Abstract: In the process of hydraulic fracturing for unconventional oil and gas development, microseismic data acquired by geophones are inevitably disturbed by noise, which affects subsequent processing and interpretation. Therefore, noise suppression is a critical step in microseismic data processing. To address the complexity and high computational cost of existing U-Net-based denoising methods, this paper proposes a lightweight attention-based U-Net (LW-AUNet) for microseismic denoising. By using depthwise separable convolution instead of conventional convolution in U-Net, the complexity of the network model is reduced and the training speed is improved through the collaboration of depthwise convolution and pointwise convolution. The lightweight Convolutional Block Attention Module (CBAM) is incorporated to emphasize key data features and efficiently extract complex background noise. Meanwhile, to prevent effective signals from being misjudged as noise due to overly early and strong attention weighting, a skip connection is added to the attention mechanism to ensure that the model can gradually learn and refine features without losing original signal characteristics. During network training, the loss function is defined using both the frequency domain and the time domain, and denoising performance is improved by integrating features from both domains. Synthetic and field data tests show that the proposed method outperforms similar algorithms in reducing model parameters and improving computational efficiency while suppressing random noise.