摘要
电力系统负荷预测是能源领域的常规问题,为解决电力系统负荷预测中普遍存在的精度不高、特征工程处理较为粗糙等问题,采用了一种基于改进GBDT(gradient boosted decision tree)算法的负荷预测新方法。该方法采用加权离散化的特征工程手段处理输入特征,合理推导GBDT算法的预测原理,并在Python环境中建模预测,输出的3个月的预测误差分别为3.91%、5.00%和4.65%,同时在同一个数据集下,和LSTM(long short term memory)模型进行对比实验,分别用不同的指标对比了二者的泛化性能。实验结果表明,所提的预测算法在运算速度以及泛化能力方面优于LSTM算法。
Power system load forecasting is a conventional issue in the energy field,where problems such as low accura⁃cy and rough treatment by feature engineering commonly exist.To solve these problems,a novel load forecasting meth⁃od based on the improved gradient boosted decision tree(GBDT)algorithm is applied,which adopts the weighted dis⁃cretization process of feature engineering to treat the input characteristics and reasonably deduces the prediction princi⁃ple based on the GBDT algorithm.In addition,a prediction model is built in the Python environment,with prediction errors of output of 3.91%,5.00%,and 4.65%in three months,respectively.At the same time,under the same data set,a contrast experiment is carried out using the proposed method and the long short term memory(LSTM)model,and the generalization performance is compared between these two models with different indexes.Experimental results show that the proposed forecasting method is superior to the LSTM algorithm in terms of computation speed and general⁃ization capability.
作者
徐永瑞
左丰恺
朱新山
李硕士
刘洪瑞
孙彪
XU Yongrui;ZUO Fengkai;ZHU Xinshan;LI Shuoshi;LIU Hongrui;SUN Biao(School of Electrical and Information Engineering,Tianjin University,Tianjin 300072,China)
出处
《电力系统及其自动化学报》
CSCD
北大核心
2021年第8期94-101,共8页
Proceedings of the CSU-EPSA
关键词
负荷预测
决策树
集成学习
特征工程
样本空间
load forecasting
decision tree
integrated learning
feature engineering
sample space
