Introduction. Maintaining high technical and economic performance in aluminum production is a pressing operational challenge. The use of artificial intelligence makes it possible to mitigate the impact of process irregularities such as the deformation of the anode bottom resulting in spike formation. The causes and consequences of spike formation have been thoroughly studied. Dozens of parameters signaling this disturbance are identified. However, existing solutions typically detect deviations only after they have occurred or predict related states. Industrial realities demand adequate, advance warning of the problem. The objective of this study is to develop and validate an intelligent model for the early prediction of spike type disturbances under aluminum production.

Materials and Methods. The study has analyzed over 70 process parameters based on average daily monitoring data from eight electrolyzers at a pilot site (2020–2025). The positive class was defined based on the pre-failure interval spanning 72 to 24 hours prior to the recorded disturbance. Ensemble methods were used for prediction: – gradient boosting (XGBoost, CatBoost, LightGBM); – FEDOT automated pipeline; – neural network ensembles (RTDL-RM, RTDL-NE, TabM).

Quality metrics: Accuracy, Precision, Recall, F1.

Results. The LightGBM model was identified as optimal, with the following metrics: Accuracy — 0.88; Recall — 0.79; F1 — 0.79. For events without disturbances, it correctly classified the “normal” state in 95% of cases (900 observations), while incorrectly identifying “disturbance” in 5% of cases (46 observations). For events involving disturbances, 62.5% of states were correctly identified (10 observations), whereas disturbances were missed in 37.5% of cases (6 observations). Validation using data from an actual technological process showed the model ability to detect spikes 24–72 hours prior to their actual registration. During the validation, six spike formation events were recorded, and the model detected signs of the developing disturbance in five of them (83.3%).

Discussion. The research results are consistent with the physical and technological characteristics of spike formation. The proposed solution complements existing methods for the local monitoring of anodes and electrolyzers through analyzing parameters derived from daily monitoring data. Validation confirms known challenges for predictive models in industrial electrolysis: difficulty in event detection, class imbalance, and changing process characteristics. The monitoring data from eight electrolyzers at the pilot site are insufficient. Further testing is required to confirm the model generalizability — over longer periods and at other sites.

Conclusion. The ability of the model to predict a disturbance 24–72 hours prior to the detection of a spike formation was confirmed, as was its sufficient effectiveness given the constraints imposed by sample size and the variability of process conditions. Future research will involve data augmentation, model validation at other production sites, and a comprehensive analysis of average daily and instantaneous monitoring parameters to localize the disturbance.