Air Conditioning Load Data Generation Method Based on DTW Clustering and Physically Constrained TimeGAN

Generating air-conditioning system load data is crucial for tasks such as power grid scheduling and intelligent energy management. Air-conditioning load data exhibit strong non-stationarity. Their load curves are influenced by seasonal variations and highly correlated with outdoor meteorological conditions, indoor activity patterns, and equipment operational strategies. These characteristics lead to pronounced periodicity, sudden shifts, and diverse data patterns. Existing load generation models tend to produce averaged distributions, which often leads to the loss of specific temporal patterns inherent in air-conditioning loads. Moreover, as purely data-driven models, they lack explicit physical constraints, resulting in generated data with limited physical interpretability. To address these issues, this paper proposes a hybrid generation framework that integrates the DTW clustering algorithm, a physically-constrained TimeGAN model, and an LSTM-based model selection mechanism. Specifically, DTW clustering is first employed to achieve structured data partitioning, thereby enhancing the model's ability to recognize and model diverse temporal patterns. Subsequently, to overcome the dependency on detailed building parameters and extensive sensor networks, a parameter-free physical constraint mechanism based on intrinsic temperature-load correlations is incorporated into the TimeGAN supervised loss. This design ensures thermodynamic consistency even in sensor-scarce environments where only basic operational data is available. Finally, to address adaptability challenges in long-term sequence generation, an LSTM-based selection mechanism is designed to evaluate and select from clustered submodels dynamically. This approach facilitates adaptive temporal fusion within the generation strategy. Extensive experiments on air-conditioning load datasets from Southeast China demonstrate that the framework achieves a local similarity score of 0.98, outperforming the state-of-the-art model and the original TimeGAN by 11.4% and 13.3%, respectively.