Combination of multiple transformer models for short-term freeway traffic speed prediction based on Bayesian model averaging

The Bayesian Model Averaging (BMA) approach is proposed as an advanced method to combine the traffic speed prediction results from three distinct Transformer models, with the aim of enhancing intelligent transportation management systems and facilitating more accurate and efficient route planning. A significant contribution of this approach is its ability to address a key limitation found in existing literature, which often fails to account for the uncertainty inherent in different traffic speed prediction models. Unlike conventional methods that rely on a single model, the BMA approach integrates multiple models to capture a wider range of possible outcomes, thus providing a more robust and reliable prediction. The Transformer models, which are equipped with a multi-head attention mechanism, are particularly well-suited for this task as they can effectively capture long-term traffic speed trends, thereby offering stable predictions over extended periods. This is crucial for handling the dynamic and often unpredictable nature of traffic conditions. To evaluate the effectiveness of the BMA approach, the model was tested using real-world traffic speed data collected from an interstate freeway in Minnesota, covering time intervals ranging from 5 to 60 minutes. The performance of the BMA approach was compared with that of the individual Transformer models, and the results showed a marked improvement in prediction accuracy, particularly for short-term freeway traffic speeds. These findings suggest that the BMA approach not only outperforms single-model predictions but also offers valuable insights for optimizing traffic management, enhancing route planning, and improving overall transportation system efficiency in real-world settings.