Macrocell radio wave propagation prediction using an artificial neural network

This paper presents and evaluates an artificial neural network model used for macrocell radio wave propagation prediction. Measurement data obtained by utilising the IS-95 pilot signal of a commercial CDMA mobile network in rural Australia is used to train the model. Simple models requiring only small amounts of training data have been used for the propagation predictions. The neural network inputs are chosen to be distance to base station and parameters easily obtained from terrain path profiles. It is concluded that a path loss predictor based on a simple neuron model generalises relatively well and requires only a few iterations in batch mode, using the Levenberg-Marquardt algorithm and early stopping, to converge to its optimum. The path loss prediction results using the neural models are favourably compared to the new semi-terrain based propagation model recommendation ITU-R P.1546, and traditional models, such as the Hata model. The statistical analysis shows that the simplistic artificial neural network approach is an alternative to traditional propagation models regarding accuracy, complexity and prediction time.