As Philadelphia strives to meet the greenhouse gas emissions targets set in its 2050 Plan, it is better to see how zoning can play a role in managing energy use in buildings. Understanding can make a city successful. Researchers at Drexel University’s School of Engineering hope that the machine learning model they developed can help predict how energy consumption will change as neighborhoods evolve, supporting these efforts.
In 2017, the city Set a goal to become carbon neutral by 2050, mainly due to the reduction of greenhouse gas emissions from building energy use. This accounted for almost three-quarters of Philadelphia’s carbon footprint at the time. But the key to meeting this standard lies not only in establishing sustainable energy use practices in current buildings, but also in incorporating energy use projections into zoning decisions that will direct future development.
Another challenge for Philadelphia, one of the nation’s oldest cities, is that the building types vary widely, with different energy usage. Planning for more efficient energy use at the city level is therefore not a problem with one-size-fits-all solutions.
“Especially in Philadelphia, where the prevalence of particular housing features and types of zoning varies widely from place to place, rather than trying to enact comprehensive policies for carbon reduction across cities and counties, , it is important to customize the energy program for each region.,” said Dr. Shmi Hawk,Professor Faculty of Engineering Led research on using machine learning to model granular energy use recently published in the journal energy and buildings.
Hoque’s team believes that existing machine learning programs, if deployed properly, will provide some clarity about how zoning decisions will affect future greenhouse gas emissions from buildings.
“We currently have an enormous amount of energy use data, but it is often inconsistent and messy to use rationally. One that corresponds to a particular housing characteristic may have usable energy estimates, but another dataset that corresponds to socioeconomic characteristics has too many values to use,” she said. rice field. “Machine learning, despite these data limitations, is well equipped to handle this challenge because it can iteratively learn and improve through the training process, reducing bias and variance. .”
To glean information from disjointed data, the team developed a process using two machine learning programs. One can extract patterns from large data tranches and use them to make predictions about future energy, and the other can identify model details. This may have had the greatest impact on forecast changes.
First, a deep learning program called Extreme Gradient Boosting (XGBoost) was run using a large set of Philadelphia commercial and residential energy usage data from the 2015 US Energy Information Residential Energy Consumption Survey and the Commercial Buildings Energy Consumption Survey. have trained City demographic and socioeconomic data from the U.S. Census Bureau’s American Community Survey for the period.
The program learned enough from the data to determine building density, population in a given area, building square footage, number of occupants, days with heating and air conditioning, and energy use for each house or building.
Deep learning models like XGBoost are very useful for making informed predictions, but given a large and inconsistent data set, their methods are limited by the complexity of the operations they perform. It can get confusing. But to be a useful tool to guide planners, the team had to develop enough so-called “black box” programs to turn those predictions into recommendations.
To do so, they employed Shapley’s additive explanatory analysis. This is an evaluation used in game theory to distribute credit among the factors that contributed to the outcome. This allowed us to infer, for example, how changes in building density or area would affect the program’s predictions.
“Machine learning models like XGBoost learn how to process datasets to perform specific tasks, such as generating reliable predictions of systems, but they don’t really understand the field relationships that underlie phenomena. We do not claim to do or represent,” said Hoque. Said. “The Shapley analysis does not tell us which features have the greatest impact on energy use, but it can tell us which features had the greatest impact on the energy use of the model. predict, which is still very useful information. ”
The team then tested the model by providing input data from a hypothetical scenario proposed by the Delaware Valley Regional Planning Commission, which estimated economic development to continue in Philadelphia through 2045. It shows the different employment and income possibilities by household, and by neighborhood across the city.
For each scenario, the model analyzes how new residential and commercial developments change greenhouse gas emissions from building energy use across 11 different parts of the city, and which variables are used to make the projections. expected to play an important role.
Looking specifically at residential energy use in the 2045 scenario, the program suggests energy use decreases in 6 out of 11 neighborhoods, mostly low-income neighborhoods. On the other hand, mixed-income areas, such as the northernmost part of the city, including Oak Lane, could see increased energy use.
According to Shapley’s analysis, the presence of single-family homes (low energy use) and single-family homes (high energy use) played an important role in the forecast. The number of rooms per building all contribute to a reduction in energy usage projections.
“Overall, the residential energy prediction model found that features associated with low building strength were relevant to the model’s energy consumption estimates. such as fewer rooms,” they wrote. “These results provide a rationale for reexamining the impact of upzoning policies, which commonly exist as affordable housing solutions in Philadelphia and other cities across the country, and subsequent changes in energy use in these areas. will give you.”
On the commercial side of the scenario, the machine learning model did not predict significant changes in energy usage under 2045 conditions. Energy use in the largest commercial buildings remained high. It was also limited to examining only six variables (area, number of employees, floors, heating degree-days, cooling degree-days, building main activity), but the data available in the training set allowed Shapley’s The analysis noted that building square footage and number of employees were the most important predictors of energy use for most types of commercial buildings.
“For the commercial sector, the study suggests that commercial buildings in the top quantiles of area and number of employees should be prime targets for energy reduction programs,” the authors write. “The study assumes a threshold of roughly 10,000 square feet of total building area, and buildings above that marker are prioritized because they disproportionately affect the model’s energy projections.”
While researchers are careful to assume direct links between model variables and changes in energy use, they do not provide planners with a high level of detail about the interplay between zoning decisions and development. suggesting that models are still very useful because of their ability to provide a broader perspective. and impact on energy use.
“We see a lot of potential in using machine learning models like XGBoost to predict increases and decreases in energy use due to new construction projects and policy changes,” said Hoque. . “For example, building a new railroad in a neighborhood could change the demographics and employment of the neighborhood, and our method is ideal for incorporating that information into the context of energy forecasting models.”
The team acknowledges that more testing is needed and that the program will improve as additional data is provided. It suggests that it is important to focus on areas of the city where there are large numbers of people and run a Shapely analysis to identify some of the factors that may be responsible for it.
“I hope this provides a resource Future researchers and policy makers don’t have to study the entire city of Philadelphia, but can focus on the areas and variables they flag as potentially important,” Hoque said. increase. “Ideally, future studies would use more interpretable methods to test whether these features actually correspond to higher or lower energy estimates for specific regions. .”
In addition to Hoque, doctoral students at Drexel’s College of Engineering, Shideh Shams Amiri and Maya Mueller, participated in the study.
Read the full paper here: https://www.sciencedirect.com/science/article/pii/S0378778823001950
