ABSTRACT In recent years, there has been a growing interest in the study of marine mammal vocalisations, particularly to understand their behavioural implications and to explore novel linguistic potentials. Among marine mammals, sperm whales (Physeter macrocephalus) have been identified to possess a highly intricate and diverse vocal repertoire, using sequences of clicks known as codas. In this research, we explore temporal dependencies between the codas within longer bouts as a first step towards understanding their longer vocal exchanges. Our results from analysing simultaneously emitted coda sequences from sperm whale pairs demonstrate that the structure of the next coda produced by the second whale is predictable and influenced by a sequence of up to 10 preceding codas from the first whale. The predictive features include the coda duration, the first and second inter-click-interval, and the duration of the delay between the codas. We explore the level of dependencies between the coda sequences by mutual information analysis. The prediction level is compared for a hidden Markov model, a multi-layer Long-Short-Term Memory (LSTM) and transformers models. Our findings provide insight into the complexity of sperm whale vocalisation interactions.

This study investigates the predictability of Elementary Cellular Automata (ECA)—simple yet dynamically rich deterministic systems—using two state-of-the-art machine learning models: Long Short-Term Memory (LSTM) networks and Transformer architectures. Although ECAs are governed by deterministic rules, their behavior often exhibits pseudo-randomness and chaotic dynamics, especially under Class III and IV rules. Our experiments assess the ability of LSTM and Transformer models to predict future configurations of ECAs based on past states. Results reveal that LSTMs excel in modeling rules with short- to mid-range dependencies, achieving accuracies above 99% in structured scenarios, while struggling with chaotic rules such as Rule 30. Conversely, Transformers demonstrate superior performance in capturing long-range dependencies, achieving perfect accuracy for rules like Rule 90 and Rule 62, but incurring higher computational costs. These findings underscore the limits of predictability in deterministic yet complex systems and highlight how different neural architectures are suited to distinct forms of structural complexity. The work contributes to a deeper understanding of machine learning’s capacity to model discrete chaotic systems and opens avenues for hybrid approaches in predictive modeling of computationally irreducible dynamics. • Cellular automata enable analysis of chaos in finite discrete systems. • LSTM excels on short-term and locally structured cellular automata rules. • Transformer captures long-range and chaotic dependencies more effectively. • Comparative results reveal trade-offs in accuracy, stability, and cost. • Hybrid models may combine LSTM and Transformer strengths for better forecasts.

Ship motion prediction is essential in marine engineering, but missing data caused by sensor faults or signal interruptions often degrades the accuracy of long short-term memory (LSTM) models. This study investigates how different missing data rates and imputation methods affect LSTM prediction performance. A ship-motion dataset under various speeds and wave conditions was used to examine model feasibility and hyperparameter sensitivity. Traditional filling strategies, including zero and mean filling, were compared under missing data scenarios. Results show that data loss significantly reduces prediction accuracy. The mean-filling method generally performs better than zero-filling, though its effectiveness decreases with higher data diversity. Proper data clustering can effectively enhance its performance.

An algorithm-enhanced stool DNA system improves the differential diagnosis of colorectal cancer versus Crohn's disease in high-risk symptomatic patients.

The rapid growth in global population necessitates efficient energy management solutions for sustainable living. Smart Building Energy Management Systems (SBEMS) play a crucial role in achieving this goal by leveraging automation and advanced analytics. This study proposes a novel Deep Learning and IoT-based SBEMS approach to predict energy consumption, classify buildings into energy-demand clusters, and optimize the monitoring and operation of electrical equipment. While traditional statistical methods have been widely used for load forecasting, recent advancements in deep learning provide robust alternatives to address the inherent complexity of nonlinear energy consumption patterns. This research employs regression analysis and state-of-the-art neural network architectures, including Single-Step and Multi-Step Dense Models, Convolutional Neural Networks (CNNs), and Long Short-Term Memory (LSTM) networks, to enhance prediction accuracy. Additionally, K-means clustering is introduced to segment buildings into distinct energy-demand categories, ensuring optimal energy utilization. Unlike prior studies that often lack a comprehensive approach, this work integrates all critical features under a unified framework. By applying these advanced methodologies to a unique dataset, the proposed system demonstrates improved accuracy in energy load forecasting and clustering, providing a significant contribution to the field of smart building energy management. The experimental results demonstrate that CNN and LSTM models significantly outperform conventional statistical approaches in capturing nonlinear energy consumption patterns. These outcomes support proactive energy scheduling, peak-demand mitigation, and scalable smart building energy management, offering practical value for facility managers, utility operators, and policymakers.

Projecting Actual Evapotranspiration (AET) is critical for the survival of semi-arid Pinus sylvestris var. mongolica forests but remains difficult due to climate uncertainties. We bridged this gap by developing a hybrid framework in Inner Mongolia. We trained a Long Short-Term Memory (LSTM) network using SEBAL-derived AET as a physics-based proxy target. This approach achieved high accuracy in simulating historical dynamics (R2 = 0.926, RMSE = 13.56 mm/month). Crucially, our model relies on high-resolution data from 2021 to learn intra-annual seasonality; consequently, our projections represent responses to mean climatological shifts rather than interannual variability. To assess future risks, we drove this validated model with an ensemble of five bias-corrected CMIP6 General Circulation Models (GCMs) for the 2030–2050 period under SSP2-4.5 and SSP5-8.5 scenarios. The ensemble projections reveal a robust increase in total annual AET, driven by a predicted ‘warmer and wetter’ climate (+2.4°C temperature, +12% precipitation). However, this increase is uneven, showing a significant intensification specifically during the growing season (May–August). This seasonal spike indicates a heightened risk of rapid soil moisture depletion due to soaring atmospheric demand, paradoxically creating water stress despite higher annual rainfall. These findings challenge conventional views and highlight the urgent need for adaptive management focused on seasonal vulnerability.

Significant research has been directed towards cyberattack detection through reactive assistive techniques, utilizing pattern-matching algorithms for scanning system logs and network traffic to identify known signatures. While effective machine learning (ML) models have been developed to automate detection processes-including identifying, tracking, and blocking malware and intruders-comparatively less attention has been given to cyberattack prediction, particularly for time scales extending beyond daily observations. Long-term attack prediction approaches are highly valued, offering defenders extended periods for the development and dissemination of defensive strategies and tools. Long short-term memory (LSTM) networks are frequently chosen among ML models for their strong performance in time series data analysis. This paper presents a novel methodology combining LSTM models with flexible sliding window techniques to improve prediction outcomes and enhance training efficiency. Experiments also incorporate convolutional filters and combined bivariate performance measures, similar to methodologies used in enhancing stock forecasting systems. These substantial contributions are anticipated to offer valuable insights for researchers in this field.

Generative modeling of clinical time series via latent stochastic differential equations.

The diagnosis of Alzheimer's disease (AD) has progressively depended on sophisticated neuroimaging methods alongside cognitive assessments. This study combines volumetric feature analysis with computational modeling techniques, focusing on spatial and temporal analysis, to categorize individuals as cognitively normal (CN), mild cognitive impairment (MCI), or AD using magnetic resonance imaging (MRI) data. In the initial phase, volumetric changes, comprising cortical thickness, white matter, grey matter, cerebrospinal fluid, and total intracranial volume, were derived from the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset utilizing the CAT12 toolbox in statistical parametric mapping (SPM). Linear regression was utilized on these variables over time to create slopes that reflect volumetric change rates, which then served as inputs for machine learning classifiers. The slopes of cortical thickness exhibited the greatest classification accuracy, reaching 82.5% with a random forest model for differentiating AD from CN individuals. During the second phase, a deep learning methodology was utilized, relying solely on the MRI scans and excluding the outcomes from the first phase. A pre-trained 3D ResNet-101 convolutional neural network (CNN) model extracted spatial characteristics from MRI volumes, whereas long short-term memory (LSTM) networks recorded temporal dynamics across subsequent annual scans. This hybrid CNN-LSTM design markedly improved classification performance, attaining 96.7% accuracy for AD against CN and enhancing the distinction of MCI cases. Nonetheless, discrepancies in MCI categorization were chiefly ascribed to the restricted access to annual MRI data and the model's pre-training on CN and AD cohorts. These findings highlight the potential of integrating volumetric statistical analysis with deep learning for automated AD categorization. This work enhances neuroimaging diagnostic methods by utilizing both spatial and temporal MRI data, enabling early diagnosis and better evaluation of disease development.

Hybrid propulsion systems increase ship energy efficiency by allowing the sharing of power between diesel engines and battery energy storage systems. However, the long-term efficiency of these types of systems depends on accurately estimating the Remaining Useful Life (RUL) of lithium-ion batteries to allow effective charge scheduling, maintenance planning, and reliable navigation. This study uses nine data-driven algorithms, including ensemble methods, recurrent neural networks, and linear models, to examine the RUL of a lithium-ion battery pack installed on a hybrid cargo ship. A 5-fold cross-validation structure was used to preprocess, normalize, and analyze actual operational data gathered during the vessel's service life. To improve the accuracy of predictions, hyperparameter optimization was performed out. Long Short-Term Memory (LSTM), which reduced MAE from 2.87 to 1.46 and RMSE from 12.57 to 6.34 after optimization while retaining a high coefficient of determination (R² = 0.9999), performed the best among the models that were evaluated. The results obtained indicate that condition-based maintenance and energy utilization methods on hybrid ships can be effectively supported by data-driven RUL estimation. In order to enhance generalization and assess integration with real-time propulsion control systems, future research will expand the analysis to multi-vessel datasets.

Air pollution remains a primary global health concern, with one of its most hazardous components, fine particulate matter PM2.5, being recognized as a significant contributor to premature mortality worldwide. Thus, forecasting PM2.5 concentrations is essential for policymaking and mitigating their harmful effects. Low-cost sensor (LCS) technology has advanced air quality monitoring by providing real-time PM2.5 measurements at finer spatial and temporal granularity due to their higher deployment density, enabling more comprehensive data for predictive modelling. However, these sensors can experience significant data gaps, necessitating effective imputation methods to improve the accuracy of PM2.5 predictions. In this study, we focus on reconstructing missing PM2.5 readings by evaluating the impact of different imputation strategies on PM2.5 forecasting accuracy using data from eight LCSs deployed across Edmonton, Alberta, from January 2023 to December 2024. The dataset includes 7 features: temperature, relative humidity, wind speed, wind direction, latitude, longitude, and altitude, integrated from both LCS sensors and adjacent air quality monitoring stations. We evaluate the efficacy of Kriging as a spatial imputation method for PM2.5 forecasting, comparing its performance along with mean imputation and k-nearest neighbours, against a no-imputation benchmark to assess whether imputation techniques provide meaningful improvements in forecasting accuracy. Additionally, we compare its performance across data with varying levels of missingness. Our comparative analysis examines how these imputation techniques affect various prediction models, including Random Forest, Extreme Gradient Boosting, Convolutional Neural Networks, Long Short-Term Memory (LSTM), and LSTM with an Attention mechanism. Our findings indicate that although differences among imputation methods were modest, Kriging consistently yielded the best results, particularly at 90% data availability for the deep learning models. The LSTM-AM architecture produced the best overall results in a short-horizon setting, [Formula: see text] = 0.9602, MAE = 1.7815μg/m[Formula: see text], and RMSE = 4.5208μg/m[Formula: see text], suggesting an optimal combination of Kriging imputation with LSTM-AM modeling for near-term PM2.5 forecasting applications.

The emergence of the Unified Payments Interface (UPI) has transformed the digital payment ecosystem by enabling seamless, instant, and interoperable financial transactions across India. Its widespread acceptance has accelerated the shift toward cashless payments and increased the availability of digital financial services for millions of users. However, the rapid growth in transaction volume and user adoption has also created new opportunities for cybercriminals to exploit vulnerabilities within the digital payment infrastructure.Financial fraud associated with UPI platforms has become increasingly sophisticated, involving techniques such as phishing campaigns, fraudulent QR codes, identity impersonation, account hijacking, social engineering attacks, and the misuse of mule accounts. These evolving threats generate complex transaction patterns that are often difficult to detect using conventional rule-based security mechanisms. As fraud strategies continue to change, static detection systems struggle to provide accurate and timely identification of suspicious activities.Recent progress in Artificial Intelligence, Machine Learning, and Deep Learning technologies has significantly enhanced the capability of fraud detection systems. One of these developments has caught a lot of interest from researchers is hybrid deep learning methods that are able to integrate the benefits of several computational models. Several techniques can be combined in a unified approach for enhancing fraud detection, such as Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, Autoencoders, Graph Neural Networks (GNNs), Explainable Artificial Intelligence (XAI), and Federated Learning. These architectures enable real-time analysis, adaptive learning, privacy protection, and efficient management of large-scale transactional data. This review is based on the recent research and studies regarding the detection of fraudulent activities in UPI-based payment environment using hybrid deep learning techniques. It offers detailed evaluations of the current models, their strengths and weaknesses, explores current issues and challenges, and suggests areas of future work. These results suggest that hybrid deep learning architectures are promising to improve the security of digital payment systems through better detection accuracy, fewer false alarms, safeguarding sensitive user data, and greater transparency of automated decision-making. Thus, these wise frameworks are a valuable basis for developing safe, reliable, and trustworthy digital financial ecosystems

Traditional SONAR systems are widely used for underwater object detection and navigation; however, they suffer from high energy consumption, noise interference, and signal degradation in varying aquatic conditions. Inspired by biological echolocation, we propose a novel machine learning-enhanced sonar model to improve accuracy, robustness, and energy efficiency in underwater sensing applications. The proposed model employs bio-inspired echolocation principles, where transmitted acoustic pulses dynamically adjust in response to environmental conditions. Deep learning techniques, specifically Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks, are employed to classify sonar echoes and enhance object detection. Additionally, digital signal processing (DSP) techniques, including Butterworth filtering, wavelet decomposition, and adaptive thresholding, are integrated to mitigate noise and improve signal clarity. Experimental evaluations demonstrate that the proposed sonar model, SonarNet, achieves 92.7% classification accuracy, outperforming conventional SONAR-based methods, which achieved 85.4% accuracy. The integration of adaptive signal processing leads to a 20.8% reduction in energy consumption compared to standard SONAR models, improves the Signal-to-Noise Ratio (SNR) by 15.3dB, and reduces the false positive rate by 18.6%. Additionally, the sound velocity profile (SVP) correction mechanism improves depth estimation accuracy by 12.4%. All performance results reported in this work were obtained using a simulation environment parameterized by real-world oceanographic datasets.

Land subsidence, a complex phenomenon with extensive environmental and economic consequences, presents a significant challenge in the Silakhor Plain, Lorestan Province. This study investigates the effective factors on subsidence, including land use, drainage density, soil properties, and groundwater level changes, using machine learning and deep learning models. The results indicate that soil properties, excessive groundwater exploitation, and human infrastructure are the primary drivers of subsidence, with the central and southeastern parts of the plain facing the highest risk due to the high density of water wells and proximity to active faults. Among the four advanced models utilized: Random Forest (RF), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Convolutional Neural Network (CNN), the RF model demonstrated the best performance in predicting subsidence intensity, exhibiting the highest coefficient of determination (R2 = 0.9880) and the lowest error (MSE = 0.0516). Furthermore, in the classification of subsidence-prone and non-subsidence areas, RF showed significant superiority with a precision of 0.9929 and a recall of 0.9999. The deep learning models also demonstrated strong performance. The LSTM model, with an R-squared of 0.9792, and the CNN model, with an R-squared of 0.9695, showed excellent capability in predicting subsidence intensity and identifying subsidence-prone areas (with a Recall of 1 for both models). However, their specificity was lower compared to the RF model. In contrast, the RNN model provided poorer performance in both prediction and classification compared to other models. SHAP analysis further confirmed these findings, indicating that soil properties, surface water resources, and human infrastructure-related factors play the most significant roles in subsidence occurrence. The findings of this research can serve as a basis for future planning in water and land resource management, as well as for developing effective strategies to mitigate subsidence risks in the Silakhor Plain and similar regions.

Ovarian cancer patients requiring intensive care unit (ICU) admission face particularly grave prognosis, yet current prognostic models rely on static baseline characteristics and generic severity scores, neglecting the rich temporal dynamics of vital signs that may better capture physiological deterioration patterns. To develop and validate an interpretable Long Short-Term Memory (LSTM) neural network integrating hourly vital signs and static features for predicting 28-day mortality in ICU-admitted ovarian cancer patients. This retrospective multicenter study utilized MIMIC-IV (2008-2022) and SICdb (2013-2021) databases. Adult ovarian cancer patients with first ICU admission, length of stay ≥ 24h, and complete hourly measurements of six vital signs (blood pressures, heart rate, respiratory rate, oxygen saturation) were included. MIMIC-IV was split into training (n = 269) and internal validation (n = 115) sets, while SICdb (n = 154) served as external validation. A dual-pathway LSTM network integrated temporal and static features (age, BMI, SOFA score). Model performance was assessed using AUROC, calibration metrics, and compared against traditional ICU scores and five static-only machine learning baselines. Five-fold cross-validation and 100 repeated temporal ablation trials assessed robustness. SHapley Additive exPlanations (SHAP) analysis quantified feature importance and temporal dynamics. The LSTM model achieved AUROC of 0.845 (95%CI: 0.789-0.901) in training, 0.785 (95%CI: 0.655-0.915) in internal validation, and 0.767 (95%CI: 0.617-0.917) in external validation, significantly outperforming SOFA (AUC: 0.683-0.733), SAPS II (AUC: 0.743), SAPS 3 (AUC: 0.678), and OASIS (AUC: 0.659 - 0.569). Five-fold cross-validation confirmed stability (mean AUROC: 0.785 ± 0.018). Negative predictive values exceeded 94% across all cohorts. The LSTM achieved AUC improvements of 0.10-0.17 over static baselines, with repeated permutation trials confirming significant temporal contributions (all P < 0.001). Temporal ablation identified the 12-24h window as most critical (ΔAUC = 0.062). SHAP analysis revealed static features dominated prediction, while respiratory rate and diastolic blood pressure were leading temporal contributors. This interpretable LSTM model accurately predicts mortality in ICU-admitted ovarian cancer patients by capturing temporal vital sign dynamics, significantly outperforming traditional severity scores and enabling personalized risk stratification for clinical decision-making.

Epileptic seizures represent the most widespread neurological disorder in the world. The efficient automatic detection of seizures is crucial for timely diagnosis and prompt treatment. In this paper, we propose a novel framework, IIRB-HybridFormer, that integrates a hybrid CNN transformer framework with time frequency representation for epileptic seizure detection from Electroencephalogram (EEG) signals. The proposed Inception-based Inverted Residual Bottleneck (IIRB) block utilizes a multi-branch feature extraction mechanism in conjunction with an inverted residual structure, ensuring efficient feature representation with minimal computational overhead. In the subsequent stage, Long Short-Term Memory (LSTM) is integrated with the Hybrid Transformer (HybridFormer) to eliminate the need for positional encoding, thereby reducing the computational complexity. The LSTM-derived feature vector is embedded in HybridFormer, which leverages Hybrid Pooled Multi-head Attention (HPMA) and an Inverted Residual Bottleneck-based Feed-forward Network (IRBFN) to enable robust local and global feature extraction. The algorithm's effectiveness is evaluated on two publicly available EEG databases of varying sizes, namely the Bonn and CHB-MIT datasets. Experimental results demonstrate classification accuracy of 99.538% and 97.801% on the Bonn and CHB-MIT, respectively. The CHB-MIT dataset is also evaluated for event-based seizure detection, achieving a sensitivity of 100% and a false detection rate of 2.474/h. The proposed architecture exhibits a superior performance compared to the existing state-of-the-art models. It also surpasses other pre-trained models with its lightweight framework, along with reduced training and testing time. The results validate that the proposed IIRB-HybridFormer framework is robust and reliable in highlighting its strong potential for effective detection of epileptic seizures.

The ongoing global energy transition has intensified the need for precise modeling of renewable energy systems, especially photovoltaic–thermal (PV/T) systems that have the ability to produce both electrical and thermal energy. Improving the efficiency and reliability of PV/T systems is a key enabler of the transition toward sustainable energy. Accurate forecasting of their thermal performance is therefore essential to maximize renewable energy use and reduce energy losses. A deep learning-based method is proposed in this study for the prediction of the thermal efficiency of an air-based PV/T system. More specifically, temporal deep learning architectures are investigated to exploit the complex nonlinear relationships and temporal dependencies governing the thermal behavior of the PV/T collector. A comprehensive comparative analysis is conducted using four state-of-the-art architectures, namely Temporal Convolutional Network (TCN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Transformer. Furthermore, the influence of sequence length is examined through a sensitivity analysis considering forecasting horizons of 1 h, 6 h, 12 h, and 24 h. The models are evaluated using the Coefficient of Determination (R2), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE). The results demonstrate that forecasting performance is strongly influenced by the selected temporal horizon. Among the investigated configurations, the 24-h horizon provided the most informative temporal context for thermal efficiency prediction. Under this common forecasting horizon, the LSTM model achieved the highest predictive accuracy, reaching an R2 of 0.9952, an RMSE of 0.5975, and an MAE of 0.2364, outperforming the TCN, GRU, and Transformer architectures. The residual error and convergence analyses further highlighted the effectiveness of recurrent neural networks in capturing the thermal dynamics of the investigated PV/T system. By enabling accurate and reliable thermal efficiency forecasting, the proposed framework supports improved energy management, higher energy efficiency, and a stronger integration of renewable energy systems, thus contributing to more sustainable operation of hybrid solar energy technologies.

Stock price prediction remains a complex challenge due to market volatility and non-linear dynamics. This study introduces a novel hybrid framework, AdapRandOpt_ARIMA-LSTM, which integrates an upgraded ARIMA model with a tailored Long Short-Term Memory (LSTM) network. The model employs Adaptive Random Search Optimization (AdapRandOpt) to precisely calibrate hyperparameters and determine optimal weight distributions for the ensemble. The framework was evaluated on sixteen leading NSE-listed companies: ASIANPAINT, BEL, CIPLA, DMART, ETERNAL, FACT, GLAXO, HINDLCO, INDIGO, JSWENERGY, KOTAKBANK, LTFOODS, MANKIND, NESTLEIND, RELIANCE, and TATAMOTORS. Beyond standard Root Mean Square Error (RMSE) metrics, the model's efficacy was validated through profitability ratios and directional accuracy, ensuring its practical utility for traders. Results confirm that the hybrid ensemble significantly outperforms standalone models, demonstrating that AdapRandOpt effectively enhances forecasting robustness and predictive precision. This approach provides a computationally efficient, high-accuracy solution for navigating the intricacies of the Indian financial market. The AdapRandOpt_ARIMA-LSTM framework supports SDG 8 (Decent Work and Economic Growth) by providing a high-precision tool that fosters financial stability and informed decision-making within the Indian equity market. Furthermore, by integrating profitability ratios and directional accuracy, the model promotes SDG 12 (Responsible Consumption and Production) through the encouragement of sustainable investment practices and transparent financial resource management.

This study explores the potential of explainable artificial intelligence to advance our understanding of allergen peptides in the context of dermatology and cosmetics. We present a hybrid deep learning framework that integrates Temporal Convolutional Networks (TCN) and stacked Long Short-Term Memory (LSTM) architectures, enhanced with Evolutionary Scale Modeling (ESM) embeddings, to decode allergenic motifs embedded in peptide sequences. The ESM embeddings allow the model to capture both the evolutionary context and structural nuances of amino acids, enabling accurate classification of allergenic potential. Beyond classification, the framework emphasizes interpretability using state-of-the-art explainability tools such as Anchor, LIME, and SHAP. Anchor identifies minimal and decisive motifs responsible for allergenic activity, while LIME and SHAP provide a distributed importance map of k-mers, highlighting synergistic contributions across the peptide sequence. These methods ensure that the model output is not only precise, but it may also be biologically meaningful, opening a path to rational peptide modification strategies. In dermatology and cosmetic science, this approach represents a transformative tool, providing a scalable and transparent means to screen peptide-based formulations for allergenic risks, facilitates the design of safer bioactive peptides for therapeutic or cosmetic use, and supports regulatory compliance by grounding computational predictions in mechanistic explanations. Ultimately, the study bridges cutting-edge machine learning with dermatology and cosmetics, fostering the development of innovative, safe, and patient-centered cosmetic products.

The growing use of renewable energy sources (RES), particularly wind power (WP), in deregulated electricity markets has resulted in large operating uncertainties because of the erratic and uncertain behavior of wind resources. Differences between actual wind velocity (AWV) and forecasted wind velocity (FWV) cause a large price imbalance (PIMB) that has implications for market execution scheduling, operational reliability, and economic profitability. To solve these problems, this paper presents a comprehensive techno-economic model comprising Long Short-Term Memory (LSTM)-based wind forecasting coupled with an American Zebra Optimization (AZO) algorithm for profit maximization in wind-integrated deregulated power systems. The proposed approach is validated on the IEEE 14-bus test system by collecting real wind datasets from Mina Sultan Qaboos, Muscat, Oman, and Duqm, Oman. This work shows that the LSTM model works in modelling nonlinear temporal dependencies in wind speed data, leading to a significant improvement in forecasting accuracy relative to traditional forecasting techniques and Random Forest (RF)-based approaches. Using Locational Marginal Pricing (LMP), imbalance settlement mechanisms, generator operational limits, and transmission constraints, the forecasted outputs are integrated into an Optimal Power Flow (OPF)-based market framework. The proposed LSTM-based framework has achieved a significant reduction in imbalance pricing by 30-35% in Muscat and 28-29% in Duqm when compared to traditional forecasting techniques. Moreover, the AZO algorithm outperforms Sequential Quadratic Programming (SQP) in convergence performance and overall system profitability by approximately 6%. Profit improvement of the obtained result of wind potential is 5.9% for Muscat and 6.4% for Duqm. The proposed framework successfully reduces uncertainty in the market and enhances renewables integration capacity, as well as offers a robust decision-support framework for the electricity market operation and a decision support tool under the RES-driven high-level market conditions to ensure economic competition for competitive electricity market operations.