摘要
准确的短期剩余负荷预报对电网稳定运行和新能源消纳具有重要意义。为提高区域电网短期剩余负荷预报精度,提出基于逐日归一化方法、完整继承经验模态分解(CEEMDAN)和长短期记忆神经网络(LSTM)的短期剩余负荷混合预报模型,以湖南省电网为案例开展应用研究。首先利用2020~2023年湖南省逐日24点负荷、新能源出力、区域内外输出功率等数据提取剩余负荷序列,并按日进行归一化处理;然后采用CEEMDAN将归一化序列分解为多个固有模式函数(IMF),并对每个IMF以及日最值序列建立独立的LSTM模型进行预报;最后,通过聚合还原操作得到剩余负荷预报结果。研究结果表明,预见期为24 h时,基于逐日归一化的预报模型的多个评价指标均优于基于传统全局归一化的同类模型,逐日归一化的CEEMDAN-LSTM模型表现最佳,测试期的决定系数R2为0.83,较全局归一化的LSTM、EMD-LSTM和CEEMDAN-LSTM分别提升45.6%、9.2%和5.0%;平均绝对误差MAE和均方根误差RMSE分别为1 209 MW和1 604MW,较全局归一化的CEEMDAN-LSTM分别降低18.3%和9.0%。研制的混合预报模型能显著提升短期剩余负荷预报精度,为电网稳定运行和新能源消纳提供技术支撑。
Accurate short-term forecasts of residual loads are of great significance for the stable operation of the power grid and the integration of renewable energy.To improve short-term forecast accuracy of residual loads,this study proposes a hybrid forecasting model by integrating daily normalization method,Complete Ensemble Empirical Mode Decomposition with Adaptive Noise(CEEMDAN)and Long Short-Term Memory(LSTM)neural network.The Hunan provincial power grid was selected as case study.Firstly,the residual load series were extracted using the daily 24-point loads,new energy output,and output power both inside and outside Hunan Province from 2020 to 2023,and normalized on a daily basis.Then,CEEMDAN was used to decompose the normalized sequence into multiple intrinsic mode functions(IMFs),and an independent LSTM model was established to predict each IMF and daily maximum value series.Finally,the aggregation operation was utilized to create the results of residual load forecasts.The results indicate that the daily normalization-based model outperforms traditional global normalization models across multiple evaluation metrics for a 24-hour forecast period.The daily normalized CEEMDAN-LSTM model has the best performance.The R2 value of the daily normalized CEEMDAN-LSTM model is 0.83 in the testing stage,which is 45.6%,9.2%and 5.0%higher than those of LSTM,EMD-LSTM and CEEMDAN-LSTM models,respectively.The Mean Absolute Error(MAE)and Root Mean Square Error(RMSE)values of the daily normalized CEEMDAN-LSTM model are 1209 MW and 1604 MW,respectively,which are 18.3%and 9.0%lower than those of the basic CEEMDAN-LSTM model.The proposed hybrid forecasting model can significantly improve the short-term forecasting accuracy of residual loads,providing technical support for stable operation of power grid and the absorption of new energy.
作者
韦溢龙
周研来
汤艳
汤纯
李彦倩
WEI Yi-long;ZHOU Yan-lai;TANG Yan;TANG Chun;LI Yan-qian(State Key Laboratory of Water Resources Engineering and Management,Wuhan University,Wuhan 430072,Hubei Province,China)
出处
《中国农村水利水电》
北大核心
2026年第3期228-233,共6页
China Rural Water and Hydropower
基金
国家重点研发计划(2024YFC3212700)资助。
