Forecasting Multiple Groundwater Time Series with Local and Global Deep Learning Networks.

Area of Science:

• Environmental science
• Data science
• Machine learning

Background:

• Environmental monitoring stations generate time series data often analyzed individually.
• Recent machine learning advances suggest benefits of 'global' models using multiple related time series from a network.
• Global models offer larger datasets, shared information, improved generalizability, and solutions for small or missing data.

Purpose of the Study:

• To present a case study on analyzing multiple groundwater monitoring well time series using global machine learning models.
• To compare and contrast different aggregation strategies (single, subsets, all series) within a global model framework.
• To investigate the prediction performance benefits of various machine learning approaches.

Main Methods:

• Analysis of 165 groundwater time series from the Namoi region, Australia.
• Comparison of aggregation methods: single time series, subsets, and all time series.
• Application of machine learning models: multilayer perceptron (MLP), self-organizing map (SOM), long short-term memory (LSTM), and DeepAR.

Main Results:

• Global models incorporating multiple time series generally show improved prediction performance compared to individual analyses.
• DeepAR, an LSTM extension, demonstrated effectiveness in handling multiple time series with autoregressive terms.
• Challenges identified include differing measurement frequencies and temporal pattern variations across the network.

Conclusions:

• Utilizing networks of environmental data in global machine learning models enhances analytical capabilities.
• This approach offers significant opportunities for improving resource-related decision-making.
• Further research should address challenges like data heterogeneity for optimal network analysis.