Deep reservoirs have shown an advantage in synergistic oil production, geothermal development, and carbon sequestration, with great economic potential. To solve the problems of high-dimensional variables and poor convergence in the collaborative optimization of these three processes, this paper proposes a new algorithm called GRNN-assisted Surrogate Optimization Algorithm (GSOA) based on the generalized regression neural network (GRNN) classifier and the surrogate model co-optimization. GSOA enhances search space diversity via GRNN hierarchical prescreening and combines a Radial Basis Function surrogate model with Differential Evolution for efficient global optimization, mitigating computational inefficiency and local optima in high-dimensional problems. A numerical model for medium-high temperature reservoirs (56 × 45 × 7 grid, 115 °C, 45 MPa) was established, optimizing bottom hole pressures for net present value (NPV) maximization while balancing oil recovery, CO2 storage, and geothermal energy production. Simulations demonstrate that GSOA outperforms traditional surrogate-assisted optimization (SOA), achieving a 3.9% increase in cumulative oil production (284.9 × 103 m3 vs 274.3 × 103 m3), 5.2% higher CO2 storage (65.3 × 106 m3 vs 62.1 × 106 m3), 0.9% enhanced heat extraction (2.72 × 1013 J vs 2.69 × 1013 J), and 3.2% greater NPV (1.21 × 108 CNY vs 1.17 × 108 CNY). GSOA converges to peak NPV in 55 generations—52% faster than SOA (115 generations)—validating its robustness in CCUS (Carbon Capture, Utilization and Storage)-integrated geothermal systems. This research advances methodologies for CO2-enhanced oil recovery-geothermal synergy and supports carbon-neutral resource utilization.