Hydraulic support pillar pressure prediction has been a pivotal basis for decision-making in the mining process. It has been one of the fundamental pieces of information for ensuring the stability of the surrounding rock. However, although the pressure of hydraulic support pillars followed certain patterns, it couldn’t be predicted using simple mathematical models. Additionally, during the mining process, issues such as the support detaching the roof, roof fragmentation, and sensor detection errors introduced a significant amount of random noise, turning the pressure data into a non-stationary time series. This significantly complicated the pressure prediction. Based on the Transformer model, this paper proposed a differencing non-stationary Transformer model, which introduced differencing normalization and de-normalization operations in the Transformer′s Encoder and Decoder, respectively, to enhance the stationarity of the series. At the same time, a de-stationary attention mechanism was deployed within the Transformer to calculate the correlations between sequence elements, which thereby enhanced the model′s predictive capabilities. Comparative experiments on a real coal mine support pillar dataset showed that the differencing non-stationary Transformer model proposed in this paper achieved a prediction performance of 0.674, which was significantly better than LSTM, Transformer, and nonstationary Transformer models.