Research On SCUC Decision-making Approach Based On Deep Learning Method

The Security Constrained Unit Commitment(SCUC)problems are treated as the foundation of dispatching electric power will be facing a lot of challenges.The reason is that the large amount of electric technologies has been used in power system with the putting forward of the electric market reform in recent years.The traditional physical-model-based UC dispatching method seems uable to fit the highspeed developing requirement of the power system.There is a worldwide significant problem needs to solved is the physical-model-based UC dispatching method cannot face the latest challenges.In this work,the author is starting with the conceptions of data-driven,that is using the deep learning model to learn the mapping relations between historical data.In this approach,it is directly building the mapping model which will be used for decision-making.The ultra-short-term wind power combination forecasting method,the intelligent UC decision-making approach based on LSTM & GRU and the intelligent UC decision-making method based on Seq2 Seq are the key research of this work.1)This work proposes a ultra-short-term wind power combination forecasting method based on noise assisted complex empirical mode decomposition and Elman Neural network.The white noise series are firstly adding in the raw data sequence to set up complex data series.Then,the wind power series will be decomposed by the NACEMD according to the different scale of wave.After that,a group of relatively stable components that has different time-frequency characteristics were setting up.And then,using the Elman neural network is used to establish the mapping model.Finally,according to the different time-frequency characteristics of the components,the superposition of the predicted results of each component is taken as the ultimate forecasting wind power.This part of works will provide the ultra-short-term wind power forecasting value for SCUC modelling in Chapter 3.It will also be the theoretical validation and pilot research for data-driven model modeling in Chapters 3 and 4.2)In this study,a data-driven intelligent security-constrained unit commitment dispatching method with self-learning ability is proposed.In this method,the Long-Short-Term Memory and Gated Recurrent Unit neural network is utilized to construct a deep learning model,which is focus on UC decision-making.Firstly,the K-means algorithm is used to do a cluster preprocessing for the historical dispatching data.Then,the unit commitment deep learning model is constructed by deep learning method.The mapping model between system load and dispatching results is constructed by historical data training.After that,the all above process is used as a foundation to do the UC decision-making.The model will be kept revised through the accumulation of historical data,which may give it the abilities of self-evolution and self-learning.A series of simulation results based on the standard calculation example and the practical grid data indicate some information.Compared with traditional physical-model-driven methods,this approach can not only improve the accuracy and efficiency of the method in practical usage,but also adapt to different kinds of UC problems.3)In this study,a data-driven UC intelligent decision-making approach has been proposed which could handle the non-cluster historical data.It is based on the Sequence to Sequence technology,which is building a composite Gated Recurrent Unit architecture.Firstly,the sample coding technology is used to compress the dimension of historical sample.Then the composite architecture is used to build the UC deep learning model.The model will be trained by the massive historical data which may make it the mapping model which can reflect the mapping relations between the daily load and UC dispatching method.The mapping model will be used to do UC decision-making.A series of simulations based on standard numerical examples indicate that,compared with data-driven methods based on gated recurrent unit neural network,this approach do not have to make a cluster preprocessing for the historical dispatching data.At the same time,it can effectively compress the dimensions of sample data with higher training and decision-making efficiency.