基于STL与自适应残差多分支融合网络的短期电力负荷预测

Short-Term Power Load Forecasting Based on STL and Adaptive Residual Multi-Branch Fusion Network

摘要: 新型电力系统建设中分布式光伏、储能装置的大规模接入及用户侧柔性负荷的多元增长，使得区域用电时序有高度非线性、强时变非平稳的特征，短期负荷预测质量直接关联电网运行调度、备用及市场交易的经济效率。目前主流深度学习方法尚不能同时处理长周期负荷模式识别和由短期气候扰动所导致的负荷异常，所用多尺度特征提取手段往往有失均衡。针对工程实际需要，文中将时序分解方法同异构深度网络加以结合，提出STL-APRFNet类短期电力负荷预测方案。所用鲁棒STL时序分解法对原负荷序列做加性拆分，把长期趋势、日内循环、气象相关残差划为各自独立的分量，由此缓解初始时序非平稳的困难；按三类分量不同性质分别设计特征提取通道：以多层感知机分析负荷慢变的长期基线，利用PatchTST的切块注意机制考察周或日级长程联系，另以一维卷积作跨模态设计，同时考虑气象变量及残差序列，以局部滑动方式将气象急变与负荷瞬态结合于同一特征空间。多通道所得特征在融合时应用可调缩放因子，使STL分解所提示的先验物理知识与网络拟合结果达到平衡，不因深度训练而忽略负荷变化基本规律。依据GEFCom2014公开电力数据集完成基准对照试验，STL-APRFNet的均方根误差、平均绝对误差是12.72 MW、8.88 MW，较典型LSTM基准模型分别减少25.2%、27.8%；模型平均百分比误差仅5.754%，R² (决定系数)可取到0.937。从实测角度考察，文中方法所建模型的误差控制效果更优、长期预测稳定度高，对日前调度具有明确帮助意义，是很有潜力的工程解决方案。

Abstract: The integration of distributed photovoltaics, energy storage and diversified user-side flexible loads makes power load time series highly nonlinear, time-varying and non-stationary in new power systems. Reliable short-term load forecasting is crucial for efficient grid scheduling, reserve arrangement and market trading. Existing deep learning methods fail to jointly model long-term load patterns and meteorology-caused load anomalies caused by short-term climate disturbances, with limited multi-scale feature extraction balance. In response to the practical needs of engineering, this paper proposes an STL-APRFNet short-term load forecasting model combining STL time series decomposition and heterogeneous deep networks. The robust STL method decomposes raw load data into long-term trend, intraday periodicity and meteorological residual components to mitigate time series non-stationarity. Targeted feature extraction branches are designed based on the different properties of three types of components: a multi-layer perception captures long-term load baselines, PatchTST acquires long-range temporal correlations. Moreover, 1D convolution is deployed for cross-modal modeling to jointly incorporate meteorological variables and residual sequences, merging abrupt meteorological changes and load transients into a unified feature space through local sliding operations. Adjustable scaling factors are utilized in multi-channel feature fusion to balance physical prior knowledge from STL decomposition and network fitting outcomes, so that deep training will not overlook the fundamental laws of load variation. Validated on the GEFCom2014 dataset, the proposed model achieves RMSE of 12.72 MW and MAE of 8.88 MW, reducing errors by 25.2% and 27.8% compared with LSTM. It obtains a MAPE of 5.754% and R² of 0.937. Validated from practical measurements, the model constructed by the proposed method delivers superior error control performance and high stability for long-term prediction. It provides clear support for day-ahead scheduling and represents a promising engineering solution.