Medium term prediction of power consumption of a crude oil pipeline based on a bootstrap method and support vector machine theory
ZHU Zhenyu, BAI Xiaozhong , XU Lei, HOU Lei, LIU Jinhai, GU Wenyuan, SUN Xin
1 College of Mechanical and Transportation Engineering, China University of Petroleum-Beijing, Beijing 102249, China 2 CNPC Key Laboratory of Oil & Gas Storage and Transportation, Beijing 102249, China 3 National Pipe Network Group Northern Pipeline Co., Ltd. Jinzhou Oil and Gas Transmission Branch, Jinzhou 121000, PR China
In general, accurate power consumption prediction is a very important basis for the energy consumption management of a crude oil pipeline operation This is extremely helpful for oil transportation enterprises to reasonably formulate batch scheduling, load distribution and other operation schemes. In general, traditional prediction methods such as process calculation and statistical analysis do not perform very well in processing high-dimensional and non-linear pipeline operation data. In contrast, machine learning methods have better prediction effects under these complex conditions. However, due to the very high cost of data acquisition and the existence of security and confidentiality of the pipeline data, the pipeline data set that can be obtained is often a very small sample data set, so the prediction accuracy of the model established by this method cannot meet the strict requirements of actual production. Therefore, in order to improve the prediction ability of the established prediction models in the case of small sample sets, according to the data generation theory, a pipeline operation power consumption prediction model combining a bootstrap method and a support vector machine is proposed. Firstly, the data of the original small sample set is expanded by the bootstrap method, and virtual samples are generated according to the distribution law of the original data set, and the sample information interval is filled to avoid the problem of over-fitting. Then particle swarm optimization is used to optimize the hyperparameters of the support vector machine to improve the fitting ability of the model. In this paper, a two-station model of an insulated crude oil pipeline in China is taken as an example. As expected, the prediction results show that compared to using only the original data set, most of the predicted values after adding virtual samples are closer to the real values, and when 50 groups of virtual samples were added to the two stations, the average absolute error (MAE) of its monthly
power consumption forecast results were reduced by 32.4 % and 29.7 %, thus proving that by adding the virtual samples to the original data set to expand the scale of data set, it can effectively reduce the prediction error and increase the ability of model fitting. In summary, this method provides a new way to solve the complex problem of insufficient available samples caused by the high cost of pipeline data acquisition and the importance enterprises attach to the data security.
Key words：
crude oil pipeline; energy consumption prediction; bootstrap; support vector machine; small sample; virtual sample
ZHU Zhenyu, BAI Xiaozhong, XU Lei, HOU Lei, LIU Jinhai, GU Wenyuan, SUN Xin. Medium term prediction of power consumption of a crude oil pipeline based on a bootstrap method and support vector machine theory. Petroleum Science Bulletin, 2021, 01: 127-137.
