A Machine Learning Approach for Transmitted Power Estimation in Power Beaming Systems

Currently, predictive machine learning methods are used in many areas of life — from traffic predictions to medical diagnosis. Recently, these approaches also appeared in atmoposheric studies, first of all, for estimation of turbulence parameters, these tasks are necessary for a qualitative solution of atmospheric optical communication issues. The purpose of this work is to show the possibility and prospects of using machine learning algorithms for estimation transmitted power in power beaming systems in real time under changing atmospheric conditions. Experimental data were collected over several months on long atmospheric experimental setup, among gathered data there are such meteorological parameters as pressure, temperatures, wind speed, humidity, dew point, wind direction, solar flux. The data was collected for several locations. The power of the incident radiation was estimated from the voltage on the photovoltaic receiver. The nearest neighbors method, gradient boosting and neural networks were used as machine learning algorithms, the algorithms were compared with each other in terms of the average absolute error (MAPE) and the coefficient of determination (R2). The analysis of the results showed a good predictive ability of all models and potential of using even on the basis of simple meteorological measurements. The use of traditional methods requires much more complex measurements, such as scintillometry, or empirical approximations are used. Machine learning makes it possible to get results with only integral meteorological parameters and shows good accuracy for arbitrary conditions. Gradient boosting with R2 0.951 and MAPE 0.020 on all data was chosen as the best model. The results of this model was interpreted using the SHAP method, the dependence of the result on the input data is consistent with expectations.