Scholarlink

<p>This study proposes a novel integrated deep learning framework—VMD-TCN-FECAM-Informer—for Air Quality Index (AQI) prediction, effectively addressing the challenges of non-linearity, multi-scale features, and non-stationarity in air quality time series. The model uses PM₂.₅ and PM₁₀ concentrations as input features. The framework combines four functionally complementary modules: Variational Mode Decomposition (VMD), which decomposes the AQI sequence in the frequency domain into intrinsic mode functions with clear physical significance; Temporal Convolutional Network (TCN), which captures local temporal dependencies using dilated causal convolutions; Frequency-Enhanced Channel Attention Mechanism (FECAM), which enhances feature representation by incorporating frequency-domain priors; and the Informer model, which enables efficient long-sequence modeling through the improved ProbSparse attention mechanism. Comprehensive experiments demonstrate that the proposed model significantly outperforms mainstream deep learning models such as Transformer, TimeXer, Crossformer, and FEDformer across multiple metrics, achieving an RMSE of 4.926, MAE of 3.828, MAPE of 7.312%, and RMLSE of 0.084. Ablation studies further validate the contribution and synergy of each component. In addition, the model exhibits strong generalization and robustness in cross-regional validation across ten representative cities in China, performing well at both daily and hourly prediction scales. These results indicate that the proposed framework effectively captures complex dependencies and multi-scale temporal features in AQI data, providing a reliable methodology for environmental monitoring, pollution warning, and urban management decision-making.</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

An Integrated TCN-FECAM-Informer Architecture Combining Feature Importance Analysis and Signal Decomposition for Urban Air Quality Forecasting

  • 期刊论文

  • 发表日期：

    2026年4月3日

查看全文
  • Guangyao Ma,
  • Yue Zhang,
  • Kai Xu,
  • Dongchao Chen,
  • Lanhe Zhang,
  • Zicheng Chen

摘要

This study proposes a novel integrated deep learning framework—VMD-TCN-FECAM-Informer—for Air Quality Index (AQI) prediction, effectively addressing the challenges of non-linearity, multi-scale features, and non-stationarity in air quality time series. The model uses PM₂.₅ and PM₁₀ concentrations as input features. The framework combines four functionally complementary modules: Variational Mode Decomposition (VMD), which decomposes the AQI sequence in the frequency domain into intrinsic mode functions with clear physical significance; Temporal Convolutional Network (TCN), which captures local temporal dependencies using dilated causal convolutions; Frequency-Enhanced Channel Attention Mechanism (FECAM), which enhances feature representation by incorporating frequency-domain priors; and the Informer model, which enables efficient long-sequence modeling through the improved ProbSparse attention mechanism. Comprehensive experiments demonstrate that the proposed model significantly outperforms mainstream deep learning models such as Transformer, TimeXer, Crossformer, and FEDformer across multiple metrics, achieving an RMSE of 4.926, MAE of 3.828, MAPE of 7.312%, and RMLSE of 0.084. Ablation studies further validate the contribution and synergy of each component. In addition, the model exhibits strong generalization and robustness in cross-regional validation across ten representative cities in China, performing well at both daily and hourly prediction scales. These results indicate that the proposed framework effectively captures complex dependencies and multi-scale temporal features in AQI data, providing a reliable methodology for environmental monitoring, pollution warning, and urban management decision-making.