Explore the vast range of titles available.

Equatorial plasma bubble (EPB) irregularities can significantly impact satellite‐based communication and navigation systems. Accurate prediction of EPB occurrence is essential for mitigating these impacts. Using the GNSS receiver network and ionosonde data from East and Southeast Asia during 2010–2021, and the rate of TEC change index to characterize the occurrence of EPB irregularities, we developed a novel Spatio‐Temporal deep learning model for regional EPB irregularities short‐term Prediction (STEP). The model integrates the convolutional neural network and long short‐term memory (LSTM) network, together with attention mechanisms, to capture both spatial and temporal features of regional ionospheric irregularities. The results show that for 5‐min forecast, the STEP model achieves a root mean square error (RMSE) of 0.062 TECU/min and an R2 of 0.818, reducing RMSE by 19.48% compared to LSTM and 27.06% compared to gated recurrent unit model. For 60‐min prediction, the STEP model can still achieve reasonable accuracy with an RMSE of 0.110 TECU/min and an R2 of 0.482, showing significant improvement over traditional models. The equatorial F layer height and regional TEC fluctuations were identified as the most critical factors for predicting the generation and duration of EPB irregularities, respectively. The spatial and temporal distributions of EPB irregularities, including their latitudinal variation and delayed onset after sunset, and the occurrence across different days in East and Southeast Asia, were well predicted by the STEP. It is expected that the STEP model would provide a valuable tool for improving the resilience of GNSS against ionospheric scintillations induced by EPB irregularities.

Real-time crime forecasting is important. However, accurate prediction of when and where the next crime will happen is difficult. No known physical model provides a reasonable approximation to such a complex system. Historical crime data are sparse in both space and time and the signal of interests is weak. In this work, the authors first present a proper representation of crime data. The authors then adapt the spatial temporal residual network on the well represented data to predict the distribution of crime in Los Angeles at the scale of hours in neighborhood-sized parcels. These experiments as well as comparisons with several existing approaches to prediction demonstrate the superiority of the proposed model in terms of accuracy. Finally, the authors present a ternarization technique to address the resource consumption issue for its deployment in real world. This work is an extension of our short conference proceeding paper [Wang, B., Zhang, D., Zhang, D. H., et al., Deep learning for real time Crime forecasting, 2017, arXiv: 1707.03340].

Dementia, a progressively debilitating neurodegenerative disorder, poses a growing challenge to global healthcare due to its subtle early onset and lack of accurate temporal diagnostic models. Traditional machine learning (ML) approaches for dementia prediction, including SVMs, Random Forests, and CNNs, often rely on static, cross-sectional data that fail to capture the progressive cognitive decline observed in real-world patient trajectories. To address these limitations, this study introduces a novel deep learning-based framework TCBiNet (Temporal Convolutional Bidirectional Attention Network) designed to model both short- and long-term symptom evolution using sequential clinical data. The architecture strategically integrates Temporal Convolutional Networks (TCN) to identify localized trends, BiLSTM for bidirectional temporal analysis, and a Temporal Attention mechanism to highlight critical symptom intervals. Implemented in Python using TensorFlow 2.11, the model processes longitudinal data from 2,149 patients aged 60–90, with cognitive, behavioral, and functional metrics segmented into 30-day intervals. Compared to conventional models such as CNN-LSTM and BiLSTM-DRL, TCBiNet demonstrated superior performance, achieving an accuracy of 99.51%, F1-score of 99.35, and AUC-ROC of 0.990 reflecting an improvement of over 4.4% in accuracy and 6.2% in F1-score relative to the best-performing existing models. This significant gain is attributed to the model’s capacity for temporal pattern mining and attention-guided symptom weighting. The proposed system not only improves prediction accuracy but also enhances interpretability and clinical relevance. By offering a robust, sequence-aware diagnostic tool, this research paves the way for proactive interventions in dementia care, encouraging further exploration into longitudinal neurocognitive modeling for early disease detection.

Surface water quality forecasting is crucial for pollution early warning and sustainable water resource management. However, accurate prediction of key water quality indicators remains challenging due to the highly nonlinear spatio-temporal dynamics and complex inter-variable relationships. Traditional statistical models and conventional machine learning approaches often struggle to effectively capture these couplings, leading to limited predictive performance. In this study, we propose a novel hybrid deep learning framework, termed MLA-Mamba, which integrates an improved Mamba-based sequence modeling network with a Multi-Head Local Attention (MLA) mechanism, optimized through a Gradient Reparameterization Optimization (GRPO) strategy. The Mamba module is designed to extract long-range temporal dependencies from water quality time series via a state-space modeling paradigm, while the MLA mechanism captures localized spatial correlations among multiple monitoring stations. To the best of our knowledge, this study represents one of the first explorations of applying Gradient Reparameterization Optimization (GRPO) to water quality prediction tasks. Furthermore, a multi-task learning scheme is incorporated to jointly predict multiple key indicators, including permanganate index (CODMn), ammonia nitrogen (NH3–N), total phosphorus (TP), and total nitrogen (TN), thereby exploiting inter-variable dependencies to enhance overall forecasting accuracy. The proposed GRPO strategy dynamically adjusts learning rates during training to accelerate convergence and improve model stability. Experimental evaluations on two real-world surface water datasets demonstrate that the proposed MLA-Mamba model achieves consistent performance improvements over the evaluated baseline methods across multiple error metrics. In addition, predictive uncertainty is quantified via Monte Carlo dropout, enabling the estimation of confidence intervals to support risk-aware water quality assessment. These results highlight the effectiveness of integrating advanced sequence modeling, attention-driven spatial feature extraction, and adaptive optimization for robust environmental time series forecasting.

Accurate short-term load forecasting is vital for the reliable and efficient operation of smart grids, particularly under the uncertainty introduced by variable renewable energy sources (RESs) such as solar and wind. This study introduces ST-CALNet, a novel hybrid deep learning framework that integrates convolutional neural networks (CNNs) with an Attentive Long Short-Term Memory (LSTM) network to enhance forecasting performance in renewable-integrated smart grids. The CNN component captures spatial dependencies from multivariate inputs, comprising meteorological variables and generation data, while the LSTM module models temporal correlations in historical load patterns. An embedded attention mechanism dynamically weights input sequences, enabling the model to prioritise the most influential time steps, thereby improving its interpretability and robustness during demand fluctuations. ST-CALNet was trained and evaluated using real-world datasets that include electricity consumption, solar photovoltaic (PV) output, and wind generation. Experimental evaluation demonstrated that the model achieved a mean absolute error (MAE) of 0.0494, root mean squared error (RMSE) of 0.0832, and a coefficient of determination (R2) of 0.4376 for electricity demand forecasting. For PV and wind generation, the model attained MAE values of 0.0134 and 0.0141, respectively. Comparative analysis against baseline models confirmed ST-CALNet’s superior predictive accuracy, particularly in minimising absolute and percentage-based errors. Temporal and regime-based error analysis validated the model’s resilience under high-variability conditions such as peak load periods, while visualisation of attention scores offered insights into the model’s temporal focus. These findings underscore the potential of ST-CALNet for deployment in intelligent energy systems, supporting more adaptive, transparent, and dependable forecasting within smart grid infrastructures.

Developmental psychology and affective computing have placed great emphasis on identifying children’s emotional cues in recent times. In this study, a novel Spatio-Temporal Multimodal Emotion Recognition Network (ST-MERN) for child emotion classification is proposed. Dense feature embeddings of the EmoReact dataset and temporal video sequences are utilized for the study. The proposed method uses 115 continuous frames per visual signal instance, e.g., rotational-translational vectors, facial keypoints, and pose predictions. With steady performance on each frame and a mean confidence of 0.967, this ensures the system maintains good detection fidelity. In order to track subtle emotional changes, our method captures dynamic data like scale variation and frame-to-frame variation (r x , r y , r z , t x , t y ). Latent features (p24–p33) provide a profound explanation of emotional states. The model is designed to preserve spatiotemporal consistency and improve emotion recognition by combining these features. Curiosity, uncertainty, excitement, happiness, surprise, disgust, fear, frustration, and valence are the nine categories on which the system categorizes children’s emotional states. Preliminary results show that our system effectively captures expressive nuances, with stable pose data and low feature variability across sequences. The model surpassed earlier models such as LSTM and TCN in generalization, with a high validation accuracy of 93.6% and test accuracy of 94.3% for the BiLSTM-based architecture. The BiLSTM model had enhanced classification capacity for different emotional states with an F1-score of 0.92. The TCN model is well-suited to real-time deployment since it recorded a competitive test accuracy of 91.7% with quick inference times of ~ 0.8 s per clip, even though it was slightly slower than the BiLSTM. With an F1-score of 0.89 and test accuracy of 90.2%, the LSTM model performed robustly; it trained faster than the BiLSTM and TCN, although its accuracy was slightly lower. By providing strong and interpretable classification that is sensitive to the dynamic nature of children’s emotional displays, this technique improves emotion detection in children. Our work provides the foundation for socially sensitive systems, therapy treatments, and affect-conscious education materials.

This paper proposes a method for forecasting surface solar irradiance (SSI), the most critical factor in solar photovoltaic (PV) power generation. The proposed method uses 16-channel data obtained by the GEO-KOMPSAT-2A (GK2A) satellite of South Korea as the main data for SSI forecasting. To determine feature variables related to SSI from the 16-channel data, the differences and ratios between the channels were utilized. Additionally, to consider the fundamental characteristics of SSI originating from the sun, solar geometry parameters, such as solar declination (SD), solar elevation angle (SEA), and extraterrestrial solar radiation (ESR), were used. Deep learning-based feature selection (Deep-FS) was employed to select appropriate feature variables that affect SSI from various feature variables extracted from the 16-channel data. Lastly, spatio-temporal deep learning models, such as convolutional neural network–long short-term memory (CNN-LSTM) and CNN–gated recurrent unit (CNN-GRU), which can simultaneously reflect temporal and spatial characteristics, were used to forecast SSI. Experiments conducted to verify the proposed method against conventional methods confirmed that the proposed method delivers superior SSI forecasting performance.

Falls among older adults are a major cause of injury and loss of independence, yet most existing systems detect falls only after onset or provide very limited warning time. This study presents a synthetic benchmarking framework for early fall-risk prediction using multimodal wearable-inspired time-series data and compares classical and temporal machine learning architectures under a realistic evaluation protocol. A synthetic dataset of 1000 sequences was generated to emulate normal activity, slip events, and pre-fall instability using biomechanical, physiological, and contextual variables. Eight baseline models and two augmented temporal variants were trained and evaluated using subject-wise splits to reduce leakage. Performance differed substantially by model family and evaluation protocol. Classical baselines achieved the strongest overall macro-F1 scores, whereas temporal models showed more modest discrimination. Under a fixed alerting rule, operational early-warning behavior varied considerably: some models failed to trigger alerts, while others achieved higher pre-fall trigger rates at the cost of increased false alarms. These findings show that apparent performance depends strongly on partitioning strategy, calibration, and alert design. The proposed framework provides a reproducible basis for benchmarking early-warning fall-risk models and supports future validation using real-world cohorts and deployment-oriented calibration strategies.

Land cover change (LCC) studies are increasingly using deep learning (DL) modeling techniques. Past studies have leveraged temporal or spatiotemporal sequences of historical LC data to forecast changes with DL models. However, these studies do not adequately assess the association between neighborhood size and DL model capability to forecast LCCs, where neighborhood size refers to the spatial extent captured by each data sample. The objectives of this research study were to: (1) evaluate the effect of neighborhood size on the capacity of DL models to forecast LCCs, specifically Temporal Convolutional Networks (TCN) and Convolutional Neural Networks (CNN-TCN), and (2) assess the effect of auxiliary spatial variables on model capacity to forecast LCCs. First, each model type and neighborhood setting configuration was assessed using data derived from multitemporal MODIS LC for the Regional District of Bulkley-Nechako, Canada, comparing subareas exhibiting different amounts of LCCs with trends obtained for the full region. Next, outcomes were compared with three other study regions. The modeling results were evaluated with three-map comparison measures, where the real-world LC for the next timestep, the real-world LC for the previous timestep, and the forecasted LC for the next year were used to calculate correctly transitioned areas. Across all regions explored, it was observed that increasing neighborhood sizes improved the DL model’s capabilities to forecast short-term LCCs. CNN–TCN models forecasted the most correct LCCs for several regions while reducing error due to quantity when provided additional spatial variables. This study contributes to the systematic exploration of neighborhood sizes on selected spatiotemporal DL techniques for geographic applications.

: Peripheral nerve ultrasound is used to evaluate nerve enlargement in immune-mediated neuropathies; however, assessment can be challenging because the distribution and severity of nerve enlargement vary among patients and are often accompanied by indistinct nerve boundaries and heterogeneous echogenicity. Although deep learning-based segmentation has been reported, most studies have focused on limited regions or single anatomical sites, primarily in compressive neuropathies. This study aimed to evaluate the performance of temporal deep learning-based segmentation for assessing diffuse or focal nerve enlargement in immune-mediated neuropathies using continuous ultrasound scans. : Twenty-five healthy participants and five patients with immune-mediated neuropathy and nerve enlargement were included. Continuous ultrasound scanning from the wrist to below the elbow was performed. A static DeepLabV3+ model and temporal models incorporating convolutional long short-term memory (ConvLSTM) or Temporal Mamba were constructed and compared. : In healthy participants, segmentation performance was comparable across models. In contrast, in patients with nerve enlargement, temporal models demonstrated higher Dice coefficients and reduced frame-to-frame variability. The ConvLSTM-based model showed the highest performance, with mean Dice coefficients ranging from 0.87 to 0.92. : Temporal deep learning showed potential for nerve segmentation in selected cases with nerve enlargement associated with immune-mediated neuropathies. Temporal models achieved improved segmentation performance and reduced frame-to-frame variability in these preliminary cases. This approach may facilitate more consistent quantitative ultrasound evaluation and warrants further validation in larger cohorts.