Study On Semi-supervised Constrained Clustering By Fast Search And Find Of Density Peaks And Its Application On Air-condition Control System

In recent years,with the rapid development of machine learning,many machine learning technologies are introduced into feed-forward control.Two of the most commonly used machine learning methods in air conditioning systems are Neural Network and Support Vector Machine.Most of machine learning methods require pre-knowledge of the tag information.However,it's hard to label massive industrial data with little label information.In this paper,a semi-supervised clustering algorithm is studied to label industrial data and applied into air conditioning control algorithm.This paper proposed two semi-supervised clustering methods named Semi-supervised constraint ensemble clustering by fast search and find of density peaks(SiCE-CFDP)and Constraint-based clustering by fast search and find of density peaks with enhanced class structure mining(CCFDP).The two proposed algorithms are supposed to have better clustering performance and wider application fields than Clustering by fast search and find of density peaks(CCFDP).Several open-source datasets and air-condition system datasets are selected as experiment datasets.Then,CCFDP is applied to label data and forecast system load in the air-condition system,which forming part of the improved control scheme of the air-condition system.Finally,a feedforward control scheme based on semi-supervised clustering method is proposed in this paper with the use of CCFDP.The main achievement of this paper are as follows:1)CFDP algorithm is a new density clustering algorithm first published on Science.CFDP has several inconveniences in real applications.First,it is hard to correctly select the cluster centers.Besides,the determination of its class numlber depends on empirical knowledge,which may be scarce in some applications scenarios.To solve these problems,the paper tries to introduce semi-supervised information and ensemble learning approach into CFDP.Compared to CFDP,the SiCE-CFDP proposed in this paper can make full use of finite constraint information without affecting the accuracy of algorithm.Experiment results of large-scale datasets show SiCE-CFDP's superior performance over other well-known constraint-based clustering algorithms.2)To further improve SiCE-CFDP's performance in handling small size dataset,a CCFDP is proposed in this paper.CCFDP simplifies the time complexity of SiCE-CFDP algorithm.Different from SiCE-CFDP,CCFDP exploits the heuristic information to discover the hierarchical relationships among data instead of using emsemble learning strategy.CCFDP adopts Support Vector Machine to replace Logistic Regression to judge the optimal separating plane.CCFDP is boosted with combining three concepts of density clustering,hierarchical clustering and semi-supervised clustering.In CCFDP,a more efficient enhanced class structure is adopted to replace ensemble learning.Thus,CCFDP is faster than other well-known semi-supervised clustering algorithms.Experiment results show that,compared to SiCE-CFDP,CCFDP can achieve a slightly higher clustering accuracy when given small size datasets and much higher accuracy when given large-size datasets.Accept for the improvement on accuracy,CCFDP also shows a faster speed over other algorithms.3)A feedforward control scheme based on semi-supervised clustering method is proposed as an improvement of basic control algorithm going through the structural adjustment of the traditional control algorithm.The new control scheme uses the semi-supervised clustering algorithm learn data distribution,so as to label the air-condition data.Therefore,the new control scheme could give out a feed forward load prediction by comparing the new coming air-conditioning operation data with data distribution.The feed forward control structure allows the control scheme to adjust the actuator in advance once detecting building load changes.Experiments of the air-condition system datasets show that the improved feedforward control scheme has a better control performance and saves more energy than the improved control algorithm without feedforward control.