Scholarlink

<p>In traffic flow prediction tasks, traffic flow data are characterized by strong non-stationarity, multi-scale periodicity, and complex nonlinearity. To address these challenges, we design a novel characteristic-driven adaptive time series prediction model, called CATS. The model consists of three stages: decomposition, prediction, and fusion. In the decomposition stage, we introduce an adaptive mechanism for determining the optimal number of Variational Mode Decomposition (VMD) components, called AVMD. This mechanism is developed using a collaborative dual-criterion optimization approach. By combining a center-frequency spacing criterion with a reconstruction-error criterion, AVMD autonomously identifies the optimal number of components. In the prediction stage, we propose a statistically characteristic-driven adaptive prediction framework (SCD-APF). This framework establishes a multi-dimensional statistical characterization system that includes stationarity testing, linearity assessment based on linear regression <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(R^2\)</EquationSource> </InlineEquation>, autocorrelation structure analysis, and spectral feature extraction. Based on these characteristics, a characteristic-to-model mapping rule is devised, along with a heterogeneous model library comprising LSTM, linear regression, gradient boosting machines, and multilayer perceptrons. This enables a differentiated “model-per-component” prediction strategy tailored to each decomposed component. In the fusion stage, we introduce a dual attention-based fusion mechanism, DABiLSTM. This mechanism incorporates component-level attention to dynamically estimate the contribution weights of predicted intrinsic mode functions (IMFs), while temporal attention captures critical time-step information. These innovations enhance both predictive accuracy and model interpretability. Experimental results on the PeMS traffic flow dataset demonstrate that the proposed method outperforms existing approaches, confirming its effectiveness and relevance.</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

CATS: a characteristic-driven adaptive prediction model for traffic flow prediction

  • 期刊论文

  • 发表日期：

    2026年6月30日

查看全文
  • Bo Liang,
  • Jiaru Deng,
  • Qi Wu,
  • Fujiang Yuan

摘要

In traffic flow prediction tasks, traffic flow data are characterized by strong non-stationarity, multi-scale periodicity, and complex nonlinearity. To address these challenges, we design a novel characteristic-driven adaptive time series prediction model, called CATS. The model consists of three stages: decomposition, prediction, and fusion. In the decomposition stage, we introduce an adaptive mechanism for determining the optimal number of Variational Mode Decomposition (VMD) components, called AVMD. This mechanism is developed using a collaborative dual-criterion optimization approach. By combining a center-frequency spacing criterion with a reconstruction-error criterion, AVMD autonomously identifies the optimal number of components. In the prediction stage, we propose a statistically characteristic-driven adaptive prediction framework (SCD-APF). This framework establishes a multi-dimensional statistical characterization system that includes stationarity testing, linearity assessment based on linear regression \(R^2\) , autocorrelation structure analysis, and spectral feature extraction. Based on these characteristics, a characteristic-to-model mapping rule is devised, along with a heterogeneous model library comprising LSTM, linear regression, gradient boosting machines, and multilayer perceptrons. This enables a differentiated “model-per-component” prediction strategy tailored to each decomposed component. In the fusion stage, we introduce a dual attention-based fusion mechanism, DABiLSTM. This mechanism incorporates component-level attention to dynamically estimate the contribution weights of predicted intrinsic mode functions (IMFs), while temporal attention captures critical time-step information. These innovations enhance both predictive accuracy and model interpretability. Experimental results on the PeMS traffic flow dataset demonstrate that the proposed method outperforms existing approaches, confirming its effectiveness and relevance.