Application of artificial neural network in modeling illumination distribution in three-dimensional space

The paper deals with modeling three-dimensional illumination distribution from light sourсes using artificial neural network (ANN). This problem is especially relevant in greenhouse irradiation systems due to the design features, namely large dimensions of the phytoirradiators relative to the small suspension height, which makes it impossible to represent them as point sources. The aim of the research is to develop a methodology for modeling illumination distribution in threedimensional space, which includes obtaining experimental data and processing them using an artificial neural network. A dataset comprising 2,100 records was obtained experimentally for the developed phytoirradiator. The illumination values E were measured at the nodes of the coordinate grid with a step of 10 cm within the range from 0 to 120 cm (along the x coordinate) and from 0 to 100 cm (along the y coordinate) in all four quadrants, at different suspension height of the phytoirradiator h at levels of 30, 50, 70 and 90 cm. The TensorFlow and Keras libraries of the Python programming language were used to build the model. The results showed that the created neural network effectively describes the flux distribution on the irradiated surface, taking into account the real geometry of the light source. The key advantage of the method is the ability to calculate illumination for any combination of suspension height and point coordinates on the plane, which overcomes the limitations of conventional lighting calculations based on the inverse square law. The mean absolute error of the neural network model is 0.04 klx, the value of the determination coefficient R2 = 0.9967 with a 95 % confidence interval of [0.09953, 0.9977], which is a good result. Mean absolute prediction error is 7.5 %, this value can be improved by regularization and data augmentation. It has been established that ANN-based method is applicable to the design of energy-efficient lighting systems, adaptable to the spectral characteristics and reflected light. After retraining, the model can be used for phytoirradiators of arbitrary design, which expands its practical significance in plant lighting.