DEVELOPMENT AND COMPARATIVE ANALYSIS OF QUOTIENTS REGRESSION BASED EMPIRICAL AND ARTIFICIAL NEURAL NETWORK BASED MODELS FOR PATH LOSS PREDICTION

By

Author

Presented To

Department of Engineering

ABSTRACT
The aim of this study is to develop and analytically compare Quotients Regression based empirical models with artificial neural network (ANN) based models for path loss prediction. The terrains considered as case study include i) the rural terrain between Jos and Abuja ii) the Urban terrain (Abuja), and iii) the semi-urban terrain (Maiduguri). The empirical models considered include the Okumura, Hata-Okumura, COST 231 Hata and the COST 231Walfisch-Ikegami, while the two types of ANN include the Multilayer Perceptron Neural Network (MLP-NN) and the Generalised Radial Basis Function Neural Network (GRBF-NN). A Quotients Regression Technique (QRT) for empirical model adaptation was developed and used to selectively adapt these empirical models to the terrains, based on path loss measurements obtained from Base Transceiver Stations (BTS) situated within the terrains. The adaptation accuracy of the QRT was determined through comparisons with two existing adaptation techniques: i) The Okumura GAREA (Gain due to type of environment) correction factor, and ii) The Root Mean Squared Error (RMSE) Adaptation Technique (RAT). The comparative analysis of the QRT adapted empirical models with ANNbased models was based on three distinct approaches: i) Splitting path loss data into 60% Training, 10% Validation and 30% Testing, ii) Splitting path loss data into 50% Training Set and 50% Testing Set. iii) Random training with path loss data from one BTS and testing with data from another. The following results were obtained: i) The QRT has the highest adaptation accuracy, based on combined RMSE (Root Mean Squared Error) value of 2.1dB across the three terrains, the RAT technique has 5.66dB across the three terrains, while the Okumura GAREA has 8.95dB across the rural terrain alone, ii) The ANN-based models have the highest path loss prediction accuracy, based on an RMSE value of 3.98dB, followed by the QRT adapted empirical models with 4.49dB. The RAT adapted empirical models have 5.83dB, while the empirical models are the least accurate with 7.07dB. By implication, the ANN-based models only slightly outperform the QRT adapted empirical models by 0.51dB. However, the QRT adapted empirical models have the highest R2 (coefficient of determination) value of 0.81, and by implication, the best fit resulting from the best correlation with the measured path loss data. On the other hand, the QRT adapted empirical models offer an improvement of about 1.34dB over the RAT adapted empirical models, as well as an improvement of 2.58dB over existing Empirical Models. The proximity in performance of the QRT adapted empirical models to the ANN-based models can be attributed to the efficiency of the QRT.