Estimação de consumo elétrico individual utilizando temporal convolutional network

Every year, companies in the electricity distribution sector suffer losses due to problems in acquiring consumption data for billing. These problems range from human error to customer fraud. Thus, estimating monthly energy consumption is a problem of great interest in the context of electricity dist...

ver descrição completa

Detalhes bibliográficos
Autor: LEMOS, Victor Henrique Bezerra de
Tipo de documento: dissertação
Estado:Versão publicada
Data de publicação:2021
País:Brasil
Recursos:Universidade Federal do Maranhão (UFMA)
Repositório:Biblioteca Digital de Teses e Dissertações da UFMA
Idioma:português
OAI Identifier:oai:tede2:tede/3534
Acesso em linha:https://tedebc.ufma.br/jspui/handle/tede/3534
Access Level:Acceso aberto
Palavra-chave:Consumo elétrico
Séries temporais
Predição
Temporal convolutional network
Energy consumption
Forecasting
Time series
Ciência da Computação
Descrição
Resumo:Every year, companies in the electricity distribution sector suffer losses due to problems in acquiring consumption data for billing. These problems range from human error to customer fraud. Thus, estimating monthly energy consumption is a problem of great interest in the context of electricity distribution companies, as a way of mitigating reading problems. A forecast that minimizes error plays an important role in identifying inconsistencies in the monthly billing process. For this, electric companies have invested in the use of prediction to define limits, between upper and lower, where a reading is considered normal. Thus, in this context, this work presents a method for predicting individual monthly electrical consumption. A method based on a Temporal Convolutional Networl (TCN) network was developed, combined with the application of an Optimization of the Hyperparameters of the proposed architecture. A pre-processing workflow was also created, able to alleviate some of the problems that can be found in the clients’ historical series, in addition to helping to generate a better representation of the time series for the predictive model. The proposed approach proposed SMAPE total 16.86 %, being superior in six of the eight consumer classes, which correspond to 98.23 % of the dataset, when compared to other methods found in the literature, such as: Autoregressive Integrated Moving Average (ARIMA), Simple Exponential Smoothing (SES),HOLT, Stocastic Gradient Descent (SGD), Long short term memory (LSTM) and the TCN network itself. In general, showing a proposed methodology capable of performing as expected, in the most different scenarios.