光伏并网发电系统中的短期功率预测与储能容量优化配置研究

Short-term power forecast of photovoltaic power generation system is complex, dynamic and disturbing, which have posed serious challenges to sustainable stability of photovoltaic grid-connected power generation. Given that current research cannot effectively solve these problems, theoretically based on feature selection and machine learning, this project will carry out the systematic study in four stages successively: (1) Establish comprehensive feature information databases based on meteorological institutions and photovoltaic power stations, then accordingly construct and test feature correlation analysis and preprocessing methods; (2) Take advantage of square root error loss function and non-convex penalty function, establish different neural networks and select the relevant features; (3) Make good use of existing machine learning models, improve and generalize stagewise hierarchical feature selection methods, then construct short-term power forecast models which can select relevant features, their associated two-way and three-way interaction features, finally realize the complete adaptation for complex weather; (4) Establish photovoltaic energy storage capacity optimization model to compensate short-term power forecast errors, then design expert system intervention mechanism to adjust the target function and constraints appropriately. This project has high value of application, which cannot only enrich short-term power forecast theory of photovoltaic power generation, but also provide dependable references on the rational allocation of energy storage system for photovoltaic power stations and enhance the ability to consume photovoltaic power generation for power grids.

光伏发电短期功率预测存在复杂性、动态性、多扰量等问题，这给光伏并网发电的持续稳定带来严峻挑战。鉴于现有研究不能切实解决这些问题，本项目将以特征选取与机器学习为理论基础，分四个阶段依次进行系统研究：（1）建立基于气象机构与光伏电站综合特征的信息库，据此构建并测试特征关联分析与预处理方法；（2）结合平方根误差损失函数和非凸惩罚函数的优势，建立不同神经网络预测模型并确定影响短期功率的一次项特征；（3）充分利用现有机器学习模型的优势，完善并推广阶段性层次结构特征选取方法，构建能够选取一次项、二次与三次交互项特征的短期功率预测模型，最终实现对复杂天气的完整适应；（4）构建基于弥补短期功率预测误差的储能容量优化配置模型，并为其设计专家系统干预机制以适当调整目标函数与约束条件。本项目不仅可以丰富光伏发电短期功率预测理论，也为光伏电站合理配置储能系统提高电网消纳光电能力提供参考依据，具有重要的应用价值。

太阳能的间歇性、波动性与随机性给光伏发电并网电力系统的安全性、稳定性、供电质量及电力供需平衡带来了极大挑战，准确预测光伏发电是解决问题的有效途径。本项目致力于构建能够挖掘重要特征的可靠预测模型，不仅可以提高光伏发电预测的准确性，还可以降低模型的复杂度，有效提高电力系统消纳光电的能力，为光伏发电并网电力系统的成本节约提供重要参考。项目组从预处理缺失或异常特征数据、构造新型复合惩罚函数、选取一次项与交互项特征、动态寻找最优参数、优选模型、优化配置等多方面建立了适应复杂数据结构的多种预测模型，同时通过构造非凸惩罚函数和阶段性层次结构的特征选择方法选取重要特征并简化机器学习模型，再根据预测误差的统计规律，构建储能容量优化模型，实验结果显示提出的特征选择方法和预测模型具有稀疏性、稳定性、鲁棒性、普适性和较好的预测效果。本项目是预测理论与方法、统计特征选择技术、能源经济与管理的交叉研究，对完善能源预测理论和推广特征选择技术具有一定的理论意义和实践价值。通过项目资助项目组共发表SCI检索的期刊论文10篇，其中以项目负责人为第一作者的8篇，第二通讯作者的2篇，获得实用新型专利1项，协助培养留学博士研究生1名，培养硕士研究生10名。