基于生成对抗网络的综合能源负荷场景生成方法

综合能源负荷场景生成是研究能源计量、规划运行等领域问题的基础,具有重要意义。但由于数据采集困难、综合能源负荷多能耦合等因素的限制,综合能源负荷场景的多样化生成仍是一大难题。提出了一种基于生成对抗网络(generative adversarial networks, GAN)的综合能源负荷场景生成方法。首先建立梯度惩罚优化的Wasserstein生成对抗网络模型,解决综合能源负荷的高随机性可能带来的不收敛或模式崩溃问题。其次,基于深度长短期记忆(long short-term memory, LSTM)的循环神经网络构建生成对抗网络的生成器和判别器,使模型更适用于复杂综合能源负荷数据生成。算例结果表明,所提模型的生成负荷场景在概率分布、曲线标志性特征和冷热电负荷之间相关性等方面相较于蒙特卡洛法和原始生成对抗网络均获得了较好结果,可以在不同模式下生成具有多样性且逼真的负荷场景。

Load Scenario Generation of Integrated Energy System Using Generative Adversarial Networks

Load scenario generation is the basis of studying energy measurement, operation scheduling and other fields, which is of great significance. Due to difficulty of data collection and multi-energy coupling of integrated energy system, it is still a big challenge to generate load data with diversity. A novel multi-load scenario generation method based on generative adversarial network (GAN) is proposed in this paper. Firstly, the Wasserstein generative adversarial network model with gradient penalty optimization is established to overcome the misconvergence and mode collapse caused by high randomness of load. Secondly, on the basis of the recurrent neural network with deep long-term and short-term memory, the generator and discriminator in the GAN are constructed to be more suitable for load data generation of complex integrated energy system. The result shows that the scenarios generated by proposed model achieves better results in probability distribution, curve signature features and correlation in cooling, heating and power load than original GAN and Monte Carlo method. The model can generate realistic load scenarios with diversity in different modes.