The rare earth extraction process has significant time delay characteristics, making it challenging to identify the time delay and establish an accurate mathematical model. This paper proposes a multi-delay identification method based on improved time-correlation analysis. Firstly, the data are preprocessed by grey relational analysis, and the time delay sequence and time-correlation data matrix are constructed. The time-correlation analysis matrix is defined, and the norm quantifies the correlation degree of the data sequence. Thus the multi-delay identification problem is transformed into an integer optimization problem. Secondly, an improved discrete state transition algorithm is used for optimization to obtain multi-delay. Finally, based on an Neodymium (Nd) component content model constructed by a wavelet neural network, the performance of the proposed method is compared with the unimproved time delay identification method and the model without an identification method. The results show that the proposed algorithm improves optimization accuracy, convergence speed, and stability. The performance of the component content model after time delay identification is significantly improved using the proposed method, which verifies its effectiveness in the time delay identification of the rare earth extraction process.
Given the complex and prolonged industrial processes involved in rare earth production, including the extraction and separation stages, the utility of short-term price predictions is limited due to the extensive times required to adjust production schedules. Consequently, accurately forecasting the long-term price trends of rare earth products is a pressing challenge. To address this, this paper introduces a VMD-SRF hybrid model tailored for Long Sequence Time-Series Forecasting (LSTF). To simplify the complexity of the initial data and improve the model's predictive accuracy, Variational Mode Decomposition (VMD) is first employed to analyze the periodicity and random components in price time series; Then, it combines the Series Random Forest model, which is improved based on the Random Forest (RF) algorithm. Series Random Forest (SRF) model uses Dynamic Time Warping (DTW) distance as heuristic information to address the deficiencies of Random Forest in long time series forecasting. This hybrid approach, leveraging the strengths of both VMD and SRF, enhances the handling of LSTF issues. An experimental comparative analysis using four representative datasets of rare earth product prices indicates superior prediction accuracy of the proposed method. These advancements present a promising and applicable strategy for addressing LSTF challenges in various practical settings.
Long short-term memory (LSTM) networks demonstrate superior time-series feature extraction capabilities and have exhibited significant advantages in the soft sensing of key indicators in complex industrial processes. However, conventional LSTM networks rely solely on the output information from forward propagation through network units, neglecting the residual information between the LSTM cell outputs and the key indicators. Moreover, unidirectional LSTM networks fail to fully exploit the inherent bidirectional temporal dependencies in industrial data. These issues lead to excessive redundancy in the features learned by the network and suboptimal prediction efficiency. This paper proposes a novel dual residual-enhanced deep bidirectional LSTM (DResBiLSTM) framework that integrates bidirectional temporal modeling and dual residual learning for the soft sensing of key variables in complex industrial processes. Firstly, residual information derived from the discrepancy between previous network outputs and key indicators is introduced into the input of the traditional LSTM cell, thereby constructing a residual bidirectional LSTM (ResBiLSTM) network. Secondly, a deep neural architecture is established using residual structures to incorporate input variable residuals, enabling effective soft sensing of key industrial indicators. This framework simultaneously extracts and utilizes latent features characterized by nonlinearity and dynamics from both process and quality variables, significantly enhancing prediction performance. Finally, through both numerical simulations and experimental validations employing real-world operational data from the LaCe/PrNd solvent extraction process, the proposed method demonstrates superior predictive accuracy and better practical effectiveness compared to existing soft sensing approaches.
24 June 2025 in International Journal of Machine Learning and Cybernetics