Introduction

The utilization of Mobile Edge Computing (MEC) has become a disruptive strategy in response to the growing need for high-throughput, low-latency applications in wireless networks. MEC brings computation closer to end- users, thereby reducing latency and diverting traffic away from centralized data centers. Thus, discussing modern cities and urban infrastructure it is impossible to ignore the concept of sustainability. In the current world, smart buildings, and smart city solutions are relatively common, and that boosts the need for efficient energy solutions. This study finds that Mobile Edge Computing (MEC) presents a way to undertake efficient resource management in these urban environments, where computation is closer to the end-user, thereby decreasing latency and energy consumption. In smart buildings, MEC can be used in building management systems commonly referred to as BMS that are basic for metering energy utilization, controlling mechanical and electrical systems and preserving sustainable practices. Advanced load balancing strategies applied to MEC can play a major role in achieving energy efficiency and sustainability of smart cities and urban infrastructure. However, the constantly changing and diverse nature of mobile environments presents substantial hurdles in ensuring efficient resource utilization and delivering optimal service [1, 2, 3, 4, 5, 6, 7]. One of the pivotal challenges in MEC lies in load balancing, which entails distributing computational tasks among edge servers to minimize response time, maximize resource utilization, and ensure equitable treatment of users. Conventional load balancing techniques frequently fail to adjust to the dynamic nature of mobile networks, wasting energy and producing less than ideal performance [8, 9, 10, 11, 12]. Given the escalating concerns surrounding energy consumption in MEC deployments, there is an urgent need for innovative load balancing strategies that prioritize energy efficiency without compromising performance. In recent years, machine learning (ML) has emerged as a potent tool for optimizing resource allocation and enhancing system performance across various domains. By leveraging historical data and real-time insights, ML algorithms can dynamically adjust resource allocation policies in response to changing network conditions and workload patterns [13, 14, 15]. Within the realm of MEC, ML-based approaches present promising avenues for devising intelligent load balancing strategies that optimize energy consumption while upholding quality of service (QoS) standards. From the user’s standpoint, a crucial scenario in Mobile Edge Computing (MEC) involves computation offloading, which can either conserve energy or accelerate the computation process. Essentially, the key decision in computation offloading is whether to execute tasks locally or to offload them. If offloading is chosen, another important consideration is determining the extent and nature of the offloading [16, 17, 18, 19, 20, 21, 22, 23, 24]. The decision on computation offloading can lead to three outcomes:

a)Local execution: In this situation, all computational processes are carried out directly on the User Equipment (UE), completely avoiding any offloading to the MEC infrastructure. Local execution might occur for various reasons, such as inadequate computational resources at the MEC or when the advantages of offloading are not deemed significant enough to justify the associated expenses. Although local execution reduces reliance on communication with the MEC, it could result in prolonged task completion durations and potentially increased energy usage, particularly if the UE has limited resources.

b)Full offloading: In full offloading, the MEC infrastructure handles the complete computation task, transferring and processing it entirely. This strategy proves beneficial when the MEC provides abundant computational resources and maintains low-latency communication links. By delegating the entire task, users stand to gain faster execution times and lower energy consumption on their personal devices.

c)Partial offloading: Partial offloading adopts a mixed method wherein some segments of the computation task are processed locally at the UE, while the rest are transferred to the MEC. This approach seeks to find a middle ground between the advantages of local execution, like decreased latency and energy usage, and the computational prowess of the MEC infrastructure. Determining which segments to offload and which to execute locally hinges on factors such as task attributes, network circumstances, and resource accessibility. Partial offloading provides adaptability and optimization prospects, enabling users to customize their offloading tactics according to current conditions and preferences.

The choice of computation offloading in MEC settings can result in three clear results: local execution, full offloading, or partial offloading. Each option comes with its own benefits and drawbacks concerning energy efficiency, latency, and resource management. Making informed decisions regarding computation offloading is vital for enhancing user satisfaction and optimizing the effectiveness of MEC systems.

This paper is organized as follows: Section II provides an overview of related work in the field of load balancing and energy efficiency in MEC. Section III define the problem formulation. Section IV outlines the proposed machine learning-based framework, including its architecture, algorithms, and key components. Section V presents experimental results and performance evaluation, followed by a discussion of findings in Section VI. Finally, Section VII concludes the paper with insights into future research directions and concluding remarks.

Related work

In recent years, the integration of MEC in sustainable urban development and smart building management has attracted increasing attention from both academia and industry. Efficient resource allocation and load balancing techniques in these contexts can significantly enhance energy efficiency, reduce operational costs, and contribute to the overall sustainability goals of smart cities. Studying MECfor load balancing holds immense importance. Significant research efforts have been devoted to this area in recent years, reflecting the growing interest and recognition of its potential impact.

Al-Khafajiy et al. [18] introduce a collaborative edge offloading technique enabling fog node collaboration for processing big data, leveraging predetermined fog parameters. This approach works well for quickly processing data at the edge level because all the necessary information about the capabilities of the fog nodes—like processors—is known in advance. Nonetheless, this method is not energy-efficient since it ignores the fact that fog nodes consume energy.

A compute offloading policy designed for a multi- user, multi-cloudlet Mobile Edge Computing (MEC) environment is introduced by Mazouzi et al. [19]. Its main goal is to reduce the offloading cost function, which depends only on energy consumption and execution time. Interestingly, this cost function ignores radio resources and the processing resources of Mobile Edge Servers (MES). However, because it’s heuristic in nature, this method could not work well for complex and dynamic applications. On the other hand, a number of academics use dynamic programming approaches like the Markov Decision Process (MDP) to determine the best course of action for offloading computing in MEC. This approach, however, requires a preset state transition probability matrix.

In order to promote power efficiency in data centers, Nguyen et al. [20] presented a VM consolidation technique based on the forecast of multiple resource use. Moreover, the relevance of load balancing in MEC extends beyond traditional computing environments to applications in smart grids, energy-efficient buildings, and sustainable city planning. For instance, Nguyen et al. [20] have demonstrated how MEC can be used to enhance power efficiency in data centers, which directly supports the sustainability objectives of urban development projects. Prediction errors nevertheless maintained the possibility of Service Level Agreement (SLA) violations and resource waste. Stochastic load balancing, which incorporates the probabilistic distribution of prediction errors into the expected output, was introduced by Yu et al. [21] as a solution to this problem. Furthermore, hotspots—overloaded servers—were identified in order to address improper VM migrations caused by overloaded servers and inefficient resource utilization. A heuristic algorithm was then proposed in order to facilitate VM migrations from hotspots to underutilized servers and fairly distribute VMs among available servers.

Later, in order to mitigate SLA violations resulting from prediction errors, an energy-efficient framework for VM prediction and migration was proposed. This framework used the Wiener filter to predict resource usage and added safety margins based on Exponentially Weighted Moving Average (EWMA) to the predicted output. Furthermore, in order to increase cloud provider revenue, an online virtual machine placement strategy was put forth in. This strategy used the Decreased Density Greedy algorithm to handle SLA violations and First-Fit and Harmonic Algorithms (HA) for online VM placement, with HA demonstrating better performance than First-Fit for energy- efficient VM placement [25, 26, 27, 28, 29, 30].

A resource allocation model based on workload and security threat predictions was more recently introduced by Saxena et al. [3]. From the perspective of sustainable urban development, efficient load balancing in MEC is not just a technical requirement but also a key enabler of reducing the environmental impact of large-scale urban infrastructure.

Furthermore, they identified hotspots—overloaded servers—and developed a heuristic technique for VM migration from these hotspots to underutilized servers in order to address problems resulting from poor resource utilization and erroneous VM migrations. This strategy sought to guarantee equitable virtual machine distribution among accessible servers. They then put forth an energy-efficient methodology for VM movement and prediction in [29]. This framework included exponentially weighted moving average (EWMA) based safety margins to the expected output and used a Wiener filter to estimate resource utilization. The purpose of this innovation was to stop SLA violations brought on by inaccurate predictions.

To strengthen the related work section by comparing the proposed approach with recent load balancing methods in MEC systems, consider the following enhancements:

A comparison with Collaborative Edge Offloading [18] show that the proposed solution is more efficient.

This method relies with cooperation between the fog nodes for big-data processing; there are specific fog parameters that are set to increase data processing rate. Nevertheless, it does not consider energy factors as it does not consider the energy level of fog nodes in its algorithms. On the other hand your proposed framework for energy-aware operations and real time resource prediction has eradicated energy wastage in a big way.

Here, a comparison of the effectiveness of each method towards the Compute Offloading Policy developed in [19] is made.

This policy tries to do this by trying to minimize the offloading cost based on energy and time of execution as shown by the equation above, without regard to radio resources or the dynamic application needs. On average, evaluation metrics such as Precision and Recall are heuristic in nature which will not be sufficient for complex search tasks. The framework for yours however employs machine learning algorithms for dynamic adjustments for the workload and they include end to end resource optimization.

Problem Formulation

In Mobile Edge Computing (MEC) settings, diverse devices produce fluctuating workloads, causing uneven usage of resources among edge servers. This non-uniform distribution can cause certain servers to be underutilized, while others may be overwhelmed, resulting in decreased performance and higher energy usage. For urban infrastructure and smart buildings, efficient load balancing is necessary to ensure that energy resources are optimally used while maintaining the required performance levels for smart systems like building management systems (BMS) and other IoT-enabled services. Balancing these demands across MEC servers is key to reducing energy waste in smart urban environments. In the context of smart cities, where devices and sensors in smart buildings constantly generate data, the need for an energy-efficient load balancing strategy becomes even more urgent. Without efficient load management, urban infrastructure systems can suffer from resource overuse, leading to increased operational costs and a higher environmental impact. Current load balancing methods frequently struggle to adjust to shifting workload trends and may overlook energy efficiency considerations. The scheduler’s effectiveness is measured using a metric referred to as Loss, which is defined for each scheduling interval. A lower Loss value indicates superior performance of the scheduler. Specifically, we denote the loss of interval SLi as Lossi.

Therefore, the scheduler, known as the Model, functions as follows: Statei leads to Actioni. The Lossi within an interval hinges on the task allocation to hosts, namely, Actioni determined by the Model. Consequently, to achieve an optimal Model, the problem can be articulated as depicted in eq (1).

Consequently, this paper aims to tackle the main issue of devising a proficient load balancing approach for MEC systems that reduces energy usage while still meeting Quality of Service (QoS) criteria. Therefore, the problem addressed in this paper is the development of a load balancing model that not only optimizes resource utilization but also minimizes energy consumption in smart buildings and urban infrastructure. The model must meet sustainability goals while maintaining the performance required by modern urban environments.

Assumptions:

Predictable Workloads: Dependent on accurate scheduling of loads on the basis of historical information.

Homogeneous Servers: Expected energy efficiency when the levels of server load are constant.

Real-Time Monitoring: Relies on proper and up-to- date resource consumption information.

Impact on Applicability:

Beneficial for organizations operating under systemic conditions, for instance, smart cities with standardized…

May not be easily used in dynamic or heterogeneous scenarios such as disaster management.

Outlining these assumptions helps to define where and how the model works and where it should be developed further.

Proposed Framework

In the context of smart cities and sustainable urban development, an efficient load balancing framework is essential to manage the vast number of IoT devices and infrastructure components that continuously generate data. The proposed framework aims to enhance energy efficiency, optimize resource utilization, and reduce operational costs in smart urban environments and buildings. We present the tiered approach to the compute offloading framework in this subsection. The computationally difficult part of the code can run locally or remotely, depending on the amount of processing required. As previously mentioned, the suggested framework is based on a three-layer compute architecture that consists of the edge, cloud, and smart device layers. The proposed load balancing framework mainly based on MEC is very flexible and potentially applicable to other smart cities’ infrastructures like smart transportation, and smart waste disposal systems. In smart transportation, it can enhance resource management in traffic signals, self-driving cars, and routing, and enhance resource utilization while lowering the energy consumption rate. In waste collection planning the framework within the Internet of Things technology can improve the way that data is collected from bins and collection trucks which can also improve the efficient collection time and consumption on energy. Furthermore, it is equally suitable to be adopted in energy grid, water distribution and public safety systems with a resultant enhanced energy efficiency in the systems. This versatility underlines the strength of the studied framework as the means to promote sustainability and efficiency in multiple smart city solutions. Figure 1 shows the overall architecture of this framework. The suggested system, known as OCLD-MEC, consists of four stages: resource prediction, load balancing and optimization, job segmentation and resource management, and service level agreement (SLA) measurement based on user expectations. The integration of machine learning techniques in the proposed framework allows for real-time prediction of resource usage, which is particularly important for smart buildings and urban infrastructure where energy consumption needs to be carefully monitored and managed. The proposed MEC-based load balancing framework employs several machine learning (ML) algorithms, with renowned optimization capabilities for resources in dynamic settings, hence, energy efficiency. RL utilized for online decisions which is used in optimizing long-term reward and slightly vary environment gives it an aptness to be implemented in dynamic MEC systems. For effective distribution of the load, it applies support vector machines (SVM) for classifying system states according to performance attributes such as utilisation rates and energy use. Resource consumption and task distribution forecasts use Random Forest because of the many cross-relationships involved. Neural networks RNN in particular can be used for Sequential data analysis, for example forecasting future usage of resources given past usage. Last is the K-Means Clustering which clusters like tasks, so that they can best be fit within the cramming schedules hence the best utilization of resources. These algorithms collectively help make the framework adaptive and to cope up with changing conditions to achieve better energy efficiency, low response time and better conformity to SLA in MEC systems.As shown in Figure 1, these stages taken together offer a tangible and ideal answer to the problem of elastic resource management in a mobile edge computing environment. They communicate with one another and work together. By their interactions, these four stages function as a team. The first phase consists of the processing of applications {Application1, Application2,..., ApplicationM} that have been submitted by users {User1, User2,..., UserM}. These applications are separated into smaller units known as tasks {T1, T2,..., Tz}, and they are given to computing instances {VM1, VM2,..., VMQ} for processing. At the second stage, virtual machines (VMs) carry out activities and forecast resource usage for the upcoming interval at the same time. Every running virtual machine (VM) has a unique online predictor that is configured to estimate resource consumption. In addition, every server has its own load analyzer, which aggregates estimated resource utilization data from every virtual machine and predicts instances of over or under load. In the third phase, activities are triggered to alleviate the impact of overload, which is identified if the expected resource demand exceeds the server’s capacity.

Figure 1.

Proposed framework.

In the third stage, load balancing is achieved by assigning virtual machines and migrating them in order to improve resource usage by efficiently controlling overloaded or under-loaded servers. In smart cities, this framework can be used to optimize the management of critical infrastructure such as energy grids, public transport systems, and smart buildings, ensuring that resources are utilized efficiently while maintaining high service quality.

The cluster manager receives periodic updates from the load analyzer regarding the expected load on each server within the cluster. This data is used in multi- objective VM placement optimisation for VM allocation on the selected server. The online predictor detects overload or underload conditions, and when it does, it starts the appropriate VM migrations to distribute the burden evenly throughout the cluster. In order to save electricity, any idle servers are also shut down.

In the fourth phase, various metrics such as execution delay, energy consumption, accuracy, and throughput are evaluated to meet user expectations. If the user’s expectations are not met, task partitioning is revisited, and resource management is rescheduled based on the user’s demands. By implementing this energy-efficient framework, smart cities and sustainable buildings can minimize their carbon footprint, reduce energy consumption, and ensure the optimal performance of IoT-based services and building automation systems.

The load balancing algorithm aims to allocate tasks dynamically across Mobile Edge Computing (MEC) servers while optimizing resource utilization, minimizing energy consumption, and reducing SLA violations. Below is a step-by-step breakdown of the algorithm followed by the pseudo code.

Step-by-Step Breakdown:

Task Evaluation and Load Analysis: The algorithm begins with estimating load on each of the MEC servers during each time slot. This load depends with the currently active resources for instance the CPU and memory and also with the incoming tasks. For each server the system calculates a load which represents the current load with respect to the available resources as well as the expected future load.

Task Offloading Decision: Groups of tasks are supposed to be given to servers depending on the load they can be assigned to. If a server has low utilization, some of the applications or work load are migrated from servers that are heavily utilized. The offloading of other tasks is consciously done when looking for other servers to switch to, depending on parameters such as current load, energy and response time.

Load Transfer Optimization: The system also quantifies the Total Utilization Load Transfer Overhead (TULTO) which is used to estimate the transfer of tasks across servers. The first tasks are of those which have least overhead so as to avoid wastage of resources. Server capacity is also crosschecked. When a server gets too busy it transfers its workload to other servers with openings.

Load Balancing Iteration: For each time interval, hence, load balancing is continuously operational to handle task distribution across servers. There is an aim to provide a balanced load or degree of server usage, which will indirectly affect the effectiveness of the system.

Service Level Agreement (SLA) Compliance: The algorithm also attempts to check the ad hoc constraints and determine whether the tasks meet the specified SLA. Where there is a breach on Service Level Agreement (SLA) for instance, with regards to response time, the algorithm gives priority to a task if it is more urgent so as to contain the breaches.

Energy Optimization: Next, the load is balanced to also allow the algorithm to adjust the resources used in order to reduce energy usage. Spare servers are switched off to save energy, and optimal performance of tasks is valued.

Experimental Results

To assess the applicability of the proposed framework in sustainable urban development and smart cities, we conducted experiments that simulate real-world conditions in smart buildings and urban infrastructure. These experiments focus on optimizing resource allocation and energy efficiency in Mobile Edge Computing (MEC) environments, which are essential components of smart city infrastructure.

We will evaluate the suggested approach to solving the computation offloading issue using the conducted prediction in this part. In the context of smart buildings and urban systems, our experimental setup evaluates the ability of the framework to reduce energy consumption, optimize load balancing, and ensure the efficient operation of IoT devices and building management systems (BMS).The pros and cons of offloading are then demonstrated by contrasting metrics between situations with “Local Execution” and “Remote Execution”. The experiments were designed to simulate conditions typical in smart cities, including the fluctuating resource demands of IoT-enabled systems in buildings, transportation, and energy management. By evaluating the framework in these scenarios, we aim to demonstrate its potential for improving energy efficiency in sustainable urban environments. The overall workflow of the proposed framework is depicted in Figure 2, illustrating the interactions between resource prediction, optimization, and task segmentation.

Figure 2.

Flow chart of proposed work.

Steps for Load Balancing

The detailed steps of the proposed load balancing algorithm are outlined in Table 1, providing a clear breakdown of the methodology for task offloading and resource allocation.

Step 1: Every Small Base Station (SBS) evaluates the load of users in its coverage area during each time slot, producing the Lmean value, which is the mean load line. The set of Resource Load Level-Mobile Edge Computing (RLL-MEC) servers, indicated by S, can be determined using this Lmean value. The total of the load of the job being requested during that time slot and the current load (Lcurrent) is the actual load of a Mobile Edge Computing (MEC) server (Lapply). RHL-MECs (Resource Heavy Load-MECs) with actual loads greater than Lmean make up set M.

Step 2: The task offloading instruction value X(n,m) equal to m indicates that a task (n,m) requested in this time slot will be executed in an Individual MEC (IMEC) if it originates from a MEC server that is not part of set M.

Step 3: For every job (n,m) from MEC servers in set M, the Total Utilization Load Transfer Overhead (TULTO) is computed. It is established what its load capacity (Lcapacity) is for each MEC server in set S. As not every User Equipment (UE) connected to the RLL-MEC may request tasks at the same time, Lcapacity denotes the amount of load that may be delivered to an RLL-MEC to maintain long-term load balancing. To avoid transferring an excessive amount of load to the RLL-MEC, this precaution is required.

Step 4: The main idea behind the load balancing algorithm is to take the task with the lowest TULTO from the RHL-MECs in set M, whose actual load is the greatest of all the MECs in set M, and move it to the RLL-MEC in set S that has the lowest real load each iteration. Next, the index of the chosen RLL-MEC is assigned to the task offloading instruction Xn,m. Then, parameters are updated, such as the RLL-load MEC’s capacity and the real loads of the RLL-MEC and RHL-MEC. Load transfer only happens when the RHL-load MEC’s is greater than Lmean; if not, the loop ends. Moreover, the RLL-MEC is taken out of set S before it gets empty to end the loop if its load capacity is not high enough to support the task’s load lNn,m, signalling possible load instability.

Step 5: Even after the iteration, some tasks in set A might not receive instructions to offload and should be completed in their respective IMECs since they have Time to Task Load Overhead (TTLO).

Simulation Environment

Python is used as the simulation tool in this work, and it makes use of crucial libraries like Numpy, sklearn, keras, and pandas to forecast important metrics like request latency, request energy consumption procedures, offloading decision-making, and uplink decision-making selection. Twenty percent of the input dataset is used for testing the accuracy of the model, and the remaining eighty percent is used for training in order to visualize the associated offloading metrics. Using the HMM decision system, the offloading decision-making problem is visualized. This study aims to forecast the CPU usage of twenty virtual machines that are chosen at random. A CPU utilization dataset from Bitsbrain, a platform that offers real-time CPU consumption information for every virtual machine every five minutes across more than 1000 nodes spread across 700 locations worldwide, is used to test the suggested model. The Bitbrain dataset includes actual CPU usage collected from CPU logs viewed in more than 700 locations of over 1,000 VMs with an update every five minutes. It simulates practical use-cases related to smart cities such as smart buildings and IoT and therefore is suitable for researching MEC energy-efficient load balancing. Bitbrain was particularly chosen for it used in different workloads and acceptance in research and thereby supports good evaluations. In this study, for evaluation and validation of the proposed framework, energy consumption, response time and SLA violation were considered where it was used for training and testing of models. The Bitsbrain VM traces, as shown in Figure 3, provide insights into CPU usage patterns across data centers for various workloads, highlighting the challenges of resource management in MEC.

Figure 3.

Bitsbrain VM traces.

To ensure credibility and reproducibility, detail how energy efficiency is calculated:

Metrics:

Energy Consumption (EC): Total energy, kWh.

Energy Savings (ES): Rounded percentage reduction compared to baseline methods due to the limitation of space:

Measurement: Include real datasets such as Bitsbrain data set for energy consumption data. Imitate MEC environments with different loads dynamically and measure energy during task load and transitions.

Baseline Comparison: To decide on ‘best fit’ also compare it with other fitting methods like the First Fit (FF) and Random Fit (RF).

Reproducibility: Make all the Datasets splits, model parameters, configurations the experiments were based on accessible to be reproducible.

Figure 4 shows resource usage trend with oversubscription. while Figure 5 illustrates usage without oversubscription, emphasizing the efficiency of the proposed model under different workload scenarios.

Figure 4.

Trends in resource usage by data center for various workloads in Bitsbrain virtual machine traces with oversubscription.

Figure 5.

Trends in resource usage by data center for various workloads in Bitsbrain virtual machine traces without oversubscription.

The power consumption per data center under oversubscription conditions is detailed in Figure 6, demonstrating significant energy savings achieved through the proposed framework.

Figure 6.

Bitsbrain VM traces show the power consumption per data center for various workloads with oversubscription.

We chose the First Fit (FF), Best Fit (BF), and Random Fit (RF) heuristics as our baseline algorithms in order to assess different VM placement strategies. Modifications to these algorithms have led to the development of numerous better VM placement techniques. For comparison, we thus used FF, BF, and RF with and without oversubscription (or prediction). Our suggested load balancing (LB) approach with oversubscription (P1) is compared to FF, BF, and RF with oversubscription (designated as F1, B1, and R1, respectively) in the upcoming figures. Furthermore, FF, BF, and RF without oversubscription—abbreviated as F2, B2, and R2—are contrasted with our suggested LB without oversubscription (P2).

The results show that the proposed framework significantly improves energy efficiency and reduces response times in smart city environments, making it an ideal solution for managing urban infrastructure and smart building systems. By efficiently balancing computational loads, the framework supports the sustainable operation of critical urban infrastructure while minimizing energy waste. The load balancing schemes employing First Fit (FF), Best Fit (BF), and Random Fit (RF) heuristics adhere to the resource utilization pattern, with BF being the highest, followed by FF, and then RF for each workload. In contrast, Figure 7 shows power consumption without oversubscription, revealing the potential inefficiencies in traditional load balancing strategies.

Figure 7.

Bitsbrain VM traces show the power consumption per data center for various workloads without oversubscription.

It’s noted that the percentage of power savings is highest in data centers with 200 VMs, gradually decreasing as the data center size increases. Eventually, power savings stabilize between 35-45% regardless of data center size. Less active servers and virtual machine migrations made easier by the suggested architecture in overcrowded cloud environments are the causes of this trend. Surprisingly, BB workloads show the biggest power usage reduction. These findings highlight the potential of the proposed framework to enhance the sustainability of urban systems by reducing the carbon footprint of smart buildings and improving the energy efficiency of IoT-based services.

This is explained through Table 2 of Performance comparison that our online resource forecasting engine correctly forecasted that between 40 and 45 percent of the virtual machines (VMs) in the BB dataset would have CPU usage percentages below 1%. Because of this accurate prediction, power usage decreases as the number of active servers decreases. The suggested model has 32.7% fewer SLA breaches than the REINFORCE

policy, as illustrated in Figure 8. Once more, this is the result of fewer migrations and smart work scheduling to avoid huge loss values from SLA breaches. Furthermore, Figure 9 demonstrates that the suggested model, which is 6.74% faster than the previously conducted model and the best among the baseline algorithms, offers the lowest average response time out of all the scheduling policies. Figure 10 displayed the mean squared error over number of iterations. Lower the MSE of the model higher will be the accuracy of the model.

Figure 8.

Fraction of SLA Violations.

Figure 9.

Average Response Time.

Figure 10.

Mean squares error of the proposed model.

In order to compare the performance of the proposed MEC-based load balancing framework with the traditional methods of load balancing the following table is made to summarize the results. The table highlights key metrics such as energy efficiency, response time, and SLA violations:

This table also presents quantitative details of the comparative analysis of the proposed efficient framework with the conventional approach and underlines the gains made by the former in terms of energy efficiency, response time, percentage of SLA violation and model accuracy.

Conclusion

This study demonstrates the potential of integrating Mobile Edge Computing (MEC) with machine learning techniques to enhance energy efficiency and optimize resource utilization in sustainable urban environments and smart buildings. The proposed load balancing framework provides an effective solution to the growing energy demands of IoT-based services, particularly in urban infrastructure. This study tackles resource management challenges through the introduction of a load balancing framework. The framework aims to leverage oversubscribed cloud environments efficiently to decrease power consumption, enhance resource utilization, and mitigate performance degradation brought on by overloads. “By applying this framework in smart cities, we can significantly reduce energy consumption in critical urban infrastructure such as transportation systems, energy grids, and building management systems (BMS), supporting long-term sustainability goals. Extensive simulation experiments using the real-world Bitbrain dataset show that our approach reduces energy consumption by 13.4%, response time by 6.74%, SLA violations by 32.7%, and increases model accuracy by 1.06% compared to previous models.” The integration of energy-efficient load balancing techniques into urban systems not only reduces operational costs but also minimizes the environmental footprint of smart cities, making this approach vital for sustainable urban development. Moreover, the framework contributes to reducing the energy consumption of overcrowded data centers by balancing the computational load among underused servers, which further minimizes energy waste. Our results demonstrate that the proposed framework reduces energy consumption by 13.4%, decreases response time by 6.74%, and minimizes SLA violations by 32.7%, making it a valuable tool for optimizing the operation of urban infrastructure and smart building systems. The proposed architecture might be extended in the future to include other goals, such as putting in place a VM allocation system based on dependability and trust. Moreover, jobs might be grouped based on expected resource usage, allowing proactive VM auto-scaling to improve cloud data center performance. Prospective real-world application of MEC-based load balancing in smart cities offers a number of challenges. First, it can interfere with working integrated with other legacy systems because integrated applications may not always be compatible with preexisting software. This can be worked around by creating adapter or middle ware modules that integrate new framework with the old stacks for the data throughputs. Another issue is scalability because coordinating numerous, various, and large loads at multiple edge servers may be demanding on resources. This can be sorted out by use of the hierarchical load balancing and clustering to ensure the load is well balanced. They are important factors mainly due to the sensitivity of data and the fact that real time data is shared across networks. A firm can protect data transactions and Its processing through proper encryption and access control. Moreover, frank resource tracking is crucial for decision-making, and the simultaneous use of modern real-time tracking tools, and other types of predictive analysis tools, will reduce such deviations or slowness. Lastly, achieving the objective of increasing efficiency while at the same time sustaining performance is an important task of the company. Optimizing the already developed energy aware algorithms and integrating renewable energy resources in the system can sustain the system. With such strategies focused back at these challenges, it will be possible to provide optimal loads of MEC computation where feasible in smart city architectures.