Generation Power Prediction Method Based On Stacking Ensemble And Hyperparameter Optimization

With the acceleration of industrialization,the types of power plants have become more abundant,environmental problems have gradually emerged,and the use of new energy sources such as wind,solar and nuclear power,and the recycling of industrial wastewater and domestic sewage need to be solved.In the power system,the plan is to schedule the generation capacity of the gensets for the next few days according to the load forecast.In addition to meeting the technical constraints,the plan should also ensure that the system has sufficient spinning reserve capacity.Uncertainty of environmental factors will bring instability of power output.Effective prediction of power generation is conducive to the integration and dispatch of power,reducing the cost of spinning reserve capacity and realizing rational utilization of energy.Due to the instability of variable conditions,the difficulty of power prediction is increased.The performance of traditional single algorithm may be better in a stable environment,but not ideal in a complex environment,and each single algorithm has different advantages and disadvantages.Compared with a single model,the ensemble model can build a power prediction model suitable for different scenarios after mining the useful information of a single algorithm,which can reduce the occurrence of great errors,so it can be used as a power prediction method under multivariable conditions in different types of power plants.This study proposes to optimize the hyperparameters of the sub-model on the basis of the stacking ensemble prediction model,and to add signal processing technology for unstable environmental variables to improve the ensemble prediction effect.In view of the complex correlation among the data of multiple environmental variables during the operation of combined cycle power plant units,this study uses several single models with large differences as sub-models,and uses GS optimization method to select the optimal combination of hyperparameters for the sub-models.Taking 9568 recorded data of a combined cycle power plant as an example,a stacking-GS prediction model is established after preprocessing.The results show that the proposed method can effectively reduce the error of the combined cycle power plant power generation prediction compared with the prediction method of the traditional machine learning model.For wind power plants,due to the intermittency and randomness of wind,the existing power generation prediction models cannot well capture the actual dynamic change trend of wind power plants,resulting in low prediction accuracy.In order to solve this problem,a novel stacking ensemble method is proposed,which takes the NWP meteorological data of the wind power plants and the rolling historical power data decomposed by DWT as the input of the model.The stacking ensemble is used to fuse multiple LSTM models,and the GA is used for hyperparameter optimization.Multi-step forecasting of wind power using the actual power data of a wind power plant in a time series in Northwest China.The results show that the prediction performance of the proposed ensemble model is better than the stacking ensemble model with RNN,DNN and SVM as sub-models and other traditional single model and ensemble model.Therefore,this method can effectively improve the multi-step prediction performance of wind power.