Dataset of "Comparison of Localization Methods for Internet of Things in 5G Cellular Networks: A Wide-scale Assessment"

As the 3rd generation partnership project (3GPP) organization pushes out new releases,positioning in heterogeneous mobile networks enables the achievement of the accuracy requiredin the majority of industrial applications without dependence on global navigationsatellite systems (GNSS). This study presents the results gathered during an extensive measurementcampaign related to the practical applicability of localization in next-generationheterogeneous networks. We present an accuracy comparison of basic timing advance (TA)localization with the k-nearest neighbor (KNN), decision tree-based random forest (RF),extreme gradient boosting (XGBoost), and long short-term memory (LSTM) recurrent neuralnetwork. Our results demonstrate that TA cannot be considered an optimal solutionfrom the perspective of localization accuracy because the error roughly corresponds to theaverage separation distance from the base station (BS) to the end device (ED). In addition,we found that the LSTM approach is not optimal for the outdoor localization of movingED because of the combination of multiple factors, with sparse deployment being the mostimportant. The median value of the location error of the LSTM was more than 200m higherthan that of the TA for the self-validation dataset. However, a simple KNN regression showssolid results for 5G New Radio (NR) operating in the non-standalone (NSA) mode. KNNprovided the most accurate results of all methods, with median error values of approximately12 (k=3) and 82 (k=5) m for the self-validated and cross-validated datasets, respectively.