Figure 2 – MDN-RNN dynamics model: latent-state forecasting conditioned on action and exogenous histories. Probabilistic digital twin (CVAE + MDN-RNN): from observed windows to multihorizon forecasts. 6. RESULTS After training and validation, the proposed CVAE+MDN-RNN digital twin was evaluated on previously unseen continuous segments from January 2026, a period that includes changes in ore type at the SAG mill feed (Figure 3). Figure 3 – Ore-type transitions at the SAG mill feed during the January 2026 test period (unseen data). The objective of this evaluation is to assess whether the learned representation and latent dynamics remain reliable across different operating scenarios and ore regimes. Across the 13 modeled variables, most signals achieve MAPE below 10% (Table 1), indicating that the probabilistic twin provides plant-relevant accuracy for short-horizon forecasting and scenario comparison. Variables with higher error remain informative as trend indicators and may still contribute to downstream decision workflows depending on their operational role. To complement aggregate metrics, we visualize representative 3-hour test excerpts (31/01/2026, approximately 20:10–23:10), where the plant transitions between two ore types (Figure 4). For clarity in the visual comparison, a moving-average smoothing is applied to the real signals for throughput and pressure in these plots. Table 1 - Multihorizon forecasting performance on smoothed signals (validation-calibrated): MAE and MAPE for 5, 15, and 30 minutes ahead.
RkJQdWJsaXNoZXIy MTM0Mzk2