Diffraction Loss Prediction using Neural Networks

Abstract

The radio propagation prediction is essential for communication coverage calculations, and also plays an important role in wireless communication network designing, cell planning, etc. In mountainous areas, there are natural obstacles in the propagation path caused by the terrains and losses due to the diffraction. This diffraction effect must be considered for prediction accuracies of propagation losses. Various methods for predicting the diffraction losses have been studied. These methods can be divided into empirical methods based on the knife-edge-approximation and deterministic methods that take into account various parameters such as the material's conductivity and the detailed information of the topography in the location where the diffraction occurs. In the latter case, it is possible to predict more precisely, but it is not practical because it is difficult to obtain detailed information on the real mountainous terrains. In addition, it is inefficient in that it requires considerably more complicated computations. Therefore, in this dissertation, three methods considering a knife-edge approximation are proposed for predicting the diffraction loss, which can be applied to mountain areas with Digital Terrain Model (DTM). The first method calculates the diffraction losses using the Bullington model (widely used to predict the diffraction loss) and the Artificial Neural Network (ANN). The second method calculates the diffraction losses by using the ANN only. In both methods, the measured data are divided into training sets and the test sets. Training sets are used only for the ANN training. The test sets are used to compare the performances of the two proposed methods with the traditional methods (Bullington and Deygout). As an additional experiment, a method that can tune the parameters of the Lee model with the diffraction loss models is proposed. Futhermore, the parameters of this proposed method are adjusted using the variable step size Least Mean Square (LMS) algorithm. Experiment results show that better prediction performances are achieved by the proposed methods.