A Comparison of Neural Network Models for Water Quality Forecasting

This project has the objective of making a comprehensive comparison of state-of-the-art machine learning methods for the prediction of water temperature, dissolved oxygen and chlorophyll-a concentration in limnological time series. The time series that are used to evaluate the models are taken from the National Ecological Observatory Network's (NEON) Ecological Forecasting Challenge, an open challenge where teams can submit forecasts to data that is collected and made publicly accessible by NEON. A common finding across the challenge is that a day of year historical mean model (also referred to as the climatology model) commonly produces top scoring forecasts; thus, a primary consideration in this project has been to compare the performance of machine learning models to the climatology model. Using Darts, a Python library that implements various machine learning (ML) models for time series forecasting like Long Short Term Networks (LSTM) to more contemporary models like Temporal Fusion Transformers and Temporal Convolutional Networks, we evaluate the probabilistic forecasts from 8 different neural network models for water temperature, dissolved oxygen and chlorophyll-a concentration at 34 different sites across North America over the course of a year. This research is unprecedented in that we use machine learning models that have not been examined previously for limnological time series forecasting, and few studies have had a similar geographical scale. We found that all except one of the machine learning models performed as well or better than the climatology model across all three target variables according to the continuous ranked probability score. From our early explorations, we learned that missing data imputation greatly affected the accuracy of the machine learning models; thus, we designed a bespoke method to fill in missing data using historical data to achieve our results. Our work supports that machine learning models can accurately forecast limnological time series and are deserving of further development.

Primary Presenter: Marcus Lapeyrolerie, UC Berkeley (marcuslapeyrolerie@me.com)

Authors: