The utilization of renewable energy sources significantly increased in response to the growing global energy demand. Rising concerns about the environment, the photovoltaic (PV) systems emerged as prominent and widely adopted among all other renewable energies. The installation cost of PV systems is more however, the recent advancements in PV technology have made them feasible for a wide range of applications. In addition to module and inverter efficiencies, the overall efficiency of a PV system also depends upon the efficiency of the tracking method. The conventional and intelligent methods are experiencing low tracking efficiency under different irradiation and temperature conditions. PV grid integration can be done in two methods. The single stage grid connected systems can experience problems related to power quality and stability of the DC link voltage at low irradiance levels. To limit this problem, two-stage systems are preferable for maintaining a stable DC-link voltage. This research evaluates several MPPT strategies, including conventional approaches (perturbation and observation, incremental conductivity) and intelligent techniques (fuzzy logic controllers, ANFIS), and proposes a hybrid AGORNN controller, that is based on the Adaptive Grasshopper Optimization Recurrent Neural Network strategy. The performance of each method is assessed through parameters such as PV maximum power (Pmpp), DC link voltage (Vdclink), MPPT tracking efficiency η_MPPT (%), Vdc Ripple Vdcr (%), response time (Tres), overshoot Osh (%), utilization efficiency η_ut (%), THD, cost analysis, and payback period (PB). The results of simulations, which were examined using MATLAB/Simulink (R2023b), and compared with existing hybrid MPPT control methods. It exhibits that the proposed AGORNN controller is more effective than the standard and intelligent MPPT. The novel hybrid MPPT controller was examined under the conditions of STC and different irradiance and temperature, which were referred to as PTC. In STC tracking and conversion efficiencies are 99.86 and 96.55, respectively. The efficiencies under PTC are 96% and 91.50%, which ensures a promising solution to high efficiency grid-connected PV systems. The LCOE and Payback Period are also estimated and compared with other methods.

Corporate financial distress typically emerges through a gradual accumulation process, rendering crisis prediction inherently dynamic and path-dependent. However, many existing studies continue to rely on static cross-sectional data or short-term observations, which limits their ability to capture the temporal evolution of financial risk. To address this issue, this study develops a time-series financial crisis early warning framework based on Recurrent Neural Networks (RNNs) and systematically evaluates the incremental value of temporal information in corporate distress prediction. Using annual data of Chinese A-share listed companies from 2019 to 2023, we construct both single-year cross-sectional datasets and a five-year multi-period time-series dataset under a unified experimental protocol. Within this dual-framework setting, RNNs are compared with Random Forest (RF), Support Vector Machine (SVM), and Backpropagation Neural Network (BPNN) using identical feature sets, training–testing splits, and evaluation criteria. Model performance is assessed through multiple metrics, including Accuracy, Precision, Recall, F1 score, and AUC, complemented by statistical validation using McNemar tests, loss-based comparisons, and bootstrap confidence intervals. The empirical results show that while RF and BPNN exhibit strong robustness in static, single-period prediction tasks, RNNs achieve consistently superior performance when multi-period temporal information is explicitly modeled. Statistical tests indicate that the observed performance advantages of RNNs are systematic and stable, though moderate under the current sample size. This study provides empirical evidence that incorporating temporal structures into financial crisis prediction can substantially enhance predictive effectiveness under constrained labeled data. The findings highlight the importance of time-series modeling for early warning applications and offer practical guidance for selecting appropriate predictive frameworks across different data structures.

Abstract Solar flares are among the most energetic phenomena in the solar system, and their forecasting remains a major challenge in space weather research. In this study, we present a deep learning framework that predicts the occurrence of ≥C5.0-class flares within a 2 hr forecast horizon by integrating multimodal, multiwavelength observations of solar active regions. The model combines sequences of Solar Dynamics Observatory (SDO)/Helioseismic and Magnetic Imager (HMI) line-of-sight magnetograms, continuum intensity images, and SDO/Atmospheric Imaging Assembly extreme-ultraviolet observations (193 and 304 Å) with physically derived Space-weather HMI Active Region Patch parameters—total unsigned magnetic flux and current helicity—within a unified end-to-end architecture. A convolutional encoder extracts spatial representations from images, which are fused with physical parameters and processed through a recurrent neural network to capture temporal evolution. Using time sequences spanning 36 minutes, our best-performing configuration achieves an accuracy of 92.7%, a recall of 97.6%, and a true skill statistic of 0.86 on an independent test set, significantly outperforming single-modality baselines. Our findings demonstrate that combining physically meaningful parameters with multiwavelength imaging substantially enhances model performance while maintaining good calibration. The proposed framework demonstrates the capability of deep learning to model the spatiotemporal and physical complexity of solar active regions, thereby enabling more accurate and physics-informed flare forecasting.

Predicting cryptocurrency price is challenging owing to high volatility, less historical data, and the impact of external parameters like news, public sentiment, and regulatory announcements. This challenge is tackled in this research by employing models of deep learning like Recurrent Neural Network (RNN), Bidirectional Long Short-Term Memory (BiLSTM) and Gated Recurrent Unit (GRU)—to predict Bitcoin's OHLC prices daily. Based on historical time-series data of Coin Codex, the research uses an autoencoder-based feature extraction method with five-day sliding window method for sequence generation. Hyperband optimization is used to tune hyperparameter of each model. The result shows that BiLSTM performs better than all the other models with minimum Mean Squared Error (MSE = 0.001183), Mean Absolute Error (MAE = 0.026090), and maximum R² score (0.980596) after optimization. The results emphasize the significance of deep learning in capturing nonlinear dynamics in time series of financial applications and bear testimony to the effectiveness of hyperparameter tuning in enhancing model accuracy. The study enhances the development of prediction tools for digital asset markets and enables more informed investment decisions.

Deep generative models have emerged as powerful computational engines for de novo molecular design, enabling efficient exploration of a vast chemical space that remains inaccessible to traditional experimental approaches. This review provides a comprehensive survey of machine learning-driven molecular generation, systematically organizing the field across three foundational pillars: molecular representations, model architectures, and evaluation frameworks. We present a detailed taxonomy of state-of-the-art generative models, including Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), Recurrent Neural Networks (RNNs), Transformers, Diffusion Models, Normalizing Flows, and Hybrid Architectures, analyzing their underlying mechanisms, comparative strengths, and inherent limitations. Critically, we depart from purely descriptive surveys by systematically examining algorithmic failure modes and practical deployment challenges across model families. We discuss core applications spanning distribution learning and goal-directed generation. Special attention is given to challenging therapeutic domains such as Central Nervous System (CNS) drug discovery, where stringent constraints like blood-brain barrier (BBB) permeability and neurotoxicity mitigation demand multiparameter optimization. We critically evaluate the gap between computational benchmarks and practical medicinal chemistry, addressing synthetic feasibility and experimental validation. Subsequently, we highlight persistent theoretical, computational, and empirical challenges that currently limit widespread deployment, and outline promising future opportunities, including physics-informed architectures, large language models, and autonomous laboratories. This review aims to provide actionable insights for both machine learning researchers and medicinal chemists engaged in next-generation drug discovery.

The generation of realistic network traffic is a critical requirement for testing, simulation, and security evaluation in ZigBee-based IoT systems. In this study, we propose a novel framework that extracts sample ZigBee traffic packets and generates semantically meaningful and protocol-compliant synthetic traffic using Large Language Models (LLMs) such as GPT-4.1 and GPT-5. Unlike traditional rule-based or statistical generators, our approach is data-driven and incorporates sample-based few-shot learning, prompt engineering, and a human-in-the-loop feedback mechanism. To evaluate the effectiveness of the proposed framework, we conduct two sets of experiments. The first focuses on generating unidirectional traffic to emulate typical device-to-hub communication, while the second extends this setup to bidirectional exchanges, capturing realistic request-response dynamics and interaction patterns. The realism of the generated traffic is assessed using a multi-dimensional evaluation framework that includes statistical similarity measures, such as Jensen-Shannon Divergence, as well as protocol compliance, semantic correctness, temporal consistency, and diversity metrics. In addition, we compare the performance of LLM-based generators with classical deep learning baselines, including recurrent neural networks (RNNs) and generative adversarial networks (GANs). We further analyze the computational cost and the impact of internal reasoning effort on traffic generation by systematically evaluating different GPT-5 reasoning configurations. Experimental results show that both GPT-4.1 and GPT-5 successfully learn the temporal and structural dependencies of ZigBee traffic and significantly outperform RNN and GAN baselines in terms of semantic correctness and long-duration generation. Across all experiments, GPT-4.1 consistently generates traffic that more closely resembles real ZigBee traffic while requiring substantially lower computational resources. These findings highlight that low-latency, non-reasoning LLMs can be particularly well suited for highly structured, protocol-constrained network traffic generation tasks, and demonstrate the potential of LLM-based approaches for realistic IoT traffic generation in research and security evaluation.

The prediction of epileptic seizures can significantly improve patients' quality of life by enabling timely preventive interventions. However, realizing automated real-time prediction on edge hardware remains challenging due to high computational complexity, inefficient temporal signal processing, and the von Neumann bottleneck. In this work, we propose a memristor-based multi-stage reservoir computing architecture that jointly addresses algorithmic and hardware limitations. Volatile memristors are employed in reservoir modules to perform nonlinear temporal feature extraction, avoiding error accumulation issues commonly observed in recurrent neural networks. Non-volatile memristor crossbar arrays are further integrated to implement in-memory analog multiply-accumulate operations, significantly reducing data movement and improving hardware efficiency. Owing to the proposed multi-stage structure, high prediction accuracy is achieved with only 1,700 trainable parameters. Moreover, comprehensive hardware-aware evaluations are conducted, including input noise injection, device-to-device and cycle-to cycle variations to assess robustness against memristor non-idealities. Results demonstrate that the proposed system achieves over 97% accuracy in simulation and exceeds 95% accuracy in hardware experiments, while maintaining stable performance under substantial noise, making it a promising low-power solution for real-time seizure prediction on edge platforms.

Wind energy plays a critical role in the European Union’s decarbonization strategy, including Romania’s growing renewable energy capacity. This study proposes a deep learning-based method for forecasting hourly wind energy production in Romania using feedforward neural networks (FFNNs) and recurrent neural networks (RNNs), trained on a dataset spanning from 1 January to 31 December 2023. The dataset includes hourly wind energy output data (mean = 850.6 MW, std = 694.0 MW) and 13 meteorological variables (e.g., average wind speed = 4.7 km/h, temperature = 14.4 °C). A total of 1296 models were trained and evaluated, with the best-performing RNN model achieving a coefficient of determination of R2 = 0.9680 and a mean absolute error (MAE) of 81.03 MW. The top three models all exceeded R2 = 0.966, demonstrating strong generalization on unseen data. The models were also validated using two external time intervals outside the training/testing sets, confirming robustness. These results show that deep learning models can provide highly accurate, data-driven predictions of wind energy output, supporting grid stability and informed decision-making amid renewable energy variability.

The integration of renewable energy sources into modern power systems requires advanced control strategies to maintain stability, reliability, and efficiency. This paper presents a comprehensive review of the application of artificial intelligence techniques, including machine learning, deep learning, and reinforcement learning, in improving droop control for renewable energy integration. These artificial intelligence-based methods address key challenges such as frequency and voltage regulation, power sharing, and grid compliance under conditions of high renewable penetration. Machine learning approaches, such as support vector machines, are used to optimize droop parameters for dynamic grid conditions, while deep learning models, including recurrent neural networks, capture complex system dynamics to enhance the stability of distributed energy systems. Reinforcement learning algorithms enable adaptive, autonomous control, improving multi-objective optimization within microgrids. In addition, emerging directions such as transfer learning and real-time data analytics are explored for their potential to enhance scalability and resilience. Overall, this review synthesizes recent advances to demonstrate the growing impact of artificial intelligence in droop control and outlines future pathways toward more intelligent and sustainable power systems.

Deciphering the structure of variable sensory input is key to building an accurate model of one's environment. Humans can accumulate evidence from sequences of stimuli to estimate their sensory statistics, predict the timing of upcoming stimuli, but also discover rules governing sequence generation. However, whether these three forms of inference operate independently or synergistically remains untested. Here, we report selective interactions between sensory integration, temporal prediction, and rule discovery in humans. Participants were exposed to rhythmic sequences of 10 stimuli governed or not by a latent rule-a predictable change in stimulus statistics after five stimuli-and then asked to predict the 10th stimulus from incomplete sequences. Individual differences in sensory integration timescale for rule-free sequences predicted efficient rule discovery. Conversely, discovering the latent rule shaped the timescale and format of sensory integration for rule-based sequences. Tampering with the rhythmicity of stimulus presentation impaired rule discovery without affecting sensory integration accuracy. Selective perturbations of recurrent neural networks trained in the same conditions confirmed these specific interactions. Together, these findings provide insights into the flexibility of human inferences based on variable yet predictable sensory input.

This paper systematically reviews the latest advances in deep learning-based path planning for autonomous mobile robots, addressing the limitations of traditional methods (e.g., A*, RRT) in dynamic, high-dimensional, and unstructured environments. We comprehensively analyze five major deep learning model categories: Convolutional Neural Networks (CNNs) for spatial feature extraction, Graph Neural Networks (GNNs) for multi-agent collaboration, Recurrent Neural Networks (RNNs) for temporal modeling, Transformers for long-range dependency and complex instruction understanding, and generative models (e.g., GANs, Diffusion Models) for creative path generation. Our analysis covers technical principles, advantages, limitations, application scenarios, and development trends of these methods. The review reveals that deep learning has fundamentally transformed path planning from perception enhancement to decision substitution, from isolated agents to multi-agent collaboration, and from search-based to generative paradigms. Key findings indicate significant performance improvements: GNN-based distributed planning triples multi-robot collaboration efficiency, and generative models increase complex instruction planning success rates to 78.1%. Future directions include cross-modal integration, lightweight deployment, simulation-to-reality transfer, and verifiable safety assurance, which will be crucial for advancing next-generation intelligent mobile robot navigation systems.

Abstract A study of neural network architectures for the reconstruction of the energy deposited in the cells of the ATLAS liquid-argon calorimeters under high pile-up conditions expected at the HL-LHC is presented. These networks are designed to run on the FPGA-based readout hardware of the calorimeters under strict size and latency constraints. Several architectures, including Dense, recurrent (RNN), and convolutional (CNN) neural networks, are optimised using a Bayesian procedure that balances energy resolution against network size. The optimised Dense, CNN, and combined Dense+RNN architectures achieve a transverse energy resolution of approximately 80 MeV, outperforming both the optimal filtering (OF) method currently in use and RNNs of similar complexity. A detailed comparison across the full dynamic range shows that Dense, CNN, and Dense+RNN accurately reproduce the energy scale, while OF and RNNs underestimate the energy. Deep evidential regression is implemented within the Dense architecture to address the need for reliable per-event energy uncertainties. This approach provides predictive uncertainty estimates with minimal increase in network size. The predicted uncertainty is found to be consistent, on average, with the difference between the true deposited energy and the predicted energy.

What are the mechanisms that enable organisms to detect and respond to the actions of others? Social contingency, or the degree to which one's actions reliably elicit timely and relevant responses from another, underlies adaptive behaviour and social interaction across species. In order to investigate general principles underlying this phenomenon, we trained and analysed populations of embodied recurrent neural networks engaged in the perceptual crossing task, a minimal social interaction experiment in humans. Through extensive robustness and performance testing, we isolated a subset of 111 circuits. Analysis revealed several shared principles among the robust subset. First, despite uniform performance, we found four distinct behavioural strategies that agents would switch to depending on state history and the strategy of their partner. Next, we found that social contingency does not depend on a single feature of feedback but rather on a scaled relationship between feedback parameters. Finally, using dynamical systems analysis, we identified a shared mechanism for social contingency across all successful circuits. Specifically, it was necessary for the nervous system to couple a contingency cue, a specific temporal pattern in the sensor's activation that distinguishes social from non-social interactions, with a method of conditional stability, a way of structuring the nervous system such that interactions are stable only if the appropriate temporal cue is present. This article is part of the theme issue 'Mechanisms of learning from social interaction'.

Electromyography (EMG) signals play a pivotal role in biomedical applications, such as prosthetic control and human-computer interaction, where advanced classification methods for accurate muscle activity translation are essential. This study evaluates the performance of three neural network architectures: Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Bidirectional Recurrent Neural Network (BRNN) for EMG signal classification. The EMG signals were preprocessed using Digital Wavelet Transform (DWT) with Daubechies 2 wavelet to extract time-frequency features. Experiments were conducted on a large-scale training dataset comprising 672 subject recordings across six hand gestures, enabling a robust, data-driven comparison. The best classification accuracy of LSTM, GRU, and Bidirectional RNN was achieved, corresponding to cD7, with values of 93.04±1.52, 92.72±1.26, and 91.59±0.97, respectively. However, these models exhibited varying degrees of sensitivity to additive noise, particularly at deeper DWT levels. The findings highlight the trade-offs between accuracy and noise tolerance, providing insights for optimizing EMG-based gesture recognition systems in real-world applications. The final analysis confirms that LSTM outperforms the other models for real-time EMG classification, while GRU and BRNN offer a favourable balance between accuracy and computational efficiency. Looking ahead, the effective handling of large-scale, high-dimensional EMG data yields significant improvements in performance, particularly for prosthetic and real-time control applications.

Intelligent extraction of coal seam gas constitutes a crucial development direction for managing underground gas disasters. Building on an established mathematical model, this study develops an intelligent control model for gas extraction. In this model, controlled variables include gas extraction concentration, gas extraction flow rate, negative pressure, and extraction pump efficiency ratio, while control variables are defined as the valve opening of extraction boreholes and the power of extraction pumps. The ideal curve of the controlled quantity with time is obtained by using the recurrent neural network (SimpleRNN), and the controlled quantity is intelligently controlled by the model predictive control (MPC) algorithm so that the actual value of controlled quantity approaches the reference value at the corresponding time of its ideal curve. Taking the simulated gas extraction data as an example, an algorithm simulation experiment is performed. The experimental results show that the ideal reference curve of the controlled quantity obtained by the cyclic neural network has a good data fitting degree. The dynamic control of the controlled quantity by the model predictive control algorithm can overcome the interference of environmental and nonlinear factors and achieve a better control effect, which provides a certain reference for the intelligent control of gas drainage.

This work addresses a key challenge in the effective adoption of Recurrent Neural Networks (RNNs) by reducing inference time and expanding the scope of a prediction. It introduces compressed learning, a novel approach that integrates Context-Free Grammar (CFG) and online tokenization into the training and inference of RNNs for streaming inputs. Through a hierarchical compression algorithm, it compresses an input sequence to a CFG and makes predictions based on the compressed sequence. Its algorithm design employs a set of techniques to overcome the issues from the myopic nature of online tokenization, the tension between inference accuracy and compression rate, and other complexities in sequence compression and prediction. Its effectiveness is theoretically analyzed and empirically validated on 16 real-world sequences, including program function calls, memory traces, and system logs. Empirical results demonstrate that compressed learning can successfully recognize and leverage repetitive patterns in input sequences, and effectively translate them into dramatic (1–1,762 \(\times\) ) inference speedups as well as much (1–7,830 \(\times\) ) expanded prediction scope, while keeping the inference accuracy satisfactory.

Brain tumors often require treatment and multiple biopsies. They are the third most common cancer among young adults in both incidence and mortality. The expression of the O6-methylguanine-DNA methyltransferase (MGMT) gene plays an important role in predicting tumor behavior. It affects how patients respond to chemotherapy and may reduce the need for invasive procedures. Machine learning can help make accurate medical predictions, but it requires large and diverse patient datasets. These datasets are difficult to access due to privacy and legal restrictions. This article proposes a Federated Learning (FL) framework to address these challenges. FL allows different institutions to train a shared model without exchanging raw data. A hybrid deep learning model combining recurrent neural networks (RNNs) and convolutional neural networks (CNNs) is developed to analyze magnetic resonance imaging (MRI) scans from the BraTS 2021 dataset. The model aims to detect glioblastoma and predict MGMT gene expression. Two swarm intelligence algorithms, the Bayesian Search Optimization Algorithm and the Sparrow Search Optimization Algorithm, are used to optimize the model’s hyperparameters. The FL system was tested across ten universities. It performed similarly to models trained on centralized data. The proposed model, BrainGeneDeepNet, achieved high performance: 0.9758 accuracy, 0.0769 loss, 0.9980 AUC, 0.9770 recall, and 0.9782 precision. These results show that federated learning is a secure and effective approach for medical imaging and biomarker prediction.

Spatiotemporal flows of neural activity, such as traveling waves, have been observed throughout the brain since the earliest recordings; yet there is still little consensus on their functional role. Recent experiments and models have linked traveling waves to visual and physical motion, but these observations have been difficult to reconcile with standard accounts of topographically organized selectivity and feedforward receptive fields. Here, we introduce a theoretical framework that formalizes and generalizes the connection between 'motion' and flowing neural dynamics in the language of equivariant neural network theory. We consider 'motion' not only in physical or visual spaces, but also in more abstract representational spaces, and we argue that recurrent traveling-wave-like dynamics are not just useful but necessary for accurate and stable processing of any signal undergoing such motion. Formally, we show that for any non-trivial recurrent neural network to process a sequence undergoing a flow transformation (such as visual motion) in a structured equivariant manner, its hidden state dynamics must actively realize a homomorphic representation of the same flow through recurrent connectivity. In this ''spatiotemporal perspective on dynamical computation'', traveling waves and related flows are best understood as faithful dynamic representations of stimulus flows; and consequently the natural inclination of biological systems towards such dynamics may be viewed as an innate inductive bias towards efficiency and generalization in the spatiotemporally-structured dynamical world they inhabit.

Spatiotemporal flows of neural activity, such as traveling waves, have been observed throughout the brain since the earliest recordings; yet there is still little consensus on their functional role. Recent experiments and models have linked traveling waves to visual and physical motion, but these observations have been difficult to reconcile with standard accounts of topographically organized selectivity and feedforward receptive fields. Here, we introduce a theoretical framework that formalizes and generalizes the connection between 'motion' and flowing neural dynamics in the language of equivariant neural network theory. We consider 'motion' not only in physical or visual spaces, but also in more abstract representational spaces, and we argue that recurrent traveling-wave-like dynamics are not just useful but necessary for accurate and stable processing of any signal undergoing such motion. Formally, we show that for any non-trivial recurrent neural network to process a sequence undergoing a flow transformation (such as visual motion) in a structured equivariant manner, its hidden state dynamics must actively realize a homomorphic representation of the same flow through recurrent connectivity. In this ''spatiotemporal perspective on dynamical computation'', traveling waves and related flows are best understood as faithful dynamic representations of stimulus flows; and consequently the natural inclination of biological systems towards such dynamics may be viewed as an innate inductive bias towards efficiency and generalization in the spatiotemporally-structured dynamical world they inhabit.

Modeling typhoid dynamics using recurrent neural networks with Bayesian regularization.