Regional Energy Simulation Methods Article Swipe

In the United States, buildings account for approximately 40% of primary energy use and 75% of electrical energy use on an annual basis (U.S. Department of Energy 2017), making them an important target for energy reduction. However, it is believed that the rate at which building energy is being reduced could be improved. The energy efficiency gap is a hypothetical phenomenon which identifies that the reduction of building energy consumption may be occurring more slowly than possible due to a number of market barriers (Jaffe and Stavins 1994). While a number of top-down or disaggregative approaches and tools, such as Scout (Harris, et al. 2016), have been developed in an attempt to promote energy conservation measure (ECM) adoption in buildings and reduce the energy efficiency gap, it is posed that bottom-up approaches may enable more rapid change (Koopmans and Willem te Velde 2001). However, bottom-up approaches rely heavily upon the use of virtual building stocks: databases or information sets representing a set of actual buildings or variations of actual buildings in order to investigate topics such as energy consumption, waste production, or building energy demand (Kohler and Hassler 2002). Unfortunately, virtual building stocks are difficult to develop due to a lack of openly-available, comprehensive data. Furthermore, virtual building stocks are considered computationally costly and time consuming to execute since building performance simulation (BPS) is required to estimate the complex interactions associated with ECM implementation and the heat replacement effect (Wittchen and Aggerholm 2000) on these virtual building stocks. To overcome the lack of characterization data and computational cost, a number of shortcuts have been developed: the use of representative energy models (REMs), reduced order methods (RMs), and heuristic ECM search methods. As opposed to a single building energy model (SBEM) which is a single model representing a single building, a representative energy model (REM) is a single model representing two or more buildings, reducing relative computational costs but potentially incorporating prediction error. RMs such as log-addition (Surana, et al. 2012) are more rapid surrogates for BPS, but they have not been proven to reliably and accurately predict energy savings relative to BPS. While heuristic ECM search methods can be intelligent and efficient replacements for exhaustive enumeration, ensuring they identify global optimal solution sets can be challenging (De Jong 2007). It is unclear from literature how these shortcuts contribute to the market barrier of imperfect information, and what level of uncertainty they generate when used in bottomup approaches. This dissertation reviewed existing surrogates for BPS and identified a new reduced order method for predicting the energy savings of combinations of ECMs. Termed the log-additive decomposition (LAD) method, it was benchmarked against existing reduced order methods proposed in literature in a testbed representing medium offices in the Greater Philadelphia Region. The testbed utilized two REMs based on CoStar data (CoStar Realty 2012), seventy ECMs, and was built around jEPlus (Y. Zhang 2012). Conducting an exhaustive enumeration via BPS for all possible ECM combinations, which consists of 57,286,656 possible combinations, would require multiple years to simulate. In place of conducting an exhaustive enumeration, LAD was compared to BPS for all possible two-way ECM interactions and a more manageable 30,000, randomly selected ECM combinations. Results indicated that LAD requires 0.018% of the computational cost of a typical BPS (EnergyPlus) but has average prediction error in the range of 10%, a large improvement over other methods which have average errors in the range of 20% - 50%. Utilizing the same testbed to refine and hone the LAD method, LAD was used in a hybrid ECM search method and compared to an ad-hoc ECM search method (commonly used in practice), based on professional judgment (Wen, et al. 2013), and a heuristic ECM search method (commonly used in research), using jEPlus+EA (Y. Zhang 2018). Since exhaustive enumeration was not possible, the ECM search methods were compared in a permutation space where the global optimal solution was unknown. Using both identified performance assessment metrics (Knowles, Thiele and Zitzler 2006) and new metrics, results indicated that the ad-hoc ECM search method results in highly sub-optimal results. Alternately, the LAD and heuristic ECM search methods performed well with similar results. Overall, a hybrid ECM search method using LAD performed well, providing consistent results, and overcoming the knowledge required with applying and tuning optimization parameters (De Jong 2007). Following refinement and verification of the LAD method, this method was implemented in a new testbed to quantify the impact of using REMs, instead of SBEMs, on projected regional retrofit costs, projected regional energy savings, and differences in ECMs identified as cost-effective for the geography under evaluation. Using combined geospatial data (PASDA 2017) and ENERGY STAR(r) data (U.S. EPA and U.S. DOE 2018), eighty-five SBEMs and two REMs identified based on TwoStep clustering (IBM 2012) were generated and simulated in CityBES (Chen, Hong and Piette 2017) in conjunction with a library of thirty ECMs. Simulation of the cost-effective ECM curves identified for the two REMs with LAD indicated LAD prediction error was on average less than 1%, with a maximum error less than 4%. Statistical comparison of the cost-effective curve shapes generated using LAD identified that REMs may not serve as good approximations for their SBEM constituents. Furthermore, the ECMs identified and ECMs prevalent along the REM costeffective curves were also found to differ statistically from their SBEM constituents. However, it was unclear if these statistical differences in cost-effective curves would result in quantifiable differences in regional retrofit costs and regional energy savings when the virtual building stocks were used to estimate ECM package adoption across the region. To understand if the differences between the REM and SBEM-based virtual building stocks resulted in quantifiable differences for regional retrofit costs and regional retrofit savings, adoption models that were based on simple payback period (SPP) and gathered from stakeholder surveys (Hamilton 2013) were utilized to simulate stakeholder adoption of ECM packages in the region. A Monte Carlo process was used to estimate ECM package adoption for the virtual building stocks and was repeated 10,000 times to develop estimated distributions of aggregate retrofit cost, aggregate energy cost savings, and resulting SPP for the REM-based and SBEM-based virtual building stocks. Overall, results indicate that a REM-based approach may not adequately reflect actual building stakeholder decisions. In this dissertation specifically, a REM-based approach significantly underpredicts not only aggregate retrofit costs, but more importantly, aggregate retrofit savings. From this thesis, the LAD method has been benchmarked as an alternative to BPS to estimate the energy savings of combinations of ECMs. When used to support ECM searches and optimization, LAD could be considered more rapid and exhaustive than other existing reduced order methods but may introduce prediction error on the order of 10% relative to BPS. When used to support the comparison of an REM-based virtual building stock against an SBEM-based virtual building stock, results indicate that the common shortcut of using REMs is that it may not adequately represent decisionmaking and ECM adoption at the individual building level. As a result, using REMs together with ad-hoc or heuristic ECM search methods to inform policies and guidelines may be resulting in imperfect information that does not reflect actual ECM attractiveness and adoption at the individual building level. However, it is important to acknowledge that the results presented in this dissertation are just the beginning of an exploration. Additional investigation is needed to validate the methods in this dissertation and their findings. While the LAD method has enabled the generation of these preliminary results, future work needs to assess how virtual building stock development can be improved in all stages, including: documenting better building characterization data, improving and testing clustering strategies, and quantifying how results and conclusions change with improved energy model generation procedures. In the end, results indicate that current practices due confirm the notion that an energy efficiency gap exists, but a continuation to refine and implement the methods presented in this thesis could reduce the size of such a gap.