基于关键特征排序的可解释碳排放预测模型  

Interpretable carbon emission prediction model based on key feature ranking

作　　者：张向阳[1,2,3] 刘树仁 刘宝亮[4] 李长春 付占宝 ZHANG Xiangyang;LIU Shuren;LIU Baoliang;LI Changchun;FU Zhanbao(The Institute of Computer Technology,Northwest Branch of China Petroleum Exploration and Development Research Institute,Lanzhou 730020,China;The Key Laboratory of Internet of Things for China National Petroleum Corporation,Lanzhou 730020,China;School of Mechanical Science and Engineering in Northeast Petroleum University,Daqing 163318,China;Heze Maternal and Child Health Care Hospital,Heze 274000,China)

机构地区：[1]中国石油勘探开发研究院西北分院计算机技术研究所,甘肃兰州730020 [2]中国石油天然气集团有限公司物联网重点实验室,甘肃兰州730020 [3]东北石油大学机械工程学院,黑龙江大庆163318 [4]菏泽市妇幼保健院,山东菏泽274000

出　　处：《中国石油大学学报（自然科学版）》2024年第4期190-197,共8页Journal of China University of Petroleum(Edition of Natural Science)

基　　金：国家自然科学基金面上项目(52374067)。

摘　　要：提出基于关键特征排序的可解释碳排放预测模型(EEMD-LSTM-ATT),选取人口总数、城镇化率、第一产业国内生产总值、第二产业国内生产总值、第三产业国内生产总值与进出口贸易总额这6个变量,以非线性预测能力强的长短时记忆网络为基线模型,采用注意力机制提取影响因素与时间属性的权重信息。结果表明:该模型一方面能够抑制模态混叠的产生,减少数据非线性对于模型预测带来的影响;另一方面能够解释不同时间属性与不同影响因素对于碳排放的重要性程度,使得预测结果具备可解释性;将影响因素与时间属性的权重信息加入模型的训练过程能够促进碳排放影响因素与模型预测有机结合;本文方法可实现高精度碳排放预测,均方根误差为3.772,平均绝对误差为3.416,拟合优度为0.880。An interpretable carbon emission prediction model(EEMD-LSTM-ATT)based on the key feature ranking was proposed,where six variables were selected,i.e.the total population,the urbanization rate,the primary industry GDP,the secondary industry GDP,the tertiary industry GDP,and the total import/export trade.Using the long and short-term memory network,which has a strongly nonlinear prediction ability,as the baseline model,innovatively adopts the attention mechanism to extract the influencing factors and the total amount of trade.The attention mechanism was used to extract the weight information of the influencing factors and time attributes.The results show that,on one hand,the model can inhibit the generation of modal overlap and reduce the influence of data nonlinearity on the model prediction;on the other hand,it can explain the importance of different time attributes and different influencing factors on the carbon emission,which makes the prediction results interpretable.In addition,the weighting information of influencing factors and time attributes is added to the training process of the model,which can promote the organic combination of carbon emission influencing factors and model prediction.The method of this paper can achieve a high-precision carbon emission prediction,with the R MSE being 3.772,the R MAE being 3.416,and the R 2 being 0.880.

关 键 词：集合经验模态分解 长短期记忆模型 注意力机制 预测模型 

分 类 号：X192[环境科学与工程—环境科学]