Research on Trajectory Prediction and Risk Control for Distributed-Drive Electric Vehicles Based on a Spatiotemporal Transformer and an Improved DRA Algorithm

The road ahead is no longer just a path to be traveled; for the modern vehicle, it is a complex, shifting puzzle of moving parts and invisible forces. As cars become more autonomous and electric, the challenge of keeping them safe has shifted from simply reacting to danger to anticipating it. This requires a system that can look at the chaotic dance of traffic—cars accelerating, braking, and weaving—and predict exactly where each vehicle will be a few seconds from now. This is the realm of trajectory prediction. For a specific type of vehicle known as a distributed-drive electric car, where each wheel is powered by its own motor, the physics are even more intricate. These cars can twist and turn with incredible precision, but that same flexibility makes their movements harder to forecast, especially when they are interacting with other drivers in a crowded city. If a computer cannot accurately guess where a car is going, it cannot warn the driver or the car itself in time to prevent a crash.

To solve this, researchers at the Hunan Institute of Traffic Engineering have developed a new way to teach computers how to see the future of traffic. They combined two powerful tools: a sophisticated pattern-recognition system inspired by how humans process language, and a clever search method that helps the computer find the perfect settings for its brain. The result is a system that can predict the path of these advanced electric vehicles with startling accuracy, even in the most confusing traffic situations. By understanding the subtle ways cars influence one another, this new method offers a clearer window into the future of the road, providing the reliable information needed to keep drivers safe before a danger even materializes.

The core of this research lies in addressing a specific weakness in how computers currently learn to predict movement. Traditional methods often struggle to keep up with the rapid, non-linear changes in how electric cars move, especially when multiple vehicles are interacting. The researchers turned to a deep learning model called a Transformer, which is excellent at understanding long sequences of events and the relationships between them. However, like a complex instrument, this model requires precise tuning to work correctly. If the settings are slightly off, the predictions become unreliable. Finding the perfect settings is like searching for a needle in a massive, multi-dimensional haystack, a task that is notoriously difficult for standard computer algorithms. They often get stuck in local solutions, thinking they have found the best answer when a much better one exists just around the corner, or they take so long to search that they are useless for real-time safety systems.

To overcome this, the team introduced an improved version of an optimization algorithm they call the Improved Divine Religions Algorithm. While the name sounds abstract, the logic is grounded in how groups of people interact. Imagine a community where individuals constantly adjust their beliefs based on the influence of leaders and their neighbors. The algorithm mimics this social dynamic to search for the best solution. The researchers enhanced this process by adding two key features. First, they created a mechanism to skip unnecessary calculations for individuals that are not improving, saving valuable computing time. Second, they allowed different groups within the search to occasionally swap members, ensuring that the search does not get trapped in one area but continues to explore the entire landscape of possibilities. This hybrid approach allowed the system to find the optimal settings for the prediction model much faster and more accurately than previous methods.

The team tested this new system using a custom-built "Multi-Vehicle Interaction Trajectory Dataset for Distributed-Drive Electric Vehicles." The data covered a wide range of scenarios relevant to the study's optimization and validation goals, including high-frequency recordings of vehicle positions, velocities, and torques from eight experimental vehicles equipped with distributed-drive systems. The researchers fed this rich history of movement into their new model, which then had to predict future states based on the input sequences. The results were striking. In tests involving complex interactions and dynamic scenarios, the new method predicted the future path of the vehicles with an average error of just three centimeters. This is a massive improvement over existing methods, which often missed the mark by over a meter or more.

The study explicitly showed that older, more common methods like those based on simple neural networks or graph-based models were insufficient for these complex, high-speed interactions. In scenarios requiring precise trajectory modeling, the older models often predicted a path that was too slow or drifted away from the actual trajectory. The new model, however, tracked the vehicle's intended path with remarkable fidelity, capturing both the smooth curves and the sudden adjustments. The researchers measured this success using several standard metrics, finding that their approach reduced prediction errors by more than 95 percent compared to the next best method. This level of precision is not just a statistical win; it means the system can identify a potential collision risk much earlier, giving the vehicle's safety systems enough time to intervene.

By combining a powerful pattern-recognition engine with a highly efficient search strategy, this research demonstrates a significant step forward in making autonomous vehicles safer. The ability to predict the future motion of a car with such accuracy, even when it is being driven by a human or another autonomous system in a chaotic environment, provides a solid foundation for active safety. It moves the technology from simply reacting to what has already happened to understanding what is about to happen. For the drivers of tomorrow, this means a road where the car understands the flow of traffic as well as a seasoned human driver, ready to act before a mistake becomes an accident. The work confirms that with the right mathematical tools, we can teach machines to see the road not just as it is, but as it will be.