Combining demand classification and forecasting models in spare parts inventory management for the energy sector

Industries are going through a changing scenario, in which they need to constantly seek to improve their processes and increase productivity. One of the factors that contribute to improving the performance of companies is inventory management, which starts with forecasting demand. This research aimed to compare the Machine Learning models, Random Forest and XGBoost, and the classic Exponential Smoothing prediction models, Holt, Holt-Winters and SARIMA Seasonal Autoregressive Integrated Moving Average. Identify which model generates the best forecasts for the items of a company in the energy sector in southeastern Brazil. The database consisted of 2,244 items. The process of classifying the demand was carried out in four types: Irregular, Intermittent, Erratic and Regular. The process of adding financial value was applied in three groups: A, B and C. The prediction models were implemented using Spyder (Python 3.11), performing 13,464 iterations. To measure the performance of the prediction models, it was, the Criteria Importance Through Intercriteria Correlation (CRITIC) multicriteria analysis method was used to determine the score of the models based on the error measures. The Augmented Dickey-Fuller tests were performed (ADF), Kwiatkowski-Phillips-Schmidt-Shin (KPSS), Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC). The result of the analysis identified that, in general, the model that presented the best predictions was Random Forest, followed by ExpSmoothing. The model Random Forest presented better forecasts for items with irregular and intermittent demand, Holt-Winters performed better for regular items, while items with erratic demand did not present models with significant prominence. It also identified that the 82 items with value classification A are subject to a more detailed analysis, due to the high representativeness of the value.