Prediction of salinity concentration in Hichirippu-numa based on long short-term memory

(Consideration of the applicability of data assimilation flow analysis results to training data)

Abstract

This paper describes a prediction method of the salinity concentration at the sea urchin farm in Hichirippu-numa, Hokkaido, using long short-term memory (LSTM). LSTM is one of the machine learning methods and also solves long term time series forecasting tasks. One of the main factors affecting the accuracy of the machine learning model is the data quality, although the observations at remote locations are used for real-time salinity concentration forecasting of sea urchin farm in the Hichirippu-numa. In order to improve the data quality, we focus on a data assimilation flow analysis based on Kalman filter FEM (KF-FEM). KF-FEM can estimate the water elevation at the specific location in computational domain. By using the results of KF- FEM considering precipitation, we generated more reliable training data for LSTM. In this study, we proposed to apply the results of the data assimilation flow analysis into training data of LSTM and confirmed proposed method is better accuracy of salinity concentration forecasting than conventional method.