Deep Learning Based Throughput Prediction model for 802.11ac WLANs

The accurate prediction of network throughput in IEEE 802.11ac wireless local area networks (WLAN) is essential for efficient dynamic resource allocation and optimization. Nonetheless, there are difficulties in achieving accurate performance prediction in indoor environments due to the presence of Non-Line-of-Sight (NLOS)environments in which physical obstacles lead to non-linear signal attenuation and multipath fading. Although conventional deterministic models and deep learning (DL) algorithms strive to capture these characteristics, they often face difficulty in processing the heterogeneity of tabular datasets generated from the actual deployment environments. In order to bridge the above gap, we present a novel throughput prediction approach that incorporates the physical environmental restrictions. This paper presents a feature engineering pipeline, which captures the spatial-signal characteristics from the physical indoor layout, in particular varying numbers of walls and doors in the environment, as composite mathematical interaction features. Using a dense grid of measurements, the performance of six state-of-the-art DL models is compared against four classical machine learning methods. Experiments show that the attention-based TabTransformer is the best suited model for capturing the heterogeneous feature set by mapping the complex spatial-signal relationships. The TabTransformer achieves an unprecedented throughput prediction accuracy with R^2 = 0.9689 and RMSE of 35.03 Mbps.