Solar Insolation Prediction Based on Machine Learning: Error Analysis on Data Similarity and Predictive Model Uncertainty

This study aims to develop a reliable solar radiation prediction model by analyzing the similarity of data to reduce the uncertainty of Machine Learning (ML) based predictions.Variable selection was conducted using the Pearson Correlation Coefficient (PCC), and the similarity of the selected variables was measured through Cosine similarity, Euclidean distance, and Manhattan distance analyses.8 scenarios were established to analyze hourly errors and Q1, Q2, and Q3 statistics.As a result, it was found that higher data similarity leads to increased prediction accuracy and a reduction in the range of error values.Normalization methods such as QuantileTransformer contributed to improving model performance by being less sensitive to data outliers.Notably, Scenario 8 recorded the lowest average prediction error, maintaining consistency in predictions.