Multistep-ahead prediction of daily water temperature for Poyang Lake, China, using monthly monitoring data

Poyang Lake (PYL), the largest freshwater lake in China. The scarcity of high-frequency data in large lakes significantly hinders the ability of deep learning (DL) models to forecast daily water temperature (WT) for near-term events management. This study proposes a novel framework that integrates a physically based (PB) model with DL models to address this issue. The PB model first converts monthly observations into daily simulations and extend daily simulations to ungauged areas. After validation against daily WT from Baisha station and monthly WT from seven other stations, the resulting daily WTs are used to train the DL model for one-, three-, and seven-day ahead forecasts. The results indicate that the PB model successfully converted monthly WT into daily WT, effectively capturing both long-term trends and intra-annual daily variations. The DL model, trained with the recalibrated daily WT, produced competitive predictions at both gauged and ungauged stations. Specifically, for ungauged stations, the DL model achieved minimum correlation coefficients (R) exceeding 0.97 for one-day-ahead forecasts and remaining above 0.92 for lead times of three and seven days. Overall, this framework offers a promising solution for near-term daily WT prediction in PYL and maximizes the utility of monitoring data across various spatiotemporal scales. • This framework integrates a physically based model and deep learning to predict daily water temperature in data-scarce lakes. • The physically based model converts monthly data to daily resolution and extends to ungauged areas. • The deep learning model achieves reliable multistep-ahead daily forecasts trained on simulated data.