A Hybrid LSTM Network for Long-Range Vehicle Trajectory Prediction Based on Adaptive Chirp Mode Decomposition

Vehicle trajectory prediction is essential for intelligent transportation systems and smart cities, but the prediction of long-range trajectories is still challenging. Since the speed of a moving vehicle is relatively high, coordinates of discrete trajectory points may vary from 0 to 10 km in long-range trajectory prediction. Moreover, the vehicle trajectory contains random noises due to the instability of GPS signals in urban areas and the error accumulation effect of the other sensors for vehicle positioning. To address this issue, the discrete trajectory data is treated as a signal, and multiple signal components are extracted from the original data through the adaptive chirp mode decomposition (ACMD) algorithm to capture spatial information at different spatial scales and filter out chaotic noises caused by other assisted positioning sensors simultaneously. To overcome the error propagation of long short term memory (LSTM), since the quality of long-range predictions relies on short-range prediction accuracies, a hybrid LSTM model is proposed based on a combination network of forward LSTM, bidirectional LSTM (BILSTM), and backward LSTM. Taking advantage of an improved whale optimization algorithm (IWOA) to optimize the hyperparameters of the hybrid-LSTM model, the embedding layer decomposes the input trajectories and further learns and trains through the hybrid LSTM layer to effectively overcome the large prediction errors. Extensive experiments based on our dataset and a public taxi trajectory dataset show that the hybrid LSTM model based on ACMD and IWOA outperforms the existing state-of-the-art methods in terms of accuracy and stability.