Abstract: Stacking algorithm is good at alleviating over fitting problem in the prediction of photovoltaic power generation, but with drawbacks of long computation time and less sample data. To solve the problem, this paper proposes an improved 3-layer stacking algorithm based on new vector representation and cross validation accuracy weighting. The first and second layers are the primary layer, which use random forest, SVR and XGboost3. The third layer is the secondary layer, and uses LightGBM to learn the output of the second layer again to reduce noise. A new vector representation method is used to increase the sample size and sample distribution density of input and output data between levels to ensure that the data dimension will not increase with the increase of the number of primary level learners. At the same time, the results are weighted according to the difference in the prediction accuracy of different prediction models in the primary layer under cross-validation. Practical analysis is demonstrated by using the power generation data of a photovoltaic power station. Compared with random forest model and Stacking model, the prediction performance of the proposed model has been greatly improved in MAE, MSE and R-Squared.

Key words: Stacking algorithm, cross validation, vector representation, regression prediction algorithm, photovoltaic power generation forecast