Due to users demand for high mobile bandwidth, operators increasingly spend money in routine optimization activities, adding cell sites, and deploying new technologies. This is mainly to improve capacity and quality of congested cells in network hotspot areas. In hotspots, Quality of Service (QoS) is highly degraded. Dynamically managing and optimizing network resources improves degraded QoS in real time and decreases capital and operational costs. For dynamic optimization, proactive and accurate identification of hotspot variations is required. Some studies have modeled hotspot prediction based on data traffic usage and others based on user density. However, this may not accurately identify congested cells since cells have different capacity configurations, and network QoS such as throughput are affected by factors such as radio conditions.rnThis study focuses on developing a model to accurately identify congested cells based on number of users and user throughput by utilizing cell counters collected from Long-Term Evolution (LTE) network. Data preprocessing techniques such as replacing missing values using per cell per hour historical mean, and resampling to reduce class imbalance are applied on the collected data. Long Short-Term Memory (LSTM), a deep recurrent neural network, is used to model hotspot prediction, and performance of the model is evaluated using metrics such as accuracy, precision, and F1 score.rnExperimental results show the model performs with an accuracy of 89.13% and F1 score of 85.5%, and predicts for four future hours. Therefore, the model achieves acceptable performance, and it can help operators predict hotspots for dynamic optimization to improve QoS in real time.