Current machine learning models to predict lake temperature are often limited by both paucity of labeled data and the inability to output values consistent with the laws of physics. Though the recent deluge of sensor data in water resources has created new opportunities for knowledge discovery with the use of deep learning, the vast majority of lakes have no temperature observation data and have thus been inaccessible for traditional deep learning models. We introduce a physics-guided meta transfer learning (PGMTL) framework that uses data from nearby lakes with similar limnological properties and other contextual data to predict a given lake’s temperature using no direct observations from that lake. This framework also leverages natural science domain knowledge to output solutions physically consistent with conservation laws and properties of water density, while also reducing the error from the more traditional deep learning frameworks such as Long Short Term Memory (LSTM).
Results/Conclusions
We find that PGDTL can often significantly outperform the state of the art calibrated General Lake Model (GLM) without needing any labeled data from the lake itself.