Accurate and efficient urban traffic flow prediction can help drivers identify road traffic conditions in real-time, consequently helping them avoid congestion and accidents to a certain extent. However, the existing methods for real-time urban traffic flow prediction focus on improving the model prediction accuracy or efficiency while ignoring the training efficiency, which results in a prediction system that lacks the scalability to integrate real-time traffic flow into the training procedure. To conduct accurate and real-time urban traffic flow prediction while considering the latest historical data and avoiding time-consuming online retraining, herein, we propose a scalable system for Predicting short-term URban traffic flow in real-time based on license Plate recognition data (PURP). First, to ensure prediction accuracy, PURP constructs the spatio-temporal contexts of traffic flow prediction from License Plate Recognition (LPR) data as effective characteristics. Subsequently, to utilize the recent data without retraining the model online, PURP uses the nonparametric method k-Nearest Neighbor (namely KNN) as the prediction framework because the KNN can efficiently identify the top-k most similar spatio-temporal contexts and make predictions based on these contexts without time-consuming model retraining online. The experimental results show that PURP retains strong prediction efficiency as the prediction period increases.

Time series forecasting has attracted wide attention in recent decades. However, some time series are imbalanced and show different patterns between special and normal periods, leading to the prediction accuracy degradation of special periods. In this paper, we aim to develop a unified model to alleviate the imbalance and thus improving the prediction accuracy for special periods. This task is challenging because of two reasons: (1) the temporal dependency of series, and (2) the tradeoff between mining similar patterns and distinguishing different distributions between different periods. To tackle these issues, we propose a self-attention-based time-varying prediction model with a two-stage training strategy. First, we use an encoder-“decoder module with the multi-head self-attention mechanism to extract common patterns of time series. Then, we propose a time-varying optimization module to optimize the results of special periods and eliminate the imbalance. Moreover, we propose reverse distance attention in place of traditional dot attention to highlight the importance of similar historical values to forecast results. Finally, extensive experiments show that our model performs better than other baselines in terms of mean absolute error and mean absolute percentage error.