413 DIGITAL TWIN OF A SAG MILL: MULTIHORIZON FORECASTING (5-30 MINUTES) WITH QUANTIFIED UNCERTAINTY (10TH-90TH PERCENTILE BANDS) *Renato Javier Martinez Ordonez, J. Campos, Alfonso Munoz Veron Minera Chinalco Peru S.A., Peru (*Presenting author: rmartinez@chinalco.com.pe) ABSTRACT This paper presents a probabilistic digital twin based on deep learning for a semi-autogenous grinding (SAG) mill. The twin provides multihorizon forecasts (5–30 minutes) and what-if scenario analyses with quantified uncertainty via 10th–90th percentile bands (P10–P90). The approach combines a convolutional variational autoencoder (CVAE) to learn nonlinear latent process states from multivariate plant observations, and a mixture-density recurrent neural network (MDN-RNN) to model latent dynamics conditioned on recent history and operator setpoints (e.g., mill speed, percent solids, and water inflow). In industrial practice, short-horizon consequences of setpoint changes are rarely quantified before acting and deterministic tools predominate; our model delivers probabilistic forecasts suitable for anticipative decision support and for integration with existing expert logic. Key novelties are: (i) a representation-to-dynamics split (CVAE → MDN-RNN) that improves robustness to noise and collinearity; (ii) operational use of P10– P90 bands to compare alternatives before acting; and (iii) a deployment-oriented validation protocol with chronological training/validation/test splits, per-type normalization, and affine calibration. On unseen test data (1-minute sampling; 30-minute effective input history), at a 30-minute forecast horizon we achieve MAPE ≤ 5% for throughput (TPH), feed fines/intermediates, percent solids, fill level, SAG power, feed pressure, and discharge pressure; MAPE 5–10% for the coarse fraction; and MAPE > 10% for impact intensity, load width, toe position, and pebble tonnage. Variables with MAPE < 10% are deemed suitable for short-term decision support within 5–30 minutes, while higher-error variables remain useful for trend monitoring. A what-if case reflecting a common manual operating adjustment (−2% solids, +2% water inflow, −5% SAG speed, and +3% throughput setpoint) yields an estimated ΔTPH ≈ +46 t/h with P10–P90 [+19, +73] t/h and a discharge pressure change of −83 kPa with P10–P90 [−142, −28] kPa relative to a nochange baseline, communicating both expected impact and uncertainty. The current scope corresponds to pre-deployment forecasting and what-if evaluation; the same probabilistic twin can be coupled to the plant’s expert logic to provide model-based predictive capabilities for closed-loop operation in follow-up work. KEYWORDS SAG mill; digital twin; multihorizon forecasting; uncertainty quantification; P10-P90; what-if scenario analysis; convolutional variational autoencoder; mixture density network; probabilistic forecasting. 1. INTRODUCTION Grinding circuits are among the highest energy-consuming and economically critical units in concentrator plants. In particular, SAG mills exhibit strong nonlinearities, multivariable coupling, and timevarying behavior driven by ore variability, liner wear, and downstream constraints. Operationally, the key objective is to maximize (or sustain) throughput while respecting limits on pressure, power, pebble load, impacts, and product size distribution; however, the feasible optimum shifts as ore and plant conditions evolve. Decision support is often built on deterministic, offline models (e.g., operating envelopes or expertsystem tuning). When deployed online, these models can degrade because they learn regime-specific correlations. Prior work on NARX-based SAG digital twins reports horizon-dependent error growth due to autoregression and the practical need for smoothing, disturbance detection, and retraining to maintain performance under changing conditions [4]. Other approaches target critical-condition prediction (e.g.,
RkJQdWJsaXNoZXIy MTM0Mzk2