基于机器学习的柴油机性能预测

Performance Prediction of Diesel Engine Based on Machine Learning

随着愈发严峻的能源形势和日趋严格的排放法规,柴油机的经济和排放性能研究已成为近年来国内外的重要课题。目前柴油机性能研究主要基于各种试验和仿真方法,但各种新技术的开发应用不断提高柴油机系统的非线性复杂程度,试验和仿真方法越来越难以满足研究需求,同时研究成本也大幅攀升。因此,基于实际台架试验数据,利用机器学习方法对柴油机的有效燃油消耗率和NOx排放进行预测研究,对支持向量回归、决策树、随机森林及反向传播神经网络算法进行测试,确定合适的算法参数,建立了四种不同的预测模型,并通过对比筛选出预测效果较好的反向传播神经网络模型,利用遗传算法对其进一步改进得到遗传算法-反向传播神经网络模型。预测结果表明,遗传算法-反向传播神经网络模型对柴油机有效燃油消耗率和NOx的预测误差分别为1.78%,1.86%,较反向传播神经网络模型预测效果提升了15%左右,具有较好的预测精度和泛化能力。

With the increasingly severe energy situation and stringent emission regulations, the study of economy and emission performance of diesel engine has become an important topic in recent years.At present, the research of diesel engine performance is mainly based on various test and simulation methods. However, the development and application of various new technologies continuously improve the nonlinear complexity of diesel engine system. The test and simulation methods are increasingly difficult to meet the research needs and the research cost is also rising sharply. Therefore,based on the actual bench test data, the representative indexes of diesel engine economy and emission,brake specific fuel consumption and NOx, are predicted by machine learning method. The suitable parameters of support vector regression, decision tree, random forest and back propagation neural networks machine learning algorithms are selected for testing, and four different prediction models are established. The back propagation neural networks model with better prediction effect is screened by comparison, and the genetic algorithm-back propagation neural networks prediction model is further improved from back propagation neural networks model by genetic algorithm. The prediction results show that the prediction errors of the genetic algorithm-back propagation neural networks model for Brake Specific Fuel Consumption and NOx are 1.78% and 1.86%, respectively, which are about 15% higher than those of the back propagation neural networks model. So, the genetic algorithm-back propagation neural networks model has good prediction accuracy and generalization ability.