基于改进深度学习模型的干旱区城市PM2.5浓度预测与应用研究

doi:10.12118/j.issn.1000-6060.2025.100

摘要/Abstract

摘要：

随着全球城市化进程加速，干旱区城市PM_2.5污染因其独特的地理气候条件，呈现强非平稳性与复杂时空特征，传统预测模型难以有效捕捉其动态规律。针对这一挑战，构建了“自适应噪声完备经验模态分解-常春藤优化算法-柯尔莫哥洛夫-阿诺德网络-双向长短期记忆神经网络”（CEEMDAN-IVY-KAN-BiLSTM）混合预测框架，以提升PM_2.5浓度的预测精度。该框架通过降噪分解与参数优化联合提取多尺度特征，融合KAN-BiLSTM模型的强非线性拟合与双向时序建模能力，有效提升预测性能。结果表明：（1） 2021—2024年乌鲁木齐市PM_2.5浓度呈显著季节性波动，冬季因燃煤供暖和逆温层影响均值达41.97 μg·m^-3，夏季因大气对流增强，浓度降至全年低位附近，且整体呈逐年下降趋势。（2）对数据进行重要性排序可知PM_2.5与空气质量指数、PM₁₀、CO、NO₂呈强正相关，与气温、露点温度呈负相关，表明燃煤排放、机动车尾气及气象扩散条件是主要影响因素，且模型有效分离了数据中PM_2.5序列的高频波动（如沙尘事件）与低频趋势（季节性变化），降低数据非平稳性影响。（3）实验基于2021—2024年乌鲁木齐市逐日空气质量数据开展，结果显示本框架在决定系数、平均绝对误差与均方根误差指标上分别达到0.991、1.391和1.881，均显著优于传统机器学习和常见深度学习模型。验证了“分解-优化-集成”的深度学习框架在干旱区城市PM_2.5预测中的适用性。

关键词: PM_2.5浓度预测, CEEMDAN分解, IVY优化算法, KAN-BiLSTM模型, 深度学习, 干旱区城市PM_2.5

Abstract:

With the acceleration of global urbanization, the severe PM_2.5pollution in arid zone cities, owing to their unique geographical and climatic conditions, exhibits strong non-stationarity and complex spatiotemporal characteristics, making it difficult for traditional prediction models to effectively capture its dynamic patterns. To address this challenge, this study proposes a hybrid prediction framework of an “adaptive noise complete ensemble empirical mode decomposition-Ivy optimization algorithm-Kolmogorov Arnold network-bidirectional long short-term memory neural network” (CEEMDAN-IVY-KAN-BiLSTM), aiming to enhance the prediction accuracy of PM_2.5 concentrations. This framework jointly extracts multi-scale features through noise reduction decomposition as well as parameter optimization and integrates the strong nonlinear fitting and bidirectional time series modeling capabilities of the KAN-BiLSTM model, effectively improving the prediction performance. The results reveal that the PM_2.5 concentration in Urumqi City from 2021 to 2024 shows significant seasonal fluctuations, with an average of 41.97 μg·m^-3 in winter due to coal heating and the influence of the inversion layer and drops to 14.04 μg·m^-3 in summer due to enhanced atmospheric convection. Moreover, it shows an overall decreasing trend annually. Moreover, the importance ranking of the data indicates that PM_2.5 is significantly positively correlated with air quality index, PM₁₀, CO, and NO₂, and negatively correlated with temperature and dew point temperature, suggesting that coal emissions, vehicle exhaust, and meteorological diffusion conditions are the main influencing factors. Moreover, the model effectively separates the high-frequency fluctuations (such as sandstorm events) and low-frequency trends (seasonal changes) in the PM_2.5 sequence, reducing the impact of data non-stationarity. Finally, the experiments were based on daily air quality data in Urumqi City from 2021 to 2024, results of which demonstrate that this model achieves the coefficient of determination, mean absolute error, and root mean square error values of 0.991, 1.391, and 1.881, respectively, significantly outperforming conventional machine learning and common deep learning models. This verifies the applicability of the “decomposition-optimization-integration” deep learning framework in the prediction of arid zone cities.

Key words: PM_2.5 concentration prediction, CEEMDAN decomposition, IVY optimisation algorithm, KAN-BiLSTM model, deep learning, urban PM_2.5 concentration in arid zones

王圣节, 张庆红, 桑铭键. 基于改进深度学习模型的干旱区城市PM_2.5浓度预测与应用研究[J]. 干旱区地理, 2026, 49(2): 343-355.

WANG Shengjie, ZHANG Qinghong, SANG Mingjian. Prediction and application of urban PM_2.5 concentration in arid zones based on improved deep learning models[J]. Arid Land Geography, 2026, 49(2): 343-355.

图/表 11

图1

图2

图3

表1

图4

图5

表2

图6

图7

图8

表3

参考文献 30

[1]	时雨, 武迪, 依里帕·依力哈木, 等. 乌鲁木齐市呼吸系统疾病空气质量健康指数的构建[J]. 环境与职业医学, 2024, 41(3): 276-281.
	[Shi Yu, Wu Di, Yilihamu Yilipa, et al. Construction of air quality health index for respiratory diseases in Urumqi[J]. Journal of Environmental and Occupational Medicine, 2024, 41(3): 276-281.]
[2]	于志翔, 于晓晶, 李霞, 等. 基于中国高分辨率数据集(CHAP)的近20 a新疆地区PM_2.5时空变化特征[J]. 干旱区资源与环境, 2025, 39(4): 132-144.
	[Yu Zhixiang, Yu Xiaojing, Li Xia, et al. Spatiotemporal variations of PM_2.5 in Xinjiang during 2000—2022 revealed by the China High Air Pollutants (CHAP) reanalysis dataset[J]. Journal of Arid Land Resources and Environment, 2025, 39(4): 132-144.]
[3]	闫劲烨, 马正权, 孙萱萱, 等. 2015—2023年“乌-昌-石”城市群PM_2.5与PM₁₀时空变化及潜在源分析[J]. 干旱区地理, 2025, 48(3): 405-420. doi: 10.12118/j.issn.1000-6060.2024.249
	[Yan Jinye, Ma Zhengquan, Sun Xuanxuan, et al. Spatiotemporal variations and potential sources of PM_2.5and PM₁₀inthe “Urumqi-Changji-Shihezi” urban agglomeration from 2015 to 2023[J]. Arid Land Geography, 2025, 48(3): 405-420.] doi: 10.12118/j.issn.1000-6060.2024.249
[4]	马正权, 闫劲烨, 帕丽达·牙合甫, 等. 乌鲁木齐市和喀什市近地面O₃变化特征及潜在源研究[J]. 干旱区研究, 2024, 41(12): 2004-2014. doi: 10.13866/j.azr.2024.12.03
	[Ma Zhengquan, Yan Jinye, Yahufu Palida, et al. Characteristics and potential sources of surface O₃ variation in Urumqiand Kashgar[J]. Arid Zone Research, 2024, 41(12): 2004-2014.] doi: 10.13866/j.azr.2024.12.03
[5]	梅承志, 李斌, 何友江, 等. 2015—2023年天山北坡区域PM_2.5与O₃-8 h变化趋势及影响因素分析[J]. 环境科学学报, 2025, 45(1): 388-399.
	[Mei Chengzhi, Li Bin, He Youjiang, et al. Variation trends and influencing factors of PM_2.5 and O₃-8 h on the north slope of Tianshan Mountain during 2015—2023[J]. Acta Scientiae Circumstantiae, 2025, 45(1): 388-399.]
[6]	王嘉敏, 杨文柱, 焦燕, 等. 内蒙古沿黄流域城市群PM_2.5区域输送及源解析[J]. 中国环境科学, 2025, 45(10): 5338-5356.
	[Wang Jiamin, Yang Wenzhu, Jiao Yan, et al. Regional transport and source analysis of PM_2.5 in urban agglomerations along the Yellow River Basin in Inner Mongolia[J]. China Environmental Science, 2025, 45(10): 5338-5356.]
[7]	潘梦瑶, 任瑛, 王思源, 等. 基于梯度提升算法和SHAP的石家庄PM_2.5和臭氧浓度预测及影响因素分析[J]. 环境科学学报, 2024, 44(7): 402-409.
	[Pan Mengyao, Ren Ying, Wang Siyuan, et al. Prediction of PM_2.5 and ozone concentration in Shijiazhuang and analysis of influencing factors based on gradient boosting algorithm and SHAP[J]. Acta Scientiae Circumstantiae, 2024, 44(7): 402-409.]
[8]	杨国亮, 余华声, 黄聪. 基于优化组合模型的PM_2.5浓度预测[J]. 计算机工程与设计, 2023, 44(10): 3132-3137.
	[Yang Guoliang, Yu Huasheng, Huang Cong. PM_2.5 concentration prediction based on optimization combination model[J]. Computer Engineering and Design, 2023, 44(10): 3132-3137.]
[9]	康俊锋, 黄烈星, 张春艳, 等. 多机器学习模型下逐小时PM_2.5预测及对比分析[J]. 中国环境科学, 2020, 40(5): 1895-1905.
	[Kang Junfeng, Huang Liexing, Zhang Chunyan, et al. Hourly PM_2.5 prediction and its comparative analysis under multi-machine learning model[J]. China Environmental Science, 2020, 40(5): 1895-1905.]
[10]	Bi J Z, Knowland K E, Keller C A, et al. Combining machine learning and numerical simulation for high-resolution PM_2.5 concentration forecast[J]. Environmental Science & Technology, 2022, 56(3): 1544-1556. doi: 10.1021/acs.est.1c05578
[11]	江雨燕, 傅杰, 甘如美江, 等. 改进灰狼算法优化GBDT在PM_2.5预测中的应用[J]. 安全与环境学报, 2024, 24(4): 1569-1580.
	[Jiang Yuyan, Fu Jie, Gan Rumeijiang, et al. Application of improved Gray Wolf Algorithm to optimize GBDT in PM_2.5 prediction[J]. Journal of Safety and Environment, 2024, 24(4): 1569-1580.]
[12]	Zhou S, Wang W, Zhu L, et al. Deep-learning architecture for PM_2.5concentration prediction: A review[J]. Environmental Science and Ecotechnology, 2024: 100400, doi: 10.1016/j.ese.2024.100400.
[13]	朱菊香, 谷卫, 罗丹悦, 等. 基于注意力机制的CNN-ILSTM地铁站PM_2.5预测建模[J]. 中国测试, 2024, 50(7): 53-62.
	[Zhu Juxiang, Gu Wei, Luo Danyue, et al. Prediction modeling of PM_2.5 in subway station based on attention mechanism and CNN-ILSTM[J]. China Measurement & Test, 2024, 50(7): 53-62.]
[14]	彭豪杰, 周杨, 胡校飞, 等. 基于深度学习与随机森林的PM_2.5浓度预测模型[J]. 遥感学报, 2023, 27(2): 430-440.
	[Peng Haojie, Zhou Yang, Hu Xiaofei, et al. A PM_2.5 prediction model based on deep learning and random forest[J]. Journal of Remote Sensing, 2023, 27(2): 430-440.]
[15]	范昕炜, 宣泉森. 基于神经网络的PM_2.5质量浓度预测模型[J]. 安全与环境学报, 2022, 22(6): 3499-3507.
	[Fan Xinwei, Xuan Quansen. Prediction model of PM_2.5 mass concentration based on neural network[J]. Journal of Safety and Environment, 2022, 22(6): 3499-3507.]
[16]	刘恩海, 付英健, 张智, 等. 联合Transformer注意力机制的PM_2.5浓度预测网络研究[J]. 安全与环境学报, 2023, 23(10): 3760-3768.
	[Liu Enhai, Fu Yingjian, Zhang Zhi, et al. PM_2.5 recurrent prediction network combined with transformer attention mechanism[J]. Journal of Safety and Environment, 2023, 23(10): 3760-3768.]
[17]	顾阔, 焦瑞莉, 薄宇, 等. 基于复合LSTM模型的PM_2.5浓度预测[J]. 中国环境监测, 2023, 39(1): 170-180.
	[Gu Kuo, Jiao Ruili, Bo Yu, et al. PM_2.5 concentration prediction based on thecomposite LSTM model[J]. Environmental Monitoring in China, 2023, 39(1): 170-180.]
[18]	Shi P, Fang X, Ni J, et al. An Improved attention-based integrated deep neural network for PM_2.5 concentration prediction[J]. Applied Sciences, 2021, 11(9): 4001, doi: 10.3390/app11094001.
[19]	Lin M D, Liu P Y, Huang C W, et al. The application of strategy based on LSTM for the short-term prediction of PM_2.5 in city[J]. Science of the Total Environment, 2024, 906: 167892, doi: 10.1016/j.scitotenv.2023.167892.
[20]	杨长春, 聂倩倩. 面向PM_2.5预测的时间序列分解与机器学习融合模型[J]. 安全与环境学报, 2023, 23(12): 4600-4608.
	[Yang Changchun, Nie Qianqian. Fusion model of time series decomposition andmachine learning for PM_2.5 forecasting[J]. Journal of Safety and Environment, 2023, 23(12): 4600-4608.]
[21]	魏疆, 赵彩欣, 王国华, 等. 乌鲁木齐市城区大气PM_2.5中水溶性离子组分特征及来源解析[J]. 干旱区地理, 2025, 48(4): 623-631. doi: 10.12118/j.issn.1000-6060.2024.358
	[Wei Jiang, Zhao Caixin, Wang Guohua, et al. Characteristics and sources of water-soluble ion components in PM_2.5 in the urban area of Urumqi City[J]. Arid Land Geography, 2025, 48(4): 623-631.] doi: 10.12118/j.issn.1000-6060.2024.358
[22]	魏疆, 宋丹丹, 赵丽莉, 等. 乌鲁木齐市PM_2.5中碳组分污染特征及来源解析[J]. 干旱区研究, 2024, 41(8): 1323-1330. doi: 10.13866/j.azr.2024.08.06
	[Wei Jiang, Song Dandan, Zhao Lili, et al. Analysis of carbon component pollution characteristics and sources of PM_2.5 in Urumqi City[J]. Arid Zone Research, 2024, 41(8): 1323-1330.] doi: 10.13866/j.azr.2024.08.06
[23]	Torres M E, Colominas M A, Schlotthauer G, et al. A complete ensemble empirical mode decomposition with adaptive noise[C]// IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), May 22-27, 2011, Prague, Czech Republic. New York: IEEE Signal Processing Society, 2011: 4144-4147.
[24]	Ghasemi M, Zare M, Trojovský P, et al. Optimization based on the smart behavior of plants with its engineering applications: Ivy algorithm[J]. Knowledge-Based Systems, 2024, 295: 111850, doi: 10.1016/j.knosys.2024.111850.
[25]	Liu Z, Wang Y, Vaidya S, et al. Kan: Kolmogorovarnold networks[J]. arXiv Preprint arXiv: 2404.19756, 2024, doi: 10.48550/arXiv.2404.19756.
[26]	谢琪, 夏飞, 袁博. 基于CEEMDAN-SE-BiLSTM模型的西安市PM_2.5浓度预测[J]. 环境工程, 2024, 42(8): 105-115.
	[Xie Qi, Xia Fei, Yuan Bo. Prediction of PM_2.5 concentration in Xi’an based on CEEMDAN-SE-BiLSTM model[J]. Environmental Engineering, 2024, 42(8): 105-115.]
[27]	张纯曦, 阿丽亚·拜都热拉, 刘丽, 等. 乌鲁木齐市PM_2.5分布特征及其预测模型研究[J]. 石河子大学学报(自然科学版), 2020, 38(5): 648-654.
	[Zhang Chunxi, Baidurela Aliya, Liu Li, et al. Temporal and spatial distribution characteristicsand prediction model of PM_2.5 in Urumqi[J]. Journal of Shihezi University (Natural Science), 2020, 38(5): 648-654.]
[28]	张士杰, 窦燕. 基于XGBoost与SHAP模型的特征分析在PM_2.5浓度预测中的应用[J]. 赤峰学院学报(自然科学版), 2022, 38(12): 10-17.
	[Zhang Shijie, Dou Yan. Application of feature analysis based on XGBoost and SHAP models in PM_2.5 concentration prediction[J]. Journal of Chifeng University (Natural Science Edition), 2022, 38(12): 10-17.]
[29]	丁成亮, 郑洪波. 基于改进机器学习的PM_2.5浓度预测模型研究[J]. 大连理工大学学报, 2024, 64(4): 353-360.
	[Ding Chengliang, Zheng Hongbo. Study of PM_2.5 concentration prediction modelbased on improved machine learning[J]. Journal of Dalian University of Technology, 2024, 64(4): 353-360.]
[30]	李明明, 王小兰, 岳江, 等. 融合EOF分解及CNN-LSTM神经网络的PM_2.5预测[J]. 环境科学, 2025, 46(2): 715-726.
	[Li Mingming, Wang Xiaolan, Yue Jiang, et al. PM_2.5 prediction based on EOF decomposition and CNN-LSTM neural network[J]. Environmental Science, 2025, 46(2): 715-726.]

统计指标	PM_2.5/μg·m^-3	AQI	PM₁₀/μg·m^-3	O₃/μg·m^-3	NO₂/μg·m^-3	CO₂/μg·m^-3	CO/μg·m^-3	TEMP/℃	SLP/hPa	WDSP/m·s^-1	PRCP/mm	DEWP/℃
总样本	1277	1277	1277	1277	1277	1277	1277	1253	1253	1253	1253	1253
平均值	41.97	75.73	78.58	62.55	34.67	6.59	0.75	47.58	1020.81	4.78	0.28	27.00
标准差	44.56	51.14	66.02	25.13	17.83	1.68	0.46	25.30	12.85	28.16	4.89	16.22
最小值	4.90	18.52	10.28	15.83	7.01	3.25	0.26	-11.10	995.20	1.50	0.00	-16.80
下四分位数	14.04	42.61	39.05	39.41	21.64	5.43	0.42	25.30	1010.20	3.20	0.00	14.00
中位数	20.91	55.26	58.27	64.99	30.16	6.33	0.54	50.60	1020.10	3.90	0.00	27.60
上四分位数	54.14	91.28	95.78	82.77	43.99	7.41	0.94	70.60	1030.30	4.60	0.00	40.30
最大值	316.68	343.68	1190.53	136.39	91.89	14.80	2.58	90.20	1059.60	999.90	99.99	59.80

影响因子	PM_2.5	AQI	PM₁₀	O₃	NO₂	SO₂	CO	TEMP	SLP	WDSP	PRCP	DEWP
PM_2.5	1.000
AQI	0.950^**	1.000
PM₁₀	0.752^**	0.886^**	1.000
O₃	-0.529^**	-0.420^**	-0.299^**	1.000
NO₂	0.824^**	0.754^**	0.556^**	-0.678^**	1.000
SO₂	0.041	0.088^**	0.110^**	0.039	0.171^**	1.000
CO	0.933^**	0.847^**	0.607^**	0.657^**	0.907^**	0.119^**	1.000
TEMP	-0.647^**	-0.540^**	-0.376^**	0.822^**	-0.663^**	0.139^**	-0.714^**	1.000
SLP	0.524^**	0.424^**	0.289^**	-0.787^**	0.569^**	-0.141^**	0.595^**	-0.930^**	1.000
WDSP	-0.006	-0.010	-0.021	-0.001	-0.008	-0.028	0.007	-0.011	0.011	1.000
PRCP	0.037	0.031	0.012	-0.063^*	0.043	0.001	0.047	-0.069^*	0.066^*	-0.003	1.000
DEWP	-0.602^**	-0.524^**	-0.390^**	0.715^**	-0.609^**	0.135^**	-0.639^**	0.872^**	-0.835^**	-0.010	-0.063^*	1.000

模型	R²	MAE	RMSE
CEEMDAN-IVY-KAN-BiLSTM（本文模型）	0.991	1.391	1.881
CEEMDAN-IVY-KAN-LSTM*	0.983	3.064	4.846
CatBoost	0.813	10.244	19.710
Gradient Boosting Model	0.806	9.860	19.902
Random Forest	0.803	9.436	19.306
IVY-KAN-BiLSTM*	0.801	10.697	16.961
XGBoost	0.787	11.803	19.159
Linear Regression	0.784	10.227	19.278
KNN	0.772	10.857	19.350
IVY-KAN-LSTM *	0.745	9.770	16.797
CNN-LSTM	0.706	11.051	21.711
AdaBoost	0.689	19.307	20.876
Decision Tree	0.647	13.315	28.083
Grid-KAN-BiLSTM*	0.626	16.333	22.935
SVR	0.571	14.764	26.950
LSTM	0.559	9.702	16.067