Short-Term Spinning Reserve Requirement Estimation Using TimeGAN-Based Synthetic Data Augmentation and a Hybrid LSTM–XGBoost Model

Accurate short-term forecasting of spinning reserve requirements is essential for ensuring frequency stability, operational reliability, and economic efficiency in modern power systems. However, the increasing penetration of renewable energy resources and the limited availability of high-quality operational data make reliable forecasting a challenging task. This study proposes a novel hybrid forecasting framework that integrates synthetic data generation, deep learning, and machine learning to overcome these limitations. To the best of the author’s knowledge, this is the first study that integrates TimeGAN-based synthetic data generation with a hybrid LSTM–XGBoost model specifically for short-term spinning reserve forecasting. Given the limited availability of real-world spinning reserve datasets, this study employs a TimeGAN-based synthetic data generation approach trained on multivariate power system variables (load, renewable generation, and frequency) to construct a realistic and representative dataset for model development. First, TimeGAN is employed to generate realistic synthetic time-series data that preserve the temporal dynamics of load, renewable generation, frequency deviations, and spinning reserve patterns. This synthetic data is combined with real operational records to enhance the diversity and volume of the training set. Then, a Long Short-Term Memory (LSTM) model is used to capture long-range temporal dependencies, while XGBoost is applied to learn nonlinear and feature-driven relationships within the data. Finally, a hybrid fusion strategy based on both weighted blending and stacking regression combines the strengths of the two models. Experimental evaluations demonstrate that the hybrid model significantly outperforms individual models across all metrics. The stacking-based hybrid approach achieves the best performance with RMSE = 12.84 MW, MAPE = 4.21%, and R² = 0.965, outperforming LSTM and XGBoost by substantial margins. Additionally, the integration of TimeGAN reduces forecasting errors by up to 18% and improves generalization, highlighting its effectiveness in addressing data scarcity and privacy constraints. The results confirm that the proposed TimeGAN–LSTM–XGBoost framework provides a robust, scalable, and highly accurate solution for short-term spinning reserve forecasting, with strong potential for real-world deployment in power system operation and energy markets.