灰色预测与SVR组合模型实现空气质量预测

Realizing Air Quality Prediction with the Combined Model of Grey Prediction and SVR

摘要: 气象因素不仅直接影响空气质量，还增加了空气质量预测(AQP)的难度。然而，现有方案在学习同时涉及时间序列空气污染物和气象因素的模式方面效果不佳。本研究首先探讨了灰色预测与SVR组合模型在AQP中的优越性。为了分析气象因素在AQP中的作用，我们选取PM2.5、NO₂、PM10等7项核心指标作为特征变量，通过Pearson相关性分析与Lasso回归完成特征筛选，保留所有有效特征以保障预测全面性。模型构建中，利用灰色预测GM (1, 1)模型处理数据不确定性与小样本特性，挖掘整体变化趋势；结合SVR算法精准捕捉数据中复杂非线性关联，实现优势互补。同时采用训练集R²得分、历史数据R²得分、平均绝对误差、中值绝对误差、可解释方差五项评价指标对模型进行全面性评估。实验结果显示，模型训练集与历史数据R²得分均达0.9745，平均绝对误差为3.1439，中值绝对误差为2.0044，可解释方差值达0.9747，且预测值与真实值对比曲线高度吻合。结果表明，该组合模型拟合精度高、抗干扰能力强、泛化性能稳定，能有效捕捉空气质量波动规律，该模型在预测空气质量方向起着重要作用。

Abstract: Meteorological factors not only directly affect air quality but also increase the difficulty of air quality prediction (AQP). However, existing schemes perform poorly in learning patterns that simultaneously involve time-series air pollutants and meteorological factors. This study first explores the superiority of the combined model of grey prediction and SVR in AQP. To analyze the role of meteorological factors in AQP, we select seven core indicators such as PM2.5, NO₂, and PM10 as feature variables. Feature selection is completed through Pearson correlation analysis and Lasso regression, retaining all effective features to ensure comprehensive prediction. In model construction, the grey prediction GM (1, 1) model is utilized to handle data uncertainty and small sample characteristics, mining overall trends; combined with the SVR algorithm, it accurately captures complex nonlinear associations in the data, achieving complementary advantages. Meanwhile, five evaluation metrics including R² score of the training set, R² score of historical data, mean absolute error, median absolute error, and explained variance are used to comprehensively evaluate the model. The experimental results show that both the R² score of the model training set and the historical data reach 0.9745, with an average absolute error of 3.1439, a median absolute error of 2.0044, and an explained variance of 0.9747. Furthermore, the curve comparing the predicted values and the actual values is highly consistent. The results indicate that this combined model has high fitting accuracy, strong anti-interference capability, and stable generalization performance, effectively capturing the fluctuation patterns of air quality. This model plays a significant role in predicting air quality.