基于K-means聚类+灰色理论+BP神经网络的区域天然气长期需求预测

A combined approach for forecasting regional long-term natural gas demand integrating K-means clustering,grey theory,and BP neural network

【目的】不同区域影响天然气需求量的因素存在差异,数据集包含的数据特征也不尽相同,同时天然气长期需求预测存在样本数据少的问题,因此较难构建各区域通用的需求预测模型。【方法】选取山东省11个地级市为研究对象,根据天然气年度消费量、GDP、人口等影响天然气需求量的主要因素,将多个地区、多时间跨度的数据作为总样本库,使用皮尔逊相关系数对样本特征进行初筛,利用K-means聚类算法对各样本数据进行聚类,选取能源消费结构相似的3个样本点,并将样本点对应的下一时间点的天然气需求量作为数据样本的新特征;同时,将灰色理论预测输出结果作为BP神经网络的输入样本,基于新的样本数据特征与BP神经网络构建组合预测模型。【结果】基于K-means聚类+灰色理论+BP神经网络的预测方法有效利用了相似能源结构的城市天然气历史需求量,并结合灰色理论预测模型在小样本数据上鲁棒性高的优点,预测得到山东省11个地级市天然气长期需求预测的平均绝对百分比误差为0.57%~6.41%。与传统的灰色理论预测模型、BP神经网络模型、K-means聚类+BP神经网络相比,新预测方法在模型误差、预测结果的稳定性方面均有明显改进。【结论】新预测方法的使用不局限于某一地区,不仅可以为中国各城市未来的天然气需求量预测提供技术支撑,还可以为各级政府及企业开展天然气资源配置计划决策提供参考。(图5,表3,参22)

[Objective] Factors impacting the demand of natural gas vary across regions in China,resulting in distinct data characteristics within corresponding datasets.The availability of adequate sample data for long-term forecasting is limited.Therefore,the development of a generalized demand forecasting model applicable to diverse regions remains a significant challenge.[Methods] This study focused on analyzing eleven prefecture-level cities in Shandong Province.Data from these regions,spanning multiple time periods,were gathered to establish a general database,incorporating key factors influencing natural gas demand,such as annual consumption,GDP,and population.Pearson's correlation coefficient was initially employed to preliminarily screen and identify sample characteristics,followed by the utilization of the K-means clustering algorithm for data clustering.Subsequently,three sample points exhibiting a similar energy consumption structure were selected,and their corresponding natural gas demands at the next time points were used as new sample characteristics.Furthermore,forecast outputs derived from the grey theory were utilized as input samples for a Back Propagation(BP) neural network.As a result,a combined forecasting model was developed by integrating the new sample data with the BP neural network.[Results] The proposed forecasting method,which integrated K-means clustering,the grey theory,and a BP neural network,leveraged the historical natural gas demands of the cities with similar energy structures.By taking advantage of the grey theory's high robustness in dealing with small sized sample data,the proposed forecasting method kept the mean absolute percentage errors of the resulting long-term natural gas demand forecasts for the 11 prefecture-level cities in Shandong Province within the range from 0.57% to 6.41%.Comparative analysis with traditional grey forecasting models,BP neural network models,and K-means clustering+BP neural network models revealed the superiority of the proposed approach,yielding improved prediction results with smaller errors and higher stability.[Conclusion] The proposed forecasting method proves effective in providing technical support for analyzing future natural gas demands in cities across China,surpassing regional limitations.Moreover,it serves as a valuable tool for assisting governments and enterprises at all levels in making decisions on allocation plans of natural gas resources.(5 Figures,3 Tables,22 References)