Data-Driven Estimation of Spatiotemporal Performance Maps in Cellular Networks

For a large class of non-delay-critical applications (e.g., buffered video streaming or data transfer from cloud services to local devices end-to-end throughput becomes the most crucial key performance indicator (KPI). In cellular networks, the achievable end-user throughput (the maximum throughput a user will get when attempting to download as much data as possible) is a spatiotemporal function, and its estimation poses a challenging and as-yet unsolved problem. The ability to accurately predict achievable throughput in a given location and time interval would, for example, allow mobile operators to further optimize their networks and design more personalized offers for the customers, or allow end-users with mobile broadband modems to make more informed decisions when selecting a provider. This work investigates the impact of individual parameters on the end-user achievable throughput in cellular networks and analyzes the feasibility and limitations of constructing realistic spatiotemporal KPI maps. To allow for the modeling of dependencies between individual KPIs, I conducted multiple static real-world and lab-measurement campaigns in LTE downlink and combined them with data provided by a mobile network operator. I propose and evaluate a model that estimates cell load from reference signal received quality (RSRQ), and another model that predicts reference signal-signal to interference plus noise ratio (RS-SINR) from a serving cell?s reference signal received power (RSRP) and interfering cells? RSRPs and cell loads. Further, I simulate the influence of the user population on resource sharing in the serving cell and its impact on the achievable throughput. End-to-end training of a throughput model that jointly considers all input parameters would be hindered by the ?curse of dimensionality.? Hence, I propose a separation into multiple sequential model components. Despite this separation, the individual components still require sufficient data from different times and locations to reconstruct realistic spatiotemporal maps. One possible approach for obtaining large amounts of heterogeneous data is crowdsensing?a method in which anonymous end-users contribute distributed measurements. However, my analysis shows that current crowdsensing solutions lack essential KPIs on which achievable throughput depends. I, therefore, focus in greater detail on the spatiotemporal modeling of RS-SINR. My temporal analysis of crowdsourced RSRP measurements confirms that RSRP can be modeled as time-invariant in a static location. RS-SINR spatiotemporal maps can therefore be obtained by fusing time-invariant RSRP maps and location-independent (if a specific cell is considered) cell-load time series. Because small-scale RSRP characteristics do not even out over time, mobile measurements are necessary to capture path loss and shadowing. Repeated drive tests are crucial for estimating the reliability of the measured RSRP. Due to GPS localization errors and varying speed profiles of the individual drives, the alignment of multiple repetitions introduces another challenge, which I solve by proposing a modification of the dynamic time warping (DTW) algorithm. The modified DTW allows for optimum alignment of the repeated measurements along a predefined path. Repeated DTW-aligned drive tests yield multiple samples per location. Therefore, I derive an efficient solution of Gaussian process regression (GPR) with repeated training locations, the computational complexity of which scales cubically with the number of distinct training locations, instead of the total number of training samples. Finally, I apply the proposed modified DTW and GPR to self-conducted drive tests and walk tests in an operational LTE network.