2015 Volume 20 Issue 2 Pages 53-60
Field measurement methods for chlorophyll a concentrations (Chl. a) have been developed, enabling data collection in a short time scale because of recent technological advancements in optical sensors. Such time-series data provides important findings in analyses and predictions of aquatic environments targeting eutrophic water areas. In this study, the water quality dynamics in a eutrophic reservoir in a flat low-lying agricultural area was analyzed from the viewpoint of short-time prediction of time series data using artificial intelligence to assess the water environmental dynamics related to a phytoplankton. Specifically, we proposed a short-term prediction method for Chl. a with the chaos recurrent neural network based on continuous observation data. This study aimed to improve the degree of prediction accuracy with observation noise processing, which uses wavelet analysis, and then examined the effectiveness of the prediction method. As a result, by introducing the noise processing, which uses wavelet analysis and reinforcement learning with supplemental training data, the accuracy of the predictions improved significantly and the practicality of the method proposed in this study increased, as evidenced by the fact that the lead time exceeded 48 h. In particular, it is suggested that predications are feasible with the same degree of accuracy for each lead time within the limit lead time. Yet distinctive variation patterns emerge with regard to changes in the algae species composition caused by artificial effects during the period in which water level management, which varies significantly from regular management, takes place. Therefore, it was difficult to conduct real-time predictions for chronological changes that share no similarities with such learning data.
