Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 8, No. 6, 7878-7901 2024 Publisher: Learning Gate DOI: 10.55214/25768484.v8i6.3717 © 2024 by the authors; licensee Learning Gate © 2024 by the authors; licensee Learning Gate * Correspondence: byeon@inje.ac.kr Optimizing quality of service forecasting in mobile networks through modified walrus optimization and multivariate approaches Bandu Uppalaiah1, D. Mallikarjuna Reddy2, Vediyappan Govindan3, Haewon Byeon4* 1Department of Mathematics, Hyderabad Institute of Technology and Management, Hyderabad, Telangana, 501401, India; upendar.sourav@gmail.com (B.U.). 2Department of Mathematics, GITAM (Deemed to be University), Hyderabad, Telangana, 502329, India; drmallikreddyd@gmail.com (D.M.R.). 3Department of Mathematics, Hindustan Institute of Technology and Science, Chennai; vedimalawi@gmail.com (V.G.). 4Department of AI Big data, Inje University, Gimhae, 50834, Republic of Korea; byeon@inje.ac.kr (H.B.). Abstract: This paper presents Ensemble-based Service Quality Prediction (EAQP), an automated method for predicting service quality under changing mobile network conditions. EAQP incorporates data preparation methods such as transformation, purification, & imputation, and then performs feature extraction utilizing statistical, geographical, as well as temporal approaches. An improved feature selection method, using a unique weighting approach and optimized by a modified Walrus Optimization Algorithm, improves the accuracy of predictions. EAQP utilizes a variety of prediction models such as support vector regression, recurrent neural network models, bi-directional short-term long-term memory networks, extreme learning machines, along with multi-layer perceptron neural networks to enhance predictive accuracy. EAQP uses complex optimization algorithms and ensemble learning approaches to provide precise and dependable predictions about service quality in real-time. This helps in proactive network management as well as improvement. This comprehensive approach shows potential for boosting network efficiency, optimizing the distribution of resources, and enhancing the end-user experience when using mobile communications systems. Keywords: Ensemble-based prediction, Feature extraction, Recurrent neural networks, Service quality prediction, Walrus optimisation algorithm. 1. Introduction In the ever-evolving landscape of mobile networks, the provision of high-quality services, particularly in the do- main of Voice over Internet Protocol (VoIP) traffic, stands as a critical challenge. As the demand for seamless communicationexperiencescontinuestosurge, theneedforaccuratecharacterizationandforecastingof VoIPtraffic becomes paramount to ensure optimal Quality of Service (QoS)[1][2]. This paper delves into the multifaceted realm of multivariate time series analysis to characterize and forecast VoIP traffic in real mobile networks, with the ultimate goal of enhancing QoS prediction. The proliferation of mobile devices and the ubiquity of high-speed data networks have transformed the way individuals and businesses communicate [3]. VoIP technology, leveraging the Internet as a medium for voice communication, has become a corner- stoneinthisparadigmshift.However,ensuringaconsistentandhigh-quality VoIPexperienceposesaconsiderable challenge,giventhedynamicandunpredictablenatureofmobilenetworkconditions.Inthiscontext,thecharacte rization and forecasting of VoIP traffic through multivariate time series analysis emerge as indispensable tools[4]. By understanding the intricate patterns and interdependencies within the time series data, network operators can make informed decisions to optimize QoS parameters. This significance is amplified in cellular 7879 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 8, No. 6: 7878-7901, 2024 DOI: 10.55214/25768484.v8i6.3717 © 2024 by the authors; licensee Learning Gate environments,wheretheheightenedunpredictabilityofvariables,suchasinterference,concurrentreal- timesessions, and the dynamic load of mobile network nodes, presents intricate challenge[5]. We address these challenges by employing a multivariate predictive time series analysis of Voice over Internet Protocol (VoIP) traffic within an urban Long-Term Evolution Advanced (LTE-A) environment. Currently, LTE standsasthepredominantbroadbandtechnology,encompassing57%ofglobalusers.Legacytechnologieslike2 G and 3G persistently find use, constituting about 38% of subscriptions. In contrast, 5G comprises approximately 5% of subscriptions, primarily due to its market infancy [6][7]. Notably, the prevalent deployment method is Non-Standalone (NSA) 5G, where a significant portion of the LTE core network is repurposed to implement voice services like Voice over LTE (VoLTE) [8]. The widespread adoption of LTE has spurred numerous studies exploring Quality of Service (QoS) and Quality of Experience (QoE) metrics, covering aspects such as deployment strategies, resource allocation, probabilistic models, and coexistence with other technologies. However, our primary contribution lies in the multivariate time series characterization of the dynamic (time-varying) behavior of crucial VoIP metrics, elucidating their mutual influence [9]. Multivariate time series analysis allows for a holistic exploration of the intricate dynamics inherent in VoIP traffic within real mobile networks. Unlike univariate analysis, which focuses on a single variable, the multivariate approach considers multiple interrelated variables simultaneously [10]. This includes parameters such as network latency,jitter,packetloss,andotherrelevantmetricsthatcollectivelyinfluencetheQoSexperiencedbyVoIPuser s. Through the utilization of advanced statistical and machine learning techniques within the multivariate time series framework,itbecomespossibletocapturethecomplexrelationshipsanddependenciesamongthesevariables [11][12]. This comprehensive understanding is vital for constructing accurate predictive models that can forecast future trends in VoIP traffic and, consequently, anticipate changes in QoS systems[13]. In Section III, we delve into background information, with a specific emphasis on current methodologies, including discussions on CNN, GRU, LSTM, and Random Forest. Section IV is dedicated to comprehensive evaluations and comparisons of forecasting performances. Lastly, Section V functions as the concluding segment, presenting a summary of findings and offering insights into potential avenues for future research in this domain. 2. Related Work Several researchers have investigated the forecasting of wireless traffic usage through a variety of methods and approaches, elucidating diverse techniques. The following outlines some of these methodologies. [14]addressedissuesinmultivariatetimeseriesgenerativemodellingbypresentingauniquetechnique that integrates state-space models (SSMs) with transformer architectures. This technique, unlike previous SSMs, uses attention processes to capture complicated non-Markovian dynamics, avoiding the requirement forre- currentneuralnetworks.Theexperimentalfindingsrevealedthattheyoutperformbaselinesinavarietyoftasks and datasets. [15]tackled Forecasting using multivariate time series issues by Convolutional network with spatial and temporal components (STCTN) is a novel model based on the Transformer library. The model’s use of continuous positional encoding improves predictions muchfurther. [3]Predictingthebehavioroftrafficinrealtimeinmobilitysituationsmightassistoperators inproperlyplanningtheirnetworkinfrastructureandoptimizingresourceallocation.Asaresult,theauthorsad vocatedinthispaperthatapredictivestudyofcriticalQoS/QoEcharacteristicsofVoIPtrafficthatisinarealmobi le context be performed. [16] examined 6.2 million real network time series of long-term evolution (LTE) data traffic as well as associated parameters, such as eNodeB-wise Physical Resource Block (PRB) utilization, with the goal ofdevelopingatrafficforecastingmodelusingmultivariatefeatureinputsalongwithdeeplearningalgorithms. [17]an end-to-end generative model known as E2GAN was suggested for estimating missing values during multivariate time series. Missing values, which occur in the majority of multivariate time series, 7880 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 8, No. 6: 7878-7901, 2024 DOI: 10.55214/25768484.v8i6.3717 © 2024 by the authors; licensee Learning Gate obstructfurtheranalysisofmultivariatetimeseriesinformation.Existingimputationmethodsincludedeletio n,statisticalattribution,machinelearning-basedimputation,andgenerativeimputation.[18]Multivariable time series prediction is built as a sequence to sequence scenarios for non-periodic datasets in this paradigm. It is suggested to use multichannel residual blocks together with an asymmetric structure based on a deep convolution neuralnetwork. [19]proposed a novel approach for predicting both univariate as well as multivariate time series using a mix of clustering, classification, and forecasting techniques. The proposed algorithm’sprimary purpose is to first use a clustering technique to group frames of time series data with similarpatterns. 2.1. Research Gap The existing literature on wireless traffic forecasting reveals a gap in improving accuracy and efficiency in predicting network performance, especially in mobile networks. This research aims to address this gap by introducing a novel approach that combines Modified Walrus Optimization (MWO) with multivariate forecasting techniques. This approach aims to optimize the prediction of Quality of Service (QoS) metrics, such as network throughput, latency, and packet loss, in mobile networks. The research incorporates MWO, a metaheuristic optimization algorithm inspired by walrus behavior, to enhance the accuracy and efficiency of QoS forecasting. The integration of multivariate forecasting techniques allows for the consideration of multiple input variables, such as network traffic data, user mobility patterns, and environmental factors, in predicting QoS metrics. This comprehensive approach enables a more holistic understanding of network dynamics and facilitates more accurate predictions of QoS performance in real-time mobile network environments. The proposed research contributes to wireless traffic forecasting by introducing a novel methodology that enhances the accuracy and efficiency of QoS predictions, ultimately benefiting network operators in improving network performance and user experience. 3. Background 3.1. Ensemble Learning Inensemblelearning,aparticularcomputerintelligenceissueissolvedbysystematicallygeneratingandco mbining a number of models, including classifiers or experts. To improve a model’s performance (in classification, prediction, linear regression, etc.) or reduce the possibility of making a poor model selection unintentionally, ensemble learning is often used. Additionally, ensemble learning is used to instill a degree of trust in the model’s selection, chooseoptimal(ornear- optimal)features,fusedata,learnincrementally,non-stationarity,andcorrectforerrors. Improved prediction performance, including such reduced regression error as well as high classification accuracy,isachievedviatheapplicationofensemblelearning.Bymixingseveralmodels,ensemblelearningmay boost machine learning performance. When compared to using only one model, this strategy yields far more accurate predictions.Thecentralconceptistoeducateapanelofexperts(classifiers),whowillthencastafinalvote[20]. Each model’s predictions are counted as a "vote" in the competition. The bulk of the models’ predictions are usedtoformthefinalforecast.Forexample,inregressionproblems,averagingmaybeusedtogeneratepredictio ns, andinclassificationissues,itcanbeusedtocalculateprobabilities.Oneofthesimplestwaystocombinetheresults of several machine learning approaches is via the use of votes. The voting classifier encapsulates a suite of several classifiers which are trained as well as assessed in parallel to capitalize on the strengths of each method. The final result of a prediction is decided by a vote made by one of twomethods. Hardvoting/majorityvoting:In its simplest form, majority voting, or "hard voting," is the method most often used. The category with the most votes, Nc (yt), will be chosen. Through averaging the results of all classifiers, we make a prediction for the y-class label. 1 2ˆ arg max( ( ), ( ),...., ( ))n c t c t c ty N y N y N y= 7881 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 8, No. 6: 7878-7901, 2024 DOI: 10.55214/25768484.v8i6.3717 © 2024 by the authors; licensee Learning Gate Let's pretend we've decided to combine different classifiers that label a training sample with in following ways: • Classifier 1 -> class 0 • Classifier 2 -> class 0 • Classifier 3 -> class 1 y^=mode {0,0,1} =0 By a large margin, we have decided that this sample belongs in "class 0." 3.2. Walrus Optimization Algorithm (WaOA) The Walrus Optimization Algorithms (WaOA) is a metaheuristic that is population-based, with its population members represented by walruses. In WaOA, these walruses symbolize potential solutions to the optimization problem, and their positions Define the specifications for issue variables in the search space. As a result, each the walrus is seen as a vector, and the whole community of the walrus is mathematically represented expressed as a population matrix. Initially, walrus populations are formed at random during the introduction of WaOA. The WaOA population matrix's construction is precisely defined using equation (1). 1,1,1 1, 1 ,,1 , ,1 , , * * j m i ji i m i N N j N m N N M N m xx x X xx x X X x x x X              = =                (1) In the given context, the population of walruses is represented as 𝑋, where each individual walrus, denoted as 𝑋𝑖, stands for a candidate solution. Within this framework, 𝑥, signifies the value proposed by the 𝑖th walrus for the 𝑗th decision variable. The population comprises 𝑁 walruses, and the problem involves 𝑚 decision variables[21]. Each walrus serves as a potential solution to the issue, as well as recommended values for variables to consider allow us to calculate the objective function[22]. The estimated objective function values resulting from the contributions of these walruses are defined in equation (2). 1 1 *1 *1 ( ) ( ) ( ) i i N NN N F F X F F F X F F X                = =                 (2) Here, F is the vector of goal operations, with each element represented as Fi, which represents the individual The desired function's value is obtained from the inputs that are provided of the ith walrus. Algorithm: pseudocode of WaOA Start WaOA 1. Input entire optimisation problem data 2. Set The total amount of iterations (T) and the number of walruses (N) 3. Locations of walruses are initialised. 4. For t=1:T 5. Update strongest walrus based on objective function value criterion 6. For i=1:N 7. Phase1: Feeding strategy (exploration) 7882 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 8, No. 6: 7878-7901, 2024 DOI: 10.55214/25768484.v8i6.3717 © 2024 by the authors; licensee Learning Gate 8. Determine the new position of the jth walrus with ( )1 , , , , ,. . P i j i j i j j i j i jx x rand SW I x= + − 9. Update the ith walrus location using 1 1, , , , P P i i i i i X F F X X else   =   10. Phase2: Migration 11. Select the ith walrus's immigration destination. 12. Find the jth walrus's new location with  ( ) ( ) 2 , , , , , , , , , , . . , , . , , P i j i j i j k j k j i j k i i j i j i j k j x x rand x I x F F x rand x x else = + −  + − 13. Update the ith walrus location using  2 2, , , P P i i i i i X X F F X else =  14. Phase 3: Escaping and fighting against predators 15. Calculate a new position in the neighbourhood of the ith walrus using ( )( )3 , , , , , , P t t t i j i j local j local j lacal jx x lb ub rand lb= + + −   , , : jt local j jt local j lb Local bounds lb t ub ub t = = 16. Update the ith walrus location using 3 3{ , , , , P P i i i i i X X F F X else =  17. end 18. Keep the best possible answer so far 19. end 20. Provide WaOA's most effective quasi-optimal solution for the given problem 21. End WaOA 7883 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 8, No. 6: 7878-7901, 2024 DOI: 10.55214/25768484.v8i6.3717 © 2024 by the authors; licensee Learning Gate Figure 1. Flowchart for WaOA. From the above algorithm of modified WaOA enhances the generalized Walrus Optimization Algorithm in many ways. Adaptive step sizes, dynamic migration computations, & perturbation enhance exploration, migration, and escape. Considering factors outside objective function values as well as dynamically improving parameter selections improves selection. The modified WaOA may additionally use hybridization via other optimization techniques or problem-specific information to increase performance, converging speed, solution quality, & robustness across optimization difficulties. 4. Methodology Thisresearchtakesanewapproachtothepredictionofservicequality.Theproposedstrategyisdividedintoman y phases.Fortheapproachtowork,firstadatasetofnetworktrafficiscollected.Theincomingdatasetisthenplaced through a pre-processing procedure that makes use of methods including data transformation, data purification, and data imputation. After the data has been pre-processed, it is used to extract features. Temporal feature extraction, Statistical and spatial techniques are used to extract important characteristics from the data. To optimize the featureselectionprocess,eachfeatureisindividuallyoptimizedthroughthenewlydevelopedWalrusOptimizat ion Algorithm.Onceoptimalfeaturesareobtained,theyareinputtedintoanEnsemble- 7884 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 8, No. 6: 7878-7901, 2024 DOI: 10.55214/25768484.v8i6.3717 © 2024 by the authors; licensee Learning Gate basedpredictionmodelforclassifying network traffic data, ultimately facilitating the prediction of quality of service. The Ensemble prediction incorporatesfunctionalitiesfromGRU,LSTM,andRandomForest.Everypredictionapproach’shyper- parameters are adjusted by the use of the Walrus Optimisation Algorithm. This suggested methodology’s examination shows that it has a high degree of accuracy when forecasting the network traffic’s quality of service. The integration of advancedoptimizationtechniquesenhancestheefficiencyoffeatureselectionandmodeltuning,contributingt othe overall effectiveness of the predictivemodel. 4.1. Dataset Description The network traffic in the dataset was collected in an authentic cellular environment in and around Salerno, Italy, which is categorized as a medium-density city (around 2000 people/Km2). As of March 2023, over 100 radio towers covering a combination of LTE/LTE-Advanced (roughly 97%) and 5G-NSA (roughly 3%) technologies service this region (information obtained from https://www.nperf.com/en/map/IT/). 4.1.1. Data Pre-Processing Theinitialstageoftheproposedmodelinvolvestheutilizationofapre- processingtechniqueappliedtothecollected network traffic data, denoted as Az. This technique aims to eliminate unnecessary attributes, thereby improving overall performance. Common issues addressed during pre-processing include outliers, missing values and redundant data. The enhancement of model accuracy is achieved through data imputation, data cleansing, and data transformation. • DataImputation: In order to deal with missing values in the input data Az, data imputation is used. Data points that are missing are substituted with comparable values, such zero or the sample mean. As an alternative, imputation might include giving the missing data the closest value, with the imputed data being shown as imp zA . • DataCleansing: Data cleansing is a technique designed to identify and eliminate errors and inconsistencies in the imputed data imp zA . Input data often contain noise, outliers, unwanted attributes, and irrelevant information. The presence of such elements can lead to increased computational time and errors in analysis. Data cleansing resolves these issues by removing redundant data, enhancing performance accuracy, and reducing computation time. The resulting cleansed data is denoted as cle zA . • DataTransformation: The transformed data, denoted as cle zA , undergoes data transformation through normalization and aggregation. Given that ambient data encompasses various particle types (solid, liquid, gas), data transformation plays a crucial role in converting one format into another. This transformation facilitates easier prediction of air quality and enhances performance analysis. The outcome of data transformation is denoted as tra zA . The culmination of the pre-processing steps yields the final pre-processed data, denoted as tra zA . This refined data is then forwarded to the feature extraction stage, contributing to the subsequent phases of the proposed model. 4.2. Feature Selection and Feature Extraction Using an Improved Optimization Algorithm 4.2.1. Feature Extraction In the proposed model, the identification of specific attributes from the pre-processed data, denoted https://www.nperf.com/en/map/IT/ 7885 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 8, No. 6: 7878-7901, 2024 DOI: 10.55214/25768484.v8i6.3717 © 2024 by the authors; licensee Learning Gate as tra zA , is essential. The primary aim of feature extraction is to enhance prediction analysis and address overfitting challenges, especially when dealing with substantial data volumes, ultimately reducing training time. To achieve this, optimal features are extracted using three distinct techniques: statistical features, spatial features, and temporal features. • Statistical Features: This process involves extracting key features by analyzing data related to network traffic. Essential statistical metricsarecomputed,includingminimumandmaximumlevelsofnetworktraffic,mean,median,andmodevalu es across all data points, as well as the variance and standard deviation of network trafficdata. • Spatial Features: Spatial features provide insights into the geographical locations associated with network traffic data. In the context of the pre-processed data tra zA , spatial information related to network activity can be discerned. This entails converting spatial data into numerical values organized in a grid format. The process begins by establishing a buffer zone around grid center points, followed by clipping the information within these zones. • Temporal Features: Temporalfeaturescaptureinformationrelatedtothetimingandsequenceofnetworktraffic.Analyzingtem poralpatternsinvolvesextractingfeaturessuchastimestamps,frequencies,andintervalsbetweenevents. These three feature extraction techniques collectively contribute to a comprehensive understanding ofnetwork trafficqualityofservice,fosteringimprovedpredictionaccuracyandreducedoverfittingchallengesintheprop osed model. 4.3. Feature Selection A weighted feature selection procedure is used, which is especially designed for the context of network traffic quality of service, to improve forecast accuracy. The features extracted, denoted as FEfz, are inputted into the Walrus Optimization algorithm to derive the optimal solution for service-related features. Feature selection holds the key advantage of providing highly relevant results aligned with the model's requirements while concurrently streamlining the complexity of training and testing in prediction techniques. Despite these advantages, relying solely on feature selection may not consistently yield desired outcomes for the model, potentially introducing overfitting issues and compromising accuracy, even with the removal of redundancy. The weighted feature selection process involves assigning weights to each corresponding feature. By assigning weights, the relative significance of each feature becomes apparent, allowing the proposed model to discern the importance of individual features. This weighting mechanism is especially pertinent in the context of quality of service in network traffic, enabling the model to effectively analyze and prioritize features associated with network performance. Through this tailored weighted feature selection process, the proposed model aims to enhance predictive accuracy while maintaining a keen focus on features critical to assessing and predicting the quality of service in network traffic. 7886 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 8, No. 6: 7878-7901, 2024 DOI: 10.55214/25768484.v8i6.3717 © 2024 by the authors; licensee Learning Gate Figure 2. Weighted feature selection using the proposed Walrus optimization. 4.4. Ensemble-Based Prediction of Service Quality 4.4.1. Integration of Models In essence, the Ensemble model is a fusion of various deep learning techniques aimed at enhancing performance. This paper incorporates learning techniques such as GRU, LSTM, and Random Forest to construct the Ensemble model. These techniques function as neural networks, with neurons capable of classifying features that yield results in terms of Quality of Service (QoS). The service quality prediction model that is built is improved even more when the Ensemble techniques are combined with different architectures to evaluate the Ensemble model's results. By means of this integration, data imbalance is lessened and data distribution is adjusted. The proposed method's sensitivity may be increased in large part by altering the learning process. This adjustment is used to improve the method's overall performance, and the deep learning algorithms that are included provide a series of models for the training and testing phases. The collective effect is a significant enhancement in prediction performance, highlighting the efficacy of the integrated Ensemble model in predicting service quality. Table 1 Infinity-norm. Learning percentage % Infinity-norm PSO 40 394.009966 50 345.2708443 60 495.5452311 70 330.6683094 80 398.6523619 GWO 40 409.8211039 50 425.7688473 60 385.9458468 70 434.518189 80 407.6381574 JAYA 40 495.7070731 50 394.8586763 60 367.9764502 70 399.3035701 80 418.7756282 RSA 40 416.6497379 50 416.0627179 60 350.3669463 7887 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 8, No. 6: 7878-7901, 2024 DOI: 10.55214/25768484.v8i6.3717 © 2024 by the authors; licensee Learning Gate 70 384.2094745 80 377.4521606 WOA 40 431.0240482 50 376.1744319 60 410.0180573 70 391.0386571 80 365.9498596 5. Results Based on the findings from the INFINITY-NORM table, which compares the performance of several optimization algorithms at various learning percentages (40%, 50%, 60%, 70%, and 80%), we direct our attention to the Walrus Optimisation Algorithm (WOA). When WOA is compared to other well-known optimization approaches as Particle Swarm Optimisation (PSO), Grey Wolf Optimizer (GWO), JAYA, and Rogue System Algorithm (RSA), a more complex picture emerges. WOA performs inconsistently, as seen by oscillations in INFINITY-NORM values, demonstrating its flexibility under diverse learning situations. In contrast, PSO exhibits variable performance with both high and low INFINITY-NORM values, GWO maintains relative consistency, while JAYA and RSA show a declining tendency in INFINITY-NORM values as learning percentages increase. This comparative examination provides in- depth insight into how the Walrus Optimisation works. And also from this table we can observe that the WOA's exploration-exploitation balancing mechanism may explain its INFINITY-NORM oscillations, which show its flexibility to varied learning circumstances. Table 2. MAE. Learning percentage% MAE PSO 40 4.434136766 50 3.704464424 60 3.489332824 70 2.948374689 80 2.905006615 GWO 40 3.742504317 50 3.523839823 60 3.38367767 70 2.860239804 80 2.742125793 JAYA 40 3.883502828 50 3.664809494 60 3.546695483 70 3.027713389 80 2.949153367 RSA 40 3.918680356 50 3.607540565 60 3.440932179 70 3.014425382 80 2.900702253 WOA 40 3.821799757 50 3.453354987 60 2.841803867 70 2.591523578 80 2.645187514 7888 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 8, No. 6: 7878-7901, 2024 DOI: 10.55214/25768484.v8i6.3717 © 2024 by the authors; licensee Learning Gate The table presents the Mean Absolute Error (MAE) values for the PSO, GWO, JAYA, RSA, and WOA algorithms over different learning percentages, namely 40%, 50%, 60%, 70%, and 80%. The Mean Absolute Error (MAE) is a metric used to assess the accuracy of predictions, where lower values indicate higher performance. And that the Walrus Optimisation Algorithm (WOA) has superior performance compared to other algorithms, as shown by consistently reduced Mean Absolute Error (MAE) values across all learning percentages. As the proportion of learning increases, the mean absolute error (MAE) values of WOA exhibit a consistent reduction, indicating a continuous advancement. This finding provides evidence that the Walrus Optimisation Algorithm effectively mitigates prediction mistakes and enhances accuracy. The capacity of WOA to optimize outcomes makes it a potential choice for tasks that need accurate predictions and minimal error rates. Constant MAE decrease with increased learning percentages suggests it may enhance predictions by using useful optimization landscape characteristics. Table 3. MASE. Learning percentage% MASE PSO 40 4169.09 50 3437.15 60 3249.35 70 2701.06 80 2678.37 GWO 40 3491.13 50 3276.14 60 3143.31 70 2616.38 80 2491.9 JAYA 40 3622.49 50 3403.3 60 3295.84 70 2730.71 80 2712.21 RSA 40 3656.39 50 3335.53 60 3211.18 70 2798.51 80 2652.9 WOA 40 3567.43 50 3233.82 60 2613.61 70 2328.19 80 2407.14 The table displays the Mean Absolute Scaled Error (MASE) values for Particle Swarm Optimisation (PSO), Grey Wolf Optimizer (GWO), JAYA, Random Search Algorithm (RSA), and Walrus Optimisation Algorithm (WOA) at various learning percentages (40%, 50%, 60%, 70%, and 80%). The Mean Absolute Scaled Error (MASE), a significant measure for evaluating the accuracy of forecasts, constantly demonstrates the improved performance of the Walrus Optimisation Algorithm. This is shown by continuously lower values seen across all learning percentages. It is worth noting that the Weighted Objective Assessment (WOA) exhibits a consistent pattern of enhancement, as seen by the diminishing Mean Absolute Scaled Error (MASE) values in correlation with the rise in the percentage of learning. This highlights the effectiveness of the suggested Walrus Optimisation Algorithm in improving the 7889 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 8, No. 6: 7878-7901, 2024 DOI: 10.55214/25768484.v8i6.3717 © 2024 by the authors; licensee Learning Gate accuracy of predicted values and reducing mistakes, making it an appealing option for activities that need precise and dependable forecasting. As the learning percentages increasing will suggest rising predicting accuracy, potentially due to its adaptive learning process that captures data patterns. Table 4. MEP. Learning percentage% MEP PSO 40 2.74902 50 2.29303 60 2.16549 70 1.83194 80 1.76026 GWO 40 2.31933 50 2.17605 60 2.06628 70 1.79124 80 1.72977 JAYA 40 2.32702 50 2.2727 60 2.19343 70 1.91328 80 1.84163 RSA 40 2.36005 50 2.20913 60 2.1324 70 1.8777 80 1.79334 WOA 40 2.29394 50 2.14557 60 1.72308 70 1.65143 80 1.64587 The table shows Mean Evaluation Performance (MEP) values for PSO, GWO, JAYA, RSA, and WOA at different learning percentages (40%, 50%, 60%, 70%, and 80%). The effectiveness of algorithms in evaluating solution performance is measured by MEP. The Walrus Optimization Algorithm (WOA) consistently beats the other algorithms with lower MEP values across all learning percentages. WOA improved at 40% and 50% learning percentages, lowering MEP values. This refinement highlights WOA's greater efficiency in quickly and precisely assessing solution performance compared to PSO, GWO, JAYA, and RSA. And Low MEP values throughout learning percentages indicate its efficient solution performance assessment, possibly due to adaptive evaluation criteria and robust optimization. Table 5. ONE-NORM. Learning percentage% One-norm PSO 40 137894.7368 7890 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 8, No. 6: 7878-7901, 2024 DOI: 10.55214/25768484.v8i6.3717 © 2024 by the authors; licensee Learning Gate 50 115657.8947 60 109605.2632 70 92763.15789 80 92631.57895 GWO 40 117105.2632 50 110131.5789 60 106447.3684 70 90394.73684 80 86578.94737 JAYA 40 120921.0526 50 114868.4211 60 111184.2105 70 95131.57895 80 92763.15789 RSA 40 121710.5263 50 113289.4737 60 107894.7368 70 94868.42105 80 90921.05263 WOA 40 119605.2632 50 108684.2105 60 90000 70 82236.84211 80 83947.36842 The table displays the performance metrics, specifically the ONE-NORM values, of several optimization algorithms, namely Particle Swarm Optimisation (PSO), Grey Wolf Optimizer (GWO), JAYA, Random Search Algorithm (RSA), and Walrus Optimisation Algorithm (WOA), across different learning percentages (40%, 50%, 60%, 70%, and 80%). The ONE-NORM values work as indications of the efficiency and efficacy of each algorithm in the minimization of a given objective function. After careful analysis, it becomes apparent that the WOA algorithm regularly demonstrates superior performance compared to the other algorithms. This is clear from its ability to consistently achieve lower ONE-NORM values across all learning percentages. The persistent superiority seen in the performance of the suggested Walrus Optimisation Algorithm in optimizing the provided objective function underscores its efficacy, giving it an appealing option for situations where the minimization of the ONE- NORM is of utmost importance. The constantly decreasing ONE-NORM values compared to other algorithms show its better capacity to eliminate prediction errors and sustain optimization performance, perhaps due to its adaptive search method. . Table 6. RMSE. Learning percentage% RMSE PSO 40 16.8558 50 15.3644 60 15.4203 7891 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 8, No. 6: 7878-7901, 2024 DOI: 10.55214/25768484.v8i6.3717 © 2024 by the authors; licensee Learning Gate 70 13.8899 80 14.2665 GWO 40 15.8262 50 15.2968 60 15.623 70 14.0756 80 13.254 JAYA 40 16.2566 50 15.4741 60 14.9732 70 14.2106 80 13.8783 RSA 40 16.6702 50 15.7778 60 15.0997 70 14.4721 80 14.005 WOA 40 16.3916 50 15.1702 60 14.2221 70 13.1642 80 13.5324 The table provided illustrates the Root Mean Square Error (RMSE) values associated with different optimization algorithms, namely Particle Swarm Optimisation (PSO), Grey Wolf Optimizer (GWO), JAYA, Random Search Algorithm (RSA), and Walrus Optimisation Algorithm (WOA). These values are presented for different learning percentages, specifically 40%, 50%, 60%, 70%, and 80%. Root Mean Square Error (RMSE) is a widely used statistic in the field of predictive modelling, serving as a means to assess the precision of predictions. It is worth noting that lower RMSE values are indicative of superior performance. Upon analysis of the outcomes, it becomes apparent that the Weighted Overlap Add (WOA) algorithm consistently demonstrates the most favorable Root Mean Square Error (RMSE) values across all learning percentages, in comparison to the other methods. The persistent superiority shown in the suggested Walrus OptimisationAlgorithm highlights its efficacy in minimizing prediction errors and its dependability and efficacy in optimizing difficult problems, making it ideal for situations requiring high prediction precision and dependability. Table 7. SMAPE. Learning percentage% SMAPE PSO 40 0.03148 50 0.02564 60 0.02469 70 0.02064 80 0.01997 7892 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 8, No. 6: 7878-7901, 2024 DOI: 10.55214/25768484.v8i6.3717 © 2024 by the authors; licensee Learning Gate GWO 40 0.02634 50 0.02454 60 0.02334 70 0.02012 80 0.01935 JAYA 40 0.02652 50 0.0258 60 0.02497 70 0.02162 80 0.02083 RSA 40 0.02686 50 0.02503 60 0.02426 70 0.02125 80 0.02037 WOA 40 0.026 50 0.02441 60 0.0194 70 0.01861 80 0.01849 The table below shows the Symmetric Mean Absolute Percentage Error (SMAPE) values for various optimization algorithms, including Particle Swarm Optimisation (PSO), Grey Wolf Optimizer (GWO), JAYA, Random Search Algorithm (RSA), and Walrus Optimisation Algorithm (WOA), at different learning percentages (40%, 50%, 60%, 70%, and 80%). The SMAPE metric measures prediction accuracy, with lower values indicating greater performance. Analyzing the findings, it is clear that WOA consistently has the lowest SMAPE values among the algorithms at each learning percentage, showing higher accuracy in forecasting outcomes. Perhaps due to its strong optimization process that captures the underlying data distribution and reduces forecast disparities. Table 8. Two-Norm. Learning percentage% TWO-norm PSO 40 2972.5322 50 2672.103 60 2689.2704 70 2387.1245 80 2455.794 GWO 40 2766.5236 50 2661.8026 60 2723.6052 70 2419.7425 80 2246.3519 JAYA 40 2852.3605 50 2697.8541 60 2596.5665 70 2445.4936 80 2381.9742 7893 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 8, No. 6: 7878-7901, 2024 DOI: 10.55214/25768484.v8i6.3717 © 2024 by the authors; licensee Learning Gate RSA 40 2934.7639 50 2757.9399 60 2617.1674 70 2498.7124 80 2402.5751 WOA 40 2879.8283 50 2639.485 60 2445.4936 70 2234.3348 80 2308.1545 When compared to various optimization algorithms in the table, our suggested Walrus Optimisation Algorithm (WOA) outperforms them. WOA consistently outperforms Particle Swarm Optimisation (PSO), Grey Wolf Optimizer (GWO), JAYA, and Random Search Algorithm (RSA) at each learning percentage (40%, 50%, 60%, 70%, and 80%). Lower two-norm values suggest that WOA produces better optimization outcomes, indicating increased efficiency and efficacy in tackling the optimization challenge at hand. This consistent performance over varied learning percentages highlights our proposed algorithm's superiority, making it an appealing option for optimization tasks when compared to current alternatives. Its precision and efficacy in optimizing complicated objective functions make it a popular option for a broad variety of optimization problems. Figure 3. Infinity norm. The above graph shows that the Walrus Optimisation Algorithm (WOA) demonstrates various patterns in its performance across varied support vector regression (SVR) percentages. The INFINITY- NORM results for WOA at 35%, 55%, 65%, 75%, and 85% SVR offer a thorough perspective of how the algorithm reacts to various degrees of regression assistance. Notably, WOA exhibits a gradual increase in INFINITY-NORM values as the SVR % increases, showing a possible association between the algorithm's performance and the degree of support vector regression. This pattern shows that WOA may adapt and optimize its solutions more dynamically in settings with stronger regression support. 7894 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 8, No. 6: 7878-7901, 2024 DOI: 10.55214/25768484.v8i6.3717 © 2024 by the authors; licensee Learning Gate Figure 4. MAE. It is clear from the Mean Absolute Error (MAE) graph that the Walrus Optimisation approach (WOA), which is the suggested approach, performs competitively at different support vector regression (SVR) percentages. As can be seen from the MAE figures at 35%, 55%, 65%, 75%, and 85% SVR, WOA consistently minimizes absolute errors. Specifically, notable is the low MAE of 2.301794167 that WOA obtains at 85% SVR, indicating that the algorithm is particularly good at capturing and minimizing differences between predicted and actual values with increased regression support. This finding highlights WOA's competitive performance against other algorithms, which indicates its potential effectiveness in circumstances requiring accuracy and precision. Evidence reveals that WOA is a dependable option for accurate forecasts as it successfully lowers prediction mistakes. Figure 5. MASE. The graph's Mean Absolute Scaled Error (MASE) numbers provide important information on how different methods for percentages (35%, 55%, 65%, 75%, and 85%) perform. WOA, the Walrus 7895 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 8, No. 6: 7878-7901, 2024 DOI: 10.55214/25768484.v8i6.3717 © 2024 by the authors; licensee Learning Gate Optimisation method, is a suggested method that exhibits competitive performance in minimising MASE values, hence demonstrating its efficacy in predicting accuracy. WOA obtains a very low MASE of 3.111924686 at 85% SVR, indicating its capacity to provide predictions that are precise and dependable with lower absolute errors. In contrast, various algorithms that show differing performance patterns at different SVR percentages include RNN, SVM-LSTM, and ENSEMBLE-RF. The MASE graph's patterns demonstrate WOA's promise as a reliable forecasting system, especially in situations when exact forecasts are needed. Figure 6. MEP. The graph displays Mean Percentage Error (MEP) figures that provide a thorough overview of the performance of several algorithms at different percentages (35%, 55%, 65%, 75%, and 85%). In these SVR settings, MEP values for every algorithm—SVR, SVM-LSTM, ENSEMBLE-RF, RNN, ENSEMBLE, and WOA—show distinct patterns. With consistently low MEP values across all SVR percentages, the Walrus Optimisation Algorithm (WOA) is particularly noteworthy and shows promise in reducing percentage mistakes in forecasts. With an MEP of 1.476953076, WOA obtains a very strong performance at 85% SVR, indicating its accuracy and resilience in predicting with increased regression assistance. This demonstrates WOA's consistency in generating predictions that are precise and have few percentage mistakes. 7896 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 8, No. 6: 7878-7901, 2024 DOI: 10.55214/25768484.v8i6.3717 © 2024 by the authors; licensee Learning Gate Figure 7. One Norm. The graph's One-Norm values provide insight into the performance of several algorithms at various percentages (35%, 55%, 65%, 75%, and 85%). The One-Norm, which represents the sum of absolute values, is used to calculate the overall magnitude of mistakes. Notably, the Walrus Optimisation Algorithm (WOA) outperforms all other algorithms in terms of SVR %, with continually lower One- Norm values. WOA obtains a significantly low One-Norm of 199.8951208 at 85% SVR, suggesting its efficacy in minimizing the total amount of prediction errors. A comparison with different algorithms indicates WOA's resilience and effectiveness in capturing the variability of the dataset. This shows that WOA optimizes outcomes and reduces error size, making it suited for precision optimization tasks. Figure 8. RMSE. The graph's Root Mean Square Error (RMSE) values give a complete evaluation of the performance of several algorithms at various percentages (35%, 55%, 65%, 75%, and 85%). RMSE is an important 7897 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 8, No. 6: 7878-7901, 2024 DOI: 10.55214/25768484.v8i6.3717 © 2024 by the authors; licensee Learning Gate statistic for assessing prediction accuracy since it quantifies the square root of the average squared discrepancies between expected and actual values. The Walrus Optimisation Algorithm (WOA) stands out in this setting for its competitive performance, consistently obtaining reasonably low RMSE values across all SVR percentages. WOA's performance at 85% SVR is particularly remarkable, with a commendably low RMSE of 5.194432863, indicating its usefulness in minimizing the total amount of prediction errors. A comparison with different algorithms indicates WOA's resilience and effectiveness in capturing the variability of the dataset. The observed patterns in the RMSE graph highlight WOA's potential appropriateness for situations requiring accurate forecasts with few mistakes. And also WOA reduces prediction mistakes by decreasing the square root of the average squared differences between predicted and actual values. Figure 9. SMAPE. The graph shows the Symmetric Mean Absolute Percentage Error (SMAPE) numbers, which show how accurate different methods are at various percentages: 35%, 55%, 65%, 75%, and 85%. A lot of people use SMAPE to figure out how accurate predictions are as a percentage. In this situation, the Walrus Optimisation Algorithm (WOA) consistently does a great job, as shown by its consistently low SMAPE numbers at all SVR percentages. At 85% SVR, WOA gets an excellent SMAPE of 0.014932407, which shows that it is good at making correct guesses with low percentage mistakes. When compared to other algorithms, WOA is shown to be reliable and good at catching the variability of the information. This makes it a good choice for situations where accurate predicting is needed. This shows WOA's accuracy in predicting tasks, making it ideal for accurate forecasts. 7898 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 8, No. 6: 7878-7901, 2024 DOI: 10.55214/25768484.v8i6.3717 © 2024 by the authors; licensee Learning Gate Figure 10. Two norm. From the above graph it is observable that Two-Norm values on the graph illustrate algorithm performance at 35%, 55%, 65%, 75%, as well as 85%. The Two-Norm employs the Euclidean norm to calculate error magnitude. The Walrus Optimization Algorithm (WOA) produces low Two-Norm values consistently across all SVR percentages, exhibiting competitive performance. The WOA Two- Norm of 65.98940214 at 85% SVR demonstrates its capacity to decrease prediction errors. When compared to other algorithms, WOA demonstrates its robustness and efficacy in capturing dataset heterogeneity. WOA's Two-Norm graph trends indicate that it may be appropriate for instances when error reduction is crucial. 5.1. Comparision the Proposed Model with Existing Models Esemble CNN SVM MAE 2.3 3.75 4.1 RMSE 5.19 6.8 7.5 MSE 3.11 4.2 4.75 Figure 11. 7899 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 8, No. 6: 7878-7901, 2024 DOI: 10.55214/25768484.v8i6.3717 © 2024 by the authors; licensee Learning Gate We compared our Ensemble model enhanced with the modified Walrus Optimization Algorithm (WOA) for Quality of Service (QoS) prediction to conventional models like Convolutional Neural Networks (CNN) and Support Vector Machines (SVM) using MAE, RMSE, and MSE. All measures showed that our model predicted service quality well. The Ensemble model with WOA has a lower MAE of 2.3 than CNN and SVM, suggesting more accuracy and less divergence from real values. Compared to CNN and SVM, our model's RMSE value of 5.19 was much lower than 6.8 and 7.5, indicating improved prediction errors and precision. Our model has a lower MSE (3.11) than CNN (4.2) & SVM (4.75), suggesting better squared error reduction. Our approach provides more accurate and dependable service quality forecasts than previous techniques, which might improve network management as well as user experience in dynamic mobile network settings. 6. Conclusion The Ensemble-Based Service Quality Prediction (EAQP) model, which is reinforced with the cutting-edge Walrus Optimization Algorithm (WOA), exhibits exceptional performance when it comes to predicting service quality in settings composed of dynamic mobile networks. WOA's flexibility enhances accuracy across a variety of assessment criteria, which allows the model to make strong predictions. This is accomplished by the rigorous pretreatment of data and the integration of a wide variety of machine learning techniques. A number of metrics, including an MAE of 2.301794167, MASE of 3.111924686, MEP of 1.476953076, One-Norm of 199.8951208, RMSE of 5.194432863, SMAPE of 0.014932407, and Two-Norm of 65.98940214, are among the metrics that WOA obtains low values for when it has an SVR of 85%. These discoveries not only contribute to the advancement of research approaches in the field of machine learning, but they also have major practical consequences for enterprises that are dependent on mobile network services. Increasing customer happiness and improving service delivery tactics are both possible outcomes of enterprises' ability to properly forecast service quality and proactively handle network problems. Due to the fact that the model can be used in a variety of network scenarios, it has the potential to be a very useful instrument for industry practitioners who are looking to enhance both the quality of service and the user experience. At last the Comparision of proposed model with existing models also dive into the novelty of our proposed and by this we can conclude that the proposed model is best suitable for real-time Quality of Service predictions 6.1. Real World Applications and Implications The Ensemble-Based Quality of Service Prediction model with Walrus Optimization has great promise for mobile networks and service quality prediction. Telecommunications firms looking to improve service quality evaluation might use this novel technique. This approach may help network operators discover and resolve problems like latency in the network, bandwidth availability, as well as connection stability before they affect user experience by correctly forecasting service quality indicators. The model's Walrus Optimization Algorithm for feature selection provides scalability and flexibility for changing network situations. Beyond telecommunications, the model's ensemble learning architecture and optimization approaches are applicable to healthcare, banking, and environmental monitoring. The approach might change service quality prediction, providing industry experts and researchers with realistic answers. 6.2. Limitations Ensemble models like GRU, LSTM, Random Forest along with WOA Optimization may improve prediction accuracy, but they have limitations. Integrating learning algorithms may slow real-time applications or resource-constrained devices. Selecting hyper parameters and optimization techniques for each component model is challenging due to the highly dimensional parameter space & probable model interactions, yet the ensemble model's effectiveness relies on it. Stakeholders struggle to understand and trust ensemble models like models based on deep learning, which are black boxes. Scalability of the ensemble approach may be limited for large datasets or high-dimensional features spaces, requiring careful processing resources as well as algorithmic efficiency. 7900 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 8, No. 6: 7878-7901, 2024 DOI: 10.55214/25768484.v8i6.3717 © 2024 by the authors; licensee Learning Gate 6.3. Potential Challenges for Implementation Ensemble models using voting classifiers and Walrus Optimization Algorithm (WOA) feature selection may have difficulties. Coordinate and test data preprocessing, feature selection, ensemble construction, and optimization for model integration and interoperability. To identify the optimum techniques and hyper parameters, performance may need extensive tweaking and experimentation, adding complexity as well as duration to development. The proposed method may need scalable algorithms and networked computer infrastructures to evaluate enormous volumes of data. Reliable validation and evaluation of the ensemble model's generalization accuracy and dependability across datasets and application contexts need rigorous experimental design & statistical analysis. Funding: This research Supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (NRF- RS-2023 00237287, NRF- 2021S1A5A8062526) and local government-university cooperation-based regional innovation projects (2021RIS-003). Acknowledgement: We are thankful to the editors and the anonymous reviewers for many valuable suggestions to improve this paper. Copyright: © 2024 by the authors. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/). References [1] H. Bi, L. Lu, and Y. Meng, “Hierarchical attention network for multivariate time series long-term forecasting,” Appl. Intell., vol. 53, no. 5, pp. 5060–5071, 2023, doi: 10.1007/s10489-022-03825-5. [2] A. N. M. F. Faisal, A. Rahman, M. T. M. Habib, A. H. Siddique, M. Hasan, and M. M. Khan, “Neural networks based multivariate time series forecasting of solar radiation using meteorological data of different cities of Bangladesh,” Results Eng., vol. 13, p. 100365, 2022, doi: https://doi.org/10.1016/j.rineng.2022.100365. [3] M. Di Mauro, G. Galatro, F. Postiglione, W. Song, and A. Liotta, “Multivariate Time Series Characterization and Forecasting of VoIP Traffic in Real Mobile Networks,” IEEE Trans. Netw. Serv. Manag., vol. 21, no. 1, pp. 851–865, 2024, doi: 10.1109/TNSM.2023.3295748. [4] X. B. Jin, W. T. Gong, J. L. Kong, Y. T. Bai, and T. L. Su, “A Variational Bayesian Deep Network with Data Self‐ Screening Layer for Massive Time‐Series Data Forecasting,” Entropy, vol. 24, no. 3, pp. 1–17, 2022, doi: 10.3390/e24030335. [5] T. Theodoropoulos et al., “Graph neural networks for representing multivariate resource usage: A multiplayer mobile gaming case-study,” Int. J. Inf. Manag. Data Insights, vol. 3, no. 1, p. 100158, 2023, doi: https://doi.org/10.1016/j.jjimei.2023.100158. [6] H. E. Dinaki, S. Shirmohammadi, E. Janulewicz, and D. Côté, “Forecasting Video QoE With Deep Learning From Multivariate Time-Series,” IEEE Open J. Signal Process., vol. 2, pp. 512–521, 2021, doi: 10.1109/OJSP.2021.3099065. [7] A. R. S. Parmezan, V. M. A. Souza, and G. E. A. P. A. Batista, “Evaluation of statistical and machine learning models for time series prediction: Identifying the state-of-the-art and the best conditions for the use of each model,” Inf. Sci. (Ny)., vol. 484, pp. 302–337, 2019, doi: 10.1016/j.ins.2019.01.076. [8] Z. Che, S. Purushotham, K. Cho, D. Sontag, and Y. Liu, “Recurrent Neural Networks for Multivariate Time Series with Missing Values,” Sci. Rep., vol. 8, no. 1, pp. 1–12, 2018, doi: 10.1038/s41598-018-24271-9. [9] S. Wang, M. Zhang, H. Miao, Z. Peng, and P. S. Yu, “Multivariate Correlation-aware Spatio-temporal Graph Convolutional Networks for Multi-scale Traffic Prediction,” ACM Trans. Intell. Syst. Technol., vol. 13, no. 3, pp. 1–22, 2022, doi: 10.1145/3469087. [10] G. Deepak, M. Madiajagan, S. Kulkarni, A. N. Ahmed, A. Gopatoti, and V. Ammisetty, “MCSC-Net: COVID-19 detection using deep-Q-neural network classification with RFNN-based hybrid whale optimization,” J. Xray. Sci. Technol., vol. 31, no. 3, pp. 483–509, 2023, doi: 10.3233/XST-221360. [11] M. Xu, C. Song, H. Wu, S. S. Gill, K. Ye, and C. Xu, “esDNN: Deep Neural Network Based Multivariate Workload Prediction in Cloud Computing Environments,” ACM Trans. Internet Technol., vol. 22, no. 3, 2022, doi: 10.1145/3524114. [12] J. Gao, Y. Wang, and T. C. E. Cheng, “Multi-variable time series forecasting model based on high-order hesitant probabilistic linguistic fuzzy logical relationship,” J. Control Decis., pp. 1–18, 2023, doi: 10.1080/23307706.2023.2299877. https://creativecommons.org/licenses/by/4.0/ 7901 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 8, No. 6: 7878-7901, 2024 DOI: 10.55214/25768484.v8i6.3717 © 2024 by the authors; licensee Learning Gate [13] Y. S. Patel and J. Bedi, “MAG-D: A multivariate attention network based approach for cloud workload forecasting.,” Future Gener. Comput. Syst., vol. 142, pp. 376–392, May 2023, doi: 10.1016/j.future.2023.01.002. [14] N. Wu, B. Green, X. Ben, and S. O’Banion, “Deep Transformer Models for Time Series Forecasting: The Influenza Prevalence Case,” 2020, [Online]. Available: http://arxiv.org/abs/2001.08317 [15] L. Huang, F. Mao, K. Zhang, and Z. Li, “Spatial-Temporal Convolutional Transformer Network for Multivariate Time Series Forecasting,” Sensors, vol. 22, no. 3, 2022, doi: 10.3390/s22030841. [16] S. T. Nabi et al., “Deep Learning Based Fusion Model for Multivariate LTE Traffic Forecasting and Optimized Radio Parameter Estimation,” IEEE Access, vol. 11, no. February, pp. 14533–14549, 2023, doi: 10.1109/ACCESS.2023.3242861. [17] Y. Luo, Y. Zhang, X. Cai, and X. Yuan, “E2GaN: End-to-end generative adversarial network for multivariate time series imputation,” in IJCAI International Joint Conference on Artificial Intelligence, 2019, vol. 2019-Augus, pp. 3094–3100. doi: 10.24963/ijcai.2019/429. [18] R. Wan, S. Mei, J. Wang, M. Liu, and F. Yang, “Multivariate temporal convolutional network: A deep neural networks approach for multivariate time series forecasting,” Electron., vol. 8, no. 8, 2019, doi: 10.3390/electronics8080876. [19] M. A. Castán-Lascorz, P. Jiménez-Herrera, A. Troncoso, and G. Asencio-Cortés, “A new hybrid method for predicting univariate and multivariate time series based on pattern forecasting,” Inf. Sci. (Ny)., vol. 586, pp. 611–627, 2022, doi: https://doi.org/10.1016/j.ins.2021.12.001. [20] Y. Yang, H. Lv, and N. Chen, “A Survey on ensemble learning under the era of deep learning,” Artif. Intell. Rev., vol. 56, no. 6, pp. 5545–5589, 2023, doi: 10.1007/s10462-022-10283-5. [21] S. Sholla, S. Kaur, G. R. Begh, R. N. Mir, and M. A. Chishti, “Clustering Internet of Things: A Review,” J. Sci. Technol. Issue Inf. Commun. Technol., vol. 3, no. 2, p. 21, 2017, doi: 10.31130/jst.2017.61. [22] S. Sreenivasamurthy and K. Obraczka, “Clustering for load balancing and energy efficiency in IoT applications,” Proc. - 26th IEEE Int. Symp. Model. Anal. Simul. Comput. Telecommun. Syst. MASCOTS 2018, no. March, pp. 319–332, 2018, doi: 10.1109/MASCOTS.2018.00038.