Fifth generation (5G) and beyond cellular networks bring significant advantages over their predecessors in the RAN part of the network, as they are based on a disaggregated, cloudbased architecture, marking a major shift from the monolithic designs of previous generations. This disaggregated approach has led to the development of Virtual RAN (V-RAN) and Open RAN (O-RAN) architectures, which provide greater flexibility in deployment while optimizing costs and resource utilization. Simultaneously, the increasing demands for high-speed, lowlatency connectivity to support diverse 5G applications such as IoT, V2X, and URLLC highlight the critical need for reliable and scalable network solutions. In this paper, we address the scalability challenges of the CU-UP component, identified as a bottleneck in the 5G disaggregated RAN, particularly under high traffic loads. We propose a proactive horizontal scaling mechanism for the CU-UP, leveraging machine learning to forecast traffic demands. This involves using an xApp on the Near-RT RIC that integrates a pre-trained Long Short-Term Memory (LSTM) model using a real-world dataset which predicts traffic demands, enabling proactive scaling of CUUP instances ahead of peak periods to maintain Quality of Service (QoS). The proposed mechanism was implemented and evaluated using real-world testbeds under dynamic conditions, demonstrating its effectiveness in practical environments. Our findings emphasize the importance of integrating ML techniques to forecast network loads accurately and adapting such scaling mechanisms in 5G systems to reduce packet loss and unnecessary scaling overhead while ensuring smoother operation during highdemand periods.