Long-Short Term Memory-Based Heuristic Adaptive Time-Span Strategy for Vehicle Speed Prediction

Vehicle speed prediction plays an important role in vehicle energy saving and safety research. It can contribute to vehicle energy saving and safety assistant driving, route navigation, automatic transmission gear control, and hybrid electric vehicle predictive control. The research on vehicle speed prediction has important theoretical basis and application value. The data-driven deep learning (DL) model provides a powerful method for building an accurate speed prediction model. However, the traditional vehicle speed prediction model has some limitations in prediction efficiency and accuracy, which fails to take into account the characteristics of the time dimension of speed data. This paper proposes an Long-Short Term Memory(LSTM) vehicle speed prediction model based on heuristic adaptive time-span strategy. The model mainly includes three parts: 1. In view of the instantaneity of the time series, we add weights to the input data, increase the weight of the data near the prediction point, and accelerate the convergence speed and accuracy of the model. 2. Simulated annealing algorithm is adopted to adaptively select the most appropriate time span for the current data. Compared with the traditional vehicle speed prediction model, this approach does not fix the time span and has better data universality. 3. The basic unit of the model is the LSTM model. The time series model is used to make prediction of speed, which is in line with the law of speed data. Validation of the model using driving data from ten vehicles over a 1-year period reveals that the LSTM speed prediction model based on a heuristic adaptive time-span strategy exhibits impressive accuracy and outperforms existing state-of-the-art machine learning models.