A novel hybrid framework for combining process-based models with machine learning for streamflow prediction

Streamflow prediction is traditionally effective in mitigating water scarcity and flood protection. Current modeling approaches often rely on either “process-based” or “data-based” models, including ones based on machine learning (ML). However, numerous simplifying assumptions tend to limit the performance of process-based models. On the other hand, ML-based models tend to have limited interpretability. To address these drawbacks, we propose and test a hybrid modeling framework entitled the “process elements correction” method (PEC) that couples the Xinanjiang (XAJ) model with the long short-term memory network (LSTM). As a basis for comparison, we also test the more commonly used terminal error correction method (TEC). Our results indicate that in the TEC or PEC frameworks, significant improvements can be achieved by careful choice of the contextual variables provided as inputs to LSTM to correct the error of XAJ model simulations. Overall, both PEC and TEC significantly enhance the performance of the XAJ model, especially under low and medium flow conditions. Meanwhile, we find the newly proposed PEC approach to be superior to TEC in correcting flood volumes. Additionally, PEC and TEC generally performs better than LSTM model, although LSTM leads to the smallest error in flood volume. Overall, the PEC is an efficient hybrid model framework to improve the model simulations.