Ultra-short-term power forecasting for photovoltaic cluster based on dynamic clustering of photovoltaic stations and optimized Ns-transformer combined model

The ultra-short-term photovoltaic power forecasting provide a basis for formulating generation plans for intraday energy scheduling. However, meteorological fluctuations and forecasted weather inaccuracies pose significant challenges for further improving the power forecasting accuracy. This paper presents an improved ensemble forecasting model that takes into account the meteorological similarities of photovoltaic stations to enhance the accuracy of ultra-short-term photovoltaic power forecasting. By incorporating the spatiotemporal correlation of photovoltaic cluster into the modeling process, this study proposes a method for dynamically identifying photovoltaic cluster states based on meteorological screening using improved clustering. To achieve dynamic cluster division, the feature channels of the traditional U-Net model are expanded and combined with an advanced generative denoising probabilistic diffusion model (DDPM), which addresses the issue of imbalanced fluctuating samples, thereby preventing misjudgment of photovoltaic station states. To capture the dependency patterns in the power evolution of photovoltaic cluster and the uncertainty of non-stationary processes, proposed a combination forecasting model incorporating bidirectional long short-term memory network (BILSTM) and non-stationary Transformer (Ns-Transformer) for photovoltaic power forecasting, and applied an advanced differentiated creative search (DCS) optimization algorithm to adjust the parameters of the ensemble model and enhance its learning ability and generalizability. The proposed method is applied to a photovoltaic cluster in Gansu, China. Quantitative results demonstrate that the proposed approach yields the lowest errors, with N RMSE and N MAE lower than those of the other methods by an average of 2.25 % and 1.38 %, respectively, while showing an average R² improvement of 3.90 %, this research not only provides critical technical support for photovoltaic grid integration and secure grid operation but can also be extended to photovoltaic clusters under diverse geographical and meteorological conditions. It contributes to enhancing renewable energy absorption rates and offers a practical forecasting solution for energy system transformation under global carbon neutrality and carbon peaking objectives.