Temporal Attention Mechanism Research Articles

Data and model synergy-driven rolling bearings remaining useful life prediction approach based on deep neural network and Wiener process

Abstract Various Remaining Useful Life (RUL) prediction methods, encompassing model-based, data-driven, and hybrid methods, have been developed and successfully applied to prognostics and health management for diverse rolling bearing. Hybrid methods that integrate the advantages of model-based and data-driven approaches have garnered significant attention. However, the effective integration of the two methods to address the randomness in rolling bearing full lifecycle processes remains a significant challenge. To overcome the challenge, this paper proposes a data and model synergy-driven RUL prediction framework that includes two data and model synergy approaches. Firstly, a convolutional stacked bidirectional long and short-term memory network with temporal attention mechanism is established to construct health index. The RUL prediction is achieved based on HI and polynomial model. Secondly, a three-phase prediction model based on the Wiener process is established by considering the evolution law of different degradation phases. Then, two synergy methods are designed. Strategy 1: HI is adopted as the observation value for online updating of physics degradation model parameters under Bayesian framework, and the RUL prediction results are obtained from the physics degradation model. Strategy 2: The RUL prediction results from the data-driven and physics-based model are weighted linearly combined to improve the overall prediction accuracy. The effectiveness of the proposed model is verified using two bearing full lifecycle datasets. The results indicate that the proposed approach can accommodate both short-term and long-term RUL predictions, outperforming state-of-the-art single models.

A temporal convolution network-based just-in-time learning method for industrial quality variable prediction

Real-time acquisition of quality variables is paramount for enhancing control and optimization of industrial processes. Process modeling methods, such as soft sensors, offer a means to predict difficult-to-obtain quality variables using easily measurable process parameters. However, the dynamic nature of industrial processes poses significant challenges to modeling. For instance, conventional models are typically trained offline using historical data, rendering them incapable of adapting to real-time changes in data distribution or environmental conditions. To tackle this challenge, we introduce a novel approach termed the Residual Temporal Attention Temporal Convolution Network (RTA-TCN) and propose a just-in-time learning method based on RTA-TCN for industrial process modeling. The RTA-TCN model incorporates temporal attention into TCN, enabling the integration of previous time-step process variables into the current ones, as well as the fusion of internally relevant features among inputs. Moreover, to prevent the partial loss of original information during feature integration, residual connections are introduced into the temporal attention mechanism. These connections facilitate the retention of original feature information to a maximal extent while integrating relevant features. Consequently, the proposed RTA-TCN demonstrates significant advantages in handling the non-linearity and long-term dynamic dependencies inherent in industrial variables. Additionally, the proposed just-in-time learning method leverages RTA-TCN as a local model and updates it in real-time using online industrial data. This just-in-time learning method enables effective adaptation to varying data distributions and environmental conditions. We validate the performance of our method using two industrial datasets (Debutanizer Column and Sulfur Recovery Unit).

A multi-task spatio-temporal fusion network for offshore wind power ramp events forecasting

With the accelerated development of offshore wind farms, the impact of Wind Power Ramp Events (WPRE) on power systems has become more pronounced. Accurate prediction of WPRE is essential for mitigating their adverse effects on the grid. However, current studies on offshore WPRE are limited, especially regarding their spatio-temporal correlations and variability across large wind farm clusters. To address this, a novel forecasting approach using a multi-task spatio-temporal fusion network is presented. This method improves WPRE prediction by integrating spatial and temporal data and utilizing multitask learning to characterize ramp features. Firstly, an index system for WPRE characterization is developed, including metrics such as power change rate and duration. Based on this, an X-means classification method for WPRE is established. Secondly, a spatio-temporal encoder combining graph convolutional networks and temporal attention mechanisms has been proposed to model the spatiotemporal dependencies of meteorological changes in offshore wind farms. Finally, a multitask learning framework has been introduced to predict WPRE characteristic indices and occurrence probabilities. This approach enhances feature extraction efficiency through the joint feedback of multiple related tasks, thereby improving prediction accuracy. A case study using data from five offshore wind farms in Eastern China validates the efficacy of the proposed method.

Distributed Photovoltaic Communication Anomaly Detection Based on Spatiotemporal Feature Collaborative Modeling

As distributed photovoltaic (PV) technology rapidly develops and is widely applied, the methods of cyberattacks are continuously evolving, posing increasingly severe threats to the communication networks of distributed PV systems. Recent studies have shown that the Transformer model, which effectively integrates global information and handles long-distance dependencies, has garnered significant attention. Based on this, our research proposes a model named STformer, which is applied to the task of attack detection in distributed PV communication. Specifically, we propose a temporal attention mechanism and a variable attention mechanism. The temporal attention mechanism focuses on capturing subtle changes and trends in data sequences over time, ensuring a highly sensitive recognition of patterns inherent in time-series data. In contrast, the variable attention mechanism analyzes the intrinsic relationships and interactions between different variables, uncovering critical correlations that may indicate abnormal behavior or potential attacks. Additionally, we incorporate the Uniform Manifold Approximation and Projection (UMAP) dimensionality reduction technique. This technique not only helps reduce computational complexity but, in certain cases, can enhance anomaly detection performance. Finally, compared to classical and advanced methods, STformer demonstrates satisfactory performance in simulation experiments.

Dynamic Spatio-Temporal Modeling for Enhanced Air Quality Prediction: Implications for Information Management and Public Health Decision Support Systems

Air quality forecasting has significant implications for environmental monitoring, public health, and information management systems. This study proposes three advanced deep learning models: Graph Neural Networks with Dynamic Spatio-Temporal Attention (GNN-DSTA), Multi-Resolution Convolutional Recurrent Neural Networks (MRC-RNN), and Variational Autoencoders with Spatio-Temporal Latent Embeddings (VAE-STLE). These models address the challenges of capturing complex spatio-temporal dependencies in environments with sparse data samples. The GNN-DSTA model introduces a temporal attention mechanism that dynamically captures evolving spatial-temporal dependencies. MRC-RNN combines CNN's spatial pattern recognition with RNN's temporal modeling across multiple spatial resolutions. VAE-STLE provides a probabilistic framework for robust and interpretable forecasting. Experimental results demonstrate significant improvements in prediction accuracy: GNN-DSTA reduces RMSE by 15-20%, MRC-RNN improves accuracy by 12-15%, and VAE-STLE shows a 10-12% improvement with enhanced uncertainty estimation. These models advance AQI predictions through dynamic attention mechanisms, multi-resolution analysis, and probabilistic forecasting. Furthermore, this study explores the implications of these advanced predictions for information management and public health decision support systems. We discuss the integration of real-time data, scalability considerations for large-scale deployments, user interface design for effective communication of predictions, and ethical considerations in using AI-driven models for public health decision-making. The proposed approach not only enhances the accuracy and reliability of AQI predictions but also provides a framework for developing more effective and responsive public health interventions and environmental policies.

Local Weather and Global Climate Data-Driven Long-Term Runoff Forecasting Based on Local–Global–Temporal Attention Mechanisms and Graph Attention Networks

The accuracy of long-term runoff models can be increased through the input of local weather variables and global climate indices. However, existing methods do not effectively extract important information from complex input factors across various temporal and spatial dimensions, thereby contributing to inaccurate predictions of long-term runoff. In this study, local–global–temporal attention mechanisms (LGTA) were proposed for capturing crucial information on global climate indices on monthly, annual, and interannual time scales. The graph attention network (GAT) was employed to extract geographical topological information of meteorological stations, based on remotely sensed elevation data. A long-term runoff prediction model was established based on long-short-term memory (LSTM) integrated with GAT and LGTA, referred to as GAT–LGTA–LSTM. The proposed model was compared to five comparative models (LGTA–LSTM, GAT–GTA–LSTM, GTA–LSTM, GAT–GA–LSTM, GA–LSTM). The models were applied to forecast the long-term runoff at Luning and Pingshan stations in China. The results indicated that the GAT–LGTA–LSTM model demonstrated the best forecasting performance among the comparative models. The Nash–Sutcliffe Efficiency (NSE) of GAT–LGTA–LSTM at the Luning and Pingshan stations reached 0.87 and 0.89, respectively. Compared to the GA–LSTM benchmark model, the GAT–LGTA–LSTM model demonstrated an average increase in NSE of 0.07, an average increase in Kling–Gupta Efficiency (KGE) of 0.08, and an average reduction in mean absolute percent error (MAPE) of 0.12. The excellent performance of the proposed model is attributed to the following: (1) local attention mechanism assigns a higher weight to key global climate indices at a monthly scale, enhancing the ability of global and temporal attention mechanisms to capture the critical information at annual and interannual scales and (2) the global attention mechanism integrated with GAT effectively extracts crucial temporal and spatial information from precipitation and remotely-sensed elevation data. Furthermore, attention visualization reveals that various global climate indices contribute differently to runoff predictions across distinct months. The global climate indices corresponding to specific seasons or months should be selected to forecast the respective monthly runoff.

Trajectory Privacy-Protection Mechanism Based on Multidimensional Spatial–Temporal Prediction

The popularity of global GPS location services and location-enabled personal terminal applications has contributed to the rapid growth of location-based social networks. Users can access social networks at anytime and anywhere to obtain services in the relevant location. While accessing services is convenient, there is a potential risk of leaking users’ private information. In data processing, the discovery of issues and the generation of optimal solutions constitute a symmetrical process. Therefore, this paper proposes a symmetry–trajectory differential privacy-protection mechanism based on multi-dimensional prediction (TPPM-MP). Firstly, the temporal attention mechanism is designed to extract spatiotemporal features of trajectories from different spatiotemporal dimensions and perform trajectory-sensitive prediction. Secondly, class-prevalence-based weights are assigned to sensitive regions. Finally, the privacy budget is assigned based on the sensitive weights, and noise conforming to localized differential privacy is added. Validated on real datasets, the proposed method in this paper enhanced usability by 22% and 37% on the same dataset compared with other methods mentioned, while providing equivalent privacy protection.

A time patch dynamic attention transformer for enhanced well production forecasting in complex oilfield operations

In the field of oil production, accurately predicting well production is of great significance for optimizing production processes and enhancing economic benefits. However, oil production data generally exhibit characteristics such as high noise, complex distribution, and severe temporal fluctuations, which pose significant challenges to traditional prediction models, thereby limiting their accuracy. To address this issue, this study proposes an innovative Time Patch Dynamic Attention Transformer (TPDAT) model. This model segments the time series data into multiple time patches, extracts local features, and integrates a dynamically fused causal temporal attention mechanism to enhance the recognition of transient events and short-term fluctuations. The TPDAT model comprises the Temporal Patch Feature Extraction (TPF) module and the Dynamic Fusion Causal Temporal Attention (DFCTA) mechanism, effectively capturing multi-scale temporal features and enhancing the modeling of causal relationships during oilfield production. Experiments on the North Sea Volve oilfield dataset show that the TPDAT model significantly outperforms traditional deep learning models in well production prediction, achieves the best performance in model evaluation metrics, and demonstrates strong adaptability to complex data. The results confirm that the TPDAT model improves the accuracy and stability of well production prediction in complex oilfield environments, providing reliable support for oilfield production optimization.

GSTGM: Graph, spatial–temporal attention and generative based model for pedestrian multi-path prediction

Pedestrian trajectory prediction in urban environments has emerged as a critical research area with extensive applications across various domains. Accurate prediction of pedestrian trajectories is essential for the safe navigation of autonomous vehicles and robots in pedestrian-populated environments. Effective prediction models must capture both the spatial interactions among pedestrians and the temporal dependencies governing their movements. Existing research primarily focuses on forecasting a single trajectory per pedestrian, limiting its applicability in real-world scenarios characterised by diverse and unpredictable pedestrian behaviours. To address these challenges, this paper introduces the Graph Convolutional Network, Spatial–Temporal Attention, and Generative Model (GSTGM) for pedestrian trajectory prediction. GSTGM employs a spatiotemporal graph convolutional network to effectively capture complex interactions between pedestrians and their environment. Additionally, it integrates a spatial–temporal attention mechanism to prioritise relevant information during the prediction process. By incorporating a time-dependent prior within the latent space and utilising a computationally efficient generative model, GSTGM facilitates the generation of diverse and realistic future trajectories. The effectiveness of GSTGM is validated through experiments on real-world scenario datasets. Compared to the state-of-the-art models on benchmark datasets such as ETH/UCY, GSTGM demonstrates superior performance in accurately predicting multiple potential trajectories for individual pedestrians. This superiority is measured using metrics such as Final Displacement Error (FDE) and Average Displacement Error (ADE). Moreover, GSTGM achieves these results with significantly faster processing speeds.

A Study on Stock Price Prediction Based on Attention GCN-BiLSTM Modeling

It is of great significance to accurately predict stock price movements. By combining the R-Vine-Copula structure with the relationship between the conceptual sectors to which the stocks belong, the price correlation between stocks is measured and the adjacency matrix is constructed, and the model AGLSTM, which is constructed by combining the temporal attention mechanism, Graph Convolutional Neural Network and Bi-directional Long and Short-Term Memory Neural Network, is proposed to predict the price of 46 stocks from the constituent stocks of the SSE 50 by modeling them. The experimental results show that, compared with the baseline, the AGLSTM can predict the stock price. The experimental results show that compared with the baseline model, the MAE and RMSE results of the AGLSTM model predicted one step forward outperform all the comparison models, and the MAPE results are close to the best baseline model results. The results of MAPE are close to those of the best baseline model. Moreover, good results are also achieved in the experiment of multi-step forward prediction, which demonstrates the ability of AGLSTM model in long-term prediction and can provide some reference for investors to make investment decisions.

TABLE: Time-aware Balanced Multi-view Learning for stock ranking

Stock ranking is a significant and challenging problem. In recent years, the use of multi-view data, such as price and tweet, for stock ranking has gained considerable attention in the research field. Most existing methods are performed in (some of) the 3 steps: 1) view-specific representation learning; 2) cross-view representation interaction; 3) multi-view representation fusion. Although these methods make breakthroughs in stock ranking, they often treat all views equally. This neglects the unbalanced phenomenon in multi-view stock data, i.e., the dimension of the text view may be extremely big compared with those of other views; the price view exhibits standard and high-quality data, whereas the text view contains noise and has irregular time intervals. To solve this, we propose a Time-Aware Balanced multi-view LEarning (TABLE) method. TABLE method consists of a view-specific learning stage and a multi-view fusion stage. In the first stage, we aim to improve the quality of the low-quality text view. We achieve this by attenuating the negative impact of irrelevant texts using a hierarchical temporal attention mechanism that captures text correlations. Additionally, we explicitly model the time irregularities between sequential texts. In the fusion stage, we address the dimensions unbalance problem by establishing a multi-view decision fusion paradigm by weighted averaging the view-specific stock predictions. These weights are dynamic and determined based on the quality discrepancy between the views. Finally, we obtain the optimal stock ranking list by optimizing the point-wise regression loss and the ranking-aware loss. We empirically compare TABLE method with state-of-the-art baselines using the publicly available dataset, S&P500. The experimental results demonstrate that TABLE method outperforms the baseline methods in terms of accuracy and investment revenue.

Dual temporal attention mechanism-based convolutional LSTM model for industrial dynamic soft sensor

Abstract Deep learning is an appropriate methodology for modeling complex industrial data in the field of soft sensors, owing to its powerful feature representation capability. Given the nonlinear and dynamic nature of the process industry, the key challenge for soft sensor technology is to effectively mine dynamic information from long sequences and accurately extract features of relevance to quality. A dual temporal attention mechanism-based convolutional long short-term memory network (DTA-ConvLSTM) under an encoder-decoder framework is proposed as a soft sensor model to acquire quality-relevant dynamic features from serial data. Considering different influences of process variables for prediction at multiple time steps and various locations, ConvLSTM and temporal self-attention mechanism are utilized as the encoder to adaptively fuse spatiotemporal features and capture long-term dynamic properties of process in order to capture the trends of industrial variables. Furthermore, a quality-driven temporal attention mechanism is employed throughout the decoding process to dynamically select relevant features to more accurately track quality changes. The encoder-decoder model meticulously analyses the interactions between process and quality variables by incorporating dual-sequence dynamic information to improve the prediction performance. The validity and superiority of the DTA-ConvLSTM model was validated on two industrial case studies of the debutanizer column and sulfur recovery unit. Compared to the traditional LSTM model, the proposed model demonstrated a substantial improvement with the accuracy R2 up to 97.3% and 94.9% and the root mean square error reducing to 0.122 and 0.022.

ConvTransformer Attention Network for temporal action detection

Boundary detection is a challenging problem in Temporal Action Detection (TAD). While transformer-based methods achieve satisfactory results by incorporating self-attention mechanisms to model global dependencies for boundary detection, they face two key issues. Firstly, they lack explicit learning of local relationships; this limitation results in imprecise boundary detection when subtle appearance changes occur between adjacent clips. Secondly, transformer-based methods lead to feature convergence across multiple actions due to the self-attention mechanism’s tendency to distribute focus across the entire input video, resulting in the prediction of imprecisely overlapping actions. To address these challenges, we introduce the ConvTransformer Attention Network (CTAN), a novel framework comprised of two primary components: (1) The Temporal Attention Block (TAB), a temporal attention mechanism designed to emphasize critical temporal positions enriched with essential action-related features. (2) The ConvTransformer Block (CTB), which employs a hybrid structure for capturing nuanced appearance changes locally and action transitions globally. Facilitated with these components, CTAN is adept at focusing on motion features between overlapping actions, and precisely capturing both local differences between adjacent clips and global action transitions. The extensive experiments on multiple datasets, including THUMOS14, MultiTHUMOS, and ActivityNet, confirm the effectiveness of CTAN.

Developing Deep LSTMs With Later Temporal Attention for Predicting COVID-19 Severity, Clinical Outcome, and Antibody Level by Screening Serological Indicators Over Time.

The clinical course of COVID-19, as well as the immunological reaction, is notable for its extreme variability. Identifying the main associated factors might help understand the disease progression and physiological status of COVID-19 patients. The dynamic changes of the antibody against Spike protein are crucial for understanding the immune response. This work explores a temporal attention (TA) mechanism of deep learning to predict COVID-19 disease severity, clinical outcomes, and Spike antibody levels by screening serological indicators over time. We use feature selection techniques to filter feature subsets that are highly correlated with the target. The specific deep Long Short-Term Memory (LSTM) models are employed to capture the dynamic changes of disease severity, clinical outcome, and Spike antibody level. We also propose deep LSTMs with a TA mechanism to emphasize the later blood test records because later records often attract more attention from doctors. Risk factors highly correlated with COVID-19 are revealed. LSTM achieves the highest classification accuracy for disease severity prediction. Temporal Attention Long Short-Term Memory (TA-LSTM) achieves the best performance for clinical outcome prediction. For Spike antibody level prediction, LSTM achieves the best permanence. The experimental results demonstrate the effectiveness of the proposed models. The proposed models can provide a computer-aided medical diagnostics system by simply using time series of serological indicators.

DRLSTM-DA: a novel aeroengine remaining useful life prediction method based on deep optimization of multiple sensors information

Abstract One of the crucial tasks in Remaining Useful Life (RUL) prediction is to effectively extract key sensor information from numerous sensor signals. In this article, a novel dual-attention enhanced deep residual LSTM (DRLSTM-DA) is developed to deeply optimize multidimensional sensor signals. First, two LSTM layers are designed to compress and reconstruct representative degradation information from input multidimensional time series data, to generate a new feature space. Second, a novel channel adaptive soft threshold module is designed to assign different weights according to the importance of different sensor information, and simultaneously eliminate the noise information in the signal. Thirdly, a temporal attention mechanism is designed to automatically highlight moments containing important decay information while suppressing unimportant moments. Finally, compared with other SOTA methods, our architecture achieves RMSE of 11.55, 13.74, 11.25, and 14.19 on four sub-datasets of the C-MAPSS dataset, with scores of 234.24, 465.49, 202.23, and 537.66, respectively. Meanwhile, in real aeroengine operation dataset, our architecture achieved the smallest RMSE (8.62). These results all validate the good predictive performance of our model.

A Transformer-based visual object tracker via learning immediate appearance change

Transformer has shown its great strength in visual object tracking due to its effective attention mechanism, but most prevailing transformer-based trackers only explore temporal information frame by frame, thus overlooking the rich context information inherent in videos. To alleviate this problem, we propose a transformer-based tracker via learning immediate appearance change information in videos, called IAC-tracker. The proposed tracker enhances the perception of the immediate motion state to improve the performance of single target tracking. IAC-tracker contains three key components: a spatial information extractor (SIE) with a superior attention mechanism to progressively extract spatial information, a temporal information extractor (TIE) with a designed temporal attention mechanism to progressively learn target immediate appearance change, and a novel spatial–temporal context enhanced fusion module integrating the information from SIE and TIE to prepare for the final prediction head. Comparison experiments with state-of-the-art trackers on six challenging datasets demonstrate the superior performance of IAC-tracker with real-time running speed.

Correction: Chu et al. Temporal Attention Mechanism Based Indirect Battery Capacity Prediction Combined with Health Feature Extraction. Electronics 2023, 12, 4951

There was an error in the original publication [...]

Water Quality Prediction Model for the Pearl River Estuary Based on BiLSTM Improved with Attention Mechanism

To improve the accuracy and stability of water quality prediction in the Pearl River Estuary, a water quality prediction model was proposed based on BiLSTM improved with an attention mechanism. The feature attention mechanism was introduced to enhance the ability of the model to capture important features, and the temporal attention mechanism was added to improve the mining ability of time series correlation information and water quality fluctuation details. The new model was applied to the water quality prediction of eight estuaries of the Pearl River, and the prediction performance test, generalization ability test, and characteristic parameter expansion test were carried out. The results showed that:① The new model achieved high prediction accuracy in the water quality prediction of the Zhuhaidaqiao section. The root-mean-square error (RMSE) between the predicted value and the measured value was 0.004 1 mg·L-1, and the coefficient of determination (R2) was 98.3 %. Compared with that of Multi-BiLSTM, Multi-LSTM, BiLSTM, and LSTM, the results showed that the new model had the highest prediction accuracy, which verified the accuracy of the model. ② Both the number of training samples and the number of forecasting steps affected the prediction accuracy of the model, and the prediction accuracy of the model increased with the increase of the training samples. When predicting the total phosphorus of the Zhuhaidaqiao section, more than 240 training samples could obtain higher prediction accuracy. Increasing the number of prediction steps caused the prediction accuracy of the model to decline rapidly, and the reliability of the model prediction could not be guaranteed when the number of prediction steps was greater than 5. ③ When the new model was applied to the prediction of different water quality indexes in eight estuaries of the Pearl River, the prediction results had high precision and the model had strong generalization ability. The input data of upstream water quality, rainfall, and other characteristic parameters associated with the section prediction index of the object could improve the prediction accuracy of the model. Through many tests, the results showed that the new model could meet the requirements of precision, applicability, and expansibility of water quality prediction in the Pearl River Estuary and thus is a new exploration method for high-precision prediction of water quality in complex hydrodynamic environments.

Temporal Attention for Few-Shot Concept Drift Detection in Streaming Data

Concept drift describes unforeseeable changes in the underlying distribution of streaming data over time. Concept drift is a phenomenon in which the statistical properties of a target domain change over time in an arbitrary way. These changes might be caused by changes in hidden variables that cannot be measured directly. With the onset of the big data era, domains such as social networks, meteorology, and finance are generating copious amounts of streaming data. Embedded within these data, the issue of concept drift can affect the attributes of streaming data in various ways, leading to a decline in the accuracy and performance of models. There is a pressing need for new models to re-adapt to the changes in streaming data. Traditional concept drift detection algorithms struggle to effectively capture and utilize the key feature points of concept drift within complex time series, thereby failing to maintain the accuracy and efficiency of the models. In light of these challenges, this study introduces a novel concept drift detection method that incorporates a temporal attention mechanism within a prototypical network. By integrating a temporal attention mechanism during the feature extraction process, our approach enhances the capability to process complex time series data, preserves temporal locality, strengthens the learning of key features, and reduces the amount of labeled data required. This method significantly improves the detection accuracy and efficiency of small sample streaming data by better capturing the local features of the data. Experiments conducted across multiple datasets demonstrate that this method exhibits comprehensive leading performance in terms of accuracy and F1-score, with excellent recall and precision, thereby validating its effectiveness in enhancing concept drift detection in streaming data.

Water Flow Prediction Based on Improved Spatiotemporal Attention Mechanism of Long Short-Term Memory Network

The prediction of water plant flow should establish relationships between upstream and downstream hydrological stations, which is crucial for the early detection of flow anomalies. Long Short-Term Memory Networks (LSTMs) have been widely applied in hydrological time series forecasting. However, due to the highly nonlinear and dynamic nature of hydrological time series, as well as the intertwined coupling of data between multiple hydrological stations, the original LSTM models fail to simultaneously consider the spatiotemporal correlations among input sequences for flow prediction. To address this issue, we propose a novel flow prediction method based on the Spatiotemporal Attention LSTM (STA-LSTM) model. This model, based on an encoder–decoder architecture, integrates spatial attention mechanisms in the encoder to adaptively capture hydrological variables relevant to prediction. The decoder combines temporal attention mechanisms to better propagate gradient information and dynamically discover key encoder hidden states from all time steps within a window. Additionally, we construct an extended dataset, which preprocesses meteorological data with forward filling and rainfall encoding, and combines hydrological data from multiple neighboring pumping stations with external meteorological data to enhance the modeling capability of spatiotemporal relationships. In this paper, the actual production data of pumping stations and water plants along the East-to-West Water Diversion Project are taken as examples to verify the effectiveness of the model. Experimental results demonstrate that our STA-LSTM model can better capture spatiotemporal relationships, yielding improved prediction performance with a mean absolute error (MAE) of 3.57, a root mean square error (RMSE) of 4.61, and a mean absolute percentage error (MAPE) of 0.001. Additionally, our model achieved a 3.96% increase in R2 compared to the baseline model.