Search

• DDPG is used to dynamically select the Pareto solution of MOICA. • TCN and SRU are used as the main predictors. • The proposed model is compared with nineteen mainstream forecasting models. Traffic congestion is a difficult problem that restricts the construction of urbanization. Spatiotemporal traffic speed forecasting technologies can provide effective technical support for alleviating traffic congestion and ensuring vehicle travel safety. The ensemble learning algorithm is a hot topic in traffic speed modeling. In this field, previous ensemble learning methods mainly adopt the principle of static modeling, which limits the learning ability of the model to dynamic features. To solve this problem, in this paper, a new dynamic ensemble deep deterministic policy gradient recursive network is presented for traffic speed forecasting, which comprises three main modeling steps. In step I, the simple recursive network (SRU) and temporal convolution network (TCN) methods are used as the main predictors to build the traffic speed forecasting model. In step II, the multi-objective imperialist competitive algorithm (MOICA) integrates these neural networks by optimizing the weight coefficients and generating the Pareto solution set. In step III, the deep deterministic policy gradient (DDPG) method dynamically selects the Pareto optimal solution of the MOICA according to the changes in the traffic speed data. The MOICA and DDPG dynamically integrate the forecasting results from the SRU and TCN to obtain the final results. Based on the experimental analysis results, several conclusions can be given as follows: (a) the model presented in this paper can obtain accurate traffic speed forecasting results with MAPE values below 4% on all data sets. (b) the proposed model can achieve better results than thirteen alternative models and four proposed models from other researchers. (c) the proposed model can improve the prediction performance of traditional predictors by about 6%. [ABSTRACT FROM AUTHOR]