To enhance the accuracy and stability of photovoltaic power forecasting under typical weather conditions, this paper proposes a forecasting model that integrates an improved spectral clustering method optimized by the NSGAⅡ multi-objective algorithm with a BiLSTM network enhanced by a multi-head attention mechanism （MHA）. Firstly, outlier detection and preprocessing are performed on meteorological and historical PV data, and key influencing features are identified. Then, the construction of the degree matrix in spectral clustering is improved using dynamic time warping （DTW）, and NSGAⅡ is employed to optimize the sparsity of the similarity matrix and the Gaussian kernel parameter, yielding an optimal clustering model that categorizes weather into sunny, cloudy, and rainy types. Finally, optimal NSGAⅡ-BiLSTM-MHA models are established for each weather type and compared with four baseline models. Results show that, under three weather conditions, the proposed model achieves 50.74%-62.95% lower RMSE and 55.85%-60.09% lower SDEX than that of SVR, while improving the R² by 8.99%-17.07%.

multiobjective optimization  /  photovoltaic power  /  prediction models  /  dynamic time warping  /  multi-head attention mechanism