Quality-Aware AIoT-Enabled Wireless Urban Sensor Networks Using Deep Learning

This paper investigates the relationship between data quality and deep learning model performance for predicting pedestrian flows in smart cities, using data from the UK's Newcastle Urban Observatory. We evaluate eight forecasting models (including multivariate LSTMs and Transformers) across 110 pedestrian sensors. Crucially, by preserving genuine data gaps, we reveal that the structure of missing data-specifically the length of consecutive sequences-is a more significant determinant of model performance than overall data completeness. Segmented regression identifies a sharp breakpoint at three hours of average sequence length, below which forecast error deteriorates rapidly. After controlling for sequence length, data completeness has no independent predictive power. Multivariate LSTMs and Transformers outperform all baselines, with Diebold-Mariano tests confirming statistical significance. A random forest meta-model successfully predicts per-sensor error (R2=0.81) using basic quality metrics, highlighting sequence fragmentation as the dominant predictor. Finally, we show that sensors in transit-dense urban cores exhibit stronger diurnal signals and lower relative error, acting as spatial proxies for data quality. These findings establish key data quality requirements for forecasting urban dynamics and provide a framework for robust, quality-aware predictive systems.