In contrast to land transport, water transport lacks well-defined fixed routes, granting ships a greater degree of freedom in their movement. Consequently, the management of water transport is inherently more intricate. Ship collisions are an unfortunate occurrence during waterborne navigation, often stemming from factors like equipment malfunctions and human errors. Thus, real-time monitoring of the ship’s course becomes indispensable for timely detection of anomalous behavior and the reduction of the risk of ship accidents. To support the safety of ship navigation, the water traffic management authorities employ a suite of modern information equipment, including the Automatic Identification System (AIS) and Global Positioning System (GPS) [1].

AIS allows real-time monitoring of vessels within the base station’s coverage area by receiving and processing information transmitted by shipboard AIS equipment, leveraging unique AIS equipment codes for automatic ship identification. While AIS data discloses the location of the most recent AIS message sent, it carries inherent transmission delays. Additionally, there’s a time lag associated with data transmission, parsing, loading, and display, hindering real-time ship position representation. Moreover, equipment failures and signal interference can lead to the loss of AIS data. These challenges in the domain of AIS data complicate the task of ensuring ship safety.

To address these issues, we present a real-time ship trajectory prediction model (RSTPM) designed to enable real-time ship trajectory monitoring and prompt detection of irregular behavior. This model offers valuable applications in ensuring the safety of ship navigation.

According to the relevant research based on AIS data at home and abroad, we summarize both the advantages and disadvantages of various ship trajectory prediction methods, as indicated in Table 1.

Simulation methods involve creating physical models to simulate real ship behavior. This method is rarely used alone in ship trajectory prediction; it is generally combined with other methods to form a hybrid method for prediction.

The ESM is used to predict the location, course, and speed of the ship; meanwhile, the actual collision scene of the ship is analyzed. This method has been shown to achieve the prediction of ships’ behavior [2]. Mazzarella proposed a Bayesian algorithm based on particle filters that uses KNN to match the current trajectory sequence of the ship, enabling the prediction of ship trajectories when traffic route data are available [3].

The statistical-based approach assumes that the historical trajectory of a ship and the predicted trajectory have a certain similarity, and the prediction is achieved by fitting the ship trajectory.

Ju et al. [4] proposed a multi-layer architecture interactive-aware Kalman neural network to solve the problem of mutual interaction in the transportation system. A differential ARIMA model was used to predict ship trajectories from ship AIS data, which is applicable to the detection of ship collision avoidance [5]. The trajectory sequence is transformed into column vectors through wavelet transform, which are then used as the input for HMM. This is an algorithm (HMM-WA) that increases the accuracy of ship trajectory prediction [6]. Rong et al. [8] presented a model based on Gaussian processes and uncertain acceleration, designed to achieve real-time monitoring of ships during navigation.

Unsupervised learning mainly focuses on the clustering and dimensionality reduction of data, while supervised learning has a broader range of applications. For example, KNN and SVM can predict ship trajectories by learning the motion characteristics of ship trajectories.

Duca et al. [9] proposed a model for trajectory prediction based on a KNN classifier, considering five characteristics of ships: longitude, latitude, heading, speed, and type. The model’s prediction accuracy was verified. Liu et al. [10] developed an online multioutput model based on a selection mechanism. The model can achieve high prediction accuracy with small samples. Additionally, an SVR-based trajectory prediction model was proposed, but the sample data and parameters required for the model cannot be changed during model training [11]. Gan et al. [12] used the clustered ship trajectory and other known factors, such as ship speed, to establish an ANN model for predicting the ship’s trajectory.

Since the advent of deep learning, it has demonstrated excellent performance in many tasks, including the prediction of ship trajectories. The advantage of LSTM over BP neural networks lies in its ability to process and analyze time series and sequence data. The gate structure in LSTM’s internal network enables it to mine trends and correlations in sequence data, resulting in a better prediction effect when applied to time series data, such as traffic and location. Moreover, LSTM’s prediction accuracy is better than that of BP neural networks [16,17,18,28], making it applicable to long-term prediction. Gao et al. [29] studied a bi-directional LSTM (Bi-LSTM) network, aiming to enhance the memory ability of historical data and the correlation between future time series data. Liu et al. [19] integrated convolutional transformations into a Bi-LSTM based on an attention mechanism in order to achieve long-term prediction. In another study, ship trajectory sequence features extracted by CNN are input into the LSTM model for prediction [22]. In addition to LSTM, GRU, CNN, and GAN, there are also DNN-based frameworks for predicting the trajectory of merchant ships (such as tankers and container ships). However, the DNN module is prone to overfitting and may not achieve high accuracy during training [23].

The hybrid method focuses on combining the advantages of various methods to enhance the efficiency of trajectory prediction tasks. Murray & Perera [24] proposed an algorithm that combines GMM, KNN, and bilinear automatic coding to build the prediction model. Schöller et al. [26] first used kernel density estimation to convert historical AIS data into probabilistic heat maps and then used a convolutional autoencoder for further coding. They constructed a model based on GAN and LSTM to achieve ship trajectory prediction. Additionally, Suo et al. [27] constructed a hybrid model based on DBSCAN and GRU to achieve real-time prediction.

Moreover, the clustering methods of ship trajectory are roughly divided into distance-based [30,31,32,33], density-based [34,35,36], and statistical-based [37,38] methods. Most studies have achieved efficient clustering of ship trajectories by combining the advantages of various clustering methods [30,31,35,36,37,39]. In this study, a new approach to cluster ship trajectory, called the combination of DBSCAN and DTW (CDDTW), is proposed, which combines the optimized DBSCAN algorithm (based on density) and the improved DTW algorithm (based on distance) to cluster ship historical trajectories. The proposed method also extracts the regional navigation characteristics of ship trajectory based on the clustering results. An RSTPM based on an LSTM is constructed.