Functional forecasting of dissolved oxygen in high-frequency vertical lake profiles

摘要

Predicting dissolved oxygen (DO) in lakes is important for assessing environmentalconditions as well as reducing water treatment costs. High levels of DOoften precede toxic algal blooms, and low DO causes carcinogenic metals to precipitateduring water treatment. Typically, DO is predicted from limited datasets using hydrodynamic modeling or data-driven approaches like neural networks.However, functional data analysis (FDA) is also an appropriatemodelingparadigm for measurements of DO taken vertically through the water column.In this analysis, we build FDA models for a set of profiles measured every 2hours and forecast the entire DO percent saturation profile from 2 to 24 hoursahead. Functional smoothing and functional principal component analysis areapplied first, followed by a vector autoregressive model to forecast the empiricalfunctional principal component (FPC) scores. Rolling training windows adaptto seasonality, andmultiple combinations of window sizes,model variables, andparameter specifications are compared using both functional and direct rootmean squared errormetrics. The FPCmethod outperforms a suite of comparisonmodels, and including functional pH, temperature, and conductivity variablesimproves the longer forecasts. Finally, the FDAapproach is useful for identifyingunusual observations.