How climate change erodes short-term lake-temperature predictability: informing climate resilient lake forecasting

Abstract

Climate warming threatens short-term environmental forecast skill, yet its effect on water quality predictability is largely unquantified. Here, we demonstrate a new approach for assessing climate change effects on lake forecasts. Random forest (RF) and gated recurrent unit network models were trained on data from the Inter-Sectoral Impact Model Intercomparison Project (ISIMIP) Local Lakes Sector (five central-European lakes, four hydrodynamic models) and then used to forecast daily lake surface temperature 14 days ahead for 2060 - 2100 under four climate scenarios. We then varied (i) sensor-sampling interval (3, 7, 14 days) and (ii) training-set length (1 - 30 years). Under the strongest forcing (SSP585), the summer mean absolute error (MAE) of worst-affected lake, Esthwaite, rose by 0.14 °C (from 1.75 to 1.89 °C), driven by higher day-to-day temperature volatility (R² = 0.78). For this lake, extending the training set from 5 to 20 years or shortening sampling from 14 to 3 days reduced summer MAE by 0.11 and 0.17 °C, effectively offsetting the volatility caused by climate change. In winter, forecast error declined for four lakes because warmer, more stratified conditions simplified surface-layer dynamics. Thus, modest investments in monitoring cadence or historical record length can preserve forecast skill, even under extreme climate change. More broadly, this highlights a largely unexplored potential use for climate scenario projections: informing the design of climate resilient lake monitoring systems.