The Kuroshio is a strong western boundary current of North Pacific wind-driven subtropical gyre. It originates from the bifurcation of the Pacific North Equatorial Current east of the Philippines and flows northeasterly through the East China Sea and the south coast of Japan. At about 35°N, the Kuroshio separates from the coast of Japan and forms the Kuroshio Extension. Interestingly, the Kuroshio south of Japan exhibits a peculiar bimodal path on the interannual time scale: the large meander (LM) and the non-large meander (NLM) paths (Yoshida, 1964). The NLM path was further distinguished into the nearshore and offshore NLM path (Kawabe, 1995). As shown by Figure 1, the LM path takes a large southward meander to the west of the Izu Ridge, whereas the NLM path is close to the southern coast of Japan. Notably, the LM and nearshore NLM (nNLM) paths pass through the deep-water gate (around 34°N) north of the Izu Ridge, while the offshore NLM (oNLM) path flows over the relatively shallow ridge (around 31-33°N).

The Kuroshio volume transport (KVT) in the East China Sea has enormous impacts on navigation, circulation structure, ecological environment, and local climate. In this study, we aim to forecast the daily variability of the KVT at three different sections using the deep learning method. We train the deep learning model using data from 1982 to 2008 and validate the model with data of 2009–2010 and subsequently test it with data of 2011–2015. Four deep learning models, including Artificial Neural Network, Temporal Convolutional Network, Gated Recurrent Unit, and Long Short‐Term Memory (LSTM) models, are first tested to choose the best prediction model. As a result, the LSTM has the best performance for the KVT prediction at each section. We then employ a multivariate causal analysis method to identify the factors affecting the KVT at the current section, such as upstream and downstream KVT, regional mean wind stress, sea surface height and temperature and combine this method with the LSTM model to construct an information flow causality‐based LSTM (IFC‐LSTM) model for predicting daily KVT variability. The results indicate that IFC‐LSTM has the highest forecast skill compared to the standard LSTM (only input KVT at the current section into the LSTM), ALL‐LSTM (input all nine variables into the LSTM), multiple linear regression, and persistence model, which can forecast the KVT variability from 23 to 27 days in advance at the three sections with relative improvement rates of 12.5%–50%.

The Kuroshio is the well-known western boundary current formed by the North Equatorial Current (NEC) flowing northward. When the Kuroshio flows along the southern coast of Japan, its path exhibits three remarkable states (Kawabe, 1995;Taft, 1972): the typical large meander (tLM) path, the offshore nonlarge meander (oNLM) path and the nearshore nonlarge meander (nNLM) path, as shown in Figure 1. Usually, according to whether the LM occurs, these paths south of Japan are classified into two categories: the large meander (LM) path and the nonlarge meander (NLM) path. Long-term observation data indicate that Kuroshio path variations have significant interdecadal characteristics. Once the Kuroshio is in a certain path, it will persist from a year to a decade, but the transition between two paths occurs over several months (Kawabe, 1995). Kuroshio path variations impose important effects on climate, fisheries, and navigational safety in the North Pacific (Hayasaki et al., 2013;Nakamura et al., 2012;Xu et al., 2010). Therefore, it is important to explore its physical dynamics and accurately predict the Kuroshio path.