Abstract: Abstract: For a plug-in four-wheel-drive hybrid electric vehicle (4WD PHEV), there are 3 power components which can work independently or cooperatively. Therefore, it has many work modes and the energy management control is relatively complicated. As the calculation of the torque request of the driver by the gas pedal travel only is not precise and that method can't reflect the driver's intention, especially the intensity of the acceleration, thus rendering bad power performances and fuel consumption. And to overcome the inherent defects of fuzzy logic and fuzzy PID (proportion integration differentiation) that they relied on prior knowledge to set the parameters and it was difficult to realize good control effect, it was put forward in this paper that the torque identification coefficient could be obtained through RBF (radial basis function) neural network, whose inputs were the gas pedal travel and its change rate, and the output was the torque identification coefficient. The parameters were obtained through experiments and the neural network model was trained to achieve a better accuracy. After the training, the output error was 0.063, which indicated that it was within the required range. The torque request calculation formula was put forward which took account of the torque identification coefficient. Considering the transient characters of the power components, the dynamic models of the power components and the vehicle were built. A control strategy in which the engine should work at the best efficiency when it worked was adopted. The work modes of the car were classified into EV (electric vehicle) mode, series mode, E-charge (engine drive and charge) mode, parallel mode, and 4WD mode. In addition, the rules for each mode and the dynamic functions were briefly presented. The Stateflow diagram of the control strategy was built in Simulink and was shown in the paper. Afterwards, the dynamic programming (DP) was adopted to obtain the optimal output sequence of the power components with the profile given, which was used as a comparison of the control effects with the rule-based strategy. The principle of DP was briefly introduced with a schematic diagram. The state variable of the DP was the state of charge (SOC) of the battery, and the control variables were the output torque of the integrated starter generator (ISG motor), rear motor and the ratio of continuously variable transmission (CVT). The state transition equation, the constraints and cost function were presented. Furthermore, a correctional DP-based (CDP) strategy was put forward to reduce the calculation time and the optimization steps were briefly introduced, as well as the principle of CDP. As a result, the running time of the program under a cycle of 1 180 s was reduced to 14.29% of that before correction, from 14 368 to 2 053 s. In the end, a hardware-in-the-loop test bench, which consisted of the power components, battery and sensors, etc, was built in Simulink/ Motohawk, and the control strategies were compiled into executive codes with D2P (development to product) tools and were tested on the hardware-in-the-loop test bench. The results validated that the rule-based control strategy and the correctional DP-based strategy could realize good control effects. With RBF neural network torque request identification, the velocity error was obviously reduced and the fuel consumption was improved by 4.54%, while with the correctional DP-based strategy, the fuel consumption had another 14.04% reduction. The research methods in this paper can work as a reference when dealing with the control strategies of similar complicated hybrid electric vehicles.