By Colm Gallagher*
Most M&V practitioners generally have a go-to set of independent variables to employ when modelling each different energy system. For example, when performing M&V on a chilled water optimisation project in a commercial building, the outside air temperature will most often be sufficient to model the systems energy consumption, as this approach has been applied time and time again with proven effectiveness. The same can be said for residential buildings. In contrast to this are the many industrial buildings that pose quite a substantial problem. The complex energy systems in these buildings are influenced by many different variables. This means it can often be difficult to accurately model the consumption you are interested in using the common independent variables. As a result, more simplistic approaches can hinder the accuracy of a project.
This is an issue that has come under increased focus due to the Energy Efficiency Directive. The Directive sets out a 20% energy efficiency savings target for EU member states to achieve by 2020. If this target is to be achieved, it will require the effective implementation of multiple energy conservation measures (ECMs) across all sectors. Industrial buildings will constitute the highest quantity of savings. Critically, M&V has a role to play in accurately verifying the performance of these ECMs. This has resulted in M&V practices coming under the spotlight and subsequently, the difficulties of M&V in industrial applications have come to the forefront. Couple this with end-users' reluctance to invest in M&V and there exists a challenge. How can M&V be performed using minimal resources while still achieving acceptable accuracy?
At the Intelligent Efficiency Research Group in University College Cork, we believe the answer lies in the extensive quantities of data that already exist in industrial buildings. ISO 50001 has resulted in most certified facilities having a monitoring and targeting (M&T) system in place. In addition to this, the current trend of Industry 4.0 practices are data dependent, meaning we will see more metering installed as we move towards this new level of industrial maturity. The data available to us contains valuable knowledge for the facility it is monitoring. The challenge is to unlock the power of this data without making M&V a resource intensive task. We believe machine learning will play a key role in overcoming this challenge. This offers an exciting opportunity, upon which M&V practitioners can capitalise.
Machine learning, a sub-field of artificial intelligence, includes techniques that are powerful enough to extract valuable knowledge contained within data sets. The application of these techniques in the context of M&V is currently being explored with a view to overcoming the challenges encountered in the transition to M&V 2.0. In recent years, published research in this area is very much focused on the residential and commercial building sectors, while the industrial counterparts lag behind.
In an effort to improve the practices applicable to industrial buildings, we have chosen to expand the typical boundary of analysis and include the large data sets available in the analysis. These data sets contain all data recorded by the M&T system and no data are removed until the statistical significance of each variable is quantified. A significant increase in the quantity of data in the analysis results in typical approaches being inefficient and resource intensive. One such problem is posed in identifying independent variables with which to construct a baseline energy model. To this end, an automated and computationally efficient feature selection algorithm has been developed to ensure hundreds of potential variables are reviewed, allowing only the most beneficial be included in the final analysis. There is no limit set on the allowable number of independent variables, allowing all statistically significant variables be included in the analysis.
In a case study application, over 500 variables from across a facility were input into the algorithm, which resulted in 10 being selected as beneficial for the modelling task. Following the identification of the most significant independent variables, the modelling algorithms commonly applied at present are a significant limiting factor in achieving low uncertainty in savings. Advancing the algorithms used for energy modelling in industrial buildings is the first step in evolving to M&V 2.0 practices in the industrial sector. The residential and commercial building sectors have highlighted the potential benefits of advanced machine learning algorithms and we believe they can be of even more use in complex industrial facilities. Algorithms such as artificial neural networks, support vector machines and k-nearest neighbours can increase the performance of a model compared to that of a ordinary least squares linear regression approach. As always the resources required to carry out M&V must be kept to a minimum; hence, a means of automating the application of these algorithms is required.
In our research, we have assessed the suitability of machine learning to solve the challenges faced in M&V by comparing traditional linear regression modelling approaches that use typically applied independent variables, with these more advanced modelling algorithms applied to all available data. In one case, we found that a 51% reduction in model error was achievable after the optimal algorithm and parameters were identified.
Following promising initial results, our team have developed a methodology which provides a prescriptive approach for applying these machine learning techniques. The methodology is designed to minimise the resources required to carry out the baseline modelling task. This M&V 2.0 approach is technology agnostic and hence, can be implemented in an automated fashion by any practitioner. To date, we have automated this using the open-source programming language R and tested the approach with an application under real-world conditions. This case study quantified the savings resulting from the optimisation of the control logic of a HVAC system with only 8.7% uncertainty associated with the final savings at a confidence interval of 95%.
The overall objective of this research is to develop a cloud-based software solution which implements the methodology supported by machine learning and automates the process. This will be capable of real-time savings verification through efficient deployment of the models constructed. Thus, offering a reliable and accurate M&V 2.0 solution for industrial buildings; something which is lacking at present. As we move beyond our 2020 energy efficiency targets, the focus on accurate M&V is set to continue with a new energy efficiency target for 2030 being decided upon at present. Conservatives in the European Parliament are backing a 30% target, while progressives are in favour of a 40% binding target. With this, M&V will continue to cement its place as a valuable tool in the transition to low carbon economies. M&V 2.0 signals an evolution to practices that are less resource intensive and more accurate. We aim to aid this evolution with our research.
(*) Colm Galagher is a PhD candidate at the Intelligent Efficiency Research Group, University College Cork, National University of Ireland, Ireland. For more information on the work at University College Cork, please visit www.ucc.ie/en/ierg/ or. You may also contact the author at c.v.gallagher@umail.ucc.ie.
