A Non-Reliant Meteorological Forecast Method for Ultra-Short-Term Photovoltaic Power Prediction

Introduction: Accurate short-term photovoltaic (PV) power prediction is crucial for ensuring the safety of the power grid and promoting the consumption of PV. However, due to factors such as Internet access conditions and information security control, some industrial parks are unable to obtain real-time weather forecast information, and there is an urgent need to develop a PV power prediction method that does not rely on weather forecast information. Method: In this paper, we propose an innovative architecture for the Morphological Filtering Empirical Wavelet Transform (MFEWT) and the Frequency Decoupled Multi-Cycle Fusion Network (FDMFNet). Firstly, MFEWT is used to decompose the historical photovoltaic (PV) power generation into a number of frequency sub-sequences, which are then classified into major and minor components based on their frequency energies. For the primary component, amplitude features are extracted using the Hilbert transform and supplemented with feature columns using the Signal Cycle- Based Adaptive Historical Feature Set Construction Method (CAF), which are then predicted by a convolutional neural network. For the secondary component, the FDMFNet is designed to accurately capture the trend of the high-frequency secondary components through frequency-domain decoupling. Finally, the major and minor components are linearly superimposed to obtain the final PV power prediction. Result: Experimental validation based on photovoltaic (PV) datasets from two industrial parks in China, Wuhan and Qingdao, confirms the effectiveness of the proposed method. In terms of overall prediction accuracy, the MAE of the proposed method is 0.2108, the RMSE is 0.3297, and the R2 is 0.8708. Discussion: The MFEWT proposed in this study effectively suppresses mode mixing. Combined with Hilbert Transform and FDMFNet, it achieves multi-scale feature mining of PV power. This fusion mechanism significantly enhances the model's capability to characterize the intrinsic timefrequency properties of PV sequences. Conclusion: The PV prediction method introduced in this paper, which operates without weather forecast information, demonstrates favorable accuracy and stability in actual ultra-short-term PV power prediction scenarios. It provides a novel solution for PV power forecasting under conditions where meteorological information is lacking.