OPTIMIZING UNIVARIATE TIME SERIES IMPUTATION USING RANDOM FOREST REGRESSION AND LSTM FOR ACCURATE FORECASTING

Abstract

Indonesia possesses high solar radiation potential, making solar energy a strategic pillar for the national clean energy transition. However, its utilization is hindered by incomplete data due to instrument failure, which significantly reduces prediction accuracy. Starting from this problem, this study aims to evaluate the performance of the Machine Learning Based Univariate Time Series Imputation-Random Forest Regression (MLBUI-RFR) method by comparing it with the Mean Imputation method and evaluating it through Long Short-Term Memory (LSTM) forecasting. The methodology begins with data preprocessing using the MLBUI-RFR scheme to handle missing values, which are then used as input for the LSTM architecture to forecast solar radiation in Indonesia. The findings demonstrate that the use of the MLBUI-RFR method contributes significantly to improving data quality, where the LSTM model trained with MLBUI-RFR imputed data achieves higher accuracy compared to Mean Imputation. The evaluation results show a lower error rate, with an NRMSE of (15.68%) and a MAPE of (18.98%), whereas the Mean Imputation method yields an NRMSE of (16.06%) and a MAPE of (19.15%) proving that the proposed method is more effective in capturing non-linear patterns in the data. However, this study is based exclusively on data obtained from the Gorontalo Climatology Station in Gorontalo Province, Indonesia. The contribution of this study lies in evaluating the integration of MLBUI-RFR and LSTM for solar radiation forecasting, demonstrating how machine learning based univariate time series imputation can improve data quality and subsequently enhance forecasting performance on solar radiation data.