Abstract: Predicting how much power solar panels will produce is a major hurdle for stable energy grids, largely because solar output is notoriously unpredictable, nonlinear, and constantly shifting. To tackle this, we’ve developed a hybrid deep learning framework designed for more accurate short-term forecasting. Our approach uses "CEEMDAN" to break down complex solar data into manageable components, paired with Iterative Filtering (IF) to pull out the most meaningful hidden patterns. These refined insights are then processed through a Bidirectional Long Short-Term Memory ("BiLSTM" ) network. By combining these tools, our model captures the deep, time-sensitive relationships in solar generation that traditional methods often miss. Each component is then modeled using "BiLSTM" networks to capture bidirectional temporal dependencies, and the final forecast is obtained through signal reconstruction. The proposed model is evaluated using real-world PV datasets collected from Tindouf (Algeria), and its performance is compared with benchmark deep learning and decomposition-based approaches. Testing proves that this framework markedly boosts both the accuracy and reliability of solar power predictions. With an R^2 of 99.06 %, an "RMSE" of 158.20 "kW" , and an "nRMSE" of 6.15%, the model consistently leads the way over standard LSTM models and other hybrid setups. These outcomes validate the idea that combining "CEEMDAN" decomposition, precise time-frequency feature extraction, and bidirectional deep learning creates a powerful, dependable solution for short-term forecasting. By blending these specific techniques, the system manages to capture the nuance of solar data more effectively than previous methods.

Keywords: Photovoltaic Power Forecasting, "CEEMDAN", Instantaneous Frequency, "BiLSTM", Hybrid Deep Learning, Time-Series Decomposition

Pages: 519-528