Short-term Memory Research Articles

FOXG1 Improves Cognitive Function in Alzheimer's Disease by Promoting Endogenous Neurogenesis.

Strategies aimed at enhancing the capacity of neural stem cells (NSCs) to generate multipotential, proliferative, and migratory cell populations capable of efficient neuronal differentiation are crucial for structural repair following neurodegenerative damage. The role of Forkhead-box gene 1 (FOXG1) in pattern formation, cell proliferation, and specification has been established. However, its involvement in Alzheimer's disease (AD) remains largely unknown. Here, we investigated the association between Foxg1 gene variants and AD-like behavioral deficits, amyloid-β (Aβ) aggregate formation, as well as p21 expression. Furthermore, we explored whether targeting the FOXG1-regulated cell cycle contributes to the promotion of adult neurogenesis in the context of AD. In this study, we successfully induced overexpression of FOXG1 in the hippocampus of AD brains through adeno-associated virus-Foxg1 infusion. Activation of FOXG1 rescued spatial learning disabilities, short-term memory deficits, and sensorimotor gating impairments observed in AD transgenic animals. By inhibiting p21 WAF1/cyclin-dependent kinase interacting protein 1 (p21cip1/waf1)-mediated cell cycle arrest, FOXG1 facilitates the activation and proliferation of NSCs. Additionally, the Foxg1 gene promotes an increase in precursor population size and enhances neuroblast differentiation. These combined effects on proliferation and differentiation lead to the generation of postmitotic neurons within the hippocampus in AD animals. Together, these findings demonstrate the importance of cooperation between FOXG1 and p21 for maintaining NSC self-renewal while facilitating neuronal lineage progression and contributing to endogenous neurogenesis during AD. Elevating levels of FOXG1 either pharmacologically or through alternative means could potentially serve as a therapeutic strategy for treating AD.

Leveraging explainable artificial intelligence with ensemble of deep learning model for dementia prediction to enhance clinical decision support systems

The prevalence of dementia is growing worldwide due to the fast ageing of the population. Dementia is an intricate illness that is frequently produced by a mixture of genetic and environmental risk factors. There is no treatment for dementia yet; therefore, the early detection and identification of persons at greater risk of emerging dementia becomes crucial, as this might deliver an opportunity to adopt lifestyle variations to decrease the risk of dementia. Many dementia risk prediction techniques to recognize individuals at high risk have progressed in the past few years. Accepting a structure uniting explainability in artificial intelligence (XAI) with intricate systems will enable us to classify analysts of dementia incidence and then verify their occurrence in the survey as recognized or suspected risk factors. Deep learning (DL) and machine learning (ML) are current techniques for detecting and classifying dementia and making decisions without human participation. This study introduces a Leveraging Explainability Artificial Intelligence and Optimization Algorithm for Accurate Dementia Prediction and Classification Model (LXAIOA-ADPCM) technique in medical diagnosis. The main intention of the LXAIOA-ADPCM technique is to progress a novel algorithm for dementia prediction using advanced techniques. Initially, data normalization is performed by utilizing min–max normalization to convert input data into a beneficial format. Furthermore, the feature selection process is performed by implementing the naked mole‐rat algorithm (NMRA) model. For the classification process, the proposed LXAIOA-ADPCM model implements ensemble classifiers such as the bidirectional long short-term memory (BiLSTM), sparse autoencoder (SAE), and temporal convolutional network (TCN) techniques. Finally, the hyperparameter selection of ensemble models is accomplished by utilizing the gazelle optimization algorithm (GOA) technique. Finally, the Grad‐CAM is employed as an XAI technique to enhance transparency by providing human-understandable insights into their decision-making processes. A broad array of experiments using the LXAIOA-ADPCM technique is performed under the Dementia Prediction dataset. The simulation validation of the LXAIOA-ADPCM technique portrayed a superior accuracy output of 95.71% over existing models.

Lightweight and universal deep learning model for fast proton spot dose calculation at arbitrary energies

Objective.To better integrate into processes like rapid adaptive planning and quality assurance, this study aims to propose a lightweight and universal proton spot dose calculation model suitable for arbitrary energies.Approach.Given the alignment between the characteristics of proton spot dose deposition and the sequence learning capabilities of the long short-term memory (LSTM) network, the lightweight model, multi-energy dose LSTM (MED-LSTM), is proposed. To comprehensively investigate the effectiveness of model, we trained and evaluated it on prostate, nasopharynx, and lung cases consistently.Main results. Regarding the results for spot dose, the prostate, nasopharynx, and lung models achieved average gamma passing rates of 99.93%, 99.81%, and 99.89% respectively under the (1%, 3 mm) criterion. Under the (1%, 1 mm) criterion, the rates were 99.06%, 97.18%, and 98.32%, respectively. For the intensity-modulated proton therapy plan dose, the prostate model achieved optimal performance with gamma passing rates of 99.88% and 98.52% under the (1%, 3 mm) and (1%, 1 mm) criteria, respectively. Following this, the lung model achieved rates of 99.22% and 93.41%. The nasopharynx model exhibited the poorest performance, with rates of 99.56% and 88.95%, respectively. It is evident that the MED-LSTM model demonstrates extremely high dose calculation accuracy in most cases. However, visible deviations occur in some spot samples for the nasopharynx and lung cases due to structural tissue differences.Significance.The MED-LSTM model can rapidly and accurately determine the proton spot dose at any energy with relatively low number of parameters. This exciting outcome holds broad prospects for applications and research directions.

Advanced smart assistance with enhancing social interaction and daily activities for visually impaired individuals using deep learning with modified seagull optimization

Visually impaired individuals face daily challenges in social engagement and routine activities due to limited access to real-time environmental information. Damage detection is a common approach in infrastructure that combines steel and concrete reinforcement to achieve optimal durability and structural strength. These bridges, designed to withstand diverse loads such as seismic forces, traffic weight, and environmental factors, are significant for maintaining structural integrity. Damage detection comprises applying advanced structural health monitoring methods to identify and assess potential deterioration or damage in concrete bridge components. Machine learning (ML) models, pattern detection, and statistical analysis are extensively adopted to identify subtle changes and process sensor information in structural response that might indicate corrosion, cracks, or other structural problems. Earlier detection and continuous monitoring of damage enable prompt intervention, ensuring longevity and safety while reducing the need for extensive repairs or the risk of unexpected failures. This study proposes an Automated Damage Detection using a Modified Seagull Optimizer with Ensemble Learning (ADD-MSGOEL) method for visually impaired people. The ADD-MSGOEL method is designed to enhance the social life and daily functioning of visually impaired people by accurately detecting damage and potential hazards in their surroundings. Initially, the ADD-MSGOEL method utilizes contrast enhancement (CLAHE) to enhance the image quality. Next, the features are extracted using the Dilated Convolution Block Attention Module with EfficientNet (DCBAM-EfficientNet) module, which derives the intrinsic and complex features. Moreover, the MSGO model is employed to choose the optimal parameter for the DCBAM-EfficientNet module. At last, an ensemble of three models, namely long short-term memory (LSTM), bidirectional gated recurrent unit (BiGRU), and sparse autoencoder (SAE) models, are implemented for the classification and detection of the damages. To demonstrate the effectiveness of the ADD-MSGOEL technique, a series of experiments were conducted using the CODEBRIM dataset. The experimental validation of the ADD-MSGOEL technique portrayed a superior accuracy value of 97.59% over existing models.

Integrating personalized shape prediction, biomechanical modeling, and wearables for bone stress prediction in runners

Running biomechanics studies the mechanical forces experienced during running to improve performance and prevent injuries. This study presents the development of a digital twin for predicting bone stress in runners. The digital twin leverages a domain adaptation-based Long Short-Term Memory (LSTM) algorithm, informed by wearable sensor data, to dynamically simulate the structural behavior of foot bones under running conditions. Data from fifty participants, categorized as rearfoot and non-rearfoot strikers, were used to create personalized 3D foot models and finite element simulations. Two nine-axis inertial sensors captured three-axis acceleration data during running. The LSTM neural network with domain adaptation proved optimal for predicting bone stress in key foot bones—specifically the metatarsals, calcaneus, and talus—during the mid-stance and push-off phases (RMSE < 8.35 MPa). This non-invasive, cost-effective approach represents a significant advancement for precision health, contributing to the understanding and prevention of running-related fracture injuries.

Neuromorphic Light-Responsive Organic Matter for in Materia Reservoir Computing.

Materials able to sense and respond to external stimuli by adapting their internal state to process and store information, represent promising candidates for implementing neuromorphic functionalities and brain-inspired computing paradigms. In this context, neuromorphic systems based on light-responsive materials enable the use of light as information carrier, allowing to emulate basic functions of the human retina. In this work it is demonstrated that optically-induced molecular dynamics in azopolymers can be exploited for neuromorphic-type of data processing in the analog domain and for computing at the matter level (i.e., in materia). Besides showing that azopolymers can be exploited for data storage, it is demonstrated that the adaptiveness of these materials enables the implementation of synaptic functionalities including short-term memory, long-term memory, and visual memory. Results show that azopolymers allow event detection and motion perception, enabling physical implementation of information processing schemes requiring real-time analysis of spatio-temporal inputs. Furthermore, it is shown that light-induced dynamics can be exploited for the in materia implementation of the unconventional computing paradigm denoted as reservoir computing. This work underscores the potential of azopolymers as promising materials for developing adaptive, intelligent photo-responsive systems that mimic some of the complex processing abilities of biological systems.

A modernized approach to sentiment analysis of product reviews using BiGRU and RNN based LSTM deep learning models

With the advent of Web 2.0 and popularization of online shopping applications, there has been a huge upsurge of user generated content in recent times. Leading companies and top brands are trying to exploit this data and analyze the market demands and reach of their products among consumers using opinion mining. Sentiment analysis is a hot topic of research in the e-commerce industry. This paper proposes such a novel sentence level sentiment analysis approach for mining online product reviews using natural language processing and deep learning techniques. The proposed model consists of various stages like web crawling and collecting product reviews, preprocessing, feature extraction, sentiment analysis and polarity classification. The input reviews are preprocessed using natural language processing techniques like tokenization, lemmatization, stop word removal, named entity recognition and part of speech tagging. Feature extraction is done using bidirectional gated recurrent unit shortly called as BiGRU feature extractor and the sentiments are classified into three polarities such as positive, negative and neutral using a hybrid recurrent neural network based long short-term memory classifier. The specific combination of techniques employed here and applying it to a new kind of online product review is making the proposed model to be novel. Performance evaluation metrics such as accuracy, precision, recall, F measure and AUC are calculated for the proposed model and compared with many existing techniques like deep convolutional neural network, multilayer perceptron, CapsuleNet and generative adversarial networks. The proposed model can be used in a variety of applications like market research, social network mining, recommendation systems, brand analysis, product quality management etc. and is found to generate promising results when compared to prevailing models.

Multi-Dimensional Feature Fusion and Enhanced Attention Streaming Movie Prediction Algorithm

Aiming at the challenges of multiple influencing factors, complex data characteristics, and limited data in streaming movie prediction, a feature fusion long- and short-term memory enhanced attention network (FFLSTMEA) was developed to achieve the short-term prediction of key indicators such as streaming movie revenue and to support business decisions. To address issues such as single data dimensions, difficulty in focusing on key information, limited data scale, and lack of diversity, several improvements were introduced. First, a feature fusion strategy was designed to integrate multi-dimensional features, including holiday factors, movie characteristics, principal component analysis (PCA) for time series dimensionality reduction, and platform exclusivity. These features were combined with a long- and short-term memory network to explore their internal correlations. Second, an attention mechanism was applied to dynamically assign importance to different time steps and features, enabling the model to focus on the most critical information based on time periods and movie types. Finally, the model’s capacity to capture data structures and variations was improved by using data augmentation techniques, such as flipping and scaling operations, to increase the dataset’s size and diversity. The experimental results show that the proposed algorithm FFLSTMEA achieves better prediction results with an average absolute error (MAE) of 3.50, a root mean square error (RMSE) of 5.28, and a coefficient of determination (R-squared) of 0.87 in the evaluation index. And compared with convolutional networks (CNN) class, long short-term memory (LSTM) class and Transformer class prediction methods, it performs better in terms of accuracy and stability, providing a more reliable basis for the operation and promotion of online movies.

Optimizing Space Heating in Buildings: A Deep Learning Approach for Energy Efficiency

Building energy management is crucial in reducing energy consumption and maintaining occupant comfort, especially in heating systems. However, achieving optimal space heating efficiency while maintaining consistent comfort presents significant challenges. Traditional methods often fail to balance energy consumption with thermal comfort, especially across multiple zones in buildings with varying operational demands. This study investigates the role of deep learning models in optimizing space heating while maintaining thermal comfort across multiple building zones. It aims to enhance heating efficiency by developing predictive models for building temperature and heating consumption, evaluating the effectiveness of different deep learning architectures, and analyzing the impact of model-driven heating optimization on energy savings and occupant comfort. To address this challenge, this study employs Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Transformer models to forecast area temperatures and predict space heating consumption. The proposed methodology leverages historical building temperature data, weather station measurements such as atmospheric pressure, wind speed, wind direction, relative humidity, and solar radiation, along with other weather parameters, to develop accurate and reliable predictions. A two-stage deep learning process is utilized: first, temperature predictions are generated for different building zones, and second, these predictions are used to estimate global heating consumption. This study also employs grid search and cross-validation to optimize the model configurations and custom loss functions to ensure energy efficiency and occupant comfort. Results demonstrate that the Long Short-Term Memory and Transformer models outperform the Gated Recurrent Unit regarding heating reduction, with a 20.95% and 20.69% decrease, respectively, compared to actual consumption. This study contributes significantly to energy management by providing a deep learning-driven framework that enhances energy efficiency while maintaining thermal comfort across different building areas, thereby supporting sustainable and intelligent building operations.

A Spatial Long-Term Load Forecast Using a Multiple Delineated Machine Learning Approach

Maintaining a balance between electricity generation and consumption is vital for ensuring grid stability and preventing disruptions. Spatial load forecasting (SLF) predicts geographical electricity demand, thereby aiding in power system planning. However, conventional methods like multiple linear regression and autoregressive integrated moving average struggle to capture the complex spatiotemporal patterns in historical data. Advanced methods like spatiotemporal graph transformers, graph convolutional networks, and improved scale-limited dynamic time warping better capture these dependencies, thereby enhancing prediction accuracy. Despite the advancements, challenges persist, particularly in developing economies with limited reliable data. This paper presents a novel SLF approach that divides the grid into predefined clusters based on regional characteristics and economic activity. For each cluster, a customized long short-term memory (LSTM) model captures unique spatiotemporal dependencies for more accurate predictions. The proposed method was tested across five load clusters using real-world data from Cameroon’s National Electricity Transmission Company and Energy Utilities. The results, compared against a linear regression model, demonstrated the superior performance of the LSTM approach across metrics like the mean absolute percentage error, root-mean-square error, mean absolute error, and R2 score. This approach highlights the potential for enhanced, localized load forecasting in regions with data constraints.

The analysis of artificial intelligence knowledge graphs for online music learning platform under deep learning

This work proposes a personalized music learning platform model based on deep learning, aiming to provide efficient and customized learning recommendations by integrating audio, video, and user behavior data. This work uses Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) networks to extract audio and video features, while using multi-layer perceptrons to encode user behavior data. To further improve the recommendation accuracy, this work constructs a knowledge graph that integrates key entities and their relationships in the music field, and fuses them with the extracted feature vectors. The knowledge graph provides the platform with rich semantic information and relational data, helping the model better understand the correlation between user needs and music content, thereby improving the accuracy and personalization of recommendation results. Experimental analysis based on different datasets shows that the proposed music recommendation platform performs well in multiple key performance indicators. Especially under different TOP-K conditions, the accuracy reaches 0.90, significantly exceeding collaborative filtering and content-based recommendation methods. In addition, the platform can maintain high accuracy when processing sparse data, demonstrating stronger robustness and adaptability. The platform has significant advantages in overall performance, providing users with more reliable and efficient recommendation services.

Enhanced Propaganda Detection in Public Social Media Discussions Using a Fine-Tuned Deep Learning Model: A Diffusion of Innovation Perspective

During the COVID-19 pandemic, social media platforms emerged as both vital information sources and conduits for the rapid spread of propaganda and misinformation. However, existing studies often rely on single-label classification, lack contextual sensitivity, or use models that struggle to effectively capture nuanced propaganda cues across multiple categories. These limitations hinder the development of robust, generalizable detection systems in dynamic online environments. In this study, we propose a novel deep learning (DL) framework grounded in fine-tuning the RoBERTa model for a multi-label, multi-class (ML-MC) classification task, selecting RoBERTa due to its strong contextual representation capabilities and demonstrated superiority in complex NLP tasks. Our approach is rigorously benchmarked against traditional and neural methods, including, TF-IDF with n-grams, Conditional Random Fields (CRFs), and long short-term memory (LSTM) networks. While LSTM models show strong performance in capturing sequential patterns, our RoBERTa-based model achieves the highest overall accuracy at 88%, outperforming state-of-the-art baselines. Framed within the diffusion of innovations theory, the proposed model offers clear relative advantages—including accuracy, scalability, and contextual adaptability—that support its early adoption by Information Systems researchers and practitioners. This study not only contributes a high-performing detection model but also delivers methodological and theoretical insights for combating propaganda in digital discourse, enhancing resilience in online information ecosystems.

Fatigue life predictor: predicting fatigue life of metallic material using LSTM with a contextual attention model.

Low-cycle fatigue (LCF) data involve complex temporal interactions in a strain cycle series, which hinders accurate fatigue life prediction. Current studies lack reliable methods for fatigue life prediction using only initial-cycle data while simultaneously capturing both temporal dependencies and localized features. This study introduces a novel deep-learning-based prediction model designed for LCF data. The proposed approach combines long short-term memory (LSTM) and convolutional neural network (CNN) architectures with an attention mechanism to effectively capture the temporal and localized characteristics of stress-strain data from acquisition through a series of cycle strain-controlled tests. Among the models tested, the LSTM-contextual attention model demonstrated superior performance (R 2 = 0.99), outperforming the baseline LSTM and CNN models with higher R 2 values and improved statistical metrics. The analysis of attention weights further revealed the model's ability to focus on critical timesteps associated with fatigue damage, highlighting its effectiveness in learning key features from LCF data. This study underscores the potential of deep-learning-based methods for accurate fatigue life prediction in LCF applications. This study provides a foundation for future research to extend these approaches to diverse materials with varying fatigue conditions and advanced models capable of incorporating non-linear fatigue mechanisms.

Abnormal traffic detection based on image recognition and attention-residual optimization

With the advancement of Internet of Things (IoT) technology, the continuous growth of IoT systems has resulted in the accumulation of massive amounts of data. Consequently, there has been a sharp increase in network attacks, highlighting the need for enhanced network security methods. Network intrusion detection systems play a crucial role in network security. Compared to the traditional approach of using single time-series models to process traffic data, this study innovatively proposes an RMCLA (Residual Network and Multi-scale Convolution Long Short-Term Memory with Attention Mechanisms) network intrusion detection system optimized with attention and residual mechanisms. This model converts traffic data into feature images and enhances the feature contrast through histogram equalization. It then utilizes the powerful performance of convolutional networks to extract abnormal feature points. The attention module and residual network enhance the focus on abnormal points, reducing feature loss and redundancy, thereby achieving effective classification of traffic image processing. We conducted experiments on the CIC-IDS2017 and UNSW-NB15 datasets and compared our model with the latest research models. This study highlights the potential of combining deep learning techniques with advanced attention and residual networks to enhance network security in IoT environments. The results show that combining image recognition with attention-residual optimization can effectively improve network intrusion detection capabilities.

A Variational Autoencoder Model Toward Molecular Structure Representation Learning of Fuels

Abstract In this work, a Variational Autoencoder (VAE)-based data-driven modeling framework is developed with the overarching goal of enabling fuel design. The VAE model is trained on a large dataset with several chemical species to learn a compressed latent space molecular representation. Chemical structure in the form of Simplified Molecular Input Line Entry System (SMILES) string is fed as input, encoded into the VAE latent space, and decoded back to the SMILES string using Long Short-Term Memory (LSTM) networks. Complexities of the VAE training loss function are thoroughly examined by varying the weightage (beta (𝜷) parameter) of the latent space regularization term, thereby assessing the balance between reconstruction accuracy and validity, and focusing on both accurate molecular structure reconstruction and latent space consistency. Two different strategies for 𝜷 variation are evaluated: linear annealing and cyclic annealing. In addition, the impact of total correlation adjustment and hierarchical priors is also studied with regard to the balance between reconstruction fidelity and latent space regularization, and potential issues such as posterior collapse, over-regularization, and poor disentanglement of latent variables. Overall, the best performance of the model is achieved with hierarchical priors and incrementally increasing 𝜷 from 0 to a threshold value of 0.25 over 75 epochs. The generative VAE model can be readily coupled with Quantitative Structure–Property Relationship (QSPR) analysis to develop an integrated end-to-end framework for fuel-property prediction and molecular design of novel promising fuels.

Energy-Efficient Fall-Detection System Using LoRa and Hybrid Algorithms

Wearable fall-detection systems have received significant research attention during the last years. Fall detection in wearable devices presents key challenges, particularly in balancing high precision with low power consumption—both of which are essential for the continuous monitoring of older adults and individuals with reduced mobility. This study introduces a hybrid system that integrates a threshold-based model for preliminary detection with a deep learning-based approach that combines a CNN (Convolutional Neural Network) for spatial feature extraction with a LSTM (Long Short-Term Memory) model for temporal pattern recognition, aimed at improving classification accuracy. LoRa technology enables long-range, energy-efficient communication, ensuring real-time monitoring across diverse environments. The wearable device operates in ultra-low-power mode, capturing acceleration data at 20 Hz and transmitting a 4-s window when a predefined threshold in the acceleration magnitude is exceeded. The CNN-LSTM classifier refines event identification, significantly reducing false positives. This design extends operational autonomy to 178 h of continuous monitoring. The experimental and systematic evaluation of the prototype achieved a 96.67% detection rate (sensitivity) for simulated falls and a 100% specificity in classifying conventional Activities of Daily Living as non-falls. These results establish the system as a robust and scalable solution, effectively addressing limitations in power efficiency, connectivity, and detection accuracy while enhancing user safety and quality of life.

Enhancing Computational Fluid Dynamics Simulations with Machine Learning: Techniques, Challenges, and Future Prospects

Recent studies have shown that the combination Machine Learning (ML) with the Computational Fluid Dynamics (CFD), can be considered as a revolutionary solution for the resolution of the well-known difficulty in fluid simulation such as the high computational costs and a complexity related to the use of traditional solvers. This study examines the levels of accuracy, efficiency, and scalability of CFD simulations that can be obtained from different ML models such as Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and Physics-Informed Neural Networks (PINNs). We evaluate each model in terms of mean squared error, structural similarity, inference time, and physical consistency, such as drag and lift coefficient prediction based on the benchmark datasets for steady and unsteady flows. CNNs achieved the highest balance between speed and accuracy for steady flows, but LSTMs evidenced the capacity of capturing temporal dynamics though they accumulated error over time. PINNs, although slower, offered long-term stability and generalization by incorporating physical laws in the learning process. The results suggest that although ML is not a complete substitute for traditional CFD, it provides significant tools for speeding up simulations and making possible real-time applications when used appropriately. Building upon the aforementioned discussion, this paper further explores the implications, limitations, and future directions of ML enhanced CFD presenting insights into the requirement of hybrid architectures, interpretability, and how data management strategy would be needed to implement these models in the mainstream engineering practices.

Developing an explainable deep learning module based on the LSTM framework for flood prediction

Long short-term memory (LSTM) networks have become indispensable tools in hydrological modeling due to their ability to capture long-term dependencies, handle non-linear relationships, and integrate multiple data sources but suffer from limited interpretability due to their black box nature. To address this limitation, we propose an explainable module within the LSTM framework, specifically designed for flood prediction across 531 catchments in the contiguous United States. Our approach incorporates a simplified gated module, which is interposed between the input data and the LSTM network, providing a transparent view of the module’s pattern recognition process. This gated module allows for easy identification of key variables and optimal lookback windows, and clusters the gated information into four categories: short-term and long-term impacts of precipitation and temperature. This categorization enhances our understanding of how the module utilizes input data and reveals underlying mechanisms in flood prediction. The modular design of our approach demonstrates high correlation with Saliency method, validating the credibility of its explanatory mechanisms, providing comparable interpretability features to LSTMs while illuminating key variables and optimal lookback windows considered most informative by hydrological models, and opening up new avenues for AI-assisted scientific discovery in the field.

Study on the Evolution of Groundwater Level in Hebei Plain to the South of Beijing and Tianjin Based on LSTM Model

This study addresses the limitations of machine learning in regional groundwater dynamics research, particularly the insufficient integration of the hydrogeological background and low simulation accuracy. Focusing on the shallow groundwater in the Hebei Plain south of Beijing and Tianjin, we integrate static data, including hydrogeological parameters, with the commonly used time-series data. A novel regionalization strategy based on depositional systems is proposed to enhance the model’s spatial adaptability. The Long Short-Term Memory (LSTM) model, augmented with an attention mechanism, adjusts the dynamic model weights using static data to reflect geological impacts on groundwater dynamics. Comparative results show that the refined regionalization and the inclusion of static data significantly improve the accuracy of the model. Based on the fitting results, the comparison of shallow groundwater level prediction between 2023 and 2040 under two mining conditions shows that the continuous implementation of the pressure mining policy has accelerated the recovery of water level, and the rise in groundwater level is obviously different between regions. The alluvial fan in the piedmont has the largest rise, and the marine sedimentary plain has the smallest rise. This study provides a new method for analyzing groundwater dynamics under complex hydrogeological conditions and provides a basis for regional groundwater management and sustainable utilization.

Predictive optimization using long short-term memory for solar PV and EV integration in relatively cold climate energy systems with a regional case study

The global shift toward sustainable energy and electric mobility addresses environmental concerns related to fossil fuels. While these alternatives are increasingly utilized in residential and commercial sectors, integrating renewable energy in building systems presents significant challenges. This is particularly evident in cold regions where unpredictable resource availability complicates energy reliability. The study emphasizes the need for innovative approaches to address these complexities and ensure consistent energy performance in dynamic conditions. This research explores the energy dynamics within a residential community located in a relatively cold climate region (Tabriz). The study begins by estimating the energy requirements of individual buildings, including the additional demand generated by electric vehicles. It then evaluates the potential for solar energy generation from photovoltaic systems. Finally, a machine learning-based approach (i.e., LSTM, Long Short-Term Memory) is employed to optimize the management of energy supply and demand across the community. This study demonstrates that heating demands in a cold climate are substantially higher than cooling needs, with solar energy providing sufficient (~ 32.1%) coverage during warmer months but requiring grid support in colder seasons. The prediction of EV charging patterns using LSTM models achieved over 93% accuracy, enabling improved energy demand forecasting and load management. These findings highlight the potential for optimizing renewable energy use, reducing grid dependency, and enhancing energy efficiency through effective production-demand management.