Scholarlink

<p>Accurate multivariate time series forecasting remains a fundamental challenge across critical domains including traffic management, environmental monitoring, and agricultural planning. While recent advances in deep learning have shown promise, existing approaches face inherent limitations: convolutional models struggle with long-range dependencies and cross-variable interactions, whereas pure attention-based architectures often overlook crucial local temporal patterns and suffer from quadratic computational complexity. To address these complementary weaknesses, we propose a novel hybrid architecture that systematically integrates Temporal Convolutional Networks (TCNs) for efficient local feature extraction with Transformer multi-head attention mechanisms for global dependency modeling. Our TCN-Transformer model employs dilated causal convolutions to capture hierarchical temporal patterns across multiple scales, followed by multi-head attention layers that learn cross-variable dependencies and long-range temporal relationships. We validate our approach across three diverse real-world domains: traffic volume prediction, air quality forecasting, and wheat productivity estimation across five Egyptian governorates. Experimental results demonstrate substantial improvements over established baselines, with the TCN-Transformer achieving coefficient of determination (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(R^2\)</EquationSource> </InlineEquation>) values ranging from 0.91 to 0.96 across all datasets, representing improvements of 5-13% over the next-best baseline. For traffic forecasting, our model reduces Mean Absolute Error (MAE) by 4.5% and Root Mean Squared Error (RMSE) by 16.7% compared to standalone TCN, while maintaining 25-50% lower prediction errors than traditional recurrent architectures. On air quality data, we achieve a 42.1% reduction in RMSE compared to standalone TCN with <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(R^2 = 0.96\)</EquationSource> </InlineEquation>. The model outperforms seven baseline approaches including statistical methods (VAR, SVR), recurrent networks (LSTM, GRU, BiLSTM), and individual deep learning components (TCN, Transformer), demonstrating robust generalization across temporal resolutions (7-day and 30-day windows), data characteristics, and application domains.</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

Enhancing time series forecasting: a hybrid TCN-transformer approach

  • 期刊论文

  • 发表日期：

    2026年4月22日

查看全文
  • Amal Mahmoud,
  • Ammar Mohammed

摘要

Accurate multivariate time series forecasting remains a fundamental challenge across critical domains including traffic management, environmental monitoring, and agricultural planning. While recent advances in deep learning have shown promise, existing approaches face inherent limitations: convolutional models struggle with long-range dependencies and cross-variable interactions, whereas pure attention-based architectures often overlook crucial local temporal patterns and suffer from quadratic computational complexity. To address these complementary weaknesses, we propose a novel hybrid architecture that systematically integrates Temporal Convolutional Networks (TCNs) for efficient local feature extraction with Transformer multi-head attention mechanisms for global dependency modeling. Our TCN-Transformer model employs dilated causal convolutions to capture hierarchical temporal patterns across multiple scales, followed by multi-head attention layers that learn cross-variable dependencies and long-range temporal relationships. We validate our approach across three diverse real-world domains: traffic volume prediction, air quality forecasting, and wheat productivity estimation across five Egyptian governorates. Experimental results demonstrate substantial improvements over established baselines, with the TCN-Transformer achieving coefficient of determination ( \(R^2\) ) values ranging from 0.91 to 0.96 across all datasets, representing improvements of 5-13% over the next-best baseline. For traffic forecasting, our model reduces Mean Absolute Error (MAE) by 4.5% and Root Mean Squared Error (RMSE) by 16.7% compared to standalone TCN, while maintaining 25-50% lower prediction errors than traditional recurrent architectures. On air quality data, we achieve a 42.1% reduction in RMSE compared to standalone TCN with \(R^2 = 0.96\) . The model outperforms seven baseline approaches including statistical methods (VAR, SVR), recurrent networks (LSTM, GRU, BiLSTM), and individual deep learning components (TCN, Transformer), demonstrating robust generalization across temporal resolutions (7-day and 30-day windows), data characteristics, and application domains.