A hybrid Deep Forest surrogate model and Rime optimization algorithm (RIME) framework for groundwater contamination source identification (GCSI)

When groundwater pollution occurs in operating-enterprises, it is often difficult to determine the source of pollution in a timely manner. In addition, the complex migration and transformation of pollutants in groundwater, coupled with the potential sparsity and noise of monitoring data, result in highly nonlinear characteristics of monitoring data, which reduces the accuracy of source identification. In order to improve the efficiency and accuracy of GCSI, we propose an integrated framework that combines Deep Forest surrogate model with the RIME, creating a high-accuracy, low-cost simulation-optimization system for precise GCSI in operating-enterprises. We establishes a high-precision groundwater numerical simulation model using practical case data, which was then used to generate the dataset required for GCSI. The objective of this work is to establish a framework, including Deep Forest surrogate model and RIME, and evaluate the efficiency and accuracy through comparing with Backpropagation Neural Networks (BPNN), Bidirectional Long Short-term neural networks (BiLSTM), and Genetic Algorithm (GA). The results indicate that the accuracy of the Deep Forest is slightly higher than that of BiLSTM, with RMSE value of 70.4469, MAE value of 37.1714, and R² value of 0.9793. BPNN requires the least amount of time, at 17.4103 s, but has the worst accuracy. It is worth noting that compared to BiLSTM, Deep Forest reduces computation time by 9.8 %. When inputting Deep Forest identified contamination source results into the groundwater numerical model, the relative error between simulated and observed contaminant concentrations is ≤15 %.