Surrogate-assisted optimisation of a dual-source geothermal organic rankine cycle under CO2/NCG variability: optimal split-ratio and setpoints

Geothermal Organic Rankine Cycle (ORC) plants are increasingly required to operate flexibly to support power systems with high penetrations of variable renewables. Off-design performance is strongly influenced by non-condensable gas (NCG) disturbances and operating constraints that introduce non-smooth response regions associated with phase-change limits and heat-transfer bottlenecks. This study develops a surrogate-assisted offline optimisation framework for a dual-source (flash-binary) geothermal ORC and quantifies how optimal setpoints shift under varying CO 2 /NCG disturbance intensity. A physics-consistent DWSIM (open-source process simulation) model generates a 5,000-point dataset spanning feasible operating regimes; a Random Forest surrogate achieves R 2 = 0.959, root mean square error (RMSE) = 283.4 kW and mean absolute percentage error (MAPE) = 0.39%, capturing the non-smooth thermodynamic response surface. The surrogate is embedded in a Powell-Conjugate Hybrid Evolution Strategy (PCH-ES) combining Differential Evolution for global exploration with Powell refinement for local exploitation, producing a pre-computed disturbance-indexed setpoint scheduling table. Results identify an optimal split-ratio (SR*) band: in the low-CO 2 regime the optimum lies near SR* = 0.671; above approximately 1.8% CO 2 it shifts to SR* = 0.710, providing a disturbance-aware design rule for distributing working fluid between steam- and brine-side heat recovery. Net power declines 13% across the studied disturbance range (27,612 to 23,921 kW) as a thermal-capacity effect. The optimisation also identifies a thermodynamic knee in working-fluid circulation near 315 to 320 kg/s, beyond which additional pumping yields diminishing returns. The resulting offline setpoint lookup provides actionable scheduling guidance for disturbance-aware supervisory operation without requiring real-time modelling software.