Real-time prediction of chlorophyll-a time series in a eutrophic agricultural reservoir in a coastal zone using recurrent neural networks with periodic chaos neurons

To assess the water environmental dynamics related to a phytoplankton, the water quality dynamics in a eutrophic reservoir in a flat low-lying agricultural area were analyzed from the viewpoint of short-time prediction of time series data using artificial intelligence. A recurrent neural network model with periodic chaos neurons was used for the real-time prediction of chlorophyll-a time series on the basis of on-site continuous observation data. These data consisted of the chlorophyll-a from four algae classes, Chlorophyceae, cyanobacteria, diatom/dinoflagellates, and cryptophytes, measured by a submerged fluorometer. The results suggest that study of a neural network could be performed sufficiently for teaching data of which a value of the fractal dimension calculated by the Higuchi method was smaller. In addition, it is possible to conduct a short-time prediction of chlorophyll-a time series such that an upper limit of lead time could be beyond 12h when there is an analogous time-frequency characteristic between the teaching and the predicting data. In conclusion, a recurrent chaotic neural network could be an effective analysis tool for short-time prediction of water quality on the basis of continuous observations, and the potential for prediction can be determined quantitatively using Higuchi's fractal dimension and time-frequency maps.