A Hybrid Framework for Missing Value Imputation in an Air Quality Sensor Network

Abstract

Accurate and continuous air quality monitoring is essential for sustainable urban management, environmental protection, and public health. However, dense networks of low-cost sensors frequently suffer from missing observations caused by communication failures, sensor malfunctions, or maintenance operations. This paper proposes a hybrid graph convolutional network–temporal convolutional network (GCN-TCN) framework for imputing missing PM2.5 measurements in dense sensor networks. The model utilizes spatial relationships between neighboring monitoring stations and temporal dependencies within sensor time series. The proposed approach was evaluated using data from a network of 20 sensors deployed across the academic campus area. Three representative missing-data scenarios were considered, including isolated missing observations, continuous missing sequences, and a hybrid combination of both patterns. The proposed model achieved the lowest reconstruction errors and the highest coefficient of determination among the evaluated methods (R2≈ 0.91–0.93). Standard GCN and TCN networks achieved lower R2 values (∼0.81–0.82 and ∼0.71–0.80, respectively), while recurrent models and classical statistical methods performed substantially worse (R2≈0.6). These findings indicate that integrating graph-based spatial learning with temporal convolution robustly reconstructs incomplete environmental observations, improving the reliability of low-cost sensor systems for sustainable air quality monitoring and management.