Forecasting power consumption consists in predicting its future values, using past ones and/or other features of interest. With the exponential growth of urban population during the last 20 years, and climate change-related power issues, the subject seems to be a hot topic. A special case is that of academic campuses with particular consumption behaviors, partly linked to a younger population. However, accurately predicting the power consumption in an academic context is a difficult task, because of the necessity to have a complete smart metering installation, as well as efficient prediction algorithms adapted to different consumption patterns (e.g., classrooms, student rooms). In this paper, we propose (1) an Advanced Metering Infrastructure (hardware) for measuring the power consumption in an Academic Campus; (2) an approach for multi-profile power forecasting, using a single machine learning algorithm with automatic hyper parameter tuning (software). Our network is composed of 18 Smart Meters strategically placed at different locations in our Engineering School’s Campus, connected to a central Supervisory Control and Data Acquisition (SCADA) station. In our approach, we propose an automatic hyper parameter selection technique for an eXtreme Gradient Boosting (XGBoost) prediction model using the Genetic Algorithm Meta-heuristic. The experimental results are quite good with a Mean Absolute Percentage Error (MAPE) as low as 8% for the first cluster, compared to 16% for other popular machine learning algorithms (Support Vector Regression, Long Short-Term Memory).