Downlink Throughput Prediction In Cellular Networks Using Multimodal Deep Learning Model

Accurate prediction of user throughput is critical for optimizing cellular network performance and enhancing user experiences. It enables network operators to identify and address bottlenecks, improve resource allocation, and optimize base station deployment. These actions collectively enhance network coverage, reduce congestion, and provide enriched end-user experiences, a key milestone in advancing cellular network management. Traditional methods for modeling user throughput rely on statistical averaging, which often lacks accuracy and completeness. To overcome these limitations, this paper proposes a novel multimodal approach that integrates numerical features from drive test data, network Key Performance Indicators (KPIs), and satellite images. By combining these diverse data sources with advanced Machine Learning (ML) and Artificial Intelligence (AI) algorithms, the proposed model offers a more accurate and comprehensive solution for modeling downlink user throughput. Satellite images provide valuable environmental information, such as land cover, topography, and urbanization, which significantly influence radio conditions. These, along with numerical features and KPIs, are used to train a predictive model. Once trained, the model predicts user throughput without requiring additional drive tests, making it a cost-effective and efficient solution. Evaluation using real-world datasets demonstrated the effectiveness of the multimodal approach, achieving a Mean Absolute Error (MAE) of 2.84 Mbps and an R2 score of 0.975. This represents a 64% improvement over numerical-only models, highlighting the importance of environmental factors and multimodal data integration in achieving accurate user throughput predictions.