Machine Learning-Based Groundwater Level Simulation for Industrial Water Management: A Comparative Study of ARIMA, BP-ANN and Random Forest

Timely and accurate simulation of regional groundwater level dynamics is critical to guarantee stable water supply for agricultural production and manufacturing sectors in arid semi-arid loess regions. This study adopts Variable Projection Importance (VPI) to screen feature variables controlling groundwater fluctuation, and constructs three data-driven models: Autoregressive Integrated Moving Average (ARIMA), Backpropagation Artificial Neural Network (BP-ANN) and Random Forest (RF). The three models are applied to simulate monthly groundwater dynamics across seven county-level administrative divisions of water-deficient Dingxi City with scattered agricultural processing and light manufacturing enterprises. Training and independent validation results show the overall simulation accuracy ranking follows RF > BP-ANN > ARIMA. For one-year groundwater level prediction, RF provides higher fitting precision, reliability and better regional adaptability, which can be prioritized for groundwater dynamic simulation to support industrial and agricultural water allocation in Dingxi. In terms of feature contribution, lagged historical groundwater level (L1) contributes over 80% to model outputs for most districts, followed by air humidity and precipitation; human interference factors including total groundwater exploitation and manufacturing water intake occupy low weight. Taking county-level administrative units (basic industrial water management units) as independent modeling objects, this paper compares the performance of three mainstream data-driven models, quantifies the contribution of manufacturing groundwater extraction to aquifer fluctuation, and provides a complete, operable modeling process for groundwater simulation in similar arid areas with small-scale manufacturing clusters. The research results can offer quantitative support for balancing groundwater exploitation limits among manufacturing, farming and ecological protection in Dingxi.