Probabilistic Approaches for Surrogate Modeling with High-Dimensional Data to Predict Short-Term Meso-scale Surface Temperature

Abstract:

From its effects on individual comfort to significant mortality rates, surface temperature plays an important role in urban areas. Recognized as potentially vulnerable places due to their dense population, urban areas are also characterized by elevated trapped heat within the building canopy (Urban Heat Island; UHI). Such distinctive heat pattern of cities increases the surface temperature and is well-known to have a nonlinear synergy with extreme heat wave amplifying temperature-induced risks. Thus, adequate weather forecast models play an important role in detailed and/or real-time risk analyses. However, physics-based models, the golden standard for highly accurate and reliable forecasts, are computationally expensive to run.

Probabilistic approaches have been adopted as an alternative way to reduce computational cost, taking advantage of achievements in geo-statistics and machine learning. However, the parameter space of models for predicting surface temperature is usually high-dimensional, and the available datasets are rarely large enough to uniquely identify a best-fitting model. In this research, we propose probabilistic models, which fairly limit the degree of freedom to predict meso-scale surface temperature while maintaining enough flexibility, and analyze their explanatory power with respect not only to forecast accuracy, but also to the number of parameters and computational cost to attain statistical inference.

To calibrate and test our statistical model, a numerical experiment is also conducted using a physical model framework coupling the Princeton Urban Canopy Model (PUCM) to the Weather Research and Forecast (WRF). Three summers, from 2016 to 2018, in New York City NY and City of Pittsburgh PA are reanalyzed to acquire 1km spacing temperature, which is a high enough resolution for city-scale risk assessment. Then, the proposed surrogate models are calibrated with the simulated data, and the performance assessment is performed for the individual models. All the proposed models show adequate accuracy when forecasting short-term meso-scale surface temperatures, largely reducing the computation time relative to the physics-based models. This performance assessment provides a guide to future applications and expanded probabilistic risk analyses.