Location-Aware Transfer Learning for Air Quality Time-Series Prediction

Accurate short-term PM10 forecasting is difficult because pollutant dynamics vary across monitoring locations, while sufficient target-station history is not always available due to new deployments, sensor outages, or quality-control filtering. Spatial cross-station transfer learning addresses this problem by pre-training a temporal model on data-rich source stations and fine-tuning it on a data-scarce target station. However, ordinary transfer learning may suffer from source–target domain mismatch and often does not explicitly condition the transferred model on station-specific spatial context, whereas graph-based spatio-temporal models typically require a predefined station graph, synchronized network-level inputs, or assumptions about spatial connectivity. This study therefore examines whether location-aware conditioning improves LSTM-based cross-station transfer learning for one-step-ahead PM10 forecasting under different target-data budgets. The proposed HybridLocLSTM extends a two-layer LSTM backbone with station-identity and geographic-coordinate embeddings, which are fused with the temporal representation. We evaluate seven approaches across 21 Slovenian PM10 monitoring stations and six target-data budgets. The results show that location-aware conditioning improves transfer learning relative to plain LSTM transfer across all evaluated scarcity levels achieving the lowest mean MAE and the best average rank. These findings indicate that explicit station-level spatial conditioning provides the most consistent performance across data regimes, particularly when target-station data are limited.