In response to the issues of low accuracy and high computational resource consumption in long-term groundwater level prediction by neural network models, a groundwater level forecasting method based on dilated filtering attention residual network is proposed. This method maintains the integrity of hydrological time series data, possesses a larger receptive field than convolutional neural networks, and can accurately capture the complex spatiotemporal relationships of groundwater levels. Firstly, the hydrological data is preprocessed, then the dilated causal convolution method is used as a feature extractor, combined with a filtering attention mechanism for feature extraction of hydrological data. Finally, residual connections are introduced to alleviate the problems of deep network layers, gradient vanishing, and gradient explosion during the model training process. The changes in hydrological data of Petrignano in Italy are predicted using the method of this paper, LSTM, GRU, CNN-GRU, and TCN methods respectively. The method of this paper has the shortest training time and the most accurate prediction of groundwater level changes, which verifies the reliability of the method proposed in this paper.