Climate risk assessments that inform how the built environment might accommodate future weather rely heavily on climate-model predictions derived from historical trends and data. AI presents an exciting opportunity to explore new ways to model these conditions.
Using 70-year continuous historical datasets recorded by the National Oceanic and Atmospheric Administration, we trained ML models to predict an hourly temperature dataset that can be used in building analytics to estimate monthly heating and cooling demand.
Incorporating additional information, like additional data stations, dry and wet bulb temperatures, dew point and relative humidity, we trained the model to predict hourly temperature data for 2020. The predicted and measured data track each other closely.
Our AI models, overall, produced less error than the common practice (University of Southampton’s CCWorldWeatherGen), which underpredicts heating demand and overpredicts cooling load. Further refinement of our approach should lead to improved building design and economic modeling of heating/cooling demands.