| 引用本文: | 郑若楠,李国杰,韩蓓,汪可友,彭道刚.基于加权扩展日特征矩阵的分布式光伏发电日前功率预测[J].电力自动化设备,2022,42(2): |
| ZHENG Ruonan,LI Guojie,HAN Bei,WANG Keyou,PENG Daogang.Day-ahead power forecasting of distributed photovoltaic generation based on weighted expanded daily feature matrix[J].Electric Power Automation Equipment,2022,42(2): |
|
| 摘要: |
| 分布式户用光伏发电系统的精确日前功率预测可为智能家庭优化运行提供依据，但历史数据量少和缺乏精确辐照预报数据的问题增大了预测难度。为此，将邻近多用户数据融合以扩充样本规模，提出一种考虑功率关联性和相关度权重的相似日搜索方法，并基于长短期记忆(LSTM)神经网络实现日前预测。分析光伏发电功率的影响因素及其内在相关性，基于天气类型统计数据划分日类型，并利用气象信息、相同日类型的历史功率信息和皮尔逊积矩相关系数构造加权扩展日特征矩阵。提取历史数据中与待预测日特征矩阵欧氏距离最小相似日的光伏功率，将其与关键气象特征共同输入LSTM神经网络模型进行预测。以北美丹佛市多个用户的实测数据验证了所提方法的有效性，该方法能够适用于历史数据受限的场景，且在多种天气类型下显著降低了预测误差。 |
| 关键词: 分布式光伏 日前功率预测 相关性 日特征矩阵 相似日 神经网络 |
| DOI：10.16081/j.epae.202112023 |
| 分类号:TM615 |
| 基金项目:国家自然科学基金资助项目(51877133) |
|
| Day-ahead power forecasting of distributed photovoltaic generation based on weighted expanded daily feature matrix |
|
ZHENG Ruonan1, LI Guojie1, HAN Bei1, WANG Keyou1, PENG Daogang2
|
|
1.Key Laboratory of Control of Power Transmission and Conversion, Ministry of Education, Shanghai Jiao Tong University, Shanghai 200240, China;2.School of Automation Engineering, Shanghai University of Electric Power, Shanghai 200090, China
|
| Abstract: |
| Accurate day-ahead power forecasting of distributed household photovoltaic generation system can provide a basis for optimal operation of smart houses, but the problems of lack of historical data and precise irradiance forecasting data increase forecasting difficulty. Therefore, the sample scale is enlarged by integrating data from multiple users in the nearby area, a similar day selection method considering power correlation and relevant weight is proposed, and the day-ahead forecasting is realized based on LSTM(Long Short-Term Memory) neural network. The influencing factors of photovoltaic generation power and their internal correlation are analyzed, the day types are classified based on the statistical data of weather type, and the meteorological information, historical power information and Pearson product-moment correlation coefficient are used to construct the weighted expanded daily feature matrix. The photovoltaic power of similar day with minimum Euclidean distance of feature matrix of the day to be forecasted is selected from historical data, and it is input LSTM neural network model together with key meteorological features for forecasting. The validity of the proposed method is verified by the measured data of multiple users in Denver City of North America, the proposed method can be applied in the scene with limited historical data and can significantly reduce the forecasting error in multiple weather types. |
| Key words: distributed photovoltaic day-ahead power forecasting correlation daily feature matrix similar day neural network |
