摘要
准确的中长期电力负荷预测对电力系统的规划和运行至关重要。由于传统方法在非线性特性处理和时序依赖建模方面存在局限,难以全面捕捉负荷数据的复杂特征,因而提出了一种基于残差网络和注意力机制的RA-LSTM模型。模型通过引入残差连接,缓解梯度消失问题,增强了模型对长时序依赖特征的捕捉能力;同时融合注意力机制,增强了对关键时间点和特征的敏感性。以山西省为案例,构建了融合时间特征和气象要素的数据集,对RA-LSTM模型进行了全面评估。实验结果表明,RA-LSTM模型在均方根误差(RMSE)、平均绝对误差(MAE)、平均绝对百分比误差(MAPE)及决定系数(R^(2))等指标上均显著优于基准BP模型和传统LSTM模型。RA-LSTM模型的MAPE、MAE较BP模型的分别降低了41.8%、40.9%,显著提升了模型的预测精度和稳定性。显著性检验结果进一步验证了RA-LSTM模型预测结果的科学性,为中长期电力负荷预测提供了一种高效且稳健的解决方案,并为未来探索多特征融合和模型优化提供了理论和实践基础。
Accurately forecasting medium-and long-term power load is crucial for the planning and operation of power systems.Because traditional methods have limitations in handling nonlinear characteristics and modeling temporal dependencies,and it is difficult to fully capture the complex features of load data,this paper constructs a RA-LSTM model based on residual networks and attention mechanisms.By introducing residual connections,the model alleviates the vanishing gradient problem and enhances its ability in capturing long-term temporal dependencies.Additionally,the integration of attention mechanisms improves the model’s sensitivity to key time points and features.Using Shanxi Province as a case study,this paper builds a dataset that integrates temporal characteristics and meteorological factors,and comprehensively evaluates the RA-LSTM model.Experimental results demonstrate that the RA-LSTM model significantly outperforms the baseline BP model and the traditional LSTM model in key metrics,including root mean square error(RMSE),mean absolute error(MAE),mean absolute percentage error(MAPE),and coefficient of determination(R^(2)).The RA-LSTM model reduces MAPE by 41.8% and MAE by 40.9% compared to the BP model,significantly improving prediction accuracy and stability.Significance tests further validate the scientific reliability of the predictions of the RA-LSTM model.The achievement of this study could provide an efficient and robust solution to medium-and long-term load forecasting and lay a theoretical and practical foundation for future exploration of multi-feature integration and model optimization.
作者
周绍妮
吴优
窦雨菡
郑奕扬
Zhou Shaoni;Wu You;Dou Yuhan;Zheng Yiyang(School of Economics and Management,Beijing Jiaotong University,Beijing 100044,China)
出处
《气象与环境科学》
2025年第1期78-87,共10页
Meteorological and Environmental Sciences
基金
国家自然科学基金项目(72372007)
大学生创新创业训练计划项目(202510004169)。
作者简介
周绍妮(1972-),女,山东烟台人,教授,博士,从事大数据财务分析与决策支持研究.E-mail:snzhou@bjtu.edu.cn。
