A Data-Driven Approach for Regional PV Power Plant Cluster Output Prediction and Uncertainty Analysis

To improve the accuracy of Photovoltaic (PV) power generation forecasting and ensure the safe scheduling and stable operation of power systems, we propose a combined forecasting method based on Variational Mode Decomposition (VMD) and the Rime Optimization Algorithm (RIME) to optimize an attention mechanism-based Convolutional Bidirectional Long Short-Term Memory Neural Network (CNN-BiLSTM-AT). Addressing the issues of nonlinearity and non-stationarity in the time series signals affecting PV power generation, we select inclined irradiance, horizontal irradiance, temperature, and humidity as feature input variables. The VMD algorithm is employed to decompose these variables into several distinct modes, with each component modeled separately. Furthermore, the RIME optimization algorithm is utilized to optimize the hyperparameters of the CNN-BiLSTM-AT algorithm, effectively resolving uncertainties in parameter selection and slow convergence rates. Finally, we conduct a case analysis using measured data from distributed PV sites in Australia. The results demonstrate that compared to traditional neural network forecasting models and fixed-parameter combined forecasting models, the VMD-RIME-CNN-BiLSTM-AT model improves the prediction errors for 72 hours, 24 hours, and 4 hours by at least 0.83%, 1.11<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">%</sup>, and 1.41%, respectively. This method exhibits higher predictive accuracy and stability.