Chiller Data Research Articles

Outlier detection via multiclass deep autoencoding Gaussian mixture model for building chiller diagnosis

In HVAC systems, the application of fault detection and diagnosis (FDD) for chillers has brought many benefits to building utilities, i.e., indoor thermal comfort, energy saving, and energy consumption management. Although recent studies have achieved great progress on chiller fault diagnostic problem, some serious issues have been raised in these studies and need to be resolved comprehensively. In that context, two problems are listed and studied in this paper namely outliers detection and the insufficiency of labeled data. To deal with outliers mixed in chiller data, this paper proposes a supervised multiclass deep autoencoding Gaussian mixture model (S-DAGMM) algorithm which is an ensemble model of individual unsupervised DAGMMs. This mechanism helps S-DAGMM to detect ambiguous outliers in both training and testing data. In addition, this paper proposes to use DAGMM to pretrain a deep neural network (DNN). Since DAGMM can learn the data distribution from the unlabeled data, it can prevent DNN from getting stuck on local optima. Comprehensive experimental results in this study have proved the outstanding performance of the proposed approach, compared with state-of-the-art rival methods.

Artificial neural networks applications in building energy predictions and a case study for tropical climates

This study presents artificial neural network (ANN) methods in building energy use predictions. Applications of the ANN methods in energy audits and energy savings predictions due to building retrofits are emphasized. A generalized ANN model that can be applied to any building type with minor modifications would be a very useful tool for building engineers. ANN methods offer faster learning time, simplicity in analysis and adaptability to seasonal climate variations and changes in the building's energy use when compared to other statistical and simulation models. The model herein is presented for predicting chiller plant energy use in tropical climates with small seasonal and daily variations. It was successfully created based on both climatic and chiller data. The average absolute training error for the model was 9.7% while the testing error was 10.0%. This indicates that the model can successfully predict the particular chiller energy consumption in a tropical climate. Copyright © 2005 John Wiley & Sons, Ltd.

Information Content of Incoming Data During Field Monitoring: Application to Chiller Modeling (RP-1139)

Monitoring and verification (M&V) of energy savings in installed systems as well as non-intrusive automated model-based fault detection and diagnosis (FDD) of HVAC&R systems both rely on the ability to identify model parameter estimates from monitored field data using either off-line or on-line techniques. Most HVAC&R systems operate in a fairly repetitive manner from day to day, while being subject to a long-term annual variation due to climatic changes and the manner in which the systems are designed to respond to them. Consequently, the benefit of collecting field data in order to identify a performance model of the system or equipment is affected by both of the two above considerations, with the benefit generally decreasing over time. We demonstrate, using monitored field data from one chiller, that the presence of a strong temporal correlation in the incoming data, along with ill-conditioning of the regressor matrix of a specific chiller model, can result in a sequence of about 300 field-monitored data points essentially having the same information content as 20 “independent” data points. We provide a brief discussion of notions relating to “information content” of data in various disciplines and how to evaluate whether a new datum is providing additional information to an already available data set. Subsequently, we select and discuss two mathematical definitions of information content relevant to our specific application and apply these to two field-operated chiller data sets differing both in time scale of data collection (15-minute and 1-hour) as well as duration (14 days and 5 months) in the framework of three different linear chiller models. We specifically discuss: (a) how they can provide insight into the initial length of data set needed to initiate adaptive on-line model training, (b) how they can be used to determine when the monitored data do not provide any new information likely to modify the parameter estimates of the linear models, and (c) how this initial data length depends on the model used. Though the results of the analyses presented in this paper are specific to the data sets used, the underlying notions and concepts can be of practical relevance as to how experimental measurements can be digested toward ascertaining statistically meaningful information. This is important in view of the increasing number of non-intrusive field monitoring projects currently being performed by the HVAC&R community in the context of M&V and FDD.