Scholarlink

<p>Proper air quality forecasting constitutes the key to successful public health planning and minimization of pollution, especially in densely populated areas supported by high-density Internet of Things (IoT) sensor networks. Nevertheless, time-series of air quality data recorded using IoT networks is inherently non-homogeneous, noisy and non-stationary, which effectively challenges traditional sequence modeling methods. In order to overcome these shortcomings, this paper proposes a new N-Space Transformer structure to multi-horizon Air Quality Index (AQI) prediction that specially focuses on integrating heterogeneous feature representations and modeling the short-term dynamics and long-horizon temporal patterns. The architecture includes a contextual fusion component that combines sensor measurements with temporal and exogenous features. On top of that, a learnable N-Space transformation pulls together all those different inputs and maps them into one shared high-dimensional latent space. This step lines up features, even when they’re on different scales or types, so attention-based representation learning stays steady. Then there’s an Attention Memory module sitting between the encoding and decoding parts. It manages how information flows and cuts down on errors stacking up when forecasting far into the future. The model gets tested on the AirIoT dataset, which covers close to three years of air pollution data from 50 IoT-based monitoring stations around Hyderabad, India. The results show that the proposed approach outperforms the baseline Seq2Seq model by a wide margin. For 31-day forecasting, it reaches an R² of 0.965, way higher than Seq2Seq’s 0.456. It also drops MAE from 42.52 to 6.91 and RMSE from 66.55 to 16.93. So, this architecture performs relatively better at handling long-range temporal patterns, cuts down error spread, and offers more reliable multi-horizon AQI forecasts.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Neural space alignment for enhanced multi-horizon air quality time series forecasting using transformers

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月6日

查看全文
  • Arunkumar Bongale,
  • Satish Kumar,
  • C. Hemanth Kumar,
  • Prajwal Birwadkar,
  • Harish Kumar M

摘要

Proper air quality forecasting constitutes the key to successful public health planning and minimization of pollution, especially in densely populated areas supported by high-density Internet of Things (IoT) sensor networks. Nevertheless, time-series of air quality data recorded using IoT networks is inherently non-homogeneous, noisy and non-stationary, which effectively challenges traditional sequence modeling methods. In order to overcome these shortcomings, this paper proposes a new N-Space Transformer structure to multi-horizon Air Quality Index (AQI) prediction that specially focuses on integrating heterogeneous feature representations and modeling the short-term dynamics and long-horizon temporal patterns. The architecture includes a contextual fusion component that combines sensor measurements with temporal and exogenous features. On top of that, a learnable N-Space transformation pulls together all those different inputs and maps them into one shared high-dimensional latent space. This step lines up features, even when they’re on different scales or types, so attention-based representation learning stays steady. Then there’s an Attention Memory module sitting between the encoding and decoding parts. It manages how information flows and cuts down on errors stacking up when forecasting far into the future. The model gets tested on the AirIoT dataset, which covers close to three years of air pollution data from 50 IoT-based monitoring stations around Hyderabad, India. The results show that the proposed approach outperforms the baseline Seq2Seq model by a wide margin. For 31-day forecasting, it reaches an R² of 0.965, way higher than Seq2Seq’s 0.456. It also drops MAE from 42.52 to 6.91 and RMSE from 66.55 to 16.93. So, this architecture performs relatively better at handling long-range temporal patterns, cuts down error spread, and offers more reliable multi-horizon AQI forecasts.