The Effect of the Reference Generator Utilization on the LSTM Based Speed and Load Torque Estimation of IMs

Deep learning (DL)-based solutions have been utilized in different research areas due to their capabilities, high performance, and reliability. The electrical vehicle application is one of the research areas in which the DL method is used. Since the performance of these methods depends on the training process, the training dataset should include all conditions or at least as many as possible. In some cases, generating a training dataset, including all conditions, is not likely for user-defined scenarios. Therefore, a random process with appropriate user-defined limit and time values can be used to obtain a comprehensive training dataset. From this point of view, to examine the effect of a reference generator ($R G$) on the estimation performance, firstly, an RG for induction motor drives is designed in this paper. Using the RG, the training and test datasets are obtained to be used in the long short-term memory (LSTM)based estimator. Here, the LSTM network utilizes the rotating axis components of the stator currents and voltages to estimate rotor mechanical speed and load torque. To demonstrate the effect of the utilization of the RG on the performance of the LTSM, it is compared to the LSTM estimator trained by user-defined dataset. Although both LSTM estimators are in the same structure, the simulation results show the superiority of the LSTM trained by the RG-based dataset.