首页|期刊导航|山东电力技术|基于SHAP的空调负荷预测模型特征标签构建方法

基于SHAP的空调负荷预测模型特征标签构建方法OA

A Feature Label Construction Approach for Air Conditioning Load Prediction Models Using SHAP Value Analysis

中文摘要英文摘要

针对楼宇空调负荷预测中影响因素复杂、特征选择困难的问题,提出了一种基于沙普利加性解释(shapley additive explanations,SHAP)的特征标签构建方法.该方法以"先建模,后筛选"为思路,在基于长短期记忆-时序注意力机制(long short-term memory-temporal attention mechanism,LSTM-TAM)的空调负荷预测模型的基础上,提出基于SHAP特征标签构建方法,量化各输入特征对预测结果的贡献度,通过阈值筛选和重训练验证,实现关键特征的筛选与标签体系构建.以镇江办公楼宇为例,分析SHAP值分布,并构建特征标签优化模型.结果表明,所构建的特征标签显著提升了空调负荷预测精度和时间效率,均方根误差(root mean square error,RMSE)和平均绝对误差(mean absolute error,MAE)显著降低,决定系数R2提升超10%,训练时间缩短40%.

A feature label construction approach is introduced,leveraging SHAP(shapley additive explanations)to tackle the challenges of complex influencing factors and difficult feature selection in building air conditioning load prediction.This method adopts the principle of"model first,filter later"and proposes a SHAP-based feature label construction framework grounded in the long short-term memory with temporal attention mechanism(LSTM-TAM)air conditioning load prediction model.It quantifies the contribution of each input feature to the prediction outcomes,enabling the identification of key features and the establishment of a label system through threshold-based screening and retraining verification.Taking office buildings in Zhenjiang as a case study,this research analyzes the distribution of SHAP values and constructs a feature label optimization model.The results demonstrate that the feature labels developed through this method significantly enhance the accuracy and time efficiency of air conditioning load prediction.Specifically,the approach markedly reduces the root mean square error(RMSE)and mean absolute error(MAE),increases the coefficient of determination(R2)by over 10%,and cuts training time by 40%.

任禹丞;王雨薇;郑杨;胡涵天;邵平

国网江苏省电力有限公司,江苏 南京 210000国网江苏省电力有限公司镇江供电分公司,江苏 镇江 212002国网江苏省电力有限公司镇江供电分公司,江苏 镇江 212002东南大学,江苏 南京 210096国网江苏省电力有限公司镇江供电分公司,江苏 镇江 212002

信息技术与安全科学

特征标签SHAP溯源分析长短期记忆时序注意力

feature labelSHAPtraceability analysislong short-term memorytemporal attention

《山东电力技术》 2026 (8)

100-110,11

国家电网有限公司科技项目"基于关口量测的典型公共建筑空调负荷分解及节电潜力评估技术研究与应用"(5400-202318574A-3-2-ZN). Science and Technology Foundation of State Grid Co.,Ltd."Research and Application of Load Decomposition and Energy-saving Potential Assessment for Air Conditioning Load Based on Gate Measurement in Typical Public Buildings"(5400-202318574A-3-2-ZN).

10.20097/j.cnki.issn1007-9904.250556

评论