基于条件扩散模型的空调负荷场景生成方法与气象响应特性研究

Air Conditioning Load Scenario Generation Method Based on Conditional Diffusion Model and Research of Meteorological Response Characteristics

摘要:

空调负荷作为典型的气象敏感型负荷，其日内演变过程受气温、湿度等气象因素显著影响，具有较强的时序性与不确定性，准确建模高温负荷波动对保障电力系统安全调度具有重要意义。现有基于生成对抗网络和变分自编码器的方法存在训练不稳定或生成结果过于平滑等局限。为此，提出了一种基于条件扩散模型的空调负荷场景生成方法。该方法通过引入气象与日期等先验信息，结合扩散过程与Transformer结构的去噪网络，构建具备可控性与物理一致性的负荷场景。以某小区2023年夏季的负荷与气象数据为研究对象，将所提方法与条件生成对抗网络（conditional generative adversarial networks, CGAN）和条件变分自编码器（conditional variational auto encoder, CVAE）进行对比，从样本精度、分布匹配性、不确定性刻画以及时空相关性等维度开展性能评估。实验结果表明，所提方法在多项指标上均显著优于CGAN和CVAE，且生成的负荷场景在时空相关性和气象响应特性上更接近真实数据，能够更准确地刻画高温条件下的空调负荷特性。

Abstract:

Air conditioning load, as a typical weather-sensitive load, exhibits the pronounced diurnal variations primarily influenced by meteorological factors such as temperature and humidity, which also demonstrates distinct temporal patterns and inherent uncertainty. So, accurate modeling of high-temperature load fluctuations is essential for maintaining secure and reliable power system operation. The existing methods based on generative adversarial networks (GAN) and variational autoencoder (VAE) suffer from training instability and overly smoothed outputs. To address these limitations, a method for generating air conditioning load scenarios based on conditional diffusion model (CDM) is proposed. This method constructs load scenarios that are both controllable and physically consistent by introducing prior information such as meteorological data and calendar dates, and combined with diffusion processes and Transformer-based denoising network. Taking load and meteorological data from a residential community during the summer of 2023 as research object, the proposed method is compared with conditional generative adversarial networks (CGAN) and conditional variational autoencoder (CVAE). And the performances are evaluated across multiple dimensions, including sample accuracy, distributional similarity, uncertainty characterization and spatiotemporal correlation. Experimental results show that the proposed method significantly outperforms CGAN and CVAE across all major metrics, and its generated scenarios exhibit stronger spatiotemporal correlation and meteorological responsiveness, closely aligning with real data, which can enable a more accurate characterization of air conditioning load behavior under high-temperature condition.