Net load forecasting for high renewable penetration Grids: Hybrid decomposition and Multi-Scale temporal convolutional network

To address the high volatility and non-stationarity of net load series in power grids with high renewable energy penetration, a novel deep forecasting framework based on a hybrid signal decomposition strategy and an attention gated multi-scale temporal convolutional network is proposed. By integrating a two-stage decomposition strategy with a sliding window mechanism, this framework achieves the decomposition and feature reconstruction of high-frequency random noise and low-frequency trend modes from the original series, while strictly preventing future information leakage. Subsequently, the proposed model integrates modules such as multi-scale convolutional networks and component attention reconstruction to realize adaptive feature extraction and dynamic fusion, thereby enhancing the model’s responsiveness to abrupt load changes. Empirical evaluation is conducted using actual data from a provincial power grid in China. The results demonstrate that the proposed model outperforms baseline models across all evaluation metrics on datasets covering four seasons; for instance, the root mean square error on the summer dataset is reduced by approximately 3.4% compared to the second-best model, effectively mitigating seasonal fluctuation differences. Furthermore, attention visualization reveals the model’s hierarchical decision-making mechanism, which utilizes the original series as a global basis while employing multi-scale modes for local fine-grained refinement. In terms of practical application, the model reduces prediction uncertainty during high-risk ramping periods and theoretically decreases the demand for spinning reserve capacity, providing robust support for the safe and economic operation of the power grid.