Cross-Scale Temporal Representation Learning for Short-Term Wind-Solar Energy Generation Regime Classification
- Authors
-
-
Haoyang Lu
Columbia University, New York, USA
Author
-
- Abstract
-
The increasing penetration of wind and solar energy introduces substantial variability and uncertainty into modern power systems, creating a growing demand for reliable identification of future renewable-energy generation states. This study proposes a dual-branch cross-scale temporal learning framework for short-term wind-solar energy generation regime classification. Rather than directly predicting continuous power values, the proposed approach transforms future renewable-energy production into low-, moderate-, and high-generation regimes, providing a more interpretable representation for operational decision support. To preserve the heterogeneous temporal characteristics of wind and solar power, two source-specific branches are employed to independently encode their historical sequences. A cross-scale temporal learning mechanism is then introduced to capture short-term fluctuations and longer-range dependencies, while a cross-energy interaction module explicitly models the complementary relationship between wind and solar generation before final classification. Experiments are conducted on a public multi-year hourly wind and solar energy production dataset under chronological data partitioning to avoid information leakage. The proposed framework achieves competitive overall performance across accuracy, precision, recall, and F1 score, while ablation experiments confirm the effectiveness of source-specific encoding, cross-scale temporal modeling, and cross-energy interaction. In addition, the training and validation loss curves demonstrate stable convergence, and the t-SNE visualization shows clear separability among different generation regimes in the learned feature space. These results indicate that the proposed method can effectively extract discriminative temporal representations from heterogeneous renewable-energy sequences and provides a practical deep-learning solution for renewable-energy generation state identification.
- Downloads
- Published
- 2025-03-01
- Issue
- Vol. 5 No. 3 (2025)
- Section
- Articles