An Adaptive Learning Rate Based Novel Recurrent Neural Network Modeling and Control of Complex Non‐Linear Dynamical Systems

ABSTRACT A Double Internal Loop Recurrent Neural Network (DILRNN) with an adaptive learning rate is proposed for the modeling and control of non‐linear dynamical plants. The structure of DILRNN is an alteration of the Fully Connected Recurrent Neural Network (FCRNN). DILRNN contains three feedback loops taken primarily from the context layer to the hidden layer, the time delay of the output layer to the hidden layer, and the time delay of output to output. The parameters of DILRNN are updated using the gradient descent‐based dynamic Back‐Propagation (BP) algorithm. The Adaptive Learning Rate (ALR) scheme is implemented to ensure that the learning rate value is determined properly in each iteration and improves the performance of the learning algorithm. The effectiveness of the suggested strategy is assessed by considering both the Multiple Input Single Output (MISO) and Multiple Input Multiple Output (MIMO) systems and comparing them with state‐of‐the‐art methods. The simulation outcomes show that the proposed model outperforms the Feed Forward Neural Network (FFNN) and other Recurrent Neural Network (RNN) models, which were taken into evaluation in terms of their output and error. Next, a DILRNN controller is implemented using the proposed model to control non‐linear dynamical systems. The controller's response is evaluated both in the absence and presence of a disturbance signal to assess the recovery capability of the controller. The response and error of the proposed controller are compared with other neural network controllers and elementary PID controllers.