In recent years energy planners consider district heating to be the most sustainable option for the future of urban heating. The most important characteristic of DH is the use of an energy source that provides a significant cost differential in generating heat compared with conventional heating using boilers or direct electric heating. Mainly due to the flexibility of the primary energy sources that could be fully reliable on renewables, district heating has a significant potential to be one of the most promising heat supply options in the future.
At the same time, the growing trend of energy consumption in the building sector has been observed for the last few decades. The building sector in the EU consumes approximately 40% overall final energy, while heating and hot water preparation alone account for almost 80% of total energy use. The EU has set a common legal framework in order to ensure that the Union’s 2020 headline targets on energy efficiency of 20% and its 2030 headline targets on energy efficiency of at least 32.5 % are met and paves the way for further energy efficiency improvements beyond those dates.
According to the Energy Efficiency Directive, Member States shall ensure that, for district heating, district cooling, and domestic hot water, final customers are provided with competitively priced meters that accurately reflect their actual energy consumption. Measuring heat delivered to each apartment can be easily performed with heat cost allocators (HCA). Introduction of HCA is a technique which can be used to ensure that the consumer pays only for the heat supplied to an individual apartment. Although numerous studies and analyses have confirmed that by installing HCAs at the building level annual heat savings can be expected from 15% to 30%, in the case of Croatia these goals were not reached as expected. This indicated that there might be a significant impact of behavioural and other non-technical factors on the energy consumption in district heating. Recently, a growing research focus has been placed on the non-technical factors affecting energy consumption.
The research goal of this thesis is to evaluate the impact of HCA instalment in district heating systems on energy savings, develop energy prediction models for the purpose of prediction of energy consumption and evaluation of impact of energy efficiency measures on apartments’, buildings’, district heating systems’ and nationwide level with machine learning algorithms. Models developed by machine learning methods thus have a dual function, they are predictive
and serve to predict outcomes based on past data, but they are also descriptive because they serve to acquire new knowledge about the process of household heat consumption.
Although it is custom to use numerous simulation tools developed for calculating building energy loads these simulation tools are often costly, require a great deal of prior knowledge for efficient use and lack in terms of prediction accuracy of actual energy consumed. Several previous researches indicated that machine learning methods can mitigate these obstacles and develop models with high accuracy of prediction when considering energy consumption. The research results showed that machine algorithms are superior in building energy calculations in comparison to conventional statistical and engineering methods (simulation tools), while there are ready to use software and programming languages for implementation like Matlab, Python or R. Analysis made in this research are developed using R programming language.
Within this research chosen machine learning algorithms were implemented on a data set consisting of 3,854,000 observations of 50 variables of actual billing data in the period of 7 years for two district heating systems in Croatia. Additionally, a on-field research of behavioural aspects of final consumers in district heating was undertaken on a set of three multiapartment buildings in Zagreb via questionnaires and interviews.
The models were developed by data pre-processing, learning on the training data subset, and accessing model accuracy on the verification set, following by the interpretation of the influence of individual parameters on the heat consumption for each model. Given the type of data available in this analysis models were developed using these machine learning methods: regression analysis, regression trees, random forest, and support vector machine (support vector machines, SVM). The results obtained were interpreted in such a way that the prediction accuracy of each method was evaluated, and a recommendation was made as to which model is the most suitable for the prediction of consumption in district heating systems.
The development of high accuracy models for the prediction of heat consumption will also stimulate an increase in the integration of renewable energy sources (RES) for the production of thermal energy in district heating systems, thus enabling the integration of such a developed model with existing models for the prediction of energy consumption. This primarily refers to the conversion of electricity from renewable sources to thermal energy during periods of high availability of intermittent RES such as solar or wind, which would contribute to greater integration of RES into the energy system
Objective and hypothesis of the research - The objective of the research is to develop machine learning models of heat consumption in district heating systems, for the purpose of energy planning and evaluating the impact of energy efficiency measures. From this goal, the following research hypotheses emerge:
1. Available data are sufficiently representative, and by appropriate pre-processing, using descriptive multivariate statistical analysis and exploratory analysis, it is possible to obtain a more detailed insight into the character of the data, their representativeness and their interrelation and the appropriate assessment of suitability for incorporation into the predictive model.
2. Based on previous analyses, it is possible to build a machine learning based models that will predict energy consumption more accurately than current models.
Scientific contribution - The scientific contribution of the research is the development of models for predicting energy consumption and assessing the impact of energy efficiency measures in district heating systems based on machine learning methods. The model has been developed at several levels, sequentially (i) at the level of the individual apartment in the district heating system, (ii) at the level of the individual building connected to the district heating system, (iii) at the level of each distribution area, and (iv) at country level.
Also, evaluation of each machine learning algorithms considered for this application is given, namely multiple linear regression, regression trees, random forests, and support vector machine as supervised methods, and grouping as an unsupervised method.
Additionally, as a result of the undertaken machine learning analyses, the level of influence of each factor of the model will is determined and its influence on the heat consumption in district heating systems is interpreted.
Conclusion - The developed regression models at the apartment level show that apartments with built-in HCAs can expect 40% less energy consumption than apartments that did not install them, in
case they do not account more than 4% of all impulses in the building on an annual basis. An additional interpretability is obtained by the random forest model, which shows that the lowest average annual consumption is in those apartments that have installed HCAs and which account for less than 3.6% of all impulses in the building annually. There is a total of 10% of such apartments, and their specific consumption ranges from 35 to 88 kWh/m2. The highest consumption in the group of apartments that have installed HCA are those that count more than 8.1% of all HCA impulses on an annual basis and their average consumption is around 125 kWh/m2. Such apartments make 3% of all analysed apartments. At the same time, the specific consumption of those apartments which did not install HCAs, and which are in the buildings where the installation was partly carried out, ranges from 97 to 176 kWh/m2. Therefore, the existing allocation method allows some of the apartments in the same building that did not install HCA to be allocated less heat than the apartments that have installed them.
The reason for the large number of HCAs' impulses in some apartments may be due to behavioural characteristics of the consumers or in the technical defects of the heating system in a building or in an apartment, which should be separately identified for each such case. As part of this research, surveys with questionnaires and interviews were undertaken to determine the impact of behavioural parameters. The analysis of data obtained from the questionnaire was performed by the clustering method, an unsupervised machine learning method. The results indicated that there is an influence of behavioural parameters on specific heat consumption, especially the parameters such as the desired level of thermal comfort, daily occupation time and number of unheated rooms.
At the building level, all analysed buildings achieved savings after HCAs were installed. The resulting regression model gives an indication that 100% installation of HCAs in the building will result with an absolute specific savings of 46 kWh/m2 annually on the building level. The model of regression trees gives additional interpretability on the building level, indicating that the lowest consumption (93 kWh/m2) is in those buildings that have a rate of installation of HCAs higher than 62 % of heated area. The prediction accuracy of each of the 4 models at the apartment and building levels is high compared to conventional simulation methods. The highest accuracy is obtained by the random forest method (apartments +/- 4.27 kWh/m2, buildings 13.24 kWh/m2) and the lowest by multiple linear regression (apartments +/- 16.44 kWh/m2, buildings 16.59 kWh/m2).
When selecting a model, a trade-off between level of interpretation and accuracy should be found. If the goal of modelling and forecasting is to obtain as accurate model as possible with a certain level of interpretability, then it is suggested to use a regression tree model. On the other hand, if our sole purpose of forecasting is to obtain highest accuracy, as it would be in the case of evaluating the impact of energy efficiency measure of HCA instalment, it is suggested to use a random forest as a machine learning algorithm.
Considering the degree of influence of individual variables, it is concluded that in the apartments the variables related to the metering of energy consumption (ratio of counted impulses in the apartment and counted impulses in the building, and metered heat on the building level) are dominant along with variables that describe whether the apartment has installed HCA and how high is the rate of installation in a building. If the target is to reduce the average energy in all apartments, these variables should be affected. In order to maximise the energy savings it is recommended to install HCA in all apartments (measures related to the implementation of the secondary energy efficiency measure), to implement all cost-effective measures on the building envelope (measures related to the implementation of primary energy efficiency measures), and to implement end-user training measures.
The positive effects of HCA installation are inevitable at the building level, DH systems and broader national contexts, but at the apartment level, large scattering and appereance of outliers occur mostly due to the behavioural parameters identified in this research. The models developed in this research can serve as a basis for forecasting consumption in district heating systems with high-accuracy and can be used as a tool to evaluate the performance of some of the energy efficiency measures.