Dynamic Multi-Scale Convolution and BiTCN Enhanced by CPO and Attention for Accurate Ultra-Short-Term PV Power Forecasting

Due to the complexity of photovoltaic (PV) systems and their susceptibility to diverse meteorological conditions and nonlinear dynamic characteristics, PV power forecasting poses significant challenges. To address these issues, this study proposes a hybrid forecasting framework that integrates Individually Fair K-means (IF-Kmeans), Crown Porcupine Optimization (CPO), Dynamic Multi-Scale Convolution (DMSC), Bidirectional Temporal Convolutional Networks (BiTCN), and the Attention mechanism for high-precision PV power forecasting. IF-Kmeans optimizes data allocation with fairness constraints, ensuring more reasonable feature grouping and reducing model complexity. CPO dynamically adjusts model parameters to enhance adaptability under complex weather conditions. The DMSC module captures both local and global features of time-series data through multi-scale convolutions, further improving feature representation. The BiTCN module comprehensively models bidirectional temporal dependencies by combining forward and backward dynamics in time-series data, while the Attention mechanism assigns higher weights to critical features, enhancing overall model performance. Experimental results demonstrate that the proposed CPO-DMSC-BiTCN-Attention model achieves outstanding predictive accuracy under various weather conditions, with average MAPE, RMSE, and R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> values of 0.643, 2.088, and 0.861, respectively, significantly outperforming baseline models. This framework provides a robust and accurate solution for PV power forecasting.