Short-term Memory Network Research Articles

Enhancing Bearing Fault Diagnosis With Deep Learning Model Fusion and Semantic Web Technologies

Given the limited accuracy of a singular deep learning model in bearing fault diagnosis, this study seeks to investigate and validate the efficacy of a deep learning model fusion strategy. It also aims to enhance the performance of deep learning models in bearing fault diagnosis using semantic web technology. Utilizing a publicly available bearing dataset, we employ semantic web to represent data in a structured format that is easily interpretable by machines. We then train separate convolutional neural network (CNN) and long short-term memory network (LSTM) models, and implement three distinct fusion strategies: voting fusion, weighted fusion, and stacking fusion. The experimental results indicate that the model fusion strategy significantly improves both the accuracy and robustness of fault diagnosis. Notably, the stacking fusion strategy outperforms the others in recognizing complex fault patterns. This study offers novel ideas and methodologies for the application of deep learning in the realm of bearing fault diagnosis.

Advancing Mental Health and Ethics Education for College Students

This study explores the implementation of the DeepPsy framework at a university in Hebei, China, assessing its impact on student mental health and ethical behavior. By integrating a 2D convolutional neural network (2D-CNN) with long short-term memory (LSTM) networks, DeepPsy analyzes complex online behavior patterns and academic performance data. The methodology involves diverse data collection techniques, providing a comprehensive understanding of students' mental health. Results indicate improved identification of at-risk students, leading to timely interventions and enhanced mental health outcomes. This study highlights the potential of advanced data analytics in fostering supportive educational environments that promote student mental health and ethical decision making.

Flow prediction via adaptive dynamic graph with spatio-temporal correlations

Flow prediction has shown its significance in optimizing and smoothing resource allocation. However, it still faces challenges due to inherent temporal variability and dynamic spatial correlations of flow data. Although many methods have been introduced to solve the problem, they often rely on predefined graphs to capture shared patterns, which is likely to overlook the critical spatiotemporal dependencies that are necessary for mid-term and long-term prediction tasks. To address the issue, a novel flow prediction model, termed Dynamic Graph Multi-Step Prediction via Long Short-Term Memory Network and Adaptive Learning (DMLSA), is proposed. DMLSA employs Node Adaptive Learning (NAL) to learn node-specific patterns without relying on predefined static adjacency matrices. Moreover, DMLSA utilizes Graph Adaptive Generation (GAG) to extract dynamic features from node attributes and automatically infer interdependencies among flow sequences, thus enhancing the ability to predict complex flow dynamics. Additionally, DMLSA integrates multi-head self-attention mechanisms with gated recurrent units, effectively capturing both long-term and short-term node dependencies, therefore allowing for more precise identification of spatiotemporal patterns at the individual node level. Extensive experiments on five datasets demonstrate that DMLSA outperformed existing baseline models in terms of accuracy and stability for multi-step prediction tasks.

Improved decision-making through life event prediction: A case study in the financial services industry

Life event prediction is an important tool for customer relationship management (CRM), because life events shift customers’ preferences towards different products and services. Existing life event research mainly uses cross-sectional data, whereas in the CRM field, incorporating longitudinal data is increasingly common. Because longitudinal data can capture the dynamics of customer behavior, opportunities arise to benchmark the power of longitudinal customer data for predictions of cross-sectional versus longitudinal life events. Therefore, this study compares statistical and machine learning (SaML) classifiers, such as logistic regression, random forest, and XGBoost, with long- and short-term memory networks (LSTM), using data represented in both cross-sectional and longitudinal setups for life event prediction. Through a real-life longitudinal customer data set from a European bank, the authors represent the longitudinal data in a cross-sectional data format, using featurization in the form of aggregation. The available data cover 42 end-of-month snapshots for 760,438 unique customers. For marketing decision-making literature, this article (1) introduces three novel life events (i.e., primary, secondary, and rental residence purchases) to life event predictions; (2) offers guidance for how to leverage longitudinal customer data, according to the comparison of various featurization approaches and benchmarking SaML classifiers against LSTM; and (3) clarifies the importance of features and timing for improving marketing decision-making dynamically. The results show that aggregating features over time is preferable as a featurization approach for cross-sectional modeling using SaML classifiers. Furthermore, LSTM can capture behavioral changes over time, unlike SaML classifiers. It also performs significantly better than SaML classifiers on the area under curve and F1 metrics. Insights into the uses of integrated gradients reveal that feature importance changes over time. An integrated gradients method can assist decision-makers in their efforts to plan effective communication with customers in advance, such as by allocating more resources to customers who exhibit high probabilities of a particular life event occurrence.

A Hybrid Deep Learning Approach with Generative Adversarial Network for Credit Card Fraud Detection

Credit card fraud detection is a critical challenge in the financial industry, with substantial economic implications. Conventional machine learning (ML) techniques often fail to adapt to evolving fraud patterns and underperform with imbalanced datasets. This study proposes a hybrid deep learning framework that integrates Generative Adversarial Networks (GANs) with Recurrent Neural Networks (RNNs) to enhance fraud detection capabilities. The GAN component generates realistic synthetic fraudulent transactions, addressing data imbalance and enhancing the training set. The discriminator, implemented using various DL architectures, including Simple RNN, Long Short-Term Memory (LSTM) networks, and Gated Recurrent Units (GRUs), is trained to distinguish between real and synthetic transactions and further fine-tuned to classify transactions as fraudulent or legitimate. Experimental results demonstrate significant improvements over traditional methods, with the GAN-GRU model achieving a sensitivity of 0.992 and specificity of 1.000 on the European credit card dataset. This work highlights the potential of GANs combined with deep learning architectures to provide a more effective and adaptable solution for credit card fraud detection.

The Use of eXplainable Artificial Intelligence and Machine Learning Operation Principles to Support the Continuous Development of Machine Learning-Based Solutions in Fault Detection and Identification

Machine learning (ML) revolutionized traditional machine fault detection and identification (FDI), as complex-structured models with well-designed unsupervised learning strategies can detect abnormal patterns from abundant data, which significantly reduces the total cost of ownership. However, their opaqueness raised human concern and intrigued the eXplainable artificial intelligence (XAI) concept. Furthermore, the development of ML-based FDI models can be improved fundamentally with machine learning operations (MLOps) guidelines, enhancing reproducibility and operational quality. This study proposes a framework for the continuous development of ML-based FDI solutions, which contains a general structure to simultaneously visualize and check the performance of the ML model while directing the resource-efficient development process. A use case is conducted on sensor data of a hydraulic system with a simple long short-term memory (LSTM) network. Proposed XAI principles and tools supported the model engineering and monitoring, while additional system optimization can be made regarding input data preparation, feature selection, and model usage. Suggested MLOps principles help developers create a minimum viable solution and involve it in a continuous improvement loop. The promising result motivates further adoption of XAI and MLOps while endorsing the generalization of modern ML-based FDI applications with the HITL concept.

Forecasting Lake Nokoué Water Levels Using Long Short-Term Memory Network

The forecasting of hydrological flows (rainfall depth or rainfall discharge) is becoming increasingly important in the management of hydrological risks such as floods. In this study, the Long Short-Term Memory (LSTM) network, a state-of-the-art algorithm dedicated to time series, is applied to predict the daily water level of Lake Nokoué in Benin. This paper aims to provide an effective and reliable method to enable the reproduction of the future daily water level of Lake Nokoué, which is influenced by a combination of two phenomena: rainfall and river flow (runoff from the Ouémé River, the Sô River, the Porto-Novo lagoon, and the Atlantic Ocean). Performance analysis based on the forecasting horizon indicates that LSTM can predict the water level of Lake Nokoué up to a forecast horizon of t + 10 days. Performance metrics such as Root Mean Square Error (RMSE), coefficient of correlation (R2), Nash–Sutcliffe Efficiency (NSE), and Mean Absolute Error (MAE) agree on a forecast horizon of up to t + 3 days. The values of these metrics remain stable for forecast horizons of t + 1 day, t + 2 days, and t + 3 days. The values of R2 and NSE are greater than 0.97 during the training and testing phases in the Lake Nokoué basin. Based on the evaluation indices used to assess the model’s performance for the appropriate forecast horizon of water level in the Lake Nokoué basin, the forecast horizon of t + 3 days is chosen for predicting future daily water levels.

A new methodology for establishing an SOC content prediction model that is spatiotemporally transferable at multidecadal and intercontinental scales

Quantifying and tracking the soil organic carbon (SOC) content is a key step toward long-term terrestrial ecosystem monitoring. Over the past decade, numerous models have been proposed and have achieved promising results for predicting SOC content. However, many of these studies are confined to specific temporal or spatial contexts, neglecting model transferability. Temporal transferability refers to a model’s ability to be applied across different periods, while spatial transferability relates to its applicability across diverse geographic locations for prediction. Therefore, developing a new methodology to establish a prediction model with high spatiotemporal transferability for SOC content is critically important. In this study, two large intercontinental study areas were selected, and measured topsoil (0–20 cm) sample data, 27,059 cloudless Landsat 5/8 images, digital elevation models, and climate data were acquired for 3 periods. Based on these data, monthly average climate data, monthly average data reflecting soil properties, and topography data were calculated as original input (OI) variables. We established an innovative multivariate deep learning model with high spatiotemporal transferability, combining the advantages of attention mechanism, graph neural network, and long short-term memory network model (A-GNN-LSTM). Additionally, the spatiotemporal transferability of A-GNN-LSTM and commonly used prediction models were compared. Finally, the abilities of the OI variables and the OI variables processed by feature engineering (FEI) for different SOC prediction models were explored. The results show that 1) the A-GNN-LSTM that used OI as the input variable was the optimal prediction model (RMSE = 4.86 g kg−1, R2 = 0.81, RPIQ = 2.46, and MAE = 3.78 g kg−1) with the highest spatiotemporal transferability. 2) Compared to the temporal transferability of the GNN, the A-GNN-LSTM demonstrates superior temporal transferability (ΔR2T = −0.10 vs. −0.07). Furthermore, compared to the spatial transferability of LSTM, the A-GNN-LSTM shows enhanced spatial transferability (ΔR2S = −0.16 vs. −0.09). These findings strongly suggest that the fusion of geospatial context and temporally dependent information, extracted through the integration of GNN and LSTM models, effectively enhances the spatiotemporal transferability of the models. 3) By introducing the attention mechanism, the weights of different input variables could be calculated, increasing the physical interpretability of the deep learning model. The largest weight was assigned to climate data (39.55 %), and the smallest weight was assigned to vegetation (19.96 %). 4) Among the commonly used prediction models, the deep learning model had higher prediction accuracy (RMSE = 6.64 g kg−1, R2 = 0.64, RPIQ = 1.78, and MAE = 4.78 g kg−1) and spatial transferability (ΔRMSES = 1.43 g kg−1, ΔR2S = −0.13, ΔRPIQS = −0.50, and ΔMAES = 1.09 g kg−1), and the linear model had the higher temporal transferability (ΔRMSET = 1.46 g kg−1, ΔR2T = −0.14, ΔRPIQT = −0.45, and ΔMAET = 1.29 g kg−1). 5) The deep learning models necessitated the OI, whereas the linear and traditional machine learning models necessitated FEI to achieve higher prediction accuracy. This study presents an important step forward in integrating multiple deep learning models to build a highly spatiotemporal transferability SOC prediction model.

Multi-temperature capable enhanced bidirectional long short term memory-multilayer perceptron hybrid model for lithium-ion battery SOC estimation

In this study, we propose a novel hybrid modeling framework for State of Charge (SOC) estimation across a broad temperature spectrum. First, we build a hybrid model to optimize stacked layers of stacked bidirectional long short term memory networks by introducing dropout mechanisms. At the same time, we also optimize the traditional multi-layer perceptron model to ResMLP, which is improved by introducing residual linkage, and then integrate the two optimization models. Finally, the synergistic effect and attention mechanism of genetic algorithm and particle swarm optimization are used to optimize its parameters. We then rigorously tested the model on nine datasets, including HPPC, DST and BBDST, at different temperatures of 5 °C, 15 °C and 35 °C. Using MAE, RMSE and MAXE benchmarks, our research results show that the proposed hybrid model outperforms the benchmark algorithm, achieving significantly enhanced performance and higher accuracy, and the maximum SOC estimation error is kept below 4.53 %. In addition, experimental evaluation at different temperatures shows the robustness and adaptability of the proposed algorithm.

Memristor-based circuit design of BiLSTM network

The bidirectional long short-term memory (BiLSTM) network involves significant amount of parameter computations. This paper proposes the memristor-based bidirectional long short-term memory (MBiLSTM) network, with its capability of in-memory computing and parallel computing, can accelerates the parameter computations speed. The MBiLSTM network circuit is composed of normalization circuit, two memristor-based long short-term memory (LSTM) circuits, memristor-based resnet circuit, memristor-based dense circuitand winner-take-all (WTA) circuit. The voltage signals are scaled to the setting range by normalization circuit, memristor-based LSTM circuit is responsible for extracting features from the dataset, memristor-based resnet circuit can enhance the overall performance of the network, memristor-based dense circuit ensures that the final outputs dimension matches the dimension of the target signals, WTA circuit outputs the maximum voltage of memristor-based dense circuit. The effectiveness of the MBiLSTM network is validated through gait recognition experiment and handwritten digit recognition experiment. The stability, robustness and potential errors in the manufacturing process of memristance are analyzed.

Classification and Recognition of Lung Sounds Using Artificial Intelligence and Machine Learning: A Literature Review

This review explores the latest advances in artificial intelligence (AI) and machine learning (ML) for the identification and classification of lung sounds. The article provides a historical overview from the invention of the electronic stethoscope to the auscultation of lung sounds, emphasizing the importance of the rapid diagnosis of lung diseases in the post-COVID-19 era. The review classifies lung sounds, including wheezes and stridors, and explores their pathological relevance. In addition, the article deeply explores feature extraction strategies, measurement methods, and multiple advanced machine learning models for classification, such as deep residual networks (ResNets), convolutional neural networks combined with long short-term memory networks (CNN–LSTM), and transformer models (transformer). The article discusses the problems of insufficient data and replicating human expert experience and proposes future research directions, including improved data utilization, enhanced feature extraction, and classification using spectrograms. Finally, the article emphasizes the expanding role of AI and ML in lung sound diagnosis and their potential for further development in this field.

Knowledge Representation Learning Method Based on Semantic Enhancement of External Information

Background: Knowledge representation learning aims at mapping entity and relational data in knowledge graphs to a low-dimensional space in the form of vectors. The existing work has mainly focused on structured information representation of triples or introducing only one additional kind of information, which has large limitations and reduces the representation efficiency. Objective: This study aims to combine entity description information and textual relationship description information with triadic structure information, and then use the linear mapping method to linearly transform the structure vector and text vector to obtain the joint representation vector. Methods: A knowledge representation learning (DRKRL) model that fuses external information for semantic enhancement is proposed, which combines entity descriptions and textual relations with a triadic structure. For entity descriptions, a vector representation is performed using a bi-directional long- and short-term memory network (Bi-LSTM) model and an attention mechanism. For the textual relations, a convolutional neural network is used to vectorially encode the relations between entities, and then an attention mechanism is used to obtain valuable information as complementary information to the triad. Results: Link prediction and triadic group classification experiments were conducted on the FB15K, FB15K-237, WN18, WN18RR, and NELL-995 datasets. Theoretical analysis and experimental results show that the DRKRL model proposed in this paper has higher accuracy and efficiency compared with existing models. Conclusion: Combining entity description information and textual relationship description information with triadic structure information can make the model have better performance and effectively improve the knowledge representation learning ability.

Sparse Graph Structure Fusion Convolutional Network for Machinery Remaining Useful Life Prediction

Effective prediction of machinery remaining useful life (RUL) is prominent to achieve intelligent preventive maintenance in manufacturing systems. In this paper, a sparse graph structure fusion convolutional network (SGSFCN) is proposed for more accurate end-to-end RUL prediction of machine. A novel node-level graph structure called time series shapelet distance graph (TSSDG) is designed to convert the time series to node feature. The SGSFCN model is proposed to learn degradation information from the graph structure. In SGSFCN, a sparse graph structure (SGS) layer and a fusion graph structure (FGS) layer preceding the graph convolutional network (GCN) are designed to learn the SGS from node representation and fuse the original graph structure, enabling the graph structure and node update iteratively in subsequent layers. Concurrently, a bidirectional long short-term memory network (BiLSTM) layer is integrated to capture the global temporal dependencies. The method is validated by two test rig data, and results demonstrate that the proposed method offers significantly higher prediction accuracy of RUL compared to several state-of-art methods.

Stock Price Prediction and Investment Simulation - Based on SSA-CNN-LSTM Method

Stock price fluctuations, which are time-series in nature. At the same time, based on the machine learning Long Short-Term Memory Network (LSTM) has excellent processing ability in predicting long time series, this paper proposes a stock price prediction method using CNN-LSTM optimized for LSTM hyperparameters using Sparrow algorithm. The data used in this study covers a total of 3403 trading days from March 1, 2010 to March 29, 2024 and about 40 technical indicator factors were selected. CNN model was firstly used to obtain features in the data. And later, CNN-LSTM network using SSA for parameter optimization uses the extracted features for stock price prediction and simulated trading based on the predicted up and down signals. The experimental results show that the SSA-CNN-LSTM network is able to provide better prediction accuracy than the CNN-LSTM network for price derivation. This prediction method not only provides a better op-eration idea for actual trading, but also provides practical experience for scholars to study time se-ries in finance.

Neural Machine Translation Between English And Bangla: A Philosophical Survey Observing Architectures And Performances

Neural Machine Translation (NMT) is a state of the art of machine translations that uses neural network models to translate text from one language to another. Unlike traditional machine translation methods, which rely on statistical models and rule-based systems, NMT leverages the power of deep learning, specifically using techniques such as recurrent neural networks (RNNs), long short-term memory networks (LSTMs), or more recently, transformers. Neural Machine Translation represents a significant dive forward in the field of automatic translation, leveraging the latest advancements in deep learning to deliver more accurate and natural translations. The progress in NMT has been remarkable, driven by advancements in model architectures, multilingual training, pretraining techniques, increased computational power, and the availability of vast multilingual datasets and improved evaluation metrics. These developments have collectively enhanced the quality, efficiency, and applicability of NMT systems across various languages and domains. In this article we will explore the technologies and performances of NMT between English and Bangla Language pair from its launching periods to present.

An omics-driven computational model for angiogenic protein prediction: Advancing therapeutic strategies with Ens-deep-AGP

Angiogenic proteins (AGPs) play a critical role in both pathological and physiological activities, making them key therapeutic targets in diseases like cancer, heart disease, and stroke. Traditional methods for identifying AGPs are labor-intensive and time-consuming, creating a need for more efficient approaches. This study addresses this challenge by developing a novel computational model, Ens-Deep-AGP, designed to enhance AGP prediction. The model introduces innovative feature engineering techniques, including Position Specific Scoring Matrix-Decomposition-Discrete Cosine Transform (PSSM-DC-DCT) and Position Specific Scoring Matrix-Auto-Cross-Discrete Wavelet Transform (PSSM-ACC-DWT), which capture comprehensive protein sequence information. The ensemble feature set of these approaches are then fed into Multi-headed Ensemble Residual Convolutional Neural Network (MERCNN), a robust deep learning architecture. Ens-Deep-AGP achieved remarkable accuracy rates of 99.79 % on training dataset and 92.97 % on testing dataset, surpassing Convolutional Neural Networks (CNN), Gated Recurrent Units (GRU), and Bidirectional Long Short-Term Memory Networks (BiLSTM). The successful prediction of AGPs is crucial for accelerating drug development, discovering novel therapeutic targets and deepen our understanding of AGPs' complex roles in healthcare.

Leak detection in water supply pipeline with small-size leakage using deep learning networks

Pipeline transportation technology has witnessed remarkable development in recent years, while the issue of leakage poses a serious limitation to this technology’s progress, especially for the small-scale leakage in water pipelines that exhibits weaker leakage signals. To address those challenges, this study investigates leak detection in water supply pipeline with an inner pipe diameter of 100 mm in the operating range of volume flow Q = 40–80 m3/h, leakage flow Ql = 0.65–1.33 m3/h and pressure p = 130–220 kPa. The characteristics and propagation of leakage pressure signals are analyzed. An adaptive improved variational mode decomposition (VMD) based on the white shark optimizer (WSO) algorithm is proposed, utilizing sample entropy as the fitness function. Additionally, tent chaotic mapping is employed for parameter initialization, mean normalization is used to minimize the influence of the DC component, and signal relative energy serves as a boundary condition. The processed data is utilized to establish leakage detection models using long short-term memory network (LSTM), gated recurrent unit (GRU), convolutional neural network (CNN) and SqueezeNet framework. The performance of leakage detection models is evaluated and compared using 962 sets’ data using evaluation indicators and confusion matrix. The results indicate that SqueezeNet-GRU demonstrates superior performance in pipeline leak detection and localization applications, achieving 97.40 % accuracy for the total dataset.

Dynamic mode decomposition and short-time prediction of PM2.5 using the graph Neural Koopman network

Analyzing and accurately predicting the spatiotemporal dynamics of PM2.5 remain challenging. The existing spatiotemporal prediction approaches are associated with high model complexity and limited interpretability. Conventional methods combining Koopman theory and deep learning often neglect spatial correlations in spatiotemporal data. This study used the hourly PM2.5 dataset of the Beijing-Tianjin-Hebei region to reveal its spatiotemporal hierarchy using Koopman mode decomposition to identify the key dynamic modes. Furthermore, a Spatial Physics Constrained Learning (SPCL) model utilizing a graph representation learning method was proposed to combine the graph topological information of the PM2.5 spatial features with the Koopman feature function. The results showed that PM2.5 has growth, decay, and oscillation modes as well as daily, weekly, monthly, and yearly periods. SPCL achieved mean absolute error, root mean square error (RMSE), correlation r, and index of agreement values of 9.678, 13.922, 0.864, and 0.921, respectively. The average RMSE at 12 h improved by 16.1%, 12.7%, 0.9%, and 3.5% compared with using Long short-term Memory, Graph Convolutional Networks and Long Short-Term Memory Networks, Spatio-Temporal Graph Convolutional Networks, and Dynamic Spatiotemporal Graph Convolution Network, respectively. By discretizing the neural network hidden layers, the explanatory key of PM2.5 modes was elucidated, which demonstrated enhanced stability.

Hourly PM2.5 concentration prediction for dry bulk port clusters considering spatiotemporal correlation: A novel deep learning blending ensemble model

Accurate prediction of PM2.5 concentrations in ports is crucial for authorities to combat ambient air pollution effectively and protect the health of port staff. However, in port clusters formed by multiple neighboring ports, we encountered several challenges owing to the impact of unique meteorological conditions, potential correlation between PM2.5 levels in neighboring ports, and coupling influence of background pollutants in city zones. Therefore, considering the spatiotemporal correlation among the factors influencing PM2.5 concentration variations within the harbor cluster, we developed a novel blending ensemble deep learning model. The proposed model combined the strengths of four deep learning architectures: graph convolutional networks (GCN), long short-term memory networks (LSTM), residual neural networks (ResNet), and convolutional neural networks (CNN). GCN, LSTM, and ResNet served as the base models aimed at capturing the spatial correlation of PM2.5 concentrations in neighboring ports, the potential long-term dependence of meteorological factors and PM2.5 concentrations, and the effects of urban ambient air pollutants, respectively. Following the blending ensemble technique, the prediction outcomes of three base models were used as the input data for the meta-model CNN, which employs the blending ensemble technique to produce the final prediction results. Based on actual data obtained from 18 ports in Nanjing, the proposed model was compared and analyzed for its prediction performance against six state-of-the-art models. The findings revealed that the proposed model provided more accurate predictions. It reduced mean absolute error (MAE) by 10.59 %–20.00 %, reduced root mean square error (RMSE) by 13.22 %–17.11 %, improved coefficient of determination (R2) by 10 %–35.38 %, and improved accuracy (ACC) by 3.48 %–7.08 %. Additionally, the contribution of each component to the prediction performance of the proposed model was measured using a systematic ablation study. The results demonstrated that the GCN model exerted the most substantial influence on the prediction performance of the GCN-LSTM-ResNet model, followed by the LSTM model. The influence of urban background pollutants can significantly enhance the generalizability of the complete model. Moreover, a comparison with three blended ensemble models incorporating any two base models demonstrated that the GCN-LSTM-ResNet model exhibited superior prediction performance and was particularly excellent in predicting the occurrence of high-concentration events. Specifically, the GCN-LSTM-ResNet model improved MAE and RMSE by at least 12.3% and 9.2%, respectively, but reduced R2 and ACC by 26.1% and 6.8%, respectively. The proposed model provided reliable PM2.5 concentration prediction outcomes and decision support for air quality management strategies in dry bulk port clusters.

Impacts of LULC changes on runoff from rivers through a coupled SWAT and BiLSTM model: A case study in Zhanghe River Basin, China

Changes in river runoff have a significant impact on the sustainable use of water resources in a watershed, and these changes are closely linked to variations in land use/land cover (LULC). This research explores an innovative approach in the Zhang River Basin (ZRB), China, by coupling a concept-based hydrological model, the Soil and Water Assessment Tool (SWAT), with a deep-learning model, the Bidirectional Long Short-Term Memory Network (Bi-LSTM), to improve the accuracy of river runoff simulations. By analyzing LULC changes in 2002, 2012, and 2022, this study developed three SWAT models and three coupled SWAT-BiLSTM models to quantitatively assess the impacts of these changes on river runoff through eight LULC scenarios. The findings revealed significant LULC changes from 2002 to 2022, with cropland and grassland areas decreasing while forest and urban land areas increased. The total area of grassland, forest, and cropland made up over 93 % of the basin, indicating active land type conversions. Calibration and validation results demonstrated that the SWAT-BiLSTM model outperformed the conventional SWAT model, yielding higher accuracy in runoff simulations. Specifically, the SWAT-BiLSTM model achieved R2 values of 0.89 and 0.90 during calibration and validation, compared to the SWAT model's R2 values of 0.76 and 0.79. Scenario analyses indicated that expansions in farmland, grassland, and urban areas were correlated with increased river runoff, while an expansion in forested areas led to reduced runoff. Notably, urban land changes had the most pronounced impact on runoff, emphasizing the need for careful runoff management and flood risk mitigation in urban planning. By combining SWAT and Bi-LSTM models, this study provides an innovative assessment of the impact of LULC changes on water resources in the ZRB. The results offer valuable insights for water resource management, LULC optimization, and flood risk management, highlighting the potential application of deep learning techniques in hydrological simulation. This research serves as a scientific basis for policy-making and sustainable land use planning in the ZRB and similar regions.