The accurate prediction of the groundwater level is of vital importance for water resource management and disaster prevention and control. In view of the complex characteristics of the groundwater level time series, such as nonlinearity, multi-scale fluctuations and long-range dependence, traditional prediction models often fail to achieve satisfactory results. To improve the prediction accuracy of complex groundwater level time series, this paper proposes a prediction framework that combines specific data preprocessing with the CNN-Informer hybrid model. This framework first performs data cleaning, wavelet denoising and an adaptive resampling algorithm combining extremum and data volatility of the data to optimize the original high-frequency data. Subsequently, the CNN-Informer model is constructed. Local features are extracted using CNN, long-range dependencies are captured by the Informer encoder, and comprehensive predictions are made through the weighted fusion mechanism. Based on the verification of the measured data from Well Su 18 in Dantu, Jiangsu Province, the proposed CNN-Informer model significantly outperforms baseline models such as LSTM, CNN and individual Informer in key indicators such as MAE, RMSE and NSE, demonstrating its high accuracy and robustness in multi-step prediction. This research provides an effective new approach for the precise prediction of groundwater levels.