traffic forecasting

To address the challenge of effectively modeling multi-scale, nonlinear, and highly noisy temporal patterns in traffic flow data—features poorly captured by single-model approaches—this paper proposes a hybrid forecasting framework based on Seasonal and Trend decomposition using Loess (STL). The original time series is decomposed into trend, seasonal, and residual components, each modeled by a specialized algorithm: LSTM for long-term dependencies, ARIMA for periodic patterns, and XGBoost for nonlinear residuals. Final predictions are obtained via multiplicative ensemble integration. Compared to conventional single-model baselines, the proposed framework achieves superior prediction accuracy, enhanced interpretability, and improved robustness. Experiments on New York City traffic flow data demonstrate reductions in MAE and RMSE exceeding 15%, alongside an R² improvement of approximately 0.08, validating the efficacy of multi-component decomposition coupled with heterogeneous model collaboration.

Existing deep spatiotemporal models for traffic flow forecasting assume temporally independent prediction errors, leading to inaccurate uncertainty quantification. To address this, we propose a dynamic error modeling framework that explicitly captures the spatiotemporal dependencies of forecast errors via a Matrix-Autoregressive (MAR) model—relaxing the independence assumption. We introduce, for the first time, an anisotropic error likelihood into the loss function, enabling joint optimization of the primary forecasting model and the error dynamics. Our approach further incorporates interpretable autoregressive coefficients and a structured spatiotemporal covariance matrix. Evaluated on diverse real-world speed and flow datasets, the method significantly outperforms state-of-the-art baselines: it improves point forecast accuracy while delivering well-calibrated, physically interpretable quantile forecasts and rigorous uncertainty quantification.

This paper addresses the potential and challenges of leveraging large language models (LLMs) for traffic mobility analysis—specifically time-series forecasting. Methodologically, it presents a systematic survey and framework development: (1) it comprehensively maps the research landscape of LLMs in traffic forecasting, clarifying integration paradigms with conventional time-series models, data adaptation bottlenecks, and domain-knowledge injection mechanisms; (2) it proposes the first taxonomy of LLM-based methods for traffic mobility prediction, covering multi-source data encoding, prompt engineering, and model-cooperative modeling. Key limitations identified include inadequate semantic-temporal alignment, limited real-time inference capability, and weak interpretability. The work further outlines scalable technical pathways to overcome these constraints. Collectively, the study establishes a theoretical foundation and practical roadmap for deeply integrating LLMs into intelligent traffic forecasting systems.

To optimize radio resource allocation in 5G networks, this paper addresses the dual challenge of uplink traffic intensity forecasting and application-layer classification (e.g., web browsing, VoIP/video calls, video streaming). We propose a unified multitask learning framework based on a shared Long Short-Term Memory (LSTM) architecture that jointly models temporal traffic dynamics and semantic application categories. Our approach integrates time-series feature engineering with a sliding-window prediction scheme and introduces— for the first time—a novel metric: burst occurrence probability estimation, enhancing network adaptability to sudden traffic surges. Evaluated on real-world 5G base station traces, the model achieves 92.3% application classification accuracy, an average absolute error of less than 8.7% in packet-volume forecasting, and an AUC of 0.89 for burst probability prediction—outperforming dedicated single-task baselines across all metrics.

Existing traffic flow forecasting models overly rely on periodic patterns, rendering them ineffective at modeling aperiodic events—such as accidents—thus compromising prediction robustness. To address this, we propose a dual-branch decoupled architecture that decomposes traffic time-series signals into intrinsic spatiotemporal patterns and exogenous environmental context containing aperiodic events. We further design a cross-temporal attention module to jointly model periodic and aperiodic dynamics while capturing high-order spatiotemporal dependencies. Our method is plug-and-play compatible with mainstream models (e.g., STGCN, AGCRN). Extensive experiments on multiple real-world datasets demonstrate an average 9.6% reduction in prediction error, significantly improving responsiveness to突发事件 and enhancing generalization robustness.

Existing statistical and shallow machine learning approaches struggle to effectively model the complex spatiotemporal dependencies inherent in multivariate network traffic. To address this limitation, this work proposes two deep learning methodologies: first, a tailored graph attention network (GAT) that jointly captures topological structure and temporal dynamics; and second, a novel framework that integrates clustering-based preprocessing with fine-tuned multimodal large language models (LLMs) for traffic forecasting—an approach not previously explored in this domain. Experimental results on real-world datasets demonstrate that the LLM-based method achieves superior overall accuracy and generalization performance, while the GAT significantly reduces prediction variance across diverse sequences and time horizons, thereby overcoming key constraints of conventional time-series models.

Traditional traffic signal control struggles to adapt to dynamic demand and lacks interpretability. This work proposes a hierarchical control framework that integrates short-term traffic state prediction using LSTM, candidate phase generation, structured reasoning via a large language model (LLM), and safety-constrained action filtering. Notably, the LLM is employed for the first time as a high-level decision-support module under explicit safety constraints, rather than as a low-level controller. Extensive experiments across multiple scenarios in the SUMO simulation platform demonstrate that the proposed approach significantly improves traffic throughput while achieving zero constraint violations through the safety filter, thereby balancing performance and interpretability.

This study addresses the challenges of high model complexity and poor interpretability in traffic time series forecasting by proposing TSNN, a novel framework that introduces non-parametric methods to this domain for the first time. TSNN employs a parameter-free multi-layer architecture to decouple temporal patterns, explicitly models periodicity, and constructs a memory bank from the training set to enable prediction through similarity matching. Requiring no trainable parameters, TSNN achieves both high interpretability and competitive predictive performance. Experimental results on four real-world traffic flow datasets demonstrate that TSNN attains forecasting accuracy comparable to state-of-the-art deep learning models, while its decision logic is validated through intuitive visualizations.

This study addresses the limited generalizability of existing deep learning approaches for traffic forecasting, which typically rely heavily on dataset-specific training and fine-tuning. We systematically evaluate the zero-shot forecasting capability of Chronos-2—a general-purpose time series foundation model—across ten diverse real-world traffic scenarios, including highway traffic volume, urban road network speed, bike-sharing demand, and EV charging station usage. Our results demonstrate that Chronos-2, without any task-specific fine-tuning, serves as a strong universal baseline, matching or surpassing the accuracy of specialized models on most datasets, with particularly pronounced advantages in long-horizon prediction. Furthermore, its native support for uncertainty quantification yields reliable and sharp probabilistic prediction intervals, enhancing its practical utility in real-world deployment.

This study addresses limitations in existing air travel demand forecasting models, which typically rely solely on temporal dynamics while neglecting the dual dynamics of intra-flight booking accumulation and inter-flight demand shifts, and often fail to generalize when aircraft types change. To overcome these issues, this work proposes a dual-stream LSTM architecture that integrates self-attention, cross-attention, and hybrid attention mechanisms to jointly model intra-flight booking sequences and fixed-offset patterns across flights for the first time. The model further incorporates residual connections and a gated fusion strategy to enhance generalization. Evaluated on real-world data from Biman Bangladesh Airlines, the proposed approach achieves a mean absolute error (MAE) of 2.8167 and an R² score of 0.9495, significantly outperforming baseline methods, and has already been deployed in operational settings.