Research on predicting solar thermal systems has primarily utilized artificial neural networks (ANNs), which heavily rely on numerical weather prediction (NWP) data. However, ANN models struggle to capture temporal dependencies in time series data, and discrepancies between NWP data and actual observations can reduce prediction accuracy. This study proposes a long short-term memory (LSTM)-based prediction model that integrates NWP, numerical weather observation (NWO), and numerical system observation (NSO) data. The model predicts three output variables and was developed into three configurations: LSTM 1, LSTM 2, and LSTM 3, each based on different input setups. LSTM 3 outperformed LSTM 1 in predicting supplied energy, reducing the mean absolute error (MAE) and root mean square error (RMSE) by 28.2% and 25.3%, respectively, while improving the R² value by 34.5%. For predicting the top storage tank temperature, LSTM 3 also achieved reductions in MAE and RMSE of 27.3% and 23.5%, and improved R² by 12.2%. In terms of the prediction of acquired energy, solar irradiation was the most significant variable, with an importance score of 0.068, followed closely by supplied energy at 0.067 and top storage tank temperature at 0.057. These results indicate that the combined influence of environmental and system variables is crucial for enhancing prediction accuracy.