Event-triggered Relearning Modeling Method for System with Non-stationary Variable Operating Conditions

This paper proposes an innovative event-triggered relearning strategy for neural network modeling, to enhance prediction accuracy under the non-stationary and variable conditions frequently encountered in complex industrial environments. Firstly, the proposed approach divides the input and output datasets across different operational conditions, by applying a sliding window algorithm along with entropy to identify the boundaries of the input data. Following this, we utilize the prediction error from the current model, which is relevant to variable operational states, to construct an array of event-triggered conditions. Once these conditions are triggered, we employ the corresponding dataset to re-establish the model and predict the data under this operating condition. Compared with the traditional neural network model, the proposed model can produce a better estimation accuracy. Finally, numerical simulation tests are provided to verify the effectiveness of the proposed method. The results show that the proposed method can effectively improve the prediction accuracy of neural network model under non-stationary variable working conditions.