| dc.description.abstract | Faced with constant and significant changes in production processes, in which the manufacturing industries of goods and services are increasingly challenged, both by the complexity caused by the variation in demand and by the fierce competition imposed by the market, manufacturers need to offer products of high quality, short term, and high customization. In this scenario, Production Planning and Control is an important task of the production system, where developing a reliable demand forecasting process is the first step in searching for optimal plans. This research proposes a demand forecasting system that helps in production planning based on different formalisms of time series and the use of metaheuristics in the optimization of formalisms. Starting from a systematic literature review, summarized from text mining techniques, the work seeks to understand how production planning and demand time series analysis tools interact, enabling their improvement. It was also sought to identify the underlying time series formalisms, the combination techniques, the performance metrics used in forecasting demand for production planning, and the types of manufacturing industry applied. The study allowed, among other things, the elaboration of the conceptual framework on the forecast system to be proposed. In order to assess the quality of the artifact, the demand forecasting system was applied in a goods manufacturing industry in the sector of hygiene, cleaning, and sanitizing products for domestic and professional use. The unit of analysis of this research is demand histories, that is, univariate time series. The forecast was carried out considering individual and combined formalisms, as well as the methodology used by the company. The results of the forecasts were evaluated based on the application of performance metrics and compared to the methodology used by the manufacturing industry (called JUA). The results showed that the demand forecasting system formed with the help of combiners had a superior performance than the company’s forecasting model. Several of the models of the proposed forecast system proved to be adequate for predicting all the time series of the analyzed product family. The results of the application of the artifact suggest that the combination formalisms based on the simple mean (cSA), on artificial neural networks (cANN), on support vectors (cSVR) and on minimum variance (cMV), have superior quality to the JUA methodology. The models can then be used as a forecasting tool for the cleaning product manufacturing industry, in order to reduce the problems of underproduction and overproduction, which cause a substantial and continuous increase in resources tied up in stock, reduced productivity and increased workload of overtime nature. Linking to operational research, forecasts were obtained through a one-step-ahead horizon, which allows managers to optimize the allocation of resources, materials, and costs, focusing on the planning and scheduling of short-term production of the company object of study. | en |
