://doi.org/10.70693/202609178002

参考文献

[1] World Health Organization. WHO global air quality guidelines: particulate matter (PM2.5 and PM10), ozone, nitrogen dioxide, sulfur dioxide and carbon monoxide[R]. Geneva: World Health Organization, 2021.
[2] Box G E P, Jenkins G M, Reinsel G C, et al. Time series analysis: forecasting and control[M]. 5th ed. Hoboken: John Wiley & Sons, 2015.
[3] Breiman L. Random forests[J]. Machine Learning, 2001, 45(1): 5-32.
[4] Chen T, Guestrin C. XGBoost: a scalable tree boosting system[C]//Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. New York: ACM, 2016: 785-794.
[5] Hochreiter S, Schmidhuber J. Long short-term memory[J]. Neural Computation, 1997, 9(8): 1735-1780.
[6] Cho K, van Merriënboer B, Gulcehre C, et al. Learning phrase representations using RNN encoder-decoder for statistical machine translation[C]//Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing. Stroudsburg: ACL, 2014: 1724-1734.
[7] Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need[C]//Advances in Neural Information Processing Systems 30. Red Hook: Curran Associates, 2017: 5998-6008.
[8] Bahdanau D, Cho K, Bengio Y. Neural machine translation by jointly learning to align and translate[C]//Proceedings of the 3rd International Conference on Learning Representations. San Diego: ICLR, 2015.
[9] Shi X, Chen Z, Wang H, et al. Convolutional LSTM network: a machine learning approach for precipitation nowcasting[C]//Advances in Neural Information Processing Systems 28. Red Hook: Curran Associates, 2015: 802-810.
[10] Zhou H, Zhang S, Peng J, et al. Informer: beyond efficient transformer for long sequence time-series forecasting[C]//Proceedings of the AAAI Conference on Artificial Intelligence. Palo Alto: AAAI Press, 2021, 35(12): 11106-11115.
[11] Wu H, Xu J, Wang J, et al. Autoformer: decomposition transformers with auto-correlation for long-term series forecasting[C]//Advances in Neural Information Processing Systems 34. Red Hook: Curran Associates, 2021: 22419-22430.
[12] Salinas D, Flunkert V, Gasthaus J, et al. DeepAR: probabilistic forecasting with autoregressive recurrent networks[J]. International Journal of Forecasting, 2020, 36(3): 1181-1191.
[13] Oreshkin B N, Carpov D, Chapados N, et al. N-BEATS: neural basis expansion analysis for interpretable time series forecasting[C]//Proceedings of the 8th International Conference on Learning Representations. Addis Ababa: ICLR, 2020.
[14] Lim B, Arık S Ö, Loeff N, et al. Temporal fusion transformers for interpretable multi-horizon time series forecasting[J]. International Journal of Forecasting, 2021, 37(4): 1748-1764.
[15] Qi Y, Li Q, Karimian H, et al. A hybrid model for spatiotemporal forecasting of PM2.5 based on graph convolutional neural network and long short-term memory[J]. Science of the Total Environment, 2019, 664: 1-10.
[16] Freeman B S, Taylor G, Gharabaghi B, et al. Forecasting air quality time series using deep learning[J]. Journal of the Air & Waste Management Association, 2018, 68(8): 866-886.https://gotu.wisvora.com/article/1354