<p>In the financial sector, research on realized volatility is highly regarded for its critical role in risk management and asset pricing. This paper constructs an innovative hybrid model and proposes an improved Crested Porcupine Optimizer (iCPO), which optimizes the initialization method of CPO. Simulation experiments conducted on five different representative functions demonstrate that the iCPO exhibits superior parameter search capabilities compared with CPO and four other innovative optimization algorithms. To enhance both computational efficiency and the ability to search for the global optimum, iCPO is employed to optimize the parameters of the Support Vector Regression (SVR) model, while a Generalized Regression Neural Network (GRNN) is introduced to model the error series and mitigate the lag effect in time series forecasting. The performance of the proposed hybrid model is then examined through a two-stage validation: first, using simulated data under different volatility scenarios to assess robustness; and second, by forecasting the realized volatility of five representative stocks from various sectors in the Chinese stock market, where it is compared with eight benchmark models, including HAR-RV, LSTM, RF, KNN, GCRA-SVR-GRNN, SBOA-SVR-GRNN, IVYA-SVR-GRNN, and AO-SVR-GRNN. The results show that the proposed model consistently outperforms the competing methods, highlighting its practical value in realized volatility forecasting.</p>