Improving Power Delivery of Grid-Connected Induction Machine Based Wind Generators under Dynamic Conditions Using Feedforward Linear Neural Networks

In the conventional grid-connected Wind Energy Conversion System (WECS), the generator side inverter is typically controlled via Field Oriented Control (FOC), while Voltage Oriented Control (VOC) controls the grid side inverter. However, robust operation cannot be guaranteed during sudden changes in wind speeds and weak grid connections. This paper presents a novel method to improve the overall dynamic performance of on-grid induction machine-based wind generators. An online mechanical parameter estimation technique is devised using Recursive Least Squares (RLS) to compute the machine inertia and friction coefficient iteratively. An adaptive feedforward neural (AFN) controller is also proposed in the synchronous reference frame, which is constructed using the estimated parameters and the system's inverse. The output of the neural controller is added to the output of the speed PI controller in the outer loop of the FOC to enhance the speed response of the wind generator. A similar approach is taken to improve the classical VOC structure for the grid-side inverter. In this case, the RLS estimates the equivalent Thevenin's grid impedance in real-time. As for the adaptive action, two identical neural networks are integrated with the inner loop direct and quadrature axis current PI controllers. Under nominal operating conditions, it is observed that the PI+AFN provides a faster settling time for the generator's speed and torque response. Upon being subjected to variations in the wind speed, the PI+AFN outperforms the classical PI controller and attains a lower integral-time error. In addition, the proposed PI+AFN controller has a better ability to maintain the grid-side inverter stability during stochastic variations in grid impedance. One significant advantage of the proposed control approach is that no data for training or validation is required since the neural network weights are directly the output of the RLS estimator. Hardware verification for the improved FOC for wind generators using the adaptive controller is also made using the DSPACE 1007 AUTOBOX platform.