Abstract: In order to realize the accurate prediction of surface subsidence and conduct in-depth interpretation of the prediction
results to guide the actual project,a SHAP interpretable model of surface subsidence based on improved generative adversarial
network (WGAN-GP) and Transformer is proposed. The model is used to predict the surface subsidence,the tangent
of the influence angle and the offset of the inflection point,so that the prediction parameters are combined with the probability
integral method to establish the surface subsidence formula. Firstly,the Wasserstein distance and gradient penalty strategy are
used to improve the traditional generative adversarial network to enhance the surface subsidence data and enrich the training
set. Then,the Transformer architecture based on the multi-head self-attention mechanism is used to perform deep learning on
the enhanced data,and the hyperparameters are optimized by Bayesian optimization. Finally,based on the SHAP method,the
prediction process and results are comprehensively analyzed and explained to reveal the influence of different characteristics on
the prediction parameters. The results show that WGAN-GP-Transformer has excellent prediction ability for subsidence,tangent
of influence angle and offset of inflection point on the test set,indicating that the model can effectively capture the complex
nonlinear characteristics of predicting surface subsidence and effectively deal with the scene of data scarcity. It is revealed that
there are significant differences in the characteristic contribution intensity and direction of action that affect the three prediction
parameters. The thickness of the loose layer has the greatest influence on the predicted subsidence,and the mining depth has
the greatest influence on the predicted influence angle tangent and the inflection point offset. The practical application of the
model in the 3301 working face of a mine in southwest Shandong shows that the predicted subsidence curve is highly consistent
with the actual situation,which verifies its reliability and generalization performance in practical engineering.

Key words: surface subsidence prediction,Transformer,WGAN-GP,SHAP,deep learning