关键词: 海岛微电网, 负荷预测, 鲁棒经验模态分解, 细节增强卷积网络, 双向门控循环单元, 评价因子重构, 多任务学习

Abstract: A load forecasting method based on evaluation factor reconstruction using Robust Empirical Mode Decomposition (REMD) combined with Detail-Enhanced Convolutional Network (DECN) and Bidirectional Gated Recurrent Unit (BiGRU) is proposed for the complex characteristics of strong nonlinearity, non-stationarity, and multi-source coupling in island microgrid loads. Firstly, abnormal data are cleaned using a sliding window Z-score method, and key meteorological factors are screened by integrating Spearman, Pearson, and Kendall correlation coefficients. REMD is used to suppress mode mixing and decompose load sequences into multi-scale Intrinsic Mode Functions (IMFs). Evaluation factors are constructed based on sample entropy (complexity) and zero-crossing rate (frequency characteristics) to reconstruct critical components, removing noise and retaining multi-time-scale features. In the prediction model, DECN extracts local temporal difference features, BiGRU captures bidirectional temporal dependencies, and multi-task learning optimizes the coupling relationships of multiple loads. Validation with real data from the Taishan Island microgrid in Fuding, Ningde, Fujian, shows that the proposed model reduces the Mean Absolute Percentage Error (MAPE) to 1.18%, Mean Absolute Error (MAE) to 7.37 kW, and Root Mean Square Error (RMSE) to 9.69 kW compared with single DECN, BiGRU, and undecomposed models, significantly improving prediction accuracy and robustness in complex environments. This study provides an effective solution for accurate load forecasting and optimal power resource scheduling in island microgrids.

Key words: island microgrid, load forecasting, Robust Empirical Mode Decomposition, Detail-Enhanced Convolutional Network, Bidirectional Gated Recurrent Unit, evaluation factor reconstruction, multi-task learning