The escalating energy and environmental crises underline the imperative for sustainable cities and societies. For effective and real-time energy management, this study proposes an enhanced building energy consumption prediction system. It introduces a novel concept named region-wide occupant energy-use behavior probability and incorporates it into the input system, which better reflects real-time and complex energy-occupant-environment interactions in buildings. In addition, it integrates the squeeze-and-excitation attention mechanism, sparrow search algorithm, and convolutional neural network processes for optimizing data processing and hyperparameter selection. Validation in seven sample buildings demonstrates the proposed prediction system has a better balance between time and accuracy, reducing 36.32% MAPE and 31.20% CV-RMSE on average compared to all other prediction systems, only with 118.354s of extra time consumption increase compared to the least time-consuming method. Furthermore, this study discusses methods for selecting suitable input systems and algorithms based on building type, data collection conditions, accuracy, and time consumption. Finally, the enhanced prediction is applied to forty-five buildings in a university community, yielding a 12.35% MAPE and a 0.1707 CV-RMSE on average, reaffirming its superiority and practicality.