TimeGAN-based Approach for Power Load Timing Data Enhancement

As the new power system accelerates to clean and low-carbon transition, the rise of new energy penetration and the complexity of load structure lead to short-term load forecasting facing the dual challenges of data noise interference and small sample characteristics. Traditional data enhancement methods are difficult to capture the dynamic dependence characteristics of time series data, and are prone to destroying the distribution consistency, which restricts the prediction model accuracy. To this end, this paper proposes a time-series data enhancement method based on conditional TimeGAN, aiming to solve the key problems of high heterogeneity and low enhancement efficiency of multi-industry load data. The application strategy of TimeGAN is optimized for the characteristics of high-dimensional time series data: the generator adopts bidirectional LSTM to capture the time series dependency, the discriminator introduces a multi-scale feature fusion module to strengthen the temporal logic constraints, and the high-fidelity synthetic data are generated by jointly optimizing the antagonistic loss, reconstruction loss and hyperspherical loss. The experiment adopts the measured load data of a provincial power grid, which contains the typical industrial, commercial, residential and other industries with 5 days and 96 hours time series samples. Comparison experiments show that compared with traditional interpolation methods such as mean padding and SMOTE, TimeGAN-generated data show better cluster structure consistency in PCA and t-SNE low-dimensional visualization, and the generated samples are highly compatible with the original data in terms of autocorrelation coefficients, power spectral densities and other statistical features. In this study, the conditional generation adversarial mechanism is innovatively introduced into the power load time series enhancement, and the proposed “group enhancement” framework provides a new paradigm for the enhancement of high-dimensional time series data. The constructed synthetic dataset not only effectively alleviates the small sample problem, but also provides high-quality data support for the subsequent construction of robust load forecasting models and optimization of power system economic dispatch.