Deep Learning-Based Prediction of Resource Block Usage Rate for Spectrum Saturation Diagnosis

Strict restrictions on spectrum utilization and the rapid increases in mobile users have brought fundamental challenges for mobile network operators in securing sufficient spectrum resources. In designing reliable cellular networks, it is essential to predict spectrum saturation events in the future by analyzing the past behavior of base stations, especially their frequency resource block (RB) utilization states. This paper investigates a deep learning-based forecasting strategy of the future RB usage rate (RBUR) status of hundreds of LTE base stations deployed in Seoul, South Korea. The dataset consists of real measurement RBUR samples with a randomly varying number of base stations at each measurement time. This poses a difficulty in handling variable-length RBUR data vectors, which is not trivial for state-of-the-art deep learning estimation models, e.g., recurrent neural networks (RNNs), developed for handling fixed-length inputs. To this end, we propose a two-step RBUR estimation approach. In the first step, we extract a useful feature of the RBUR dataset that accurately approximates the behavior of the top quantile base stations. The feature parameters are carefully designed to be fixed-length vectors regardless of the dimensions of the raw RBUR samples. The fixed-length feature parameter vectors are readily exploited as the training dataset of RNN-based prediction models. Thus, in the second step, we propose a feature estimation strategy where the RNN is trained to predict the future RBUR from the input feature parameter sequences. With the estimated RBUR at hand, we can easily predict the spectrum saturation of the future LTE systems by examining the resource utilization states of the top quantile base stations. Numerical results demonstrate the performance of the proposed RBUR estimation methods with the real measurement dataset.