Inspiration
The initial inspiration was from living on camps at UCSD, where we observed that UCSD has a real heat problem, specifically certain paths, plazas, and open walkways get dramatically hotter than others. The core question became, why are some spots so much hotter, and what could actually fix it. We combined UCSD's campus heat sensor network with solar infrastructure and shade coverage data, finding that the hottest spots on campus aren't random they're gaps in shade and solar coverage that could be fixed with strategically placed solar canopies. Instead That local insight led us to zoom out and incorporate national data from the EIA, EPA, and ZenPower, revealing that the same pattern of underinvestment exists at scale across the country.
What it does
Heat Trace is an energy data dashboard with a ML predictor with four main components, the first component is an interactive Choropleth map showing all 50 US states colored to show energy consumption and green bubbles for solar production. Second is our State Explorer tab with a 60 year time series for energy, year-over-year change, and solar production estimates across 21 states. Our insights page shows charts analyzing the zen power permit data, permit count, average system size by city (revealing the residential vs. commercial split), and adoption trends over time. Finally, our ML model allows us to predict annual kilowatt-hour output for a proposed solar installation given 3 inputs: system capacity (kW), average temperature, and solar irradiance.
How we built it
We built the frontend as an interactive dashboard using Streamlit, pulling in data that we cleaned and merged using pandas across three public sources EIA, EPA, and ZenPower. For the predictive model, we used a Random Forest Regressor to estimate annual kilowatt-hour output from system capacity, temperature, and solar irradiance, alongside a physics-based formula as a check against real-world expectations. The time series visualizations on the State Explorer page are powered by an LSTM to capture long-term energy consumption trends across six decades of data.
Challenges we ran into
Challenges we faced was faced was sourcing our data, multiple of our datasets were across different sources and figuring out what was compatible and meaningful to join together. Ultimately we had to adapt our scope based on what was available. Our ML model challenge was to determine the right structure, specifically whether the relationship between our inputs and solar output was linear or required a more complex approach. Getting that right took experimentation and validation before we were confident in the predictions.
Accomplishments that we're proud of
Honestly, we didn't expect the ML model to perform as well as it did given how little time we had that was a nice surprise. We're also proud of getting three completely separate datasets to actually talk to each other by wrangling EIA, EPA, and ZenPower data into one clean pipeline to make it possible to display something meaningful.
What we learned
We learned to research into the datasets more extensively as not having enough data to prove our conclusions led us to adapt to project mid scope to work with whats available.
What's next for Heat Trace
We'd love to expand the dataset to cover more states and integrate more real-time data sources as they become available. On the model side, there's room to improve predictions by feeding in additional features things like roof orientation, shading, and local utility rates that would make estimates more precise and personalized.
Log in or sign up for Devpost to join the conversation.