Abstract Missing or incomplete data in short or long intervals is a common problem in measuring air pollution. Severe issues may arise when dealing with missing data for time-series prediction schemes or mean analysis. This study aimed to develop a new regression imputation framework to impute missing values in the hourly air quality data set of Tehran and enhance the applicability of Tehran Air Pollution Forecasting System (TAPFS). The proposed framework was designed based on three types of features including measurements of other stations, WRF and CAMx physical models. In this framework, elastic net and neuro-fuzzy networks were efficiently combined in a two-layer structure. The framework was applied on Tehran's air pollution monitoring network. The hourly imputing results of the suggested method were seen to be superior to existing methods according to statistical criteria such as RMSE, MAE and R-values. Average R-values of 0.88, 0.73, 0.76 and 0.79 were obtained for O3, NO, PM2.5 and PM10, respectively. The measurements of other stations had the main predictive power with a modest increase for the two physical models. The benefit of the models was somewhat higher for stations on boundaries of monitoring network. In addition, the central stations had better performance than the boundary stations and an approximately 0.05 increase was obtained in average R-value.

中文翻译：

---

使用 WRF 和 CAMx 模型输出的德黑兰空气污染监测网络的新型回归插补框架

摘要 短期或长期数据缺失或不完整是空气污染测量中的常见问题。在处理时间序列预测方案或均值分析的缺失数据时，可能会出现严重问题。本研究旨在开发一种新的回归插补框架，以插补德黑兰每小时空气质量数据集中的缺失值，并增强德黑兰空气污染预测系统 (TAPFS) 的适用性。提议的框架是基于三类特征设计的，包括其他台站的测量、WRF 和 CAMx 物理模型。在这个框架中，弹性网络和神经模糊网络有效地结合在一个两层结构中。该框架已应用于德黑兰的空气污染监测网络。根据 RMSE、MAE 和 R 值等统计标准，建议方法的每小时插补结果优于现有方法。O3、NO、PM2.5 和 PM10 的平均 R 值分别为 0.88、0.73、0.76 和 0.79。其他台站的测量具有主要的预测能力，两个物理模型略有增加。对于监测网络边界上的站点，这些模型的好处要高一些。此外，中心站的性能优于边界站，平均 R 值增加了约 0.05。其他台站的测量具有主要的预测能力，两个物理模型略有增加。对于监测网络边界上的站点，这些模型的好处要高一些。此外，中心站的性能优于边界站，平均 R 值增加了约 0.05。其他台站的测量具有主要的预测能力，两个物理模型略有增加。对于监测网络边界上的站点，这些模型的好处要高一些。此外，中心站的性能优于边界站，平均 R 值增加了约 0.05。