Advanced Deep Learning Approach for Solar Radiation Prediction in Medan: CNN-LSTM Hybrid Model

Abstract

The availability of solar energy in Indonesia is hindered by small-scale variations in solar radiation that occur in tropical climates. For maintaining grid stability and for ensuring the efficient operation of solar power systems, accurate short-term prediction is necessary. In this study, a hybrid Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) model is developed in order to predict minute-scale solar radiation in Medan City. The model makes use of minute-resolution meteorological data (including temperature, humidity, dew point, and wind speed), combining CNNs for the extraction of spatial features with LSTMs for temporal learning. Four different architectural setups were evaluated, and Model 4 — consisting of four LSTM layers with 50, 100, 150, and 200 units, respectively — achieved the best results: an RMSE of 39.88 W/m², a MAE of 26.89 W/m², and an R² of 0.955, which represents a 20.82% improvement on the baseline models. The feature importance analysis found that temperature, dew point, and relative humidity were the most significant predictive factors. The model is able to capture both the daily cycles and the rapid fluctuations typical of tropical climates, thus providing accurate predictions that are important for the stability of the electricity grid and for the optimisation of solar power plants. The hybrid architecture that has been developed provides a transferable framework which can be applied to other tropical cities in Indonesia and thus contributes to the country's energy transition objectives and to the development of sustainable renewable energy. This research promotes advances in deep learning methods for forecasting renewable energy while tackling the real-world problems associated with integrating solar energy under changing atmospheric conditions.