within plant-approved boundary conditions. Performance is reported against key KPIs: Cu recovery, raffinate Cu concentration, electrolyte-purity metrics, and pH behavior. The first principle -based SX model produces baseline outputs; deviations vs. plant measurements are learned by a residual model; a compensation layer applies feasible corrections to yield hybrid predictions. Inputs: plant operating variables, assays, historian data. Outputs: corrected hybrid predictions with uncertainty and diagnostics. Figure 2 – Digital twin workflow for training 3.4 Real-Time Inference in Plant (a) Data acquisition and connectivity. Live plant inputs (flows, O/A, temperatures, online pH, laboratory assays as available) are streamed via open industrial protocols (e.g., OPC UA, MQTT; historian APIs) into the first principle model and the residual model. (b) First principle baseline and residual inference. The first principle model computes baseline predictions consistent with constraints and extrapolates safely beyond the historical domain. The residual model, now in inference mode, estimates deviations given current inputs and selected first principle features. (c) Residual and final digital twin outputs. The compensation layer combines first principle outputs and predicted residuals to produce hybrid predictions in real time: ℎ = ( ℎ , ), applying feasibility checks to maintain operational safety (e.g., electrolyte composition limits, equipment envelopes). (d) Optimization Layer evaluates the multiple operating scenarios and send to Residual layer to Hybrid outputs are published to the Operator Assistant (Section 3), enabling scenario analysis (what-if O/A, flow splits), ranked recommendations (expected KPI deltas with confidence), and alerts when conditions deviate from expected corridors. The human-in-the-loop review ensures safe execution and accountability (e) Edge execution and resilience. The twin executes at the edge layer to minimize latency and to remain resilient to network interruptions. Buffered operation handles laboratory assay latencies; when assays arrive, the model performs light recalibration (e.g., residual bias correction) without interrupting inference.
RkJQdWJsaXNoZXIy MTM0Mzk2