Short-term Load Forecasting Based On An Integrated Model Of Holistic And Discrete Learning

Short-term load forecasting is an important work of power distribution system.Accurate forecasting method plays a key role in stable operation of power distribution system and low energy consumption.At present,most of the researches on short-term load forecasting at home and abroad focus on optimizing the learning performance of algorithms.A variety of serial parallel integration algorithms and parameter optimization methods are proposed,and excellent forecasting results are achieved.However,in the process of seeking higher forecasting accuracy,the problem of poor adaptability of forecasting model is ignored.Therefore,it is of great significance to explore the model construction approach of bi-directional optimization of accuracy and adaptability for short-term load forecasting.In order to comprehensively improve the accuracy and adaptability of short-term load forecasting model,this paper proposes an integrated forecasting method which can fully reflect the inherent characteristics of each power system load.Firstly,the inherent characteristics of load are deeply analyzed in this paper.Aiming at the problem of nesting between load recent trend correlation and external factors cause fluctuation,holistic and discrete processing methods for two characteristics are proposed to model respectively.And the two models run side by side independently to fully learn the relationship between the two characteristics and load.The variational mode decomposition(VMD)algorithm is used to decompose the recent low frequency trend components as training samples,and the long-short term memory(LSTM)network is used to predict the recent load trend of the system and build the holistic forecasting model.The method of practical experience combined with mutual information is used to reduce the dimension of external factor data,and the similar external factor information are extracted as training samples by weighted center clustering algorithm.Then,the e Xtreme gradient boosting(XGBoost)algorithm is used to perceive the load fluctuation caused by external factors,and build the discrete forecasting model.Then,the holistic and discrete integrated learning model is constructed based on the improved Stacking algorithm,and Elman network is used as the meta learner to fit the weight of the holistic and discrete model,to realize the complementary advantages of the holistic and discrete model.Finally,an example of a certain area in eastern China is used to verify the effectiveness of the proposed model.The simulation results show that the improved Stacking algorithm can fully integrate the advantages of the holistic and discrete model,and has strong small-sample learning ability and special-load adaptability.The holistic and discrete integrated forecasting method can comprehensively improve the prediction accuracy for short-term load and the adaptability for each power system.Then the stability of the distribution system is strengthened,and the energy efficiency is improved.The paper has 32 figures,4 tables and 70 references.