Improving Deep Learning Performance on Small Datasets via Similar Pretraining Data Generation

With the increasing global emphasis on environmental sustainability and renewable energy, solar power has emerged as a prominent method of energy generation. Accurate forecasting of solar energy output is essential for optimizing energy management and grid scheduling. However, in real-world applications, data availability is often limited due to constraints such as equipment costs and time-consuming data collection, making it challenging to train reliable predictive models. This study proposes a data augmentation framework tailored for small-scale time series datasets. The framework generates synthetic time series with similar statistical properties to the target data for model pretraining, overcoming learning limitations from insufficient real data to improve predictive performance. Additionally, this study leverages a channel-independence strategy proposed in prior research, allowing the proposed method to flexibly handle both univariate and multivariate forecasting tasks. A fusion module is also designed to integrate related features, further enhancing the model's ability to capture temporal dynamics. To validate the proposed approach, we conduct experiments on a real-world solar power generation dataset. Results demonstrate that, under data-scarce conditions, the pretraining method significantly improves performance across various model architectures, highlighting its potential and practical value in small-scale time series forecasting scenarios.