Reduction of Training Samples in Solar Insolation Prediction Under Weather and Climatic Changes

Abstract

The paper considers the problem of forecasting solar insolation. Due to a large number of factors that are difficult to predict, this problem is complex and difficult like other problems where parameters depend on weather or climate. Despite such factors do not significantly affect the parameters under study, they create an essential bias. The problem is also characterized by a significant amount of information, which should be processed to obtain a reliable forecast. The paper discusses relevant machine learning methods and analyzes methods for reducing the training sample to make quality and reliable predictions of solar insolation under the weather and climatic changes. Reducing the size of the training sample allowed testing a significant number of models and optimizing the hyperparameters of these models, which made it possible to identify the most accurate model, which graph is close to ideal. The effect of reducing the size of the training sample has been measured related to the speed and accuracy of machine learning algorithms. It is shown that the use of clustered data sampling for models that prone to overfitting can improve the accuracy of these models. It is shown that using the method of clustering to reduce data samples can also reduce variance in the training dataset. The reduced data sample helps to perform optimization of hyperparameters because the accuracy of different models is preserved over a wide range of input data. The results obtained can be successfully applied in control systems for large objects of solar energy generation under conditions of fast and frequent weather changes as well as slow climatic changes.KeywordsSolar insolationMachine learningReduction of training samplesOptimizing the hyperparameters