Forecasting groundwater levels at unseen wells without per-well training: GeoAI-GWFM-KR, a time-series foundation model adapted to the Korean monitoring network

Groundwater-level forecasts at monitoring wells are still built mostly by training a recurrent network on each well’s own record, so every new well needs its own model and a well with a short record gets a weak one. Pretrained time-series foundation models forecast series they have never seen, which raises a practical question: does training a network for each well still buy accuracy? We tested this at 30 monitoring stations (46 wells) of the Korean National Groundwater Monitoring Network that were held out before any model was trained, with weekly forecasts for 2023–2024 and every forecaster given the same input variables, the observed levels and the daily rainfall. Chronos-2 used without any training showed no significant difference from per-well long short-term memory (LSTM) networks at both 30 and 90 days, and was less accurate than them at 7 days. Adapted once with a low-rank adapter trained on 188 wells at other stations, the model, released as GeoAI-GWFM-KR, reached a median skill against persistence of 0.41, 0.48 and 0.52 at 7, 30 and 90 days, against 0.37, 0.30 and 0.33 for the per-well LSTM, and was the more accurate at 34, 39 and 40 of the 46 wells; at wells installed since 2019 it was the more accurate at 13 of 13 at 30 days. A physics-informed loss did not make the per-well networks more accurate. For 7- to 90-day forecasts with a rainfall forecast at wells with at least a year of record, one adapted pretrained model matched or exceeded per-well training, without any training at the new well.