A Machine Learning Approach for Path Loss Prediction Using Combination of Regression and Classification Models.

Area of Science:

• Wireless Communication Engineering
• Radio Propagation Modeling
• Machine Learning Applications

Background:

• Accurate radio path loss prediction is crucial for effective radio link planning.
• Existing prediction methods often lack a balance between accuracy, generality, and computational efficiency.
• The 433 MHz frequency band is vital for various wireless systems like IoT and LPWAN.

Purpose of the Study:

• To develop a generalized and computationally efficient machine learning model for radio path loss prediction.
• To improve the accuracy of path loss prediction across various terrains and propagation conditions (line-of-sight and non-line-of-sight).
• To create a flexible model applicable to different antenna heights and environmental types.

Main Methods:

• A novel compound machine learning model combining two regression models and one classifier was developed.
• The model utilizes only five input parameters: distance, antenna heights, and terrain/obstacle statistics.
• A classification model probabilistically combines outputs from regression models trained for line-of-sight and non-line-of-sight conditions.

Main Results:

• The proposed machine learning approach achieved a low root mean square error of 7.3 dB.
• A high coefficient of determination (R-squared) of 0.702 was recorded, indicating strong predictive performance.
• The model demonstrated excellent performance across diverse terrains (flat, hilly, mountain) and areas (rural, urban, suburban).

Conclusions:

• The developed compound machine learning model offers a superior balance of accuracy, generality, and low computational complexity for path loss prediction.
• The model's flexibility allows for application in various radio link planning scenarios, including those for IoT and LPWAN systems.
• While validated at 433 MHz, the model is adaptable for other frequencies within the decimeter wavelength range.