Development of input variable selection and structural optimization algorithm for recurrent neural network

With the rapid development of industrial field data acquisition and storage systems, neural networks have been widely used to handle complex nonlinear processes and establish data-driven models. In these situations, the time-dependent correlation between process variables and target variables is essential. The performance of data-driven models will worse when the time-series characteristics are omitted. Considering this problem, the paper proposes an input variable selection and structure optimization algorithm for recurrent neural network (RNN) with nonnegative garrote (NNG). The recurrent neural network deals with the strong nonlinearity and dynamic characteristics of the industrial process by obtaining time information from the sequence data. The NNG algorithm shrinks the input weights of RNN and then selects the input variables and optimizes the hidden layers. By taking an artificial data set as an example and comparing it to other algorithms, the effectiveness of the developed algorithm is verified.