A Spatio-temporal Prediction Methodology Based on Deep Learning and Real Wi-Fi Measurements

TL;DR

The paper tackles predicting future Wi-Fi AP metrics by exploiting spatial correlations among neighbouring APs. It formulates the problem as learning a function $f$ that maps the past $L$ observations from an AP (and, when available, its $M_i$ correlated neighbours) to the next value, with neighbour sets determined by a correlation threshold $Ths$. The methodology uses offline training and online prediction, evaluates DL models including SRNN, CNN, GRU, LSTM, Transformer, and several hybrid CNN-RNN architectures on measurements from $I=100$ APs over $D=75$ days, and reports high predictive accuracy (e.g., $R^2$ up to about 0.93 for some targets) with modest online computation times. The results support proactive network management, enabling load balancing and resource allocation with practically real-time PCTs while training is performed offline, and show that including spatial information yields gains over purely temporal predictions.

Abstract

The rapid development of Wi-Fi technologies in recent years has caused a significant increase in the traffic usage. Hence, knowledge obtained from Wi-Fi network measurements can be helpful for a more efficient network management. In this paper, we propose a methodology to predict future values of some specific network metrics (e.g. traffic load, transmission failures, etc.). These predictions may be useful for improving the network performance. After data collection and preprocessing, the correlation between each target access point (AP) and its neighbouring APs is estimated. According to these correlations, either an only-temporal or a spatio-temporal based prediction is done. To evaluate the proposed methodology, real measurements are collected from 100 APs deployed in different university buildings for 3 months. Deep Learning (DL) methods (i.e. Convolutional Neural Network (CNN), Simple Recurrent Neural Network (SRNN), Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), Transformer) are evaluated and compared for both temporal and spatio-temporal based predictions. Moreover, a hybrid prediction methodology is proposed using a spatial processing based on CNN and a temporal prediction based on RNN. The proposed hybrid methodology provides an improvement in the prediction accuracy at expenses of a slight increase in the Training Computational Time (TCT) and negligible in Prediction Computational Time (PCT).

Figures (4)

• Figure 1: General architecture of proposed methodology.
• Figure 2: Input and output data in the spatio-temporal prediction process.
• Figure 5: Location of APs in study and class area.
• Figure 6: Comparing the Real and Prediction traffic load of target AP(XSFA4PS205) using LSTM in four days of the last week.