Convolutional Long Short-term Memory Research Articles

Video Motion Magnification and Subpixel Edge Detection-Based Full-Field Dynamic Displacement Measurement

Noncontact measurement techniques in structural dynamics field have progressed significantly in the past few decades. Vision-based measurement techniques are unique in that they have the ability to achieve full-field measurement and possess the typical advantages associated with noncontact measurement techniques. Recently, vision-based techniques have also been applied to streaming of videos for structural dynamic displacement measurement. The most recent trends in vision-based measurements include target tracing, digital image correlation, and target-less approaches. There are, however, some shortcomings of the vision-based techniques such as susceptibilities to image noise, prevailing light conditions, and limit in measurement resolution. To reduce these shortcomings, a method known as video motion magnification (MM) can be used to amplify small structural motions. Using the phase-based motion magnification (PBMM) and subpixel edge detection methods, the full-field dynamic displacements of the structure can be obtained. The deep convolutional long short-term memory (ConvLSTM) network is applied to aid in the selection of the frequency band for magnification in the PBMM algorithm. To achieve higher measurement accuracy, the displacement results with and without MM are combined with the finite impulse response (FIR) filter which can reduce the error caused by the PBMM procedure. In the tests, plastic optical fiber (POF) displacement sensors are introduced and used as reference measurements to compare the dynamic displacement results from the proposed vision-based method. Compared with the measured displacements with POF sensors, the proposed method offers high level of accuracy for full-field displacement measurement.

Dynamic personalized human body energy expenditure: Prediction using time series forecasting LSTM models

Dynamic human energy expenditure (EE) consists of energy costs for resting metabolism, food digestion, physical activity, and thermoregulation. Currently, multiple models predict EE mainly concerning physical activity, thus, discarding other factors contributing to the dynamic variation of EE. This paper aimed to demonstrate that (i) a dynamic human body EE prediction requires the time series approach, (ii) personalization of input features and models can outperform the generalized approach. To achieve these objectives, data were collected from 3 sets of experiments with 6 test subjects wearing multiple sensors. The analysis of features' importance showed that the selection of features varies for activity-dominated and non-activity-dominated cases and also varies between individuals. Long-Short Term Memory (LSTM) networks were used to develop personalized models such as a simple LSTM, a convolutional LSTM (CNN-LSTM), and also an ensemble model combining CNN-LSTM with a Gradient Boosting algorithm (LSTM-LGBM). A personalized autoregressive linear model and a generalized approach of the LSTM-LGBM method were also developed to have a base of comparison. The results show that the personalized models provide good prediction accuracy, with the mean absolute percentage error (MAPE) mostly lying in the range of 5–15 %. The CNN-LSTM outperforms a simple LSTM model by 3–5 % in MAPE values, and the ensemble model outperforms the by 5–8 % the simple LSTM. The personalized modeling approach with LSTM has shown the potential to improve the prediction accuracy of dynamic EE and capture the non-activity-related effects such as thermoregulation and postprandial thermogenesis.

A new long short-term memory based approach for soil moisture prediction

Water scarcity is becoming more severe around the world as a result of suboptimal irrigation practices. Effective irrigation scheduling necessitates an estimation of future soil moisture content. This study presents deep learning models such as CNN-LSTM, a hybrid Deep Learning model that predicts future soil moisture using climate and soil information, including past soil moisture content. The study also investigates the appropriate number of observations and data sampling rate required to predict the next day’s soil moisture value. In terms of MSE, MAE, RMSE, and R 2 , the hybrid CNN-LSTM model is compared to standalone LSTM and Bi-LSTM models. The LSTM model achieved an MSE of 0.2471, MAE of 0.1978, RMSE of 0.4971, and R 2 of 0.9714. The LSTM model outperformed the Bi-LSTM model, which had an MSE of 0.3036, MAE of 0.3248, RMSE of 0.5510, and R 2 of 0.9614. With an MSE of 0.1348, MAE of 0.1868, RMSE of 0.3672, and R 2 of 0.9838, the hybrid CNN-LSTM model outperformed the LSTM. Our findings suggest that deep learning models, particularly the Convolutional LSTM, hold great potential for predicting soil moisture accurately. The Convolutional LSTM model’s superior performance can be attributed to its ability to capture spatial dependencies in soil moisture data. Furthermore, the results show that for better prediction, sub-hourly data samples from the previous three days should be considered.

Traffic flow prediction under multiple adverse weather based on self-attention mechanism and deep learning models

To improve the accuracy of traffic flow prediction under adverse weather, a deep hybrid attention (DHA) model including the traffic and weather blocks is proposed. The traffic block introduces the convolutional neural network (CNN) and the gated recurrent unit (GRU) neural network to capture the spatio-temporal rules of the traffic flow data. To consider the impacts of adverse weather on traffic flow, the weather block is introduced. The weather block utilizes the convolutional long short-term memory (ConvLSTM) neural network to extract the relationship between the weather and traffic flow data. The self-attention mechanism is embedded in the two blocks. Four cases involving rainy, foggy and windy weather are used to verify the DHA model. The experiments reveal that: the DHA model shows the excellent performance under multiple adverse weather; different adverse weather has various impacts on the rules of traffic volume and speed; the prediction accuracy of each model reduces with the increase of the severe degree of each type of adverse weather.

Image segmentation in marine environments using convolutional LSTM for temporal context

Unmanned surface vehicles (USVs) carry a wealth of possible applications, many of which are limited by the vehicle's level of autonomy. The development of efficient and robust computer vision algorithms is a key factor in improving this, as they permit autonomous detection and thereby avoidance of obstacles. Recent developments in convolutional neural networks (CNNs), and the collection of increasingly diverse datasets, present opportunities for improved computer vision algorithms requiring less data and computational power. One area of potential improvement is the utilisation of temporal context from USV camera feeds in the form of sequential video frames to consistently identify obstacles in diverse marine environments under challenging conditions. This paper documents the implementation of this through long short-term memory (LSTM) cells in existing CNN structures and the exploration of parameters affecting their efficacy. It is found that LSTM cells are promising for achieving improved performance; however, there are weaknesses associated with network training procedures and datasets. Several novel network architectures are presented and compared using a state-of-the-art benchmarking method. It is shown that LSTM cells allow for better model performance with fewer training iterations, but that this advantage diminishes with additional training.

Spatio-temporal DeepKriging for interpolation and probabilistic forecasting

Gaussian processes (GP) and Kriging are widely used in traditional spatio-temporal modelling and prediction. These techniques typically presuppose that the data are observed from a stationary GP with a parametric covariance structure. However, processes in real-world applications often exhibit non-Gaussianity and nonstationarity. Moreover, likelihood-based inference for GPs is computationally expensive and thus prohibitive for large datasets. In this paper, we propose a deep neural network (DNN) based two-stage model for spatio-temporal interpolation and forecasting. Interpolation is performed in the first step, which utilizes a dependent DNN with the embedding layer constructed with spatio-temporal basis functions. For the second stage, we use Long-Short Term Memory (LSTM) and convolutional LSTM to forecast future observations at a given location. We adopt the quantile-based loss function in the DNN to provide probabilistic forecasting. Compared to Kriging, the proposed method does not require specifying covariance functions or making stationarity assumptions and is computationally efficient. Therefore, it is suitable for large-scale prediction of complex spatio-temporal processes. We apply our method to monthly PM2.5 data at more than 200,000 space–time locations from January 1999 to December 2022 for fast imputation of missing values and forecasts with uncertainties.

Spatio-Temporal Coherence of mmWave/THz Channel Characteristics and Their Forecasting Using Video Frame Prediction Techniques

Channel state information in millimeter wave (mmWave) and terahertz (THz) communications systems is vital for various tasks ranging from planning the optimal locations of BSs to efficient beam tracking mechanisms to handover design. Due to the use of large-scale phased antenna arrays and high sensitivity to environmental geometry and materials, precise propagation models for these bands are obtained via ray-tracing modeling. However, the propagation conditions in mmWave/THz systems may theoretically change at very small distances, that is, 1 mm–1 μm, which requires extreme computational effort for modeling. In this paper, we first will assess the effective correlation distances in mmWave/THz systems for different outdoor scenarios, user mobility patterns, and line-of-sight (LoS) and non-LoS (nLoS) conditions. As the metrics of interest, we utilize the angle of arrival/departure (AoA/AoD) and path loss of the first few strongest rays. Then, to reduce the computational efforts required for the ray-tracing procedure, we propose a methodology for the extrapolation and interpolation of these metrics based on the convolutional long short-term memory (ConvLSTM) model. The proposed methodology is based on a special representation of the channel state information in a form suitable for state-of-the-art video enhancement machine learning (ML) techniques, which allows for the use of their powerful prediction capabilities. To assess the prediction performance of the ConvLSTM model, we utilize precision and recall as the main metrics of interest. Our numerical results demonstrate that the channel state correlation in AoA/AoD parameters is preserved up until approximately 0.3–0.6 m, which is 300–600 times larger than the wavelength at 300 GHz. The use of a ConvLSTM model allows us to accurately predict AoA and AoD angles up to the 0.6 m distance with AoA being characterized by a higher mean squared error (MSE). Our results can be utilized to speed up ray-tracing simulations by selecting the grid step size, resulting in the desired trade-off between modeling accuracy and computational time. Additionally, it can also be utilized to improve beam tracking in mmWave/THz systems via a selection of the time step between beam realignment procedures.

Body-Pose-Guided Action Recognition with Convolutional Long Short-Term Memory (LSTM) in Aerial Videos

The accurate detection and recognition of human actions play a pivotal role in aerial surveillance, enabling the identification of potential threats and suspicious behavior. Several approaches have been presented to address this problem, but the limitation still remains in devising an accurate and robust solution. To this end, this paper presents an effective action recognition framework for aerial surveillance, employing the YOLOv8-Pose keypoints extraction algorithm and a customized sequential ConvLSTM (Convolutional Long Short-Term Memory) model for classifying the action. We performed a detailed experimental evaluation and comparison on the publicly available Drone Action dataset. The evaluation and comparison of the proposed framework with several existing approaches on the publicly available Drone Action dataset demonstrate its effectiveness, achieving a very encouraging performance. The overall accuracy of the framework on three provided dataset splits is 74%, 80%, and 70%, with a mean accuracy of 74.67%. Indeed, the proposed system effectively captures the spatial and temporal dynamics of human actions, providing a robust solution for aerial action recognition.

Learning Temporal and Spatial Features Jointly: A Unified Framework for Space-Time Data Prediction in Industrial IoT Networks

Industrial Internet of Things (IIoT) networks, as the application of IoT networks to modern industry, are growing rapidly with the digital transformation accelerates. Mobile sensors are widely adopted in monitoring key information of massive IIoT networks in an ad hoc fashion or retrofitting on to an existing infrastructure. Hence, owing to the characteristics of scalability, unstable locations and highly unstructured, achieving perfect prediction for mobile multi-sensor is a challenging problem. In this paper, a dynamic space-time prediction algorithm combining temporal convolutional network and graph convolutional network (TCN-GCN) is proposed to provide reliable multiple nodes prediction for detection and maintenance in IIoT networks. In particularly, an adaptive learning graph process is exclusively designed according to the dynamic characteristics of mobile sensors. Moreover, the improved dilated time convolution network and dynamic graph convolution network are combined to effectively capture the time dependence and topology information. Furthermore, an advanced loss function is applied to avoid overfitting. Numerical simulations show that the proposed TCN-GCN prediction algorithm exhibits higher effectiveness in accuracy and complexity than convolutional long short-term memory (ConvLSTM), temporal graph convolutional network (T-GCN) and residual graph convolutional long short-term memory (Graph-ResLSTM).

GATE: A guided approach for time series ensemble forecasting

In this article, a new ensemble learning model called GATE is proposed to improve the accuracy and stability of time-series forecasting, which is a crucial aspect of modern engineering practices. Despite the promise of deep learning (DL) models in this area, their performance can be volatile due to the diversity of time series data. To address this, the GATE model combines the strengths of recurrent neural networks (RNN), long short-term memory network (LSTM), and convolution-LSTM (Conv-LSTM) structures and utilizes an unsupervised learning strategy to steer the ensemble output using a guided network. To prevent overfitting in DL models, GATE optimizes the sample loss function and the weight updating function for each individual model within the ensemble structure. A comprehensive evaluation of the proposed GATE method on four real-world datasets is presented. The experimental results unequivocally demonstrate that GATE surpasses state-of-the-art ensemble methods and individual models, exhibiting the best performance in terms of testing errors. Notably, GATE outperforms existing single models in addressing long-term prediction tasks. To validate the effectiveness of GATE, ablation studies is carried out, comparing different ensemble combinations involving two distinct models. Through systematic analysis, these studies provided valuable insights into the performance of various ensemble configurations, further confirming the effectiveness and superiority of the proposed GATE method.

Applying deep learning to predict SST variation and tropical cyclone patterns that influence coral bleaching

South Penghu Marine National Park, Taiwan experienced severe coral bleaching in 2020 that may lead to coral mortality. Corals are important animals that enrich marine biodiversity, but several studies have shown that the range of sea surface temperature (SST) corals can survive is narrow, and recovery time can be as long as 10 years. The influences of climate change induced extreme weather should not be underestimated, and it is critical to accurately predict SST variations. In this study, a deep learning model, convolutional long short-term memory (ConvLSTM), which is a newer model suitable for estimating autocorrelated sequences, was built to predict daily SST for the next day, next week, and next two weeks, with the highest achieved 99.22% accuracy and a RMSE values of 0.2579. On the other hand, Penghu was not directly impacted by tropical cyclones in 2020; hence the lack of typhoon-induced SST cooling could be one of the reasons for extensive coral bleaching. The wind intensity, speed, as well as the thickness of ocean mixed layer all affect the variation of SST cooling. The results show that the degree heating weeks (DHWs) near Penghu achieved its highest level in 2020, and according to the analysis of the study, there is a clear association between climate change and coral bleaching risk increase in recent years. Therefore, we applied Random Forest to establish a machine learning model to classify whether a typhoon can trigger significant cooling. Model accuracy achieved 92.90% in the testing set, and empirical orthogonal function (EOF) analysis contributed the most to the SST cooling pattern among typhoons. This study provides a predictive contribution to global coral conservation for relevant institutions to know the risk beforehand, and they can discuss corresponding strategies and early warnings.

Sentiment Analysis on Large-Scale Covid-19 Tweets using Hybrid Convolutional LSTM Based on Naïve Bayes Sentiment Modeling

Millions of lives were affected rapidly throughout the world when the Covid-19 outbreak spread by leaps and bounds. During this catastrophic period, people used to express their condolence as well as emotions through different social networks. In order to analyze the public comments on Twitter, an experimental approach is developed based on popular words regarding this pandemic. In this paper, various NLP-based research works are discussed on sentiment analysis, trend prediction, topic modeling, learning mechanisms, etc. Furthermore, the hybrid deep learning models are developed based on the Naïve Bayes sentiment model to predict the sentiment from the collected huge number of Coronavirus-related tweets. After performing the n-gram analysis, the Covid-19 specific words are extracted based on their popularity. The public sentiment trend has been analyzed using the extracted topics related to Covid-19 and the tweets are classified according to their sentiment scores. The distinguished sentiment ratings are assigned to the collected tweets based on their sentiment class. Then Convo-Sequential and Convo-Bidirectional long-short term networks are trained using fine-grained sentiment-rated tweets to categorize Covid-19 tweets into five different sentiment classes. Finally, our proposed Convo-Sequential and Convo-Bidirectional LSTM models achieved 84.52% and 85.03% of validation accuracy respectively for the first phase dataset whereas using the second phase dataset the models obtained the validation accuracy of 86.58% and 87.22% respectively.

ITF-WPI: Image and text based cross-modal feature fusion model for wolfberry pest recognition

As one of the necessary cash crops in China and many other countries, wolfberry is parasitized by multiple pests, and its yield is highly susceptible to being affected. On the other hand, agricultural pest backgrounds are complex. When identifying them, single-modal models cannot utilize diverse data types across modalities, resulting in low identification accuracy and data utilization. Traditional unimodal identification models can no longer meet the needs of multimodal data development in agriculture. To overcome these challenges, the ITF-WPI cross-modal feature fusion model is proposed, which consists of CoTN and ODLS for parallel processing of images and text, respectively. We incorporate the Transformer structure (CoT), which focuses on contextual feature extraction, into CoTN to make full use of the rich static and dynamic linear fusion contexts between adjacent keys and improve the 4-stage network of CoTN using Pyramid Squeezed Attention (PSA) to improve the extraction of multi-scale feature structure information and effectively promote the interaction of in-depth features with multi-scale spatial information. The ODLS network constructed by introducing 1D convolutional and bidirectional LSTM stacking has been shown to have more robust text feature acquisition than other advanced convolutional neural network-long short-term memory (CNN-LSTM) models from experimental results, with a 30% reduction in MACCs compared to the optimal model. The results showed that ITF-WPI performed well in accuracy, F1 score, model size, and MACCs with 97.98%, 93.19%, 52.20 MB, and 7.828 G compared to the classical state-of-the-art (SOTA) model, lightweight SOTA model and advanced Transformer neural network synthesis, respectively. The model has critical practical applications for promoting the development of cross-modal models in agriculture and research on wolfberry pest control and improving wolfberry yields. The code and dataset for this study will be posted on GitHub (https://github.com/wemindful/Cross-modal-pest-Identifying) as soon as the study is released, and new data will be updated in the future.

Comparison and Enhancement of Machine Learning Algorithms for Wind Turbine Output Prediction with Insufficient Data

As the penetration of renewable energy sources into a power system increases, the significance of precise short-term forecasts for wind power generation becomes paramount. However, the erratic and non-periodic nature of wind poses challenges in accurately predicting the output. This paper presents a comprehensive investigation into forecasting wind power generation for the following day, using three machine learning models: long short-term memory (LSTM), convolutional neural network-bidirectional LSTM (CNN-biLSTM), and light gradient boosting machine (LGBM). In addition, this paper proposes a method to improve the prediction performance of LGBM by separating data according to the distribution of features, and training and testing each separated dataset with a distinct model. This study includes a comparative analysis of the performance of the proposed models in predicting wind turbine output, offering valuable insights into their respective efficiencies. The results of this investigation were analyzed for two geographically distinct wind farms (Korea and the UK). The findings of this study are expected to facilitate the selection of efficient prediction models within the forecast accuracy auxiliary service market and assist grid operators in ensuring reliable power supply for the grid.

Soil Moisture and Heat Level Prediction for Plant Health Monitoring Using Deep Learning with Gannet Namib Beetle Optimization in IoT.

Plant health monitoring is crucial in ensuring a constant food supply to satisfy the growing demand for food. Hence, it is essential to monitor plant health to maximize the yield and minimize the risk of various diseases. Soil moisture and temperature are of critical importance in plant growth, and predicting them enables farmers to take preventive actions, thereby mitigating the issues affecting plant health. This work presents a plant health monitoring approach by forecasting soil moisture and heat levels by collecting data in an Internet of Things (IoT) environment. Here, for transmitting the soil data acquired by the IoT nodes, a cluster head (CH) selection and routing technique using Gannet Namib beetle optimization (GNBO) is used. The data is routed to a prediction module, wherein soil moisture and heat levels are predicted by Convolutional long short term memory (Conv-LSTM). Furthermore, the hyperparameters of the Conv-LSTM are optimized by the GNBO algorithm. The efficiency of the GNBO-Conv-LSTM is examined based on link life time (LLT), energy, delay, distance, negative predictive value (NPV), positive predictive value (PPV), and true negative rate (TNR) and is observed to have achieved values of 0.675, 0.478 J, 0.092 ms, 50.200 m, 0.885, 0.882, and 0.875, correspondingly.

An Integrated INS/GNSS System With an Attention-Based Deep Network for Drones in GNSS Denied Environments

We propose a neural network-based approach to assist positioning in an integrated Inertial Navigation System/Navigation Satellite System (INS/GNSS) during GNSS interruptions. One can aid a navigation system during GNSS outages by substituting the GNSS measurements with a well-trained Neural Network (NN). Outputs of this NN can then be used in a Kalman filtering scheme to acquire the best consistent and concise estimates according to the INS measurements. Since this problem has inherent spatial and temporal aspects, the proposed NN should account for both aspects simultaneously. Convolutional Long Short-term Memory (CLSTM) structure is an excellent candidate to satisfy the aforementioned requirement. The designed CLSTM architecture uses the angular rates and specific forces measured by INS to output the pseudo GNSS position increment covering for the lost GNSS signal. An attention mechanism is added to the final layer of the CLSTM to counter the gradient vanishing problem in long time-series prediction (ACLSTM). A field test utilizing a fixed-wing Unmanned Aerial Vehicle (UAV) is arranged to evaluate the proposed architecture's performance. The field test result shows a significant 3-dimensional positioning improvement during GNSS outages.

Deep Reinforcement Learning for Load Shedding Against Short-Term Voltage Instability in Large Power Systems.

We introduce an innovative solution approach to the challenging dynamic load-shedding problem which directly affects the stability of large power grid. Our proposed deep Q-network for load-shedding (DQN-LS) determines optimal load-shedding strategy to maintain power system stability by taking into account both spatial and temporal information of a dynamically operating power system, using a convolutional long-short-term memory (ConvLSTM) network to automatically capture dynamic features that are translation-invariant in short-term voltage instability, and by introducing a new design of the reward function. The overall goal for the proposed DQN-LS is to provide real-time, fast, and accurate load-shedding decisions to increase the quality and probability of voltage recovery. To demonstrate the efficacy of our proposed approach and its scalability to large-scale, complex dynamic problems, we utilize the China Southern Grid (CSG) to obtain our test results, which clearly show superior voltage recovery performance by employing the proposed DQN-LS under different and uncertain power system fault conditions. What we have developed and demonstrated in this study, in terms of the scale of the problem, the load-shedding performance obtained, and the DQN-LS approach, have not been demonstrated previously.

Attention U-Net Based on Bi-ConvLSTM and Its Optimization for Smart Healthcare

As an important part of cyber–physical–social intelligence, artificial intelligence (AI)-driven smart healthcare is committed to promoting the application of human–machine hybrid augmented intelligence in the medical field, including AI-assisted medical image analysis and lesion recognition. Among them, deep learning models represented by fully convolutional networks (FCNs) have achieved excellent performance in medical image segmentation. However, limited by the complex structure of segmentation networks and the inherently redundant characteristics of convolutional operation, the scale of these models is extremely large. To further promote the application of machine intelligence in the field of medical image analysis, we propose an attention U-Net based on Bi-ConvLSTM (AUBC-Net) for accurate segmentation of medical images in this article. Different from classical U-Net, the proposed model deals with the potential association between decoding features and encoding features by bidirectional convolution LSTM. Furthermore, for the inherent redundancy characteristics of FCNs, we propose a lightweight feature generation strategy and optimize the calculation process of Bi-ConvLSTM based on tensor multilinear algebra, which can greatly reduce the number of network parameters. In addition, we have conducted the image segmentation experiments on two benchmark medical datasets, and the experimental results demonstrate that the proposed model can not only achieve better performance than existing methods, but also effectively compress network parameters while ensuring performance, which greatly facilitates AI-driven smart medical applications.

Novel attention-based convolutional autoencoder and ConvLSTM for reduced-order modeling in fluid mechanics with time derivative architecture

To construct reduced-order models, we propose a convolutional autoencoder and a convolutional LSTM (CAE-ConvLSTM). Using convolutional LSTMs, our model preserves spatial information at the level of representative features, which are extracted using convolutional autoencoding. A block attention algorithm was proposed for improving the convolutional layer to capture long-range dependency in features. According to the results, the concatenated version of this attention mechanism can improve the performance of the model. A first- and second-order time derivative architecture is employed to modify CAE-ConvLSTM. Future forecasting in this approach is not only based on current conditions but also on dynamic representations of input that enable better features to be captured. The results indicate that the model’s accuracy has improved significantly. On the basis of the obtained results, more than 3000 and 30 times better accuracy were achieved on cylinder and dam break test cases, as well as the ability to predict the next 20 steps.

Ship Trajectory Prediction: An Integrated Approach Using ConvLSTM-Based Sequence-to-Sequence Model

Maritime transportation is one of the major contributors to the development of the global economy. To ensure its safety and reduce the occurrence of a maritime accident, intelligent maritime monitoring and ship behavior identification have been drawing much attention from industry and academia, among which, the accurate prediction of ship trajectory is one of the key questions. This paper proposed a trajectory prediction model integrating the Convolutional LSTM (ConvLSTM) and Sequence to Sequence (Seq2Seq) models to facilitate simultaneous extraction of temporal and spatial features of ship trajectories, thereby enhancing the accuracy of prediction. Firstly, the trajectories are preprocessed using kinematic-based anomaly removal and Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) to improve the data quality for the training process of trajectory prediction. Secondly, the ConvLSTM-based Seq2seq model is designed to extract temporal and spatial features of the ship trajectory and improve the performance of long-time prediction. Finally, by using real AIS data, the proposed model is compared with the Seq2Seq and Bidirectional LSTM based on attention mechanism (Bi-Attention-LSTM) models to verify its effectiveness. The experimental results demonstrate that the proposed model achieves excellent performance in predicting turning trajectories, good predictive accuracy on straight line motions, and greater improvement in prediction accuracy compared to the other two benchmark models. Overall, the proposed model represents a promising contribution to improving ship trajectory prediction accuracy and may enhance the safety and quality of ship navigation in complex and volatile marine environments.