Karpenko O. The system for monitoring photovoltaic stations with intelligent decision support system based upon prognostic models

The thesis deals with solution of the actual scientific and technical problem to maintain the required conditions of the balance operation by energy network and energy market through integration of adequate simulation methods and temporal series of photovoltaic energy generation, using photovoltaic stations with controlling and information and measurement systems monitoring all the necessary information. The thesis has analyzed features of the system, controlling photovoltaic processes, involving intelligent support to make decisions based upon prognostic models, and the process of photovoltaic conversion of solar energy as an object of the automated predicting. The analysis results have helped make the decision concerning the development of exogenous model taking into consideration both retrospective data and simultaneous with the data vector of parameters influencing the generation level. Meteorological data as well as data characterizing timely the sun position in the sky relative to geolocation of the photovoltaic station are among the parameters. The exogenous model predictor is based upon mathematical description of the photovoltaic process. In this regard, the mathematical description development has involved a hybrid approach. According to the approach, the prognostic model includes the two functional blocks: analytical predicting block and statistical predicting block with the use of machine learning methods. The analytical predicting block calculates the value of solar radiation energy, achieving photovoltaic station panels if the sky is cloudless, for each hour of each day and a year month depending upon the sun position. The statistical predicting block calculates the predicted hourly electric energy generation by the photovoltaic station taking into consideration the predicted atmospheric state. Available approaches have been considered as for the development of a prognostic model blocks. It has been mentioned that lack of efficient mechanism, protecting predicting results from false signs of the meteorological state of the environment in the retrospective databases used for machine learning models, is the key disadvantage of the considered approaches to develop such a statistical predicting block. Due to the false signs, a model is learnt incorrectly which results in significant future deviations of the forecast from reality even if meteorological forecasts at the predicted hour are qualitative and unmistakable. To overcome the drawback, it is proposed to develop such a model structure, which would apply reverse (reflexive) mathematical transformation (i.e. physical use of photovoltaic modules as solar radiation sensors) for automatic correction of selective retrospective meteorological data in accordance with functional relation between them, other data, and actual level of electric energy generation. It has been noted that polynomial prognostic model, developed using the least square method where dependence of the generation station capacity upon influencing factors is obtained through the machine learning results in the form of a polynom (i.e. analytically) matches the problem most of all. Hence, it becomes possible to derive in an explicit–form influence functions of meteorological factors on the level of photovoltaic energy generation and use the available mathematical apparatus relative to the polynomial model transformations. The model in the work has been developed relying upon the proposed hypothesis to factorize influence function of meteorological factors on the generated photovoltaic energy level. According to the hypothesis, the general influence function can be represented in the form of product of partial influence functions which of them takes into consideration the influence of one of the factors. The hypothesis has been supported in the work through a method of correlation analysis of photovoltaic transformation of the solar energy at the operating network photovoltaic stations using Pearson coefficient of determination as the criterion.