Enhancing grid dispatch reliability using hybrid CNN-LSTM solar power forecasting

Short-term solar power forecasting (STSPF) is critical for the integration of renewable energy into smart grids, yet its accuracy is often compromised by the stochastic nature of meteorological variables. This study proposes a hybrid deep learning framework that integrates 1D-Convolutional Neural Networks (CNN) with Long Short-Term Memory (LSTM) networks to capture both spatiotemporal features and long-term temporal dependencies. Using a multivariate dataset comprising DC power, irradiation, and temperature, the model was optimized via a sliding window approach with a 6-hour look-back period (24 steps at 15-minute resolutions). Statistical descriptive baseline validation using the Augmented Dickey-Fuller (ADF) test provided a preliminary check of target series stationarity over short horizons ($p < 0.05$). To ensure robust evaluation, the proposed framework was benchmarked against an expanded suite of baseline models, including classical statistical frameworks (ARIMA), machine learning models (SVR, Random Forest), and deep recurrent architectures (standalone GRU, standalone LSTM, and Bidirectional LSTM). The proposed Hybrid CNN-LSTM model achieved a Coefficient of Determination ($R^2$) of 0.9181, and a Mean Absolute Percentage Error (MAPE) of 9.84\%, outperforming all evaluated baseline models. Furthermore, a one-step-ahead Diebold-Mariano test using squared error loss confirmed the statistical significance of these improvements over the standalone LSTM ($p = 1.89 \times 10^{-10}$). With a low computational inference speed of 9.55 ms per sample, the framework demonstrates potential viability for real-time grid management and energy dispatch pilot testing.