Informer Network Fusing Interpretability and Dynamic Frequency Denoising Without Information Leakage for Predicting Complex Systems

To accurately predict the output of complex systems with input noise, a deep Informer network is innovatively designed, which combines signal decoupled denoising and interpretable functions. ELasticNet is employed for fitting evaluation and principal component feature selection. The dynamic variational mode decomposition (VMD) technique is established to decompose the input sequence. The high-frequency signal with a certain weight is combined with the low-frequency signal to realize decoupling reconstruction and weaken noise. The sliding window strategy is constructed to regularly decompose and update the newly obtained data online, so as to overcome the information leakage problem. Informer is applied to reasonably divide and reconstruct the principal component feature sequence. Encoder and Decoder are used to realize feature capture under Embedding framework. In the Encoder layer, the correlation of sequence signals is extracted and activated by Multi-head Prob-Sparse Attention and wavelet activation function, respectively. The Feedforward Neural Network (FNN) is utilized to map the extracted features by combining with the intermediate output of Decoder. The combined results are analyzed globally using Multi-head Attention. In the Decoder layer, the masked Attention and one-dimensional convolution are combined to decode features, and the fully connected layer is utilized to obtain the prediction output. The Integrated Gradients (IG) is applied to analyze the global and local interpretability of the prediction results to reveal the differential preferences of the proposed models in capturing key features. Finally, the accuracy and applicability of the proposed network are verified in complex industrial systems by comparing with the existing networks.