Multiscale Temporal-Variable Patches Network for Long-Term Prediction of Industrial Key Parameters

In the face of industrial time-delay characteristics, long-term prediction of key parameters is pivotal for industrial optimization and control. However, traditional soft sensor models encounter difficulties in addressing long-term prediction. In industrial processes, long-term prediction typically encounters three challenges. First, the complex patterns resulting from multiple factors and stages pose modeling difficulties. Second, random errors in measurements hinder the extraction of reliable long-term dependencies. Last, the misalignment of variables in the temporal dimension constrains the utility of other variables in assisting prediction. To address these challenges, we propose a multiscale temporal-variable patches network (MS-TVPNet). First, resampling is conducted to extract multiscale dynamic features. Subsequently, for each scale, complex seasonal features are separated and treated as individual patches, which are then fed into Multihead Attention to adaptively capture global dependencies among variables. Notably, a novel cross-scale redundancy optimization mechanism is proposed to mitigate information overlap. Furthermore, variable-centered features are segmented into patches, and both interpatch and intrapatch features are captured using bidirectional gated recurrent unit (BiGRU) and convolutional neural network (CNN), respectively, for modeling long-term trends and short-term fluctuations. Finally, multilayer perceptron (MLP) is employed to extract trend features directly, which are then combined with seasonal features to obtain prediction results. MS-TVPNet uses patches to model in both the variable and temporal dimensions, effectively leveraging the continuity and correlation of industrial data to mitigate interference from random noise. Two industrial datasets are used for extensive experiments to demonstrate the effectiveness of the proposed method.