数据驱动下的日前市场发电商智能决策方法  

Data-driven Intelligent Decision-making Method for Day-ahead Market Power Generation Companies

作　　者：李宏仲 吕勇荡 LI Hongzhong;Lü Yongdang(College of Electrical Engineering,Shanghai University of Electric Power,Yangpu District,Shanghai 200090,China)

机构地区：[1]上海电力大学电气工程学院,上海市杨浦区200090

出　　处：《电网技术》2023年第3期1056-1065,共10页Power System Technology

摘　　要：双碳背景下日前市场逐步开放,新能源发电接入电力系统中,增加了发电商报价决策的风险。考虑新能源出力的不确定性给发电商带来额外的成本,将新能源发电商也视为市场价格的制定者,建立了发电商的报价博弈模型,提出了一种深度学习的数据驱动型发电商智能决策方法。该方法通过学习历史的市场交易结果,不需要利用发电商的成本等隐私参数就能求解最优报价决策。首先,利用Nash均衡条件搜索历史数据中的最优报价标签,并与负荷功率、天气等因素作为数据驱动模型的训练数据。其次,构建了基于自动编码器(auto encoder, AE)与长短期记忆(long short-term memory,LSTM)网络结合的AE-LSTM深度学习模型。考虑到含新能源的电力市场报价样本较少难以满足深度学习的训练要求,利用大量的无标签数据对LSTM网络做无监督预学习,再用少量含标签的样本数据对该网络微调整。利用训练后LSTM模型在线决策得到发电商的最优报价,并递交至独立系统运行商完成出清。最后,通过仿真算例验证了所提出的数据驱动模型在不同场景下报价决策的有效性。With the gradual opening of the day-ahead markets under the background of double carbons, a lot of new energy generation grid connection increases the risks of bidding decision-making of Generation Companies(GENCOs)obviously. This paper regards the new energy producer as one of the price makers and considers that the uncertainty of new energy output brings about extra costs to the producers, thus a bidding game model for the producers is established and an intelligent data-driven decision-making method based on deep learning for the GENCOs is proposed. By studying the historical market transaction results, this method is able to solve the optimal bidding without referring to the GENCOs’ private information such as their costs. Firstly, the optimal bid labels in the historical data are selected by using the Nash equilibrium conditions. These labels are treated as the training data for the data-driven model with load power, weather and other factors. Then, a deep learning model which combines the Auto Encoder(AE) with the Long/Short-Term Memory(LSTM)network is proposed. Since there are few bidding samples in the new energy market, which is unable to meet the training requirements of deep learning, the LSTM network is allowed to pre-learn with a large amount of unlabeled data and is fine-tuned with some labeled samples. After the training, the LSTM model makes the decision online and submits the optimal bidding of GENCOs to the independent system operators for clearing. Finally, the simulation results show the effectiveness of the proposed data-driven model in different scenarios.

关 键 词：日前市场 发电商 报价决策 数据驱动 长短期记忆 自动编码器 

分 类 号：TM721[电气工程—电力系统及自动化]