To address the issues of complex flood generation and concentration processes, sudden flood occurrence, and low prediction accuracy of physical mechanism models in small watersheds, two kinds of hybrid methods, namely the errorcorrected hybrid models mainly based on physical mechanism models and the mechanismguided hybrid model mainly based on deep learning models were constructed by using the HYMOD, GR4J, and LSTM models. The simulation performance of different hybrid methods was explored, and the adaptive bandwidth kernel density estimation (ABKDE) was also proposed for flood interval forecasting with different lead times. With the typical small watershed of the Heihe River in Shaanxi Province as an example, the flood forecasting performance of each model was evaluated. The results show that the single models, including the HYMOD, GR4J, and LSTM models can provide reliable forecast results, and the deep learning model LSTM is superior to the physical mechanism models HYMOD and GR4J, while the simulation performance of the HYMOD model is more stable than that of the 〖JP2〗GR4J model. The hybrid models not only retain the interpretability of the physical model, but also improve the accuracy of flood forecasting, with the Nash efficiency coefficient increasing by 3.66% to 70.51%, demonstrating a significant improvement in forecasting performance compared to the single models. The errorcorrected hybrid models have better forecasting performance than the mechanismguided hybrid models, among which the errorcorrected hybrid model HYMODLSTM has the best forecasting effect. The prediction interval coverage probability of the HYMODLSTM model at the 90% confidence level exceeds 92%, demonstrating excellent performance of the model. The HYMODLSTM model can effectively reflect the uncertainty of the forecasted flood process, and the results of flood interval forecasting based on ABKDE are reasonable and reliable, reflecting the good adaptive adjustment ability of ABKDE.