Scholarlink

With the rapid advancement of urbanization, the security landscape of cities has become increasingly complex. Conventional crime time-series forecasting models face inherent limitations in capturing long-term dependencies, detecting abrupt events, and modeling interactions among multiple heterogeneous variables, making them inadequate for dynamic, real-time security applications. To address these challenges, this study proposes a novel forecasting framework, TCN-MUDA-Inverted Transformer. The Temporal Convolutional Network (TCN) module captures multi-scale temporal dependencies through dilated causal convolutions and residual connections; the Multi-head Uncertainty-aware Differential Attention (MUDA) module employs a differential attention mechanism to enhance sensitivity to critical information and incorporates uncertainty awareness to improve robustness; and the inverted transformer architecture enhances multivariate interaction modeling via variable embedding and feature projection. Extensive experiments on the Chicago Crime dataset demonstrate that the proposed model significantly outperforms state-of-the-art baselines, including Long Short-Term Memory networks (LSTM), Informer, and NeuralProphet, in both prediction accuracy and generalization capability. Evaluation metrics such as Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and the coefficient of determination (R \(^2\) ) validate the robustness and reliability of the framework, indicating its strong potential for supporting real-time crime monitoring and early warning systems.

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

A Multivariate Time Series Forecasting Framework Based on Multi-scale Convolution and an Inverted Transformer with Differencing Mechanism

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Mengbo Fan,
  • Na Liang,
  • Qingyan Ding,
  • Qian Wang

摘要

With the rapid advancement of urbanization, the security landscape of cities has become increasingly complex. Conventional crime time-series forecasting models face inherent limitations in capturing long-term dependencies, detecting abrupt events, and modeling interactions among multiple heterogeneous variables, making them inadequate for dynamic, real-time security applications. To address these challenges, this study proposes a novel forecasting framework, TCN-MUDA-Inverted Transformer. The Temporal Convolutional Network (TCN) module captures multi-scale temporal dependencies through dilated causal convolutions and residual connections; the Multi-head Uncertainty-aware Differential Attention (MUDA) module employs a differential attention mechanism to enhance sensitivity to critical information and incorporates uncertainty awareness to improve robustness; and the inverted transformer architecture enhances multivariate interaction modeling via variable embedding and feature projection. Extensive experiments on the Chicago Crime dataset demonstrate that the proposed model significantly outperforms state-of-the-art baselines, including Long Short-Term Memory networks (LSTM), Informer, and NeuralProphet, in both prediction accuracy and generalization capability. Evaluation metrics such as Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and the coefficient of determination (R \(^2\) ) validate the robustness and reliability of the framework, indicating its strong potential for supporting real-time crime monitoring and early warning systems.