Channel prediction through machine learning techniques : A deep learning approach of forecasting channel state information over time

Understanding the characteristics of a wireless radio channel is a fundamental and challenging aspect of wireless communication. Through the years, numerous conventional signal processingtechniques have been developed to predict Channel State Information (CSI). This thesis introduces a new approach for forecasting the dynamic Rayleigh fading channel by harnessing the power of deep learning techniques, namely a Long Short-Term Memory (LSTM) neural network and a Convolutional Neural Network (CNN). By leveraging the inherent ability of neural networks (NNs) to identify non-linear patterns in data, the proposed method offers an innovative solution. To evaluate the performance of the neural network-based channel predictors, computer simulations are conducted in both noise-free and noisy, time-variant scenarios. The results are compared against the channel prediction capabilities of a linear filter, a Normalized Least Mean Squares (NLMS) filter. The goal was, to assess the prevailing assertion that machine learning predictors excel in discovering complex non-linear patterns in data and generally outperform traditional model-based linear solutions. Remarkably, the NNs yield impressive results, even surpassing the benchmark of -8 dB at half awavelength set by previous research papers. These finding confirmed the hypothesis that NNs can effectively identify patterns in radio channel data that linear predictors struggle to detect. However, it is noted that while the NNs achieved a prediction accuracy of -10 dB at half a wavelength, the author cannot conclusively claim that the NNs are the superior predictors. Their enhanced performance may be attributed to their higher computational complexity, and there is also a risk that the predictors are subjected to overfitting. In conclusion, this theses research highlights the potential of employing deep learning techniques, specifically LSTM and CNN, for forecasting time-variant Rayleigh fading channels. These findings contribute to the growing body of evidence supporting the ...