Accurate occupancy prediction in smart buildings is crucial for optimizing energy management, improving occupant comfort, and effectively controlling building systems, particularly for short- and long-term horizons. Recently, deep learning-based occupancy prediction methods have gained considerable attention. However, the full potential of these methods remains under explored in terms of model architecture variations and prediction horizons. This study introduces cascaded LSTM and cascaded Bi-LSTM models for multi-horizon predictions from 10 minutes to 24 hours, integrating a modified activation function, additional input features, and optimized hyper-parameters using OPTUNA. Traditional performance metrics and various other analyses were conducted to compare the models. Both models performed well for short- and long-term predictions, with minimal differences in the results. Nevertheless, analysis focusing on non-zero data errors (accounting for approximately 11% of occupied periods) and occupancy-wise errors showed a significant performance gap between the two models. The cascaded Bi-LSTM model demonstrated consistent performance across various prediction horizons and occupancy variations, with accuracy approximately 10-15% higher than the cascaded LSTM model, highlighting its superior capability in capturing complex dataset dynamics through a bidirectional process. This study highlights the importance of additional input features, data feature analysis, and multi-perspective result analysis to select the most suitable model for occupancy prediction, validated with pre- and post-modeling feature importance analysis.