Base station dormancy strategy based on user traffic prediction in heterogeneous cellular network

Abstract

The invention discloses a base station dormancy strategy of a double-layer heterogeneous cellular network composed of a macro base station and a micro base station. The method includes strategically integrating workloads distributed in the network onto fewer base stations. The core of the strategy is a bidirectional long and short term memory (BLSTM) neural network which is used for predicting future traffic of each user. The predicted traffic data provides information for a suggested base station dormancy policy that transfers a user connection from a micro base station with a lower utilization rate to other base stations, and then the idle micro base station is turned off while the macro base station remains in an on state all the time. When the strategy is executed, service quality maintenance can be fully considered, and it is ensured that the signal to interference plus noise ratio threshold value or the flow demand rate of each user cannot be reduced through switching.

Description

Technical Field

[0001] The invention relates to the field of mobile cellular networks, in particular to a method for predicting mobile cellular network traffic. Background Art

[0002] Mobile network services, such as mobile payments, online maps, and real-time video, have become an integral part of people's daily lives. With the widespread use of smartphones, tablets, wearable devices, and Internet of Things (IoT) applications, the number of devices connected to the network has increased significantly. At the beginning of 2022, the number of IoT devices reached 14.4 billion, an increase of 18% from 2021[1]. The growth of connected devices has led to a significant increase in data traffic. By the end of 2021, global mobile data traffic had reached approximately 84EB per month and is expected to increase by approximately 4.2 times to 368EB per month by 2027[2].

[0003] To keep up with the explosive growth of data traffic, the construction and deployment of fourth-generation and fifth-generation wireless networks have been promoted and popularized, and the number of base stations (BSs) as well as the capacity and coverage of mobile cellular networks have been improved. However, since information and communication technology (ICT) accounts for 1.8%-2.8% of global greenhouse gas emissions[3] and base stations consume more than 80% of the energy in mobile cellular networks[4], energy efficiency (EE) has become a critical issue. Optimal base station switching and base station sleep strategies[5],[6] can significantly improve the energy efficiency of mobile cellular networks, reduce greenhouse gas emissions, and save a lot of costs. Therefore, these strategies and solutions are crucial for the design of future mobile cellular networks.

[0004] However, the unprecedented increase in the number of mobile device connections in the network and the increasing diversity of data service types make the traditional homogeneous cellular network (HoCN) unable to meet the complex requirements of future networks. Homogeneous cellular networks are composed of base stations with similar working modes and transmission modes [7]. To solve this problem, heterogeneous cellular networks (HCN) are introduced to achieve flexible and low-cost deployment by adding micro base stations (MiBS) within the coverage area of ​​macro base stations (MaBS), such as pico base stations and femto base stations with different working modes and transmission modes. This approach improves data transmission rate and supplements coverage area, such as Figure 1 As shown in the figure, the carrier frequencies of micro base stations are different from those of macro base stations, and they will not interfere with macro base stations in the same coverage area.

[0005] In heterogeneous cellular networks, micro base stations can be shut down or switched to sleep mode when appropriate, greatly reducing the excessive energy consumption of base stations. This feature makes heterogeneous cellular networks more suitable for base station sleep than earlier, more rigid homogeneous cellular networks, because in heterogeneous cellular networks, users of shut-down micro base stations can connect to macro base stations covering their locations. In contrast, in homogeneous cellular networks, if a base station is shut down, its users may not be able to obtain the required quality of service (QoS) by connecting to active, distant base stations. The flexibility of heterogeneous cellular networks enables efficient network management and better utilization of network resources, making it a promising solution for future mobile networks.

[0006] Researchers have focused on analyzing and predicting data traffic in mobile cellular networks to find the best base station sleep schedule. Studies such as [8] and [9] have found that traffic in mobile cellular networks exhibits strong self-similarity and follows the general law of spatiotemporal distribution. The predictability of mobile cellular network traffic has made it a topic of exploration for researchers. In addition, there are significant differences in traffic patterns during the day and night on each weekday [8],

[10] . Therefore, by analyzing historical traffic data for a specific period of time in the same area, we can effectively predict the network traffic demand for a specific period of time in the future in the area based on the spatiotemporal distribution of user mobile data traffic.

[0007] In specific scenarios, traffic prediction methods for mobile cellular networks have been studied to schedule base station sleep in different time periods. In

[11] ,

[12] ,

[13] , base station sleep strategies based on predicted traffic were proposed to reduce energy consumption, but these strategies are all targeted at homogeneous cellular networks. For example, Lin et al.

[14] predicted data traffic in HeCN and proposed a base station sleep strategy based on the predicted traffic, but did not consider the differences between traffic demands of different users in the network and the impact of the locations of different base stations on the performance of the sleep strategy. Donevski et al.

[15] proposed a method based on neural network traffic prediction to determine the optimal base station sleep time in heterogeneous cellular networks, but did not specify user reallocation. Specifically, for the association between users and base stations, since the transmission power of macro base stations in heterogeneous cellular networks is much greater than that of micro base stations, users are more willing to choose macro base stations that can provide better service quality without special intervention, resulting in micro base stations not being fully utilized or even unloaded. Therefore, the base station sleep strategy must consider the impact of different base station deployment methods on the performance of the sleep strategy and control the user's choice so that the micro base stations in heterogeneous cellular networks can be fully utilized during the sleep period. These important improvements to base station sleep strategies will help achieve load balancing and energy saving among base stations.

[0008] References

[0009] The following references are cited herein by their numbers in parentheses. The entire contents of these references are hereby incorporated by reference for all purposes.

[0010] [1] M.Hasan. "State of IoT 2022: Number of connected IoT devices growing18% to 14.4billion globally." IoT Analytics.May 2022.Accessed:Mar.10,2023.[Online].Available:https: / / iot-analytics.com / numberconnected-iot-devices /

[0011] [2] R. "Ericsson mobility report." Ericsson.Rep.,Jun.2022.Accessed:Mar.10,2023.[Online].Available:https: / / www.ericsson.com / 49d3a0 / assets / local / reports-papers / mobility-report / do cuments / 2022 / ericsson-mobility-report-june-2022.pdf

[0012] [3] C. Freitag, M. Berners-Lee, K. Widdicks, B. Knowles, GS Blair, and A. Friday, "The real climate and transformative impact of ICT: A critique of estimates, trends, and regulations," Patterns, vol.2, no.9, 2021, Art.no.100340.

[0013] [4] T.Yang, F.Heliot, and CHFoh, "A survey of green scheduling schemes for homogeneous and heterogeneous cellular networks," IEEE Commun.Mag., vol.53, no.11, pp.175-181, Nov.2015.

[0014] [5]J.Wu,S.Zhou,and Z.Niu,“Traffic-aware base station sleeping controland power matching for energy-delay tradeoffs in green cellular networks,”IEEE Trans.Wireless Commun.,vol.12,no.8,pp.4196-4209,Aug.2013.

[0015] [6]Z.Niu,X.Guo,S.Zhou,and P.R.Kumar,“Characterizing energy-delaytradeoff in hyper-cellular networks with base station sleeping control,”IEEEJ.Sel.Areas Commun.,vol.33,no.4,pp.641-650,Apr.2015.

[0016] [7]S.Salim,C.H.Oey,and S.Moh,“Are heterogeneous cellular networkssuperior to homogeneous ones?”in Proc.9th IFIP Int.Conf.Netw.Parallel Comput.(NPC),2012,pp.61-68.

[0017] [8]U.Paul,A.P.Subramanian,M.M.Buddhikot,and S.R.Das,“Understandingtraffic dynamics in cellular data networks,”in Proc.30 th IEEEConf.Comput.Commun.(INFOCOM),2011,pp.882-890.

[0018] [9]E.M.Rezende Oliveira,A.C.Viana,K.P.Naveen,and C.Sarraute,“Measurement-driven mobile data traffic modeling in a large metropolitanarea,”in Proc.13th IEEE Int.Conf.Pervasive Comput.Commun.(PerCom),2015,pp.230-235.

[0019]

[10] Y.Wang,M.Faloutsos,and H.Zang,“On the usage patterns ofmultimodal communication:Countries and evolution,”in Proc.32 nd IEEEConf.Comput.Commun.(INFOCOM),2013,pp.3135-3140.

[0020]

[11] Q.Wu,X.Chen,Z.Zhou,L.Chen,and J.Zhang,“Deep reinforcementlearning with spatio-temporal traffic forecasting for data-driven basestation sleep control,”IEEE / ACM Trans.Netw.,vol.29,no.2,pp.935-948,Apr.2021.

[0021]

[12] G.Jang,N.Kim,T.Ha,C.Lee,and S.Cho,“Base station switching andsleep mode optimization with LSTM-based user prediction,”IEEE Access,vol.8,pp.222711-222723,2020.

[0022]

[13] R.Shinkuma,N.Kishi,K.Ota,M.Dong,T.Sato,and E.Oki,“Smarter basestation sleeping for greener cellular networks,”IEEE Netw.,vol.35,no.6,pp.98-103,Nov. / Dec.2021.

[0023]

[14] J.Lin,Y.Chen,H.Zheng,M.Ding,P.Cheng,and L.Hanzo,“A data-drivenbase station sleeping strategy based on traffic prediction,”IEEETrans.Netw.Sci.Eng.,early access,Sep.2,2021,doi:10.1109 / TNSE.2021.3109614.

[0024]

[15] I.Donevski,G.Vallero,and M.A.Marsan,“Neural networks for cellularbase station switching,”in Proc.38th IEEE Conf.Comput.Commun.Workshops(INFOCOM WKSHPS),2019,pp.738-743.

[0025]

[16] C.Liu,Y.Wan,L.Tian,Y.Zhou,and J.Shi,“Base station sleepingcontrol with energy-stability tradeoff in centralized radio access networks,”in Proc.34th IEEE Glob.Commun.Conf.(GLOBECOM),2015,pp.1-6.

[0026]

[17] M.Wildemeersch,T.Q.Quek,C.H.Slump,and A.Rabbachin,“Cognitivesmall cell networks:Energy efficiency and trade-offs,”IEEE Trans.Commun.,vol.61,no.9,pp.4016-4029,Sep.2013.

[0027]

[18] A.Ghosh et al.,“Heterogeneous cellular networks:From theory topractice,”IEEE Commun.Mag.,vol.50,no.6,pp.54-64,Jun.2012.

[0028]

[19] H.ElSawy,E.Hossain,and M.Haenggi,“Stochastic geometry formodeling,analysis,and design of multi-tier and cognitive cellular wirelessnetworks:A survey,”IEEE Commun.Surveys Tuts.,vol.15,no.3,pp.996-1019,3rdQuart.,2013.

[0029]

[20] W.Guo,S.Wang,X.Chu,J.Zhang,J.Chen,and H.Song,“Automated small-cell deployment for heterogeneous cellular networks,”IEEE Commun.Mag.,vol.51,no.5,pp.46-53,May 2013.

[0030]

[21] H.T.Cheng,A.Callard,G.Senarath,H.Zhang,and P.Zhu,“Step-wiseoptimal low power node deployment in LTE heterogeneous networks,”in Proc.76thIEEE Veh.Technol.Conf.(VTC Fall),2012,pp.1-4.

[0031]

[22] B.Yang,G.Mao,X.Ge,M.Ding,and X.Yang,“On the energy-efficientdeployment for ultra-dense heterogeneous networks with NLoS and LoStransmissions,”IEEE Trans.Green Commun.Netw.,vol.2,no.2,pp.369-384,Jun.2018.

[0032]

[23] Z.Mlika,E.Driouch,and W.Ajib,“Energy-efficient base stationoperation and association in HetNets:Complexity and algorithms,”IEEETrans.Wireless Commun.,vol.17,no.4,pp.2690-2702,Apr.2018.

[0033]

[24] S.Ak,H.Inaltekin,and H.V.Poor,“A tractable framework for theanalysis of dense heterogeneous cellular networks,”IEEE Trans.Commun.,vol.66,no.7,pp.3151-3171,Jul.2018.

[0034]

[25] C.S.Chen,V.M.Nguyen,and L.Thomas,“On small cell networkdeployment:A comparative study of random and grid topologies,”in Proc.76thIEEE Veh.Technol.Conf.(VTC Fall),2012,pp.1-5.

[0035]

[26] C.Coletti,P.Mogensen,and R.Irmer,“Deployment of LTE in-band relayand micro base stations in a realistic metropolitan scenario,”in Proc.74thIEEE Veh.Technol.Conf.(VTC Fall),2011,pp.1-5.

[0036]

[27] C.Coletti et al.,“Heterogeneous deployment to meet traffic demandin a realistic LTE urban scenario,”in Proc.76th IEEE Veh.Technol.Conf.(VTCFall),2012,pp.1-5.

[0037]

[28] L.Hu,I.Z.Kovács,P.Mogensen,O.Klein,and W.Stormer,“Optimal newsite deployment algorithm for heterogeneous cellular networks,”in Proc.74thIEEE Veh.Technol.Conf.(VTC Fall),2011,pp.1-5.

[0038]

[29] A.Cao,Y.Qiao,K.Sun,H.Zhang,and J.Yang,“Network traffic analysisand prediction of hotspot in cellular network,”in Proc.6th IEEE Int.Conf.Netw.Infrastruct.Digit.Content(IC-NIDC),2018,pp.452-456.

[0039]

[30] F.Xu et al.,“Big data driven mobile traffic understanding andforecasting:A time series approach,”IEEE Trans.Services Comput.,vol.9,no.5,pp.796-805,Sep. / Oct.2016.

[0040]

[31] K.Sultan,H.Ali,and Z.Zhang,“Call detail records driven anomalydetection and traffic prediction in mobile cellular networks,”IEEE Access,vol.6,pp.41728-41737,2018.

[0041]

[32] M.Zhang,H.Fu,Y.Li,and S.Chen,“Understanding urban dynamics frommassive mobile traffic data,”IEEE Trans.Big Data,vol.5,no.2,pp.266-278,Jun.2019.

[0042]

[33] M.Xu,Q.Wang,and Q.Lin,“Hybrid holiday traffic predictions incellular networks,”in Proc.16th IEEE / IFIP Netw.Oper.Manage.Symp.(NOMS),2018,pp.1-6.

[0043]

[34] J.Guo,C.Yang,and I.Chih-Lin,“Exploiting future radio resourceswith end-to-end prediction by deep learning,”IEEE Access,vol.6,pp.75729-75747,2018.

[0044]

[35] S.Hochreiter and J.Schmidhuber,“Long short-term memory,”NeuralComput.,vol.9,no.8,pp.1735-1780,1997.

[0045]

[36] H.D.Trinh,L.Giupponi,and P.Dini,“Mobile traffic prediction fromraw data using LSTM networks,”in Proc.29th IEEE Annu.Int.Symp.Pers.IndoorMob.Radio Commun.(PIMRC),2018,pp.1827-1832.

[0046]

[37] A.Azari,F.Salehi,P.Papapetrou,and C.Cavdar,“Energy and resourceefficiency by user traffic prediction and classification in cellularnetworks,”IEEE Trans.Green Commun.Netw.,vol.6,no.2,pp.1082-1095,Jun.2022.

[0047]

[38] S.Wu,Z.Zeng,and H.Xia,“Coalition-based sleep mode and powerallocation for energy efficiency in dense small cell networks,”IET Commun.,vol.11,no.11,pp.1662-1670,2017.

[0048]

[39] J.Kim,W.S.Jeon,and D.G.Jeong,“Effect of base station-sleepingratio on energy efficiency in densely deployed femtocell networks,”IEEECommun.Lett.,vol.19,no.4,pp.641-644,Apr.2015.

[0049]

[40] Y.Li,H.Celebi,M.Daneshmand,C.Wang,and W.Zhao,“Energy-efficientfemtocell networks:Challenges and opportunities,”IEEE Wireless Commun.,vol.20,no.6,pp.99-105,Dec.2013.

[0050]

[41] A.D.Wyner,“Shannon-theoretic approach to a Gaussian cellularmultiple-access channel,”IEEE Trans.Inf.Theory,vol.40,no.6,pp.1713-1727,Nov.1994.

[0051]

[42] L.Saker,S.-E.Elayoubi,R.Combes,and T.Chahed,“Optimal control ofwake up mechanisms of femtocells in heterogeneous networks,”IEEE J.Sel.AreasCommun.,vol.30,no.3,pp.664-672,Apr.2012.

[0052]

[43] G.Wu,G.Feng,and S.Qin,“Cooperative sleep-mode and performancemodeling for heterogeneous mobile network,”in Proc.IEEE WirelessCommun.Netw.Conf.Workshops(WCNCW WKSHPS),2013,pp.6-11.

[0053]

[44] J.Wu,E.W.Wong,Y.-C.Chan,and M.Zukerman,“Power consumption and GoStradeoff in cellular mobile networks with base station sleeping and relatedperformance studies,”IEEE Trans.Green Commun.Netw.,vol.4,no.4,pp.1024-1036,Dec.2020.

[0054]

[45] J.Wu,Y.Zhang,M.Zukerman,and E.K.-N.Yung,“Energy-efficient base-stations sleep-mode techniques in green cellular networks:A survey,”IEEECommun.Surveys Tuts.,vol.17,no.2,pp.803-826,2nd Quart.,2015.

[0055]

[46] P.Ren and M.Tao,“A decentralized sleep mechanism in heterogeneouscellular networks with QoS constraints,”IEEE Wireless Commun.Lett.,vol.3,no.5,pp.509-512,Oct.2014.

[0056]

[47] J.Kim,W.S.Jeon,and D.G.Jeong,“Base-station sleep management inopen-access femtocell networks,”IEEE Trans.Veh.Technol.,vol.65,no.5,pp.3786-3791,May 2016.

[0057]

[48] Y.S.Soh,T.Q.Quek,M.Kountouris,and H.Shin,“Energy efficientheterogeneous cellular networks,”IEEE J.Sel.Areas Commun.,vol.31,no.5,pp.840-850,May 2013.

[0058]

[49] H.Zhang,Y.Wang,H.Ji,and X.Li,“A sleeping mechanism forcacheenabled small cell networks with energy harvesting function,”IEEETrans.Green Commun.Netw.,vol.4,no.2,pp.497-505,Jun.2020.

[0059]

[50] A.M.Alqasir and A.E.Kamal,“Cooperative small cell HetNets withdynamic sleeping and energy harvesting,”IEEE Trans.Green Commun.Netw.,vol.4,no.3,pp.774-782,Sep.2020.

[0060]

[51] X.Li and C.Zhang,“Semi-dynamic Markov approximation-based basestation sleep with user association for heterogeneous networks,”IET Commun.,vol.17,no.6,pp.704-711,2023.

[0061]

[52] H.S.Dhillon,R.K.Ganti,F.Baccelli,and J.G.Andrews,“Modeling andanalysis of K-tier downlink heterogeneous cellular networks,”IEEE J.Sel.AreasCommun.,vol.30,no.3,pp.550-560,Apr.2012.

[0062]

[53] T.-C.Hou and V.Li,“Transmission range control in multihop packetradio networks,”IEEE Trans.Commun.,vol.34,no.1,pp.38-44,Jan.1986.

[0063]

[54] B.Matérn,Spatial Variation,vol.36.Berlin,Germany:Springer,2013.

[0064]

[55] M.E.Leinonen,M.Jokinen,N.Tervo,O.Kursu,and A. “Radiointeroperability in 5G and 6G multiradio base station,”in Proc.92nd IEEEVeh.Technol.Conf.(VTC-Fall),2020,pp.1-5.

[0065]

[56] A.Baddeley,I.Bárány,and R.Schneider,“Spatial point processes andtheir applications,”Stochastic Geometry(LNMCIME 1892).Berlin,Germany:Springer,2007,pp.1-75.

[0066]

[57] B.Boots,K.Sugihara,S.N.Chiu,and A.Okabe,Spatial Tessellations:Concepts and Applications of Voronoi Diagrams.New York,NY,USA:Wiley,2009.

[0067]

[58] B.Sklar,“Rayleigh fading channels in mobile digital communicationsystems.I.Characterization,”IEEE Commun.Mag.,vol.35,no.7,pp.90-100,Jul.1997.

[0068]

[59] F.Rashid-Farrokhi,K.R.Liu,and L.Tassiulas,“Transmit beamformingand power control for cellular wireless systems,”IEEE J.Sel.Areas Commun.,vol.16,no.8,pp.1437-1450,Oct.1998.

[0069]

[60] C.E.Shannon,“A mathematical theory of communication,”BellSyst.Technol.J.,vol.27,no.3,pp.379-423,1948.

[0070]

[61] G.Auer et al.,“How much energy is needed to run a wirelessnetwork?”IEEE Wireless Commun.,vol.18,no.5,pp.40-49,Oct.2011.

[0071]

[62] J.Schmidhuber,“Deep learning in neural networks:An overview,”Neural Netw.,vol.61,pp.85-117,Jan.2015.

[0072]

[63] Y.Bengio,P.Simard,and P.Frasconi,“Learning long-term dependencieswith gradient descent is difficult,”IEEE Trans.Neural Netw.,vol.5,no.2,pp.157-166,Mar.1994.

[0073]

[64] S.Hochreiter,“The vanishing gradient problem during learningrecurrent neural nets and problem solutions,”Int.J.Uncertainty FuzzinessKnowl.-Based Syst.,vol.6,no.2,pp.107-116,1998.

[0074]

[65] J.Hu,W.Heng,G.Zhang,and C.Meng,“Base station sleeping mechanismbased on traffic prediction in heterogeneous networks,”in Proc.Int.Telecommun.Netw.Appl.Conf.(ITNAC),2015,pp.83-87.

[0075]

[66] L.Li and W.Meng,“Collaborative base station sleeping solutiondesign in heterogeneous cellular network,”in Proc.27th Asia Pac.Conf.Commun.(APCC),2022,pp.231-235.

[0076]

[67] X.Chen,Y.Jin,S.Qiang,W.Hu,and K.Jiang,“Analyzing and modelingspatio-temporal dependence of cellular traffic at city scale,”in Proc.25 th IEEE Int.Conf.Commun.(ICC),2015,pp.3585-3591.

[0077]

[68] MathWorks.“Regression output layer-MATLAB regressionLayer.”2017.Accessed:Mar.10,2023.[Online].Available:https: / / www.mathworks.com / help / deeplearning / ref / regressionlayer.html

[0078]

[69] S.Siami-Namini, N.Tavakoli, and ASNamin, "Acomparison of ARIMAandLSTM in forecasting time series," in Proc.17th IEEE Int.Conf.Mach.Learn.Appl.(ICMLA), 2018, pp.1394-1401.

[0079]

[70] S. Siami-Namini, N. Tavakoli, and ASNamin, "Acomparative analysis of forecasting financial time series using ARIMA, LSTM, and BiLSTM," 2019, arXiv:1911.09512.

[0080]

[71] PTYamak, Y.Li, and PKGadosey, "Acomparison between ARIMA, LSTM, and GRU for time series forecasting," in Proc.2nd Int.Conf.Algorithms, Comput.Artif.Intell., 2019, pp.49-55. Summary of the invention

[0081] Therefore, one aspect of the present invention is to provide a method for managing base stations in a dual-layer heterogeneous cellular network. The heterogeneous cellular network includes multiple macro base stations and multiple micro base stations. The above method includes the following steps: a) predicting the future traffic of users of the heterogeneous cellular network; b) switching the user's connection from a base station with low utilization among the multiple micro base stations to one or more of the multiple micro base stations or multiple macro base stations; and c) shutting down the base station with low utilization among the multiple micro base stations.

[0082] Preferably, step a) is performed using a bidirectional long short-term memory neural network (BLSTM).

[0083] More preferably, the bidirectional long short-term memory neural network comprises two layers, one of which is used to transmit information in chronological order, and the other layer is used to transmit information in reverse chronological order.

[0084] In a variant of a preferred embodiment, in step b), the QoS requirements of the user are guaranteed.

[0085] Preferably, in step b), the user's signal to interference plus noise ratio (SINR) threshold is not violated.

[0086] More preferably, step b) further includes: e) identifying a macro base station with the highest signal to interference plus noise ratio from multiple macro base stations; and f) reallocating user connections from underutilized micro base stations among multiple micro base stations to the macro base station with the highest signal to interference plus noise ratio.

[0087] Alternatively or additionally, step b) further includes: g) identifying a second micro base station with the highest load from multiple micro base stations; and reallocating the user's connection from a micro base station with low utilization among the multiple micro base stations to the second micro base station.

[0088] In a variant of a preferred embodiment, during step b), the user's maximum flow demand rate is always met.

[0089] In a variation of a preferred embodiment, step b) further includes: i) identifying a third base station from a plurality of macro base stations and a plurality of micro base stations that can meet the maximum traffic demand rate during low load periods; and reallocating the user's connection from a micro base station with low utilization among the plurality of micro base stations to the third base station.

[0090] In a variation of a preferred embodiment, the method further comprises modeling the deployment of the plurality of micro base stations and the plurality of macro base stations as a random point process.

[0091] Preferably, the random point process includes a Poisson point process (PPP) and a Markov hard core point (MHCPP) process.

[0092] In a variation of a preferred embodiment, none of the plurality of macro base stations are powered off at any time.

[0093] In a variation of a preferred embodiment, the method further comprises repeating steps a)-c) for different users.

[0094] In a variation of a preferred embodiment, the method further comprises repeating steps a)-c) until there is no more macro base stations that can be shut down.

[0095] According to another aspect of the present invention, a non-transitory computer-readable storage recording medium is provided, on which computer instructions are recorded. When executed on one or more processors, the computer instructions cause the one or more processors to perform operations according to any of the above methods or variations thereof.

[0096] According to another aspect of the present invention, there is provided a computing system comprising one or more processors and a memory, wherein the memory contains instructions. When executed on the one or more processors, the instructions cause the computing system to perform operations according to any of the above methods or variations thereof.

[0097] Therefore, it can be seen that the embodiments of the present invention are more efficient and superior to other baseline strategies. Recognizing that real-time traffic in a cellular network fluctuates and often exhibits tidal patterns, such as diurnal traffic, the present invention exploits this property to minimize energy consumption. The method involves strategically consolidating workloads dispersed across the network onto fewer base stations.

[0098] The above summary of the invention is neither intended to define the invention of the present application nor to limit the scope of the invention in any way. BRIEF DESCRIPTION OF THE DRAWINGS

[0099] The above and further features of the present invention will be clearly described in the following description of embodiments, which are only provided as examples in conjunction with the following appendix. Figure 1 From now on:

[0100] Figure 1 An architecture for a heterogeneous cellular network is presented.

[0101] Figure 2a Demonstrated base station deployment via PPP in a 10km×10km area, with a macro base station strength of 0.2 points / km 2 , while the micro base station strength is 0.4 points / km 2 .

[0102] Figure 2b Shows the Figure 2a Base station deployment implemented by PPP and MHCPP in the PPP, where the hard-core parameters of macro base stations and micro base stations are r h They are set to 1km and 2km respectively.

[0103] Figure 3a is a block diagram of the power consumption of a macro base station in active mode.

[0104] Figure 3b This is a block diagram of the power consumption of a macro base station in sleep mode.

[0105] Figure 4a Showing the repeating module of a standard recurrent neural network (RNN).

[0106] Figure 4b Showing the repeating module of a standard Long Short-Term Memory (LSTM).

[0107] Figure 5The architecture of BLSTM is described according to an embodiment of the present invention.

[0108] Figure 6 The Minimum Load First Sleep (MiLSF) strategy in a two-layer heterogeneous cellular network is shown. * The algorithm is as follows.

[0109] Figure 7a The figure shows the hourly average traffic prediction of user 1 by BLSTM, with MAE = 75.56 and RMSE = 98.32.

[0110] Figure 7b The figure shows the hourly average traffic prediction of user 2 by BLSTM, with MAE = 55.78 and RMSE = 75.17.

[0111] Figure 7c The figure shows the hourly average traffic prediction of user 3 by BLSTM, with MAE = 99.26 and RMSE = 115.23.

[0112] Figure 8a The figure shows the RNN's prediction of the average hourly traffic of user 1, with MAE = 87.89 and RMSE = 135.52.

[0113] Figure 8b The figure shows the RNN's prediction of the average hourly traffic of user 2, with MAE = 68.07 and RMSE = 118.93.

[0114] Figure 8c The figure shows the RNN's prediction of the average hourly traffic of user 3, with MAE = 104.99 and RMSE = 179.61.

[0115] Figure 9a The ARIMA prediction for the average hourly traffic of user 1 is shown, with MAE = 257.931 and RMSE = 460.085.

[0116] Figure 9b The hourly average traffic forecast for user 2 is shown, with MAE = 221.24 and RMSE = 339.779.

[0117] Fig.9c The hourly average traffic forecast for user 3 is shown, with MAE = 221.132 and RMSE = 332.755.

[0118] Fig.10 The energy-saving performance of MiLSF and four benchmark strategies under PPP and MHCPP conditions is compared.

[0119] Fig.11 The performance comparison of MiLSF and four baseline strategies under different numbers of users is shown.

[0120] Fig.12 The performance comparison of MiLSF and four baseline strategies at different SINR thresholds γ0 is shown.

[0121] Fig.13 The performance comparison of MiLSF and four baseline strategies under different numbers of dormant micro base stations is shown. DETAILED DESCRIPTION

[0122] Real-time traffic in cellular networks varies over time, often showing tidal patterns, such as day and night traffic patterns. This property can be exploited to consolidate workloads dispersed throughout the network onto fewer base stations, thereby reducing the energy consumption of the cellular network. In one embodiment, a base station sleep strategy is proposed for a two-layer heterogeneous cellular network consisting of macro base stations and micro base stations. The sleep strategy includes consolidating the workload dispersed in the network onto fewer base stations. A BLSTM neural network is used to predict the future traffic of each user. The predicted traffic data provides information for the base station sleep strategy, which transfers the user's connection from the underutilized micro base stations to other base stations, and then shuts down the idle micro base stations, while the macro base stations are always turned on. When executing this strategy, full consideration is given to maintaining the quality of service, ensuring that the switch does not reduce the SINR threshold or traffic demand rate of each user.

[0123] Specifically, in an exemplary embodiment, a BLSTM neural network is used to predict the future traffic of each user. Based on the predicted traffic, the base station sleep strategy switches the user's connection from the underutilized micro base station to other base stations, and then shuts down the idle micro base station. However, the macro base station is never shut down. All user connections have a predefined SINR threshold, and the service quality of each user can be guaranteed, which is related to the user's traffic demand rate, and the service quality will not be reduced when switching the user's connection. Through a large number of numerical simulation experiments, the effectiveness and superiority of the base station sleep strategy of the embodiment of the present invention relative to the other four benchmarks are proved. In different scenarios, the base station sleep strategy is much better than these four benchmarks in different scenarios, demonstrating its potential to significantly reduce the energy consumption of cellular networks without affecting the quality of service.

[0124] In the embodiments of the present invention, the main advantages of the base station sleep strategy in the heterogeneous cellular network include but are not limited to the following aspects:

[0125] First, we use BLSTM to predict the traffic demand rate of each user in a heterogeneous cellular network, rather than predicting the traffic of the entire network or a single base station as in the prior art. We use real historical traffic datasets to train and test the prediction model. The predicted traffic is then used as input for the base station sleep strategy.

[0126] Second, based on random geometry, the base station deployment in heterogeneous cellular networks is abstracted as a random point process. Specifically, a two-layer heterogeneous cellular network consisting of macro base stations and micro base stations is considered, and PPP and MHCPP are used to simulate the positioning of base stations, respectively. Then, the energy-saving performance of the two-layer heterogeneous cellular network is evaluated under different base station deployment methods (i.e., PPP and MHCPP).

[0127] Third, based on the predicted user traffic demand and base station deployment results, a base station sleep strategy, namely MiLSF strategy, is designed for the dual-layer heterogeneous cellular network. This strategy attempts to reallocate users of micro base stations in the network to other base stations and switch idle micro base stations to sleep mode to minimize the overall energy consumption of the entire network. Each user's connection is considered to have a predefined SINR. By ensuring that the SINR does not violate a given threshold when switching from one base station to another, the QoS requirement of each user can be guaranteed, which is related to the user's required traffic rate. The goal is to optimize the set selection of micro base stations, where these micro base stations switch to sleep mode at the beginning of the low load period and then switch back to active mode after the low load period ends. In practice, the low load period is usually at night. This design avoids frequent base station switching changes, which is problematic because initialization is required when switching from sleep mode to active mode. Note that the macro base stations in the network will not switch to sleep mode, and the MiLSF strategy is implemented so that the QoS of all users will not be degraded after the reallocation of users.

[0128] MiLSF strategies effectively reduce energy consumption and operating costs of heterogeneous cellular networks. This in turn helps reduce carbon emissions, making it an environmentally friendly solution. When combined with artificial intelligence techniques to predict future traffic, MiLSF strategies can help ensure QoS by proactively avoiding violations of the SINR thresholds provided to each user. This proactive approach helps prevent potential client losses and revenue losses for mobile operators, thus making it a highly desirable solution for improving network performance.

[0129] Deploying micro base stations in macro base stations can improve user satisfaction, network capacity, network coverage, and reliability of heterogeneous cellular networks. However, existing green cellular network research does not consider the mobility of users and the changes in traffic load generated by these users over time. On the other hand, traffic prediction is usually achieved by analyzing network traffic patterns and extracting traffic change characteristics. Since cellular network traffic is usually presented in the form of time series, time series modeling methods can describe and predict the changes in cellular network traffic. In addition, the emergence of many neural network methods has provided new ideas for modeling and predicting cellular network traffic. LSTM is a special RNN that successfully solves the gradient vanishing problem and easily learns long-term dependency information. Specifically, LSTM feeds the output of the previous step to the input layer of the current step, which is a dynamic feedback connection used to model dependencies in time series. However, existing publications focus more on traffic prediction of base stations or the entire network without considering the traffic demand of individual users.

[0130] In heterogeneous cellular networks, micro base stations are usually subject to a sleep strategy to reduce the overall energy consumption of the network, given their high flexibility and ease of activation or deactivation. The base station sleep strategy must ensure that the active base stations can withstand the network traffic load while reducing energy consumption. The dynamic changes in network traffic load are an important reference for formulating base station sleep strategies. The sleep of base stations needs to ensure the user's quality of service and the user's association with the remaining active base stations.

[0131] In summary, the research on base station sleep strategies mainly considers the traffic load, resource allocation and user association of the base station. Researchers use greedy algorithms, heuristic algorithms or game theory to obtain energy-saving sleep strategies to maintain user satisfaction. At present, most base station sleep strategies are based on the traffic prediction of the base station, rather than on individual users with different traffic requirements in the network. Traffic prediction for different users helps to formulate base station sleep strategies that accurately meet the needs of different users. At the same time, the impact of different deployment methods of macro base stations and micro base stations in heterogeneous networks on user association and energy consumption has not been deeply studied. In an embodiment of the present invention, based on the traffic data of each user with different traffic patterns in the BLSTM prediction network, the MiLSF strategy selectively puts some micro base stations into a sleep state and considers user association issues, thereby reducing energy consumption in the network without reducing the quality of service for each user.

[0132] System Model

[0133] The introduction of different types of base stations in heterogeneous cellular networks leads to the development of a multi-layer network structure, which includes multiple types of base stations. In an exemplary embodiment of the present invention, a two-layer heterogeneous cellular network consisting of a macro base station and a micro base station is considered. The following will introduce the two-layer heterogeneous cellular network and the BLSTM-based user traffic prediction model in detail.

[0134] A. Base station deployment

[0135] Compared with the traditional cellular topology (in which macro base stations use a hexagonal grid model and micro base stations are deployed in a specific way within the coverage area of ​​the macro base stations), random geometry theory provides an effective and tractable method to study the performance of heterogeneous cellular networks from a statistical perspective

[19] ,

[52] . Specifically, PPP is an tractable and popular point process due to its independence

[53] . PPP can be used to simulate the process of base station deployment in cellular networks, while MHCPP

[54] is based on PPP but avoids the situation where the points are too dense, which can better reflect the actual base station deployment and describe the network more realistically. In an exemplary embodiment, PPP and MHCPP are considered to be deployed on a two-dimensional Euclidean plane.

[0136] PPP: If any two-dimensional compact set The number of base stations in is a Poisson random variable, then the two-dimensional Euclidean plane The base station Φ in i ; i = 1, 2, ...} can be modeled as PPP, where B i is the i-th base station. Specifically, macro base stations and micro base stations are deployed by applying different strength values ​​on PPP. Note that the deployment of macro base stations and micro base stations is independent of each other.

[0137] MHCPP: MHCPP is based on PPP, but by removing the predefined non-negative distance (also called hard core parameter r h ) to avoid interference between very close base stations, i.e.

[0138] Specifically, for any two pairs whose distance from each other is less than r h Base station B i and B j For example, if i is less than j, remove the base station with a smaller index. i .

[0139] Figure 2a-2b The PPP and MHCPP for implementing base station deployment are demonstrated. The deployment process of macro base stations and micro base stations is independent of each other. Since macro base stations have higher transmission power and larger coverage area, in actual deployment, the number of macro base stations is less than that of micro base stations. Figure 2a As can be seen in Figure 1, base stations of the same type may be located in very close proximity, resulting in considerable signal interference. Figure 2b In the example above, this situation is improved by removing base stations of the same type that are too close to each other. It should be noted that for the sake of convenience, only the signal interference between base stations of the same type is considered here. This is because, as mentioned above, macro base stations and micro base stations have completely different carrier frequencies, so the interference between base stations of different types can be ignored compared to the interference between base stations of the same type.

[0140] However, it is worth mentioning that for the actual situation of high-density base station networks of 5G or future 6G, the increase in carrier frequency and the shortening of base station wavelength accelerate the attenuation of signal propagation. Compared with 4G base stations, the coverage range of 5G and 6G base stations is greatly reduced

[55] . These characteristics indicate that while increasing the base station density in the MHCPP model, a smaller hard core constraint distance can be adopted, making the MHCPP model scalable in the actual scenario of high-density networks of 5G or even 6G architectures.

[0141] Another limitation of the MHCPP model is that it assumes a specific form of spatial dependency characterized by a correlation structure

[54] ,

[56] , which may not fully capture the complex spatial dependencies present in real-world scenarios. To address this limitation, other models, such as Voronoi Tessellation

[57] and network-based models, provide greater flexibility in capturing complex spatial dependency patterns. Voronoi tessellation divides the space into regions around each base station, ensuring that any point in each region is closer to its corresponding base station than any other point

[57] . This model can effectively capture the spatial interactions between base stations and more accurately represent coverage areas and interference patterns. It takes into account the distance between users and individual base stations and can analyze signal strength and interference levels. Network-based models explicitly consider the underlying network structure and connectivity between base stations. Graph-based models, such as random geometric graphs or random connectivity models, can capture the spatial proximity of base stations and their connectivity patterns. These models are particularly useful for simulating wireless networks with interference, signal propagation, and connectivity limitations.

[0142] B. SINR Model

[0143] In the exemplary embodiment, it is assumed that all homogeneous base stations have the same characteristics. Macro base stations and micro base stations differ in deployment strength, transmit power, number of transmit antennas included, carrier frequency and bandwidth.

[0144] Consider K users, M macro base stations and N micro base stations in a two-layer heterogeneous cellular network, where:

[0145] The transmit power of a single transmit antenna in a macro base station and a micro base station is represented by p1 and p2 respectively;

[0146] The carrier frequencies of macro base stations and micro base stations are and express;

[0147] The bandwidths of the macro base station and micro base station are represented by w1 and w2 respectively.

[0148] For ease of expression, define the set Ψ M = {1, 2, ..., M} and ΨS = {M+1, M+2, ..., M+N, , the macro base station and micro base station in the two-layer heterogeneous cellular network are marked as B i ,i∈Ψ M and B i ,i∈Ψ S Assume K ={1,2,..., K is the user set in the dual-layer heterogeneous cellular network. Since macro base stations and micro base stations have different characteristics (for example, they use different carrier frequencies and different bandwidths), a binary indicator θ(x) is introduced to clarify this, which indicates which layer the base station belongs to, and is defined as:

[0149]

[0150] Consider the base station i∈Ψ M ∪Ψ S , user k∈Ψ K , its path loss (dB) is obtained as follows:

[0151]

[0152] Where c is the speed of light, β is the path loss exponent, is the carrier frequency of base station i, and d i，k is the distance between base station i and user k. Then the received SINR from base station i to user k is (3)

[0153]

[0154] Among them, p θ(i) and w θ(i) are the transmission power and bandwidth of base station i respectively. η0 represents the noise spectral density. ρ i，k is the small-scale fading between user k and base station i, where Rayleigh fading

[58] that follows an exponential distribution is considered. In this case, if the received SINR is lower than the threshold γ0, user k will be shielded by base station i. The numerator (i.e. ) is the signal power received by user k from the associated base station i. The denominator consists of interference power and white noise power. Interference power (i.e. )) is the signal power received by user k from other base stations (i.e., base stations other than base station i).

[0155] To improve SINR performance, different interference suppression strategies are used in heterogeneous cellular networks, such as beamforming techniques

[59] . In beamforming, an antenna array is used instead of a single antenna to transmit or receive signals. Due to different signal paths, the signals received by multiple antennas have different phases (and amplitudes). By introducing appropriate delays, different signals can enhance each other, thereby achieving a channel gain that is much greater than that achievable with a single antenna. In addition, other techniques, such as power control and resource allocation, can also be used to mitigate interference and improve signal quality. However, according to an exemplary embodiment, a base station sleep strategy is a strategy that effectively saves energy while ensuring quality of service. This paper considers a simple and general SINR model, but does not delve into detailed interference management techniques. These mitigation techniques can be combined with the base station sleep strategy by providing different SINR thresholds to the algorithm.

[0156] C. Base station power consumption model

[0157] For user k connected to base station i whose SINR is not less than the threshold γ0, that is, S i，k ≥γ0, according to the Shannon-Hartley theorem

[60] , the bandwidth of base station i required by user k at time t is:

[0158]

[0159] where r k (t) is the required rate of user k at time t. Formula (4) describes the relationship between the required bandwidth resources and the user's required traffic rate. Note that the expression S in formula (4) i，k is a linear power ratio, not a logarithmic decibel. Let the set of users connected to base station i at time t be Then the load of base station i at time t is

[0160]

[0161] Note that since no base station is allowed to be overloaded, μ i (t)∈[0,1]. For the convenience of study, two modes of the base station are considered: sleep mode and active mode, respectively. Figure 3a and Figure 3b Here are some definitions.

[0162] · and They are defined as the sleep power of macro and micro base stations, respectively. The sleep power is considered to be independent of time and load, because a base station in sleep state does not bear any traffic load.

[0163] For activity mode, and are defined as the active power of macro base stations and micro base stations, respectively. The power of an active base station can be divided into two parts, namely, transmission power and circuit power

[61] . The former is related to the load of the base station, while the latter contains the energy consumption of necessary circuit components and can be regarded as a constant. Specifically, and are defined as the circuit power of the macro base station and the micro base station, respectively. Assume that each macro base station and micro base station consists of multiple transmitting antennas, and the number of transmitting antennas of each macro base station and micro base station is α1 and α2, respectively. In this case, the linear approximation model of

[61] is used to represent the total power of base station i at time t, as shown below.

[0164]

[0165] D. BLSTM model

[0166] The traffic patterns of users in cellular networks vary over time

[30] . In recent years, RNNs have gained widespread attention in processing sequence data as input and performing recursion along the direction of sequence evolution

[62] . In RNNs, all nodes (i.e., artificial neurons) are interconnected so that the output of some nodes can affect the subsequent input of the same node. However, for very long sequences, even with careful parameter selection, RNNs are generally unable to handle long-term dependencies in reality

[35] .

[0167] LSTM is a type of RNN designed to handle long-term dependencies that standard RNNs cannot handle

[35] . LSTM allows the model to learn how to forget previous hidden states and update the current state, making it well suited for capturing long-term temporal dependencies in sequences. The LSTM architecture includes forget, input, and output gate mechanisms that enable the model to selectively forget or retain information. Figure 4a-4b shows the standard RNN (see Figure 4a ) and the standard LSTM ( Figure 4b ) in terms of the repeated model. At time t, They represent input, internal state, candidate state, hidden state, forget gate output, input gate output, and output gate output respectively. The gate structure includes sigmoid and tanh activation functions represented by σ(·) and τ(·) respectively. The sigmoid function compresses the input to between 1 and 0, while the tanh function compresses the input to between 1 and -1. This property enables LSTM to forget information by multiplying zeros and remember some or all of the information by multiplying non-zero numbers.

[0168] For the repetitive model at time t described in Figure 4(b), LSTM introduces the internal state Used to linearly transmit cyclic information and nonlinearly output cyclic information to hidden states Internal state c t and the hidden state h t Obtained by the following formula:

[0169]

[0170] h t =o t ☉τ(c t ), (8)

[0171] Among them, ⊙ is the point-by-point multiplication operation, is the candidate state obtained by the following formula:

[0172]

[0173] For the outputs of the three gates, the following are provided:

[0174] The output of the forget gate is f t ∈[0, 1] D , used to control the last cell state c t -1 The amount of information that needs to be forgotten,

[0175] f t =σ(W f x t +U f h t-1 +v f ) (10)

[0176] Input gate output i t ∈[0, 1] D , used to determine the amount of information that needs to be remembered for the candidate state

[0177] i t =σ(W i x t +U i h t-1 +v i ), (11)

[0178] And the output gate output o t ∈[0, 1] D , which determines the internal state c t How much information needs to be output to the hidden state h t ,

[0179] o t =σ(W o x t +U o h t-1 +v o ), (12)

[0180] Where W * ,U * , and v * ,U * *∈{i,f,o,c} are learnable network parameters. In short, formulas (9)-(12) can be described as

[0181]

[0182] in, is the input at time t, and are the network parameters of LSTM. Formula (13) together with (7)-(8) fully describes the repetitive model of the standard LSTM in Figure 4(b)

[0183] While the standard LSTM described above has been widely used in the existing literature, not all LSTMs are the same. Different publications may use slightly different versions of the LSTM network, and the temporal output may be related to past and future information. In an exemplary embodiment of the present invention, a BLSTM is used to enhance the ability of the LSTM network to more accurately predict user traffic data. A specific network layer is added that transmits information in reverse time order, allowing the BLSTM to handle past and future context when predicting. This modification improves the BLSTM's ability to capture temporal correlations in the data and produces more accurate predictions than the standard LSTM. See Figure 5 . Figure 5 The two layers in have different information transmission directions, where the first layer is in chronological order and the second layer is in reverse chronological order. Figure 5 The internal state outputs of the two layers in and hidden state output Compared with the standard LSTM t and h t The same can be obtained by formulas (7), (8) and (13). Then, the internal state output and hidden state output of BLSTM at time t can be obtained by the following formulas:

[0184]

[0185]

[0186] in, It is a concatenation operation.

[0187] The computational complexity of the BLSTM model depends on factors such as the length of the input sequence, the number of model layers, and the number of LSTM units in each layer. For long input sequences and models with multiple layers and many LSTM units, the time complexity of the BLSTM model can be high. It is worth mentioning that there are some techniques to reduce this complexity. For example, during the training process, truncated backpropagation through time (TBPTT) can be used to limit the number of time steps considered, thereby reducing the overall computational burden

[63] ,

[64] .

[0188] Compared with other types of RNN, the BLSTM model also has some disadvantages, such as computational complexity, higher memory requirements, risk of overfitting, and sensitivity to hyperparameters. However, considering that traffic prediction accuracy is very important in technical problems and may have a significant impact on the performance of the sleep strategy, the BLSTM model is selected in this example embodiment, which performs better than RNN on the given data set. The accuracy comparison of the BLSTM and RNN models on the given data set shows the superiority of the BLSTM model, which will be described in detail below.

[0189] Problem Description

[0190] The goal is to reduce the total energy consumption of all base stations in a two-layer heterogeneous cellular network by selectively putting micro base stations into sleep mode during periods of low network load. To achieve this goal, three constraints need to be met: first, ensure that the SINR of all users remains above a threshold γ0; second, ensure that the QoS of each user does not degrade, which is closely related to the demand traffic; third, ensure that the load of each base station does not exceed 1.

[0191] The motion vector It is defined as the state vector associated with the policy Φ, that is

[0192] ·if Then base station i at time t is in active mode;

[0193] Otherwise, base station i at time t is in sleep mode.

[0194] In this case, the set of all active base stations and dormant base stations is and set up is the set of maximum user demand rates in the low-load time period [T1, T2]. In order to mathematically define the feasibility of strategy φ in the low-load time period [T1, T2], the following constraints are introduced on the action variables:

[0195]

[0196]

[0197] Formula (16) ensures that each user in the network can access at least one active base station, where the active base station can maintain an SINR above the threshold γ0 while also meeting the user's maximum traffic demand rate during the low load period [T1, T2].

[0198] The different traffic demands of each user connected to an active base station together constitute the load on the base station. Since no base station is allowed to be overloaded, therefore:

[0199]

[0200] Recall that only dormant operations of micro base stations are considered, given their high flexibility and ease of activation and deactivation. This can be solved by the following formula:

[0201]

[0202] The goal is to minimize the energy consumption of base stations in a two-layer heterogeneous cellular network during the low-load period [T1, T2]. Specifically, let

[0203]

[0204] t∈[T1,T2], (19)

[0205] is the total power of all base stations in the two-layer heterogeneous cellular network at time t, then the optimization problem is:

[0206]

[0207] It depends on formulas (16), (17) and (18). Let Φ denote the set of all policies subject to formulas (16), (17) and (18).

[0208] MiLSF Strategy

[0209] This section introduces the user traffic demand r based on BLSTM prediction. k (t) and base station load μ i(t). Please note that due to the high flexibility and easy operation of micro base stations, only their sleep operation is considered. In addition, frequent switching of the state of the base station will cause additional network delays and additional base station initialization energy consumption. Therefore, for the low load time period [T1, T2], only one sleep operation of the micro base station is considered, which occurs at the beginning time of the low load period, that is, T1. Before T1, it is assumed that all M macro base stations and N micro base stations are in active mode. Since the low load period is usually at night, there are fewer human activities and the user's mobility is also low. For simplicity, it is assumed that the mobile terminal or user is stationary during this period.

[0210] The detailed implementation process of the MiLSF strategy is as follows.

[0211] a) Sorting: According to the load situation at time T1, the active micro base stations are sorted from low to high.

[0212] b) Sleep. According to the sorting order of step a), each micro base station is evaluated to determine whether it can be switched to sleep mode. If there is an active base station (including macro base stations and micro base stations) that can meet the maximum traffic demand rate of each user currently connected to the micro base station during the low load period [T1, T2], the micro base station currently connected can be switched to sleep mode and then enter step c). Otherwise, the method will continue to evaluate the next micro base station until a micro base station that can be switched to sleep mode is found.

[0213] c) Reassignment of users. Active base stations including macro base stations and micro base stations are selected, which can meet the maximum traffic demand rate of each user during the low load period [T1, T2] of the micro base station switched to sleep mode in step b). Then, the users are reallocated to the selected active base stations. Each user is preferentially reallocated to the macro base station with the highest SINR among the selected base stations. If there is no qualified macro base station among the selected base stations, the user is reallocated to the micro base station with the highest load among the selected base stations.

[0214] d) Repeat. Repeat steps a) and b) until no more micro base stations can be switched to sleep mode.

[0215] In step c), since the goal is to put as many micro base stations as possible into sleep mode to save energy without reducing the user service quality, and not to put any macro base station into sleep mode, priority is given to the active macro base station with the highest SINR among the selected base stations. Since all macro base stations have the same power consumption characteristics, assigning users to the macro base station with the highest SINR can achieve the best spectrum efficiency.

[0216] Assume that there is no active macro base station among the base stations selected in step c). In this case, the user will be reallocated to the micro base station with the largest load, which helps to speed up the execution process and reduce complexity, because the method always tries to put the micro base station with the smallest load into sleep state first in each cycle of the above execution.

[0217] Figure 6 The algorithm shown in the figure proposes the MiLSF strategy φ * If argmax returns multiple arguments, then the tie rules in lines 8 and 20 of the algorithm can be arbitrary, and card(·) returns the cardinality of the set.

[0218] Numerical results

[0219] This section numerically demonstrates the effectiveness of the MiLSF strategy by comparing it with the following four other baseline strategies in different scenarios.

[0220] 1) Random Sleep (RS). Randomly select active micro base stations (see step b) in the “MiLSF Strategy” section above) and try to put them into sleep mode until there are no remaining micro base stations that can be switched to sleep mode. The user reallocation process in RS is the same as that in MiLSF (see

[0221] Step c) in the “MiLSF strategy” section.

[0222] 2) Random reallocation of users (RRU). The selection process of active micro base stations is the same as that of active micro base stations in MiLSF (see step b) in the "MiLSF Strategy" section), but under RRU, users will be randomly reallocated to active base stations that meet the SINR threshold and can guarantee the user traffic demand rate, which is different from step c) in MiLSF.

[0223] 3) Closest User Reallocation (CUR)

[65] . The CUR strategy always reallocates the user to the nearest available base station (including macro base stations and micro base stations) that can guarantee the maximum traffic demand rate of the user with an SINR greater than the threshold during low load periods, while the selection of micro base stations follows the same principle as the MiLSF strategy (see step b) in the “MiLSF strategy” section).

[0224] 4) Closest Base Station Priority Sleep (CBSSF). In the CBSSF strategy, the sleep of the micro base station closest to the macro base station is given priority, and the user reallocation process is the same as MiLSF (see step c) in the "MiLSF strategy" section). 5)

[0226] In Part A of this section, we will introduce the traffic data source and demonstrate the use of BLSTM to predict user traffic demand rate. In Part B of this section, we will demonstrate that the MiLSF strategy outperforms RS and RRU in different scenarios. According to the Student's t distribution, the energy saving percentage results of MiLSF, RS, and RRU relative to the non-sleep strategy are within 2% of the observed mean within the 95% confidence interval.

[0227] Part A: User traffic data prediction by BLSTM

[0228] The user traffic data comes from

[67] , which is obtained by analyzing the traffic data of a medium-sized city with a large population in China for a week. In order to predict the traffic demand of each user on the next day, their historical traffic demand rate in the past week is used for model training. Specifically, the hourly average traffic rate dataset of each user for eight consecutive days is divided into two parts: the first seven days are used for training, and the eighth day is used for testing.

[0229] A BLSTM neural network is used, which consists of an input layer for receiving user traffic data sequences, two LSTM layers with 500 hidden units each, a fully connected layer whose output size matches the input size, and a regression layer for calculating the semi-mean square error loss of the regression task

[68] . When training the model, RMSProp is used as the optimizer, and the parameters (including optimization algorithm, learning rate, and batch size) are set to the default values ​​of the optimizer.

[0230] In order to train the BLSTM neural network on the above-mentioned real user traffic data from

[67] , it is observed that user traffic exhibits daily periodic characteristics. Therefore, a special training method is designed to exploit this feature. First, a model is trained on the user traffic of the first two days, and the model is used to predict the traffic of the third day. Then the error between the predicted traffic and the actual traffic is recorded, and the model parameters (such as the gate parameter) are modified to minimize the modified error. Using the modified model, the user traffic of the fourth day is predicted based on the traffic conditions of the last two days (i.e., the second and third days), the error is recorded and the model parameters are modified. These steps are repeated until all traffic data for seven days are used, which marks the end of one round of training. The above steps are repeated until the average error of this round is lower than the given threshold, thus completing the model training.

[0231] To evaluate the performance of the BLSTM neural network, two commonly used metrics were used: mean absolute error (MAE) and root mean square error (RMSE). Specifically, MAE is calculated by the average of the absolute prediction errors, while RMSE is the square root of the average of the squared prediction errors.

[0232] A 20% holdout validation set is used to evaluate the performance of the BLSTM, RNN, and ARIMA models. Specifically, 20% of the user traffic data in

[67] is reserved for validation, and the remaining 80% is used for training. Table 1 lists the validation results of the three models on the validation set, among which BLSTM shows a significant advantage over RNN and ARIMA in user traffic prediction, with the lowest MAE and RMSE.

[0233]

[0234] Table 1: 20% holdout validation for BLSTM, RNN, and ARIMA models

[0235] exist Figure 7a-9c In Figure 1, hourly traffic predictions for three users with different traffic patterns are shown for BLSTM, RNN, and ARIMA models, respectively. The results presented in each sub-figure are based on the performance metrics MAE and RMSE. These figures are used to illustrate the prediction accuracy of each model for different traffic patterns.

[0236] It can be seen that Figure 9a-9c As shown in Figure 1, the difference between the prediction results of ARIMA and the actual values ​​is the largest, which can be observed from the MAE and RMSE indicators. This is because ARIMA is a classic linear model that assumes that time series data has a linear relationship, but the actual data may contain nonlinear relationships, which may lead to inaccurate prediction results for nonlinear time series

[69] ,

[70] ,

[71] . Nonlinear time series analysis requires the use of corresponding nonlinear models or methods to model and predict data, such as RNN and LSTM in neural network models.

[0237] It is observed that the mean square error (MAE) of the prediction results of RNN and BLSTM for the three users is similar (BLSTM is slightly better than RNN), indicating that the average prediction accuracy of the two methods is comparable. Figure 7a-7c and Figure 8a-8c As shown in Figure 2, BLSTM is significantly better than RNN in predicting peak rate, as BLSTM has much lower root mean square error (RMSE) for three users. Therefore, BLSTM is used to predict the future traffic of each user in a two-layer heterogeneous cellular network. Based on the user traffic predicted by BLSTM, the MiLSF strategy is implemented and proved to be superior to RS and RRU strategies in different scenarios.

[0238] Part B: Effectiveness of MiLSF

[0239] Here, the MiLSF strategy is compared with four benchmark strategies, RS, RRU, CUR, and CBSSF, in different scenarios based on the user traffic data predicted by BLSTM. Unless otherwise specified, the simulation parameter values ​​in Table 2 are used. Although different users have different traffic demands at different times, most users usually have lower traffic demands at night, resulting in a lower total network load during this period. This provides an opportunity for the base station sleep strategy to save energy, so it can be demonstrated that MiLSF outperforms the four benchmark strategies considering the night time. In addition, it is assumed that users do not move during the low-load period at night, because user mobility is greatly reduced during this period. The low-load period is considered to be the sleep period between 10 pm and 6 am the next day. That is, a group of selected micro base stations enter the sleep mode at 10 pm and enter the active mode at 6 am, with the goal of optimizing the selection of these micro base stations. PPP is used to determine the location of the users. Each user is randomly assigned to a base station that meets the SINR threshold and the user traffic demand rate for initialization. MiLSF, RS, RRU, CUR, and CBSSF strategies are then implemented during the low-load period and compared with the non-sleep strategy for energy-saving performance. Numerical results show that MiLSF is more effective and superior than four baseline strategies.

[0240] Scenario I: In this scenario, the energy-saving performance of MiLSF and four benchmark strategies under two different base station deployment methods is compared. Specifically, for each strategy, two different deployment processes, PPP and MHCPP, are used to determine the locations of all base stations. It is worth noting that the deployment processes of macro and micro base stations are independent and have different strengths, as shown in Table 2. By comparing the energy-saving performance of each strategy under different deployment methods, we can gain insight into the effectiveness of each strategy under different base station deployment methods.

[0241]

[0242]

[0243] Table 2: Simulation parameters

[0244] Fig.10The impact of different base station deployment methods on the energy-saving performance of the base station sleep strategy was observed in Figure 2. The results show that MiLSF outperforms all four baseline strategies under PPP and MHCPP deployments. However, when MHCPP is used instead of PPP for base station deployment, the energy saving percentage of the same strategy increases significantly. For example, the energy saving percentage of the MiLSF strategy under PPP deployment is 5.82%, but the energy saving percentage under MHCPP deployment is 11.26%, almost twice that of PPP deployment. The disadvantage of PPP is that there are severe signal interference due to the presence of base stations of the same type that are too close to each other. This reduces the number of base stations that can be selected by the user in step c), resulting in a significant reduction in the number of micro base stations that can eventually switch to sleep mode and a lower energy saving percentage. Therefore, in scenarios II, III, and IV, base stations are deployed using MHCPP.

[0245] Scenario II: In this scenario, by changing Fig.11 The user density in the graph is used to observe the impact of user load on the energy-saving performance of MiLSF and four benchmark strategies.

[0246] Fig.11 It is shown that MiLSF outperforms all four baseline strategies and achieves higher energy saving percentages at different numbers of users, especially when the number of users is moderate. This is because, unlike RS, MiLSF always selects the micro base station with the lowest load and reallocates connected users to other base stations. It is easier to put the micro base station with the lowest load to sleep than to put the micro base station with high load to sleep, because users connected to the micro base station with the lowest load are more likely to be "satisfied" and be assigned to other base stations. Prioritizing the allocation of users of the micro base station with the lowest load to other base stations can also reduce the possibility of problems with other micro base stations during subsequent sleep operations. In addition, MiLSF prioritizes the allocation of users to the macro base station with the largest SINR, thereby achieving a higher energy saving percentage than RRU, where users are reallocated to randomly active base stations that may not fully utilize the bandwidth resources of MBS.

[0247] In the CBSSF strategy, the micro base stations closest to the macro base station are always given priority for sleep. Therefore, the users previously associated with these micro base stations are most likely to be reallocated to the nearest corresponding macro base stations. This characteristic leads to over-emphasis on macro base stations during user reallocation, which triggers premature load saturation of macro base stations. At the same time, micro base stations that cannot transition to sleep mode (if a user only relies on a specific micro base station and cannot be assigned to other base stations) remain in a low-load working state and cannot enter sleep mode. This situation reduces the total number of micro base stations that can eventually switch to sleep mode, and the energy saving rate is greatly reduced compared to the MiLSF strategy.

[0248] As for the CUR strategy, it simply reallocates users to the base stations closest to them and guarantees the maximum user traffic demand and SINR greater than the threshold. However, it ignores the load distribution of each base station in the entire network and the different SINRs that different base stations can provide. The MiLSF strategy fully considers the load information and optimal SINR of all base stations in the network, while the CUR strategy does not achieve effective global energy-saving optimization in the entire network like the MiLSF strategy.

[0249] also, Fig.11 It is shown that the energy saving percentages of the five strategies are similar when the number of users is zero or close to 400. This is because, under all strategies, when the number of users is small, all micro base stations are in sleep mode, and when the number of users is large, they are all active, so the energy saving performance is similar. In addition, the more users there are in the network, the lower the energy saving percentages of all five strategies, because fewer micro base stations can be switched to sleep mode.

[0250] Scenario III: In this scenario, the effect of SINR threshold on MiLSF and Fig.11 The results show that MiLSF outperforms all four benchmarks at different SINR thresholds. In addition, the energy saving percentage of all five strategies gradually decreases with the increase of SINR threshold, because the number of base stations that each user can choose to reallocate to is reduced, which reduces the possibility of successfully reallocating all users to each MiBS, resulting in a reduction in the number of micro base stations that can be switched to sleep mode.

[0251] In addition, when the SINR threshold is moderate, the difference in energy saving percentage between MiLSF and the four baseline strategies is the largest. When the SINR threshold is too large (i.e., γ0 = 0 dB), only very few active base stations can reach the SINR threshold for user reallocation in step c), which makes it difficult for all five strategies to successfully reallocate users. Therefore, the difference in energy saving percentage is small. On the contrary, when the SINR threshold is too small (i.e., γ0 = -12 dB), almost all active base stations can become qualified active base stations for user reallocation in step c), which makes it easier for all five strategies to reallocate all users connected to the micro base station and achieve higher energy saving percentages.

[0252] When the SINR threshold is moderate, the user reallocation in step c) cannot be too "random" because each user may have only a few (or even one) eligible active base stations. In this case, MiLSF prioritizes the macro base station with the highest SINR, maximizes the demand traffic of all users, and puts as many micro base stations into sleep state as possible. Therefore, when the SINR threshold is moderate, MiLSF has significant energy-saving advantages over RS, RRU, CUR and CBSSF.

[0253] Scene IV: In this scene, Fig.13 The relationship between the number of dormant micro base stations and the energy saving percentage under the five strategies is shown. The results show that there is a linear positive correlation between the number of dormant micro base stations and the energy saving percentage. Note that the energy saving percentages of the five strategies are very different when the number of dormant micro base stations is the same. This is because under MiLSF and the four benchmark strategies, different micro base stations may enter the dormant state, resulting in different user reallocation situations (users may be reallocated to different macro base stations or micro base stations). The small difference in energy saving performance between MiLSF and the four benchmark strategies shows that it is challenging to significantly reduce energy consumption if only different base station dormancy strategies or user reallocation are considered without switching more micro base stations into dormant mode. Therefore, a base station dormancy strategy that can put more micro base stations into dormant mode is always preferred. In this case, Fig.10 , Fig.11 and Fig.12 The results in show that under given network parameters, the MiLSF strategy can always achieve a higher energy saving percentage than the four baseline strategies by putting more micro base stations into sleep state, thus proving its effectiveness and superiority over RS, RRU, CUR and CBSSF.

[0254] in conclusion

[0255] Therefore, the above exemplary embodiment of the present invention proposes a new base station sleep strategy, namely MiLSF, for a two-layer heterogeneous cellular network, which is based on the predicted traffic demand rate of each user. The underutilized micro base station is switched to sleep mode, and the user is reallocated to other active base stations that meet the SINR threshold and can guarantee the maximum demand traffic rate of the user during low load. BLSTM is used to predict the future traffic demand rate of each user based on the historical traffic data of the past week, and its prediction accuracy is higher than RNN, especially in terms of peak rate. The MiLSF strategy has been implemented during low load at night to explore its energy-saving performance in different scenarios. Specifically, the MiLSF strategy always tries to put the micro base station with the least load into sleep mode and reallocate users to the macro base station with the highest SINR to ensure their traffic demand rate. Through a large number of numerical simulations, the effectiveness and superiority of the MiLSF strategy over the four benchmarks of RS, RRU, CUR and CBSSF are demonstrated. In addition, although only one sleep operation of a two-layer heterogeneous cellular network is considered, the MiLSF strategy can be extended to any multi-layer heterogeneous cellular network to perform multiple sleep operations during various low load periods.

[0256] Although the above exemplary embodiments focus on improving energy efficiency while guaranteeing the QoS of each service at a predefined SINR ratio, load balancing is an important aspect to consider in network design. Load balancing helps to evenly distribute network traffic and improve the quality of service for end users. Therefore, combining MiLSF strategies with load balancing technology can further improve the overall performance and efficiency of the network while ensuring the QoS of each deployed service.

[0257] Compared with existing technologies and products on the market, the above exemplary embodiments of the present invention have the following advantages:

[0258] More accurate prediction: Unlike existing studies that predict the traffic of the entire network or a single base station, this embodiment uses BLSTM to predict the traffic demand rate of each user in a heterogeneous cellular network. This prediction method is more accurate because it takes into account the differences in traffic demand between users.

[0259] Superior base station sleep strategy: The MiLFS strategy outperforms existing base station sleep strategies. By switching the lightest-loaded micro base stations to sleep mode first, the strategy can minimize the overall energy consumption of the network.

[0260] Better network management: This implementation considers the impact of different base station deployment methods on the performance of the sleep strategy and fully utilizes micro base stations in heterogeneous cellular networks during sleep by controlling user selection. This approach can achieve load balancing between base stations, resulting in more efficient network management and better utilization of network resources.

[0261] Higher energy efficiency: By optimizing base station on / off and sleep strategies, this implementation can significantly improve the energy efficiency of mobile cellular networks, thereby reducing greenhouse gas emissions and saving considerable operating costs.

[0262] ·

[0263] In summary, the present invention provides a more efficient, more accurate and more energy-efficient user traffic prediction and base station sleep strategy, which is much better than existing technologies and products on the market.

[0264] Therefore, the exemplary embodiments have been fully described. Although specific embodiments have been mentioned in the description, it will be clear to those skilled in the art that the implementation of the present invention may be modified from these specific details. Therefore, the present invention should not be interpreted as being limited to the embodiments described herein.

[0265] Although the embodiments have been described in detail in the drawings and the above description, they should be regarded as illustrative rather than restrictive, because only exemplary embodiments are shown and the scope of the invention is not limited in any way. It is understood that any feature described herein can be used in combination with any embodiment. The illustrative embodiments are not mutually exclusive, nor are they exclusive of other embodiments not mentioned in this specification. Therefore, the present invention also provides embodiments including a combination of one or more of the above illustrative embodiments. The invention described in this specification may be modified and varied without departing from its spirit and scope, and therefore, only the limitations described in the attached claims should be applied.

[0266] The functional units and modules of the systems and methods described in the embodiments of the present invention may be implemented using computing devices, computer processors or electronic circuits, including but not limited to application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), and other programmable logic devices configured or programmed according to the present disclosure. Computer instructions or software codes are run in computer devices, computer processors or programmable logic devices, and can be easily written by skilled personnel in the field of software or electronic technology based on the contents of the present disclosure.

[0267] All or part of methods consistent with embodiments may be performed in one or more computing devices, including server computers, personal computers, laptop computers, and mobile computing devices such as smartphones and tablet computers.

[0268] Embodiments include computer storage media, transient and non-transient memory devices having stored therein computer instructions or software codes that can be used to program a computer or microprocessor to perform any of the processes of the present invention. Storage media, transient and non-transient computer-readable storage media may include, but are not limited to, floppy disks, optical disks, Blu-ray disks, DVDs, CD-ROMs, magneto-optical disks, ROMs, RAMs, flash memory devices, or any type of medium or device suitable for storing instructions, codes, and / or data.

[0269] Each functional unit and module consistent with various embodiments may also be implemented in a distributed computing environment and / or a cloud computing environment, wherein all or part of a machine instruction is executed in a distributed manner by one or more processing devices connected via a communication network (such as an intranet, a wide area network, a local area network, the Internet, and other forms of data transmission media).

Claims

1. A method for managing base stations in a dual-layer heterogeneous cellular network, the heterogeneous cellular network comprising a plurality of macro base stations and a plurality of micro base stations; The method comprises the following steps: a) predicting future traffic of users of the heterogeneous cellular network; b) switching the user's connection from a base station with low utilization rate among the multiple micro base stations to one or more of the multiple micro base stations or the multiple macro base stations; and c) shutting down the base stations with low utilization among the multiple micro base stations.

2. The method of claim 1, wherein step a) is performed using a bidirectional long short-term memory neural network.

3. The method according to claim 2, wherein the bidirectional long short-term memory neural network includes two layers, one layer is used to transmit information in chronological order, and the other layer is used to transmit information in reverse chronological order.

4. The method according to claim 1, wherein in step b), the QoS requirement of the user is guaranteed.

5. The method of claim 4, wherein in step b), a signal to interference plus noise ratio threshold of the user is not violated.

6. The method according to claim 5, wherein step b) further comprises: e) identifying a macro base station having a highest signal to interference plus noise ratio from the plurality of macro base stations; as well as f) reallocating the connection of the user from the micro base station with low utilization rate among the multiple micro base stations to the macro base station with the highest signal to interference plus noise ratio.

7. The method according to claim 5, wherein step b) further comprises: g) identifying a second micro base station with the highest load from the multiple micro base stations; as well as h) reallocating the connection of the user from the micro base station with low utilization rate among the multiple micro base stations to the second micro base station.

8. The method according to claim 3, wherein during step b), the maximum traffic demand rate of the user is always met.

9. The method according to claim 8, wherein step b) further comprises: i) identifying a third base station from the plurality of macro base stations and the plurality of micro base stations that can meet the maximum traffic demand rate during a low load period; as well as j) reallocating the user's connection from the micro base station with low utilization rate among the multiple micro base stations to the third base station.

10. The method according to claim 1 also includes modeling the deployment of the multiple micro base stations and the multiple macro base stations as a random point process. The method according to claim 10 , wherein the random point process comprises a Poisson point process and a Markov hard core point process.

12. The method of claim 1, wherein at any time, none of the plurality of macro base stations is shut down.

13. The method of claim 1, further comprising repeating steps a)-c) for different users.

14. The method according to claim 13, further comprising repeating steps a)-c) until there is no more macro base stations that can be shut down.

15. A non-transitory computer-readable storage recording medium having computer instructions recorded thereon; when executed on one or more processors, the computer instructions cause the one or more processors to perform operations according to the method of claim 1.

16. A computing system comprising: one or more processors; as well as A memory containing instructions; when executed on one or more processors, the instructions cause the computing system to perform operations according to the method of claim 1.

Images