Surrogate Modeling Based on Machine Learning Approaches for Hospitals as Microgrids

Abstract

This thesis investigates the use of machine-learning surrogates for short-term simulation of hospital microgrids, where reliable operation is critical due to high energy demand, local PV generation, battery storage, and strict availability requirements. A synthetic stateful reference simulator was developed to generate 15-minute resolution time-series data, which was used to train and evaluate Persistence, Linear Regression, LSTM, and Transformer surrogate models. Beyond forecasting accuracy, the thesis introduces a decorator-based runtime architecture that allows surrogates to act as replaceable simulation components while enabling trust monitoring, simulator fallback, re-entry, and lightweight online recalibration. The results show that the proposed approach can reduce simulator calls while maintaining reliable predictions, with the strongest contribution being the controlled and transparent integration of ML surrogates into simulation workflows.