Fusion of Multiple Coarse-Grained Data for Time Series Forecast and Its Application in Joint Power Forecast of Source-Load

Time series prediction holds significant value in energy management, yet directly modeling fine-grained data (e.g., minute-level resolution) often introduces substantial errors due to noise interference and long-term dependencies. To address this, we propose a multi-granularity fusion framework. First, fine-grained data are decomposed into coarse-grained groups (e.g., hourly resolution) via signal decimation to capture complementary temporal patterns. Subsequently, a cross-granularity attention module and a granularity-based attention module are designed to adaptively map these groups into a semantic space, leveraging nonlinear equations to distinguish noise and extract purified features, thereby learning the prior distribution of raw data. Applied to joint power forecasting of wind, photovoltaic, and load, our method demonstrates superior performance over traditional models (reducing Root Mean Square Error(RMSE) by 30%-50%, achieving Coefficient of Determination(R²) up to 0.96) while maintaining robustness in 2-10 hour multi-step predictions. The results validate the effectiveness of multi-granularity strategies in noise suppression and trend enhancement, offering a novel approach for precise forecasting in complex energy systems.