A GRU-based deep learning framework for IoT-based measurement and prediction of ambient noise levels in Taiwan

• GRU model outperformed LSTM and CNN in hourly urban noise prediction. • IoT-based noise sensors enabled fine-scale ambient noise monitoring. • Noise data collected across 16 diverse sites in four Taiwanese districts. • GRU achieved R 2 = 0.77 and RMSE = 1.36 dB across multiple urban settings. • Deep learning fills spatial and temporal gaps in noise monitoring networks. Accurately predicting urban noise is crucial for proactive mitigation, but it remains challenging with conventional models. We deployed low-cost IoT-based noise sensors at 16 sites representing four land-use types (provincial and county roadways, urban roads, and environmental areas) in four Taiwanese districts, recording per-second sound pressure levels over seven days and aggregating them to hourly L eq . Three deep-learning architectures, namely, 1D CNN, LSTM, and GRU, were trained to predict the next-hourly L eq . Across all sites, the GRU model yielded the highest accuracy (RMSE = 1.36 dB, MAE = 2.09 dB, R 2 = 0.77), outperforming LSTM (R 2 = 0.64) and 1D CNN (R 2 = 0.50). GRU excelled on structured, high‑traffic roadways while retaining robustness in variable environmental settings. These results demonstrate that GRU-based models can reliably fill spatial and temporal monitoring gaps, providing urban planners with a scalable tool for data-driven noise management and exposure assessment.