Browsing by Author "LHR TP 7896"

Now showing 1 - 1 of 1

Nowcasting of RSSL in wireless communication channel over the sea using machine learning algorithms
(COMSATS University Islamabad Lahore Campus, 0023) Farwa Jafar,; FA20-REE-005; Dr. Khurram Zaidi; LHR TP 7896
The presence of naturally occurring evaporation duct (ED) phenomenon is very high in the tropical/equatorial regions of the world. Although, refractivity estimation of EM and Radio waves in ED is well studied in the literature, still, the signal propagation through ED over-the-horizon needs to be thoroughly researched to help determine the received-signal-strength-level (RSSL) for a reliable wireless communication link. In order to accurately predict RSSL in ED, we have acquired RSSL (avg.) per-minute data for three months over-the-horizon distance of 50 km (Tx-Rx) from onshore-to-offshore Oil & Gas Platform. This data was collected using fixed antenna heights in ED. Applying deep learning algorithms on real-time RSSL data, we have nowcasted the future RSSL values for next 5 seconds timescale in this thesis. A thorough comparison is made between the CNN and LSTM deep learning methods for real-time series prediction analysis. These deep learning networks are linked with numerous convolution layers to grasp the nonlinear mapping between measured and future RSSL values. Coding and Simulation work is performed in Python 3.9 environment and results are generated in Kaggle Notebook. CNN and LSTM networks have never been used earlier for predicting “signal strength” over-the-horizon and over-the-sea under ED environment. The contribution of this research is to bridge this gap and examine the accuracy of LSTM and CNN for nowcasting RSSL data. According to what we've discovered, both of these neural network models are capable of achieving adequate to high prediction power given that the "datasets" are suitably big. Both methods, when taken as a whole, are reliable with regard to their hyperparameters. However, with increasing number of training courses, LSTM didn’t improve its performance, whereas CNNs performed correspondingly more accurate each time. For 3rd training, CNN has given the most optimal fitting of training data as compare to test data. The RMSE achieved for CNN after third training was 4.47 which is the least of all simulations. Hence, CNNs proved to be superior, since they operate one order of magnitude quicker than LSTM. We proposed that the early predictive capability, speed, and resilience of CNN open its door to nowcasting’s future.