Research On Power Load Forecasting Method Based On Weighted Combination And Kernel Density Estimation

Load forecasting is critical to power systems,it is a research hotspot in the process of intelligent construction of power grids.Short-term load forecasting can provide an important reference for power companies to formulate short-term power generation plans,unit start-stop,energy dispatching and other tasks.Medium term load forecasting is of great significance to the purchase of raw materials for power generation and the signing of energy supply contracts.Nowadays,accurate load forecasting technology is one of the key technologies to ensure the normal and stable operation of power grid.Through the research and analysis of the predecessors in the field of load forecasting,we know that machine learning method is the mainstream method in this field.In this thesis,the whole point load and meteorological data in New England are used as experimental simulation data.The load characteristics and influencing factors in this area are analyzed in detail,and two prediction models are proposed.A combined model for forecasting electricity consumption at every hour in the next week;An interval model for forecasting electricity consumption at every hour of the next four years.Firstly,due to the short period of short-term load forecasting,we selected training samples with moderate time span.The random forest and BP networks were used to predict the short-term electricity consumption in the region.Due to the large difference in the number of working days and rest days in the sample,and the sampling mode of random forest is Bootstrap.This leads to poor predictive effect of random forests on rest days,but BP neural network can fit the complex relationship between predicted value and input characteristics through repeated training.Next,the prediction results of the two models are combined by means of the reciprocal square error ratio method.The experimental results show that the combined model can integrate the advantages of two single models,the prediction result is more accurate,the maximum error is smaller,the MAPE of the combined model is 2.282%,and the number of samples with percentage error exceeding 6% is 4.This shows that the combined model is more applicable.Secondly,we established a point prediction model based on long and short-term memory neural networks for forecasting electricity consumption at every hour in the next four years.Because of the long time span,there are always some deviations between the predicted value and the actual value.Interval prediction model can provide more uncertain reference information,considering the difference in the error distribution of different size loads,a piecewise error statistics method based on kernel density estimation is proposed in this thesis.First,dividing the interval by the size of the load value,calculating the error probability density curve of each interval separately.Finally,the confidence intervals of errors under different confidence levels are calculated.Based the confidence intervals of errors,the confidence interval of load forecasting value is transformed.The experimental results show that the interval coverage of the interval prediction model under 80% confidence is 76.5%,this verifies the validity of the model.