Answers are grounded in this paper and cite their sources.
Highly irregular and heterogeneous microgrid loads challenge energy management systems (EMSs) and limit the generalizability of widely used time-series models, making accurate multi-step forecasting difficult. To address this, we propose a seasonal and trend decomposition using loess (STL)-based multi-model framework that matches tailored network architectures to decomposed load components according to their characteristics. We decompose the load series into three components: 1) a trend component, where a long short-term memory (LSTM) network with monthly labels captures long-term variations; 2) a periodic component, utilizing a dual-branch (workday/non-workday) Transformer with intra-interval labels and attention to extract daily/weekly cycles; and 3) a residual component, modeled by a multi-scale convolutional neural network (CNN) using ultra-short-term lagged loads, proximal feature extraction, and time-decay attention. The key novelty lies in explicitly matching tailored predictors to interpretable load components, rather than forcing a single model to learn heterogeneous dynamics or relying solely on empirical decomposition. Evaluated on a full-year practical dataset, our method significantly outperforms monolithic baselines including LSTM and temporal fusion transformer (TFT), and surpasses complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN). Our approach reduces mean absolute percentage error (MAPE) by 2.3%∼8.0% and mean absolute error (MAE) by 16.1%∼66.1%. It also performs well on public datasets (e.g., 29.3% lower MSE and 31.2% lower MAE than literature-reported approaches). When integrated into our EMS, these enhanced forecasts reduce daily electricity costs by 2.4%∼6.5% and carbon emissions by 1.5%∼3.6%. These findings highlight the broader significance of forecast improvement for reliable and environmentally friendly microgrid operation.
No takes yet. Share an insight, caveat, or question.
Gan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: