Speed Estimation in Induction Motor Drive using Long Short-Term Memory-Sensorless Model Reference Adaptive System (LSTM-SMRAS) Control

Accurate rotor speed estimation is significant for achieving high-performance control in sensorless induction motor drives, especially under dynamic and low-speed operating conditions. This research proposes a new speed estimation method based on a Sensorless Model Reference Adaptive System (SMRAS) enhanced with Long Short-Term Memory (LSTM) neural networks. By integrating deep learning capabilities into the traditional MRAS approach, the proposed LSTM-SMRAS model significantly improves the robustness and accuracy of rotor speed prediction. Trained on historical motor performance data, the LSTM model captures nonlinear behaviors and transient dynamics more effectively than conventional estimators. The proposed method has been validated through extensive MATLAB/Simulink simulations across various operating scenarios, including step and ramp inputs at both high (1400 RPM) and low (200 RPM) speeds. Compared to existing MRAS techniques such as ANN-MRAS, FLC-MRAS, SMC-MRAS, and ANFIS-MRAS, the LSTM-MRAS method achieved the shortest delay time (0.47 s), rise time (0.85 s), and settling time (2.02 s) with a minimal peak overshoot of 0.21% at 1400 RPM. At 200 RPM, it maintained superior performance with a peak overshoot of only 3.70%. Under ramp input conditions, LSTM-MRAS exhibited the lowest steady-state error of 0.98%, significantly outperforming other methods. Furthermore, hardware implementation using a TMS320F28335 DSP controller confirmed the simulation results, with improved response time and reduced torque ripple. The FOC-SVPWM technique yielded a voltage total harmonic distortion (THD) of 2.2% and a current THD of 1.6%, meeting the IEEE 519 standards. These findings highlight the LSTM-MRAS approach as a reliable, sensorless control solution for precision-critical applications such as electric vehicles, robotics, and renewable energy systems.