A novel PV Power Fusion Prediction with CEEMDAN and Bayesian optimized LSTM

In the context of renewable energy systems, accurate prediction of solar energy output is crucial for mitigating the challenges posed by intermittency and volatility inherent in photovoltaic (PV) power generation. This study introduces a novel PV power forecasting model leveraging the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) technique. The model integrates multi-dimensional feature sets—including historical PV power data length, climatic variables, and solar irradiance—to decompose the original time series into multiple intrinsic mode functions (IMFs). Subsequently, Pearson correlation analysis is employed to evaluate the relationships among these components, guiding their grouping and aggregation. The resulting Co-IMFs are individually input into a Bayesian-optimized Long Short-Term Memory (BO_LSTM) network for component-level PV power prediction. The individual prediction outcomes are then re-incorporated as distinct features within the BO_LSTM framework to achieve a fused, comprehensive forecast of PV power output. Experimental evaluations demonstrate that the proposed approach effectively captures detailed prediction dynamics for each IMF component, with the fused predictions exhibiting robust performance and high accuracy over a forecast horizon of up to four days.