基于IMPSO-BPNN的天然气掺氢发动机性能预测

Performance prediction of hydrogen enriched compressed natural gas engine based on IMPSO-BPNN

为进一步提高传统粒子群算法优化反向传播神经网络(PSO-BPNN)模型的性能,基于惯性权重和加速因子对粒子群优化的影响机制,提出一种采用非线性递减惯性权重和非线性加速因子调整策略的改进粒子群算法优化反向传播神经网络(IMPSO-BPNN)方法;将IMPSO-BPNN方法应用于天然气掺氢(HCNG)发动机扭矩、等效燃料比消耗和NO_(x)比排放等性能参数的回归分析与预测,并从预测精度、泛化能力和收敛速度3个方面与其他神经网络方法进行了比较,包括PSO-BPNN、遗传算法优化反向传播神经网络(GA-BPNN)和反向传播神经网络(BPNN)方法。研究结果表明:燃空比和点火提前角均可显著影响HCNG发动机的扭矩、等效燃料比消耗和NO_(x)比排放;以扭矩为预测变量,最优IMPSO-BPNN模型的平均绝对百分比误差分别比PSO-BPNN、GA-BPNN和BPNN方法所建立的最优模型小5.85%、12.62%和17.96%,且最优IMPSO-BPNN模型的相关系数也最大,达到了0.99986,说明IMPSO-BPNN方法所建立模型的预测性能和泛化能力总体上优于其他方法;以NO_(x)比排放为预测变量,最优PSO-BPNN和最优IMPSO-BPNN模型的CPU运行时间比最优GA-BPNN模型均减少约95%,说明与GA-BPNN方法相比,PSO-BPNN和IMPSO-BPNN方法在时间维度上优越性明显。可见,本文提出的IMPSO-BPNN方法相比PSO-BPNN和GA-BPNN方法在预测性能和泛化能力方面均有显著的优势,同时能够保证较高的计算效率。

To further improve the performance of traditional particle swarm optimization back-propagation neural network(PSO-BPNN)model,based on the influence mechanisms of inertia weight and acceleration factor on particle swarm optimization,an improved particle swarm optimization back-propagation neural network(IMPSO-BPNN)method adopting non-linear decreasing inertia weight and non-linear acceleration factor was proposed.The IMPSO-BPNN method was applied to the regression analysis and prediction of performance parameters such as torque,equivalent brake specific fuel consumption,and brake specific NO_(x)emission of a hydrogen enriched compressed natural gas(HCNG)engine.It was also compared with other neural network methods in terms of prediction accuracy,generalization ability,and convergence speed,including PSO-BPNN,genetic algorithm optimized back-propagation neural network(GA-BPNN),and back-propagation neural network(BPNN)methods.Research results show that the fuel-air ratio and spark advance angle can significantly affect the torque,equivalent brake specific fuel consumption,and brake specific NO_(x)emissions of the HCNG engine.With torque as the predictive variable,the average absolute percentage error of the optimal IMPSO-BPNN model is 5.85%,12.62%,and 17.96%smaller than those of PSO-BPNN,GA-BPNN,and BPNN models,respectively,and the correlation coefficient of the optimal IMPSO-BPNN model is 0.99986,also the highest among these models,which indicates that the prediction performance and generalization ability of the model established by the IMPSO-BPNN method are generally superior to those established by other methods.With brake specific NO_(x)emission as the predictive variable,the CPU running times reduce by 95%in both the optimal PSO-BPNN model and the optimal IMPSO-BPNN model compared with the optimal GA-BPNN model,which demonstrates the superiority of PSO-BPNN and IMPSO-BPNN methods to the GA-BPNN method in terms of time dimension.Therefore,compared with PSO-BPNN and GA-BPNN methods,the proposed IMPSO-BPNN method has significant advantages in prediction performance and generalization ability,and ensures high computing efficiency.