ST-GAIN+：太陽光日射量データ欠損補完のための時空間生成敵対ネットワーク

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:Provides a robust imputation method with uncertainty estimation for solar data, useful for renewable energy forecasting research.

🏢実務担当者:Can improve solar forecasting accuracy and grid optimization by handling missing data, beneficial for renewable energy operators.

<title>Abstract</title> <p>Missing data in solar irradiance datasets significantly degrades the performance of solar forecasting models and smart grid optimization systems. Traditional imputation methods often struggle to capture complex spatiotemporal dependencies and typically lack mechanisms for uncertainty estimation. To address these challenges, this paper proposes ST-GAIN+, a hybrid spatiotemporal generative adversarial network for missing solar irradiance data imputation. The proposed framework integrates Temporal Convolutional Networks (TCNs) to capture long-range temporal dependencies, dual attention mechanisms to model spatial and temporal correlations, and Bayesian uncertainty estimation based on Monte Carlo dropout to quantify predictive uncertainty. In addition, a GAIN-based adversarial learning framework with a hint mechanism is employed to improve realistic reconstruction of missing observations. The proposed model is evaluated under Missing Completely at Random (MCAR) conditions with missing rates of 30%, 50%, 70%, and 90% using multiple real-world solar irradiance datasets, including NASA POWER, NSRDB, and Nama_Oman. Experimental results demonstrate that ST-GAIN+ achieves stable and accurate reconstruction performance across multiple solar irradiance datasets under MCAR missing rates ranging from 30% to 90%. The framework consistently maintains low RMSE and MAE values even under severe data incompleteness. Comparative and ablation analyses indicate that temporal modeling, attention-based dependency learning, and uncertainty-aware estimation contribute to robust solar irradiance reconstruction. The findings suggest that the integration of adversarial learning, spatiotemporal modeling, and uncertainty-aware estimation provides a promising framework for reliable solar irradiance data imputation in renewable energy applications.</p>

• Research Square https://doi.org/10.21203/rs.3.rs-10487457/v1first seen 2026-08-26 04:20:23 · last seen 2026-09-08 04:21:54