The energy management of parallel Hybrid Electric Vehicles (HEVs) is the problematic of determining at each instant the optimal torque split between the Internal Combustion Engine (ICE) and the Electric Motor (EM). Two opposite dynamics are related to this problem: the long dynamic of the battery State Of Charge (SOC) and the short dynamic of the ICE torque and speed variations due to the driver’s power demand. Due to estimation and control errors in the air loop, these ICE variations of torque and speed lead to an imbalance in the air/fuel mix stoichiometry, causing an increase in pollutants generation. This paper aims at solving this issue by computing online an optimal torque split, taking into account the engine transients and their impacts on pollutants generation. The proposed Model Predictive Control (MPC) strategy is based on a Single Layer Perceptron pollutants generation model, trained on experimental data. The control strategy used, when considering a prediction horizon long enough, allows to reduce respectively by 0.9% the fuel consumption, by 23% the CO generation, by 11% the HC generation and by 31% the PM generation at the cost of an S% increase in $\mathrm{NO}_{\mathrm{x}}$ generation.