Characterisation of residential energy use for heating using smart meter data

Keeping climate change actions in mind, the reduction of energy use in buildings is a matter of growing interest. The implementation of the Energy Performance of Buildings Directive leads to an increasing energy-efficiency in both new and refurbished buildings, at least in theory. While calculations of the building energy use in the design stage are useful to estimate the building energy performance in reference conditions, these theoretically estimated energy use figures often differ significantly from the actual measured energy use of the occupied building. In reality, construction deficiencies may occur, the actual operation and control of the building services may differ, and the building users interact with the building envelope and services according to their personal comfort needs and activities, which in turn may also be driven by energy-efficiency objectives. In order to estimate the actual and case-specific impact of energy-efficiency measures, there is an urgent need to better understand the actual energy use in occupied buildings. The application of energy monitoring and smart metering systems leads to an increasing availability of frequent, hourly or daily, and long-term measurements of the actual energy use in occupied buildings whereas in the past only monthly or yearly meter readings were collected. These energy use time series are the outcome of the system consisting of building, building services and occupants (the 'occupied building'), and the interaction of this system with ambient conditions such as the weather, which influences especially the energy use for space heating, but also the calendar, which may influence user behaviour. In order to characterise the energy use of the system starting from the measured data, the challenge is to describe it mathematically in function of these ambient conditions. The resulting set of parameters that characterise the energy use of the system is called the Energy Signature of the system. It is the basis for comparing the energy use of an occupied building in different periods, for comparing different occupied buildings, for separating weather-dependent and weather-independent types of energy use, or for estimating the energy use for a given set of ambient conditions. When occasional energy use measurements and weather data are available, widely applied data-driven models are the Heating Degree Day and Energy Signature methods, using basic linear regression techniques. However with the availability of energy monitoring data from smart meters, a first question is whether these classical methods are still applicable when daily or sub-daily data are used. A second question is if these methods can be improved by elaborating on the model structure or the data processing. Eventually these high temporal energy use time series often display energy use time patterns that reflect the effects of behavioural patterns, building and services characteristics in more detail than weekly or monthly measurements would do. Therefore a third question is whether the set of Energy Signature parameters can be extended towards the inclusion of these energy use time patterns. The primary application envisaged for the models is household-specific energy use feedback. In this dissertation, the characterisation of residential Energy Use Signatures is investigated applying regression models and daily or sub-daily energy use smart meter data. The energy use time series in focus are hourly gas meter readings from residential buildings in Belgium, where gas is used mainly for space heating purposes and often also for domestic hot water generation and/or cooking. The main part of the study elaborates on data from 25 dwellings where gas is used only for space heating. First, the applicability of classical Energy Signature methods on daily averaged energy use time series is investigated in terms of predictor variables, model validity, accuracy and training period length. It is found that for linear regression models, the statistical assumptions are not satisfied in about half of the cases, where especially the assumption of the independency of the error is not fulfilled because of serial correlations. A logical step to improve the model with regard to regularly recurrent serial correlations, is by adding auto-regressive terms to the linear regression model, which is then called an auto-regressive model with exogenous inputs. This results in an overall model validity for the majority of the cases. Using valid linear regression or ARX-models, the Energy Signatures for two consecutive years are compared for the 25 cases. Using an example case, the application of linear regression and ARX-models is also illustrated on 2-hourly averaged energy use time series. While linear regression models as such are not valid (due to non-linearity, dependency of errors in time), ARX models deal with the regularly recurring energy use time patterns in their auto-regressive terms. Some energy use time patterns however do not possess a regular (e.g. weekly) recurrence. Therefore, the third part of the dissertation investigates a different approach for dealing with energy use time patterns in the sub-daily series in a structural way. Using time series decomposition methods, the energy use time series is decomposed into a component that deals with the sub-daily energy use time patterns, a trend component that deals with the seasonal variations, and a residual component. The de-trended data are split into day-length time series that are subjected to cluster analysis. The resulting groups of energy use time patterns with similar shapes are described in function of weather and calendar variables using logistic regression methods. The results are case-specific energy use time patterns for different weather conditions and different days of the week, that indicate start and end of the heating season, stand-by of the system etc. The information from these sub-daily energy use time patterns is entered into the linear regression trend models that use daily average data, by means of adding a categorical predictor indicating the groups of energy use time patterns. It is found that the resulting regression models deal better with both irregular occurring and some of the regular occurring serial correlations. The Energy Signatures are expanded with energy use time patterns and again they are compared for two consecutive years. Finally, in a last part of the study, the data-set is extended with 2 cases where gas is used for both space heating and domestic hot water demand and the decomposition-based methodology is now applied, thus extending the application domain and providing insight in the weather-independent energy use.

http://ift.tt/2tP2x4u