The University of Ottawa plans to achieve carbon neutrality by 2040. The university analyzes a framework that integrates fossil fuel and renewable energy systems to realize this measurable objective. This integration is aimed at meeting a portion of the electricity demands of the campus, particularly during peak periods, with the goals of minimizing energy expenses and mitigating environmental impacts. To reach this ambitious goal, short-term energy demand forecasting of buildings is essential as it is vital in optimizing building energy management systems, increasing renewable energy sources, shaving peak loads, and curbing energy expenses. However, accurately predicting electricity consumption for peaks is challenging. This research addresses 3-h-ahead load prediction by evaluating the effectiveness of Recurrent Neural Networks (RNN) and long short-term memory networks (LSTM), both with and without the attention layer. The methodology employed in this study involves a step-by-step procedure for fine-tuning each algorithm to optimize its predictive capabilities. The results show that the performance of models almost doubled through the hyperparameter optimization.