What does this research mean for the field?

Recurrent deep-learning models, particularly Gated Recurrent Units (GRU), outperform classical statistical models like ARIMAX and persistence baselines in providing accurate one-hour-ahead Air Quality Index (AQI) forecasts. Novelty: ClaimNovelty.INCREMENTAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

The aim is to assess the reliability of deep learning and statistical models for one-hour-ahead AQI forecasting.

May 16, 2026Open Access

Applying Time-Series Statistical Learning to Civil IoT Air-Quality Data: A Case Study in Fengyuan, Taiwan

Key Points

The aim is to assess the reliability of deep learning and statistical models for one-hour-ahead AQI forecasting.
Utilized data from Taichung Fengyuan Station in Taiwan's Civil IoT for forecasting.
Developed three predictive models: LSTM, GRU, and ARIMAX, plus a persistence baseline.
Evaluated models based on prediction errors and their ability to capture short-term temporal structure.
GRU exhibited the lowest overall prediction errors compared to LSTM and ARIMAX.
Persistence baseline performance exceeded ARIMAX but was inferior to both deep learning models.
The study highlights the advantages of employing recurrent deep-learning models for precise AQI forecasting.

Abstract

This study uses data from Taichung Fengyuan Station in Taiwan’s Civil IoT to conduct short-term forecasting of the Air Quality Index (AQI). We compile multiple pollutant and meteorological features and develop three models—Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Autoregressive Integrated Moving Average with exogenous variables (ARIMAX)—together with a persistence baseline for comparison. The purpose of this study is to clarify whether deep sequence models and a classical statistical model can provide reliable one-hour-ahead AQI forecasts at the site level and to examine the practical value of such forecasts for early warning and air-quality management. Results show that GRU achieves the lowest overall prediction errors, followed by LSTM. The persistence baseline outperforms ARIMAX but remains clearly inferior to both recurrent models. In sum, the study shows that site-level AQI forecasting can benefit from recurrent deep-learning models not only in terms of numerical accuracy, but also in terms of capturing short-term temporal structure beyond a naive carry-forward baseline. These findings provide a benchmark-oriented and application-oriented reference for short-horizon AQI warning scenarios.

Applying Time-Series Statistical Learning to Civil IoT Air-Quality Data: A Case Study in Fengyuan, Taiwan

Key Points

Abstract

Cite This Study