Numerous strategies were developed over the years in order to encourage users to reduce energy consumption and bolster energy efifciency. However, with increasing levels of eficiency achieved by most household appliances, one of the most impactful approaches that remains as a means to further increase energy eficiency is attempting to encourage users to behave in an energy eficient manner. More precisely, positive behavior change can be motivated through the creation of unique social pressure and competition. Namely, the idea of the methodology presented in this paper is providing a fair, normalized, comparable ranking (benchmark) between diferent energy consumptions of diferent users. Therefore, the ranking is supposed to motivate them to either retain a leading position in the ranking or to attempt to improve their behavior and advance within the ranking.
The energy-use performance benchmarking and user behavior assessment
methodologies appear to be a relatively unexplored topic in the relevant literature of this
domain, especially when compared with other energy related topics like demand
side management or demand response optimizations. Review papers [
With regards to specific methodologies, [
Given findings in the aforementioned analysis regarding relevant papers that are dedicated to the subject of energy-use eficiency benchmarking, it can be deduced that the current state-of-the-art solutions appear to be ill-equipped to deal with an IoT future in which individual homes will be outfitted with a vast number of sensors. Therefore, this paper aims to introduce a flexible methodology that can be utilized for smart homes and that incorporates various factors pertaining to the energy consumption of the analyzed households. Furthermore, the same methodology could even be extended towards commercial objects that are suficiently covered with smart sensors. 2
With the main goal of the presented methodology being the holistic and comprehensive assessment of user behavior through multiple energy usage indicators, the user benchmarking methodology is based on four diferent elements denoted ri where each one of them depicts a diferent aspect of energy eficiency (normalized comparison with others, normalized comparison to oneself, alignment with intermittent renewable generation and engagement), as will be described in greater detail in the following sections. With respect to the weights wi of each of these criteria, the final unscaled score (rating) can be obtained as a linear combination of these factors
Runscaled =
4 X wkrk = w1r1 + w2r2 + w3r3 + w4r4.
k=1 However, these raw results are further processed before being presented to end users. Namely, linear scaling is applied to convert the obtained interval of values into the range of [40, 95]%, as specified in accordance with expert inputs, so that less eficient users are not demotivated and so that the most eficient users get the impression that there is still room for improvement. 2.1
According to [15], DEA represents a quantitative, nonparametric technique which is used in operational research and most commonly economics to establish a best practice group of decision making units (DMUs) (the so-called eficiency frontier) and to determine which units are less eficient when compared to the best practice groups and at what the magnitude of ineficiencies are.
In consistence with the related literature, let the following symbols be – j the order number of DMU, – i the order number of input used by DMUs, – r the order number of output used by DMUs, – θ the eficiency/ineficiency rating, – yrj the amount of output r by DMU j, – xij the amount of input i by DMU j, – ur the weight coeficient assigned to r-th output, – vr the weight coeficient assigned to i-th input.
Now, the DEA problem is posed as determining the maximum objective function as defined by θ j = max u1y1j + u2y2j + · · · + usyrj v1x1j + v2x2j + · · · + vmymj = max
Ps
r=1 uryrj Pim=1 vixij where s is the total number of outputs and m is the total number of inputs. The maximization is obtained under a set of constraints (∀j) u1y1j + u2y2j + · · · + usyrj v1x1j + v2x2j + · · · + vmymj =
Ps
r=1 uryrj Pim=1 vixij ≤ 1 and
u1, u2, . . . , us > 0 and v1, v2, . . . , vm ≥ 0.
However, DEA is most often implemented using linear programming, which cannot be performed with the given set of constraints and the objective function because the division between the numerator and denominator presents a non-linear operation. This issue is circumvented by modifying the given set of formulas through an additional constraint that specifies that all denominators must be equal to one. In this modified form, DEA is formulated as maximizing the objective function specified by subject to θ j = max ( s
X uryrj r=1
) (∀j)
s X uryrj − r=1 m X vixij ≤ 0 i=1 ∧ m X vixij = 1 i=1 ! while the constraints
u1, u2, . . . , us > 0 and v1, v2, . . . , vm ≥ 0 still apply.
In general, DEA is capable of considering a wide variety of diferent input parameters. For energy eficiency applications specifically, these parameters can be grouped in two diferent categories – Static parameters: • heated area, • heated volume, • outward wall (and window) area, • wall thickness, • wall material (conductivity), • number of reported tenants; – Dynamic parameters: • total energy consumed, • avg. occupancy for the household/building, • cooling/heating degree days, • dif. between indoor and outdoor temperature.
However, having in mind that for the specific use cases that will be demonstrated in the following text, buildings from the same neighborhood were considered, with all of them sharing the same construction properties and microclimate, the number of considered parameters is limited to – total energy consumed, – average total occupancy, – average absolute diference between indoor and outdoor temperature, – total heated area, as including others would add no additional information.
A resulting arrangement of users in this space with the total energy consumed as the primary output is illustrated in Figure 1 where those users on the ineficiency frontier are assigned the rating of 0 and others are given a rating r1 = θ i corresponding to their position between the origin and frontier, as dictated by the DEA approach. 2.2
The main idea of this novel approach was to exploit machine learning through models like random forests, k nearest neighbors, support vector machines, linear regression and neural networks as the estimator of the user’s expected energy usage in accordance with his previous behavior making the most of machine learning’s (ML) extraordinary estimation potential. It is intended to be used in a way which would result in rewarding on penalizing the users depending on their change in behavior. In other words, given a similar set of inputs as the DEA approach uses, it is supposed to approximate what amount of energy a user is expected to consume. Therefore, the estimated value Eˆ can then be compared with measured (real) one Emeas, and reward or penalize the user proportional to the diference that would be assigned to that user, as illustrated in Figure 2. This concept could be considered as an example of the diferential part in control theory, as it measures the diference from the previous behavior and proposes the ”control” accordingly i.e., behavioral stimulus which is in form of positive or negative rating in this particular case. The output of the discussed ML-based part of the benchmarking methodology can therefore be defined as r2 = tansig ln ˆ
E Emeas !! =
2 1 + e− ln(Eˆ/Emeas) − 1.
Namely the idea of using the logarithm function was to obtain negative result when the real measured consumption is greater than the one based on previous behavior, as a negative penalty for ineficient behavior is supposed to be assigned, and vice versa. Additionally, the tansig function has been chosen as it is an odd limited function, so for the same positive and negative behavior the same absolute penalty would be assigned and the output would be within the required limits. Having in mind the increasing penetration of RES installations with individual users, one of the key goals to their eficient usage is maximizing self-consumption, i.e., ensuring that as much of locally produced energy is also consumed locally. Achieving this objective entails that the demand profile should be well aligned with the generation profile meaning that peaks in the demand should follow the peaks for production and vice versa for valleys. However, the stif character of users’ daily habits can notably hinder this process as customs are not so easy to adapt to, for example, the production profile of PV modules which generally displays peak performance during the mid-day period when the sun is shining the brightest.
Therefore, in order to numerically quantify how well-aligned the consumption profile X is to the renewable generation profile Y , their correlation coeficient σ is calculated as r3 = σ {X, Y } =
Cov(X, Y ) pCov(X, X) · Cov(Y, Y ) and used as the third benchmarking contribution r3. 2.4
The final part of the proposed energy eficiency performance evaluation methodology considers user’s responsiveness to notifications and suggestions delivered through a companion mobile application for their smart device management in form of a reward for users that show motivation for behavioral improvements. Namely, this part of the system is supposed to additionally encourage users willing to adapt their demands in order to save energy. For example, if the household/building owner promptly reacts to suggestions about energy conservation, such behavior should be rewarded. Additionally, this factor is not meant for any penalization if the suggestions are not considered because some of the events that are being checked may not imply that energy is being wasted.
The score rewarded in this category is obtained simply by ranking the users by three equally weighted and combined factors that are considered to contribute to the overall responsiveness: the percentage of the energy conservation noticfiations that they have responded to in due time (less than 30 minutes), the total number of controls actions sent using the app (turning appliances on or of) and total number of sessions (discreet log-ins separated by more than 30 minutes).
In summary, this paper provides an outline of a benchmarking methodology for smart homes of the future with a specific goal of further increasing energy eficiency. It takes into account a multitude of diferent factors relating to the collective energy consumption of a community as well as changes in individual behavior. It also considers other factors such as integration with renewable generation and interaction with installed smart devices through a provided platform. Planned future eforts include the evaluation of the methodology on a real set of users and illustrating the link between changes in the ranking and in energy consumption. https://doi.org/10.1016/j.enbuild.2007.07.001, https://www.sciencedirect. com/science/article/pii/S0378778807001971 12. Monts, J.K., Blissett, M.: Assessing energy eficiency and energy conservation potential among commercial buildings: A statistical approach 7(10), 861–869, https://ideas.repec.org/a/eee/energy/v7y1982i10p861-869.html, publisher: Elsevier 13. Nikolaou, T., Kolokotsa, D., Stavrakakis, G.: Review on methodologies for energy benchmarking, rating and classification of buildings 5(1), 53– 70. https://doi.org/10.1080/17512549.2011.582340, https://doi.org/10.1080/ 17512549.2011.582340 14. Sharp, R.: Benchmarking energy use in schools, /paper/Benchmarking-Energy-Use-in-Schools-Terv-Sharp/ a25163f540515d0dc4f7e03c889feff49020d3e3 15. Sherman, H.D., Zhu, J.: Service Productivity Management: Improving Service Performance using Data Envelopment Analysis (DEA). Springer US. https://doi.org/10.1007/0-387-33231-6, https://www.springer.com/gp/book/ 9780387332116 16. Terry Sharp: Energy benchmarking in commercial ofice buildings. In: Proceedings of the ACEEE. vol. 7996, pp. 321–29