Benchmarking Deep Learning for PM2.5 Forecasting in IoT-Based Smart Cities: GCN Spatial Encoding and Transformer Temporal Modeling

Graph convolutional networks are widely used in air quality forecasting, yet their benefit over simpler approaches remains insufficiently validated. This study presents a fully reproducible benchmark comparing five model families (ARIMA, XG-Boost, LSTM, CNN-LSTM, and a GCN+Transformer model) on the public Beijing Multi-site Air Quality Dataset. All experiments share identical preprocessing, three random seeds (confidence intervals reported as exploratory), and a chronological split, with full code availability. ARIMA outperforms deep learning on aggregated data, due to strong temporal autocorrelation. On individual stations, Transformer-based models achieve the best performance through improved temporal modeling. A systematic multi-node analysis reveals that graph-based spatial aggregation can degrade performance under highly homogeneous conditions: when monitoring stations exhibit strong correlations, a simple linear baseline outperforms the GCN topologies tested here. Alternative spatial encoders (e.g., GAT, learnable adjacency) may behave differently and remain an open question. These findings define a practical regime in which GCN-based spatial aggregation provides no benefit over linear baselines. A preliminary compression experiment on Apple M1 reports a 42% inference-latency reduction via 8-bit quantization; this is indicative only, and validation on Raspberry Pi 4 or NVIDIA Jetson Nano is identified as required follow-up.