421 Because the focus of this paper is to demonstrate forecasting and scenario-simulation capability, the analysis follows a deployment-consistent validation logic. The 30-minute horizon is generated via a rollout from the 1-minute dynamics model; and, to isolate the effect of the scenario intervention, teaching forcing is applied to actions and exogenous variables (i.e., the model is conditioned on their known trajectories over the horizon), while the observations evolve according to the learned process dynamics. The results indicate that the twin responds with operationally coherent behavior under the imposed changes. For throughput (TPH), the what-if scenario yields a sustained increase relative to the baseline, with a positive mean ΔTPHand P10–P90 bands that quantify the uncertainty of the expected impact (Figure 10a). In parallel, discharge pressure shows a systematic reduction under the what-if changes, consistent with expected responses to solids/water and speed adjustments (Figure 10b). In both cases, the baseline vs whatif comparison provides not only an expected difference, but also a plausible range of outcomes, supporting pre-action evaluation within short horizons (5–30 min), particularly under rapidly varying operating conditions. Figure 10 – 30-min what-if vs baseline: (a) TPH forecast with ΔTPH and P10–P90 bands; (b) discharge pressure forecast with ΔP and P10–P90 bands. 7. DISCUSSION The results show that the probabilistic CVAE + MDN-RNN twin consistently captures the primary grinding dynamics over 5–30 minute horizons, with performance suitable for operational decision support on several critical variables. Although nine signals achieve MAPE < 10%, this paper discusses only two representative cases due to space constraints; the remaining metrics are reported in the Results section. Methodologically, the explicit representation → dynamics split is key in a noisy, non-stationary industrial environment: the CVAE learns a compact and robust latent state from multivariate observation windows of the primary grinding line, and the MDN-RNN models the evolution of that state conditioned on actions and exogenous inputs. This enables not only point forecasts but also state-dependent uncertainty (P10–P90 bands), which typically widens during regime transitions (e.g., changes in ore characteristics or feed conditions) and narrows during steadier periods, providing risk-aware information for decision-making. Performance varies by signal type. Variables that are more directly coupled to actions/setpoints and exhibit smoother minute-scale dynamics tend to yield lower errors (e.g., throughput, pressures, power,
RkJQdWJsaXNoZXIy MTM0Mzk2