To address the challenge of limited photovoltaic (PV) power forecasting accuracy, which is primarily attributed to the significant impacts of abrupt weather changes and the strong non-stationarity of PV power time series, this paper proposes a multi-scale PV power forecasting model based on modified Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) and a hybrid neural network. First, key meteorological features including solar irradiance and ambient temperature are screened via the Pearson correlation coefficient (PCC), and the K-means clustering algorithm is adopted to construct three weather scenario datasets for sunny, cloudy, and rainy days, which effectively mitigates cross-scenario data distribution discrepancy. Second, the noise standard deviation and number of decomposition layers of the ICEEMDAN are dynamically optimized using the Dream Optimization Algorithm (DOA), achieving optimal modal decomposition and stationarization reconstruction of PV time series features. Subsequently, the Long Short-Term Memory (LSTM) network is utilized to deeply extract the periodic and trend characteristics embedded in the time series, which is combined with the multi-head attention mechanism from the Transformer architecture to effectively capture dynamic correlation information in the global time dimension. Finally, extensive experimental results demonstrate that the proposed PV forecasting method exhibits significant outperformance in both computational efficiency and forecasting accuracy under various weather conditions compared with state-of-the-art methods.
Open Access
Journal Metrics
Journal Overview
Associated Journals
