An LSTM-Driven Deep Learning Model for Real-Time Control and Optimization of Nano-Grid Systems Using AI Techniques

Abstract

The rise of renewable energy in small decentralized grids calls for smart and flexible ways to control the system for reliable, uninterrupted and economical operation. This research proposes a new hybrid model that uses LSTM networks coupled with ST-GNNs to carry out real-time forecasting and control in nano-grids. A key feature of the suggested model is its use of LSTM for temporal analysis and ST-GNN for modeling the interaction among grid components. An input space that includes data on PV production, battery SoC, inverter status, load requirements and real-time pricing is processed and fused into a single representation using a hybrid mechanism. A Jetson Nano platform was used to implement the model and it was put through real-world workload simulations, using data from several seasons. The results show progress over the original models, with an RMSE of 1.937 kW for load forecasting, 9.62% for MAPE, 82.7% for Renewable Energy Utilization and 1.95% for Loss of Power Supply Probability. The proposed system lowered the Levelized Cost of Energy (LCOE) to ₹3.26/kWh and achieved better results than 9 comparison models, including those based on ARIMA, SVR, LSTM, Transformer, CNN-LSTM architectures and others. Real-time use of a closed-loop feedback system allowed for continual verification of prediction accuracy and correction of the dispatch policy. Therefore, the new model offers a strong, flexible and scalable way to control the dynamic operations of an AI-based nano-grid system.