Original Article

A temperature prediction method for rare earth electrolytic molten salt based on dual interference filtering and hybrid neural networks

College of Automation and Electrical Engineering, Inner Mongolia University of Science and Technology, Key Laboratory of Integrated Automation for Process Industry in Higher Education Institutions of Inner Mongolia Autonomous Region, Baotou 014010, Nei Mongol, PR China

^* e-mail: This email address is being protected from spambots. You need JavaScript enabled to view it.

Received: 5 February 2026
Accepted: 19 May 2026

Abstract

Molten salt temperature serves as a critical parameter for assessing the electrolytic reaction state in rare earth molten salt processes, and its precise prediction is fundamental to optimizing process control. Addressing challenges such as flame and material disturbances during electrolysis, as well as the insufficient accuracy of existing prediction models under dynamic disturbances, this paper proposes an intelligent molten salt temperature prediction method. This method first employs a Gaussian mixture model and Kalman filter to establish a dual “spatial-temporal” interference suppression mechanism, filtering out flame and material interference while smoothing temporal noise. It then integrates multi-color space information (RGB, HSV, Lab) to construct a 19-dimensional feature vector, comprehensively characterizing molten salt radiation properties. Building upon this foundation, a BO-Transformer-BiLSTM hybrid neural network is designed, achieving high-precision temperature prediction through self-attention mechanisms and bidirectional temporal modeling. Experiments demonstrate that the proposed GMM–Kalman preprocessing reduces the mean absolute error (MAE) from 11.6 °C to 1.85 °C, representing an approximately 84% performance improvement. The BO-Transformer-BiLSTM model further controls MAE between 0.55 °C and 0.60 °C, achieves an R² of 0.96, and attains an RPD of 8.27, with average prediction errors within 2 °C. Its overall performance significantly outperforms comparison models. This study provides an effective method for online molten salt temperature prediction in highly disturbed environments and offers a technical reference for visual temperature measurement in similar industrial scenarios.