Machine learning methods for solar power generation maximizing and forecasting

Abstract

Systems for insolation forecasting based on machine learning methods increase the efficiency of a solar plant. Therefore, these systems are relevant in accordance with the priority of the state energy policy “Energy Strategy for the period up to 2035”. The insolation has complex non-linear dynamics with uncertainties due to changes in cloudiness. Thus, it is impossible to approximate these complex dynamics with classical methods with a given accuracy, while machine learning methods provide the required accuracy. When solving solar prediction problems, intelligent methods in comparison with traditional methods provide the required accuracy of solving these problems by contributing to the safe and effective management of electric grids that integrating solar power plants. The problem of hourly forecasting of insolation for a day ahead in conditions of uncertainty was solved based on a modified fuzzy neural network. We modified the method for creating a modified fuzzy neural network which reduced the dimension of the search space for swarm particles and computational costs by simplifying optimization. The modified fuzzy neural network provides by means of recurrent neurons and the attention mechanism the effective generation and transmission of a hidden representation of information as a signal of the hidden layer of deep neural networks, on the basis of the outputs of which the predicted value of insolation is generated by the fuzzy-possible convolution algorithm. The modified fuzzy neural network effectively distinguishes from the data significant functional aspects of solar prediction, including aspects of identifying the specifics of the cloudiness of the hour. The obtained results of comparative experimental modeling of a modified fuzzy neural network when predicting insolation for a day ahead demonstrate its robustness and a decrease in the mean square error of its forecast by an average of three and six times in comparison with recurrent neural networks and a standard model of moving average autoregression in conditions of uncertainty.