Previsão de demanda para sistema de abastecimento de água
Data
2017-11-15
Autores
Título da Revista
ISSN da Revista
Título de Volume
Editor
Biblioteca Digital de Teses e Dissertações da USP
Universidade de São Paulo
Escola de Engenharia de São Carlos
Universidade de São Paulo
Escola de Engenharia de São Carlos
Resumo
Descrição
O presente trabalho de pesquisa enfoca a problemática da previsão de demandas com vistas à operação dos sistemas de abastecimento de água em tempo real, utilizando-se dados de consumo horários de água das cidades de São Carlos e Araraquara, SP, para que se identifique o modelo que produza os melhores ajustes. Foram estudadas as redes neurais artificiais Perceptron de Múltiplas Camadas (RNAs MLP), a Rede Neural Dinâmica (DAN2) e duas RNAs híbridas, sendo que estas últimas consistem em associar previsão por séries de Fourier com a RNA MLP e a DAN2, sendo denominadas respectivamente RNA-H e DAN2-H. As entradas fornecidas para os modelos de previsão foram escolhidas com base na revisão bibliográfica e por meio de análise de correlação, considerando os dados de consumo e as variáveis meteorológicas, tais como temperatura, umidade relativa do ar e ocorrência de chuva. Os melhores modelos de previsão utilizaram a DAN2, a qual se mostrou de manuseio mais fácil em relação às redes neurais de múltiplas camadas, pois dispensa o processo de tentativas e erros para se determinar a melhor arquitetura para os dados fornecidos ao modelo. Os melhores modelos de previsão para a próxima hora produziram um erro médio absoluto de 2,25 L/s (DAN2-H) para um subssetor de Araraquara, representado cerca de 8% do consumo médio, e 2,3 L/s (DAN2) para um setor de São Carlos, equivalente a 4% do consumo médio.
The present work focuses the problem of water demand forecasting for real time operation of WSS. The study was conducted using hourly consumption data from water distribution system from the cities of São Carlos, Araraquara, SP, to identify the model that fits better. It were studied the artificial neural network Multilayer Perceptron (ANN MLP), the Dynamic Neural Network (DAN2) and two hybrid ANN. The hybrid ANN is an association of the water demand prevision by series of Fourier with the ANN MLP and DAN2, which were called respectively ANN-H and DAN2-H. The inputs provided to the forecasting models were chosen based on literature review and correlation analysis, considering consumption data and meteorological variables, such as temperature, air relative humidity and rain occurrence. The best forecasting models were based on DAN2, which showed easy handling compared to other neural network with multiple layers, because it dispenses the trial and error procedure to find the best architecture for a given data. The best forecasting model for the next hour produced an absolute medium error of 2.25 L/s (DAN2-H) for a subsector from Araraquara, representing about 8% of the average consumption, and 2.30 L/s (DAN2) for a sector from São Carlos, which correspond to 4% of its average consumption.
The present work focuses the problem of water demand forecasting for real time operation of WSS. The study was conducted using hourly consumption data from water distribution system from the cities of São Carlos, Araraquara, SP, to identify the model that fits better. It were studied the artificial neural network Multilayer Perceptron (ANN MLP), the Dynamic Neural Network (DAN2) and two hybrid ANN. The hybrid ANN is an association of the water demand prevision by series of Fourier with the ANN MLP and DAN2, which were called respectively ANN-H and DAN2-H. The inputs provided to the forecasting models were chosen based on literature review and correlation analysis, considering consumption data and meteorological variables, such as temperature, air relative humidity and rain occurrence. The best forecasting models were based on DAN2, which showed easy handling compared to other neural network with multiple layers, because it dispenses the trial and error procedure to find the best architecture for a given data. The best forecasting model for the next hour produced an absolute medium error of 2.25 L/s (DAN2-H) for a subsector from Araraquara, representing about 8% of the average consumption, and 2.30 L/s (DAN2) for a sector from São Carlos, which correspond to 4% of its average consumption.
Palavras-chave
Abastecimento de água, Previsão de demanda, Rede neural, Forecasting, Neural network, Supply system