A hybrid ensemble approach with LSTM and gradient boosting for traffic flow prediction in intelligent transportation systems

Background: Traffic flow forecasting is one of the critical challenges in modern transportation systems, with tremendous ramifications for traffic congestion alleviation, traffic light optimization, and urban mobility development. Yet, due to the dynamic and non-linear characteristics of traffic flows, as influenced by temporal states, meteorological conditions, and special events, their forecasting is an intricate endeavor. Conventional approaches, including statistical models and isolated machine learning approaches, tend to fall short in capturing the complex temporal dependencies and non-linear relationships inherent in traffic flow data. Objective: This thesis aims to overcome the limitations of traditional traffic flow prediction models by proposing a new hybrid approach that combines the strengths of Gradient Boosting, LightGBM, and Long Short-Term Memory (LSTM) networks, with the goal of achieving enhanced prediction accuracy and robustness. Methods: The Road Traffic Prediction Dataset, published by the Huawei Munich Research Center (HMRC), is used in this study. It contains temporal traffic flow measurements collected at six urban intersections over 56 days. The dataset is preprocessed and utilized to train and test a variety of models, including parametric methods, Machine Learning (ML) techniques, Deep Learning (DL) architectures, and ensemble learning models. Model performance is evaluated using standard regression metrics: Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), R-squared (R²), Mean Absolute Percentage Error (MAPE), and Explained Variance Score (EVS). Results: The suggested hybrid model outperforms all other baseline models with better accuracy and reliability. It achieves an R² of 0.9684, an MAE of 8.27, and an RMSE of 12.54, outperforming parametric models, machine learning models, and deep learning models in all evaluation metrics. Compared to individual models like Gradient Boosting, LightGBM, and LSTM, the hybrid model demonstrated better stability and reduced error levels at all times, pointing to the gains wrought by the synergy of the diverse learning paradigms. Moreover, the hybrid model ensured high performance across varying traffic patterns, such as peak and off-peak times, thereby reflecting its capacity to endure under evolving traffic conditions. Results validate that ensemble learning methods executed in conjunction with deep learning methodologies successfully model the intricate temporal and non-linear dynamics of city traffic flow patterns, thereby optimizing predictability and generalizability. Conclusion: The thesis findings have important implications for traffic flow prediction and urban transportation management. By integrating multiple modeling approaches, the proposed hybrid methodology captures a more complete and accurate representation of traffic flow behavior. Additionally, the study identifies challenges such as noise in data and real-time deployment considerations for future research. In summary, this dissertation advances the field of intelligent transportation systems (ITS) by demonstrating the effectiveness of hybrid methods for traffic prediction. The proposed hybrid model sets a new benchmark for future research in ITS.