Traffic Distribution Generation Considering Spatial Continuity for Base Station Sleep Control

Base station (BS) sleep control with reinforcement learning (RL) offers a promising way to operate 5G radio access networks efficiently with low power consumption. For the optimal control considering the channel quality, an accurate and high-resolution spatial traffic distribution is required. In this study, we propose an effective method for generating a high-resolution spatial traffic distribution using throughput and load data per BS, data typically readily available to network operators. Our approach ensures the generated spatial traffic distribution aligns closely with the actual measured throughput and load per BS. However, a challenge arises when the available measurements are fewer than the unknown variables, leading to multiple possible solutions. To overcome this, we leverage the inherent spatial correlation of traffic amounts and assume the spatial continuity of traffic, thus constraining the range of possible solutions. Our simulation studies, conducted using traffic data from Milan, confirm that the generated spatial traffic distributions closely mirror the original ones. Specifically, we observe a 10.3 % improvement in the root mean squared error in the area experiencing the highest traffic volumes, compared to traditional method. Furthermore, we evaluate the performance of BS sleep control by RL, utilizing the generated spatial traffic distributions.