Short-Term Load Forecasting Using Hybrid LSTM-ANN for Agricultural VSG-based Microgrids

The rapid electrification of agricultural activities increases the need for reliable and low-carbon rural power systems. Renewable-based agricultural microgrids integrating photovoltaic generation and battery energy storage require accurate short-term load forecasting to ensure efficient energy management and stable virtual synchronous generator operation. This study proposes a hybrid long short-term memory-artificial neural network architecture for short-term load forecasting and embeds it within a predictive three-layer framework for agricultural microgrids. The originality of the approach lies in combining recurrent temporal feature extraction with nonlinear feedforward mapping and coupling the forecasting output directly with adaptive virtual inertia and damping control. The model is trained on real operational data and validated across heterogeneous power systems to evaluate robustness and transferability. For 24-hour ahead forecasting, the proposed model achieves a Mean Absolute Percentage Error of 2.892%, outperforming artificial neural networks (4.796%) and standard long short-term memory models (3.305%), corresponding to improvements of 39.8% and 12.2%, respectively. Results demonstrate that the hybrid structure delivers improved multi-step forecasting accuracy and stable prediction trajectories.