Long Short-term Memory Model Research Articles

A novel heat load prediction model of district heating system based on hybrid whale optimization algorithm (WOA) and CNN-LSTM with attention mechanism

Machine learning models, particularly long short-term memory (LSTM) networks, have been extensively employed for heat load prediction in district heating systems (DHS). Nevertheless, the over-reliance on default hyperparameter settings in most methods hinders further enhancement of prediction accuracy. A novel load prediction model is presented, which integrates the whale optimization algorithm (WOA) to refine the hyperparameters of an LSTM model bolstered by an attention mechanism (ATT) and convolutional neural network (CNN). Three hybrid models (WOA-CNN-ATT-LSTM, PSO-CNN-ATT-LSTM and GA-CNN-ATT-LSTM) are constructed by comparing WOA with particle swarm optimization (PSO) and genetic algorithm (GA). The proposed hybrid models are evaluated against traditional LSTM models using an 1100-hour dataset from a real DHS. The outcomes reveal that the WOA-CNN-ATT-LSTM model surpasses both the PSO-CNN-ATT-LSTM and GA-CNN-ATT-LSTM models in heat load prediction accuracy, achieving improvements of 1.9% and 3.2% respectively, and attaining the highest prediction accuracy (R2=0.9962, MSE=0.0001, MAE=0.0082). Moreover, the WOA-CNN-ATT-LSTM model demonstrates superior performance across various time scales (half-day, one-day, three-days, and one-week), highlighting its robustness in heat load prediction. This novel model adaptively adjusts its hyperparameters to identify the optimal configuration, thereby significantly augmenting the overall predictive capabilities of the model.

Water Demand Forecasting in Multiple District Metered Areas Based on a Multi-Scale Correction Module Neural Network Architecture

Short-term water demand forecasting (STWDF) for multiple spatially and temporally correlated District Metering Areas (DMAs) is an essential foundation for achieving more refined management of urban water supply networks. However, due to the greater uncertainty associated with specific DMA demand compared to overall water usage, accurately predicting STWDF poses significant challenges. This study introduces an innovative network architecture—the multi-scale correction module neural network, built upon Long Short-Term Memory (LSTM) networks and Convolutional Neural Networks (CNN) enhanced with Attention mechanisms—for simultaneous STWDF with a temporal resolution of one hour over a week for 10 DMAs located in a single city in northern Italy. This framework utilizes multivariate corrections to refine and enhance the output accuracy. The results reveal that, in comparison to traditional Gated Recurrent Unit or LSTM models, the proposed model with integrated correction modules, particularly those that leverage inter-DMA correlations, improves performance across all evaluation metrics by an average of 5%-20% per DMA. Additionally, it consistently delivers superior accuracy across three scenarios: single DMA forecasting, total water demand, and extreme conditions, while maintaining stable performance throughout. Furthermore, the interpretability analysis underscores the feasibility of this innovative structure and highlights the contribution of meteorological features to the predictive model in some DMA-level STWDF. The unified input-output framework elegantly simplifies the STWDF process across multiple DMAs, providing new insights and methodologies for future research in this domain.

Detecting student attention through electroencephalography signals: a comparative analysis of deep learning models

<span lang="EN-MY">In the landscape of educational technology, understanding and optimizing student attention is important to enhance student’s learning experience. This study explores the potential of using electroencephalography (EEG) signals for discerning students' attention levels during educational tasks. With a cohort of 30 participants, EEG data were meticulously collected and subjected to robust preprocessing techniques, including independent component analysis (ICA) and principal component analysis (PCA). The research then employed different deep learning algorithm such as long short-term memory (LSTM), recurrent neural network (RNN), gated recurrent unit (GRU), multi-layer perceptron (MLP), and convolutional neural network (CNN) classifiers to predict students' attention. The results reveal notable variations in the classifiers' predictive performance. Our finding revealed that the LSTM model emerged as the top performer and achieved 96% of the accuracy. This study not only contributes to the advancement of attention detection in educational technology but also underscores the importance of preprocessing methodologies, such as ICA and PCA, in optimizing the performance of deep learning models for EEG-based applications.</span>

Adaptive compensation using long short‐term memory networks for improved control performance in real‐time hybrid simulation

AbstractReal‐time hybrid simulation (RTHS) divides structural systems into numerical and experimental substructures, providing a cost‐effective solution for analyzing structural systems, especially those that are large or complex. However, the actuation systems between these substructures inevitably introduce delays, affecting the stability and accuracy of RTHS. To address this issue, this study proposes an adaptive compensation method based on a conditional adaptive time series (CATS) compensator and a long short‐term memory (LSTM) network, termed CATS‐LSTM. The LSTM model predicts actuator responses for parameter estimation and calculates prediction errors, improving control performance and reducing delays. The effectiveness of the proposed CATS‐LSTM method and the accuracy of the LSTM prediction are validated through a series of simulations and experiments. The results indicate that the proposed CATS‐LSTM method outperforms both the CATS and phase lead (PL) methods. Compared to the CATS method, the proposed method reduces the maximum delay, root mean square error, and peak error by 3 ms, 3.66%, and 4.78%, respectively, while achieving reductions of 12 ms, 8.4%, and 10.05%, compared to the PL method. Furthermore, the CATS‐LSTM method is significantly less sensitive to initial parameter estimates, compared to the CATS method, enhancing robustness and mitigating the effects of inaccurate or varying initial parameter estimates.

Using Long Short-Term Memory (LSTM) recurrent neural networks to classify unprocessed EEG for seizure prediction

ObjectiveSeizure prediction could improve quality of life for patients through removing uncertainty and providing an opportunity for acute treatments. Most seizure prediction models use feature engineering to process the EEG recordings. Long-Short Term Memory (LSTM) neural networks are a recurrent neural network architecture that can display temporal dynamics and, therefore, potentially analyze EEG signals without performing feature engineering. In this study, we tested if LSTMs could classify unprocessed EEG recordings to make seizure predictions.MethodsLong-term intracranial EEG data was used from 10 patients. 10-s segments of EEG were input to LSTM models that were trained to classify the EEG signal. The final seizure prediction was generated from 5 outputs of the LSTM model over 50 s and combined with time information to account for seizure cycles.ResultsThe LSTM models could make predictions significantly better than a random predictor. When compared to other publications using the same dataset, our model performed better than several others and was comparable to the best models published to date. Furthermore, this framework could still produce predictions significantly better than chance when the experimental paradigm design was altered, without the need to reperform feature engineering.SignificanceRemoving the need to perform feature engineering is an advancement on previously published models. This framework can be applied to many different patients’ needs and a variety of acute interventions. Also, it opens the possibility of personalized seizure predictions that can be altered to meet daily needs.

Performance Comparison of Deep Learning Algorithm for Autism Spectrum Disorder Prediction

Autism Spectrum Disorder is a complex neurodevelopmental condition that significantly impacts social interaction and conduct in individuals. Early diagnosis and timely intervention are crucial for effectively managing the condition and enhancing overall outcomes. This study concentrates on creating predictive models through deep learning algorithms to assist in the early identification of ASD. The study employs an extensive dataset that covers various features linked to ASD traits, such as speech and language development, learning disorders, genetic factors, and behavioral characteristics. The investigation encompasses the assessment of three recurrent neural network (RNN) architectures-standard RNN, Long Short Term Memory(LSTM), and Gated Recurrent Unit(GRU) –for their effectiveness in predicting ASD. The dataset is split into 80% for training and 20% for testing. This study compares RNN, LSTM, and GRU algorithms. The classification results show that the LSTM and GRU models exhibited similar accuracy, with values of 71.03% and 70.78% respectively.

Enhancing water security through advanced modeling: integrating deep learning and a novel metaheuristic optimization algorithm for accurate pan evaporation prediction

ABSTRACT This study evaluates and enhances machine learning models for predicting pan evaporation under diverse climatic conditions. Five fundamental machine learning models were employed and tested across four different stations. Subsequent comparisons were made with advanced techniques, including long short-term memory (LSTM) networks. An innovative approach was introduced, combining LSTM with Binary Al-Biruni Earth Radius (BER–LSTM). This hybrid method was benchmarked against other optimization techniques. The BER–LSTM model consistently outperformed other models across all stations and time scales, achieving up to a 97.54% improvement in root mean square error (RMSE) compared to standard LSTM on daily time scales. Compared to simpler models like Multilayer Perceptron and Support Vector Regressor, BER–LSTM showed even more substantial improvements, with up to a 99.03% reduction in RMSE. The BER–LSTM model demonstrates superior predictive capabilities for pan evaporation across varied climatic conditions, offering significant improvements over both traditional and advanced machine learning techniques. This approach shows promise for enhancing evaporation forecasting in diverse environmental contexts.

Deciphering Membrane Proteins Through Deep Learning Models by Revealing Their Locale Within the Cell

Membrane proteins constitute essential biomolecules attached to or integrated into cellular and organelle membranes, playing diverse roles in cellular processes. Their precise localization is crucial for understanding their functions. Existing protein subcellular localization predictors are predominantly trained on globular proteins; their performance diminishes for membrane proteins, explicitly via deep learning models. To address this challenge, the proposed study segregates membrane proteins into three distinct locations, including the plasma membrane, internal membrane, and membrane of the organelle, using deep learning algorithms including recurrent neural networks (RNN) and Long Short-Term Memory (LSTM). A redundancy-curtailed dataset of 3000 proteins from the MemLoci approach is selected for the investigation, along with incorporating pseudo amino acid composition (PseAAC). PseAAC is an exemplary technique for extracting protein information hidden in the amino acid sequences. After extensive testing, the results show that the accuracy for LSTM and RNN is 83.4% and 80.5%, respectively. The results show that the LSTM model outperforms the RNN and is most commonly employed in proteomics.

Mapping Crop Types for Beekeepers Using Sentinel-2 Satellite Image Time Series: Five Essential Crops in the Pollination Services

Driven by the widespread adoption of deep learning (DL) in crop mapping with satellite image time series (SITS), this study was motivated by the recent success of temporal attention-based approaches in crop mapping. To meet the needs of beekeepers, this study aimed to develop DL-based classification models for mapping five essential crops in pollination services in Quebec province, Canada, by using Sentinel-2 SITS. Due to the challenging task of crop mapping using SITS, this study employed three DL-based models, namely one-dimensional temporal convolutional neural networks (CNNs) (1DTempCNNs), one-dimensional spectral CNNs (1DSpecCNNs), and long short-term memory (LSTM). Accordingly, this study aimed to capture expert-free temporal and spectral features, specifically targeting temporal features using 1DTempCNN and LSTM models, and spectral features using the 1DSpecCNN model. Our findings indicated that the LSTM model (macro-averaged recall of 0.80, precision of 0.80, F1-score of 0.80, and ROC of 0.89) outperformed both 1DTempCNNs (macro-averaged recall of 0.73, precision of 0.74, F1-score of 0.73, and ROC of 0.85) and 1DSpecCNNs (macro-averaged recall of 0.78, precision of 0.77, F1-score of 0.77, and ROC of 0.88) models, underscoring its effectiveness in capturing temporal features and highlighting its suitability for crop mapping using Sentinel-2 SITS. Furthermore, applying one-dimensional convolution (Conv1D) across the spectral domain demonstrated greater potential in distinguishing land covers and crop types than applying it across the temporal domain. This study contributes to providing insights into the capabilities and limitations of various DL-based classification models for crop mapping using Sentinel-2 SITS.

Dengue Prediction using Short-Term Memory

Dengue is a severe infectious disease on the rise in Malaysia, and there is a demand for artificial intelligence to support the health system. However, the application of deep learning, specifically Long Short-Term Memory (LSTM) time series forecasting, has not been explored by many in dengue prediction studies. However, considering the availability of daily weather data being collected, the ability of LSTM to capture long term dependencies can be leveraged in forecasting dengue cases. Therefore, this study investigates the performance and viability of LSTM time series forecasting on predicting dengue cases. An LSTM model is developed and evaluated to be compared to a Support Vector Regression (SVR) model by utilising the availability of a dengue dataset consisting of weather and climate data. The results indicated LSTM time series forecasting performed better than SVR, with R2 and MAE scoring 0.75 and 8.76. In short, LSTM has shown better performance and, in addition, capturing trends in the rise and fall of dengue cases. Altogether, this research could contribute to the fight against the increase of dengue cases without relying on forecasted weather data but instead, history.

Novel statistically equivalent signature-based hybrid feature selection and ensemble deep learning LSTM and GRU for chronic kidney disease classification

Chronic kidney disease (CKD) involves numerous variables, but only a few significantly impact the classification task. The statistically equivalent signature (SES) method, inspired by constraint-based learning of Bayesian networks, is employed to identify essential features in CKD. Unlike conventional feature selection methods, which typically focus on a single set of features with the highest predictive potential, the SES method can identify multiple predictive feature subsets with similar performance. However, most feature selection (FS) classifiers perform suboptimally with strongly correlated data. The FS approach faces challenges in identifying crucial features and selecting the most effective classifier, particularly in high-dimensional data. This study proposes using the Least Absolute Shrinkage and Selection Operator (LASSO) in conjunction with the SES method for feature selection in CKD identification. Following this, an ensemble deep-learning model combining long short-term memory (LSTM) and gated recurrent unit (GRU) networks is proposed for CKD classification. The features selected by the hybrid feature selection method are fed into the ensemble deep-learning model. The model’s performance is evaluated using accuracy, precision, recall, and F1 score metrics. The experimental results are compared with individual classifiers, including decision tree (DT), Random Forest (RF), logistic regression (LR), and support vector machine (SVM). The findings indicate a 2% improvement in classification accuracy when using the proposed hybrid feature selection method combined with the LSTM and GRU ensemble deep-learning model. Further analysis reveals that certain features, such as HEMO, POT, bacteria, and coronary artery disease, contribute minimally to the classification task. Future research could explore additional feature selection methods, including dynamic feature selection that adapts to evolving datasets and incorporates clinical knowledge to enhance CKD classification accuracy further.

Long short term memory networks for predicting resilient Modulus of stabilized base material subject to wet-dry cycles.

The resilient modulus (MR) of different pavement materials is one of the most important input parameters for the mechanistic-empirical pavement design approach. The dynamic triaxial test is the most often used method for evaluating the MR, although it is expensive, time-consuming, and requires specialized lab facilities. The purpose of this study is to establish a new model based on Long Short-Term Memory (LSTM) networks for predicting the MR of stabilized base materials with various additives during wet-dry cycles (WDC). A laboratory dataset of 704 records has been used using input parameters, including WDC, ratio of calcium oxide to silica, alumina, and ferric oxide compound, Maximum dry density to the optimal moisture content ratio (DMR), deviator stress (σd), and confining stress (σ3). The results demonstrate that the LSTM technique is very accurate, with coefficients of determination of 0.995 and 0.980 for the training and testing datasets, respectively. The LSTM model outperforms other developed models, such as support vector regression and least squares approaches, in the literature. A sensitivity analysis study has determined that the DMR parameter is the most significant factor, while the σd parameter is the least significant factor in predicting the MR of the stabilized base material under WDC. Furthermore, the SHapley Additive exPlanations approach is employed to elucidate the optimal model and examine the impact of its features on the final result.

Fall Detection in Q-eBall: Enhancing Gameplay Through Sensor-Based Solutions

The field of physically interactive electronic games is rapidly evolving, driven by the fact that it combines the benefits of physical activities and the attractiveness of electronic games, as well as advancements in sensor technologies. In this paper, a new game was introduced, which is a special version of Bubble Soccer, which we named Q-eBall. It creates a dynamic and engaging experience by combining simulation and physical interactions. Q-eBall is equipped with a fall detection system, which uses an embedded electronic circuit integrated with an accelerometer, a gyroscopic, and a pressure sensor. An evaluation of the performance of the fall detection system in Q-eBall is presented, exploring its technical details and showing its performance. The system captures the data of players’ movement in real-time and transmits it to the game controller, which can accurately identify when a player falls. The automated fall detection process enables the game to take the required actions, such as transferring possession of the visual ball or applying fouls, without the need for manual intervention. Offline experiments were conducted to assess the performance of four machine learning models, which were K-Nearest Neighbors (KNNs), Support Vector Machine (SVM), Random Forest (RF), and Long Short-Term Memory (LSTM), for falls detection. The results showed that the inclusion of pressure sensor data significantly improved the performance of all models, with the SVM and LSTM models reaching 100% on all metrics (accuracy, precision, recall, and F1-score). To validate the offline results, a real-time experiment was performed using the pre-trained SVM model, which successfully recorded all 150 falls without any false positives or false negatives. These findings prove the reliability and effectiveness of the Q-eBall fall detection system in real time.

Health Monitoring and Prediction Using Non-invasive Sensors in IoT Environment

Introduction: IoT's blend with non-invasive sensors has made health state prediction possible and constant real-time monitoring transformed healthcare. Using photoplethysmography (PPG) and electrocardiograms (ECG), this device evaluates important parameters including heart rate, blood oxygen saturation, and blood pressure. As healthcare systems face increasing demands, this technology provides a patient-centered and efficient method for long-term health tracking—needed for early intervention and management of chronic diseases. Objectives: The main goal is to evaluate using time-series data from PPG and ECG sensors the efficiency of an IoT-enabled health monitoring system coupled with machine learning models in predicting health conditions including arrhythmias and hypoxia. Methods: This system detects events based on temporal pattern recognition by use of Long Short-Term Memory (LSTM) networks and Random Forest classifiers. Noise reduction is preprocessed and wearable device data is labeled for training of machine learning models. The design combines wearable sensors, a smartphone app, and cloud-based analysis for continuous monitoring and real-time input. Results: Experimental data shows that the LSTM model attained a 95% accuracy while the Random Forest classifier attained a 92.5% accuracy. Both models displayed remarkable accuracy and recall even while the LSTM model excels in identifying anomalies in time-series data and the Random Forest classifier skillfully detects health events depending on certain criteria. Conclusions: This study demonstrates the capability of IoT-based health monitoring systems to enhance patient outcomes through continuous, non-invasive monitoring and predictive analytics. Despite its potential, difficulties persist, including data security, sensor reliability, and energy efficiency, necessitating additional research for broad use and clinical integration.

Intelligent water quality prediction system with a hybrid CNN–LSTM model

ABSTRACT Water pollution remains a longstanding challenge globally, prompting substantial investment in water quality protection. The integration of advanced machine learning models offers promising avenues for accurate water quality prediction, enabling proactive measures to safeguard water sources. Presently, water quality assessment relies predominantly on physical and chemical metrics. This study developed MultiLayer Perceptron (MLP), eXtreme Gradient Boosting (XGBoost), long short-term memory (LSTM), and a hybrid CNN–LSTM model to forecast pH and dissolved oxygen (DO) levels. Results demonstrated the hybrid model's superior performance, with mean squared errors (MSEs) of 0.0015 and 0.0361 for pH and DO prediction, respectively. For Water Quality Classification (WQC), Random Forest (RF), k-Nearest Neighbors (kNNs), Support Vector Machine (SVM), and Light gradient Boosting Machine (Light GBM) were employed, with SVM achieving the highest accuracy at 88.75%. The research underscores the effectiveness of the CNN–LSTM model in predicting pH and DO levels. Leveraging these predictions as inputs to the SVM model offers valuable insights, particularly in regions where conventional monitoring methods face limitations. This streamlined approach, requiring only two parameters, signifies a significant advancement in accurate water quality prediction.

Combining LiDAR, SAR, and DEM Data for Estimating Understory Terrain Using Machine Learning-Based Methods

Currently, precise estimation of understory terrain faces numerous technical obstacles and challenges that are difficult to overcome. To address this problem, this paper combines LiDAR, SAR, and DEM data to estimate understory terrain. The high multivariable-precision spaceborne LiDAR ICESat-2 data, validated by the NEON, are divided into training and validation sets. The training dataset is used as a dependent variable, the SRTM DEM and Sentinel-1 SAR data are regarded as independent variables, a total of 13 feature parameters with high contributions are extracted to construct a Multiple Linear Regression model (MLR), BAGGING model, Random Forest model (RF), and Long Short-Term Memory model (LSTM). The results indicate that the RF model exhibits the highest accuracy among the four models, with R2 = 0.999, RMSE = 0.701 m, and MAE = 0.249 m. Then, based on the RF model, the understory terrain at the regional scale is generated, and an accuracy assessment is performed using the validation dataset, yielding R2 = 0.999, RMSE = 0.847 m, and MAE = 0.517 m. Furthermore, this paper quantitatively analyzes the effects of slope, vegetation coverage, and canopy height on the estimation accuracy of understory terrain. The results show that as slope, and canopy height increase, the estimation accuracy of the RF model for understory terrain gradually decreases. The accuracy of the understory terrain estimated by the RF model is relatively stable and not easily affected by slope, vegetation coverage, and canopy height. The research on the estimation of understory terrain holds significant practical implications for forest resource management, ecological conservation, and biodiversity protection, as well as natural disaster prevention.

Non-conventional feature-based LSTM model for prediction of bearing performance degradation

Abstract Mechanical bearings are critical in electrical machines, such as motors and generators. They provide support and facilitate the movement of rotating components in these machines. They are essential to the operation of electrical machinery cause they provide support, reduce friction, ensure proper alignment, and improve overall reliability and efficiency. The appropriate selection, installation, and maintenance of these bearings allow the longevity of such machines. The maintenance and the extent of usage determine the usefulness of the bearings. In predictive maintenance and condition monitoring, accurately predicting the level of bearing degradation holds significant importance for strategizing maintenance tasks, reducing downtime, and avoiding unforeseen breakdowns. This study centers on predicting the level of bearing degradation through the application of neural networks, focusing mainly on non-conventional features, rather than the usual features like kurtosis, and skewness. The dataset utilized is sourced from NASA and includes diverse time series signals that provide insights into the entire lifespan of the bearing. In this investigation, advanced machine learning techniques, long-short term memory (LSTM) in particular, are employed to harness the intricate patterns within the bearing dataset, aiming to improve the accuracy of the predictions.

NH4 Modelling with ARIMA and LSTM

AI-Mining is a prototype designed to detect various environmental gases, including CO2, NH3, NH4, and hydrogen, alongside temperature, pressure, and humidity. This study emphasizes the importance of modeling NH4 time series data due to its critical role in environmental and health monitoring. Accurate NH4 predictions facilitate early pollution detection and timely greenhouse gas control interventions. The study investigates the effectiveness of AI-Mining in detecting and predicting gas levels, focusing on data collection and analysis. Initial data analysis employed the Autoregressive Moving Average (ARIMA) model, specifically ARIMA (1,1,1), described by the equation yt = 0.0311 - 0.0750yt-1 + 0.3842εt-1. Despite its use, ARIMA's Root Mean Square Error (RMSE) performance was found lacking compared to more advanced methods. Given the classification of the obtained data as big data and time series, the Long Short-Term Memory (LSTM) method was also applied. The LSTM model initially used two layers with tanh and relu activation functions, and its performance was further explored by adding a third layer and varying the number of neurons (64, 128, and 256). The Adam optimizer was consistently used across all LSTM variations. Results indicated that increasing layers and neurons did not significantly impact LSTM's performance, with RMSE values around 0.023. However, LSTM consistently outperformed ARIMA in prediction accuracy, highlighting its robustness and reliability. Consequently, the study recommends using LSTM for predicting other recorded data in AI-Mining, underscoring its superiority in handling complex environmental datasets.

COMPARATIVE ANALYSIS OF RF, SVR WITH GAUSSIAN KERNEL AND LSTM FOR PREDICTING LOAN DEFAULTS

This investigation elucidates the paramount endeavour of predicting loan defaults, which is imperative for the efficacious management of financial risk and the overall stability of financial institutions. Conventional statistical methodologies frequently encounter challenges in effectively capturing the nonlinear and sequential dynamics inherent in financial data, thereby necessitating the examination of more sophisticated machine learning methodologies. This research reports an experimental-based comparative evaluation of three ML and DL models—Long Short-Term Memory (LSTM) networks, Random Forest (RF), and Support Vector Regression (SVR)—to assess their efficacy in forecasting loan defaults. The models are evaluated using metrics such as Mean Squared Error (MSE), F1 score, and Accuracy, and their proficiency in addressing imbalanced datasets and elucidating intricate data relationships is highlighted. The results indicate that while the Random Forest model surpasses its counterparts in terms of accuracy and MSE, the LSTM model exhibits considerable potential in managing imbalanced data, as evidenced by its stable F1 score. Although SVR reveals competitive precision, it exhibits deficiencies in addressing class imbalance. The ANOVA analyses substantiate that the disparities in model performance are statistically significant. The research acknowledges that both the LSTM and SVR models remain in the developmental stages, with ongoing initiatives aimed at refining these models through hyperparameter optimization and advanced architectural frameworks to enhance their predictive efficacy in practical applications.

Intelligent fault diagnosis in power distribution networks using LSTM-DenseNet network

This paper introduces a novel fault diagnosis method that combines DenseNet and Long Short-Term Memory (LSTM) networks. The DenseNet utilizes its unique dense block structure to detect subtle variations in three-phase voltage and zero-sequence current signals. In addition, the Squeeze-and-Excitation (SE) module is introduced in DenseNet. The SE module enhances DenseNet's feature representation by adapting the importance of each channel in the feature map. Furthermore, integrating the LSTM model enables capturing time-domain features of fault signals, enhancing the analysis of waveform changes and trends. These extracted features are subsequently fused in a cascaded manner, leveraging the strengths of both approaches to obtain a more comprehensive information representation. To better explain the capability of feature extraction in each part of the model, t-distributed Stochastic Neighbor Embedding (t-SNE) method is used for visual analysis. The proposed method is evaluated using two distribution network models, namely the 10 kV and IEEE34 networks, in simulation. The verification results indicate that the proposed method achieves exceptionally high accuracy in fault identification for both tested distribution network models, with rates of 99.87 % and 99.82 %, respectively, while also demonstrating robust performance in noisy environments. This performance surpasses that of other related methods, underscoring the enhanced effectiveness of our approach.