基于多任务渐进式学习模型的风-光-荷功率短期预测  

Short-Term Prediction of Wind-Photovoltaic-Load Power Based on Multi-Task Progressive Learning

作　　者：李丹[1] 唐建 甘月琳 罗娇娇 黄烽云 LI Dan;TANG Jian;GAN Yuein;LUO Jiaojiao;HUANG Fengyun(College of Electrical Engineering and New Energy,China Three Gorges University,Yichang Hubei 443002,China;Hubei Key Laboratory of Operation and Control of Cascade Hydropower Station(China Three Gorges University),Yichang Hubei 443002,China;不详)

机构地区：[1]三峡大学电气与新能源学院,湖北宜昌443002 [2]梯级水电站运行与控制湖北省重点实验室(三峡大学),湖北宜昌443002 [3]不详

出　　处：《湖北电力》2024年第2期38-47,共10页Hubei Electric Power

基　　金：国家自然科学基金项目(项目编号:51807109)。

摘　　要：同一地区的风电、光伏和负荷功率与风速、辐照度和温度等气象因素密切相关,在电力系统不同运行场景下存在一定的互动耦合关系。为了捕获多源荷间的差异性和相关性,挖掘高维数据中蕴含的潜在规律,提出一种基于深度时空融合网络的多任务渐进式学习模型,实现风-光-荷联合功率预测。首先,基于深度时空融合网络设计共享信息与特有信息子网;然后,构建计及时空相关性的多任务渐进式学习模型,分别对风、光、荷功率的共享和特有时空信息进行由浅至深渐进式提取;最后,将共享信息与特有信息子网所得特征向量进行融合映射,实现对未来风-光-荷功率的联合预测。实际日前风-光-荷联合预测算例结果表明,所提模型可弥补现有多任务模型出现“负迁移”和“跷跷板”现象的不足,提高预测精度和稳健性。Wind power,photovoltaic power and load power in the same area are closely related to meteorological factors such as wind speed,irradiance and temperature,and there is a certain interactive coupling relation in different operation scenarios of power system.In order to capture the difference and correlation between multi-source loads and mine the potential rules contained in high-dimensional data,a multi-task progressive learning model based on deep spatiotemporal fusion network is proposed to realize the wind-photovoltaic-load joint power prediction.Firstly,the shared information and specific information subnet are designed based on deep spatiotemporal fusion network.Then,a multi-task progressive learning model with time-space correlation is constructed to progressively extract the shared and unique spatiotemporal information of wind,photovoltaic and load power from shallow to deep.Finally,the feature vectors obtained from shared information and specific information subnet are fused and mapped to realize the future wind-photovoltaic-load joint power prediction.The actual day-ahead wind-photovoltaic-load power short-term prediction example shows that the proposed model can compensate for the shortcomings of the existing multi-task model with“negative migration”and“seesaw”phenomena,and thus improve the prediction accuracy and robustness.

关 键 词：多源荷 多任务渐进式学习 风-光-荷联合功率预测 深度时空融合网络 电力系统 光伏发电 风力发电 

分 类 号：TM614[电气工程—电力系统及自动化]