The accurate energy consumption forecast of educational institutions is critical for sustainable energy management and intelligent campus infrastructure. Classical forecasting models fail to simultaneously capture local temporal variations and long-range dependencies, thus diminishing their reliability across different time horizons. To overcome this, we propose an effective hybrid deep learning method that combines Convolutional Neural Networks (CNNs) with Long Short-Term Memory (LSTM) networks and domain-driven feature engineering to incorporate occupancy and seasonality patterns. The dataset from an office building (research support facility) was used to perform a series of short-term (05-minutes), medium-term (daily), and long-term (weekly) forecasts. Experimental results show that the proposed forecasting approach demonstrates greater precision than the ten baselines including machine learning and deep learning methods in short, medium, and long time horizons. In particular, the model with feature engineering improved Root Mean Squared Error (RMSE) by 4.71% and Mean Absolute Percentage Error (MAPE) by 3.99% in the short term and saw even greater improvements for medium-term forecasts, with RMSE improved by 6.08% and MAPE by 5.37%, and for long-term forecasts, with RMSE improved by 2.58% and MAPE by 0.33%, respectively. The outcomes here offer significant benefits to the science community, supporting the development of such smart campuses with a specific, understandable methodology for energy forecasting. The methodology enables the application of data-driven energy management, cost optimization, and sustainability planning in educational spaces. The findings also provide valuable insights for deploying Artificial Intelligence (AI)-driven infrastructure management systems in academia and institutional energy governance.