Real-time forecasting with deep learning surrogates in minerals processing

Real-time forecasting in continuous industrial processes shifts operations from reactive to proactive. By anticipating key process changes, plants can maintain stability, prevent disturbances, and optimize performance before issues arise. When combined with a digital twin, these forecasts enable scenario analysis, predictive maintenance, and smarter decision-making, turning live data into actionable insights for improved efficiency and reliability.

In the Business Finland AIMODE project, we have developed data-driven and physics-informed deep learning surrogate models for a mineral flotation process. These models provide multivariate, multi-step time-series forecast of plant output up to 30 minutes ahead. Training is based on offline data generated by a dynamic, physics-based simulation model (digital twin) calibrated to real plant behaviour through adaptation and control parameters from the actual process. After demonstrating strong performance on offline data, these surrogate models are now being adapted for continuous operation with streaming online data coming from the digital twin and the plant.

In this study, we present a pipeline that integrates the developed surrogate models to predict real-time output grades of the flotation process using data streams from the plant’s digital twin. The pipeline handles data preprocessing to match model input requirements, incorporates uncertainty estimation for predictions across multiple time horizons, and provides an interactive visualization of real-time forecasts. Future work will focus on developing more robust uncertainty estimation methods and integrating these real-time forecasting surrogates into process optimization workflows.