Short-term Memory Research Articles

Background: Obstructive sleep apnea (OSA) is an established trigger for sudden cardiac death (SCD), yet national mortality patterns and forward-looking risk remain incompletely defined. Methodology: U.S. death data (1999–2023) were obtained from CDC WONDER, including co-listed ICD-10 codes G47.3 (obstructive sleep apnea) and I46 (cardiac arrest). Age-adjusted mortality rates (AAMRs) per million were standardized to the 2000 U.S. census. Temporal trends were assessed using Joinpoint regression to calculate average annual percent change (AAPC). Forecasting was performed using a stacked long short-term memory (LSTM) model, validated against ARIMA baselines. Bootstrapped 95% prediction intervals were derived from 1,000 resamples. Results: From 1999 to 2023, 53,105 U.S. deaths involved both obstructive sleep apnea and cardiac arrest. The national age-adjusted mortality rate (AAMR) rose sixfold from 0.29 to 1.77 per 100,000 (AAPC: 7.21%; 95% CI: 6.61–8.04; P < 0.000001). All adult age groups saw rising mortality, with the steepest increases in those aged 85+ (AAPC: 11.60%) and 75–84 (AAPC: 8.71%). Rates also rose significantly in younger groups (35–74), with AAPCs ranging from 5.38% to 7.34%. Males peaked at an AAMR of 2.49 in 2021, dropping to 2.17 by 2023, while females plateaued at ~1.05 after 2021. White individuals had the highest burden (AAPC: 8.00%), followed by Black Americans (AAPC: 6.26%). Non-metropolitan counties exceeded metro areas in 2020 (1.76 vs. 1.57 per 100,000), with both showing strong growth (AAPCs: 8.45% and 8.67%). All U.S. regions saw sharp increases during the COVID-19 era. A long short-term memory (LSTM) neural network, validated against ARIMA, projects AAMRs to reach 2.74 by 2033 (projected APC ≈ 4.3%), aligning with ARIMA's AAPC of 4.33% for 2024–2035. The highest mortality rates occurred in Mississippi, Alabama, and Oklahoma, up to four times greater than in the lowest-burden states like Hawaii, Massachusetts, and California. Conclusion: OSA-related SCD mortality has risen markedly since 1999, with disproportionate burdens in men, Black Americans, and rural residents. Machine-learning projections indicate continued escalation over the next decade despite the recent plateau, highlighting an urgent need for earlier detection and treatment of OSA.

Introduction: Despite decreases in mortality, infants with single-ventricle heart disease remain at significant risk for morbidity-associated “red flag” events during the interstage period. As patients are typically discharged home during this time, preventive care is essential through proactive nursing assessments of remote (parent entered) monitoring data. Preemptive identification of complex cyanotic complications remains challenging to identify from vital signs alone. Patient videos capturing movement, color, and status holds potential to enhance nursing assessments in asynchronous remote patient monitoring. Assessing subtle measures of risk from video data in a non-trivial task. Hypothesis: In addition to physiologic data, computer vision machine learning (ML) approaches applied to videos are a feasible method for aiding proactive, personalized video review of parent-obtained, interstage infant characteristics by a nursing care team. Methods: A retrospective multi-site cohort was obtained from the CHAMP® repository (3/2014 – 12/2022), including infants with at least one video prior to Glenn surgery or death. For each eligible video, thirty-three 3D pose landmarks of major body points were detected using MediaPipe, an open source pose mapping toolkit. Processed data was used to train a long short-term memory (LSTM) model pipeline to predict if an event occurred within 28 hours of the video upload time. Results: Infants (n=494)- demographics in Table 1- from 10 institutions with 4,858 candidate videos had a computer-vision and ML pipeline successfully applied. The team was able to extract, process, and score event risk from parent uploaded videos. Each video was ranked by the likelihood of experiencing an event and performance was evaluated using lift, focused on identifying events relative to random. However, the LSTM model, trained solely on pose landmarks, offered no improvement for identifying imminent red flag events. Conclusion(s): The ability to successfully capture pose and movement data from parent videos was confirmed and proves to be a promising adjunct to a full remote nursing assessment to augment parent-only reported red flags for high-risk congenital heart disease patients. The low predictive power of red-flag events alone encouraged current work to incorporate vitals signs, demographics, facial landmark features, respiratory effort, and skin tone to determine if the ML model can be further trained to aid in prioritizing video review.

We propose a novel deep learning approach to asset pricing that predicts individual stock returns using daily data while integrating no-arbitrage constraints and capturing market dynamics. Our model combines Long Short-Term Memory (LSTM) networks, Convolutional Neural Networks (CNN), and Generative Adversarial Networks (GAN) to model the complex relationships between stock returns and various conditioning variables, including macroeconomic indicators, technical indicators, and market sentiment data. By incorporating the no-arbitrage condition into the deep learning framework, we enhance the accuracy and stability of asset pricing. We estimate a stochastic discount factor that explains asset returns from the conditional moment constraints implied by no-arbitrage. Our method outperforms traditional multi-factor models, such as the Fama-French model, in terms of Sharpe ratio, explained variation, and pricing errors. The GAN enforces the no-arbitrage constraint by identifying portfolio strategies that contain the most pricing information. The LSTM network uncovers hidden economic states, while the feedforward network captures the non-linear effects of conditioning variables. This research provides a new direction in asset pricing by applying deep learning to integrate market dynamics and enforce no-arbitrage constraints, offering more accurate pricing and valuable insights for generating profitable investment strategies.

Introduction: Predictive analytics powered by artificial intelligence (AI) and machine learning (ML) are revolutionizing cardiovascular risk assessment. Accurate prediction of low-density lipoprotein cholesterol (LDL-C) is critical for evaluating cardiovascular disease (CVD) risk and guiding therapeutic decisions. This study evaluates deep learning (DL) models for LDL-C prediction in patients with prior cardiovascular events, comparing their performance against traditional ML methods and established LDL-C estimation formulas. Methods: We retrospectively analyzed data from 8,315 patients with documented cardiovascular events from Rhythm Heart and Critical Care. Key lipid parameters included LDL-C, triglycerides (TG), total cholesterol (TC), and high-density lipoprotein cholesterol (HDL-C). Patient CVD history was blinded during model training to ensure unbiased prediction. DL models tested included Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM) networks, and a Transformer-based architecture. These were benchmarked against Back Propagation Neural Network (BPNN) models and LDL-C formulas by Sampson and Martin. Model performance was assessed using Root Mean Squared Error (RMSE) and Mean Absolute Percentage Error (MAPE). Results: The models generated LDL-C predictions for 5,132 patients (61% of the cohort). The Transformer-based model achieved the highest accuracy with an RMSE of 10.58 mg/dL and MAPE of 7.35%, significantly outperforming BPNN (RMSE 17.16 mg/dL; MAPE 11.01%), RNN (RMSE 32.47 mg/dL), and LSTM (RMSE 32.51 mg/dL). Deep learning models also surpassed traditional LDL-C formulas in accuracy. Partial Dependence Plots (PDP) of the Transformer model revealed clinically meaningful relationships between LDL-C and predictors such as HDL-C, BMI, and thyroid hormones, supporting physiological validity and interpretability. Conclusion: This study demonstrates that DL models, particularly the Transformer-based approach, significantly outperform conventional methods in predicting LDL-C levels among patients with cardiovascular events. The model’s superior accuracy and interpretability offer a promising clinical tool for personalized risk assessment, early detection, and optimized management of CVD. Incorporation of such AI-driven models into clinical workflows could improve patient outcomes and resource allocation in cardiovascular care.

Electrocorticography (ECoG) signals provide a valuable window into neural activity, yet their complex structure makes reliable classification challenging. This study addresses the problem by proposing a feature-selective framework that integrates multiple feature extraction techniques with statistical feature selection to improve classification performance. Power spectral density, wavelet-based features, Shannon entropy, and Hjorth parameters were extracted from ECoG signals obtained during a visual task. The most informative features were then selected using analysis of variance (ANOVA), and classification was performed with several machine learning methods, including decision trees, support vector machines, neural networks, and long short-term memory (LSTM) networks. Experimental results show that the proposed framework achieves high accuracy across individual patients as well as the combined dataset, with clear separability between classes confirmed through t-SNE visualization. In addition, analysis of selected features highlights the prominent role of electrodes located near the visual cortex, providing insights into the spatial distribution of neural activity.

Globally, undiagnosed chronic kidney disease (CKD) is a prevalent asymptomatic disease that leads to a significant morbidity and early mortality burden. Current researches have shown that heart issues, identified to as Cardio-Renal Syndrome (CRS) in research, often emerge in patients with renal disease. This disease has the potential to cause sudden cardiac arrest in its later stages. Research on patients with cardio-vascular issues to determine whether their kidneys are affected is valuable, as chronic kidney disease and cardio-vascular disorders are closely linked. Early diagnosis of CKD can enable patients to slow or even reverse disease progression with the help of medicinal interventions. Therefore, in this study we developed an Enhanced Deep Learning (DL)-based technique for the automatic identification of CKD. Digitized electrocardiogram (ECG) data is gathered from Physionet Database's Fantasia (healthy individuals) and PTB (kidney patients). In order to eliminate noise from the ECG measurements, an adaptive median filter is utilized. The significant features are extracted from preprocessed ECG signals. Next, the extracted features are sent into the Enhanced Attention Mechanism with a Long Short-Term Memory (EALSTM) model to classify if a signal is abnormal (CKD) or normal (Non-CKD). To enhance the effectiveness of the EALSTM, its hyper-parameters are optimized using the Adaptive Dingo Optimization (ADO) algorithm. Based to the results of the experiment, the recommended method achieves outstanding 97.65% accuracy, 98.83% precision, 99.21% sensitivity, 98.04% specificity, 99.22% recall, and 99.02% f-measure. The results indicated that the proposed method significantly outperforms other state-of-the-art methods.

Correction To: Modelling the seismic activity of Kahramanmaraş, Türkiye with recurrent neural network (RNN) and long short-term memory (LSTM) methods

Background: Hypertension is a leading contributor to stroke related events, yet most health systems lack predictive infrastructure to identify at-risk individuals early enough for preventive action. In collaboration with Emory Healthcare’s informatics division, Guidehealth developed a risk stratification model to identify patients with hypertension most likely to experience adverse cerebrovascular events and benefit from targeted interventions. Objective: To evaluate the predictive performance and clinical utility of a novel risk stratification algorithm: (1) to identify hypertensive patients at high risk for a cerebrovascular event (2) to estimate likelihood of successful intervention based on clinical and social context. Methods: Guidehealth built machine learning models using longitudinal data from 197,967 Medicare-eligible hypertensive seniors to a feed-forward neural network with long short-term memory predict adverse cerebrovascular events. Time series were subsampled with a 24-month lookback and prediction interval over the following 6-12 months. Additive temporal encoding preserved chronicity and exposure. Features included comorbidities, medication adherence, labs/imaging, and utilization trends—capturing both static and time-varying variables from claims data. Outcomes were 6–12-month stroke admissions. Outputs prioritized outreach and modifiable drivers of risk. Results: Among flagged patients in the historical validation set, >98% a cerebrovascular event within the timeframe. Model specificity (98%) was prioritized over sensitivity (30%) due to the cost and resource allocation. A patient prioritization dashboard enabled targeted outreach and prospective monitoring. Conclusion: Whole Patient Targeting represents a powerful advance in stroke prevention and represents a shift toward anticipatory health. By identifying high-risk and high-impactability patients, the model offers a scalable method to reduce avoidable cerebrovascular events and enable more effective, person-centered care for aging populations.

Introduction: Successful resuscitation of ventricular fibrillation (VF) out-of-hospital cardiac arrest (OHCA) relies on timely defibrillation and minimally-interrupted CPR. Defibrillator shock decision analysis has traditionally required CPR interruption because CPR causes electrical artifacts in the ECG signal. Recent emerging defibrillator algorithms have been proposed to reduce or eliminate CPR interruption for shock decision analysis. However, these methods are challenged by lower sensitivity, a high proportion of indeterminates, or requirement for CPR-free rhythm confirmation. Aim: We sought to determine whether a deep learning algorithm can accurately detect shockable rhythms during CPR. Methods: We performed a retrospective cohort study of adult VF-OHCA cases in King County WA from 2006-2021. Patients were randomized into training (60%), validation (20%), and test (20%) groups. We annotated the entirety of defibrillator paddle ECG recordings from cohort patients as non-shockable (Asystole, Organized Rhythms) or shockable (VF, Ventricular Tachycardia). ECGs were segmented into non-overlapping 2.5-s clips. The presence of CPR was confirmed by review of thoracic impedance. The algorithm comprised two steps: (1) A deep convolutional neural network predicted individual clip classes based on ECG scalogram images, and (2) a deep long short-term memory recurrent network incorporated the sequence of prior clip predictions to inform each clip’s current-time prediction. Results: Of 2682 eligible patients, N=2011 (75%) with available defibrillator files were included in the cohort; 1207, 402, and 402 patients were used for algorithm training, validation, and test, respectively. A total of 1047601 2.5-s ECG clips were collected from the cohort, with 604484 (58%) collected during CPR. During CPR, algorithm sensitivity/specificity for detecting shockable rhythms in training, validation, and test data were 99.0%/99.0%, 97.4%/98.9%, and 99.1%/98.7% respectively (Table 1). Of the CPR clips, 99628 (16.5%) were predicted as indeterminate by the algorithm and not scored. When indeterminate decisions were disallowed, algorithm sensitivity/specificity values in training, validation, and test groups were 92.7%/98.7%, 90.8%/97.6%, and 91.8%/97.9%, respectively. Conclusions: A deep learning algorithm developed using >1 million ECG segments can accurately detect shockable rhythms during CPR, suggesting potential to reduce CPR interruption and improve VF-OHCA resuscitation.

The real-time, precise monitoring of physiological signals such as intracranial pressure (ICP) and arterial blood pressure (BP) holds significant clinical importance. However, traditional methods like invasive ICP monitoring and invasive arterial blood pressure measurement present challenges including complex procedures, high infection risks, and difficulties in continuous measurement. Consequently, learning-based prediction utilizing observable signals (e.g., BP/pulse waves) has emerged as a crucial alternative approach. Existing models struggle to simultaneously capture multi-scale local features and long-range temporal dependencies, while their computational complexity remains prohibitively high for meeting real-time clinical demands. To address this, this paper proposes a physiological signal prediction method combining composite feature preprocessing with multiscale modeling. First, a seven-dimensional feature matrix is constructed based on physiological prior knowledge to enhance feature discriminative power and mitigate phase mismatch issues. Second, a network architecture CNN-LSTM-Attention (CBAnet), integrating multiscale convolutions, long short-term memory (LSTM), and attention mechanisms is designed to effectively capture both local waveform details and long-range temporal dependencies, thereby improving waveform prediction accuracy and temporal consistency. Experiments on GBIT-ABP, CHARIS, and our self-built PPG-HAF dataset show that CBAnet achieves competitive performance relative to bidirectional long short-term Memory (BiLSTM), convolutional neural network-long short-term memory network (CNN-LSTM), Transformer, and Wave-U-Net baselines across Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (R2). This study provides a promising, efficient approach for non-invasive, continuous physiological parameter prediction.

The emergence of artificial intelligence (AI) in digital communication has significantly transformed content generation, particularly through its ability to produce text that closely resembles human writing. This advancement has revolutionized content creation across various online platforms, leading to the pervasive presence of AI-generated text that blurs the lines between human and machine authorship. This study examines the effectiveness of various machine learning and deep learning models in identifying AI-generated essays. A dataset comprising 44,852 labeled essays, consisting of 27,361 human-written and 17,491 AI-generated instances, was utilized for the analysis. Classical models such as Naive Bayes, Logistic Regression, Random Forest, and XGBoost were compared to deep learning models, including Feedforward Neural Networks (FNN), Recurrent Neural Networks (RNN), and Long Short-Term Memory (LSTM) networks. Furthermore, an ensemble model combining RNN and LSTM through soft voting was developed. The performance of these models was evaluated based on accuracy, precision, recall, and F1-score. Notably, the ensemble model achieved the highest accuracy, reaching 95.0%, and outperformed all individual models. The results indicated that deep learning models, particularly LSTM and RNN, significantly surpassed classical approaches. The ensemble model demonstrated superior performance across all evaluation metrics, showcasing the advantages of hybrid methodologies. This research highlights the feasibility of automated detection systems and sets a benchmark for future advancements in AI text detection.

Information and communication technology (ICT) is crucial for maintaining efficient communications, enhancing processes, and enabling digital transformation. As ICT becomes increasingly significant in our everyday lives, ensuring its security is crucial for maintaining digital trust and resilience against evolving cyber threats. These technologies generate a large amount of data that should be analyzed simultaneously to detect threats to an ICT system and protect the sensitive information it may contain. NetFlow is a network protocol that can be used to monitor network traffic, collect Internet Protocol (IP) addresses, and detect anomalies in NetFlow. The article follows the design science research (DSR) methodology to reach an objective of providing a methods for developing a set of features for NetFlow analysis with a machine learning. The sets of features were analyzed and validated by implementing anomaly detection with the K-means clustering algorithm and time-series forecasting using the long short-term memory (LSTM) method. The study provides two separate sets of features for both machine learning methods (24 features for clustering and 14 for LSTM), an overview of the anomaly detection methods used in this research and a method to combine both machine learning approaches. Furthermore, this study introduces a method that integrates the outputs of both models and evaluates the reliability of the final decision based on Bayes' theorem and previous performance of the models.

This research is based on a systematic review of machine learning (ML) approaches for the cognitive load (CL) assessment of applications for unmanned aerial system (UAS) operator training. The review synthesises evidence on how ML techniques have been applied to assess CL using diverse data sources, including physiological signals (e.g., EEG, HRV), behavioural measures (e.g., eye-tracking), and performance indicators. It highlights the effectiveness of models such as Support Vector Machines (SVMs), Random Forests (RFs), and advanced deep learning (DL) architectures such as Long Short-Term Memory (LSTM), as well as how the use of different methods affects the performance of ML models, with studies reporting accuracies of up to 98%. The findings also indicate that, compared with traditional UAS training approaches, ML approaches can enhance training by providing adaptive assessment, with methodological factors such as model selection, data preprocessing, and validation being central to ML assessment performance. These findings highlight the value of accurate CL assessment as a foundation for adaptive training systems, supporting enhanced UAS operator performance and operational safety. By consolidating the methodological insights and identifying research gaps, this review provides valuable background information for advancing ML-based CL assessment and its integration into adaptive UAS operator training systems to enhance UAS operator training.

Lithium-ion batteries are high-performance energy storage devices that have been widely used in a variety of applications. Accurate early-stage prediction of their remaining useful life is essential for preventing failures and mitigating safety risks. This study proposes a novel multiview approach for estimating the State-of-Health (SOH) of lithium-ion batteries by integrating time-domain and time–frequency features. Firstly, time-domain signals are transformed into time–frequency images using a wavelet transform. Three representative features are then selected and converted into grayscale images, which are combined into three-channel color images as inputs for a convolutional neural network (CNN) to extract spatial features. These features are subsequently passed into a long short-term memory (LSTM) network to capture spatial dependencies. In parallel, raw temporal features are processed through a two-stage attention mechanism to explore both temporal and spatial correlations, followed by another LSTM to model temporal dependencies. The outputs from the two branches are fused using weighted integration and passed through a fully connected layer to generate the final SOH estimate. Comparative experiments with four baseline models demonstrate that the proposed time–frequency fusion architecture significantly enhances prediction accuracy, and that each component makes a meaningful contribution to the overall performance.

Nonverbal tests assess cognitive ability in multicultural and multilingual children, but language-based instructions disadvantage non-proficient children. This is a growing concern worldwide due to the increasing number of multilingual classrooms. The tablet-based FLUX (Fluid Intelligence Luxembourg) test was developed within a highly multicultural and multilingual educational context to offer not only nonverbal test content but also language-fair animated video instructions. A total of 703 third graders (Mage = 8.85, SD = 0.66; 48.8% females, 51.1% males, 0.1% with no gender specified) were included in the standardisation sample and were assessed with tasks measuring figural fluid intelligence, quantitative fluid intelligence, visual processing and short-term memory. The test proved sufficiently reliable (FLUX Full-scale: McDonald’s Omega = 0.94; split-half = 0.95). Test fairness was ensured by analysing each item for Differential Item Functioning (DIF) on children’s background characteristics (language spoken at home, socioeconomic status, gender). Its factorial structure was confirmed using Confirmatory Factor Analysis (CFA). Further validity evidence was provided by determining its concurrent and criterion-related validity (correlations with a test of cognitive ability and educational achievement scores). Research implications and future prospects in promoting equal opportunities in a heterogeneous multilingual educational context are discussed.

This paper presents a real-time, output-only structural health monitoring framework that integrates damage-sensitive cepstral features with a streaming Long Short-Term Memory (LSTM) network for automated damage detection. Acceleration time histories are segmented into overlapping windows, converted into cepstral coefficients, and processed sequentially by a stacked LSTM architecture with state carry-over. This design preserves temporal dependencies while enabling low-latency inference suitable for continuous monitoring. The framework was evaluated under a strict zero-shot setting on the full-scale Z24 Bridge benchmark, in which no training or calibration data from the bridge were used. Our results show that the proposed approach can reliably discriminate staged damage states and track their progression using only vibration measurements. By combining a well-established spectral feature representation with a streaming sequence model, the study demonstrates a practical pathway toward deployable, data-driven monitoring systems capable of operating without retraining on each individual asset.

Antioxidant peptides (AOPs) have emerged as promising peptide agents due to their efficacy in counteracting oxidative stress-related diseases and their applicability in functional food and cosmetic industries. In this study, we developed a comprehensive quantitative structure-activity relationship (QSAR) utilizing a multimodal deep learning framework that integrates 6 sequence-based structure representations with stacking ensemble neural architectures–convolutional neural networks, bidirectional long short-term memory, and Transformer—to enhance predictive accuracy. Additionally, we employed a generative model to design novel AOP candidates, which were subsequently evaluated using the best-performing QSAR model. Remarkably, the stacking models using one-hot encoding achieved outstanding predictive metrics, with accuracy, AUROC, and AUPRC values surpassing 0.90, and the MCC above 0.80, demonstrating a highly accurate and robust QSAR model. SHAP analysis highlighted that proline, leucine, alanine, tyrosine, and glycine are the top five residues that positively influence antioxidant activity, whereas methionine, cysteine, tryptophan, asparagine, and threonine negatively impact antioxidant activity. Finally, 604 high-confidence AOPs were computationally identified. This study demonstrates that the multimodal framework improves the prediction accuracy, robustness, and interpretability of the AOP. It also enables the efficient discovery of high-potential AOPs, thereby offering a powerful pipeline for accelerating peptide discovery in pharmaceutical and functional applications.Supplementary InformationThe online version contains supplementary material available at 10.1186/s13321-025-01102-4.

Abstract. Long short-term memory (LSTM) networks have shown strong performance in rainfall–runoff modeling, often surpassing conventional hydrological models in benchmark studies. However, recent studies raise questions about their ability to extrapolate, particularly under extreme conditions that exceed the range of their training data. This study examines the performance of a stand-alone LSTM trained on 196 catchments in Switzerland when subjected to synthetic design precipitation events of increasing intensity and varying duration. The model's response is compared to that of a hybrid model – a model that combines conceptual hydrological approaches with the LSTM – and evaluated against hydrological process understanding. Our study reiterates that the stand-alone LSTM is not capable of predicting discharge values above a theoretical limit (which we have calculated for this study to be 73 mm d−1), and we show that this limit is below the maximum value of 183 mm d−1 in the training data. Furthermore, the LSTM exhibits a concave runoff response under extreme precipitation, indicating that event runoff coefficients decrease with increasing design precipitation – a phenomenon not observed in the hybrid model used as a benchmark. We show that saturation of the LSTM cell states alone does not fully account for this characteristic behavior, as the LSTM does not reach full saturation, particularly for the 1 d events. Instead, its gating structures prevent new information about the current extreme precipitation from being incorporated into the cell states. Adjusting the LSTM architecture, for instance, by increasing the number of hidden states and/or using a larger, more diverse training dataset, can help mitigate the problem. However, these adjustments do not guarantee improved extrapolation performance, and the LSTM continues to predict values below the range of the training data or show unfeasible runoff responses during the 1 d design experiments. Despite these shortcomings, our findings highlight the inherent potential of stand-alone LSTMs to capture complex hydrometeorological relationships. We argue that more robust training strategies and model configurations could address the observed limitations, preserving the promise of stand-alone LSTMs for rainfall–runoff modeling.

To solve the problems of the lack of economic efficiency and the short driving range of electric commercial vehicles, a hybrid system was developed in this work that uses fuel cells as a range extender. In addition, a method to solve the problem of multi-power energy management was proposed using the model predictive control as a framework. In the state of charge maintenance interval, a quadratic utility function was used to calculate the output power of the fuel cell and battery. The unknown parameters in the quadratic utility function were solved using the model prediction control. Speed prediction was performed using long short-term memory and particle swarm optimization. The demanded power sequence within the prediction horizon was calculated based on the predicted speed. The dynamic programming algorithm was used to solve the power demand sequence within the prediction horizon length, and the unknown parameters in the utility function were deduced inversely. The simulation results show that the proposed energy management strategy (EMS) is superior to conventional EMS in improving component durability and vehicle economy.

Given the critical importance of accurate energy demand and production forecasting in managing power grids and integrating renewable energy sources, this study explores the application of advanced machine learning techniques to forecast electricity load and wind generation data in Austria, Germany, and the Netherlands at different sampling frequencies: 15 min and 60 min. Specifically, we assess the performance of the convolutional neural networks (CNNs), temporal CNN (TCNN), Long Short-Term Memory (LSTM), bidirectional LSTM (BiLSTM), Gated Recurrent Unit (GRU), bidirectional GRU (BiGRU), and the deep neural network (DNN). In addition, the standard machine learning models, namely the k-nearest neighbors (kNN) algorithm and decision trees (DTs), are adopted as baseline predictive models. Bayesian optimization is applied for hyperparameter tuning across multiple models. In total, 54 experimental tasks were performed. For the electricity load at 15 min intervals, the DT shows exceptional performance, while for the electricity load at 60 min intervals, DNN performs the best, in general. For wind generation at 15 min intervals, DT is the best performer, while for wind generation at 60 min intervals, both DT and TCNN provide good results, in general. The insights derived from this study not only advance the field of energy forecasting but also offer practical implications for energy policymakers and stakeholders in optimizing grid performance and renewable energy integration.