Online Vehicle Velocity Prediction Based on an Adaptive GRNN with Various Input Signals

To improve the prediction accuracy and computational speed of vehicle velocity prediction (VVP) strategies for energy management, an online VVP strategy based on general regression neural network (GRNN) is proposed and optimized. First, a GRNN was employed to achieve online VVP, with an evaluation of the effects of order and σ on prediction accuracy. Then, the impact of various input signals on the VVP prediction effect was compared, and the minimal ARMSE was found under the input signal combination of vehicle velocity, driving motor torque, and brake pedal opening degree. Subsequently, a GRNN structure determination method (SDM) based on the Akaike information criterion (AIC) was proposed to construct an online VVP model based on adaptive-structure GRNN. Simulation results using real vehicle test data indicate that the online VVP strategy based on GRNN is feasible under various urban driving conditions. Additional case studies have demonstrated that, compared with the GRNN relying solely on historical velocity data, the optimized GRNN with adjusted structure and input signals reduced prediction error by at least 26.3%.