Transformer-based Photovoltaic Power Generation Prediction Method For Long Sequences

Photovoltaic power generation refers to the process of converting solar energy into electric energy by using the photovoltaic effect.It has the advantages of clean,safety,renewable and distributed,and is one of the important ways to realize the "double carbon" strategic goal of our country.Photovoltaic power generation forecasting refers to predicting the power generation of photovoltaic power stations in a future period of time based on historical data and meteorological information.The accurate prediction of photovoltaic power generation is of great significance for improving the operation efficiency of photovoltaic power stations,reducing the risk of grid connection,and optimizing the scheduling strategy.Since grid dispatching requires the deployment of photovoltaic power generation for a long period of time in the future,the demand for prediction length is usually high and long sequence prediction is required,and successfully achieving accurate prediction of long sequence requires that the model can effectively capture the longterm dependence between input and output.Recent work has shown that transformerbased methods have the potential to capture long-term dependencies and thus improve prediction performance.However,there are some problems in current transformerbased methods,which are difficult to be directly applied to photovoltaic power generation prediction.These problems include: 1)Transformer-based models have to use a sparse form of attention mechanism to deal with quadratic complexity,resulting in an information bottleneck;2)The Post-LN structure in the existing model has the situation that the deep network gradient disappears;3)Existing transformer-based models ignore the impact of autocorrelation errors and distribution shifts on prediction.To solve the above problems,this paper establishes EW-Infomer for the first and second problems,and proposes Ad＿Sym Norm optimization algorithm for the third problem.The main contributions of this paper are as follows:1)Aiming at the problem of information bottleneck and gradient disappearance,this paper establishes the EW-Informer model from the model level,which is: i)a Prob Sparse self-attention mechanism based on Wasserstein distance is proposed to solve the problem of information loss caused by KL divergence when two distributions do not overlap.ii)Chunk-Max-Pooling is used for self-attention distillation,so as to retain more sequence information while reducing space complexity.iii)The residual structure is modified to alleviate the gradient disappearance phenomenon caused by the Post-LN structure.2)Aiming at the problems of autocorrelation error and distribution shift,we propose the Ad＿Sym Norm optimization algorithm from the data level.The autocorrelation error coefficient is put into the model as a model parameter for learning,and the influence of second-order autocorrelation error on model prediction is eliminated.Symmetric instance normalization is used to remove the unstable factors in the training data,so as to solve the problem of data distribution changing with time.3)The model and optimization algorithm proposed in this paper are experimentally verified by using multiple photovoltaic power generation datasets,and compared with other photovoltaic power generation prediction models to verify the effectiveness of the model and optimization algorithm proposed in this paper.