A master thesis from Aalborg University
Data-driven Modeling and Black-box Optimization of MPC Parameters for Building Control
Author(s)
Term
4. semester
Education
Publication year
2024
Submitted on
2024-05-31
Pages
98 pages
Abstract
En tidligere udviklet resistiv-kapacitiv (nRnC) modelleringsmetode blev forbedret ved at inkorporere forsinkede output, en effektiv træningsmetode og en forbedret skaleringsmetode til datapreparation. Et fysikinformeret neuralt netværk (PINN) blev også udviklet ved at bruge den trænede nRnC-model til at generere træningsdata. PINN-modellens tab kombinerede ét-trins forudsigelse af de genererede nRnC-data og rekursiv forudsigelse af de originale træningsdata, med automatisk differentiering anvendt til at udregne fejlgradienter. Derudover blev en Bayesiansk optimeringsmetode implementeret til online justering af Model Predictive Control (MPC) hyperparametre, idet problemet blev behandlet som et black-box optimerings problem. Den Bayesianske optimeringsmetode modtog feedback om præstationssignaler, baseret på analysen af MPC-præstation ved foreslåede hyperparameter værdier. Testning mod den originale nRnC-model viste forbedret præstation for både den forbedrede nRnC-model og PINN på tværs af flere bygningssimulationer. Dog forbedrede hyperparametertunings systemet ikke præstationen på grund af fejltolkning af påvirkning fra vejforhold.
A previously developed resisitive-capacitive (nRnC) modeling approach was improved by incorporating delayed outputs, an efficient training method, and an improved scaling method for data preparation. A physics-informed neural network (PINN) was also developed using the trained nRnC model to generate training data. The PINN model's loss combined one-step prediction of the generated nRnC data and recursive prediction of the original training data, with automatic differentiation used for error gradients. Additionally, a Bayesian optimizer was implemented for online tuning of Model Predictive Control (MPC) hyperparameters, treating the problem as a black-box optimization. The Bayesian optimizer received performance signal feedback, based on the analysis of MPC performance for proposed hyperparameter values. Testing against the original nRnC model showed improved performance for both the improved nRnC model and the PINN across multiple building simulations. However, the hyperparameter tuning system did not improve performance due to misinterpretation of impact from concurrent ambient conditions.
Keywords
Documents
Colophon: This page is part of the AAU Student Projects portal, which is run by Aalborg University. Here, you can find and download publicly available bachelor's theses and master's projects from across the university dating from 2008 onwards. Student projects from before 2008 are available in printed form at Aalborg University Library.
If you have any questions about AAU Student Projects or the research registration, dissemination and analysis at Aalborg University, please feel free to contact the VBN team. You can also find more information in the AAU Student Projects FAQs.