| 摘 要: 精准电力负荷预测对电网调度和能源管理具有重要指导意义。针对负荷数据复杂的非线性、多尺度周期性和随机波 动 特 性,以 及 单 一 模 型 难 以 有 效 捕 捉 全 部 动 态 的 局 限,提 出 融 合 STL、长 短 期 记 忆 LSTM 与Transformer模型的基于时序分解与双分支神经网络混合模型(STL-LSTM-TransformerNetwork,STL-LTNet)。模型首先利用 STL 将 原 始 负 荷 序 列 解 耦,随 后 设 计 双 分 支 架 构 中 LSTM 分 支 捕 捉 残 差 中 的 短 期 随 机 波 动,Transformer分支则负责建立长期依赖关系。最后通过交叉注意力机制动态融合双分支特征。实验结果表明,所提模型的均方误差(MSE)、均方根误差(RMSE)、平均绝对误差(MAE)分别为0.0034、0.0583和0.0465,相较于传统模型具有更高的电力负荷预测精度。 |
| 关键词: 电力负荷预测 时序分解 双分支神经网络 交叉注意力机制 |
|
中图分类号: TP399
文献标识码: A
|
|
| Power Load Forecasting Based on Time Series Decomposition and Dual-Branch Neural Networks |
|
HU Bowen, HE Liwen
|
(School of Internet of Things, Nanjing University of Posts and Telecommunications, Nanjing 210003, China)
1223076633@njupt.edu.cn; helw@njupt.edu.cn
|
| Abstract: Accurate power load forecasting provides crucial guidance for grid dispatch and energy management.To address the complex nonlinearity, mult-i scale periodicity, and stochastic fluctuations inherent in load data, as well as the limitations of single models in effectively capturing its full dynamics, this paper proposes STL-LTNet (STL-LSTM-Transformer Network)—a hybrid model integrating Seasonal and Trend decomposition using Loess(STL), Long Shor-t Term Memory(LSTM), and Transformer based on temporal decomposition and dua-l branch neural networks. The model first employs STL to decompose the raw load series into components. Subsequently, it designs a dua-l branch architecture: the LSTM branch captures shor-t term stochastic fluctuations in the residual component, while the Transformer branch establishes long-term dependencies. Finally, a cross-attention mechanism dynamically fuses
features from both branches. Experimental results demonstrate that the proposed model achieves Mean Squared Error(MSE), Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE) values of 0.0034, 0.0583, and 0.0465,respectively, outperforming traditional models in power load forecasting accuracy. |
| Keywords: power load forecasting temporal decomposition dua-l branch neural network cross-attention mechanism |
