Abstract: The rapid expansion of nuclear power in China has led to a significant increase in both operational and under-construction reactor capacity. To ensure grid stability and efficient power consumption, nuclear power plants must engage in peak load regulation. The nuclear reactor core represents a complex, nonlinear, and multivariable coupled system. During peak load regulation, the core’s poisons oscillate heavily, leading to axial offset (AO), which in turn affects core safety. Currently, pressurized water reactors (PWRs) in China predominantly implement constant axial offset control (CAOC) strategy during peak load regulation periods. Therefore, to achieve constant AO operation, it is essential to control the AO by adjusting the control rod to ensure the safety of the reactor. However, during the current peak load regulation processes, discrepancies between the AO values calculated by core physics software and those measured in practice hinder operators from strictly adhering to predetermined regulation plans. Consequently, the actual process relies heavily on operators’ personal experience, which introduces decision-making and human-factor risks, and is not conducive to core safety. In order to reduce decision-making and human-factor risks for operators during peak load regulation in Chinese PWR nuclear power plants, and to achieve the purpose of assisting operators during peak load regulation or partial automating peak load regulation, this paper investigated a fast search method for control rod positions of PWR nuclear power plants based on convolutional neural networks (CNN). The Inception-ResNet structure was used to build the CNN, which was trained using measured AO data from the peak load regulation process and core operational parameters generated by the core physics software LOCUST/SPARK. A high-precision AO prediction model was developed, followed by a fast control rod position search method during peak load regulation. The method was validated using data from 26 peak load regulation processes of the M310 unit at a commercial PWR in China. The verification results indicate that the average deviation between the measured AO and the predicted AO values for the AO prediction model developed in this study is −0.084%. The average absolute deviation is 0.591%, the root mean square deviation is 0.836%, with the 95% confidence interval of −1.746%, 1.578%. Each AO prediction takes approximately 40 ms. When the error between the predicted AO and the target AO does not exceed 1%, the predicted R rod position closely matches the actual position, with a maximum error of only 3 steps. The proposed method demonstrates high accuracy in predicting rod positions during peak load regulation. Therefore, this fast search method for control rod positions during peak load regulation achieves both rapid and accurate results, supporting operators or enabling partial automation of peak load regulation, and shows potential for practical engineering applications.