Evaluating different statistical regression models for industrial energy measurement and verification

Abstract

The worldwide concern regarding the environmental impact of using fossil fuels as an energy source leads to the increased interest in improving energy efficiency. The energy efficiency of existing facilities can be improved through the implementation of energy saving measures (ESMs). However, to convince different stakeholders involved of the impact of ESMs, the savings need to be quantified accurately. This falls under the field of measurement and verification (M&V). An important part of the M&V process is the construction of a model to predict energy usage. Different model types are available that fall within different categories. Each of these categories provides different advantages and disadvantages. The black-box based approach of using statistical regression models (SRMs), can offer superior accuracy and ease of use. Also, combining different SRMs through model-averaging can provide even greater accuracy than a single model. Many studies have focused on evaluating different types of SRMs for energy modelling purposes. These include simple models, such as the well-known linear model, but also the more complex machine learning models. Some studies have also focused on specifically applying different types of SRMs towards M&V. However, studies tend not to give holistic consideration to the fundamental principles when evaluating models for M&V. Furthermore, there has been minimal evaluation of different types of SRMs for use of M&V within the industrial sector, with only a few isolated case studies reported in literature. There is thus a need for a comprehensive evaluation of different types of SRMs for M&V use within the industrial sector. This dissertation evaluated four different types SRMs for accuracy and consistency in the M&V process, by applying it to five industrial case studies. The models included a simple linear model and three types of machine learning models. Two model-averaging techniques (simple and weighted) were applied to the four single models to form two new models. A comprehensive evaluation procedure was developed based on well-known methods and techniques in literature. Different aspects of accuracy and consistency were considered in the procedure, including absolute error, bias, sensitivity to the data sampling process and sensitivity to drivers used. The procedure also accounted for additional complexities introduced by the machine learning models. A scoring technique was applied to evaluate the performance of model types across the five case studies. The results obtained are verified based on what is expected from literature. The study shows that no specific model type stands out with differences changing from case study to case study. Savings and errors of the different model types tend to fall within the same range. The specific energy drivers used, however, tend to influence the predicted savings and accuracy more than the type of model. This indicates that great care should be taken when considering energy drivers while the different model types can thus be used as verification and to identify anomalies.