闫乃锋, 王晨. BP神经网络在加氢裂化装置航煤性质软测量中的应用[J]. 工业催化, 2020,28(8): 65-69.
Yan Naifeng, Wang Chen. Application of BP neural network in soft sensing of kerosene properties in hydrocracking unit[J]. Industrial Catalysis, 2020,28(8): 65-69.
BP neural network with single hidden layer was constructed by using Matlab,and the soft-sensing application of kerosene properties in the medium pressure hydrocracking unit was carried out.The model was trained to predict kerosene flash point and final boiling point(FBP) with a training set of 700 sample data,and respectively a mean square error (RMSE) of 1.57 ℃ and 2.74 ℃for flash point and FBP prediction were obtained by using BP model on a validated set with 152 sample data,furtherly a generalized RMSE of 1.87 ℃and 1.98 ℃on a test set with 80 sample data was achieved. Another model was trained to predict kerosene density with a training set of 300 sample dataand a RMSE of 2.18 kg·m-3 was obtained by using BP model on a validated set with 100 sample data,furtherly a generalized RMSE of 2.72 kg·m-3 on a test set with 70 sample data was achieved respectively.The generalized RMSEs demonstrated that the BP neural network with single hidden layer could meet the requirements of industrial soft sensing of kerosene properties by reasonably selecting characteristic variables and designing network architecture.
Keyword:
BP neural network; hydrocracking; kerosene properties; soft sensing; generalization
SuXin, WuYingya, PeiHuajian, et al. Prediction of coke yield of FCC unit using different artificial neural network models[J]. China Petroleum Processing and Petrochemical Technology, 2016, 18(3): 102-109. [本文引用:1]
[3]
张笑天, 颜学峰, 钱锋. 基于多神经网络模型的石脑油干点软测量[J]. 控制工程, 2004, 11(z1): 52-54, 64. ZhangXiaotian, YanXuefeng, QianFeng. Soft sensor of PDU naphtha dry point based on multiple neural network[J]. Control Engineering of China, 2004, 11(z1): 52-54, 64. [本文引用:1]
[4]
钱欣瑞, 史彬, 鄢烈祥. 基于NARX神经网络的油品性质软测量建模[J]. 计算机与应用化学, 2017, 34(1): 64-68. QianXinrui, ShiBin, YanLiexiang. Soft-sensing modeling for crude products properties based on NARX neural network[J]. Computers and Applied Chemistry, 2017, 34(1): 64-68. [本文引用:1]
[5]
杨杰. 中压加氢改质催化剂再生及其反应性能评价[J]. 石油炼制与化工, 2017, 48(1): 67-71. YangJie. Regeneration and reaction performance of catalysts in MHUG process[J]. Petroleum Processing and Petrochemicals, 2017, 48(1): 67-71. [本文引用:1]
[6]
王庆波, 张毓莹, 胡志海, 等. 煤柴油加氢裂化装置掺炼重凝析油工艺研究[J]. 石油炼制与化工, 2010, 41(12): 1-5. WangQingbo, ZhangYuying, HuZhihai, et al. A study of MHC unit processing feedstock blended with heavy condensate oil[J]. Petroleum Processing and Petrochemicals, 2010, 41(12): 1-5. [本文引用:1]
The hydrodenitrogenation (HDN) experiments of five low-quality feeds of mixed gasoline and diesel were carried out in an 100 mL hydrogenation device under the conditions of using Ni-Mo-P/Al2O3 catalyst, with reaction temperature of 320—360 ℃, space velocity of 1.2—2.0 h-1, hydrogen to oil ratio of 350—550 and reaction pressure of 6—8.5 MPa. Based on the experimental data, the prediction models for HDN rate of these mixed feeds were established by BP neural network and RBF neural network respectively. The calculation results show that the average relative error of BP neural network in predicting HDN rate is 3.42%, which of RBF neural network is 2.58%. Both of them could meet the industrial prediction requirements, however, the prediction performance of that by RBF neural network seems better. The effects of feed properties and process conditions on HDN rate of mixed feeds are further studied with RBF neural network. The sequence of feed properties affecting HDN rate is as follows: sulfur content > density > nitrogen content > 50% distillation point > viscosity > bromine value. The sequence of process conditions affecting HDN rate is as follows: temperature > space velocity > pressure > hydrogen to oil ratio. These results are helpful to optimize the HDN process conditions for treating mixed feeds of gasoline and diesel.