| China will take forceful measures to achieve the goal of peaking its carbon emissions by2030,reaching carbon neutrality by 2060 through the large-scale development and utilization of new energy systems with PV and wind power as the main sources.As the PV generation is greatly affected by the weather,it is highly unstable.Meanwhile,with the continuous growth of PV installations capacity,the security and stability of large-scale PV power grid operations are greatly challenged.As one of the important approaches to maintain the stable operation of the power grid,accurate PV power forecast is of great significance to effectively reduce unreasonable phenomena such as some PV installations being abandoned,safe and economic operation of the power grid and reliable dispatching.Because the PV output is closely related to weather conditions,the PV features volatility and intermittent,to effectively reduce the volatility of its time sequence,most of the existing research results are only for PV power sequence mode decomposition,and the impact of the same volatility of the weather feature sequence on the forecasting results has not been considered.This will cause a large error in the prediction results.To solve the above-mentioned problems,the paper proposes a PV forecasting model based on singularity spectrum analysis and Stacking ensemble learning under various weather features.The weather feature is decomposed into multiple sub-sequences by singularity spectrum analysis.And the sub-sequences act as the input variable sequence of the base learner.Similarly,PV power is also decomposed into multiple sub-sequences as outpurt variable sequences of the base learner,which can improve the stability of the forecasting curve of the first-level forecasting model.The weather is divided into three types: clear day,overcast day as well as rainy day,and different forecast models are built according to diverse weather types.In the case of the same weather type,different single model prediction has the difference,combined with Stacking ensemble learning has the advantage of complementary advantages for the single model with a large difference.To build a Stacking ensemble learning framework with Recurrent Neural Network(RNN),Long Short-Term Memory(LSTM),Gated Recurrent Unit(GRU)as a base learner and using its prediction results as the input feature of the meta-learner.Stacking ensemble learning framework is used to construct short-term PV power forecast models for different weather types.Finally,the practical data of the Alice Springs photovoltaic power station in Australia are utilized to hold up case studies to verify the effectiveness of the proposed method. |
