Solar Radiation Prediction using Long Short-Term Memory with Handling of Missing Values and Outliers

The pyranometer sensor is an instrument for measuring Global Horizontal Irradiance (GHI) which is used as parameter for analyzing and predicting weather. GHI data which is processed into prediction model for photovoltaics is useful for determining the performance of solar power generation systems in distributed energy operations. However, GHI sensor data has weaknesses in missing values and outliers due to measurement errors. The research designed a GHI sensor data prediction model using data preprocessing by the imputation of missing values using linear, polynomial, and Piecewise Cubic Hermite Interpolating Polynomials (PCHIP) interpolation and eliminating outliers using Random Sample Consensus (RANSAC) on the dataset. Previous researches show that Long Short-Time Memory (LSTM) can improve the performance of predictions compared to machine learning. This research designs an LSTM prediction model with data preprocessing and without data preprocessing. The results of the imputation of missing values obtained the best performance in PCHIP with Mean Absolute Error (MAE) 39.708 W/m2, Root Mean Absolute Error (RMSE) 76.224 W/m2, Normalized Root Mean Absolute Error (NRMSE) 0.433, and Coefficient Determination (R2) 0.903 then imputation from outlier elimination obtained MAE 44.377 W/m2, RMSE 86.738 W/m2, NRMSE 0.500, and R2 0.886. RANSAC testing succeeded in eliminating 100% outliers. The results of LSTM with data preprocessing obtained better performance with the best evaluations on MAE, RMSE, NRMSE, and R2 for test data of 42.863 W/m2, 82.396 W/m2, 0.396 and 0.918. This study contributes to GHI prediction model that can handle missing values ​​and outliers from sensors to support solar power plants.