融合机器学习与SHAP值算法的居民需求响应个体异质性因素挖掘与应用研究  

Research on mining and applications of individual heterogeneity factors in resident demand response by integrating machine learning and SHAP value algorithm

作　　者：王兆华[1,2,3] 刘杰 王博[1,2,3] 邓娜娜[1,2,3] 聂富华[1,2,3] WANG Zhaohua;LIU Jie;WANG Bo;DENG Nana;NIE Fuhua(School of Management and Economics,Beijing Institute of Technology,Beijing 100081,China;Center for Sustainable Development and Smart Decision,Beijing Institute of Technology,Beijing 100081,China;Digital Economy and Policy Intelligentization Key Laboratory of Ministry of Industry and Information Technology,Beijing 100081,China)

机构地区：[1]北京理工大学管理与经济学院,北京100081 [2]北京理工大学可持续发展与智慧决策研究中心,北京100081 [3]数字经济与政策智能工业和信息化部重点实验室,北京100081

出　　处：《系统工程理论与实践》2024年第7期2247-2259,共13页Systems Engineering-Theory & Practice

基　　金：国家自然科学基金(72243001,72074026,72141302,72321002)。

摘　　要：本研究基于大规模居民电力需求响应(EDR:electricity demand response)实验以及家庭用电调查数据,利用机器学习和SHAP(Shapley additive explanatory)值算法从全局和个体两个层面对影响居民参与需求响应的影响因素进行了识别和异质性分析.研究发现,居民是否参与需求响应活动是外部激励,家庭结构,用电规律与习惯倾向,用电知识等因素共同作用的结果,其效应的大小和极性存在着丰富的异质性.其中,电话营销等外部激励对用户参与需求响应影响最大,其效果在年龄较大以及受教育程度较高的群体较为明显;响应时段基准用电量在1度左右的用户参与倾向较大;节能环保意识较强且具有较高节电条件的家庭参与概率更高.同时,依据SHAP值的交互以及分解性质,在后续需求响应活动中对用户进行分类营销,可以节省93.9%的营销成本,并提高46.4%的参与人数.本研究对不同群体的异质性进行了更为细致的分析研究,为未来新型电力系统下进行更为精确和智能的需求响应提供了重要支撑.Based on the large-scale residential electricity demand response(EDR)experiment and household survey data,this study uses machine learning and SHAP(Shapley additive explanatory)value algorithm to identify and analyze the influencing factors of residents’participation in demand response from the whole and individual levels.Our study found that whether residents participate in EDR activities is the result of the joint action of external incentives,family structure,electricity use habits,electricity use knowledge,and the value and polarity of the effect is varied in heterogeneity.Among them,external incentives such as telemarketing have the greatest impact on customers’participation in demand response,and the effect is more obvious among older and more educated groups;customers with a baseline electricity consumption of about 1 kW·h during the response period have a higher tendency to participate;households with a stronger awareness of energy conservation and higher conditions for saving electricity have a higher probability of participation.At the same time,according to the interaction and decomposition properties of SHAP value,classified marketing for users in subsequent EDR activities can save 93.9%of marketing costs and increase the number of participants by 46.4%.This study has carried out a more detailed analysis and research on the heterogeneity of different groups,providing an important support for more accurate and intelligent EDR to China’s new power system.

关 键 词：需求响应 因素分析 机器学习 SHAP值 

分 类 号：F423.3[经济管理—产业经济]