利用基于PSO算法的径向基人工神经网络优化重催干气脱硫

针对重催干气脱硫过程存在进料波动频繁、优化响应滞后导致能量消耗过大等问题，通过Aspen HYSYS V11软件利用Li-Mather物性方法对该系统进行全流程模拟，根据Plackett-Burman设计筛选对目标值具有显著影响的有效因素，利用基于PSO算法的径向基人工神经网络对预测模型进行训练、验证和测试，并在满足净化干气硫化氢浓度约束的前提下对其进行深度优化，以期最小化系统能耗。结果表明，重催干气流量、重催干气硫化氢浓度、贫液哌嗪质量分数、贫液N-甲基二乙醇胺（MDEA）质量分数、胺液循环量、T-3001塔底温度和E-3003贫液出口温度对系统能耗影响非常显著，当以上述因素为输入信号，以系统能耗为网络输出时，7-16-1型径向基人工神经网络预测模型经过4182次迭代后，它的训练样本、验证样本、测试样本均方误差分别为5.08×10^-6、7.78×10^-6和9.56×10^-6，均小于容许收敛误差限10^-5，而其决定系数亦高达0.981、0.975、0.969，表现出良好的相关性。当利用基于PSO算法的径向基人工神经网络对重催干气脱硫系统能耗进行优化时，经过3198次粒子进化迭代后系统能耗仅为0.0649kg_oe/h，较优化前系统能耗0.0713kg_oe/h降低了8.98%，节能效果显著。

Desulfurization optimization of reforming catalytic dry gas using radial basis artificial neural network based on PSO algorithm

To address the issues of excessive energy consumption caused by frequent feed fluctuation and retarded optimization response of desulfurization for reforming catalytic dry gas process, the flowsheet simulation was conducted through the Aspen HYSYS V11 package using Li-Mather physicochemical property calculation method. Screening the effective factors that had a significant influence on the target value was adopted according to Plackett-Burman design. The radial basis artificial neural network based on the PSO algorithm was utilized to train, validate, and test the prediction model. On the premise of satisfying the constraint of hydrogen sulfide content in purified dry gas, the deep optimization was carried out to minimize the energy consumption of the system. The results show that the flowrate and hydrogen sulfide content of reforming catalytic dry gas, the piperazine and N-methyl diethanolamine content in lean solution, circulation quantity of amine solution, the bottom temperature of T-3001, and the outlet temperature of a lean solution of E-3003 play a crucial role in energy consumption of the system. The prediction model of the 7-16-1 radial basis artificial neural network where the aforementioned factors were taken as the input signal and the system energy consumption as the network output evolves 4182 epochs. The mean square errors of training samples, verification samples, and test samples are 5.08×10^-6, 7.78×10^-6, and 9.56×10^-6 respectively, which are less than the allowable convergence error limit of 10^-5. A good correlation is presented as the determination coefficients reach 0.981, 0.975, and 0.969. When the radial basis artificial neural network with the PSO algorithm is used to optimize the energy consumption of the desulfurization system for reforming catalytic dry gas, the system energy consumption is reduced to be merely 0.0649kg_oe/h after 3198particle evolution iterations, which is 8.98% lower than 0.0713kg_oe/h before optimization, and the energy saving effect is significant.