Time Series Predictions in Unmonitored Sites: A Survey of Machine Learning Techniques in Water Resources (2308.09766v3)

Abstract: Prediction of dynamic environmental variables in unmonitored sites remains a long-standing challenge for water resources science. The majority of the world's freshwater resources have inadequate monitoring of critical environmental variables needed for management. Yet, the need to have widespread predictions of hydrological variables such as river flow and water quality has become increasingly urgent due to climate and land use change over the past decades, and their associated impacts on water resources. Modern machine learning methods increasingly outperform their process-based and empirical model counterparts for hydrologic time series prediction with their ability to extract information from large, diverse data sets. We review relevant state-of-the art applications of machine learning for streamflow, water quality, and other water resources prediction and discuss opportunities to improve the use of machine learning with emerging methods for incorporating watershed characteristics into deep learning models, transfer learning, and incorporating process knowledge into machine learning models. The analysis here suggests most prior efforts have been focused on deep learning learning frameworks built on many sites for predictions at daily time scales in the United States, but that comparisons between different classes of machine learning methods are few and inadequate. We identify several open questions for time series predictions in unmonitored sites that include incorporating dynamic inputs and site characteristics, mechanistic understanding and spatial context, and explainable AI techniques in modern machine learning frameworks.

• The paper presents a comprehensive survey of ML techniques like LSTM, transfer learning, and KGML to enhance time series predictions in unmonitored water sites.
• It demonstrates how broad-scale models that leverage diverse datasets outperform traditional process-based approaches in capturing temporal patterns.
• The survey highlights the need for future work on uncertainty quantification and explainable AI to bolster ML-driven water resource management.

Machine Learning Techniques for Time Series Predictions in Unmonitored Water Resource Sites

This paper presents a comprehensive survey of ML techniques applied to the challenge of predicting dynamic environmental variables in unmonitored water resource sites. With the inadequacy of monitoring infrastructure for global freshwater resources, there is a pressing need for effective predictive models to manage water resources amidst changing climates and increased human impacts. The authors focus on ML's capacity to outperform traditional process-based models in time series predictions and explore various applications and methodologies for water resources, such as streamflow and water quality prediction.

The paper categorizes ML methodologies into three main approaches: entity-aware broad-scale modeling, transfer learning, and knowledge-guided machine learning (KGML).

1. Entity-Aware Broad-Scale Modeling: This method primarily involves the development of models that encompass a large number of monitoring sites to account for spatial heterogeneity. The analysis finds that long short-term memory (LSTM) networks dominate this approach due to their effectiveness in capturing temporal dependencies. The review also highlights that incorporating site characteristics alongside dynamic inputs can enhance prediction performance significantly. Comparisons between models using all available data and those focusing on similar behavioral sites reveal that broad-scale models tend to perform better, suggesting a preference for leveraging diverse datasets.
2. Transfer Learning: Transfer learning addresses scenarios where target systems have no or scarce data by utilizing knowledge derived from related, well-monitored systems. The paper discusses various strategies, including meta transfer learning, which uses a metamodel to select the best source models for transfer based on site characteristics and other contextual data. This approach aligns with the need to generalize ML applications in environments lacking sufficient observations.
3. Knowledge-Guided Machine Learning (KGML): KGML integrates domain knowledge into ML frameworks to enhance predictive performance and address the limitations of black-box ML models. The survey explores various KGML methods, including using informed loss functions to steer predictions toward physically consistent outputs and differentiable process-based models, which maintain process-based structure while incorporating ML flexibility.

The paper underscores the current dominance of LSTM models for environmental time series prediction and emphasizes the potential of KGML techniques to bridge the gap between ML performance and domain-specific interpretability. The authors argue for the consideration of all available data in model training to maximize prediction accuracy, suggesting that models trained on heterogeneous datasets often generalize better to unmonitored sites.

The paper identifies several open questions for future research, such as the optimal selection of training data and input features, the role of dynamic vs. static site characteristics, and the integration of KGML to enhance the transferability and interpretability of predictions. Moreover, the paper calls for the development of uncertainty quantification techniques and advancements in explainable AI to foster trust and understanding in ML-driven water resource management.

Ultimately, this survey provides a valuable resource for researchers seeking to enhance ML applications in water resources by synthesizing state-of-the-art methodologies and highlighting areas for further investigation. The integration of ML with domain knowledge and regional data presents opportunities for more robust and generalized predictions, potentially transforming management practices for global water resources.