A Short-term Load Forecasting Method Combining Multi-algorithm&Multi-model And Online Second Learning

With the rapid and stable development of the national economy,power system load presents a significant growth year after year.However,the traditional extensive scheduling mode of power causes the problem that the shortage of electricity when it is busy and enormous waste when it is leisure because the energy storage capacity is extremely small and costly.It becomes the common goal of the world that builds a smart grid with the characterized:efficient,energy-saving,security and clean.Power load forecasting has always been a key component of smart grids.The ability to realize accurate and effective short-term load forecasting can reduce costs,implement optimal scheduling management,and ensure security with early warnings in smart grids.In this thesis,we newly propose a short-term load forecasting method combining multi-algorithm&multi-model and online second learning.First,we use the mutual information and statistical information to select the input variables and construct dataset.Then,we generate multiple training sets by performing diversity sampling with bootstrap on the original training set.We obtain multiple models using different artificial intelligence and machine-learning algorithms.Finally,we improve the offline second-learning method.We use the original dataset,the actual load,and the multi-model forecasts for recent period within the forecasted time to generate a new training set,which is trained by online second learning to obtain the final forecasting results.We applied the proposed method to the daily total load and daily peak load forecast of the power system city.We explored dataset selection,the generation of multiple training sets,the construction of multiple heterogeneous models,and online second learning by performing experiments and comparative studies.We study the daily total load and daily peak load for the past three years in Guangzhou,Guangdong Province,China.Compared to the optimal single-model,single-algorithm multi-model and multi-algorithm single-model,the results show that MAPE was reduced by 22.61%,10.98%,and 13.10%,respectively,in the daily total load forecasting,and by 15.79%,7.69%,and 11.77%,respectively,in the daily peak load forecasting.The results demonstrate that the proposed method is better for tracing the source and revealing the inner regularity and trends of load variation with the cross-border multi-source data.In the sampling space,the proposed method effectively overcomes the shortcomings of over-fitting and the limited generalization capability of a single-model by generating multiple training sets with bootstrap.In the construction of forecasting model,the proposed method uses the complementary advantages of multiple algorithms to solve the problem of the limited application of a single algorithm.However,the forecasting performance may not improve with the number of algorithms,and there is a need to understand the positive and negative fusion effect of the algorithm and data characteristics.In the construction of decision model,the proposed method reduces the forecasting errors caused by the load growth rate and the recency effect by online second learning.