Features of Precedents Space of Artificial Neural Networks for the Solar PV Station Control

Abstract

In the partial shading conditions, when photovoltaic modules in different parts of the vast area of the solar station are under different insolation, many local maxima appear on the energy characteristic. Only one among them provides the maximum power generated by the power station. Standard methods for search and maintaining the maximum power point, designed to work under uniform insolation conditions, lose the ability to detect the maximum power point with partial shading and bring the photovoltaic station mode to a local peak point with significantly less power generation. Since the configuration of insolation can change relatively quickly, special algorithms are used to control the efficiency of the photovoltaic station, providing a quick determination of the vicinity of the global maximum power point. The paper presents a new algorithm implementing a high-speed output of a photovoltaic station mode in the vicinity of a point with maximum power generation with subsequent pass of control to standard methods of maintaining a working point. The basis of the algorithm is a neural network using four nonlinear classifiers, tuned with the use of support vector method. The space of precedents of the training sample of the neural network has a dimension equal to 3, although the measurement vector includes power values at 4 points of the energy characteristic. To make the algorithm universal, it is proposed to present the characteristics of the photovoltaic modules of the power station in the form of normalized dependencies. The process of tuning the neural network is illustrated by images of dividing surfaces in the use-case space of the training sample. A tuned neural network transfers the photovoltaic station mode to maximum energy production mode in a maximum of four stages of voltage change at the output of the photovoltaic modules.