Day-Ahead Wind Power Forecasting Based On Deep Neural Network And Combined Dispatching Of Power System

Wind power has been rapidly developed in recent years by virtue of non-pollution,renewability and other characteristics.Wind power contributes to energy saving and emission reduction in the power system,but brings serious threats to safe operation of the power grid due to its uncertainty.To improve the controllability of wind power is of practical significance to safeguard stable operation of the power system.In order to solve the wind power uncertainty,this dissertation,according to actual dispatching management conditions of the power system,and studied the key techniques such as prediction method,mathematical model,solution strategy and algorithm,with wind power forecasting and combined dispatching of the wind integrated power system as the study objects,mainly including:This dissertation analyzed the periodic characteristics of the wind farm output power and summarized its change rules.Regarding the hourly change characteristics,this dissertation established two day-ahead wind power forecasting models with long short-term memory(LSTM)deep neural network structure,and analyzed the relationship between the quantity of input variables and forecasting errors of the forecasting model.The calculation results showed that the model had good generalization and stability,and its forecasting accuracy was obviously improved compared with the BP neural network model,Elman neural network model and SVM model.Considering that the single forecasting mode errors sharply increased with the forecasting step length,this dissertation proposed the day-ahead wind power forecasting model based on variational mode decomposition(VMD)and LSTM deep neural network.This model decomposed the characteristics of the wind power sequence by using VMD technique,and studied the characteristics based on the deep neural network.Regarding the selection of the VMD number,this paper proposed the method to determine the decomposition number by setting the "acceptable limits for residual errors" based on the effects of VMD residual error components on neural network learning effects.The calculation results showed that the model effectively controlled the error rate and improved the wind power forecasting accuracy.Regarding the uncertainty of wind power forecasting errors,this dissertation proposed a error uncertainty modeling method based on component-dependent multivariate normal distribution.This dissertation studied the functional relationship between forecasting step length and forecasting error,and stated that the prediction errors had good diffusion characteristics.Regarding the forecasting error distribution,this dissertation verified the one-dimensional error sequence through fitting based on normal distribution and Laplace distribution,and the results showed that normal distribution presented stronger applicability than Laplace distribution.This dissertation verified the multi-dimensional error sequence through fitting based on component-independent multivariate normal distribution and component-dependent multivariate normal distribution,and the results showed that the forecasting error sequences had correlations,and the unification between theoretical frequency and actual frequency of the forecasting error theory was achieved based on component-dependent multivariate normal distribution.Regarding invalid reserve capacity and redundant reserve capacity for the dispatching of the power grid,this dissertation proposed and established the power system combined dispatching model based on wind farm scenario envelope intervals.Regarding the wind power reserve capacity redundancy caused by broad scenario envelope intervals,the scenario envelope intervals were reasonably adjusted by the scenario reduction method.Regarding the solution of the combined dispatching model,this dissertation proposed the three-stage method:the wind farm envelope interval was determined by the scenario clustering method at the first stage;the start and stop status of hydraulic power and wind power units was determined by the dynamic programming method and priority list method at the second stage;the optimal output of the units was determined by the wind-driven algorism at the third stage.The calculation results showed that the system dispatching and operation costs were saved under the premise of guaranteeing that wind power uncertainty was controllable.The wind power prediction model and combined dispatching model of the wind integrated power system based on this were of positive significance to promote and improve wind power grid-connected operation.