应用深度学习和网格搜索的变频冷水机组节能优化策略研究  

Research on Optimal Control Strategy of Chiller System Based on Deep Learning and Grid Search

作　　者：韩林志 周镇新 方正辉 郑铁君 陈焕新[1] HAN Linzhi;ZHOU Zhenxin;FANG Zhenghui;ZHENG Tiejun;CHEN Huanxin(School of Energy and Powering Engineering,Huazhong University of Science and Technology,Wuhan 430074,Hubei,China;Shanghai Volkswagen Automotive Co.,Ltd.,Ningbo 315336,Zhejiang,China;Ningbo Hangzhou Bay New Area Xiangyuan Power Supply Co.,Ltd.,Ningbo 315336,Zhejiang,China)

机构地区：[1]华中科技大学能源与动力工程学院,湖北武汉430074 [2]上汽大众汽车有限公司,浙江宁波315336 [3]宁波杭州湾新区祥源动力供应有限公司,浙江宁波315336

出　　处：《制冷技术》2024年第1期52-57,66,共7页Chinese Journal of Refrigeration Technology

基　　金：国家自然科学基金(No.51876070)

摘　　要：为了解决变频冷水机组优化控制中的变量耦合问题,本文利用深度神经网络建立了多层解耦能耗模型。文章收集了宁波某厂的冷水机组实时运行数据,作为模型的训练集和测试集,模型的第一层通过室外环境变量和相关控制变量,分别实现了冷却水泵频率预测、冷冻水回水温度预测和冷冻水泵频率预测。模型的第二层在此基础上,进一步建立了冷水机组功耗预测模型。最后,通过网格搜索法确定了优化控制参数。经测试,所建立的能耗预测模型的误差为4.61%;此外,在典型日,优化策略运行下的冷水机组能耗下降了9.96%。Chiller systems account for a huge energy consumption.To decoupling the variables of chillers with variable frequency compressor,pumps and fan,the paper uses deep neural networks to establish a multi-layer decoupling model.The real-time operating data of chillers from a factory in Ningbo in the past two years was collected as the training set and test set of the models.The prediction of chilled water return temperature,the frequency of cooling water pumps and chilled water pumps is achieved through environment variables and related control variables on the first stage.On this basis,the second stage we further establish a chiller power consumption prediction model.Finally,the optimal control parameters are determined by the grid search method.After testing,the average relative errors of the energy consumption prediction model are 4.61%.On a typical day,the energy consumption of the chiller under the optimized strategy operation was reduced by 9.96%.

关 键 词：冷水机组 深度学习 优化控制 

分 类 号：TB611[一般工业技术—制冷工程] TQ051.5[化学工程]