Abstract: In deep coal mining, microseismic events often serve as a critical signal for early warning of mine water inrushes. Studies have shown that coal rock masses experience fractures and microseismic activities during the formation of water inrush channels, and accurate prediction of microseismic events is crucial for mine safety. However, traditional deep learning models such as long short-term memory (LSTM) and convolutional neural network (CNN) have limitations in handling the non-stationarity and complex dynamic features of microseismic data. To improve prediction accuracy and generalization capability, a microseismic time series prediction method based on a dual-layer modal decomposition and autocorrelation decomposition Transformer model is proposed. First, the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) is used to decompose the original microseismic data into multiple intrinsic mode functions (IMFs). Then, a consistent random sampling method is applied to each IMF sequence to calculate its sample entropy. By setting a sample entropy threshold, different IMF sequences are reconstructed into high-frequency, mid-frequency, and low-frequency sequences. Next, variational mode decomposition (VMD) is applied to process these three sequences, and the optimal number of decompositions is determined by calculating the center frequency for different decomposition numbers, resulting in a multi-feature time series of microseismic data. Finally, a sliding window mechanism is employed to construct an input-output dataset, where a multivariate feature time series over a period is used as input, and microseismic data from a subsequent period is used as output. The Transformer model with autocorrelation decomposition is then used for time series forecasting. Comparative studies are conducted on various time series neural network models using microseismic data from a working face in Zhang Shuanglou coal mine. Experimental results show that the proposed model has significant advantages in prediction accuracy and model stability. In single-step prediction, the model accurately predicts changes in microseismic energy intensity, with the prediction fitting the true values with an R² of 0.92 on the test set. In multi-step prediction, the model effectively predicts the evolution trend of microseismic events, with the predicted values fitting the true values with an R² of approximately 0.82 when predicting the next 12 steps based on the first 24 steps. The proposed method not only improves prediction accuracy and stability but also provides new technical support for the scientific management and risk prevention of coal mine microseismic events. It offers new insights into microseismic monitoring for water inrush in coal mines and has significant implications for the construction of smart mines.