Graph-based spatial–temporal prediction and feature interaction analysis of CO2 and occupant in large indoor space

Abstract

With increasing focus on human health, Indoor Air Quality (IAQ) research has become crucial as people spend over half of their lives indoors. Carbon Dioxide (CO2) is a key IAQ parameter associated with Sick Building Syndrome (SBS). Despite the existing connection between indoor occupants and CO2 levels, research on predicting these parameters has largely been conducted in isolation. Most existing studies have not fully analyzed the complex interplay between environmental factors and occupant presence, often relying on historical data without considering intricate feature interactions. This paper introduces a novel learning-based framework to construct spatial–temporal graphs for large indoor environments, interlinking multi-sourced time series into a single graph. To address sensor resolution disparities, we incorporate time-lag nodes, resolving the Different sEnsor at Different Time (DEDT) problem and preserving in high-resolution data integrity. Our proposed model, the Attention-based Spatial–Temporal Graph Convolutional Recurrent Unit (AST-GCGRU), incorporates an encoder–decoder structure, can dynamically capture and validate complex interactions between CO2 and occupants. The model enhances adaptability and robustness in predictive tasks, improves prediction accuracy by 28.5% over baselines. Specifically, for CO2 and occupant nodes, accuracy improvements reach 14.9% and 27.1%, respectively. Additionally, the proposed occupant-in-loop analytical framework enables a comprehensive understanding of the interplay between occupants and environmental variables. This work advances indoor environment research by providing a robust and adaptable solution for predicting and analyzing indoor environmental dynamics.