فهرست نشریات علمی دسترسی آزاد ایران

In dynamic and heterogeneous network environments, the prediction of Voice over Internet Protocol (VoIP) traffic is complicated by constantly changing conditions such as congestion, bandwidth variability, and latency. This paper proposes an intelligent framework that integrates Centroid Probabilistic Clustering (CPC) with a hybrid CNN-LSTM deep learning model for forecasting dynamic VoIP traffic. CPC utilizes fuzzy C-means clustering to estimate and classify key traffic features, delay, jitter, and packet loss, enabling adaptive and interpretable prediction. The model’s centroids guide the deep learning components, which are a CNN that extracts spatial features, an LSTM captures temporal dependencies, and a Vector Autoregression (VAR) model linear interactions, enhancing robustness. Based on this design, the experimental evaluation demonstrates superior forecasting performance across multiple metrics, achieving, including a training accuracy of 95.1%, validation accuracy of 92.8%, MSE of 0.0148, MAE of 0.093, MAPE of 6.34%, and an R² of 0.979. Incorporating VAR reduces RMSE from 0.134 to 0.108. The framework adapts efficiently under varying congestion scenarios and cluster densities, proving highly scalable, interpretable, and well-suited for next-generation intelligent network management systems. It is important to note that CPC is positioned as an empirical variant of fuzzy C-means with probabilistic interpretation. While it does not currently include a formal convergence proof, its novelty lies in its empirical robustness and removal of the fuzzifier parameter.