Calibrated uncertainty estimation for spatiotemporal data

The application of Artificial Intelligence (AI) in high-risk domains like personalized medicine and sustainable energy management is constrained by the substantial risks of predictive errors, which can lead to consequences such as harm to individuals, infrastructure damage, and significant financial losses. A major challenge lies in the complexity and dynamic spatiotemporal nature of the data (data show a spatial and temporal complexity as that provided by a sensor network). Capturing intrinsic spatiotemporal dependent dynamics is inherently difficult to model with AI, increasing the risk of predictive errors even further. This project aims to mitigate these risks by developing a method for calibrated uncertainty estimation for spatiotemporal data using Conformal Prediction (CP). By rigorously aligning predicted probability distributions with actual observed events, CP will be leveraged in the developed methods herein to enhance the reliability of AI predictions and improve decision-making robustness.