Enhancing Temperature Forecasting for Sustainable Energy Systems Using CTGAN‑Based Data Augmentation and Temporal Convolutional Networks

Abstract

Accurate temperature forecasting plays a critical role in energy management, building operation optimization, and sustainable energy system planning. Reliable short-term temperature prediction can significantly improve the operational efficiency of heating, ventilation, and air conditioning systems, data centers, and smart energy infrastructures. This paper introduces a temperature prediction method that uses multiple weather prediction datasets to establish models for the maximum and minimum temperatures. A temporal convolutional network (TCN) is used for temperature prediction and a conditional tabular generative adversarial network (CTGAN) for data augmentation. Given the challenge of fitting CTGAN-generated synthetic data to the original data and the inefficiency of hyperparameter tuning for CTGANs, this study adopts a series of data preprocessing methods — outlier removal with a one-class support vector machine, dataset normalization, and the Yeo–Johnson transformation — that bring the distribution of the synthesized data close to that of the original data. The proposed TCN + CTGAN model with preprocessing consistently outperformed the TCN baseline and the unprocessed augmentation variant across all experimental configurations. On the Seattle dataset, the model achieved R² values of 0.829 and 0.819 for the maximum and minimum temperatures, with corresponding RMSE values of 3.039 and 2.136. On the Seoul dataset, R² values of 0.441 and 0.681 were obtained, with RMSE values of 2.346 and 1.395.

Key Results

Method

CTGAN is fitted inside each cross-validation fold, so synthetic samples never see the held-out test data. The contribution is the preprocessing stage between generation and training: it aligns the synthetic distribution with the real one without any CTGAN hyperparameter search.

Datasets

Two publicly available meteorological datasets with different climates, scales, and temporal coverage. Only the target temperature variable is fed to the model — adding further weather variables degraded accuracy, as the input configuration table below shows.

Aligning Synthetic Data with Real Data

Raw CTGAN output does not match the real distribution, and training on it makes forecasts worse. Each preprocessing stage is scored with the Wasserstein distance between the synthetic and original distributions — lower is closer.

Standardisation is by far the most impactful step. The Yeo–Johnson transformation optimises for normality rather than for transport cost, so it can raise the distance slightly (Seattle T_min, Seoul T_max) while still improving the shape alignment visible in the density plots.

Prediction Quality

Pooled predictions of the final TCN + CTGAN (processed) model across all ten folds.

BibTeX

@article{KUO2026109646,
  title   = {Enhancing temperature forecasting for sustainable energy systems using
             {CTGAN}-based data augmentation and temporal convolutional networks},
  author  = {Kuo, Ping-Huan and Lin, Yu-Sian and Chiu, Yu-Chih},
  journal = {Energy Reports},
  volume  = {16},
  pages   = {109646},
  year    = {2026},
  issn    = {2352-4847},
  doi     = {10.1016/j.egyr.2026.109646},
  url     = {https://www.sciencedirect.com/science/article/pii/S2352484726006153}
}