A Novel Three-Phase Smart Inverter Based on Long Short-Term Memory Network for VAR Compensation

As the proportion of renewable energy generation in the power grid continues to increase, integrating it into the grid will have a significant impact on the existing power system. Therefore, countries around the world have introduced strict standards for renewable energy grid integration, especially when grid faults occur. Renewable energy sources must be able to inject reactive power to maintain grid voltage at required levels via smart inverters in case low voltage occurs. In the past, inverter control primarily relied on Proportional Integral (PI) controllers, and parameter tuning was often done through trial and error or empirical methods. This approach was time-consuming, costly, and lacked robustness. Long Short-Term Memory (LSTM) networks, on the other hand, offer advantages of recurrent neural networks while addressing the issues of vanishing and exploding gradients. This paper utilizes LSTM networks followed by fully connected layers as the primary controller for the inverter. The paper uses both an outer loop and an inner loop to off-line optimize the LSTM network controller, replacing traditional PI controllers. The inner loop employs the Adam optimizer to optimize the weights and biases of the LSTM network and the fully connected layers, while the outer loop uses the Particle Swarm Optimization algorithm to optimize the initial learning rate and batch size of Adam optimizer and the number of hidden units in the LSTM network. Through simulation and experimentation, the proposed method achieved smaller averages of Root Mean Squared Error (0.0243) and Mean Absolute Error (0.0106) compared to those (0.0407 and 0.0330) obtained by the traditional PI controller, respectively. These results indicate that the proposed method outperforms traditional PI controllers.