Hai-Long Yang

Predictive models play a pivotal role in industrial control systems, particularly demonstrating extensive and significant applications in data processing and data mining. Currently, both the gated recurrent unit (GRU) and the Transformer play a crucial role in building predictive models. However, most of the research methods still face the problems of imperfect data processing and simple model construction. Regarding data processing, the lack of redundant feature removal inevitably complicates the model's learning process. In terms of model construction, existing methods simply combine the two without deeper exploration and combination. To address these challenges, we present TG-FCM, a new prediction model in this study. Specifically, this paper develops a feature selection method using AGglomerative NESting (AGNES) and fuzzy c-mean (FCM), and constructs a Transformer prediction framework with memory function with the help of GRU. In the feature selection stage, the data features are first clustered using AGNES to obtain clustering results, which are then used as initial cluster centers for FCM. This combination method effectively alleviates the problem of FCM's dependence on initial values. In the prediction framework part, GRU is utilized to have the ability of storing historical data to provide memory function for Transformer. The specific situation is to add GRU-based storage memory unit and extraction memory unit to the encoder and decoder of Transformer respectively, so as to store the memory in the encoding process and extract the memory in the decoding process, thus realizing the memory function of Transformer. The experimental analysis of TG-FCM and five comparative models on eight real datasets shows that TG-FCM has high prediction performance and better generalization ability.