A Decomposition-Driven Hybrid Framework Based on STL for Accurate Traffic Flow Forecasting

Fujiang Yuan, Yangrui Fan, Xiaohuan Bing, Zhen Tian, Chunhong Yuan, Yankang Li · 2025

Read A Decomposition-Driven Hybrid Framework Based on STL for Accurate Traffic Flow Forecasting by Fujiang Yuan, Yangrui Fan, Xiaohuan Bing, Zhen Tian, Chunhong Yuan, Yankang Li (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-23668.txt.

Introduction

Fujiang Yuan and five co-authors forecast traffic flow by splitting the problem into three simpler ones. Traffic series mix slow trends, daily and weekly rhythms, and noisy leftovers, and a single model often misses at least one of those layers. Their framework first applies STL decomposition, which separates the series into trend, seasonal, and residual parts.

Each part then goes to a specialist: an LSTM network for long-term trends, an ARIMA model for the seasonal rhythm, and XGBoost for the nonlinear residual wiggles. Multiplying the three forecasts rebuilds the final prediction. On 998 traffic records from a New York City intersection in late 2015, the hybrid beat each standalone model across standard error measures. The authors credit the split itself: isolating each pattern lets every model do the job it suits, which lifts accuracy and makes the system easier to interpret and trust.

Catalog record

CREATORFujiang Yuan, Yangrui Fan, Xiaohuan Bing, Zhen Tian, Chunhong Yuan, Yankang Li
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE43 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-23668

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

A Decomposition-Driven Hybrid Framework Based on STL for Accurate Traffic Flow Forecasting Fujiang Yuan, Yangrui Fan, Xiaohuan Bing, Zhen Tian, Chunhong Yuan, Yankang Li Abstract Accurate traffic flow forecasting is crucial for intelligent transportation systems and urban traffic management. However, existing single-model approaches often struggle to capture the complex, nonlinear, and multi-scale temporal patterns inherent in traffic flow data. This study proposes a novel decomposition-driven hybrid framework that integrates Seasonal-Trend decomposition using Loess (STL) with three complementary predictive models to enhance forecasting accuracy and robustness. The STL method first decomposes the original traffic flow time series into three distinct components: trend, seasonal, and residual. Subsequently, a Long Short-Term Memory (LSTM) network is employed to model the trend component and capture long-range temporal dependencies, an Autoregressive Integrated Moving Average (ARIMA) model is applied to the seasonal component to exploit periodic patterns, and an Extreme Gradient Boosting (XGBoost) algorithm is utilized to predict the residual component and handle nonlinear irregularities. The final forecast is obtained through multiplicative integration of the three sub-model predictions. The proposed framework is validated using 998 traffic flow records collected from a New York City intersection between November and December 2015. Experimental results demonstrate that the LSTM-ARIMA-XGBoost hybrid model significantly outperforms individual baseline models including standalone LSTM, ARIMA, XGBoost, and their variants (xLSTM, sLSTM, mLSTM) across multiple evaluation metrics, including Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and coefficient of determination (R²). The decomposition strategy effectively isolates different temporal characteristics, enabling each specialized model to focus on its strength domain, thereby improving overall prediction accuracy, interpretability, and stability. This hybrid approach provides a reliable and efficient solution for real-time traffic flow forecasting in urban transportation systems. keywords Traffic flow forecasting, STL decomposition, Hybrid model, LSTM, ARIMA, XGBoost, Intelligent transportation systems † † equal-contributors: These authors contributed equally to this work. 1 Introduction In the evolution of Intelligent Transportation Systems (ITS), traffic flow prediction has played a pivotal role Li and Xu 2021 . Accurate and real-time traffic forecasting is not only a fundamental component of ITS but also a key enabler for efficient urban operation and intelligent mobility development Lv et al. 2014 ; Zheng et al. 2025 . With the rapid increase in private vehicle ownership, particularly in fast-growing economies, urban road networks have become increasingly congested, and major intersections and arterial roads often experience persistent traffic jams Hao et al. 2011 . By accurately predicting traffic flow over short time intervals at critical intersections, transportation authorities can make informed decisions on traffic control and road planning, reduce accidents and delays, and provide travelers with reasonable route recommendations, thereby alleviating traffic pressure and maximizing the utilization of road resources. Figure 1 shows the traffic flow distribution scene at a typical four-way intersection on a city road. Figure 1: Traffic flow distribution scenario at a typical four-way intersection on a city road In traditional traffic flow prediction studies, various modeling approaches have been proposed, ranging from classical time series models (such as ARIMA) to machine learning and deep learning frameworks (such as RNN and LSTM) Lin et al. 2024 . Although these single-model approaches can achieve satisfactory planning performance under controlled conditions Yuan et al. 2025a , their generalization and robustness are often limited by the highly dynamic and nonlinear nature of urban traffic systems Yuan et al. 2025b . Moreover, most existing models primarily emphasize prediction accuracy while overlooking critical aspects such as computational efficiency, adaptability, and scalability, which are essential for real-time applications in large-scale traffic networks Zheng et al. 2025 . To address the aforementioned limitations, hybrid and decomposition-based modeling approaches have attracted growing research interest. By decomposing complex traffic flow data into more structured components and assigning suitable predictive models to each part, such frameworks can more effectively capture temporal dependencies and improve both interpretability and robustness. Building on this idea, this study introduces a hybrid forecasting framework grounded in Seasonal-Trend Decomposition using Loess. In this framework, the original traffic flow series is separated into three components: (1) a trend component predicted with a Long Short-Term Memory (LSTM) network to model long-range temporal dynamics; (2) a seasonal component captured by the Autoregressive Integrated Moving Average (ARIMA) model to reflect periodic variations; and (3) a residual component estimated using the Extreme Gradient Boosting (XGBoost) algorithm to learn nonlinear patterns and irregular fluctuations. The final forecast is obtained by integrating the outputs of these three sub-models. This decomposition-based hybrid approach effectively combines the complementary advantages of neural networks, statistical models, and ensemble learning methods. It enhances prediction accuracy, robustness, and adaptability, offering a reliable solution to the complex and dynamic challenges of traffic flow forecasting in urban transportation systems. 2 Related Work The literature on traffic flow forecasting reveals a broad spectrum of methodologies designed to capture the intricate spatiotemporal dependencies inherent in traffic data. Early efforts, such as those by Li et al. Guo et al. 2025 , introduced the Spatial–Temporal Fusion Graph Neural Network (STFGNN) to address the challenges posed by complex spatial correlations and dynamic temporal variations across road networks. This line of research underscores the necessity of jointly modeling spatial and temporal information through graph neural networks to enhance forecasting accuracy. Building upon this spatiotemporal modeling paradigm, subsequent studies have incorporated more advanced deep learning architectures. Fang et al. Fang et al. 2022 proposed a model that integrates attention mechanisms with Long Short-Term Memory (LSTM) networks, demonstrating the effectiveness of attention in identifying salient temporal features. Similarly, Wang et al. Wang et al. 2022 developed an Attention-Based Spatiotemporal Graph Network, further reinforcing the role of attention mechanisms in capturing intricate dependencies within traffic data. Graph Neural Networks (GNNs) have emerged as a dominant framework in this domain. Zhang et al. Zhang et al. 2022 presented a graph-based temporal attention model tailored for multi-sensor traffic flow forecasting, highlighting the advantages of leveraging sensor data within graph structures. Extending this idea, Chen et al. Chen et al. 2021 proposed TrafficStream, a continual-learning-based GNN framework capable of adapting to evolving traffic networks, particularly in the context of expanding sensor infrastructures. Transformer-based models have also gained considerable attention due to their strong capability in modeling long-range temporal dependencies. Reza et al. Reza et al. 2022 introduced a multi-head attention transformer model and provided a comparative analysis against recurrent neural networks, illustrating the transformer’s superior forecasting potential. Complementing this, Huo et al. Huo et al. 2023 developed a hierarchical spatiotemporal model that integrates a long-term temporal transformer with spatiotemporal graph convolutional networks, achieving robust and consistent performance across multiple datasets. Hybrid neural architectures that combine complementary modeling paradigms have likewise been explored. Méndez et al. Méndez et al. 2023 employed a CNN–BiLSTM hybrid model to improve long-term traffic flow prediction, emphasizing the synergy between CNNs for spatial feature extraction and BiLSTMs for temporal sequence learning. Similarly, Djenouri et al. Djenouri et al. 2023 integrated graph convolutional networks with branch-and-bound optimization techniques to further enhance prediction accuracy through algorithmic optimization. In addition, several recent frameworks have addressed data dynamics and scalability challenges. Chen et al. Chen et al. 2021 proposed a streaming GNN framework with continual learning, capable of adapting to the long-term evolution of traffic networks. Xu et al. Zhang 2025 developed a dynamic graph

Read the complete text