A hybrid machine learning approach for optimising hydrocarbon injection control in diesel oxidation catalyst for diesel particulate filter active regeneration

Accurately predicting hydrocarbon injection quantities is critical to achieving the high-temperature conditions (585–615 °C) required for the active regeneration of diesel particulate filters. This paper uses a physical model of a diesel oxidation catalyst, employing the whale optimization algorithm to optimise the model's parameters. The accuracy of this approach has been validated through bench testing. Under transient conditions, the mean absolute error was 0.312 °C, and the root mean square error was 3.589 °C. Short-term predictions of upstream temperature and exhaust mass flow rate were first generated using the gated recurrent unit algorithm to back-calculate hydrocarbon injection. The results demonstrate that the predicted upstream temperatures closely align with the measured values, with absolute errors confined within 1 °C and maximum deviations not exceeding 5 °C. Predicted exhaust flow rate errors cluster below 0.0028 kg/s, with a maximum error of 0.0194 kg/s. Based on model validation, the prediction of hydrocarbon injection employs a deep, domain-adaptive convolutional neural network with hyperparameters optimised using a genetic algorithm. Experimental results show that the proposed method keeps prediction errors within the range of 0–40 mg/s, with a maximum deviation of no more than 80 mg/s. This outperforms support vector machines, random forests, and deep neural networks. Further validation under transient operating conditions confirms that the predicted results effectively determine the feasible range for hydrocarbon injection. The methodology presented here provides the technical basis and theoretical support for optimising control strategies in diesel oxidation catalyst systems.