A Data-driven Base Station Sleeping Strategy Based on Traffic Prediction

Abstract

Due to the rapidly increasing number of deployed base stations (BSs) in current cellular networks, energy consumption has emerged as a great challenge in network operation. In this paper, we propose a novel data-driven intelligent BS sleeping mechanism based on a wireless traffic prediction model that measures the BSs' capacity in different regions. Firstly, a spatial-temporal traffic prediction model is proposed, where a multi-graph convolutional network (MGCN) is developed to capture spatial features, and a multi-channel long short-term memory (LSTM) involving short-term, daily, and weekly periodic data is used to capture temporal features. Secondly, the capacities of macro-cell BSs (MBSs) and small-cell BSs (SBSs) with different environment characteristics are modeled, where both clustering and transfer learning algorithms are adopted to quantify the traffic supported by MBSs and SBSs. Finally, an optimal BS sleeping strategy is proposed to minimize the network power consumption. Experimental results show that the proposed MGCN-LSTM model outperforms the existing models in terms of traffic prediction, and the proposed BS sleeping strategy using an approximated non-linear model of capacity function achieves a near-optimal energy-saving performance with relatively low complexity.

In-depth study of RNTI management in mobile networks: Allocation strategies and implications on data trace analysis

“…Fig. 20(c) considers user durations that fall within the bin of (8)(9)(10)(11)(12)(13)(14)(15)(16) minutes and analyzes their characteristics in terms of per user-load. The average load of T R extracted users is significantly higher than T L extracted ones, i.e., 830 kB vs. 32 kB.…”

“…When T L instead of T R is used, non-existent users are created, that have a smaller load value than the T R extracted ones. These errors can have substantial impact [0-1] min,TR [2][3][4] min,TR [5][6][7][8] min,TR [9][10][11][12][13][14][15][16] min,TR [17][18][19][20][21][22][23][24][25][26][27][28][29][30][31][32] min,TR min,TR [0-1] min,TL [2][3][4] min,TL [5][6][7][8] min,TL [9][10][11][12][13][14][15][16] min,TL [17][18][19][20]…”

“…Data-driven approaches have become more and more popular as techniques to advance mobile network technologies and components. Relevant examples encompass deployment planning [5] and automatic parameter configuration of newly deployed base stations [6], and optimizations at all the layers of the mobile network protocol stack, such as resource allocation in cloud Radio Access Networks (RAN) [7] and resource scaling of 5G core virtualized network functions [8], network slicing [9], energy savings with intelligent BS sleeping strategies [10], improved mobility management [11] and selfconfiguration of handover parameters [12], down to the physical layer to optimize beam management of radios operating at millimeter-wave frequencies [13]. The characterization of the user behavior is a key aspect to properly analyze traffic traces that are used in data-driven network optimization.…”

Dynamic Multi-Sleeping Control with Diverse Quality-of-Service Requirements in Sixth-Generation Networks Using Federated Learning

“…Neural networks can now predict non-critical stations for sleep initiation, offering a comprehensive solution for optimizing network energy efficiency [16]. Frameworks like MGCN-LSTM have improved traffic predictions [17], and Q-learning has been employed to balance load and energy efficiency [18]. Furthermore, a new data-driven framework has been proposed to manage base station sleep modes, satisfying both energy and quality-of-service (QoS) requirements, with simulations showing significant performance advantages [19].…”

Energy-Optimal RAN Configurations for SWIPT IoT

“…The first type (labeled LLP -low load proportionality) reflects the behavior of the majority of current stand-alone BSs, and it is characterized by a 27% load proportionality (with q 1 = 1100, q 2 = 100, and q 3 = 30). Conversely, the high load proportionality (HLP) BS type (with q 1 = 482.3, q 2 = 48.23, and q 3 = 144.69) corresponds to a 75% load proportionality, achievable, e.g., through time-domain duty-cycling at the sub-system level, i.e., through micro-sleep techniques involving modules of the BS or of the BBU in cloud-RAN designs [41]. For the sake of comparison, these parameters were chosen to fit a per-BS maximum consumed power of 1500 W, typical of stand-alone macro BSs [42].…”