Machine Learning for Predictive Management: Short and Long term Prediction of Phytoplankton Biomass using Genetic Algorithm Based Recurrent Neural Networks

Abstract

In the regulated Nakdong River, algal proliferations are annually observed in some seasons, with cyanobacteria (Microcystis aeruginosa) appearing in summer and diatom blooms (Stephanodiscus hantzschii) in winter. This study aims to develop two ecological models forecasting future chlorophyll a at two time-steps (one-week and one-year forecasts), using recurrent neural networks tuned by genetic algorithm (GA-RNN). A moving average (MA) method pre-processes the data for both short- and long-term forecasting to evaluate the effect of noise downscaling on model predictability and to estimate its usefulness and trend prediction for management purposes. Twenty-five physicochemical and biological components (e.g. water temperature, DO, pH, dams discharge, river flow, rainfall, zooplankton abundance, nutrient concentration, etc. from 1994 to 2006) are used as input variables to predict chlorophyll a. GA-RNN models show a satisfactory level of performance for both predictions. Using genetic operations in the network training enables us to avoid numerous trial-and-error model constructions. MA-smoothed data improves the predictivity of models by removing residuals in the data prediction and enhancing the trend of time-series patterns. The results demonstrate efficient development of ecological models through selecting appropriate network structures. Data pre-processing with MA helps in forecasting long-term seasonality and trend of chlorophyll a, an important outcome for decision makers because it provides more reaction time to establish and control management strategies.