【作者】 周慧；

【导师】 曾杏元；

【作者基本信息】 湖南师范大学 ， 应用统计（专业学位）， 2021， 硕士

【摘要】 电力负荷预测是电网系统规划和运行过程的重要环节。通过合理的负荷预测方法实现准确的预测有助于提高电力利用效率,并取得经济和社会效益。但由于电力负荷是非线性的,易受到气象、经济等各类不确定因素的影响,给负荷预测增加了随机性与不稳定性,使得单纯的点预测或者单一模型不能满足预测精度的稳定性的需求。因此本文以区间预测、概率密度预测为方向,通过构建优化组合负荷预测方法更好的描述负荷变化范围与程度以提升模型在精度与稳定性方面的综合预测能力。本文的实证研究基于2018年国内某市实际电力负荷数据集和影响因素数据集。采用了XGBoost算法对影响因素如气温、体感温度、季节、节假日以及历史负荷等进行特征选择以去除冗余特征,为后续的模型验证做数据准备。之后,本文构建了优化组合模型:第一步,构建优化的单项模型,即采用贝叶斯优化算法对随机森林分位数回归、梯度提升分位数回归以及支持向量机分位数回归三种非线性分位数回归算法进行超参优化并得到优化模型;第二步,采用不同的组合方法将优化后的单项模型进行组合于是构建了对应的优化组合模型。对于组合方法的选择,本文考虑了定权值和变权值组合方法,包括简单平均、基于MAPE权重、最小二乘组合及约束分位数回归平均组合方法。所构建的优化组合模型和优化的单项模型都可以实现对负荷的条件分位数的预测,得到0.01至0.99分位点信息,间隔为0.01。此外,本文还在实现点预测的基础上,利用分位点信息结合Bootstrap区间估计方法实现区间预测、结合核密度估计实现概率密度预测,得到更加全面的预测信息。为评估模型性能,本文对单项模型和优化组合模型的预测效果进行了对比分析,采用了不同的指标评价模型的性能,如MAPE、MSE、CPIA及APL等。结果表明,贝叶斯优化算法有效提高了单项模型预测准确性,约束分位数回归平均优化组合模型在点预测、区间预测、概率密度预测稳定性及准确性方面表现出更好的性能。更多还原

【Abstract】 Power load forecasting is an important part of the planning and operation of the process grid system.Accurate forecasting which realized by reasonable load forecasting methods will help improve power utilization efficiency and achieve economic and social benefits.However,The power load is non-linear and easily affected by various uncertain factors such as weather and economy.As a result,that adds randomness and instability to the load forecast,making the simple point forecasting method or a single model unable to meet the demand for the stability of forecast accuracy.Therefore,this paper takes interval prediction and probability density prediction as the direction,and describes the range and degree of load change by constructing an optimized combination load forecasting method to improve the model’s comprehensive forecasting ability in terms of accuracy and stability.The empirical research in this paper is based on the actual power load data set and influencing factor data set of a certain city in China in2018.In this paper,the XGBoost algorithm is used to perform feature selection on influencing factors such as temperature,body temperature,seasons,holidays,and historical load to remove redundant features and prepare data for subsequent model verification.After that,the paper constructs the optimal combination of model: Firstly,building the optimized single model,which uses Bayesian optimization algorithm to nonlinear three kinds quantile regression algorithm(Random Forest Quantile Regression,Quantile Regression Gradient Boosting and Support Vector Quantile Regression)to optimize the hyperparameters and obtain the optimized model;Secondly,different combination methods are used to combine the optimized single models,and the corresponding optimized combination model is constructed.For the choice of combination method,this paper considers fixed weight and variable weight combination methods,including Simple Average,MAPE-based weight,Least Square and Constrained Quantile Regression Average combination methods.Both the constructed optimized combination model and the optimized single model can realize the prediction of the conditional quantile of load,and obtain 0.01 to 0.99 quantile information with an interval of 0.01.In addition,based on the realization of point prediction,this paper uses quantile information combined with Bootstrap interval estimation method to achieve interval prediction,and combined with kernel density estimation to achieve probability density prediction,and obtain more comprehensive prediction information.In order to evaluate the performance of the model,this paper compares and analyzes the prediction effects of the single model and the optimized combination model,and uses different indicators to evaluate the performance of the model,such as MAPE,MSE,CPIA,and APL.The results show that the Bayesian optimization algorithm effectively improves the prediction accuracy of the single model.Compared with the optimized single model,the Constrained Quantile Regression Average optimized combination model shows better performance in point prediction,interval prediction,and probability density prediction stability and accuracy.更多还原