| Sumario: | This master's thesis presents a short-term electricity demand forecasting model for the Duques de Soria campus in Spain, using machine learning techniques including Random Forest (RF), Multi-Layer Perceptron (MLP), and Long Short-Term Memory (LSTM) networks. Based on six years of hourly smart meter data (2019–2024), the study explores the impact of various features—such as lagged load, temporal variables, calendar indicators, and weather conditions—on prediction performance. After thorough data cleaning, feature engineering, and model optimization, the RF model, which utilized only historical load and hour features, achieved the best results with an MAPE of 9.51% and an MAE of 5.95. While MLP showed strong robustness result with 9.66% MAPE with the same senario, Though LSTM required richer sequences and more training data to perform better. The study concludes that careful feature selection is just as crucial as algorithm choice, and that temporal features at the hourly scale provide the highest predictive power. This work supports sustainable energy management on university campuses by enabling more accurate, data-driven forecasting
|
|---|
