In order to improve the drilling efficiency and prediction accuracy, this paper designs a drilling speed modeling algorithm based on real-time data stream comparison. A prediction framework oriented to real-time updating and model reuse is constructed by comparing the data streams of the main well with nine historical wells. The system firstly utilizes the sliding window mechanism to dynamically slice the data of the main well, and extracts the data of the same depth section from the nine wells in a K-nearest neighbor manner to construct the reference dataset. Then, it combines the fast Fourier transform with the spectral similarity index to realize the frequency domain comparison between the main well window and the historical data. When the similarity is higher than a set threshold, the system reuses the historical model, otherwise it instantly triggers re-modeling. The modeling process adopts the Random Forest algorithm, fuses the cumulative window data of the main well with the historical near-neighbor data, and carries out training and validation according to the method of "80% training + 20% testing". In the final modeling results, the model shows high stability and good generalization ability, with an average R2 of 0.99 and residuals distributed around zero. The system provides a real-time, adaptive and scalable modeling strategy for drilling speed prediction, which provides an important support for intelligent drilling decision-making.