Filling gaps in PM2.5 time series: A broad evaluation from statistical to advanced neural network models

Abstract

This study addressed the critical challenge of filling gaps in PM2.5 time series data from Pavlodar, Kazakhstan. We developed and evaluated a comprehensive hierarchy of 46 gap-filling methods across five representative gap lengths (5–72 hours), introducing dynamic models capable of adapting to gaps of variable duration. Tree-based models with bidirectional sequence-to-sequence architectures delivered superior performance, with XGB Seq2Seq achieving a mean absolute error of 5.231 ± 0.292 μg/m3 for 12-hour gaps, representing a 63% improvement over basic statistical methods. The advantage of multivariate models incorporating meteorological variables increased substantially with gap length, from modest improvements of 2–3% for 5-hour gaps to significant enhancements of 16–18% for 48–72 hour gaps. Dynamic multivariate models demonstrated remarkable operational flexibility by successfully processing real-world gaps ranging from 1 to 191 hours despite being trained on maximum lengths of 72 hours. Analysis of the reconstructed complete time series revealed that 61.2% of monitored hours exceeded the WHO daily threshold of 15 μg/m3, with strong seasonal patterns and pronounced diurnal cycles. This research advances environmental monitoring capabilities by providing robust methodological tools for addressing data continuity challenges that currently limit the utility of PM2.5 measurements for public health applications and scientific analysis.