Salinity forecasting for Hichirippu lagoon using LSTM networks with data assimilation results

抄録

In this paper, we investigate how data quality improvement affects the performance of salinity forecasting in Hichirippu Lagoon, Hokkaido, Japan. A Long Short-Term Memory (LSTM) network, a type of recurrent neural network, was used for salinity forecasting. Due to the limited availability of field observation data, conventional forecasts use data collected from an Automated Meteorological Data Acquisition System which is 10 km from the lagoon. It is theoretically possible to improve data quality by predicting the parameters to be used; however, little research has been conducted on the effect of data quality improvement on the performance of salinity forecasting. This prompted us to focus on water elevation as a key feature, and to conduct data assimilation analysis using the Kalman filter finite element method (KF-FEM), which combines the Kalman filter with the finite element method. KF-FEM considers climate change information, such as rainfall, by using observation data for the calculations. KF-FEM cannot be carried out in the absence of field data, so a surrogate model based on deep learning was used instead. Finally, we conducted salinity forecasting using the model, DA-LSTM (data assimilation LSTM), trained with data assimilation results. To evaluate the effect of data quality improvement, we compared the forecast results of DA-LSTM and conventional LSTM. As a result, DA-LSTM showed improved forecast performance when salinity was 30 [g/L] or below. An overall performance comparison revealed the impact of data quality improvement to be limited, but it showed the potential to contribute to improved accuracy under specific conditions. This is an element that should be considered in future research.