Power Load Forecasting Based On Variable Weight Synthesis

Power load forecasting can help the power management departments to schedule the number of generators,coordinate the transmission of electric energy,rationally allocate regional power supply,plan the deploy of infrastructure and overhaul the power system.Accurate load forecasting can not only improve the scientific management level of power grid,but also improve the benefit of the power enterprise and the social benefit of power system.Therefore,improving the accuracy of forecasting is not merely an imminent strategic task of the electric power department,but also one of the committed studies for electric power workers.In the preface,this paper briefly introduces the research background and significance of load forecasting,the characteristics and basic requirements of load forecasting.Based on the current research situation of load forecasting,this paper probes into the existing theoretical methods and finds out that there is a shortcoming of common weight which are difficult to adapt to the complex load environment.Therefore conduct the study of variable weights on power load forecasting.The application of the variable weight method in load forecasting mainly embodies the following aspects:First,A pattern classification method for power load analysis is proposed based on Threshold and cloud improved fuzzy clustering(for short,TACIFCA).Firstly,the classic FCM clustering algorithm were improved by introducing a threshold to recognize the in-homogeneous datum and atypical homogeneous datum and of each cluster and reduce their affectation on the forming of cluster center.Then,cloud description of each cluster is given and the weights of each homogeneous data in same cluster were determined by the correlation coefficient which indicates the typical degree of the sample data for the cluster.The experimental result shows that the new method has better performance than traditional fuzzy c-means clustering.Second,a power load forecasting method based on variable weight combination model is presented on basis of load type grouping combined with RBF neural network.In this method,the RBF network training algorithm is improved by using the conjugate gradient descent method,which is designed to learn the typical load classes which are obtained from the load data handled by the TACIFCA algorithm,and single Prediction sub-model which adapts to each category group is set up.Then,a variable weight weighting method based on the similarity of load type is proposed.Finally,the total forecasting results of all unit load models are integrated adaptively by using variable weighted synthesis.Third,considering the similarity of the historical day and the forecast date and combining with the new machine learning method,an similar degree improved gradient boosting decision tree forecasting method is proposed by introducing the process of similar-data selection and synthetic similarity weighted loss function.In this method,the training data for certain day's power load forecasting was selected by means of synthetic similarity degree,which is the minimum of several local partial features based similarity index,such as weather factors,time factors and trend factors etc..Then,a similar day improved gradient boosting decision tree based forecasting model was set up by introducing synthetic similarity weighted loss function and using it on the selected training data.Finally,the simulation results Indicate that both FCM clustering&RBF network adaptive Combination model and similarity data selection&improved gradient decision tree show a significant improvement in the accuracy of the load forecast.