| Power load forecasting is the premise of reliable,economical and stable operation of power system.With the development of society,people’s demand for the accuracy and stability of power load forecasting is higher and higher.Therefore,it is of great significance to study the problem of power load forecasting.Aiming at the problem of regional short-term power load forecasting,this paper presents a power load variable weight forecasting method based on multi-model variable weight synthesis under the framework of integrated learning based on the analysis of power load distribution characteristics and its influencing factors.Specific contents include:Firstly,the distribution characteristics of power load and its main influencing factors are analyzed.Based on historical data,the temporal distribution characteristics of power load and the influence of external factors such as weather,date and price on power load are analyzed from daily,weekly,quarterly and annual scales.The analysis results show that the regional power load data has the characteristics of random fluctuation,multiple quasi-periodicity,high autocorrelation,easy to be affected by various factors such as weather,and significant uncertainty.The effects of temperature,date and electricity price on power load are complex and changeable,and it is difficult to quantify accurately.Secondly,an improved gated cyclic neural network based on attention mechanism is proposed for power load prediction.In order to solve the problem that GRUs will lose sequence information when processing structure information between data when the input load sequence is long,an attention-GRUs prediction model which can highlight key features more and is not affected by sequence length is established by introducing Attention mechanism.The feasibility and superiority of the method are verified by simulation experiments.Furthermore,a multi-model integrated power load forecasting method is presented based on the Stacking integrated learning framework.Due to the common problems of a single power load forecasting model,such as low forecasting accuracy,poor generalization ability,and difficult forecasting stability,the integrated learning framework of Stacking is improved and multiple models are integrated to improve the forecasting accuracy and forecasting stability.The Attention-GRU,SVR,RF and TCN models are selected as the base learning device for the training and prediction based on the comprehensive consideration of the characteristics,advantages and differences of various common power load forecasting methods.At the second layer,the unique metaphores in the traditional stacking integrated learning method are selected and determined and multiple different metaphores are used for forecasting simultaneously.A third layer is constructed to integrate the prediction results of multiple models by weighted average.The simulation results show that the proposed multi-model integrated forecasting method of electric load based on Stacking integrated learning can further reduce the forecasting errors.Finally,to further improve the forecasting performance of the model,a power load forecasting method that improves the Stacking integrated learning based on variable weights is proposed.This method designs con stant weights that reflect the relative importance of the models based on model prediction accuracy,designs status weights based on mutual information between evaluation indicators,and constructs variable weights through the combination of constant weights and status weights,which are applied to the third layer of Stacking integrated learning.Variable weights are applied to the output data of the second layer to determine the final prediction results.After experimental verification and analysis,the results show that the power load variable weight comprehensive prediction model proposed in this paper based on Stacking integrated learning can make full use of the advantages of different model structures and significantly improve the generalization and stability of the model. |
PDF Full Text Request