摘要
在“双碳”战略目标下,区域综合能源系统 (regional integrated energy systems,RIES) 多元负荷的高精度短期概率预测是实现精细化调度的关键。针对输入特征高维冗余、多能耦合规律难以挖掘以及不确定性风险难以量化等难题,提出一种基于主客观特征优选与多任务 Transformer 联合分位数回归 (multi-task Transformer quantile regression,MT-TransQR) 的多元负荷短期概率预测方法。首先,构建基于最大互信息系数 (maximal information coefficient,MIC) 与层次分析法 (analytic hierarchy process,AHP) 的融合特征筛选策略,在剔除冗余噪声的同时融入物理先验机理,提纯关键驱动特征集;其次,构建基于硬参数共享的多任务 Transformer 网络,并行捕捉电、冷、热负荷的长时序依赖,打破独立建模壁垒以深度挖掘多能耦合特征;最后,结合分位数回归技术构建多任务联合弹球损失函数,直接生成无需先验分布假设的多置信水平负荷预测区间。实际 RIES 数据算例表明,所提方法有效克服了传统参数化假设局限,在保证高区间覆盖率的同时显著收窄了区间宽度,实现了预测可靠性与敏锐度的最优权衡,为系统的协同风险调度提供了科学的量化支撑。
Abstract
Toward the national strategic goal of “ carbon peak and carbon neutrality” , high-precision short-term probabilistic forecasting of multi-energy loads in regional integrated energy systems (RIESs) is crucial for refined scheduling. To address challenges such as high-dimensional feature redundancy, the difficulty in mining multi-energy coupling mechanisms, and the inaccurate quantification of uncertainty risks, this paper proposes a short-term probabilistic forecasting method based on subjective-objective feature selection and multi-task Transformer quantile regression (MTTransQR). First, a fusion feature selection strategy based on the maximal information coefficient (MIC) and the analytic hierarchy process (AHP) is constructed to eliminate redundant noise, integrate physical prior mechanisms, and extract key driving features. Second, a multi-task Transformer network using hard parameter sharing is established to capture the longterm dependencies of electricity, cooling, and heating loads in parallel, breaking independent modeling barriers to deeply mine multi-energy coupling features. Finally, quantile regression is combined to construct a joint pinball loss function,directly generating multi-confidence-level load prediction intervals without requiring prior distribution assumptions. Case studies on actual RIES data demonstrate that the proposed method overcomes the limitations of traditional parametric assumptions. While ensuring a high prediction interval coverage probability, it significantly narrows the prediction interval normalized average width, achieving an optimal trade-off between reliability and sharpness, and providing scientific quantitative support for system risk scheduling.
关键词
综合能源系统 /
多元负荷短期预测 /
概率预测 /
特征筛选 /
多任务学习
Key words
integrated  /
energy  /
system /
short-term  /
multi-energy  /
load  /
forecasting /
probabilistic  /
forecasting /
feature selection /
multi-task learning
翟晶晶1 , 李新琳2 , 沈观斌1 , 王 盛2.
基于主客观特征优选与 MT-TransQR 的多元负荷短期概率预测[J]. 分布式能源, 0 https://doi.org/10.16513/j.2096-2185.DE.26110225.
ZHAI Jingjing1, LI Xinlin2, SHEN Guanbin1, WANG Sheng.
Short-Term Probabilistic Forecasting of Multi-Energy Loads Based on Subjective-Objective Feature Selection and MT-TransQR[J]. Distributed Energy, 0 https://doi.org/10.16513/j.2096-2185.DE.26110225.
{{custom_sec.title}}
{{custom_sec.title}}
{{custom_sec.content}}