| With the transformation of energy consumption,the proportion of new energy in the total energy consumption is gradually rising,and the new energy power represented by wind power and photovoltaic power is developing rapidly.However,due to the influence of meteorological factors and diurnal cycle,wind power generation has strong volatility and uncontrollability.When a high proportion of wind power is connected to the power system,it will pose a great threat to the security and stability of the power grid,and increase the difficulty of dispatching and operation.Therefore,the accurate prediction of wind power has become the focus of the research.After the data analysis of wind power data,this paper improves the Transformer structure and proposes a wind power prediction architecture based on Hodrick-Prescott(HP)filter and sparse self-attention(Hodrick-Prescot-t Filter Prob-Sparse Self-attention Network,HPPSAN),which improves the wind power prediction accuracy.In addition,considering the spatial correlation of wind farms,this paper improves the model based on HPPSAN using graph convolutional neural network to further improve the wind power prediction accuracy.The specific work and innovations are as follows.First,there are multiple peaks and valleys in wind power within a day,and the difference between days is large,and there is a large volatility and uncertainty,but the calculation of its autocorrelation coefficient shows that wind power has autocorrelation,that is,there is a period term that can be decomposed.In addition,wind power is affected by a variety of factors together,and they show a complex and non-linear relationship with power,and the impact on wind power is changing over a period of time,thus affecting the fluctuation of wind power.Based on the above analysis,this paper proposes an encoderdecoder structure model HPPSAN based on HP filtering and sparse self-attentiveness for wind power prediction,which serves as a unified framework for fusing feature decomposition and prediction modeling to further optimize the decomposition results through end-to-end training,and the internal modules of the model make full use of the input features to obtain more generalized wind power features to improve The model prediction accuracy is improved.Experimental results on a publicly available wind farm dataset show that the model predicts an average NRMSE of 20.8%lower wind power compared to other deep learning algorithms.The proposed HP decomposition module reduces NRMSE by 10.4%compared to the model without this module,and the proposed decomposition-reconstruction module reduces NRMSE by 3.5%on average compared to the model without this module.Second,considering the spatial correlation of wind farms in the same region,an improved HPPSAN model based on graph convolutional neural network is proposed in this paper to construct spatially-aware and semantically-aware graphs,and further improve the model prediction accuracy by fusing the spatial information of wind power data through graph convolutional neural network.Experimental results on publicly available wind farm datasets show that the average MSE of this model is reduced by 21.8%compared with other algorithms.The above is the main work of this paper.We hope that the work of this paper can provide useful and effective theoretical reference and data basis for many businesses such as grid dispatching,maintenance planning,and new energy consumption analysis. |
