On the Issue of Reducing the Negative Impact of Erroneous Data in the Training Sequence of a Predictive Model

The paper analyzes the approach to the problem of building and machine learning of a predictive model of a photovoltaic station. Such models are designed to predict hourly electricity generation a day in advance in accordance with the requirements of the energy market. One of the main problems that need to be resolved that the predictive modeling methodology faces is errors in a number of training sequences. Meteorological errors in the retrospective database on which the model is trained lead to the fact that later, when making forecasts, even accurate weather data are interpreted by the model in the form of erroneous forecasts of photovoltaic energy generation. In this regard, the problem of developing a training predictive model of a photovoltaic station that has the ability to learn correctly in terms of limited reliability of input meteorological data poses and is being solved. A new ensemble method for learning a predictive model has been developed and tested in real conditions in terms of operating photovoltaic stations; the method makes it possible to reduce the influence of meteorological forecast errors on the accuracy of predicting the amount of electricity generation.