A common way to increase performance in Radio Access Networks (RANs) is to use carrier aggregation. To decide what carrier to aggregate, inter-frequency measurements are used. These measurements increase battery consumption of User Equipment (UE) and temporarily interrupts data transmission. The measurements could instead be estimated based on intra-frequency measurements, using Machine Learning (ML).This is called Second Carrier Prediction (SCP).

Previous work have not put an emphasis on how simulations used for data gathering differ from the realworld. This thesis shines a light on what the differences are, and how they affect the performanceof ML models when conducting SCP. Noise estimates are gathered from data collected in the real world. The SCP viability in this environment is tested with three different families of ML models, NeuralNetworks (NN), Random Forests (RF) and eXtreme Gradient Boosting (XGB). 

The results show that models trained on pure simulation data performs worse when predicting on noisy data. This highlights the importance of training models on the same data distribution that is expected tobe present during inference. Input features varies in impact, with path delay having a positive performance increase across all categories of noise. Time series, i.e. giving the model a history of changes,does not seem to have an impact on performance.