Previsão do consumo de eletricidade em instituições educacionais: uma abordagem de Cooperative Ensemble Learning aplicada a modelos de regressão e classificação

Abstract

Brazilian federal educational institutions face numerous challenges in managing their costs, such as electricity, especially in a scenario of budget constraints. Forecasting this consumption can be an effective tool for resource management; however, it is a complex process due to the influence of various factors on this expense. This dissertation’s main objective was to propose and evaluate the Cooperative Ensemble Learning approach named Weaker Separator Booster with Coyote Optimizer (COA-WSB), comparing its performance against learning models: Echo State Networks (ESN), Multilayer Perceptron (MLP), Random Forest (RF), and Extreme Gradient Boosting (XGBoost) in forecasting electricity consumption in a multi-campus educational institution (Instituto Federal do Paraná - IFPR). The methodology used monthly historical data from 24 campuses, combined with exogenous variables related to climate, the academic calendar, and course offerings. Feature analysis with SHapley Additive exPlanations (SHAP) identified the previous month’s consumption (LAG_01) as the variable with the greatest impact, followed by the COVID period indicator and other seasonal lags. The models’ hyperparameters were optimized using Simulated Annealing (SA) and Particle Swarm Optimization (PSO) metaheuristics. The study compared two training strategies (Local, per-campus, and Global, with aggregated data) across two modeling approaches (Regression and Classification) for 3, 6, and 12-month horizons.The results showed that Global Training was superior to Local Training. In regression, the average RRMSE of the best global model at 𝐻 = 12 (COA-WSB) was 0.259 (“reasonable”), surpassing the best local RRMSE (0.301). In classification, the global approach was also superior, achieving an accuracy of 73.6% (RF, 𝐻 = 3, 3 classes), but performance proved sensitive to the number of classes, dropping to 27.8% with 10 classes. The use of data from varied campuses (Global Training) was fundamental to increasing the models’ predictive performance. Notably, the proposed COA-WSB approach stood out, presenting the most accurate and consistent results in medium and long-term regression forecasts, which are essential horizons for budget planning, thus constituting a robust and useful tool for the institution’s managers. Regarding the conversion of the problem into a classification approach, the results indicate it is a viable alternative, provided that a small number of classes (3 to 5) is used; otherwise, the accuracy significantly worsens. In summary, using the classification approach with a reduced number of intervals (classes), although it may increase accuracy, might not provide a sufficient level of detail about consumption to support management decision-making processes.