Deep ConvLSTM-Inception Network for Traffic Prediction in Smart Cities

Accurate and real-time traffic prediction is essential for smart cities. However, due to the diversity of traffic point of interest in the city, the spatial correlation of urban traffic is extremely complicated. The use of ordinary spatial correlation feature extraction methods cannot cope with this complex spatial correlation. In addition, traffic flow changes also have temporal correlation. The traditional time series feature extraction method can't extract the temporal correlation of traffic data well. Therefore, we propose a ConvLSTM-Inception Network(CL-IncNet) model for traffic prediction. The CL-IncNet uses the Inception module to extract the multi-granularity spatial correlation and uses the Inception module's stack to extract the indirectly affected spatial correlation. In terms of extracting time-related features, we chose ConvLSTM structure. Compared with LSTM and GRU, this structure has fewer parameters and can guarantee the spatial relative relation of output information. The experimental results on the NYC-Taxi dataset show that our method outperforms four well-known baseline results.