LSTM-Based Packet Loss Prediction in Communication Networks: A Comprehensive Review

Packet Loss Prediction (PLP) has become a very important field of research that helps to increase performance, reliability, and efficiency of modern networks. The development of the prediction techniques went through several stages starting from classical statistical models like ARIMA and Markovian models to machine learning approaches such as Decision Trees, Support Vector Machines, Random Forest, and XGBoost. While these techniques ensure reasonable performance in case of stable networks, they usually have problems dealing with non-linear nature and time dependency of dynamic network traffic. Deep learning algorithms, especially LSTM, helped to make great improvements in PLP because of their ability to learn long-term dependencies. As this literature review has shown, LSTM is still considered the best architecture of deep learning used in packet loss prediction because of the memory-based structure, solution to the problem of the vanishing gradient, and better results in working with time series. Compared with machine learning algorithms and classical recurrent neural networks, LSTM demonstrates better prediction accuracy, RMSE and MAE values, and greater precision and recall. In addition, the use of LSTM in 5G/6G networking, Internet of Things, Software-Defined Networking, cloud and edge computing, and intelligent transport systems shows its relevance to current networking problems. However, there are some drawbacks in addition to the benefits described above. The LSTM neural networks usually require big training sets, lots of computational power, and time for training that is bigger compared to other methods. The black box algorithm also makes it difficult to interpret results, thus increasing difficulties with deploying these models in safety-critical situations. Accordingly, future research efforts should thus concentrate on the application of Explainable AI, Federated Learning, TinyML, Transformer-LSTM model architectures, Graph Neural Networks, Digital Twin Networks, and Reinforcement Learning in order to address the above-stated drawbacks. With the above-mentioned techniques having a lot of potential to enhance prediction precision, computational efficiency, scalability, privacy, and adaptiveness as well as provide intelligent network management, it is evident that efficient packet loss prediction models will be an essential area for future research. This is especially relevant in the context of the evolution of communication systems to 6G, edge intelligence, and large-scale IoT.

[1]J. Yao et al., “A Data-Driven Framework for TCP to Achieve Flexible QoS Control in Mobile Data Networks,” 2023 IEEEACM 31st Int. Symp. Qual. Serv. IWQoS, vol. null, pp. 1–11, 2023, doi: 10.1109/IWQoS57198.2023.10188765. [2]A. Das, W. Kong, R. Sen, and Y. Zhou, “A decoder-only foundation model for time-series forecasting,” 2023, doi: 10.48550/arXiv.2310.10688. [3]J. Zhou et al., “A Machine Learning-Based Framework for Dynamic Selection of Congestion Control Algorithms,” IEEEACM Trans. Netw., vol. 31, pp. 1566–1581, 2023, doi: 10.1109/TNET.2022.3220225. [4]Z. Xia et al., “A Multi-objective Reinforcement Learning Perspective on Internet Congestion Control,” 2021 IEEEACM 29th Int. Symp. Qual. Serv. IWQOS, vol. null, pp. 1–10, 2021, doi: 10.1109/IWQOS52092.2021.9521291. [5]Y. Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam, “A Time Series is Worth 64 Words: Long-term Forecasting with Transformers,” ArXiv, vol. abs/2211.14730, p. null, 2022, doi: 10.48550/arXiv.2211.14730. [6]Z. Du, J. Zheng, H. Yu, L. Kong, and G. Chen, “A unified congestion control framework for diverse application preferences and network conditions,” Proc. 17th Int. Conf. Emerg. Netw. Exp. Technol., vol. null, p. null, 2021, doi: 10.1145/3485983.3494840. [7]J. Yao et al., “A Unified Framework for DRL-Based Congestion Control to Optimize QoS Over Mobile Networks,” 2025 IEEEACM 33rd Int. Symp. Qual. Serv. IWQoS, vol. null, pp. 1–10, 2025, doi: 10.1109/IWQoS65803.2025.11143480. [8]H. Tian et al., “Achieving Fairness Generalizability for Learning-based Congestion Control with Jury,” Proc. Twent. Eur. Conf. Comput. Syst., vol. null, p. null, 2025, doi: 10.1145/3689031.3696065. [9]Q. Hong, F. Meng, and F. Maldonado, “Advancing Long-Term Multi-Energy Load Forecasting with Patchformer: A Patch and Transformer-Based Approach,” ArXiv, vol. abs/2404.10458, p. null, 2024, doi: 10.48550/arXiv.2404.10458. [10]S. Rahmatinia and S. Hosseini-Seno, “An Attention-Based Deep Learning Model for Multi-Horizon Prediction of Load, Price, and Wind Power Generation in Smart Grids,” 2025 29th Int. Electr. Power Distrib. Conf. EPDC, vol. null, pp. 1–9, 2025, doi: 10.1109/EPDC67173.2025.11278281.