Toward Decision-oriented Wind Power Generation Forecasting Methods In Power Systems

In the development of the society and economy of our country,there are many challenges,such as the growing of energy demand,the increasing pressure of environmental pollution,and the rising of dependence on foreign energy.Wind power plays an important role in solving these challenges.However,the uncertain,random,and fluctuate nature of wind power leads to challenges in integrating it into existing electricity systems.Wind power generation forecasting is considered to be one of the most cost-efficient solutions to these problems.In this thesis,following the principal of "forecasts are made to guide decisions",a decision-oriented forecasting method is first proposed,and related theory,algorithms,and applications are researched.The main contents of this thesis are:1)The effects of loss function in forecasting model are analyzed in the viewpoint of statistics,then the definition of optimal forecasting is illustrated.A point forecast of wind power generation problems using cost-oriented loss functions is developed,which can incorporate the actual costs associated with forecasting errors into a model building and forecasting process.The characters of optimal forecasting under costoriented loss function are analyzed.2)A cost-oriented boosted regression tree algorithm is proposed,making it capable of handling the cost-oriented loss function efficiently.The cost-oriented loss function is practical using this algorithm.The proposed method is tested on a wind power trading problem in electricity markets,with the aim of maximizing the benefit of wind power producers.In the test cases,compared with the traditional unbiased point forecasting method,the cost-oriented forecasting method can reduce 23.06% of the total cost,which is a promising result.3)A method for clustering high-dimensional data in wind power generation forecasting problems is proposed.The method is based on TRansformation Under Stability-reTraining Equilibrium Characterization(TRUST-TECH)technique.This method employs linear and nonlinear dimension reduction techniques to map the high-dimensional data into a low-dimensional space while the structure of the original data is kept,then uses TRUST-TECH e to find multiple clustering results of the low-dimensional data.At last,clustering ensemble technique is adopted to combine different clustering results to achieve better and more robust final clustering result.Case studies in real-world datasets demonstrate the efficiency of the proposed method.4)The reason why forecasting error is biased under quadratic loss function is analyzed.Then,a wind power generation forecasting method based on residual learning and error correction is proposed.In this method,residual learning is used to combine several forecasting model,and error correction is used to correct the forecasting results.In the error correction process,the forecasting results are partitioned to different classes based on the wind power generation curve,and the forecasting values in each class are corrected following their error distribution.The final forecasting results are unbiased,which can satisfy the demand of some applications in power system.