Ground Penetrating Radar (GPR) forward modeling is crucial in interpreting real radar data and conducting full waveform inversion. To address the limitations of traditional GPR forward simulations, such as extensive computational demands and inefficiency for real-time detection, we propose a near real-time GPR forward simulation method utilizing a machine learning framework. This method is specifically designed for underground cable detection scenarios, including cable trenches, cable banks, or pipelines. The model considers cable radius and burial depth as key parameters. We employ Principal Component Analysis (PCA) to compress the echo data, deriving principal component weight coefficients. These coefficients serve as outputs for the machine learning network. Our approach features a multi-layer cyclic network architecture, coupled with a learning strategy based on Random Forest algorithms. This design aims to thoroughly examine the relationships between model parameters and principal components, enabling swift GPR simulations through machine learning. Furthermore, we integrate a Deep Neural Network with the Random Forest approach. This combination, which takes the principal component coefficients of echo data as input, establishes a scene parameter prediction model grounded in machine learning. Through inversion simulation and error analysis, we achieved a maximum depth and radius measurement error of 2 cm, with a standard deviation under 1 cm, validating the method's effectiveness.