基于VMD-EWT-IASSP-EBILSTM的短期电力负荷预测  

Short-term Power Load Forecasting Based on VMD-EWT-IASSP-EBILSTM

作　　者：杨健 孙涛 陈小龙 苏坚 姚健 周倩 YANG Jian;SUN Tao;CHEN Xiao-long;SU Jian;YAO Jian;ZHOU Qian(State Grid Ningxia Wuzhong Power Supply Company,Wuzhong 751100,China)

机构地区：[1]国网宁夏供电有限公司吴忠供电公司,吴忠751100

出　　处：《科学技术与工程》2023年第27期11646-11654,共9页Science Technology and Engineering

基　　金：国家自然科学基金(62027801)。

摘　　要：电力系统在国家工业基础设施中起着举足轻重的作用,维持系统负荷高精度预测是保障电力系统高效供应的关键。针对负荷数据的非平稳性、随机性与非线性,负荷预测误差较大的问题,结合变分模态分解(variational mode decomposition, VMD)、经验小波变换(empirical wavelet transform, EWT)、改进的空洞卷积金字塔模块(improved atros spatial pyramid pooling, IASSP)、集成双向长短时记忆模块(ensemble BiLSTM,EBiLSTM),提出了一种短期电力负荷预测模型。为解决负荷数据的非平稳性引起的模型预测波动问题,通过变分模态分解方法与经验小波变换的结合分解为若干子序列,显著降低了原始负荷序列的复杂性;为提高模型预测精度,将分解的负荷子序列利用过零率指标划分高低频序列,在低频序列中构建一种时序依赖捕获模块EBiLSTM提取长期负荷特征,高频序列中构建特征提取模块IASSP提取局部负荷特征,最后累加各子序列的预测结果,实现电力系统负荷的短期预测。选取行业通用客观评价指标:平均绝对误差、均方根误差,在宁夏某地电站的实测数据上对比前沿算法进行仿真实验验证。结果表明,该算法平均绝对误差(mean absolute error, MAE)降低了37%~75%,具有较高的准确性与可靠性。The power system plays an important role in the national industrial infrastructure,the key to ensure the stable and efficient supply of the power system is to maintain the system load prediction with high precision.Aiming at the problem of nonstationarity,randomness and nonlinearity of load data and large error of load forecasting,combined with variational mode decomposition(VMD),empirical wavelet transform(EWT),improved atros spatial pyramid pooling(IASSP),and integrated bidirectional long and short term memory module(EBiLSTM),a short-term power load forecasting model was proposed.In order to solve the problem of model prediction fluctuation caused by the non-stationary nature of load data,the combination of variational modal decomposition and empirical wavelet transform was used to decompose into several subsequences,which significantly reduced the complexity of the original load series.In order to improve the prediction accuracy of the model,the decomposed load subsequences were divided into high and low frequency series using the zero-crossing rate index,and a time-series dependent capture module EBiLSTM was constructed in the low frequency series to extract the long-term load characteristics,and the feature extraction module IASSP was constructed in the high frequency series to extract the local load characteristics,and finally the prediction results of each subsequence were accumulated to achieve the short-term load prediction of the power system.The mean absolute error(MAE)and root mean square error(RMSE)were selected as the common objective evaluation indexes in the industry,and the simulation experiment was verified by comparing the frontier algorithm with the measured data of a power station in Ningxia.The results show that the error of the algorithm is reduced by 37%~75%,and it has high accuracy and reliability.

关 键 词：电力负荷 变分模态分解 经验小波变换 特征提取 高低频序列 

分 类 号：TK221[动力工程及工程热物理—动力机械及工程]