The quick expansion of broadband services, Internet of Things (IoT) applications, and multimedia traffic has substantially raised the need for efficient resource management in High Throughput Satellite (HTS) systems. Traditional DRM solutions rely heavily on instantaneous traffic information, leading to reactive resource distribution and limited flexibility to quickly changing user demand. To address this limitation, this paper proposes a Long Short-Term Memory (LSTM)-based user demand forecasting structure for adaptive management of resources in Multi-Beam Satellite (MBS) systems. The proposed method builds on a demand-driven beam clustering approach by employing deep learning for temporal traffic forecast. Over a 5-day observation span, a traffic dataset of 66,680 users spread across 71 beam clusters was collected, with hourly beam traffic demand combined to create multidimensional time-series data for model training and assessment. Before training, the dataset was normalized with Min-Max scaling and converted into sequential samples employing a sliding window method. The LSTM network was trained using the Adam optimizer and evaluated against a Linear Regression (LR) baseline using Mean Squared Error (MSE), Mean Absolute Error (MAE), Coefficient of Determination (R²), and Explained Variance Score (EVS). The experimental findings show that the proposed LSTM model improves the conventional LR model in terms of predicting accuracy and ability to detect nonlinear temporal shifts in beam traffic usage. The improved prediction capabilities allow proactive bandwidth distribution and intelligent DRM, increasing the versatility of future HTS systems. The proposed technique provides a solid foundation for combining demand prediction with adaptive beam hopping, dynamic resource distribution, and AI-driven optimization in future-generation satellite communications systems.