Desenvolvimento de algoritmo adaptável utilizando redes neurais artificiais para previsão de demanda

Abstract

Demand forecasting is an important tool for any job market sector, which companies manage and plan its organization based on the future prediction of their variables involved. Therefore, demand forecasting has a variety of ways and methods to obtain a prediction. For that reason, the present work has the aim of developing a demand prediction algorithm adaptable to the data of the organizations, being able to model and assign process variables that directly influence the prediction. Thereby, the prediction model was based on machine learning, developed with Artificial Neural Network (ANN) and, for its supervised learning training, were used historical data of companies in the furniture segment, US dollar and electric energy consumption in the state of Rio Grande do Sul, in which economic indicators such as PIB, SELIC, TR, among others are used, so that the algorithm can find a relationship pattern in the database and predict demand with accuracy. The dollar forecast model resulted in a smaller map error in relation to the other forecasts, with the obtained error of 1.06%. This low error was also due to the larger database, making the network adapt more effectively to the history and indicators used. The forecast of electricity consumption obtained a map error of 2.15% and for the smaller database which is the forecast of consumption of MDF sheets of the furniture factory resulted in an error of 3.85%.