Following the connection from electricity use to the people who bear its costs
An air conditioner sits inside a home. Some of the consequences of powering it occur somewhere else.
The person using it experiences the relief of a cooler room and pays the electricity bill. But the bill does not tell the whole story. Beyond the outlet lies a system of power plants, transmission lines, and decisions about how electricity is supplied. Emissions from that system can affect people who may never see the appliance or share in its benefits.
During my graduate research at Lawrence Berkeley National Laboratory, I worked on a project exploring that connection: how could changes in appliance-efficiency standards be linked to regional air quality and estimated health impacts?
Room air conditioners provided the test case. The broader ambition was to develop a modeling process that could eventually be applied to other appliances.
My part of the work involved preparing and analyzing appliance-use data and translating efficiency scenarios into electricity-demand inputs for downstream models. I also wrote the capstone documenting the project. Much of the challenge lay in the space between those tasks: deciding how to move from one source of information to another without carrying assumptions further than the evidence allowed.
An annual total leaves out part of the story
An estimate of annual electricity savings answers an important question: how much less electricity might appliances use under a different efficiency standard?
For this project, we also needed to ask when that difference would occur.
Knowing annual electricity consumption is a little like knowing how many vehicles traveled through a city over a year. The total tells you something about activity, but it cannot describe the morning commute. Timing changes what the system must do to accommodate demand.
Electricity has its own patterns. An air conditioner’s use varies with conditions and household behavior. Changes in demand at different hours can interact with different patterns of generation. To connect appliance use with emissions, we needed to preserve more of that timing than an annual total could provide.
The workflow therefore required hourly load profiles: a way of distributing electricity consumption across the hours of a year.
It also required a clear comparison. We were examining the difference between a baseline without new standards and alternative efficiency levels. The relevant saving was that difference—not the appliance’s entire electricity consumption.
Those distinctions sound straightforward when laid out in prose. Making them work together required aligning datasets that described different things: appliance-use patterns, efficiency levels, shipments, and regional appliance stocks.
The data needed more than cleaning
The work drew on residential appliance-metering data collected in Massachusetts and Pennsylvania, alongside modeled appliance-load profiles from the National Renewable Energy Laboratory’s ResStock resources.
While cleaning the metering data, I identified missing observations and brought the issue to colleagues. I presented what I had found, discussed possible approaches with the team, and implemented a multiple linear regression approach to address the gaps.
That sequence matters to how I understand the work. Identifying the problem was the beginning. I needed to make it visible to others, discuss an analytical response, and carry the agreed approach into the data preparation.
An estimated replacement for a missing observation remains an estimate. It can help make analysis possible, but it does not acquire the same status as a reading collected by a meter. The assumptions behind that estimation remain part of the analysis.
There was another problem that filling the gaps could not solve.
The observations came from particular households in particular places. The intended application was broader, including the possibility of using the data to characterize national air-conditioner use. That required us to examine how much weight those local observations could carry.
A more complete photograph of one street is still a photograph of one street. Sharpening the image does not make the rest of the country appear inside the frame.
Missing observations concern what we do not know within a dataset. Representativeness concerns what that dataset can tell us beyond the people and places it captures. Treating them as the same problem risks creating confidence where we have only created completeness.
I worked to narrow the interpretation of the metering findings, making clear that the observed behavior was local, granular, and shaped by geographic and behavioral conditions. Those corrections happened during preparation, before the modeled outputs. This was analytical work to improve the basis of the modeling, rather than a result that overturned a formal hypothesis.
That distinction also informed our subsequent choice of NREL data for benchmarking. Another source could provide a useful comparison, while leaving the limitations of the original observations in view.
Turning appliance scenarios into model inputs
The next challenge was to connect the appliance analysis with the models used by other researchers.
For the Pennsylvania test case, the workflow used appliance shipments and stock information to estimate changes in electricity consumption under alternative efficiency levels. Hourly load profiles then provided a way to distribute those changes across the year.
My contribution was in this preparation and translation: working with the appliance-use data and helping turn the efficiency scenarios into the demand inputs required by the downstream workflow.
This kind of work can disappear behind the final output. A table of estimated emissions or health impacts is easier to recognize as a research result than the work needed to prepare its inputs. Yet every stage depends on those inputs meaning what the next model expects them to mean.
Passing data between models is like passing a measurement between people using different instruments. A number can travel easily. Its units, geographic scope, time period, and assumptions need to travel with it.
The prepared inputs were handed to colleagues at the Laboratory who operated the grid and air-quality models. Those stages connected changes in electricity demand with modeled generation and emissions, and then with estimated health impacts.
The project and the ideas behind integrating the models were collaborative work led by Anna Spurlock. I authored the capstone under Sydny Fujita’s supervision. My responsibility was a substantive part of that workflow; it did not extend to operating every downstream model or independently validating the entire system.
What the initial outputs established—and what they did not
As the capstone was being finalized in spring 2024, colleagues ran the prepared inputs through the downstream workflow. The report describes initial outputs covering power-plant emissions and estimated premature mortality under different grid scenarios, along with monetary valuation of health impacts.
Those initial runs indicated that the prepared inputs could be carried through the modeling chain to produce the types of outputs the project sought.
That was a meaningful step. It connected an appliance-policy scenario with estimates of consequences beyond household electricity use. But a workflow producing results is not, by itself, proof that every estimate is accurate.
Nor are modeled health impacts observations of what happened to people after a standard changed. They are estimates conditional on the data, scenarios, and relationships used in the analysis.
The distinction matters because the language of health outcomes carries weight. A modeled change in premature mortality should not become a claim that the project observed deaths prevented. Similarly, the possibility of examining differences across communities should not be presented as an analysis already completed.
The capstone identified further work that could refine the approach and explore how impacts differ across populations. The initial work provided a basis for asking those questions more precisely.
Following the costs beyond the electricity bill
What stayed with me was how much of the project depended on deciding what the evidence could responsibly carry.
I found missing observations, discussed and implemented an approach to address them, and worked to keep local behavior from being treated as automatically representative of a larger population. I contributed to preparing the connection between appliance-efficiency scenarios and downstream modeling, then documented the broader project in my capstone.
These are connected responsibilities, but each requires a different judgment. A statistical method can address a gap without resolving a question of geographic applicability. An hourly profile can add detail without removing uncertainty. A downstream result can make a consequence visible without becoming an observed fact.
For me, that is part of the substance of applied research: making an analysis useful while keeping its boundaries legible.
It also brings me back to the question that drew me toward this work. An electricity bill records the household’s payment. It does not show everyone whose life may be affected by supplying that electricity.
Following those consequences through a modeling process gives us a fuller set of questions to bring to policy: how much energy might a standard save, when and where might those savings occur, and whose exposure could change?
I remain interested in the space between technical analysis and public decision-making because those connections are easy to lose. The more distant a consequence is from the original decision, the easier it becomes to leave it outside the calculation.
Good analysis can help bring it back into view. It must also show the reader how it got there.
View the LBNL project case study · Read the full research article · More research and writing
