Indonesian J our nal of Electrical Engineering and Computer Science V ol. 41, No. 3, March 2026, pp. 935 945 ISSN: 2502-4752, DOI: 10.11591/ijeecs.v41.i3.pp935-945 935 Satellite-based assisted-ofoading f or ener gy-constrained edge netw orks Thembelihle Dlamini 1 , Mengistu A. Mulatu 1 , Siso V ilakati 2 1 DepartmentDepartment of Electrical and Electronic Engineering, Uni v ersity of Esw atini, Kw aluseni, Esw atini 2 Department of Biostatistics, Uni v ersity of Free State, Free State, South Africa Article Inf o Article history: Recei v ed Oct 27, 2025 Re vised Jan 21, 2026 Accepted Feb 27, 2026 K eyw ords: F orecasting Green ener gy Protected areas Remote edge serv ers Satellite constellations ABSTRA CT As the need for global broadband internet connecti vity increases, there is a need to consider the use of non-terres trial netw orks (NTNs) to e xtend the netw ork co v erage to protected are as (e.g., national parks). Usually , protected areas are prohibited from ha ving po wer lines thus lacking wireless connecti vit y . T o o v er - come this challenge, ener gy can be pro vided through the use of green ener gy from a solar photo v oltaic (PV) system. Then, a green ener gy-based base station (BS) can be deplo yed within the area in order to pro vide mobile connecti vity to visitors, as well as also using the NTNs to handle e xcess traf c or tak e o v er the traf c in the e v ent the BS does not ha v e suf cient green ener gy from stor - age. In this paper , a h ybrid wireless communication system is proposed to in- clude BS sites located in a protected area and satellites in the lo w earth orbits (LEO), coupled with ne w of oading strate gies, with the main goal of optimizing the trade-of f between ener gy consumption and end-to-end delay for the green ener gy-based BS sites. F or accurac y of our simulations, we c onsider real data from a solar photo v oltaics system, traf c w orkloads, visitor’ s location data, and satellite orbits from Starlink constellations. Our results demonstrate that the co- e xistence of the BS and satellite achie v e ener gy sa vings from 59 % to 34 % , with an a v erage sys tem delay of 0 . 83 seconds and a pack et drop rate that ranges from 8 . 3 % to 2 . 7 % , when compared with our benchmark. This is an open access article under the CC BY -SA license . Corresponding A uthor: Thembelihle Dlamini Department of Electrical and Electronic Engineering, Uni v ersity of Esw atini Kw aluseni, Kw aluseni, Esw atini Email: tldlamini@unesw a.ac.sz 1. INTR ODUCTION The ur ge for sensing, learning, and communication services in future mobile node (MN) is a k e y re- search area in 6G [1]. This is moti v ated by the need to ha v e wireless communication systems in protected areas (e.g., national parks, nature reserv es) to intelligently monitor the en vironment, pro vide mobile services to visi- tors, and also to track endangered species. Ha ving the communication infrastruct ure will allo w data processing at the edge, that is, within a base station (BS) empo wered with computation capabilities or in remote clouds (satellites). The primary objecti v e of ha ving protected areas is to protect biodi v ersity and ecosys tem functions, and through interactions with natural en vironments, people deri v e a v ariety of ph ysical and psychological ben- ets [2]. Despite the benets, according to go v ernment la ws, electricity lines are not permitted thus limiting the pro visioning of MN services [3]. T o enable smart connecti vity in protected areas, in the near future, it is e xpected that the combination of edge serv ers and green-po wered BS will pro vide the capability to deplo y communication sites without requiring electrical wiring for po wer supply . In this, the BS are equipped with J ournal homepage: http://ijeecs.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
936 ISSN: 2502-4752 ener gy harv esting (EH) equipments and computing capabilities [4]. The use of green ener gy to po wer the edge systems will reduce the carbon emissions and also to e xpand netw ork co v erage within protected areas [5]. The pro vision of netw ork co v erage in unserv ed areas, similar to protected areas, is also of great im- portance despite the terrain dif culties which hinders communication to wer installations. Using non-terrestrial netw orks (NTN), which comprise satellites, unmanned aerial v ehicles (U A Vs), and high altitude platforms (HAPs), to supplement terrestrial netw orks or of fer on-demand wireless access to places without infrastruc- ture, is one potential remedy for this [6]. High tele vision (TV) to wers and lar ge antenna arrays utilizing massi v e-MIMO are proposed to pro vide connecti vity to sparse areas [7]. Here, the systems ha v e the latest emer ging antenna technologies and designs such as recongurable phased/inatable/fractal antennas realized with metasurf ace material. In addition, NTN can also of fer a connecti vity service in the e v ent of a natural disaster [8], that is, where the deplo yed netw ork infrastructure or terrestrial to wers are out of service. Similar to Multi-access Edge Computing MEC platforms, NTN can of fer communication-plus compu- tation services in addition to e xpanding netw ork co v erage [9]. Here, the y can accept of oaded delay-dependent tasks from ener gy-decient or capacity-constrained BS sites. F or e xample, within a protected area, visitors are serv ed by the BS which are located along the perimeter of the area, which in turn drains the battery of the mo- bile de vices when processing an y data due to the lar ger separation distance. This a v ails the opportunity of using NTNs. Re g arding of oading tasks to sa tellites in lo w earth orbit (LEO), Pietro [10] proposed the use of in-orbit computing to pro vide near real-time computing to areas where satellites are the only option and terrestrial con- necti vity is lacking in order to of oad jobs to LEO satellites. Here, the author presented an algorithm for a LEO satellite constellation to handle tasks from dif ferent locations through the sharing of the computing plat- form, thus relie ving ground resources from computing some of the w orkloads. Similarly , Soret [11] proposed the use of LEO satelli te constellations for of oading w orkloads and also backhauling the traf c from remote terrestrial communication sites to the core netw ork. Here, the y e v aluated their performance based on Age of Information, latenc y , and collision rate. Performance asses sments of LEO satellites are pro vided in [12], where data of oading techniques are suggested within the frame w ork of v ehicular edge computi ng. In that paper , the delay-sensiti v e task is sent from the ground de vice directly to the satellite. The use of terrestrial netw orks and NTNs for ubiquitous co v erage is proposed as a solution for multi-connecti vity in rural areas [13]. Here, the research w ork focused on smart agriculture related use case and using l atenc y as a performance metric. From the aforementioned papers, it is noted that more studies are required to measure the performanc e of the co-e xistence of LEO sate llite computing platforms and green-po wered BSs (empo wered with computing capa- bilities), in order to guarantee the e xpected end-to-end delay . Moreo v er , this researc h article dif fers from [12] as the ground mobile de vices send information to the BS rst, and then the BS transmit the information to the satellite via a locally mounted Starlink antenna. Managing green-po wered communications sites is also important as the y are dependent on the amount of green ener gy that can be harv ested per time instance. In order to handle dynamic w orkload of oading, an ef cient reinforcement learning-based resource management algorithm is proposed and in this paper green ener gy sources were inte grated into a MEC system [14]. Then, in our pre vious paper , we suggested a ne w net- w ork design where a controller manages the EH BSs [4]. T o manage the computing resources (V irtual machine (VM) and BS), we selecti v ely turn them on and of f o v er a constrained prediction horizon. By redistrib uting the netw ork load among the BSs and taking adv antage of the spatial di v ersity of the a v ailable green ener gy , the green-based load balancing technique is suggested, in order to maximize the edge system performance [15]. Here, contai ners were c o ns idered as computing resources within the MEC serv er . Ov erall, it should be stated that the aforementioned research w orks lack the consideration of of f-grid BS systems for protected areas and the use of LEO satellite to complement the BSs located in such areas. The main contrib utions of this paper are summarized as follo ws: (i) we propose a no v el edge com- puting frame w ork for ground mobile de vices that i nclude mechanisms to dynamically of oad tasks to LEO satellites via the BS if there is a guarantee of near real-time communications-plus-computation processes, or prioritize the use of a local edge platform. This frame w ork also mak e use of admission control procedures within the BS, as well as in the satellites; (ii) T o e v aluate the netw ork performance, we jointly consider the use of green-po wered BSs and LEO satellites for of oading the delay-sensiti v e tasks from a protected area. Here, our main goal it to optimize the trade-of f between ener gy consumption and the end-to-end delay , through the use of limited liabili ty compan y (LLC) principles and the use of green ener gy as a performance metric. Real-w orld harv ested ener gy , traf c load traces, visitor’ s location data, orbital traces and parameters, are used to e v aluate the performance of the proposed optimization strate gy . The numerical results obtained through Indonesian J Elec Eng & Comp Sci, V ol. 41, No. 3, March 2026: 935–945 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 937 simulation sho w that the proposed optim ization strate gies are able to ef ciently manage the communication sites, as well as allo wing processing of tasks in LEO satellites. The remainder of this article is or g anized as follo ws: Section 2 presents our sys tem model, sec tion 3 pro vides the mathematical formulation of the problem and the of oading strate gies, section 4 discuss the simulation results, and section 5 summarizes our conclusion. 2. SYSTEM MODEL This section describes the scenario section 2.1, the satellite orbit and channel model section 2.2, the ener gy consumption model for the BS site section 2.3, and the delay model section 2.4. 2.1. Scenario description Our considered scenario is depicted in Figure 1. Here, we consider n N BSs deplo yed in a protected area (in our instance, a national park is considered) and each is equipped with computing capabilities (i.e., each EH BS has a local computing platform that runs containers). A set of u ground mobile de vices, from visit ors who prefer w alking within the park, of oad t heir delay-sensiti v e or delay-tolerant tasks to BS n , (in our case direct of oading to the satellite is not allo wed). The BS are within the co v erage of LEO Starlink satellites. V isitor’ s current locat ions (mobility patterns) are kno wn through the location service application programming interf ace (LS API) [16], which is a service that supports the mobile de vice location retrie v al mechanism and then passing the information to authorized applications within the MEC platform. Here, we emulate the user equipment (UE) location lookup procedure between the edge serv er the subscribed mobile users using the location dataset from [17]. Figure 1. The BS system being complemented by a satellite The BS infrastructure is shared between mobile operators, in order to a v oid cro wding the prot ected site, and it is po wered by ener gy harv ested from a solar photo v ol taic (PV) system. Here, the use of wind turbines as a source of ener gy is ne glected due to its noise pollution which has an ef fect to wildlife [18]. The obtained ener gy then supplies the computing platform, BS communication infrastructure, as well as Starlink antenna mounted onsite to pro vide satellite-based Internet services. An y traf c that is going to be of oaded to the satellite, will pass through the mounted Starlink antenna. In addition, the battery bank of size Υ max stores the e xcess ener gy , and for the satellite we assume that there is enough ener gy within its battery bank as there e xist a direct line of sight (LOS) with t he sun. On the MEC serv er of capacity C lo c , there is an access control application that is responsible for admitting and forw arding w orkloads to its acti v e G computing resources (containers) or to a set of s visible LEO Starlink satellites that are each running G containers onboard, with each ha ving a capacity C sat , for processing the of oaded data taking into account the delay constraint σ . 2.2. Orbital and channel model Generic orbital model: The ground mobile de vices of oad their tasks to the BS, and the BS loc ation is determined by its latitude l u and longitude L u , and the satellite’ s position in space is dened by its altitude Satellite-based assisted-of oading for ener gy-constr ained edg e networks (Thembelihle Dlamini) Evaluation Warning : The document was created with Spire.PDF for Python.
938 ISSN: 2502-4752 l h , latitude l s and longitude L s . T o accurately determine the location of a Starlink satellite on its orbit, we mak e use of the tw o-line element (TLE) data from [19], which pro vides an up-to-date trajectory and orbital parameters. Similar to [12], the separation distance d (in km) from the BS and the generic satellite as follo ws, d = ( l h + r e ) " 1 + r e l h + r e 2 2 r e l h + r e cos( ψ ) # 1 / 2 (1) where r e is the radius of the earth (i.e., 6378 . 137 km), and ψ represents the angle between the BS and the satellite as observ ed from the earth’ s center , and it is related using the follo wing equation, cos( ψ ) = cos( l s ) cos( l u ) cos( L u L s ) + s i n( l u ) sin( L s ) (2) From [20], the angle of ele v ation θ is obtained from the follo wing relationship, cos( θ ) = ( l h + r e ) sin( ψ ) d (3) F or a satellite to be visible to the BS, its ele v ation angle must be abo v e some minimum v alue and it is upper bounded by some v alue, i.e., θ 81 . 3 [20]. W ir eless channel model: The channel model for the BS to satellite connecti vity can be obtained from the 3GPP specications [21], where we assume the LOS and let the signal-to-noise ratio (dB) between BS n and the visible satellite (in log ) to be, Γ n,s = EIRP n + ( G/T ) n P n k B n , (4) where EIRP n is the ef fecti v e isotropic radiated po wer of t he transmitter in W , ( G/T ) n is the recei v ed antenna g ain to noise temperature ratio (sometimes called gure of merit), P n is the path loss which constitute of free space path loss, pointing loss, polarization loss, and loss due to the atmosphere, k is the Boltzmann constant and B n is the bandwidth is Hz. The free space path loss, P FS n , is gi v en by [20], P FS n = 92 . 45 + 20 log ( f c ) + 20 log ( d ) (5) where f c is the carrier frequenc y in GHz. 2.3. BS consumption model The total ener gy consumed [J] by each BS site, denoted by β site n ( t ) , consists of the wireless com- munications processes, denoted by β bs n ( t ) , and the edge serv er processes, which includes computing, caching, and communication acti vities within itself, is denoted by β edg n ( t ) . Thus, at slot t , the ener gy consumed can be formulated as follo ws [15], β site n ( t ) = β bs n ( t ) + β edg n ( t ) (6) The transmission process drains ener gy from the BS site. Here, we let β 0 represent the operating ener gy ne- glecting w orkloads, β l d ( t ) is the task dependent transmission po wer to-and-from the ground mobile de vices and also to the satellite at a tar get rate of r 0 which guarantees the lo w latenc y threshold, β sat ( t ) is the ener gy used by the Starlink antenna for uplink (UL) and do wnlink (DL) data transfer , and β dt ( t ) is the inter -communication ener gy cost [J/byte] for passing data to the edge serv er int erf aces for processing. Since the BS transmission po wer is adapti v e, we let η ( t ) { 0 , 1 } be the s witching s tatus indicator . Thus, the BS consumption is as follo ws, β bs n ( t ) = η ( t ) β 0 + β l d ( t ) + β sat ( t ) + β dt ( t ) (7) The ener gy consumed by the softw arized computing platform (in the satellite or MEC serv er) is dependent on the total number of acti v e containers at time t . Th e CPU utilization share is denoted by β cp ( t ) , then the resource scheduler will acti v ate and de-acti v ate containers based on-demand, through a reliable intra-communication link operating at rate of r g ( t ) bits/s, in order to allocate tasks of size γ g ( t ) and such incurs a cost denoted by β sw ( t ) . Then, the tasks are queued before processing and dequeue after processing, thus the input-output Indonesian J Elec Eng & Comp Sci, V ol. 41, No. 3, March 2026: 935–945 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 939 b uf fer acti vities incurs an ener gy cost and it is denoted by β q u ( t ) . Since some of the viral Internet content can be cached locally , the caching process contrib utes an amount of ener gy drained and it is denoted by β ch ( t ) . The total ener gy drained within the computing platform is as follo ws, β bs n ( t ) = η ( t ) β 0 + β l d ( t ) + β sat ( t ) + β dt ( t ) (8) W ithin the computing platforms, the maximum per -slot communication rate is limited to a pre-assigned v alue r max , thus the follo wing hard constraint must hold: P G ( t ) g =1 r g ( t ) r max . The ener gy that is used for BS operations is from a solar PV system Figure 2 for harv ested ener gy traces from [22]) and the battery . At each time instance, the amount of ener gy dra wn from the battery , for communication-plus-computing acti vities, must be less than the amount required by the communication site and it is dened as Υ n ( t ) β site n ( t ) . Thus, the amount of the green ener gy demanded per site is as follo ws, Figure 2. Solar ener gy traces from a PV system for day 1 (D1) to day 3 (D3) υ n ( t ) = υ c n ( t ) + υ o n ( t ) (9) where υ c n ( t ) is the fractional ener gy that is for char ging the battery and υ o n ( t ) is the share that is used im- mediately for supporting local operations. Per each time slot t , it is important to note that the actual amount of ener gy (denoted by E max n ) that can be e xtracted from the en vironment is limited, t hus we ha v e an ener gy harv esting constraint as, υ c n ( t ) + υ o n ( t ) E max n , n, t. (10) During the day , the storage de vice le v el uctuates according to the follo wing equation, Υ n ( t ) = ζ n n ( t 1) β site n ( t )) + Φ( υ c n ( t )) , (11) where ζ n (0 , 1] represents the battery self-dischar ging beha vior , and Φ (0 , 1] represents the incurred losses during the char ging phase. 2.4. Delay model The input/output (I/O) queue of the system are assumed to be loss-free such that the time e v olution of the backlogs queues follo ws Lindle y’ s equations [23]. F or intra-communication within the computing platform, we note that there e xist a tw o-w ay pe r task e x ecution delay (task to-and-from the container) where each link delay is denoted by ρ g ( t ) = 2 γ g ( t ) /r g ( t ) . Then, the computation processing duration that depends on the CPU c ycles denoted by t cp = | G ( t ) | /C lo c = | G ( t ) | /C sat , queuing delay of the task on the input-output b uf fer , assuming the e xistence of a congestion handling mechanism, denoted by t q u . Thus, the delay on the computing platform, denoted by t d , is as follo ws, t d = ρ g + t cp + t q u . (12) Satellite-based assisted-of oading for ener gy-constr ained edg e networks (Thembelihle Dlamini) Evaluation Warning : The document was created with Spire.PDF for Python.
940 ISSN: 2502-4752 Since the tasks are either of oaded to the BS site for local computation or forw arded to the satellite edge computing platform, the time delay due to access decision making (accepting or forw arding) is denoted by t ac . The ground mobile de vice of oad the task to the BS, and the UL and DL delays are denoted as t u ul and t u dl , which are dependent on the transmission rate r 0 that guarantee the e xpected delay threshold and the size of the tasks as t u ul = γ g ( t ) /r 0 and t u dl = γ g /r 0 , where γ g is the computed results from the BS. Thus, for tasks of oaded to the BS for local computation, the total delay (denoted by t l oc ) is as follo ws, t l oc = t u ul + t u dl + t ac + t d + 2 t pg , (13) where t pg is the propag ation delay . If the of oaded tasks are transmitted to the selected satellite from the BS, then there is the UL and DL delay due to the transmission rate denoted by r s n ( t ) which is related to (4) through shannon capacity and the size of the transmitted tasks (denoted by γ n ( t ) ) as t ul n,s = γ n ( t ) /r s n , and t dl n,s = γ n ( t ) /r s n , where γ n is t he computed results from the satellite to BS n . Then, the total delay for tasks of oaded to the satellite (denoted by t sat ) is gi v en as follo ws, t sat = t ul n,s + t dl n,s + t d + 2 t pg . (14) T o guarantee lo w latenc y for applications in MNs, we ha v e to mak e sure that the follo wing conditions hold for delay-sensiti v e tasks: t l oc σ and t sat σ . 3. BS-SA TELLITE OFFLO ADING FRAMEW ORK Data processing in v olv es local BS computation or satellite-based computation, follo wed by tasks drop- ping if all t he options are not a v ai lable. Cases of dropping t asks are dependent on the a v ailable stored ener gy and the loading of the input-output queue. In terms of resource allocation within the computing platform , the container pro visioning and load allocation o v er them, at t , is performed similar to [15]. In general, local computation is more con v enient, pro vided that suf ci ent green-ener gy i s a v a ilable, as well as the computing resources. The ener gy to be harv ested and the tasks to be of oaded are accumulated o v er the time slot, and the y can only be kno wn at the end of it. This implies that the amount of harv ested ener gy and the tasks from the ground mobile de vices can only be estimated using the LSTM neural netw orks [24], i.e., ˆ Υ n ( t ) and ˆ γ n ( t ) . In order to manage the of oading process, using green ener gy as a performance metric, we propose a frame w ork that will identify if the BS will compute the tasks locally or it will steer part of the of oaded tasks to the satellite edge system. T o achie v e this, the communication-plus-computing interv al is dened as the ratio of the ne xt time slot a v ailable green ener gy and the e xpected total ener gy consumption (recall that the harv ested ener gy and t asks are forecasted), per BS site, as J n ( t ) = Υ n ( t +1) β site n ( t +1) 1 . F or of oading decision making, we emplo y the follo wing strate gies: Str ate gy 1 (Local of oading (LO)): If J n ( t ) 1 and ˆ t l oc < σ , the site ener gy will be suf cient to handle the e xpected tasks with the guarantee of a lo w latenc y , otherwise if J n ( t ) < 1 the communication site will not be able to handle the e xpected tasks, thus the strate gy is to of oad the tasks to the visible satellites or drop them. Here, we assume the ground mobile de vices of oad their delay-sensiti v e tasks to the BS with the highest signal strength. Str ate gy 2 (Assisted-of oading (A O)): F or J n ( t ) < 1 , the tasks will be of oaded to the satellite via the Starlink antenna. Here, the BS selects a serving satellite from a set of visible satellites. The set consist of satellites whose SNR (dened in (4)) is abo v e a set threshold Γ th . T o select the satellite that will serv e the BS from the set, the BS mak e use of a feedback mechanism that monitors the e v olution of each satellite queue system, in a round robin manner . The feedback pro vides the state of the input queues for each satellite, and then the BS estimate the queuing duration’ s { ˆ t q u } follo wed by checking if ˆ t sat < σ . Then, the satellite with the least v alue of { ˆ t q u } from the set, also fullling the latenc y constraint, of oading the tasks to the satellite, denoted by s , will be prioritized, otherwise the data will be dropped and the communication with the satellite will be deacti v ated. T o handle congestion in the computing platform input b uf fers, the proposed strate gies ma k e use of a soft-dropping polic y . Here, the tasks are dropped at the satellite edge s ystem when ˆ t sat σ and in the BS the tasks are also dropped when ˆ t l oc σ . When the tasks ha v e been dropped, the mobile de vices will back-of f for Indonesian J Elec Eng & Comp Sci, V ol. 41, No. 3, March 2026: 935–945 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 941 a random period of time denoted by , and when e xpires the de vice can retry to of oad their tasks. This allo w the input b uf fers to decongest and pre v ent the edge systems’ queues from o v erloading. In such cases, the data will be dropped with a probability of q dr op = ( ˆ t sat ) δ = ( ˆ t l oc ) δ , where δ is a parameter that describes the steepness of q dr op . The online algorithm: The distrib uted process of the online algorithm is as follo ws: The online algo- rithm starts with the initial s tate and b uilds a tree with all potential future states up to the prediction depth in a breadth-rst manner . The current system state i s then initialized to create a search set, which is then accumulated whil e the algorithm iterates through the tree, taking into consideration predictions (ener gy , traf c), accumulated w orkloads at the output b uf fer , pre vious outputs, and controls. The coll ection of states that are reached at each prediction depth, taking into account the performance metric J n ( t ) . In order to generate the ne xt set of reachable control actions, we rst esti mate the traf c load, del ay- dependent tasks, locally acceptable computational load, har v es ted ener gy , and J n ( t ) . Base on J n ( t ) , the of- oading strate gy is then selected and an y e xpected delay on the forecasted BS load and the b uf fers Ne xt, the ener gy cost β bs n ( t ) for e v ery created state is calculated. After e xamining the predi ction horizon, a series of achie v able states with the lo west ener gy use is found. The system recei v es an input control action, which corresponds to the rst state in the sequence; the others are discarded. F or e v ery time slot t , the procedure is repeated. 4. PERFORMANCE EV ALU A TION 4.1. Simulation setup A BS system empo wered with computation capabilities deplo yed in a protected area, and the S tarlink constellation, is considered in this setup. The pa rameters that were used in the simulations are listed in T able 1. Our time slot duration τ is set to 15 min and the time hori zon is set to 2 time slots. Datasets for traf c loads from [25], visitor’ s location from [17] for emulating the LS, harv ested ener gy from [22], were used in our setup. F or simulation, Python is used as the programming language. T able 1. System parameters P arameter V al ue Figure of merit ( G/T ) n 15 . 84 dB/K Carrier frequenc y , f c 30 GHz Bandwidth, B n 10 MHz EIRP Satellite antenna, EIRP n 34 . 9 dBW Earth radius, r e 6378 . 137 km Satellite height, l h 350 600 km SNR for A O polic y , Γ th 10 dB Satellite capacity , C sat 1 TB Starlink antenna po wer , β sat 50 W BSs total , N 20 BS operating po wer β 0 , 10.6 W MEC capacity , C lo c 40 GB Number of containers, G 20 Application time constraint, σ 0.8 s Battery self-char ging, ζ n 0 . 9999 Ener gy storage capacity , Υ max 100 kJ Uplink task size γ n 3 MB Do wnlink task size γ n 0 . 1 MB T ar get transmission rate, r 0 1 Mbps 4.2. Numerical r esults In Figure 3, the real and predicted v alues for the harv ested ener gy are sho wn. Here, the forecast ing algorithm tracks each v alue and predict it o v er one-step. From the obtained results, the prediction v ariations are observ ed between Υ( t ) and ˆ Υ( t ) , the obtained root mean squared error (RMSE) v alues are 0 . 050 for one-step, 0 . 070 for the second-step. The obtained accurac y is good enough for our simulation setup. Satellite-based assisted-of oading for ener gy-constr ained edg e networks (Thembelihle Dlamini) Evaluation Warning : The document was created with Spire.PDF for Python.
942 ISSN: 2502-4752 F or performance e v aluation, we compare the tw o of oading methods: (i) the ground mobile de vice directly of oads the data to the satellite (DO) without an y short-term future kno wledge, similar to [12], and (ii) our proposal of sending data via the BS for local computation (LO) or satellite-based computation (A O), utilizing short-term future kno wledge. Figure 3. Solar ener gy traces for day 1 (D1) and its predicted v alues (D1(pred)) In Figure 4, we compare the ener gy sa vings that can be obtained when partiall y some of the of oaded tasks are forw arded by the BS to the selected LEO satellite for computation (LO + A O) and other tasks com- puted locally , and cases where all the tasks are computed locally (LO). It is observ ed that in the early hours of the morning (before 8 am), there are fe w visitors in the national park (lo w acti vity), thus the BS site is not much utilized as the ener gy is used for operation acti vities. Between 9 . 00 am - 15 . 00 pm, there is high acti vity within the park. This period corresponds to periods where there is suf cient amount of ener gy that can be harv ested, as well as the arri v al and departure of visitors from the national park. After 15 . 00 pm, the visitors will start to lea v e the park, then the ener gy sa vings increases. The ener gy sa vings obtained by LO + A O ranges from 59% to 34% and LO ranges from 52% to 30%. From the obtained results, it is observ ed that partial of oading to the satellite is benecial as it relie v e t he BS site from computing e v erything locally , that is, frac tional computing is better than computing e v erything locally . Figure 4. Ener gy sa vings within the BS site Indonesian J Elec Eng & Comp Sci, V ol. 41, No. 3, March 2026: 935–945 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 943 The time-a v erage system delay cost is sho wn in F igure 5, where the computational duration is ob- serv ed o v er a number of time slots. It is observ ed that local computation (LO) bears a near real-time delay of 0 . 83 seconds when compared to the application delay requirement of 0 . 8 s econds ( σ ). In addition, we observ ed that assisted-of oading (where some of the tasks are partially of oaded to the satellite) i s benecial as it of fers delays of 0 . 82 seconds which is as close to local computation on a v erage. This is due to the f act that a portion of the tasks are of oaded to the satellite if computational resources are a v ailable on the satellite and congestion is not e xpected in the ne xt time slot. The performance for direct of oading (DO) is lo w (when compared to assisted-of oading) as the men delay is 0 . 84 seconds. By of oading some of the tasks to the satellite, via the BS, helps in that the BS (acts as relay or sink) decide on the share that can be forw arded to the satellite platform, taking into account the amount of green ener gy to be harv ested and the e xpected tasks. Figure 5. T ime a v erage system delay The drop rate from the queues is illustrated in Figure 6. Here, assisted-of oading (LO + A O) is compared with full of oading to the satellite (DO), and LO + A O achie v es a maximum drop rate of 8 . 3 % and a minimum of 2 . 6 % whereas DO achie v es a maximum drop rate of 12 . 9 % and minimum of 9 . 9 % . It is observ ed that our proposed strate gies perform better (drop rate of < 10 % ) when compared to direct of oading to the selected satellite. This is due to the loading of the input b uf fer of the satellites in the case of forw arding all the tasks to the satellite and the e xpected harv ested ener gy . Figure 6. T ask drop rate o v er a number of time-slots Satellite-based assisted-of oading for ener gy-constr ained edg e networks (Thembelihle Dlamini) Evaluation Warning : The document was created with Spire.PDF for Python.
944 ISSN: 2502-4752 5. CONCLUSIONS In this paper , we propose a h ybrid wireless communication system consisting of a base station em- po wered with computing capabilities, ener gized using green ener gy , and a LEO satellite system for pro viding mobile services to visitors within a protected area. The main goal of this research w as to minimize the ener gy consumption per communication site with a guarantee of the e xpected end-to-end latenc y . In this w ork, we put forw ard a ne w tasks of oading strate gy whereby the BS can handle some of the delay-sensi ti v e tasks locally or of oad the task to the visible satellite, using green ener gy as a performance metric. T o sa v e ener gy , the BS system forecast the short-term ener gy a v ailability and then pro v i sion the computing resources base on the fore- casted ener gy , and to guarantee the end-to-end delay the access control application on the edge serv er decide on the fraction of w orkloads to be computed locally or of oaded to the satellite. Our numerical results, obtained with real-w orld datasets, sho w via simulations that our proposed of oading strate gy (LO + A O), which mak es used of foresighted optimization in terms of green ener gy to be harv ested, pro vision of computi ng resource, and the e xpected tasks, can be able to guarantee the end-to-end delay e xpected from applications when compared with our benchmark. The ener gy sa vings obtained by LO + A O ranges from 59 % to 34 % and LO ranges from 52 % to 30 % . It is observ ed that local computation (LO) bears a near real-time delay of 0 . 83 seconds when compared to the application delay requirement of 0 . 8 seconds ( σ ). In terms of drop rate, LO + A O achie v es a maximum drop rate of 8 . 3 % and a minimum of 2 . 6 % whereas DO achie v es a maximum drop rate of 12 . 9 % and minimum of 9 . 9 % . It is observ ed that our proposed strate gies perform better (drop rate of < 10 % ) when compared to direct of oading to the selected satellite. As part of our future w ork, we will design more sophisticated of oading strate gies that include long term dependencies in terms of ener gy forecasting, other forecasting methods, and the consideration of other NTNs such as high altitude platforms. REFERENCES [1] G.-P . Nuria, et al. , “The inte grated sensing and communication re v olution for 6G: V ision, techniques, and applicat ions, Pr oceed- ings of the IEEE , v ol. 112, no. 7, pp. 676-723, 2024. https://doi.or g/10.1109/JPR OC.2024.3397609. [2] J . Y . Kim, T . K ubo, and J. Nishihiro, “Mobile phone data re v eals spatiotemporal recreational patterns in conserv ation areas during the CO VID pandemic, Scientic Reports , v ol. 13, no. 20282, 2023. https://doi.or g/10.1038/s41598-023-47326-y . [3] BEREC and RSPG joint r eport on F acilitating mobile connectivity in ”c halleng e ar eas ”, BEREC, R ¯ ıg a, Latvia, T ech. Rep., Dec 2017. [4] T . Dlamini, ´ A. F . Gamb ´ ın, D. Munaretto, and M. Rossi, “Online supervisory control and resource management for en- er gy harv esting BS sites empo wered with computation capabilities, W ir eless Communications and Mobile Computing , 2019. https://doi.or g/10.1155/2019/8593808. [5] L. Chen, S. Zhou, and J. Xu, “Ener gy ef cient mobile edge computing in dense cellular netw orks, in IEEE Internati onal Confer ence on Communications (ICC) , P aris, France, May 2017. https://doi.or g/10.1109/ICC.2017.7997128. [6] G. Marco and Z. Michele, “Non-T errestrial netw orks in the 6G era: Challenges and opportunities, IEEE Network , v ol. 35, no. 2, pp. 244–251, 2021. https://doi.or g/10.1109/MNET .011.2000493. [7] T . T aheri, R. Nilsson, and J. v an de Beek, “The potential of massi v e-MIMO on TV to wers for cellular co v erage e xtension, W ir eless Communications and Mobile Computing , 2021. https://doi.or g/10.1155/2021/8164367. [8] C. Abdelaali, et al. , “6G for bridging the digital di vide: W ireless connecti vity to remote areas, IEEE W ir eless Communications , 2021. https://doi.or g/10.1109/MWC.001.2100137. [9] N. Dinh, et al. , “6G internet of things: A comprehensi v e surv e y, IEEE Internet of Things J ournal , v ol. 9, pp. 359–383, 2022. https://doi.or g/10.1109/JIO T .2021.3103320. [10] C. Pietro, G. Alberto, M. Mario, and P . F abio, “Orbital edge of oading on me g a-LEO satellite constellations for equal access to computing, IEEE Communications Ma gazine , v ol. 60, no. 5, pp. 32–36, 2022. https://doi.or g/10.1109/MCOM.001.2100818. [11] B. Soret, I. Le yv a-Mayor g a, S. Cioni, and P . Popo vski, “5G satellite netw orks for internet of things: Of oading and backhauling, In- ternational journal on Satellite Communications and Networks , v ol. 39, no. 4, pp. 431–444, 2021. https://doi.or g/10.1002/sat.1394. [12] A. Bonora, A. T raspadini, M. Giordani, and M. Zorzi, “Performance e v aluation of s atellite-based data of oading on Starlink constellations, in IEEE W ir eless Communications and Networking Confer ence (WCNC) , Milan, Italy , Mar 2025. https://doi.or g/10.1109/WCNC61545.2025.10978515. [13] M. L ´ opez, S . Damsg aard, I. Rodr ´ ıguez, and P . Mogensen,“Connecting rural areas: an empirical assessment of 5G T errestrial-LEO Satellite Multi-Connecti vity, IEEE V ehicular tec hnolo gy confer ence (VTC ) , Florence, Italy , Aug 2023. https://doi.or g/10.1109/VTC2023-Spring57618.2023.10199206. [14] X. Jie and R. Shaolei, “Online learning for of oading and autoscaling in rene w able-po wered mobile edge computing, in IEEE Global Communications Confer ence (GLOBECOM) , W ashington, USA, De c. 2012. https://doi.or g/10.1109/GLOCOM.2016.7842069. [15] T . Dlamini and S. V ilakati, “LSTM-based traf c load balancing and resource allocation for an edge system, W ir eless Communica- tions and Mobile Computing , 2020. https://doi.or g/10.1155/2020/8825396. [16] “Mobile edge computing (MEC): Location API, ETSI, Sophia-Antipolis, France, T ech. Rep., Jul 2017. [17] K. J. Y oon, “Point location of visitor centers in National P arks and Ramsar sites in Japan (Data set)”, Zenodo: https://doi.or g/10.5281/zenodo.10066858. Indonesian J Elec Eng & Comp Sci, V ol. 41, No. 3, March 2026: 935–945 Evaluation Warning : The document was created with Spire.PDF for Python.