Data Driven Model Identification for Remote Electrical Tilt Systems

Abstract: This thesis explores the use of supervised machine learning for modelling the dynamics of Remote Electrical Tilt (RET) telecom systems. Three methodologies, including linear regressionfor linear dynamics models, Gaussian Process (GP) regression, and Recurrent Neural Networks (RNN) with Gated Recurrent Units (GRU) are proposed. Linear regression is simple to implement and highly interpretable. However, due to their simplicity, linear models fall short in modelling the complicated dynamics of RET systems. GP regression is capable of modelling non-linear dynamics and has inherent uncertainty estimation, but suffers from computational inefficiency. Lastly, RNNs provide the capability to model temporal dependencies by consideringtime-series data. However, it is the least interpretable method. The thesis provides a comparative analysis of these methods, w.r.t. the accuracy of single and multi-step ahead predictions. It was found that RNNs provided the best balance of computational cost and higher accuracy over linear and GP regressions. Future research suggestions include trying other methods that could consider interactions between antennas for more realistic models and applying methods in model-based control settings to determine their potential in optimizing RET systems.