Multilingual web traffic forecasting for network management using artificial intelligence techniques

• Developed a modular architecture for accurate web traffic forecasting using deep learning. • Proposed CNN, LSTM, and GRU models for time-series prediction of website visits. • Collected and analyzed BBC World Service traffic data in English, Arabic, and Persian. • Evaluated models using MAE, RMSE, RRMSE, and R², with CNN achieving the highest accuracy. Web traffic forecasting is critical for ensuring an effective allocation of network resources for websites to avoid potential congestion and maintain service quality. Analyzing traffic forecasts helps network operators plan for future capacity needs to accommodate unexpected traffic loads. Multilingual web platforms present unique challenges due to varying traffic patterns across several languages. In this paper, modular architecture is proposed for multilingual web traffic forecasting using deep learning models. The architecture consists of multiple processes, from collecting the traffic data of the targeted website to preparing the data to be in a suitable form to feed the deep learning models. This study explores time-series forecasting methods using three deep learning models: long short-term memory (LSTM), gated recurrent units (GRU), and convolutional neural networks (CNN). A comprehensive dataset was collected from the BBC World Service websites in three languages—English, Arabic, and Persian—spanning the period from January 2017 to July 2024. The performance of the employed deep learning models is evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), relative root mean square error (RRMSE), and R-squared (R²). Results show that the CNN model outperformed the GRU and LSTM models in accuracy and variance explanation across all three language datasets. The CNN model demonstrated superior performance, achieving R² scores of 0.980, 0.989, and 0.994; MAE values of 47,195, 5090, and 6714; RRMSE values of 1.43 %, 1.90 %, and 1.86 %; and RMSE values of 61,372, 6342, and 9376 for the English, Arabic, and Persian websites, respectively. These results highlight the potential of deep learning models in capturing complicated patterns of multilingual web platforms, making them suitable for real-world network management.