Prerpints.org logo
Preprint
Article

A TSENet Model for Predicting Cellular Network Traffic

Altmetrics

Downloads

113

Views

38

Comments

0

A peer-reviewed article of this preprint also exists.

This version is not peer-reviewed

Application of Deep Learning Method in 6G Communication Technology

Submitted:

06 February 2024

Posted:

09 February 2024

You are already at the latest version

Alerts
Abstract
Wireless Sensor Networks (WSNs) are gaining traction in the realm of network communication, renowned for their adaptability, configuration, and flexibility. The forthcoming network traffic within WSNs can be forecasted through temporal sequence models. In this correspondence, We present a method (TSENet) that can accurately predict the traffic in the cellular network. TSENet is composed of transformers and self-attention network. We have designed a temporal transformer module specifically for extracting temporal features. This module accomplishes this by modeling the traffic flow within each grid of the communication network at both near-term and periodical intervals. Simultaneously, we amalgamate the spacial features of each grid with information from its correlated grids, generating spacial predictions within the spacial transformer. Furthermore, we employ self-attention aggregation to capture dependencies between external factor features and celluar data features. Empirical assessments performed on a genuine cellular traffic dataset offer compelling evidence substantiating the efficacy of TSENet.
Keywords: 
Subject: Engineering  -   Telecommunications

1. Introduction

In recent years, the swift progress in distributed computing and micro-electromechanical systems (MEMS) has propelled wireless sensor networks into numerous domains, encompassing military [1], intelligent architecture [2] and industrial sectors [3,4]. To observe diverse phenomena like network traffic, sensors are equipped with sensing, computing, communication, and additional functionalities.
With the continuous development of wireless communication technologies, higher requirements are being imposed on wireless communication networks. Among these requirements, accurately predicting the future change trend of key performance indicators of wireless cells holds significant significance. It profoundly impacts the network planning, network optimization, and network energy-saving practices of mobile communication operators.
Predicting the change trend of the key performance index of the wireless cell is very important for the mobile communication operator, which can help the operator to plan the network. By accurately predicting future KPI changes, operators can properly plan network resource allocation, ensure stable network connection during peak hours, and avoid congestion and service quality deterioration. By learning the change trends of performance counters, operators can adjust network parameters and configurations in a timely manner to optimize network performance and provide better user experience. For example, according to the prediction result, the operator may decide whether to increase the number of base stations, adjust the power allocation, or change the channel allocation policy, thereby increasing the network capacity and coverage area.
In recent studies, Yu et al. [5] tackled the challenge of cellular data prediction with STEP, a spatiotemporal fine-granular mechanism utilizing GCGRN. The deep graph convolution network outperforms existing methods, offering efficiency and reduced resource consumption in cellular networks. Santos et al. [6] evaluated DL and ML models for short-term mobile Internet traffic prediction. LSTM exhibited excellent performance across diverse clusters in Milan. Hardegen et al. [7] present a pipeline using Deep Neural Networks (DNNs) to predict information of real-world celluar data flows. Unlike traditional binary classifications, their system quantizes predicted flow features into three classes. They emphasize the potential for flow-based routing scenarios. Furthermore, Dodan et al. [8] propose an efficient Internet traffic prediction system using optimized recurrent neural networks (RNNs). They address the challenges in large-scale networks, emphasizing the need for scalable solutions to anticipate traffic fluctuations, avoid congestion, and optimize resource allocation. Experimental outcomes showcase the enhanced efficacy of their RNN-based methodology compared to conventional approaches such as ARIMA and AdaBoost regressor.
In the realm of deep learning methodologies, both Convolutional Neural Networks (CNNs) [9] and Recurrent Neural Networks (RNNs) [10] stand as foundational structures for modeling temporal information. Cellular communication data exhibits temporal periodicity in its spatial distribution, bearing distinct characteristics [11,12]. Consequently, contemporary deep learning models aim to integrate temporal and spacial information in celluar data. Zhang et al. [13] come up with STDenseNet, a convolutional neural network approach for cellular data prediction. In RNN models, elements in the input sequence are processed sequentially, leading to inefficiencies in handling lengthy sequences. Conversely, methodologies primarily reliant on CNNs may fall short in fully capturing spatial dependencies between adjacent features. Additionally, approaches leveraging RNNs for extracting prolonged temporal relationships exhibit inherent weaknesses. These constraints may result in an insufficient description of continuous and temporal features.
In response to these challenges, we present a model named TSENet, founded on the principles of Transformer and Self-attention. Initially introduced in 2017 within the realm of natural language processing, Transformer boasts a parallel encoder-decoder architecture, later extended to various domains. Primarily, we harness its robust sequential modeling capabilities tailored. Concretely, we devise a Time Sequence Module (TSM) and a Spatial Sequence Module (SSM), adept at simultaneously and accurately extracting temporal and spatial features within short time intervals. Subsequently, we employ self-attention mechanisms to amalgamate the model’s prowess in capturing dependencies between external factors and cellular traffic characteristics. Our comprehensive experiments conducted on the bespoke dataset aim to substantiate the efficacy of our approach. Subsequent analyses and discussions will delve into a nuanced exposition of the TSENet model.
In summary, the primary contributions of this study can be delineated as follows: 1) Pioneering the introduction of a spatiotemporal traffic forecasting model, christened TSENet, which harmoniously integrates external elements. This model, predicated on harnessing spatiotemporal features, takes into account the nuanced characteristics of external factors. 2) By employing attention mechanisms, TSENet adeptly captures the multifaceted global spatiotemporal correlations within cellular traffic networks. Notably, TSENet extracts both enduring and fleeting temporal features, culminating in the establishment of a spatial nexus model among nodes. 3) Through the incorporation of self-attention mechanisms, TSENet effectively excavates the intricate interdependencies between external factors and cellular traffic, thereby further elevating the predictive precision of TSENet.

2. Related Work

2.1. Wireless Sensor Networks

Wireless Sensor Networks epitomize the avant-garde paradigm in ubiquitous computing, embodying the fusion of sophisticated sensor technologies and wireless communication. The structural backbone of these networks comprises diminutive yet potent sensor nodes. Communication among these sensors unfolds through wireless protocols, fostering seamless interconnectivity between nodes [14]. This mirrors the traditional Single-Receiver WSN (refer to Figure 1, left segment). However, this singular receiver scenario lacks scalability. Due to the substantial number of nodes, the collector accumulates more data, impeding the network’s scalability when it reaches capacity. This is a frequent scenario, especially with multiple receivers in a network (depicted on the right side of Figure 1) [15]. This decentralized approach exhibits advantages in terms of scalability, fault tolerance, and load distribution. In practical applications, wireless sensors serve diverse purposes in the realm of cellular traffic monitoring. Krishnasamy et al. [16] provided an overview of wireless sensor networks, comprising multiple sensors with sensing and communication functionalities. These sensors both receive and transmit data to designated nodes. The core purpose of a sensor node is to monitor environmental conditions, process the acquired data, and then transmit it to an analysis center. Sensor deployment locations are typically non-uniformly designed, often installed randomly in irregularly shaped specific locations based on their transmission range. Hence, it is important that the deployed algorithms can be adapted to each different geographical area.. In essence, deploying wireless sensors facilitates the analysis of traffic patterns, detection of anomalous behavior, optimization of cellular networks, and predictive modeling.

2.2. Network Traffic Prediction

Network traffic prediction has been a subject of prolonged inquiry. ARIMA [17] and Holt-Winters (HW) [18] are both univariate time series models. ARIMA leverages autoregressive, difference, and moving average components, with component order determined by Akaike or Bayesian information criteria. While statistical models exhibit commendable performance on resource-constrained mobile devices due to their modest computational requirements, their linear nature often results in inferior performance compared to machine learning models adept at capturing nonlinear relationships [19]. In contrast to statistical models, machine learning models often show superior capabilities [20]. Xia et al. [21] applied RF and LightGBM to the network traffic prediction model. Zhang et al. [22] employs a model based on Gaussian and a weighted EM method to optimize UAV deployment for reduced power consumption in downlink transmission and mobility. In recent years, neural networks have emerged as potent tools, attaining significant success across various prediction domains, notably in traffic prediction challenges. Basic yet pivotal, the Feed-forward Neural Network serves as the cornerstone for cellular traffic prediction, incorporating hidden layers and diverse activation functions like tanh, sigmoid, and ReLU to enhance learning capabilities. However, for spacial-temporal prediction challenges, the flattening calculation is required, potentially leading to the lack of spatial information and a decline in prediction performance.

2.3. Transformer

In recent years, endeavors rooted in the transformative power of transformers have burgeoned across domains such as computer vision, signal processing [23], and bioinformatics [24], yielding superlative outcomes in these realms. An array of studies has been conducted to unravel the potential of transformer networks in forecasting network traffic, demonstrating their superiority over conventional machine learning methodologies. Wen et al. [25] proffered a Transformer-based model meticulously designed for the prediction of short-term traffic congestion. This model, when juxtaposed against traditional machine learning methods like ARIMA and LSTM [26], exhibited a pronounced prowess. Qi et al. [27] expounded upon the Performer model, an innovative transformer-based predictive methodology for clustering voluminous time-series data, showcasing conspicuous efficacy in comparison to conventional methods such as random forest and SVM. Chuanting Zhang et al. [28] presented a groundbreaking framework for wireless traffic prediction, termed as Federated Dual Attention (FedDA). FedDA facilitates collaborative training among multiple edge clients to cultivate a high-fidelity predictive model, eclipsing state-of-the-art methods with substantial enhancements in average mean squared error performance. Rao et al. [29], leveraging deep learning techniques, undertook an exhaustive exploration of challenges and recent strides in cell load balancing. Feng et al. [30] propounded an end-to-end framework christened as DeepTP, meticulously designed for predicting traffic demands from mobile network data through the prism of deep learning techniques. DeepTP not only demonstrated a superiority over other cutting-edge traffic prediction models by exceeding them but also adeptly captured the spatiotemporal characteristics and external factors’ influence for a more precise predictive outcome.
Despite achieving commendable milestones in internet traffic prediction, further exploration is requisite regarding the application of cellular traffic data models based on wireless sensor networks. Diverging from prior endeavors, the utilization of self-attention mechanisms has been employed to congregate the extraction of external factors and dependencies in network traffic features, thereby judiciously harnessing external influences. Moreover, we employ the transformer to capture temporal and spatial information within concise intervals. Through the efficacy of multi-head self-attention mechanisms, we adeptly discern spatial relationships effectively interlaced between grids. This novel structure holds promise in mitigating, to some extent, the limitations inherent in existing methodologies.

3. Method

In this paper, we use self-attention mechanism to aggregate external factors and dependencies in network traffic features, and use transformer to capture temporal and spatial information in concise intervals. Through the effectiveness of multi-head self-attention mechanism, we can skillfully identify effective interlaced spatial relationships between grids. This section describes the TSENet framework, including spatial sequence module, time sequence module, external factor module and fusion output module.

3.1. The Proposed TSENet Framework

This article introduces TSENet, which comprises four primary modules: the Time Sequence Module, Spatial Sequence Module, External Factor Module, and Fusion Output Module, as illustrated in Figure 2. The Time Sequence Module and Spatial Sequence Module are designed based on the fundamental structures of the transformer. TSM extracts recent temporal trends in cellular traffic M c ( t ) and periodic features in M p ( t ) .The calculation block fuses them into temporal features. Concerning spatial feature modeling, the multi-head attention mechanism in a series of networks can effectively capture spacial relationships between network features. Real-time spatial features are captured in SSM. EFM is designed using a self-attention structure to extract external factor features. These three types of features complement each other, resulting in the generation of the final prediction through FOM. The transformer enables element of the input sequence to contribute to the result, incorporating target sequence prediction values in the decoding layer, thus achieving more effective predictions using additional information.

3.2. Spatial Sequence Module

In SSM, the cellular traffic data of the pertinent network serves as elements within the input sequence. We use a multi-head attention structure in order to engender spatial predictions. It processes information of each network and its related networks. For each network, analyzing its relationship with all other networks yields a considerably extensive grid sequence. Indeed, only a subset of networks significantly influences the traffic states of others. Hence, we select the L networks most correlated with the current network (inclusive of the network itself). The chosen traffic data from M c ( t ) is linked to a new sequence, M s ( t ) ∈ R H × L × J , serving as input to the spatial sequence block, wherein L is the length of the spatial array, and J denotes the dimension of each array.
(1) Grids Arrangement:We sort each Grid according to its correlation with other Grids and select the top L largest Grids. The value of L can be chosen based on experimental results. We employ the Pearson correlation coefficient to assess the correlation. The correlation between two grids A and B with equal traffic sequence lengths is calculated by the covariance of the variables and the product of their standard deviations: R A , B = c o v ( A , B ) σ A σ B .
We concatenate the incoming data A i n ( t ) ∈ R H × J and outgoing data A o u t ( t ) ∈ R H × J to form a matrix of dimensions H × 2 J . Subsequently, the correlation matrix P ( t ) ∈ R H × H is computed according to Equation. P i , j signifies the relevance between the i t h and j t h grids at the t t h interval.
(2) Spatial Sequence: As illustrated in Figure 1, the source sequence and the initial sequence serve as the principal inputs to the essential transformer. In Natural Language Processing (NLP), the elements of the sequence involve more informative word embeddings, but the array of cellular traffic data results in the scarcity of information. We choose the two complementary input arrays to ensure the predictive performance of the Transformer. The proximity matrix M c ( t ) signifies the cellular communication states close to the target interval. The M s ( t ) designed by M c ( t ) encompasses real-time spatial features for the sequence with a length of L, serving as the source array, and M c ( t ) ∈ R L × 1 × J is taken as the initial array with a length of 1, yielding the output of SSM Y s p a t i a l ( t ) ∈ R L × J . We omit positional encoding because there is no chronological order between spacial networks.

3.3. Time Sequence Module

As depicted in Figure 2, the TSM employs two temporal transformers to individually acquire features from near-term and periodic continuous time. The calculation block utilizes a ⊕ operation and a SSM to aggregate these two features into a predictive time sequence.
(1) Closeness Sequence: To harness additional information from the input data, M c ( t ) is extended as the source sequence of the proximity sequence. During the enhancement procedure, we use the correlation matrix P t to select the topmost U correlated networks for each grid. We combine the selected data and the grid data and create a scalar with dimension L × J × ( 1 + U ) , where U is a parameter determined through experimentation. This process enhances both the temporal features within the sequence elements and introduces a few space information in original sequence, thereby enhancing the accuracy of predictions.
In the Closeness Sequence, the averaging of the second dimension of M p ( t ) ∈ R L × G × J yields an initial sequence of dimensions L × J × 1 . We employ these averaged historical data as indicators for the traffic data levels in future time intervals. The Closeness Sequence produces the proximity time prediction, denoted as Y c ( t ) ∈ R L × J .
(2) Periodic Sequence: The Periodic Sequence is akin to the Proximity Sequence but lacks critical information. We employ M p ( t ) ∈ R L × G × J as the original sequence, with M c ( t ) ∈ R L × 1 × J serving as the initial sequence for the Temporal Sequence, thereby utilizing the compact time information included in M c ( t ) to complement the period sequence. The outcome of the periodic time prediction is expressed as Y p ( t ) ∈ R L × J .

3.4. External Factor Module

As illustrated in Figure 3, We utilize a self-attention mechanism to capture the feature representation of exterior factor ( f ∈ B S s , P O I s , S o c i a l s ). Mapping the input M f to distinct feature spaces, with query Q f = C q M f ∈ R z k × 0 , key K f = C k M f ∈ R z k × 0 , and value V f = C v M f ∈ R z v × 0 , where { C q , C k , C v } represents a set of 1 × 1 convolutional weights, and z k and z v denote channel numbers.
Utilizing matrix multiplication to compute the similarity scores for each pair of points yields: f = Q f T K f z k ∈ R O × O .
The similarity between point i and point j can be expressed as f i , j = M f , i T C q T C k M f , j Z k where M f , i and M f , j are feature scalars of shape z k × 1 . Subsequently, the index of similarity is column-normalized: α i , j = e x p ( f i , j ) ∑ k = 1 N e x p ( f j , k ) , i , j ∈ { 1 , 2 , . . . , N } .
The aggregated feature at position i is computed as the weighted sum of all positions: X i = ∑ j = 1 N e x p ( α ( C v M f , j ) ) , where C v M f , j ∈ R z v × 1 represents the j-th column of the value V f . Thus, in the fully connected layer, the input X ∈ R z v × O results in the output vector w f ∈ R p O × 1 . Finally, the vector is reshaped into an output W f ∈ R p × O . w f = σ ( C e X + a f ) , W f = R e s a h p e ( w f ) , Where C e and a f constitute the learnable parameters of the fully connected layer.

3.5. Fusion Output Module

From the above analysis, it can be discerned that the traffic data of the community is not only related to the continuous flow data but also associated with the time periods. Initially, we integrate the relevant features to capture this relationship. Subsequently, multiple DenseBlocks are employed to extract fused features, as illustrated in Figure 4. we get the predicted X through a sigmoid activation function.

4. Experiment

4.1. Dataset and Experimental Settings

Performing performance prediction by collecting engineering parameter data and performance indicator data of the target radio cell using the target performance indicator prediction algorithm can help obtain a performance prediction result of the target radio cell. Engineering parameters include the longitude and latitude of the site, site height, site type, azimuth, downtilt, power, bandwidth, and antenna model. The performance indicator time series data includes physical resource block utilization, downlink radio resource control user connection number, uplink radio resource control user connection number, and traffic volume. Employing Min-Max normalization, all data is scaled to the interval [0, 1], subsequently rescaled to its original values for a comparative assessment against the ground truth.
All training and testing processes are performed on an RTX 3080 with 10GB of memory. The training process uses the Adam optimizer. TSENet employed the widely used the polynomial learning rate decay strategy to minimize the MSE loss during training. The initial learning rate was 0.01. The learning rate decayed by factors of 10 at 50% and 100 at 75% of epochs. The primary TSENet model was trained with a batch size of 16, while smaller models (batch size: 32) underwent 500 epochs. Within the convolutional module, the last layer incorporated a 1 × 1 filter with a sigmoid activation function, aiming to minimize the Mean Squared Error loss during the training process. Additionally, the remaining layers were equipped with 3 × 3 filters comprising 16 kernels and ReLU activation functions.
Evaluation and comparison were conducted through the use of Normalized Root Mean Square Error (NRMSE), Mean Absolute Error (MAE), and Regression Score (RS). A smaller value for MAE and NRMSE, or a value closer to 1 for R2, signifies superior performance.

4.2. Inner Parameters of Transformers

We performed distinct ablative experiments, systematically varying the embedding dimension d and the number of heads h for each transformer. Table 1 presents the ultimate internal parameters of each transformer, while Section 4.4 provides comprehensive experimental outcomes for diverse input sequence combinations across transformers.

4.3. Analyze Useful Temporal Features

In TSM, we adeptly employ both hourly and daily data to extract temporal features, encompassing the recent and epochal aspects of time. Within our experimental domain, we further formulate a weekly dataset, denoted as M t ( t ) , derived from the original matrix. We analyze diverse time series to identify crucial temporal features for predicting cellular traffic on TSM. Results in Table 2 demonstrate that combining proximity and daily periodic time features yields the most effective outcomes. Longer cycle sequences tend to introduce disruptions to predictions, thereby reducing forecasting accuracy.

4.4. Analyze Useful Temporal Features

In our model, each transformer necessitates two sequences as inputs. To enhance precision, we conducted experiments with varied combinations of input sequences. In Figure 5, the optimal inputs for each transformer are depicted. Spatial transformers leverage traffic data from pertinent grids and M c ( t ) to capture spatial information and self-information. Enclosed transformers and periodic transformers benefit from input sequences that incorporate recent and cyclical characteristics.

4.5. Experiment Analysis

We conducted experiments on our in-house dataset of cellular network traffic to contrast the proposed TSENet against alternative models, and the assessment outcomes are illustrated in Table 3. As shown, ARIMA exhibits elevated MAE and RMSE on the dataset compared to alternative models, primarily because it disregards other dependencies and concentrates solely on temporal features. SLTM’s performance is inferior to deep learning approaches but surpasses statistical methods. HSTNet exclusively attends to temporal attributes, overlooking external factors such as BS information and POI distribution. Similarly, STDenseNet neglects external factors. Despite TWACNet utilizing convolution-based networks, its performance falls short of ConvLSTM-based methods. STCNet employs ConvLSTM-based networks but lacks the incorporation of self-attention and correlation layers for improved feature extraction. The aforementioned techniques heavily rely on networks for uncovering latent information, potentially inadequately modeling continuous spatiotemporal features. Compared to other methods, TSENet achieves the best performance. On one side, TSM and SSM amplify the capacity to capture continuous spatiotemporal features. Furthermore, TSENet introduces a self-attention mechanism, aggregating the capability to capture dependencies between extracted external information and network traffic information.

4.6. Validation of Key Components

1) Augmentation of Data: When setting Q to zero, it is equivalent to omitting data augmentation from TSM. As discerned from Figure 5, the augmentation of data enhances the predictive outcomes, with the minimum value of MAE observed when Q is set to 15.
2) Grid Selection: Setting K to 1 is comparable to removing grid selection from the set-top box. As evidenced in Figure 5, grid selection markedly enhances predictive performance. Optimal performance is achieved when selecting 20 grids.
3) Spatial Features in SSM: In order to substantiate the significance of extracting spatial features, we conducted experiments solely utilizing TSM. The outcomes of TSENet (TSM only) presented in Table 4 underscore the efficacy of capturing spatial relationships. Additionally, we utilized Graph Convolutional Network (GCN) [31] in lieu of SSM to represent spatial features in TSENet, assessing the spatial modeling efficacy of SSM through experiments. The results in Table 3 demonstrate that incorporating SSM yields better performance compared to GCN, enhancing the overall model performance.
4) Temporal Fusion with SSM: We also trained TSENet (without SSM fusion), where SSM is excluded from TSM. As illustrated in Table 4, TSENet (excluding SSM fusion) demonstrates superior performance in anomalous scenarios, confirming the spacial transformer’s effectiveness in capturing real-time spacial dependence.

5. Conclusions

This paper introduces a novel TSENet to simultaneously explore spatial sequences, temporal sequences, and external factor information. TSENet treats cellular data across grids within a defined time interval as a spacial sequence, facilitating the modeling of real-time spatial correlations globally. Additionally, incorporating a self-attention mechanism, this method integrates the capability to capture dependencies among external factors and network traffic features. Experimental results show the efficacy of TSENet, affirming the utility of the transformer architecture in cellular traffic prediction. This suggests that our proposed approach can enhance the accuracy of cellular network traffic forecasting. Future endeavors will involve developing more efficient transformers, unearthing additional valuable external information, and further refining predictions.

References

  1. Orfanus, D.; Eliassen, F.; de Freitas, E.P. Self-organizing relay network supporting remotely deployed sensor nodes in military operations. 2014 6th International Congress on Ultra Modern Telecommunications and Control Systems and Workshops (ICUMT). IEEE, 2014, pp. 326–333.
  2. Giménez, P.; Molina, B.; Calvo-Gallego, J.; Esteve, M.; Palau, C.E. I3WSN: Industrial intelligent wireless sensor networks for indoor environments. Computers in Industry 2014, 65, 187–199. [Google Scholar] [CrossRef]
  3. Li, X.; Li, D.; Wan, J.; Vasilakos, A.V.; Lai, C.F.; Wang, S. A review of industrial wireless networks in the context of Industry 4.0. Wireless networks 2017, 23, 23–41. [Google Scholar] [CrossRef]
  4. Gomes, R.D.; Queiroz, D.V.; Lima Filho, A.C.; Fonseca, I.E.; Alencar, M.S. Real-time link quality estimation for industrial wireless sensor networks using dedicated nodes. Ad Hoc Networks 2017, 59, 116–133. [Google Scholar] [CrossRef]
  5. Yu, L.; Li, M.; Jin, W.; Guo, Y.; Wang, Q.; Yan, F.; Li, P. STEP: A spatio-temporal fine-granular user traffic prediction system for cellular networks. IEEE Transactions on Mobile Computing 2020, 20, 3453–3466. [Google Scholar] [CrossRef]
  6. Santos, G.L.; Rosati, P.; Lynn, T.; Kelner, J.; Sadok, D.; Endo, P.T. Predicting short-term mobile Internet traffic from Internet activity using recurrent neural networks. International Journal of Network Management 2022, 32, e2191. [Google Scholar] [CrossRef]
  7. Hardegen, C.; Pfülb, B.; Rieger, S.; Gepperth, A. Predicting network flow characteristics using deep learning and real-world network traffic. IEEE Transactions on Network and Service Management 2020, 17, 2662–2676. [Google Scholar] [CrossRef]
  8. Dodan, M.; Vien, Q.; Nguyen, T. Internet traffic prediction using recurrent neural networks. EAI Endorsed Transactions on Industrial Networks and Intelligent Systems 2022, 9. [Google Scholar] [CrossRef]
  9. Mozo, A.; Ordozgoiti, B.; Gomez-Canaval, S. Forecasting short-term data center network traffic load with convolutional neural networks. PloS one 2018, 13, e0191939. [Google Scholar] [CrossRef] [PubMed]
  10. Dalgkitsis, A.; Louta, M.; Karetsos, G.T. Traffic forecasting in cellular networks using the LSTM RNN. Proceedings of the 22nd Pan-Hellenic Conference on Informatics, 2018, pp. 28–33.
  11. Wang, H.; Ding, J.; Li, Y.; Hui, P.; Yuan, J.; Jin, D. Characterizing the spatio-temporal inhomogeneity of mobile traffic in large-scale cellular data networks. Proceedings of the 7th International Workshop on Hot Topics in Planet-scale mObile computing and online Social neTworking, 2015, pp. 19–24.
  12. Zhao, N.; Ye, Z.; Pei, Y.; Liang, Y.C.; Niyato, D. Spatial-temporal attention-convolution network for citywide cellular traffic prediction. IEEE Communications Letters 2020, 24, 2532–2536. [Google Scholar] [CrossRef]
  13. Zhang, C.; Zhang, H.; Yuan, D.; Zhang, M. Citywide cellular traffic prediction based on densely connected convolutional neural networks. IEEE Communications Letters 2018, 22, 1656–1659. [Google Scholar] [CrossRef]
  14. Buratti, C.; Conti, A.; Dardari, D.; Verdone, R. An overview on wireless sensor networks technology and evolution. Sensors 2009, 9, 6869–6896. [Google Scholar] [CrossRef] [PubMed]
  15. Lin, C.Y.; Tseng, Y.C.; Lai, T.H. Message-efficient in-network location management in a multi-sink wireless sensor network. IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing (SUTC’06). IEEE, 2006, Vol. 1, pp. 8–pp.
  16. Krishnasamy, L.; Dhanaraj, R.K.; Ganesh Gopal, D.; Reddy Gadekallu, T.; Aboudaif, M.K.; Abouel Nasr, E. A heuristic angular clustering framework for secured statistical data aggregation in sensor networks. Sensors 2020, 20, 4937. [Google Scholar] [CrossRef] [PubMed]
  17. Wang, L.N.; Zang, C.R.; Cheng, Y.Y. The short-term prediction of the mobile communication traffic based on the product seasonal model. SN Applied Sciences 2020, 2, 1–9. [Google Scholar] [CrossRef]
  18. Shu, Y.; Yu, M.; Yang, O.; Liu, J.; Feng, H. Wireless traffic modeling and prediction using seasonal ARIMA models. IEICE transactions on communications 2005, 88, 3992–3999. [Google Scholar] [CrossRef]
  19. Huang, J.; Xiao, M. Mobile network traffic prediction based on seasonal adjacent windows sampling and conditional probability estimation. IEEE Transactions on Big Data 2020, 8, 1155–1168. [Google Scholar] [CrossRef]
  20. Cai, Y.; Cheng, P.; Ding, M.; Chen, Y.; Li, Y.; Vucetic, B. Spatiotemporal Gaussian process Kalman filter for mobile traffic prediction. 2020 IEEE 31st Annual International Symposium on Personal, Indoor and Mobile Radio Communications. IEEE, 2020, pp. 1–6.
  21. Xia, H.; Wei, X.; Gao, Y.; Lv, H. Traffic prediction based on ensemble machine learning strategies with bagging and lightgbm. 2019 IEEE International Conference on Communications Workshops (ICC Workshops). IEEE, 2019, pp. 1–6.
  22. Zhang, Q.; Mozaffari, M.; Saad, W.; Bennis, M.; Debbah, M. Machine learning for predictive on-demand deployment of UAVs for wireless communications. 2018 IEEE global communications conference (GLOBECOM). IEEE, 2018, pp. 1–6.
  23. Owens, F.J.; Lynn, P.A. Signal Processing Of Speech, Macmillan New electronics 1993.
  24. Baxevanis, A.; Bader, G.; Wishart, D. Bioinformatics. Hoboken, 2020.
  25. Wen, Y.; Xu, P.; Li, Z.; Xu, W.; Wang, X. RPConvformer: A novel Transformer-based deep neural networks for traffic flow prediction. Expert Systems with Applications 2023, 218, 119587. [Google Scholar] [CrossRef]
  26. Zhang, C.; Patras, P. Long-term mobile traffic forecasting using deep spatio-temporal neural networks. Proceedings of the Eighteenth ACM International Symposium on Mobile Ad Hoc Networking and Computing, 2018, pp. 231–240.
  27. Qi, W.; Yao, J.; Li, J.; Wu, W. Performer: A Resource Demand Forecasting Method for Data Centers. International Conference on Green, Pervasive, and Cloud Computing. Springer, 2022, pp. 204–214.
  28. Zhang, C.; Dang, S.; Shihada, B.; Alouini, M.S. Dual attention-based federated learning for wireless traffic prediction. IEEE INFOCOM 2021-IEEE conference on computer communications. IEEE, 2021, pp. 1–10.
  29. Rao, Z.; Xu, Y.; Pan, S.; Guo, J.; Yan, Y.; Wang, Z. Cellular Traffic Prediction: A Deep Learning Method Considering Dynamic Nonlocal Spatial Correlation, Self-Attention, and Correlation of Spatiotemporal Feature Fusion. IEEE Transactions on Network and Service Management 2022, 20, 426–440. [Google Scholar] [CrossRef]
  30. Feng, J.; Chen, X.; Gao, R.; Zeng, M.; Li, Y. Deeptp: An end-to-end neural network for mobile cellular traffic prediction. IEEE Network 2018, 32, 108–115. [Google Scholar] [CrossRef]
  31. Kipf, T.N.; Welling, M. Semi-supervised classification with graph convolutional networks. arXiv 2016, arXiv:1609.02907. [Google Scholar]
Preprints 98344 g001
Figure 1. (a): Single-Receiver WSN. (b) : Multiple-Receiver WSN.
Figure 1. (a): Single-Receiver WSN. (b) : Multiple-Receiver WSN.
Preprints 98344 g001
Preprints 98344 g002
Figure 2. Overview of the proposed TSENet.
Figure 2. Overview of the proposed TSENet.
Preprints 98344 g002
Preprints 98344 g003
Figure 3. Structures of External Factor Module.
Figure 3. Structures of External Factor Module.
Preprints 98344 g003
Preprints 98344 g004
Figure 4. Structures of Fusion Output Module.
Figure 4. Structures of Fusion Output Module.
Preprints 98344 g004
Preprints 98344 g005
Figure 5. (a) MAE values when different transformers employ distinct initial sequences, (b) MAE values for SSM and enclosed transformers with varied K and Q.
Figure 5. (a) MAE values when different transformers employ distinct initial sequences, (b) MAE values for SSM and enclosed transformers with varied K and Q.
Preprints 98344 g005
Table 1. Parameters of different transformer.
Table 1. Parameters of different transformer.
Transformer d h Source sequence Initial sequence
Spatial 64 8 Augmented M s ( t ) Average of M c ( t )
Closeness 256 8 M c ( t ) M p ( t )
Period 128 8 M p ( t ) M c ( t )
Table 2. Different combinations of time converters
Table 2. Different combinations of time converters
Combination MAE NRMSE R 2
M c (hourly) 11.3955 0.55332 0.81508
M p (hourly) 10.5194 0.57136 0.81226
M c + M p 10.4823 0.55282 0.81539
M c + M t (weekly) 17.6349 0.86621 0.57041
M p + M t 17.6375 0.77772 0.63955
M c + M p + M t 17.4319 0.74192 0.67187
Table 3. Experimental results of RMSE and MAE.
Table 3. Experimental results of RMSE and MAE.
Methods MAE NRMSE R 2
ARIMA 21.19 0.8364 0.3574
LSTM 15.341 0.7613 0.7275
DenseNet 13.243 0.6241 0.7830
TWACNet 11.22 0.6198 0.8021
HSTNet[39] 10.930 0.5549 0.8295
STCNet 10.898 0.5766 0.8271
TSENet(Ours) 9.758 0.51858 0.84302
Table 4. Transformer-based verification of key components
Table 4. Transformer-based verification of key components
Methods MAE NRMSE R 2
TSENet(TSM only) 10.3499 0.54258 0.82695
TSENet(with GCN) 10.9125 0.57528 0.80473
TSENet(w/o SSM in fusion) 9.9268 0.53369 0.83329
TSENet 9.758 0.51858 0.84302
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.
Prerpints.org logo

Preprints.org is a free preprint server supported by MDPI in Basel, Switzerland.

facebook logotwitter logolinkedin logo
MDPI Initiatives
Important Links
Subscribe

Disclaimer

Terms of Use

Privacy Policy

© 2024 MDPI (Basel, Switzerland) unless otherwise stated