The microelectromechanical system (MEMS) inertial measurement unit (IMU), characterized by its low cost and compact size, is extensively employed in the inertial navigation system (INS). However, suffering from the significant biases and noise of MEMS IMUs, the accuracy of INS is compromised. This article presents a deep learning-based method for compensating IMU errors. First, we utilize dilated convolutional neural network as the model framework to capture long sequence information dependencies in IMU data for errors compensation. Meanwhile, a dynamic receptive field mechanism is employed to automatically select information at different scales between gyroscopes and accelerometers, enhancing the regression accuracy of the model. Furthermore, according to the IMU kinematic model, we designed the IMU integral to train the IMU errors compensation model, enabling the regression and compensation of IMU integration errors. Finally, the performance of our IMU errors compensation approach is tested using the TUM-VI dataset.