Accurate and reliable forecasts of short-term demand for a microgrid and power genetration forecasts from photovoltaics are an important part of smart microgrid management, influencing technical, operational, and financial aspects of microgrid operation. Managing university campus microgrids is challenging due to their specificity. This work is a response to the need of university campuses for a reliable deep learning model supporting microgrid energy management that allows for accurate forecasting of both energy demand and photovoltaic electricity generation. In this study, sixteen forecasting models based on deep learning, utilizing various layer combinations and architectures, were developed and evaluated. The combination of a deep neural network, Kalman Filtering, and a Self-Attention model was found as a robust and reliable solution for managing energy within a university campus microgrid. This model enables precise photovoltaic power generation and load forecasting, as indicated by the high determination coefficient values (R2=0.987 for load forecasting, and R2=0.989 for PV power generation forecasting) and the low mean absolute error MAE, root mean square error RMSE, and their standard deviations (4.598 and 5.749 respectively for load forecasting and 0.953, 1.833 for PV generation). Furthermore, it effectively manages conditions that differ from those encountered during training, showcasing its high generalization capacity (R2=0.982 MAE=5.006, RMSE=1.257 for load forecasting and R2=0.992 MAE=0.634, RMSE=1.395 for PV generation). Additionally, this model achieved the lowest standard deviations. Its practical application was validated using actual energy performance data from Savona Campus microgrid in Italy over all seasons, rather than relying on simulation or small-scale research, which helped minimize errors arising from training set imperfections.