Existing imbalanced data regression algorithms require multiple calculations of distances between samples, which leads to low processing efficiency when the number of samples is large. The granular ball model can iteratively divide a sample set into multiple granular balls, reducing the sample scale. However, current granular ball models rely on sample category labels and are not suitable for regression tasks. This paper utilizes the grid division of the granular ball sample area, defines the filling degree of the granular ball, and proposes a grid self-organized granular ball model (GSOGB) that can handle both regression and classification tasks simultaneously. On this basis, an oversampling strategy for samples within the granular ball in the neighborhood is given, and an imbalanced regression method based on the grid self-organized granular ball model (GSOGB-SMOTER) is proposed. Experimental results show that the proposed GSOGB outperforms existing granular ball models on 12 classification datasets, and the proposed GSOGB-SMOTER slightly outperforms seven algorithms in the literature on the MSE metric of five equal-length target value domains of 10 imbalanced regression datasets, and has robustness and higher operational efficiency, capable of quickly processing large-scale imbalanced data for regression.