Data-driven groundwater level prediction in agricultural areas using temporal convolutional networks

Groundwater plays a critical role in the global water cycle, serving as a primary source of freshwater for agriculture, industry, and domestic use.However, overexploitation of groundwater resources, coupled with the impacts of climate variability, has led to severe consequences. In agriculturally intensive regions, groundwater pumping for irrigation constitutes a significant portion of total water use. Variations in pumping practices, crop types, and irrigation methods result in pronounced spatial and temporal differences in groundwater extraction. Inefficient irrigation practices further exacerbate water losses, underscoring the need for data-driven approaches to enhance water-use efficiency. Machine learning techniques have emerged as transformative tools for groundwater level prediction. Temporal Convolutional Networks (TCN), a deep learning model, are particularly well-suited for this purpose due to their ability to capture long-range temporal dependencies in time-series data with superior computational efficiency. This approach not only ensures improved computational performance and scalability but also makes TCN more resilient to missing or proxy data, such as using power consumption as a substitute for direct pumping volume measurements, enhancing its real-world applicability. In this study, monthly groundwater level records from 2007 to 2023 from nine monitoring wells in a high-density agricultural area were collected, along with precipitation, and pumping data. In the absence of direct pumping volume measurements, power consumption data from pumping wells were utilized as a proxy for groundwater discharge. According to the registered purposes of these wells, they were classified into 14 groundwater usage categories, including irrigation for different crops, aquaculture, and livestock. The TCN model demonstrated robust predictive performance, with RMSE, MAE, and R² values ranging from 0.938–2.966 m, 0.797–2.477 m, and 0.66–0.891, respectively, during training, and 0.523–2.697 m, 0.426–2.288 m, and 0.821–0.842, respectively, during testing. Results from SHAP analysis revealed that precipitation and groundwater pumping for rice irrigation were the dominant factors influencing groundwater level variation. These findings emphasize strong generalization capability, computational efficiency, and ability to learn complex temporal relationships of TCN model. The interpretability and adaptability of TCN model make it an invaluable tool for improving agricultural water management practices, addressing the challenges of groundwater sustainability and climate variability. Furthermore, by incorporating downscaled meteorological forecasts from IPCC AR6 into this developed model, coupled with projected power consumption patterns of pumping wells, the model can efficiently predict future groundwater level variations. This approach has significant implications for policy-making related to groundwater and surface water resource management, promoting sustainable agricultural development and resource conservation.