Poster Abstract: The Accuracy Paradox in Data-Driven Model Predictive Control

Abstract

Data-Driven Model Predictive Control (DDMPC) offers a scalable approach for optimizing building energy systems, yet its real-world adoption is hindered by the “accuracy paradox”—the disconnect between a model’s predictive accuracy and its practical control utility. This paper systematically diagnoses this paradox by evaluating a pool of 261 artificial neural networks (ANNs) and linear models within a closed-loop framework. We prove systematically that traditional metrics like Mean Absolute Error (MAE) fail to reliably predict actual control performance. By assessing models across multiple operational scenarios, we reveal that a model controlling the system optimally under standard conditions can fail catastrophically when subjected to novel scenarios, such as dynamic pricing or shifted comfort bounds. Our findings highlight that robust DDMPC requires a paradigm shift: instead of optimizing for single accuracy metrics on test datasets, model selection must prioritize area-wise validity and physical plausibility to ensure stable control behavior across the entire operational regime.

Authors 2

1. Martin Rätz Aachen Institute for Energy Efficient Buildings and Indoor Climate

   RWTH Aachen University

   Affiliation as printed

   E.ON Energy Research Center; Institute for Energy Efficient Buildings and Indoor Climate, RWTH Aachen University, Aachen, Germany

2. Dirk Müller Aachen E.ON Energy Research Center Institute for Energy Efficient Buildings and Indoor Climate

   RWTH Aachen University

   Affiliation as printed

   E.ON Energy Research Center, Institute for Energy Efficient Buildings and Indoor Climate, RWTH Aachen University, Aachen, Germany