System identification method based on residual-whitening constrained LSTM

Accurate system modeling is essential for complex industrial process control and optimization, yet strong nonlinearity and coupling pose significant challenges to both first-principles and data-driven methods. This paper proposes a system identification method based on residual-whitening constrained LSTM, embedding the classical residual-whiteness criterion as an autocorrelation regularization term to jointly minimize prediction errors and suppress residual correlations. Evaluated on three benchmarks—CE8, Silver Box, and Cascaded Tanks—against LSTM, GRU, RNN, TCN, NSI, Polynomial NFIR, Hammerstein, and Neural ODE, RW-LSTM achieves competitive or superior accuracy and residual-whiteness performance. Further experiments verify hyperparameter sensitivity and offer statistically supported finite-lag residual-whiteness validation. In addition, the method features reasonable tuning cost and good practical deployability. The results demonstrate that incorporating residual-whiteness constraints into deep learning offers a principled and effective pathway toward statistically well-behaved system identification.