Eventos Anais de eventos
ENCIT 2022
19th Brazilian Congress of Thermal Sciences and Engineering
Power Plant Performance Assessment - Physical and Machine Learning Integration Through Multifidelity Modeling
Submission Author:
Lara Werncke Vieira , RS
Co-Authors:
Lara Werncke Vieira, Augusto Delavald Marques, Renata Rech de Souza, Paulo Smith Schneider, Paola Mendes Albino
Presenter: Lara Werncke Vieira
doi://10.26678/ABCM.ENCIT2022.CIT22-0539
Abstract
Energy conversion systems have assumed a crucial role in current society. In this context, the need for a more sustainable way of electricity production is evident. The integration of renewable energy sources, improving the conversion efficiency, and controlling of power plant emissions is essential for energy transition. While a global energy transition is underway, further action is needed to reduce carbon emissions through smooth transactions from fossil fuels to clean energy. Power system optimization problems can be solved with the use of AI, which is becoming a key enabler of a complex, new and data-related energy industry, providing a key tool to increase operational performance and efficiency. The present paper proposes a multifidelity modeling approach to increase the efficiency of a 2x360MW coal-fired power plant located in Ceará, Brazil. The goal is to use multifidelity in order to ally the benefits of both the physical and machine learning modeling to increase accuracy. Multifidelity enable data associated with a high fidelity function that is costly to be evaluated or whose values are available at only a few points to be combined with another function of low-fidelity. The low-fidelity function is typically faster and can be evaluated at more points to form the set. A set of eleven experiments acquired on site is the high-fidelity data and the low fidelity data considers a Machine learning approach based on artificial neural networks (ANN). The low-fidelity model is used to generate samples globally over the range of the design parameters. The ANN considers operating data from 2018 to 2020. The proposed methodology aims safe and stable conditions to improve power plant performance during operation. The ANN results showed a relative deviation of around 6\% while the multifidelity model decreased the relative deviation to around 1\%. The increase in model accuracy and ability to represent the high-fidelity data is significant through the application of multi-fidelity.
Keywords
Power generation, Multifidelity, artificial neural network, Performance Assesment
