Abstract
With the proliferation of devices that interconnect wirelessly, the radio spectrum has become increasingly congested and scarce due to the high number of devices using the limited spectrum space, inefficient use of allocated spectrum and the Internet of Things paradigm further steeps the demand. Cognitive radio has been developed as a system to allow for more efficient use of the spectrum, where Secondary Users (SUs) sense the spectrum originally allocated to Primary Users (PUs) for unused spaces (spectrum holes) and dynamically adjust its parameters to access the unused spectrum space. The introduced spectrum sensing subsystem consumes a portion of the SU’s battery power, thereby depleting the power meant for the SU’s transmission purposes. This work aimed to solve the problem of available energy for cognitive radio devices, in two folds, one, by scavenging for and harvesting energy from the ambient environment, using deep learning algorithms to predict future energy for harvesting and objective two, is to compare the results with traditional machine learning, based on the predicted harvested energy and error metrics. This combined technique focused on the use of deep learning techniques in improving the efficiency of CRNs where the predicted energy helps the device to have a foreknowledge of the amount of energy available on the transmission channel and using that for its internal optimum energy resource management. Three deep learning models were investigated namely Long-Short Term Memory (LSTM), Convolutional Neural Network (CNN) and Convolutional Long-Short Term Memory (ConvLSTM). To compare the deep learning models with machine learning models, three traditional machine learning models were also developed, namely, Artificial Neural Networks (ANN), Support Vector Regressor (SVR) and Extreme Gradient Boost (XGBoost). The results showed that the deep learning models outperformed the machine learning models across all datasets. More specifically, the ConvLSTM model performed better than the other models at a Normalized Root Mean Squared Error (nRMSE) of 0.0632 and Mean Absolute Error (MAE) of 1.479 which is 8.80% and 9.04%, superior, respectively, compared to the best performing machine learning model which was the ANN with nRMSE of 0.0693 and MAE of 1.626.