Application Of Deep Neural Network In Electric Power Industry Modeling

In recent years,the electrical power industry in China has reached an unprecedented level.In order to constructing energy internet,the electric power industry is developing towards a highly informationized and intelligent direction.In the power generation process,huge amounts of production,operation,control and other kinds of data are generated,which can be generally characterized by massiveness,multi-source,heterogeneity and high-dimension.The accumulation of big data not only promotes the development of production technology,but also brings great challenges.Due to its limited representation ability,the traditional shallow-layer neural network cannot fully extract the information contained in the big data of electric power industry.This thesis intends to fully explore the useful information in the big data by constructing appropriate deep neural network(DNN)structure.And based on this,developing the DNN framework suitable for the electric power industry.To achieve the goal,the following works have been carried out in this thesis.(1)The unsupervised learning and supervised learning algorithms in deep neural network are studied.Firstly,the most commonly used supervised learning algorithm:gradient descent algorithm are analyzed.The three forms of gradient descent algorithm:batch gradient descent algorithm,stochastic gradient descent algorithm and mini-batch gradient descent algorithm are explained in detial.And the effectiveness of the three algorithms are verified through the example of nonlinear function approximation.Then,in order to sovle the exisiting problem in deep neural network traning using supervised learning,the training algorithm combining unsupervised learning and supervised learning are proposed.The performance of the proposed algorithm is verified by several simulation examples.(2)The stacked auto-encoder model used for the ultra-supercritical coal fired boiler-turbine unit is established.For the ultra-supercritical unit with multi-variables,great inertia and highly-nonlinear,the stacked auto-encoder is used for modeling.In the training process of this model,maximum correntropy is chosen as the loss function,since it can effectively alleviate the influence of the outliers existing in ultra-supercritical unit data.(3)The stacked denoising auto-encoder and long short-term memory network model suitable for the ultra-supercritical coal fired boiler-turbine unit is established.In order to further consider the typical characteristics of ultra-supercritical unit,such as the great inertia,the large delay and the noise of the monitoring data,a hybrid deep neural network model is established by the combining of stacked denoising auto-encoder and long short-term memory network.In this model,the stack denoising auto-encoder is used for the processing and feature selection of the noisy input data,while the long short-term memory network in charge of outputting the expected normal system behaviors along the time axis.(4)The DNN model for wind speed forecasting is proposed.A deep neural network model combining stacked denoising auto-encoder with long short-term memory network is constructed for the wind speed with highly fluctuation and intermittent.The feature selection method based on mutual information is developed to determine the most appropriate inputs for the model.With the real-time big data from the wind farm running log,the stacked denoising auto-encoder is used to extract the features of the original data layer by layer,and the long short-term memory network is responsible for generating the final output of the model.