The building field occupies a significant ratio of overall energy consumption globally, and nationally in the Moroccan case. This makes the management of building energy demand a fundamental subject for decision makers to fulfill the sustainable development challenges, in both early and late stages of the building’s lifecycle, through accurate energy prediction. Nowadays, with the Advancements in computer science, Machine Learning (ML) has proven its high performance in thermal energy prediction, leveraging large datasets to achieve high accuracy. However, obtaining an ideal-sized database for model training is not always feasible. Giving this situation, this paper conducts a comparative examination of four ML models; “Multi-layer perceptron (MLP)”, “Decision Tree (DT)”, “Support Vector Machine (SVM)” and “Linear Regression (LR)”, trained to predict heating and cooling loads, based on a small-size database, generated through Dynamic Thermal Simulations (DTS) of a single-zone building throughout Morocco’s six climatic zones. Results demonstrate that even with a limited dataset, accurate thermal energy forecasts can be achieved. Among the models, the MLP outperformed others, highlighting its suitability for even small databases in building energy prediction.