An Enhanced Time-Series-Based Solar Power Forecast for Microgrid System

The world is moving towards net-zero emissions. Forecasting solar power with precision is a key enabler for the efficient integration of renewable energy into the grid for smarter grid management, and hence optimizing the balance between supply and demand. This paper develops and rigorously evaluates five forecasting models, namely standard LSTM, attention-enhanced custom LSTM, ARIMA, rolling window ARIMA, and Prophet. All models are implemented on the same test bed using SARAH satellite data for training and ERA5 reanalysis data for validation, ensuring a controlled and reproducible benchmark. Although rolling window ARIMA is a known strategy for capturing evolving time series patterns, its application to cross-domain solar forecasting and its direct comparison with modern deep learning models under unified conditions remain limited. The analysis in this paper indicates that among all the evaluated models, the ARIMA(1,1,1) rolling window, dynamically retrained using a 30-day sliding window, demonstrates the most consistent and superior performance, surpassing both conventional statistical models and deep learning approaches. It achieves the lowest RMSE (11.19 ± 1.34), MAE (8.92 ± 1.01), and MAPE (7.2% ± 0.8%), along with the highest R² value (0.93 ± 0.02), indicating strong generalization with minimal computational complexity. In contrast to static ARIMA, diagnostic tests validate that the residuals of the rolling model are not correlated (Ljung-Box p = 0.53), nearly normal (Jarque-Bera p = 0.96; kurtosis = 3.89) and symmetric (skewness = -0.01). A comparative evaluation of rolling window ARIMA and attention-based LSTM models demonstrates forecast accuracy improvements of 71.9% and 50.9%, respectively, with results reported under 95% confidence intervals. These results support the conclusion that the rolling window ARIMA framework offers a statistically reliable and operationally efficient approach, making it an apt model for enhancing time series-based solar power forecasting in microgrid systems.