A novel hybrid model integrating CEEMDAN decomposition, dispersion entropy and LSTM for photovoltaic power forecasting and anomaly detection

Photovoltaic (PV) power generation exhibits significant non-stationary characteristics due to the influence of meteorological conditions and equipment status, which makes traditional prediction methods difficult to accurately capture its dynamic variations and abnormal behaviors. To address these limitations, a CEEMDAN-DispEn-LSTM hybrid framework is proposed for PV power forecasting and anomaly detection. Following preprocessing via the Median Absolute Deviation (MAD) method and decomposition using the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), optimal components are selected in this study through a dual-criterion approach that concurrently accounts for energy proportion and correlation coefficient. Dispersion Entropy (DispEn) is employed to quantify signal complexity, while dedicated Long Short-Term Memory (LSTM) subnetworks integrated with entropy weighting are utilized to dynamically achieve multi-scale feature fusion. Furthermore, dual deviation logic is adopted to detect non-meteorological anomalies. Experimental results confirm that the proposed framework outperforms selected benchmark models across most prediction metrics. In anomaly detection, the framework demonstrates significant effectiveness in identifying line faults and PID effects, while exhibiting preliminary capability in detecting partial shading. The latter finding points to a clear direction for future performance enhancement through multi-source data fusion. Thus, this study establishes a validated technical pathway for non-stationary time series forecasting, particularly suited for ultra-short-term power prediction and anomaly detection in distributed photovoltaic systems under temperate climates, highlighting its application potential in the operation and maintenance of such systems.