Banca de DEFESA: DANIEL FREITAS DE JESUS

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
STUDENT : DANIEL FREITAS DE JESUS
DATE: 22/03/2023
TIME: 10:00
LOCAL: https://conferenciaweb.rnp.br/events/defesa-da-dissertacao-explorando-o-potencial-da-rede-neural-em-histerese-l2p-para-previsoes-de-energias-intermitentes-nos-sistemas-de-smartgrids
TITLE:

Exploring the potential of neural networks in L2P hysteresis for intermittent energy predictions in Smartgrid systems.

PAGES: 70
BIG AREA: Engenharias
AREA: Engenharia Elétrica
SUBÁREA: Sistemas Elétricos de Potência
SPECIALTY: Medição, Controle, Correção e Proteção de Sistemas Elétricos de Potência
SUMMARY:

Using a hysteretic activation function and the L2P hysteresis model, the current study investigates intermittent energy prediction to improve neural networks. The goal of this study is to demonstrate that intermittent energy sources have a high potential to replace nonrenewable sources. Due to generation fluctuations, forecasts are required to safely and efficiently serve supply and demand markets. As distributed grids and smart grids become more important in the global energy mix, forecasting models must be developed and the importance of forecasting for smart grid generation assessed. A practical study of hysteresis reveals that its application reduces MSE errors, making it a potentially useful tool. Several natural and biological phenomena, such as biological brain networks and memory development, exhibit hysteresis, according to a review of the literature. Because the model used in its activation function is known, the results confirm the importance of prediction in a power supply scenario and describe the neural network and L2P hysteresis neurons. The neural network has four configuration parameters that can be used to change the shape of the hysteresis curve, allowing it to be more flexible when training and predicting nonlinear series. The plots contrast three common activation functions used in time series prediction. The neural network values in the experiment were lower than those of the Simoid, RELU, and Hyperbolic Tangent. However, because there is a higher processing overhead offset by future prediction, a greater number of layers are required to approximate the training values. Because the curve was modeled by previous neural network data, hysteresis provided a better approximation. Finally, more training data can help you perform better. Backpropagation and gradient descent require the network to update its weights in response to predicted data. Using a delay memory, the curve is updated based on the signal in this method.

COMMITTEE MEMBERS:
Presidente - Interno ao Programa - 1205456 - LUIZ ALBERTO LUZ DE ALMEIDA
Membro Titular - Examinador(a) Interno ao Programa - 1762419 - CARLOS EDUARDO CAPOVILLA
Membro Titular - Examinador(a) Externo à Instituição - OMAR ALEXANDER CHURA VILCANQUI
Membro Suplente - Examinador(a) Interno ao Programa - 1669196 - LUIZ HENRIQUE BONANI DO NASCIMENTO
Membro Suplente - Examinador(a) Externo à Instituição - MARCOS ANTONIO MAIA LAVIO DE OLIVEIRA - FATEC B.Pta