The formulation of advanced management strategies in applications involved with the energy market is recommended to avoid unbalances between the demand for power with the supply that may result in potential system failures. However, the uncertainties of the demand behavior as a result of the non-linear patterns may complicate the formulation of an optimization plan that reduces the energy costs as much as possible. Hence, forecasting methods from the artificial intelligence and machine learning areas should be considered to learn non-linear patterns in the training phase in order to obtain accurate predictions for unseen data. Some forecasting methods can be enumerated including Artificial Neural Networks, K-Nearest Neighbors and XGBoost. Understanding the reasons behind the obtaining of higher or lower forecasting accuracy is not easy for data scientists and machine learning professionals. Hence, the Explainable Artificial Intelligence area arises as a field that generates explanations for example of the input features contribution to the forecasting performance. Some of these methods are the following: Local Interpretable Model-Agnostic Explanations (LIME), SHapley Additive exPlanations (SHAP), impurity-based feature importance and Explainable Boosting Machine (EBM). This paper analyzes the input features (past consumptions and sensors) that contributed more or less to the forecasting performance for an office building with the support of the impurity-based feature importance and the Explainable Boosting Machine methods.