Multi-Timescale Traffic Intensity Forecasting

Traffic intensity prediction is a core task in smart city while there is fast-paced urbanization. However, traffic data usually have complex characteristics such as non-linearity, spatiotemporal dependence, and burstiness, which bring great challenges to prediction models. Therefore, this study proposes a hybrid LSTM-IMTRAN model that combines LSTM and the improved Transformer (IMTRAN). The LSTM gate mechanism dynamically captures the short-term and long-term dependencies of time series, and IMTRAN allows the extraction of global features and improves model stability. The proposed model simplifies the Transformer structure by removing the decoder module and replacing the traditional position encoding with an LSTM module, enhancing time series modeling while reducing computational complexity. Based on the Madrid traffic dataset, this study validates multi-scenario traffic intensity data for both regular days and holidays. The experimental results show that the LSTM-IMTRAN model outperforms the LSTM, STGCN, CNN-LSTM, and Transformer models in short-term (15 min, 30 min), medium-term (60 min) and long-term (1 day) predictions, with a root mean square error (RMSE) reduction of approximately 1.87% - 6.47%.