Deep Time Series Neural Network Based Ensemble Method And Application In Energy Demand Forecasting

Demand load control plays an important role in the power demand side management,while demand load forecasting with high accuracy will directly affect the reliability of power demand control,which is a guarantee to keep the balance between power supply and demand to improve the power grid stability.However,as for large industrial power users,due to equipment operating conditions and process switching,their ultra-short-term power consumption usually has strong randomness and volatility,making accurate prediction of ultra-short-term demand load become a difficult problem.Aiming at this problem,this paper uses multivariate correlation analysis,deep temporal neural network theory and ensemble learning method etc.methods to study the ultra-short-term industrial power demand prediction,and proposes a hybrid integrated learning prediction model based on deep temporal neural network.Among of them,in order to effectively analyze time series load data,the classical long short-term memory(LSTM)network is adopted as the base learner of the proposed model,and multivariate correlation analysis is performed to determine the related input attributes from multi-source heterogeneous raw data.At the same time,according to the principle of bias-variance decomposition,this thesis proposes a hybrid ensemble strategy,which combines multiple ensemble learning methods,including Bagging,Random subspace and Boosting with ensemble pruning to solve the deviation of single model and local sample,improve the generalization ability of the model.At last,the data set is structured based on collected real data to carry out the ultra-short-term demand load forecasting experiment.It is analyzed and evaluated with a variety of advanced time series prediction models and classical integrated learning algorithms.The comparison experimental results show that,in the three commonly used predictive performance evaluation indicators of MAPE,MAE and NRMSE and the absolute percentage error(PAPE)corresponding to peak load,the proposed model can achieve stable optimal effect and predict peak demand load more accurately than other models.At the same time,the proposed model is applied to a certain steel plant to carry out multi-time-steps demand load forecasting,which can accurately pre-judge and alarm of the maximum demand excess,and effectively assist the reliable implementation of demand regulation.