In the aluminum electrolysis production process, the traditional cell control method based on cell voltage and series current can no longer meet the goals of energy conservation, consumption reduction, and digital-intelligent transformation. Therefore, a new digital cell control technology that is centrally dependent on various process parameters has become an urgent demand in the aluminum electrolysis industry. Among them, the real-time online measurement of alumina concentration is one of the key data points for implementing such technology. However, due to the harsh production environment and limitations of current sensor technologies, hardware-based detection of alumina concentration is difficult to achieve. To address this issue, this study proposes a soft-sensing model for alumina concentration based on a long short-term memory (LSTM) neural network optimized by a weighted average algorithm (WAA). The proposed method outperforms BiLSTM, CNN-LSTM, CNN-BiLSTM, CNN-LSTM-Attention, and CNN-BiLSTM-Attention models in terms of predictive accuracy. In comparison to LSTM models optimized using the Grey Wolf Optimizer (GWO), Harris Hawks Optimization (HHO), Optuna, Tornado Optimization Algorithm (TOC), and Whale Migration Algorithm (WMA), the WAA-enhanced LSTM model consistently achieves significantly better performance. This superiority is evidenced by lower MAE and RMSE values, along with higher R² and accuracy scores. The WAA-LSTM model remains stable throughout the training process and achieves the lowest final loss, further confirming the accuracy and superiority of the proposed approach.

To address the challenge of the real-time monitoring of alumina concentrations during the production process, this paper employs a Deep Belief Network (DBN) within the framework of deep learning to predict alumina concentration. Additionally, the improved Grey Wolf Optimizer (IGWO) is utilized to optimize key parameters of the DBN model, including the number of hidden layer nodes, reverse iteration count, and learning rate. An IGWO-DBN hybrid model is then constructed and compared against DBN models optimized by other techniques, such as the Sparrow Search Algorithm (SSA) and Particle Swarm Optimization (PSO), to evaluate the predictive performance. The comparative analysis reveals that, in terms of predictive accuracy, the IGWO-DBN model outperforms both the SSA-DBN and PSO-DBN models. Specifically, it achieves lower root mean square errors (RMSE) and mean absolute errors (MAE), alongside a higher coefficient of determination (R²). Furthermore, the IGWO-DBN model exhibits a faster convergence rate and a lower final convergence value, indicating superior generalization ability and robustness. Furthermore, the IGWO-DBN model not only demonstrates significant advantages in prediction accuracy for alumina concentration but also substantially reduces model training time through its efficient parameter optimization mechanism. The successful implementation of this model provides robust support for the intelligent and refined management of the aluminum electrolysis industry, aiding enterprises in reducing costs, improving production efficiency, and advancing the green and sustainable development of the industry.

Dissolved alumina concentration in the electrolyte cannot be measured continuously in aluminium reduction cells. Recursive soft sensing estimates the concentration from online process variables; however, process-condition shifts and the write-back of previous estimates can change the inputs to later predictions before the next laboratory concentration measurement is available. This study develops a dual-timescale three-state output supervisor for a published 730-sample sequence from one 400 kA aluminium reduction cell, using working current, cell voltage, electrolyte level, aluminium level, and electrolyte temperature as online inputs. Unlike monitoring methods that report process or model change, the supervisor assigns a use state to each recursive estimate. Process-variable deviation, adjacent prediction change, predictor-corrector discrepancy, and a cumulative sum (CUSUM) of process deviation are combined by state rules to determine Warning and Unavailable transitions. The observed later interval (samples 601–730) was evaluated by continuous recursive replay, while separate gradual-shift tests were synthetically constructed from the observed variable ranges. The supervisor entered Warning and Unavailable after 6.2 and 19.0 samples, on average. At matched prediction availability, it reduced the out-of-tolerance rate among available predictions by 15.81 percentage points and increased the failure blocking rate, defined as the fraction of all out-of-tolerance predictions assigned Unavailable, by 15.59 percentage points relative to random blocking. The method assigns Normal, Warning, and Unavailable use states to recursive estimates to support reference-measurement requests and output management.