Search

To enhance the accuracy of wind power prediction, we introduce a novel prediction method: the IVMD-CNN-GRU-Attention model integrated with sample entropy fusion. This approach initially decomposes the raw wind power series using tailored indicators $\Gamma $ , subsequently classifying and inputting the decomposed sub-modes, based on their central frequencies and sample entropies, into a hybrid model comprising CNN, GRU, and an attention mechanism. This innovative model was rigorously tested on SCADA data from a Chinese wind farm, achieving remarkable improvements over existing methods. Specifically, average enhancements of 12.06% in R2 score, 59.43% reduction in MAE, 52.04% reduction in RMSE, and 48.40% reduction in MAPE were observed. These substantial outcomes demonstrate that our method significantly enhances wind power prediction accuracy, thereby contributing to the advancement of the wind energy industry and ensuring stable power grid operation.

This research addresses the imperative need for advanced detection mechanisms for the identification of phishing websites. For this purpose, we explore state-of-the-art machine learning, ensemble learning, and deep learning algorithms. Cybersecurity is essential for protecting data and networks from threats. Detecting phishing websites helps prevent fraud and safeguard personal information. To evaluate the efficacy of our proposed method, the top features using information gain, gain ratio, and PCA are used to predict and identify a website as phishing or non-phishing. The proposed system is trained using a dataset that covers 11,055 websites. The ensemble learning model applied achieved an impressive 99% accuracy in predicting phishing websites, surpassing previous models, and setting a new benchmark in the field. The findings highlight the effectiveness of combining deep learning architectures with ensemble learning, offering not only improved accuracy but also adaptability to emerging phishing techniques.

Fish is one of the most demanding protein sources in the food industry. However, the increasing demand must be followed by increasing production efficiency. One of the problems in fish production efficiency is an ineffective feeding method. In this paper, we address the problem of fish feeders using artificial intelligence in an aquarium. We propose fish appetite detection using multi-modal sensors resulting in the data for our AI system. The main aim is to improve the efficiency of fish feeding using a multi-modal sensor system. Our AI system consists of Residual ( $2+1$ )-Dimensional Convolutional Neural Networks (R( $2+1$ )D-CNN) and dense networks to process video and accelerometer data. The video data is split into 20 frames and processed by an R( $2+1$ )D-CNN. Furthermore, the accelerometer data is used to train several networks such as 1-dimensional Convolutional Neural Networks (CNN), Gated Recurrent Networks (GRU), and Dense Networks or Artificial Neural Networks (ANN). The system is implemented in an aquarium with two sensors i.e., a webcam camera and an accelerometer and a main board processing using Raspberry-Pi 4. Experimental results show that the proposed system outperforms other methods with validation accuracy up to 99.09% for the Zeromean dense model and up to 99.39% for the Filtered dense model. The results also show that the multi-modal sensor system increases the accuracy significantly without a significant increase in the number of parameters. The work is useful for automation and efficiency in aquaculture.

This paper aims to find that predictive performance is better when earning management predictions using deep learning technology, including litigation case information filed with companies. Unlike previous accounting forensics studies, this study designed a basic earning management prediction model using related financial variables by referring to various previous studies on earning management in accounting. To present more objective and reliable verification results, this study used two measures of accrual earning management (AEM) as proxy variables for corporate earning management. It used four deep-learning classifiers: RNN, GRU, LSTM, and Transformers. In addition, before predicting earning management through deep learning analysis, regression analysis proved that the number and amount of litigation cases filed against the company significantly reduced accrual earning management. In this study, the number of litigation cases filed with the company and the amount of litigation were used as information on the company’s litigation cases. The main findings of this paper are as follows. First, as expected, when predicting the earning management level, including the company’s litigation case information, accuracy, recall, precision, AUC (area under the ROC curve), and F1 score, all showed high overall predictive performance. Second, overall predictive accuracy was higher when predicting the level of earning management, including the number of litigation cases, than when predicting the level of earning management, including the number of litigation cases. Overall predictive performance was the highest when predicting the earning management level, including the number and amount of litigation cases. Third, among the four classifiers of deep learning, Transformers’ predictive performance was the best, followed by LSTM, GRU, and RNN.

Stock market prediction poses significant challenges due to the inherent noise and volatility of the data. These challenges are further amplified in emerging stock markets, where data volatility increases due to numerous endogenous and exogenous variables. Despite the progress made in models for stock market prediction, such as Autoregressive Integrated Moving Average (ARIMA), Support Vector Machines (SVMs), and deep learning models, there is still a need for further research in emerging stock markets. This study addresses the complexity and non-linearity of emerging stock market data by proposing a deep learning model that utilizes Gated Recurrent Unit (GRU) algorithm to predict the next-day closing price. The proposed model leverages the inclusion of exogenous variables to enhance the model’s performance. Three datasets are constructed for three main emerging market indices, specifically in Qatar, Saudi Arabia, and China. Using mean absolute percentage error (MAPE), the inclusion of exogenous variables led to a noticeable improvement over the related work results from 0.74, 1.68, and 0.72 for indices of Qatar, Saudi Arabia, and China respectively to 0.16, 0.6, and 0.2. Furthermore, the results demonstrate the appropriateness of GRU algorithm for predicting emerging stock markets.

This paper presents a novel approach combining Gated Recurrent Unit (GRU) and adaptive Proportional-Integral-Derivative (PID) algorithm employed for predicting tire slip angles (TSA). In previous studies, research has shown successful prediction of tire slip angles using Recurrent Neural Network (RNN). Research has shown that when using only RNN for prediction, the model’s accuracy is low in untrained speed conditions. For this investigation, vehicle data under constant speed double lane change (DLC) conditions and S-Turn variable speed (STVS) scenarios were generated using the CarSim software. The dataset includes variables such as tire lateral force, tire longitudinal force, tire vertical force, vehicle TSA, and time. A total of three GRU models were utilized in this research. The first GRU model is employed for primary predictions, the second GRU model is utilized to predict the discrepancies between the first model’s predictions and the actual values, and the third GRU model is employed to learn variations in the PID weight parameters, facilitating further optimization of the predictions. The training of the GRU models is conducted using the TensorFlow framework. An adaptive PID was realized by utilizing the predicted PID parameter changes from the third GRU model, with the aim of further refining the prediction errors. In theory, the output values of GRU2 can be directly subtracted from GRU1. Simulation results showed that the GRU_APID algorithm performed better compared with GRU1 minus GRU2. Finally, the predictions generated by the first GRU model are combined with the corrective values generated by the last adaptive PID to obtain the ultimate TSA prediction. Simulation results have confirmed that this approach indeed enhances the model’s performance and increases TSA prediction accuracy under untrained speed conditions.

Deep learning (DL), a branch of machine learning (ML), is the core technology in today’s technological advancements and innovations. Deep learning-based approaches are the state-of-the-art methods used to analyse and detect complex patterns in large datasets, such as credit card transactions. However, most credit card fraud models in the literature are based on traditional ML algorithms, and recently, there has been a rise in applications based on deep learning techniques. This study reviews the recent DL-based literature and presents a concise description and performance comparison of the widely used DL techniques, including convolutional neural network (CNN), simple recurrent neural network (RNN), long short-term memory (LSTM), and gated recurrent unit (GRU). Additionally, an attempt is made to discuss suitable performance metrics, common challenges encountered when training credit card fraud models using DL architectures and potential solutions, which are lacking in previous studies and would benefit deep learning researchers and practitioners. Meanwhile, the experimental results and analysis using a real-world dataset indicate the robustness of the deep learning architectures in credit card fraud detection.

Non-Intrusive Load Monitoring (NILM) represents a technique for energy monitoring that is adept at discerning the quantity, type, and operational status of appliances through the analysis of aggregate power data within a specified area. This method facilitates the disaggregation of total energy consumption into the power usage of individual appliances. Recent advancements in deep learning have led to the application of various neural networks in the context of NILM. However, the presence of multiple appliances operating concurrently in various states, coupled with significant disparities in appliance switching behavior, poses challenges to the efficacy of disaggregation models. To mitigate these challenges, this study introduces a novel load-energy disaggregation algorithm, dubbed GTA-BT, which integrates Gated Recurrent Units (GRU) and a bidirectional Transformer architecture featuring time-aware self-attention. This approach aims to enhance disaggregation accuracy for appliances that exhibit multi-stage operational patterns and low utilization rates. The bidirectional Transformer is employed to capture local feature dependencies across different segments of the power consumption sequence, while the integration of GRU facilitates effective modeling of temporal sequences. In alignment with NILM task requirements, this model incorporates a device time state variable within the self-attention mechanism to augment focus on active appliances and extends the local receptive field through convolutional layer expansion. Moreover, a novel masking strategy, predicated on the operational state of devices, has been developed to heighten the model’s responsiveness to transitions in multi-stage appliance states. Comparative evaluations of this method against leading NILM algorithms using the REDD and UK-DALE datasets demonstrate its superior performance, evidenced by a 25% reduction in Mean Absolute Error (MAE) across all appliances, a 20% enhancement in F1-Score, and overall improved disaggregation capability.

Sentiment analysis, a pivotal task in natural language processing, aims to discern opinions and emotions expressed in text. However, existing methods for sentiment analysis face various challenges such as data scarcity, complex language patterns, and long-range dependencies. In this paper, we propose MPNet-GRUs, a hybrid deep learning model that integrates three key components: MPNet, BiGRU, and GRU. MPNet, a transformer-based pre-trained language model, enhances language understanding through masked and permuted pre-training. BiGRU and GRU, recurrent neural networks, capture long-term dependencies bidirectionally and unidirectionally. By combining the strengths of these models, MPNet-GRUs aims to provide a more effective and efficient solution for sentiment analysis. Evaluation on three benchmark datasets reveals the superior performance of MPNet-GRUs: 94.71% for IMDb, 86.27% for Twitter US Airline Sentiment, and 88.17% for Sentiment140, demonstrating its potential to advance sentiment analysis.

Malaria is a mosquito-borne, life-threatening, and contagious disease that has caused thousands of fatalities in recent years. Due to inadequate detection, the inexperience of laboratory personnel, and lack of advanced point-of-care equipment, the malaria-induced mortality rate is increasing. In addition to the traditional detection mechanisms, researchers have recently been investigating microscopic malaria-infected Red Blood Cells (RBC) image analysis based on deep learning models to detect malaria parasites as a general-purpose point-of-care solution. In this paper, we develop several hybrid data-driven models by combining a convolutional neural network (CNN) to extract the relevant features and two cascaded recurrent neural networks (RNNs) classifiers. Gated recurrent unit (GRU), long short-term memory (LSTM), and Bi-directional LSTM (BiLSTM) are considered candidate RNN classifiers. The models are compared in terms of accuracy, type-I & II error rates, and inference & training computation time. Experimental results demonstrate that the CNN-LSTM-BiLSTM model outperforms the other models with a significantly higher accuracy (96.20%), less type-I error rate (2.23%), and fewer combined type-I and type-II errors (3.80%). Also, we consider the model computation time (both training time per epoch and inference time per step)‘ as an important metric for emerging distributed learning paradigms where point-of-care devices with IoT (Internet of Things) capability can jointly contribute to global model accuracy. Thus, our findings demonstrate the practicality of cascading classifiers in resource-constrained point-of-care devices.

The widespread use of the Internet and the tremendous growth of social media have enabled people to connect with each other worldwide. Individuals are free to express themselves online, sharing their photos, videos, and text messages globally. However, such freedom sometimes leads to misuse, as some individuals exploit this platform by posting hateful and abusive comments on forums. The proliferation of abusive language on social media negatively impacts individuals and groups, leading to emotional distress and affecting mental health. It is crucial to automatically detect and filter such abusive content in order to effectively tackle this challenging issue. Detecting abusive language in text messages is challenging due to intentional word concealment and contextual complexity. To counter abusive speech on social media, we need to explore the potential of machine learning (ML) and deep learning (DL) models, particularly those equipped with attention mechanisms. In this study, we utilized popular ML and DL models integrated with attention mechanism to detect abusive language in Urdu text. Our methodology involved employing Count Vectorizer and Term Frequency-Inverse Document Frequency (TF/IDF) to extract n-grams at the word level: Unigrams (Uni), Bigrams (Bi), Trigrams (Tri), and their combination (Uni + Bi + Tri). Initially, we evaluated four traditional ML models—Logistic Regression (LR), Gaussian Naïve Bayes (NB), Support Vector Machine (SVM), and Random Forest (RF)—on both proposed and established datasets. The results highlighted that RF model outperformed other conventional models in terms of accuracy, precision, recall, and F1-measure on both datasets. In our implementation of deep learning models, we employed various models integrated with custom fastText and Word2Vec embeddings, each equipped with an attention layer, except for the Convolutional Neural Network (CNN). Our findings indicated that the Bidirectional Long Short-Term Memory (Bi-LSTM) + attention model, utilizing custom Word2Vec embeddings, exhibited improved performance in detecting abusive language on both datasets.

In this work, for the first time, a machine learning behavioral modeling methodology based on gate recurrent unit (GRU) is developed and used to model and then analyze the kink effects (KEs) in the output reflection coefficient $(S_{22})$ and the short-circuit current gain $(h_{21})$ of an advanced microwave transistor. The device under test (DUT) is a 0.25- $\mu \text{m}$ gallium nitride (GaN) high electron mobility transistor (HEMT) on silicon carbide (SiC) substrate, which has a large gate periphery of 1.5 mm. The scattering (S-) parameters of the DUT are measured at a frequency up to 65 GHz and at an ambient temperature up to 200°C. The proposed model can accurately reproduce the KEs in $S_{22}$ and in $h_{21}$ , enabling an effective analysis of their dependence on the operating conditions, bias point and ambient temperature. It is worth noticing that the proposed transistor model shows also good performance in both interpolation and extrapolation test.

Recently, systems for classifying gait disorders have been of great interest. However, quantifying the progress of these disorders has been highly dependent on a physician’s judgement in classifying sick and healthy subjects. We examine the effects of gait stability analysis on gait dysfunction problems, which are impacted by the patient’s dynamic balance. The dataset in this study was collected and labelled based on the opinions of physicians at Prague Hospital; it included 84 measurements of 37 patients. A keypoint detector was applied to detect the skeletal keypoints of patients. We have prepared two different datasets from the detection and tracking results. For the proposed feature selection method, we have used statistical measurements such as the x and y coordinates for each keypoint, the distance, and the angle between two selected keypoints. Using these statistical measurements, we have prepared different subgroups with different numbers of features to examine. We have also applied ten different feature selection algorithms to obtain data from different numbers of features automatically. Then, these datasets with high-level features were used to train well-known networks, such as the long short-term memory (LSTM), gated recurrent unit (GRU), and multiple layer perceptron (MLP) networks. The study results showed that the 30 features selected by the analysis of variance (ANOVA) algorithm and used to train the GRU network ranked among the best features and resulted in a classification $F$ -score of 85%. The results also prove that the data generated by the detector method are more effective than the data generated by the tracking method due to the format of the exercises in our dataset, which were designed by physicians. Moreover, the best feature selection approaches have considerably improved the classification $F$ -score compared to manual feature generation.

In recent years, there has been a remarkable increase in the usage of Deep Neural Networks (DNNs) for addressing and solving electrical field problems. This research primarily aims to present an advanced approach to classify different motor faults based on their time-series data by implementing a new Recurrent Neural Network (RNN) model that consists of mixed Long short-term memory (LSTM), Gated Recurrent Unit (GRU), and two Fully Connected (FC) layers. The main idea of this study centers on developing one comprehensive model capable of categorizing primary motor faults. The proposed model is supposed to classify 10 different classes, extracted from the Machinery Fault Database (MaFaulDa), which are normal (no-fault), vertical misalignment, horizontal misalignment, imbalance, overhang-ball, overhang-cage, overhang-outer race, underhang-ball, underhang-outer race, and underhang-cage. Classifying 10 different situations can be considered as a notable classification problem. Additionally, the learning period did not include any data augmentation, which reflects the model’s power in training over the available data. Significantly, the accuracy of the model is enhanced by setting precise values for hyperparameters, including network structure (number of layers and neurons), learning rate, regularization, optimizer type, number of epochs, and more. The obtained train-validation-test accuracies from the proposed model are 99.87%, 99.599%, and 99.48%, respectively. The accuracy of the model represents the highest accuracy among other publications. This advanced approach offers numerous advantages, including early-stage fault detection, improved robustness in industrial maintenance, and generating fast and intelligent alerts, thereby reducing the possible damage to electrical instruments.

Precise outpatient volume prediction holds significant importance in hospital management. While the Gated Recurrent Unit (GRU) is a frequently utilized deep learning technique for forecasting hospital outpatient volumes, creating a proficient GRU model necessitates the fine-tuning of pertinent GRU parameters. The adjustment of suchparameters relies heavily on an individual’s practical experience and prior knowledge. The recently proposed Cheetah optimizer is a novel intelligent algorithm with unique optimization capabilities. The Cheetah optimizer holds significant research potential; however, additional investigations are warranted, as it may be vulnerable to issues related to local optimization. In the present study, the selection of hyperparameters for the GRU model was optimized through the utilization of the Modified Cheetah Optimization (MCO) algorithm, and a combined MCO-GRU model was established. Using the Successive Variational Mode Decomposition (SVMD) method to decompose outpatient volume sample data, the parameters of the GRU model were optimized with the MCO method to construct a hybrid forecasting model. This yielded the smallest Root Mean Square Error (RMSE) for the proposed model, with a value of 0.0843. Additionally, the results indicate that in comparison to SVMD, Long Short-Term Memory (LSTM), GRU, Particle Swarm Optimization-GRU (PSO-GRU), and Cheetah Optimization-GRU (CO-GRU), the proposed model significantly enhanced the accuracy of outpatient volume forecasting.

Human motion prediction is a popular method to predict future motion sequences based on past sequences, which is widely used in human-computer interaction. Space-time-separable graph Convolutional Network (STS-GCN) is a conventional mathematical model for human motion prediction. However, the uncertainty of human movements often leads to the problem of significant prediction error in the prediction results. This paper first proposed a Multi-scale STS-GCN (MSTS-GCN) model based on the conventional STS-GCN method to find the relevant factors that affect the prediction results. In our study, the constructed Multi-scale Temporal Convolutional Network (MTCN) decoder effectively reduced the human motion prediction error at specific time nodes. To expand the transmission and utilization performance in a larger receptive field, a Gated Recurrent Unit-TCN decoder was also designed. Finally, a new STS-GCN (NSTS-GCN) human motion prediction model was proposed, which realized the transmission and utilization of motion sequence features under a larger temporal perceptual field. To verify the effectiveness of NSTS-GCN, the Human3.6M dataset, AMASS, and 3DPW dataset were tested. The experimental results show that the MPJPE error of the proposed model for human joint prediction at each time node is reduced compared with the conventional STS-GCN model, and the mean reduction was achieved by 3.0mm. All the experimental results validated the effectiveness of the proposed NSTS-GCN model, which further improved the performance of human motion prediction.

Emotion recognition is susceptible to interference such as feature redundancy and speaker gender differences, resulting in low recognition accuracy. This paper proposes a speech emotion recognition (SER) method based on attention mixed convolutional neural network (MCNN) combined with gender information, including two stages of gender recognition and emotion recognition. (1) Using MCNN to identify gender and classify speech samples into male and female. (2) According to the output of the first stage classification, a gender-specific emotion recognition model is established by introducing coordinated attention and a series of gated recurrent network units connecting the attention mechanism (A-GRUs) to achieve emotion recognition results of different genders. The inputs of both stages are dynamic 3D MFCC features generated from the original speech database. The proposed method achieves 95.02% and 86.34% accuracy on EMO-DB and RAVDESS datasets, respectively. The experimental results show that the proposed SER system combined with gender information significantly improves the recognition performance.

In healthcare, a huge amount is paid to meet the requirements of High-Need High-Cost (HNHC) patients, also known as super-utilizers. The major aim of the proposed study is to predict HNHC patients. This paper proposes hybrid Deep Learning (DL) models that identify HNHC patients, and help the management to manage their resources and budgets accordingly. We use the strategy of the Interquartile Range (IQR) that specifies the range of HNHC. The dataset used in this work has a lot of inconsistencies that are tackled by different techniques. Using three DL models, Gated Recurrent Unit (GRU), Fully Convolutional Network (FCN) and Vanilla Recurrent Neural Network (VRNN), we proposed 2 hybrid DL models, FCN-VRNN and GRU-FCN, for the prediction of HNHC patients. The performance of these models is evaluated based on different performance metrics, Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE), R-Squared ( $R^{2}$ ) and execution time. The results depict that the proposed hybrid models show more competitive and outperforming results than the individual models. Proposed hybrid system model 1 FCN-VRNN achieves MSE of 0.4%, RMSE of 7%, MAPE of 29.8%, MAE of 3%, and $R^{2}$ of 99.4%. Proposed hybrid system model 2 GRU-FCN achieves MSE of 0.4%, RMSE of 6.5%, MAPE of 23.55%, MAE of 2.5%, and $R^{2}$ of 99.5%.

The massive digital information generated in conjunction with the ever-increasing phasor measurement data in the power grid has led to a tremendous constraint on the analysis and timely processing of real-time data. Under these conditions, leveraging Artificial Intelligence (AI) can play a crucial role in assisting more efficient data processing and analysis. In this paper an AI-assisted power grid event classification method is proposed, which aims at improving the overall power grid system performance. Furthermore, an edge cloud sharing scheme is introduced for a large-scale power grid system. To balance the load and reduce the maximum processing time, a multiple edge cloud node-based scheme is developed. The simulation results verify that the proposed AI-assisted event classification method, together with the edge cloud sharing scheme, can significantly improve the overall performance of the system.

Congestion in the world’s traffic systems is a major issue that has far-reaching repercussions, including wasted time and money due to longer commutes and more frequent stops for gas. The incorporation of contemporary technologies into transportation systems creates opportunities to significantly improve traffic prediction alongside modern academic challenges. Various techniques have been utilized for the purpose of traffic flow prediction, including statistical, machine learning, and deep neural networks. In this paper, a deep neural network architecture based on long short-term memory (LSTM), bi-directional version, and gated recurrent units (GRUs) layers has been structured to build the deep neural network in order to predict the performance of the traffic flow in four distinct junctions, which has a great impact on the Internet of vehicles’ applications. The structure is composed of sixteen layers, five of which are GRU layers and one is a bi-directional LSTM layer. The dataset employed in this work involved four congested junctions. The dataset extended from November 1, 2016, to June 30, 2017. Cleaning and preprocessing operations were performed on the dataset before feeding it to the designed deep neural network in this paper. Results show that the suggested method produced comparable performance with respect to state-of-the-art approaches.

Over the past few years, great importance has been given to wireless sensor networks (WSNs) as they play a significant role in facilitating the world with daily life services like healthcare, military, social products, etc. However, heterogeneous nature of WSNs makes them prone to various attacks, which results in low throughput, and high network delay and high energy consumption. In the WSNs, routing is performed using different routing protocols like low-energy adaptive clustering hierarchy (LEACH), heterogeneous gateway-based energy-aware multi-hop routing (HMGEAR), etc. In such protocols, some nodes in the network may perform malicious activities. Therefore, four deep learning (DL) techniques and a real-time message content validation (RMCV) scheme based on blockchain are used in the proposed network for the detection of malicious nodes (MNs). Moreover, to analyse the routing data in the WSN, DL models are trained on a state-of-the-art dataset generated from LEACH, known as WSN-DS 2016. The WSN contains three types of nodes: sensor nodes, cluster heads (CHs) and the base station (BS). The CHs after aggregating the data received from the sensor nodes, send it towards the BS. Furthermore, to overcome the single point of failure issue, a decentralized blockchain is deployed on CHs and BS. Additionally, MNs are removed from the network using RMCV and DL techniques. Moreover, legitimate nodes (LNs) are registered in the blockchain network using proof-of-authority consensus protocol. The protocol outperforms proof-of-work in terms of computational cost. Later, routing is performed between the LNs using different routing protocols and the results are compared with original LEACH and HMGEAR protocols. The results show that the accuracy of GRU is 97%, LSTM is 96%, CNN is 92% and ANN is 90%. Throughput, delay and the death of the first node are computed for LEACH, LEACH with DL, LEACH with RMCV, HMGEAR, HMGEAR with DL and HMGEAR with RMCV. Moreover, Oyente is used to perform the formal security analysis of the designed smart contract. The analysis shows that blockchain network is resilient against vulnerabilities.

Aiming at the problem that the traditional license plate recognition method lacking of accuracy and speed, an end-to-end deep learning model for license plate location and recognition in natural scenarios was proposed. First, we added an improved channel attention mechanism to the down-sampling process of the You only look once(YOLOv5). Additionally, a location information is added in the ones to minimize the information loss from sampling, which can improve the feature extraction ability of the model. Then we reduce the number of parameters on the input side and set only one class in the YOLO layer, which improves the efficiency and accuracy of the detector for locating license plates. Finally, Gated recurrent units(GRU) + Connectionist temporal classification(CTC) was used to build the recognition network to complete the character segmentation-free recognition task of the license plate, significantly shortened the training time and improved the convergence speed and recognition accuracy of the network. The experimental results show that the average recognition precision of the license plate recognition model proposed in this paper reaches 98.98%, which is significantly better than the traditional recognition algorithm, and the recognition effect is good in complex environment with good stability and robustness.

The Rainfall-Runoff (R-R) relationship is essential to the hydrological cycle. Sophisticated hydrological models can accurately investigate R-R relationships; however, they require many data. Therefore, machine learning and soft computing techniques have taken the attention in the environment of limited hydrological, meteorological, and geological data. The accuracy of such models depends on the various parameters, including the quality of inputs and outputs and the used algorithms. However, identifying a perfect algorithm is still challenging. This study develops a fuzzy logic-based algorithm called Cascaded-ANFIS to accurately predict runoff based on rainfall. The model was compared against three regression algorithms: Long Short-Term Memory, Grated Recurrent Unit, and Recurrent Neural Networks. These algorithms have been selected due to their outstanding performances in similar studies. The models were tested on the Mahaweli River, the longest in Sri Lanka. The results showcase that the Cascaded-ANFIS-based model outperforms the other algorithms. The correlation coefficient of each algorithm’s predictions was 0.9330, 0.9120, 0.9133, 0.8915, 0.6811, 0.6811, and 0.6734 for the Cascaded-ANFIS, LSTM, GRU, RNN, Linear, Ridge, and Lasso regression models respectively. Hence, this study concludes that the proposed algorithm is 21% more accurate than the second-best LSTM algorithm. In addition, Shared Socio-economic Pathways (SSP2-4.5 and SSP5-8.5 scenarios) were used to generate future rainfalls, forecast the near-future and mid-future water levels, and identify potential flood events. The future forecasting results indicate a decrease in flood events and magnitudes in both SSP2-4.5 and SSP5-8.5 scenarios. Furthermore, the SSP5-8.5 scenario shows drought weather from May to August yearly. The results of this study can effectively be used to manage and control water resources and mitigate flood damages.

Device heterogeneity can cause a detrimental impact on the classification of healthcare data. In this work, we propose the Maximum Difference-based Heterogeneity Mitigation (MDHM) method to address device heterogeneity. Mitigating heterogeneity increases the reliability of using multiple devices from different manufacturers for measuring a particular physiological signal. Further, we propose an attention-based bilevel GRU (Gated Recurrent Unit) model, abbreviated as AT2GRU, to classify multi-modal healthcare time-series data for human emotion recognition. The physiological signals of Electroencephalogram (EEG) and Electrocardiogram (ECG) for twenty-three persons are leveraged from the DREAMER dataset for emotion recognition. Also, from the DEAP dataset, the biosignals namely EEG, Galvanic Skin Response (GSR), Respiration Amplitude (RA), Skin Temperature (ST), Blood Volume (BV), Electromyogram (EMG) and Electrooculogram (EOG) of thirty-two persons are used for emotion recognition. The EEG and the other biosignals are denoised by the wavelet filters for enhancing the model's classification accuracy. A multi-class classification is carried out considering valence, arousal, and dominance for each person in the datasets. The classification accuracy is validated against the self-assessment obtained from the respective person after watching a movie/video. The proposed AT2GRU model surpasses the other sequential models namely Long Short Term Memory (LSTM) and GRU in performance. [ABSTRACT FROM AUTHOR]

In smart cities of the future, data will be generated, integrated, processed and utilized from heterogeneous sources and at varying levels of complexity. For urban traffic planning in smart cities, one of the biggest challenges is traffic congestion prediction and its avoidance. Traffic congestion is a complex phenomenon and it is a manifestation of various contributing factors. In addition to vehicular mobility, properties of road network, weather, holidays and peak hours play a significant role in traffic congestion especially on arterial roads within a city. In this paper, we proposed a hybrid GRU-LSTM based deep learning model and applied it on city-wide novel traffic data integrated from heterogeneous sources. We have devised our indigenous data pipeline that is composed of a set of algorithms dealing with map matching, sparsity handling, outlier removal, zero speed adjustments, Open Street Map (OSM) and segment mapping etc. Extensive experimentations have been carried out to demonstrate the improved performance of the proposed method. The comparative analysis reveals that our methodology yields 95 % accuracy that outperforms other deep neural network models.

Presently, the volatile and dynamic aspects of stock prices are significant research challenges for stock markets or any other financial sector to design accurate and profitable trading strategies in all market situations. To meet such challenges, the usage of computer-aided stock trading techniques has grown in prominence in recent decades owing to their ability to rapidly and accurately analyze stock market situations. In the recent past, deep reinforcement learning (DRL) methods and trading bots are commonly utilized for algorithmic trading. However, in the existing literature, the trading agents employ the historical and present trends of stock prices as an observing state to make trading decisions without taking into account the long-term market future pattern of stock prices. Therefore, in this study, we proposed a novel decision support system for automated stock trading based on deep reinforcement learning that observes both past and future trends of stock prices whether single and multi-step ahead as an observing state to make the optimal trading decisions of buying, selling, and holding the stocks. More specifically, at every time step, future trends are monitored concurrently using a forecasting network whose output is concatenated with past trends of stock prices. The concatenated vectors are subsequently supplied to the DRL agent as an observation state. In addition, the suggested forecasting network is built on a Gated Recurrent Unit (GRU). The GRU-based agent captures more informative and inherent aspects of time-series financial data. Furthermore, the suggested decision support system has been tested on several stock markets such as Tesla, IBM, Amazon, CSCO, and Chinese Stocks as well as equity markets i-e SSE Composite Index, NIFTY 50 Index, US Commodity Index Fund, and has achieved encouraging profit values while trading.

Human activity recognition (HAR) is one of the important research areas in pervasive computing. Among HAR, sensor-based activity recognition refers to acquiring a high-level knowledge about human activities from readings of many low-level sensors. In recent years, although the existing methods of deep learning (DL) have been widely used for sensor-based HAR with some good performance, they still face such challenges as feature extraction and characterization, continuous action segmentation in dealing with time series problems. In this study, a multichannel fusion model is proposed with the idea of dividing. In this proposed architecture, a multichannel convolutional neural network (CNN) is used to enhance the ability to extract features at different scales, and then the fused features are fed into the gated recurrent unit (GRU) for feature labeling and enhanced feature representation, through the learning of temporal relationships. Finally, the multichannel CNN-GRU model is designed using global average pooling (GAP) to connect the feature maps with the final classification. The model performance was conducted on three benchmark datasets of WISDM, UCI-HAR, and PAMAP2 with the accuracy of 96.41%, 96.67%, and 96.25% respectively. The results show that the proposed model demonstrates better activity detection capability than some of the reported results.

Natural language processing (NLP) has been one of the subfields of artificial intelligence much affected by the recent neural revolution. Architectures such as recurrent neural networks (RNNs) and attention-based transformers helped propel the state of the art across various NLP tasks, such as sequence classification, machine translation, and natural language inference. However, if neural models are to be used in high-stakes decision making scenarios, the explainability of their decisions becomes a paramount issue. The attention mechanism has offered some transparency in the workings of otherwise black-box RNN models: attention weights (scalar values assigned input words) invite to be interpreted as the importance of that word, providing a simple method of interpretability. Recent work, however, has questioned the faithfulness of this practice. Subsequent experiments have shown that faithfulness of attention weights may still be achieved by incorporating word-level objectives in the training process of neural networks. In this article, we present a study that extends the techniques for improving faithfulness of attention based on regularization methods that promote retention of word-level information. We perform extensive experiments on a wide array of recurrent neural architectures and analyze to what extent the explanations provided by inspecting attention weights are correlated with the human notion of importance. We find that incorporating tying regularization consistently improves both the faithfulness (−0.14 F1, +0.07 Brier, on average) and plausibility ( $+53.6\%$ attention mass on salient tokens) of explanations obtained through inspecting attention weights across analyzed datasets and models.

Water quality prediction (WQP) plays an essential role in water quality management for aquaculture to make aquaculture production profitable and sustainable. In this work, we propose hybrid deep learning (DL) models, convolutional neural network (CNN) with the long short-term memory (LSTM) and gated recurrent unit (GRU) for aquaculture WQP. CNN can effectively fetch the aquaculture water quality characteristics, whereas GRU and LSTM can learn long-term dependencies in the time series data. We conduct experiments using the two different water quality datasets and present an extensive study on the impact of hyperparameters on the performance of the proposed hybrid DL models. Furthermore, the performance of hybrid CNN-LSTM and CNN-GRU models are compared with different baseline LSTM, GRU and CNN DL models and also with attention-based LSTM and attention-based GRU DL models. The results show that the hybrid CNN-LSTM outperformed all other models in terms of prediction accuracy and computation time.

Smartphones have proven to be a transformative tool that helps users perform various tasks such as online banking, chatting, sending an email or SMS, and online shopping. However, with the growing number of available applications and people downloading new applications at a high rate, managing the performance of such a large number of applications will increasingly become a concern, which makes managing smartphones’ screens and folders complicated which led the user to spend time finding those application to perform actions such as chatting or sending a message or even finding his favorite game at evening time may take few seconds to reach. By letting the smartphone learns the user’s behavior and their interactions, to predict which app the user looking for at a specific time after a certain sequence of actions. This saves users time and increases the level of usability. This paper investigates to what extent the usage of those applications can be predicted. The proposed methodology utilizes a deep learning algorithm (long short-term memory) to accurately predict the probability of a given application to be used by the smartphone user after a sequence of applications usage. The experimental result shows that forecasting those applications’ usage performance can be correct with an achieved accuracy of approximately 80%.

The rapid development of mobile technologies has made social media a vital platform for people to express their feelings and opinions. Understanding the public opinions can be beneficial for business and political entities in making strategic decisions. In light of this, sentiment analysis plays an important role to understand the polarity of the public opinions. This paper presents an ensemble hybrid deep learning model for sentiment analysis. The proposed ensemble model comprises three hybrid deep learning models which are the combination of Robustly optimized Bidirectional Encoder Representations from Transformers approach (RoBERTa), Long Short-Term Memory (LSTM), Bidirectional Long Short-Term Memory (BiLSTM) and Gated Recurrent Unit (GRU). In the hybrid deep learning model, RoBERTa is responsible for projecting the textual input sequence into a representative embedding space. Thereafter, the LSTM, BiLSTM and GRU capture the long-range dependencies in the embedding given the class. The predictions by the hybrid deep learning model are then amalgamated by averaging ensemble and majority voting, further improving the overall performance in sentiment analysis. In addition to that, the data augmentation with GloVe pre-trained word embedding has also been applied to alleviate the imbalanced dataset problems. The experimental results show that the proposed ensemble hybrid deep learning model outshines the state-of-the-art methods with the accuracy of 94.9%, 91.77%, and 89.81% on IMDb, Twitter US Airline Sentiment dataset and Sentiment140 dataset, respectively.

In the current era, data is growing exponentially due to advancements in smart devices. Data scientists apply a variety of learning-based techniques to identify underlying patterns in the medical data to address various health-related issues. In this context, automated disease detection has now become a central concern in medical science. Such approaches can reduce the mortality rate through accurate and timely diagnosis. COVID-19 is a modern virus that has spread all over the world and is affecting millions of people. Many countries are facing a shortage of testing kits, vaccines, and other resources due to significant and rapid growth in cases. In order to accelerate the testing process, scientists around the world have sought to create novel methods for the detection of the virus. In this paper, we propose a hybrid deep learning model based on a convolutional neural network (CNN) and gated recurrent unit (GRU) to detect the viral disease from chest X-rays (CXRs). In the proposed model, a CNN is used to extract features, and a GRU is used as a classifier. The model has been trained on 424 CXR images with 3 classes (COVID-19, Pneumonia, and Normal). The proposed model achieves encouraging results of 0.96, 0.96, and 0.95 in terms of precision, recall, and f1-score, respectively. These findings indicate how deep learning can significantly contribute to the early detection of COVID-19 in patients through the analysis of X-ray scans. Such indications can pave the way to mitigate the impact of the disease. We believe that this model can be an effective tool for medical practitioners for early diagnosis.

Accurate prediction the remaining useful life (RUL) and estimation the state of health (SOH) are critical to the management of lithium-ion batteries. In this paper, a lithium battery capacity prediction method based on cuckoo search optimization variational mode decomposition (CS-VMD) and gated recurrent unit (GRU) is proposed. Firstly, the VMD algorithm is used to divide the capacity into some intrinsic mode functions (IMFs) to reduce the impact of capacity regeneration and other situations. The number of decomposition layers and the quadratic penalty factor of VMD are optimized by the CS algorithm. Then, the GRU network is introduced to capture small changes in the capacity degradation process and perform the capacity prediction of decomposed sequence. Finally, some prediction results are integrated effectively. Based on two publicly available lithium-ion battery datasets, the model proposed in this paper can significantly reduce the complexity of the sequence and have high prediction accuracy, which is better than other prediction models. The root mean square error (RMSE) is controlled within 2%, and the maximum mean absolute error (MAE) does not exceed 2%.

A new method of road roughness level identification based on the bidirectional gated recurrent unit (BiGRU) network is proposed in this paper, which is contribute to solve the problems of intelligent chassis technology such as suspension control. Firstly, the vehicle vibration response data is attained by the ride comfort simulation of two-degree-of-freedom vehicle vibration model. Then the mapping relationship between the road roughness level and the vehicle vibration responses is determined, and the road roughness level identification model is established. The Adam algorithm and mini-batch gradient descent are utilized to improve the accuracy and increase the speed of the model training process. Finally, in order to verify the feasibility and practicability of the model, the ride comfort experiments are carried out on asphalt and brick roads. The results show that the accuracy of the road roughness level identification model based on the BiGRU network reaches 95.83%, and the recognition result is reliable. Moreover, the experimental road level can be successfully identified through the road roughness level identification model, which has high engineering application value.

State-of-charge (SOC) is a crucial battery quantity that needs constant monitoring to ensure cell longevity and safe operation. However, SOC is not an observable quantity and cannot be practically measured outside of laboratory environments. Hence, machine learning (ML) has been employed to map correlated observable signals such as voltage, current, and temperature to SOC values. In recent studies, deep learning (DL) has been a prominent ML approach outperforming many existing methods for SOC estimation. However, yielding optimal performance from DL models relies heavily on an appropriate selection of hyperparameters. At present, researchers relied on established heuristics to select hyperparameters through manual tuning or exhaustive search methods such as grid search and random search. This results in lengthy development time in addition to less accurate and inefficient models. This study proposes a systematic and automated approach to hyperparameter selection with a Bayesian optimization strategy known as tree Parzen estimator (TPE) in combination with the Hyperband pruning algorithm. The TPE optimization is run on various DL models such as BGRU, GRU, LSTM, CNN, FCN, and DNN to optimize for the best hyperparameter combination for each architecture. The Hyperband pruning algorithm is used to prune unpromising trials during the TPE run which results in time and computational resource savings. Experiment results show that the best performing model, BGRU-TPE incurs a low computation cost with no compromise in model accuracy, maintaining the best balance of both. The proposed model BGRU-TPE achieves the lowest RMSE and MAE error at 0.8% and 0.56%, respectively on various electric vehicle drive cycles at varying ambient temperatures and maintains a low computational cost at 15 600 FLOPS with the model size of only 193.3 KB. [ABSTRACT FROM AUTHOR]

With the increasing demand for intelligent transportation systems, short-term traffic flow prediction has become an important research direction. The memory unit of a Long Short-Term Memory (LSTM) neural network can store data characteristics over a certain period of time, hence the suitability of this network for time series processing. This paper uses an improved Gate Recurrent Unit (GRU) neural network to study the time series of traffic parameter flows. The LSTM short-term traffic flow prediction based on the flow series is first investigated, and then the GRU model is introduced. The GRU can be regarded as a simplified LSTM. After extracting the spatial and temporal characteristics of the flow matrix, an improved GRU with a bidirectional positive and negative feedback called the Bi-GRU prediction model is used to complete the short-term traffic flow prediction and study its characteristics. The Rectified Adaptive (RAdam) model is adopted to improve the shortcomings of the common optimizer. The cosine learning rate attenuation is also used for the model to avoid converging to the local optimal solution and for the appropriate convergence speed to be controlled. Furthermore, the scientific and reliable model learning rate is set together with the adaptive learning rate in RAdam. In this manner, the accuracy of network prediction can be further improved. Finally, an experiment of the Bi-GRU model is conducted. The comprehensive Bi-GRU prediction results demonstrate the effectiveness of the proposed method. [ABSTRACT FROM AUTHOR]

Automatic fingerprint identification system (AFIS) uses fingerprint to authenticate users, which is legal if the user is enrolled. However, numerous studies reveal that it is susceptible to spoofing attacks where a third person might freely synthesize counterfeit fingerprints to trick the scanner. To resist spoofing attacks, it makes fingerprint liveness detection (FLD) highly desirable. Most of previous work was to directly input the whole fingerprints into convolutional neural network, making it impossible to fully explore the relationship of spatial ridges, especially those with the latent fine-grained minutia on fingerprint ridges. Accordingly, in this paper, we exploit the relationship of spatial ridges in fingerprints and propose a novel FLD method based on spatial ridges continuity (FLD-SRC). Several fingerprint patches are first selected utilizing ridge texture saturation, and then uniformly split into several slices and thus construct the spatial continuity between pixels and between slices. Next, the proposed FLD-SRC learns deep features from fingerprints and eliminates redundant information. After that, the extracted feature maps are treated as a sequence and analyzed the intra-continuity by cascade gated recurrent unit (GRU). A discriminant slice grouping subnetwork is then developed to model the correlation between ridges slices and implicitly discover the discriminant inter-continuity. Pruning strategy is further utilized to reduce network parameters and promote its practical application in real scenarios. Experimental results, evaluated on three publicly available datasets, show the competitiveness of our method. Furthermore, in addition to reducing computational complexity, our method also shows the best ACE performance in cross-material and cross-sensor cases. [ABSTRACT FROM AUTHOR]

Aspect-based sentiment analysis (ABSA) is crucial for exploring user feedbacks and preferences on produces or services. Although numerous classical deep learning-based methods have been proposed in previous literature, several useful cues (e.g., contextual, lexical, and syntactic) are still not fully considered and utilized. In this study, a new approach for ABSA is proposed through the guidance of contextual, lexical, and syntactic cues. First, a novel sub-network is introduced to represent a target in a sentence in ABSA by considering the whole context. Second, lexicon embedding is applied to incorporate additional lexical cues. Third, a new attention module, namely, dependency attention, is proposed to elaborate syntactic dependency cues between words in attention inference. Experimental results on four benchmark data sets demonstrate the effectiveness of our proposed approach to aspect-based sentiment analysis. [ABSTRACT FROM AUTHOR]

Forecasted global horizontal irradiation (GHI) can help for designing, sizing and performances analysis of photovoltaic (PV) systems including water PV pumping systems used for irrigation applications. In this paper, various deep neural networks (DNN) models for one day-ahead prediction of GHI at Hail city (Saudi Arabia) are developed and investigated. The considered DNN models include long-shortterm memory (LSTM), bidirectional-LSTM (BiLSTM), gated recurrent unit (GRU), bidirectional-GRU (Bi-GRU), one-dimensional convolutional neural network (CNN1D) and other hybrid configurations such as CNN-LSTM and CNN-BiLSTM. A dataset of daily GHI recordings collected during January 1, 2000 to June 30, 2020 from National Aeronautics and Space Administration (NASA) at an arid location (Hail, Saudi Arabia) is used to develop and compare the above DNN-based models. The parameters affecting the accuracy of the models have been also deeply analyzed. Only historical values of daily GHI have been used to build the DNN-based models whereas additional weather parameters such as air temperature, wind speed, wind direction, atmospheric pressure and relative humidity are not considered in this work. Keras library and Python language have been used to develop and compare the GHI forecasting models. The evaluation metrics such as correlation coefficient (r), Mean Absolute Percent Error (MAPE), Mean Absolute Error (MAE), cumulative distribution function (CDF) and standard deviation (σ) are opted to evaluate the performance of the prediction models. The obtained results showed that the DNN models have provided globally good performances with a maximum reached value of r = 96%, for daily GHI forecasting.

Spectrum occupancy prediction is a key enabling technology to facilitate a proactive resource allocation for dynamic spectrum management systems. This work focuses on the prediction of duty cycle (DC) metric that reflects spectrum usage (in the time domain). The spectrum usage is typically measured on a shorter time scale than needed for prediction. Hence, data thinning is required and we apply block averaging. However, averaging operation results in flattening the DC data and losing essential features to assist deep neural network (DNN) to predict the spectrum usage. To improve DC prediction after block averaging, a feature-based deep learning framework is proposed. Namely, long short-term memory (LSTM) and gated recurrent unit (GRU) are selected and enhanced by using features of the data, such as the variance of DC data in addition to DC data themself. The proposed model is capable of proactively predicting the spectrum usage by capturing complex relationships among various input features for the measured spectrum, thus providing higher prediction accuracy with an average improvement of 5% in RMSE compared with traditional models. Moreover, to have a better understanding of the proposed model, we quantify the effect of input features on the predicted spectrum usage values. Based on the most significant input features, a simpler and more efficient model is proposed to estimate DC with similar accuracy to when using all features.

It is difficult for existing deep learning-based satellite on-orbit anomaly detection methods to define the residual-based detection threshold and identify false anomalies. To solve the above problems, this paper proposes both a detection threshold determination and dynamic correction method and a causality-based false anomaly identification and pruning method. We use the GRU (Gated Recurrent Unit) to model and predict the telemetry parameters to obtain the residual vector; determine and dynamically correct the threshold according to the prescribed false positive rate; propose an improved multivariate transfer entropy method to identify the causal relationships between the telemetry parameters; and, based on the causality, determine whether the detected parameter anomalies are false. Experiments show that the precision, recall, and F1-score of the method proposed in this paper are superior to the current typical method, and the false positive rate is significantly reduced, demonstrating the effectiveness of the proposed method.

Conventional human-centric image compression techniques are optimized for human visual perception, and are generally evaluated by metrics such as MSSSIM and PSNR. On the other hand, task-centric image compression techniques that target deep neural networks (DNN) inference focus on understanding images, and are measured by inference accuracy. As images should be encoded and decoded differently depending on the target metric, existing deep learning based compression techniques have focused on designing an independent neural network that is optimized for either one of the two targets. However, as these techniques require the target metric to be determined during a training phase, separate DNN models should be trained for different target applications, allowing no scalability between human- and task-centric compression. In this paper, we propose a target-dependent scalable image compression pipeline, where compressed images can be differently decoded in real time depending on the interpolated target between perceptual quality and DNN inference accuracy. This dynamic scalability is supported by a reconfigurable recurrent neural network (RNN) that can dynamically change its flow according to the interpolation parameter. The experimental results show that the proposed method achieves comparable MSSSIM and classification accuracy to the compression networks individually trained for human-centric and task-centric targets, respectively, while providing scalability between the two targets depending on the interpolation parameter.

As the leading component of the financial market, the price formation mechanism of gold futures has been attracting extra attention of financiers and scholars. However, the data of gold futures price belongs to time series, and its forecast is very challenging owing to its chaotic, noisy, and non-stationary characteristics in data. A new forecast model named SGRU-AM, based on the special gated recurrent unit (SGRU) and attention mechanism (AM) is proposed in this paper to tackle these challenges. SGRU is the rectified model of gated recurrent unit (GRU) though executing the 1-tanh function on the reset gate output of the basic GRU to transform the value range of the reset gate value and adjusting the memory ratio between the current moment and the previous moment. Firstly, SGRU has the advantage of capturing long-distance information and can forecast the closing price of gold futures in the next trading day. Then, AM is introduced to adjust the SGRU time dimension’s feature expression, so that the model can obtain more comprehensive feature information, learn the importance of current local sequence features and improve the forecast accuracy. Taking China’s gold futures as an example, the gold futures data of the Shanghai Futures Exchange are selected from January 9, 2008, to May 31, 2021. Compared with the baseline methods, the experimental results show that SGRU-AM has the best performance among all baseline models in forecast efficiency and forecast accuracy.

High-resolution temperature forecasting can often prove to be challenging for conventional machine learning models as temperature is highly seasonal and varies with the time of the year as well as with passing hours of the day. In most cases, only the daily extremes or mean temperatures are provided by temperature forecasting methods. However, with the growing availability of data and the development of deep neural networks (DNNs) capable of detecting complex relationships, high-resolution temperature forecasting is becoming easier. Typically, historical temperature data along with multiple meteorological sensor data is used for temperature forecasting which increases the complexity of the system making it harder and costlier to implement physically. In this paper, high-resolution hourly temperature forecasting is performed using only historical temperature data. The paper presents a comparative analysis among four popular DNNs-simple recurrent neural network (SRN), gated recurrent unit (GRU), long-short term memory (LSTM), convolutional neural network (CNN), and two hybrid models-CNN-LSTM parallel network and GRU-LSTM parallel network trained on Beijing temperature dataset. Experimental results showed GRU-LSTM parallel network obtained the lowest RMSE (1.691° C) whereas CNN has the best computational efficiency obtaining a slightly worse RMSE (1.759° C). Additionally, a robustness analysis is performed on temperature data from four additional geographically diverse locations (Toronto, Las Vegas, Seattle, and Dallas) which reveals GRU to be the most consistent algorithm. Finally, the paper establishes a correlation between the model performance and the dataset based on their variance and mean absolute deviation with reference to the training dataset.

Spacecraft anomaly detection which could find anomalies in the telemetry or test data in advance and avoid the occurrence of catastrophic failures after taking corresponding measures has elicited the attention of researchers both in academia and aerospace industry. Current spacecraft anomaly detection systems require costly knowledge and human expertise to identify a true anomaly. Moreover, some new problems and challenges such as large volume of test data, imbalanced data distribution and the scarcity of faulty labeled samples have emerged. In this work, we propose an unsupervised anomaly detection algorithm combining Gated Recurrent Unit (GRU) based Recurrent Neural Network (RNN) and Extreme Value Theory (EVT). First, we develop a two-layer ensemble learning based predictor framework which stacks three GRU-based networks with different architectures to learn and capture the normal behavior of multiple channels of data. Then, the prediction errors are calculated and smoothed using Exponentially Weighted Moving Average (EWMA) algorithm. Next, we propose a detection rule setting anomaly threshold automatically through EVT which does not assume any parent distribution on the prediction errors. To the best of our knowledge, it is the first attempt that stacked GRU-based predictors with EVT has been employed into the spacecraft anomaly detection. Through extensive experiments conducted on public datasets as well as real data sampled from a launch vehicle, we show that the proposed detection algorithm is superior to other state-of-the-art anomaly detection approaches in terms of model performance and robustness.

This paper proposes a multi-data driven hybrid learning method for weekly photovoltaic (PV) power scenario forecast that is coordinately driven by weather forecasts and historical PV power output data. Patterns of historical data and weather forecast information are simultaneously captured to ensure the quality of the generated scenarios. By combining bicubic interpolation and bidirectional long-short term memory (BiLSTM), a super resolution algorithm is first presented to enhance the time resolution of weather forecast data from three hours to one hour and increase the precision of weather forecasting. A weather process-based weekly PV power classification strategy is proposed to capture the coupling relationships between meteorological elements, continuous weather changes and weekly PV power. A gated recurrent unit (GRU)-convolutional neural network (CNN)-based scenario forecast method is developed to generate weekly PV power scenarios. Evaluation indices are presented to comprehensively assess the quality of the generated weekly scenarios of PV power. Finally, the PV power, weather observation and weather forecast data collected from five PV plants located in Northeast Asia are used to verify the effectiveness and correctness of the proposed method. [ABSTRACT FROM AUTHOR]

Smart manufacturing, which is increasingly popular worldwide, is aided by time-series forecasting. As the volume of historical data increases, powerful forecasting techniques that reveal unknown relationships between past and future values are required to provide accurate forecasts of production and sales. Thus, in this article, a composite gate recurrent unit (GRU)–Prophet model with an attention mechanism was constructed to predict sales volume. In this composite model, Prophet model and GRU model with attention mechanism were used to capture linear and nonlinear features, respectively. The composite model was experimentally determined to be more applicable and to provide more accurate predictions than did recurrent neural network, long short-term memory, gate recurrent unit, Prophet, and autoregressive integrated moving average models. This article's composite model is suited to rapid changes in market demand and helps enterprises be more competitive in the field of smart manufacturing. [ABSTRACT FROM AUTHOR]

Wireless traffic usage forecasting methods can help to facilitate proactive resource allocation solutions in cloud managed wireless networks. In this paper, we present temporal and spatial analysis of network traffic using real traffic data of an enterprise network comprising 470 access points (APs). We classify and separate APs into different groups according to their traffic usage patterns. We study various statistical properties of traffic data, such as auto-correlations and cross-correlations within and across different groups of APs. Our analysis shows that the group of APs with high traffic utilization have strong seasonality patterns. However, there are also APs with no such seasonal patterns. We also study the relation between number of connected users and traffic generated, and show that more connected users do not always mean more traffic data, and vice versa. We use Holt-Winters, seasonal auto-regressive integrated moving average (SARIMA), long short-term memory (LSTM), gated recurrent unit (GRU) and convolutional neural network (CNN) methods for forecasting traffic usage. Our results show that there is no single universal best method that can forecast traffic usage of every AP in an enterprise wireless network. The combined models such as CNN-LSTM and CNN-GRU are also used for spatio-temporal forecasting of a single AP traffic usage. The results show that considering spatial dependencies of neighboring APs can improve the forecasting performance of a single AP if it has significant spatial correlations.

Channel state information (CSI), which enables wireless systems to adapt their transmission parameters to instantaneous channel conditions and consequently achieve great performance boost, plays an increasingly vital role in mobile communications. However, getting accurate CSI is challenging due mainly to rapid channel variation caused by multi-path fading. The inaccuracy of CSI imposes a severe impact on the performance of a wide range of adaptive wireless systems, highlighting the significance of channel prediction that can combat outdated CSI effectively. The aim of this article is to shed light on the state of the art in this field and then go beyond by proposing a novel predictor that leverages the strong time-series prediction capability of deep recurrent neural networks incorporating long short-term memory or gated recurrent unit. In addition to an analytical comparison of computational complexity, performance evaluation in terms of prediction accuracy is carried out upon multi-antenna fading channels. Numerical results reveal that deep learning brings a notable performance gain compared with the conventional predictors built on shallow recurrent neural networks.

Massive amount of water level data has been collected by using Internet of Things (IoT) techniques in the Yangtze River and other rivers. In this paper, utilizing these data to construct deep neural network models for water level prediction is focused. To achieve higher accuracy, both the factors of time and locations of data collection sensors are considered to perform prediction. And the network structures of gated recurrent unit (GRU) and convolutional neural network (CNN) are combined to build a CNN-GRU model in which the GRU part learns the changing trend of water level, and the CNN part learns the spatial correlation among water level data observed from adjacent water stations. The CNN-GRU model that using data from multiple locations to predict the water level of the middle location has higher accuracy than the model only based on GRU and other state-of-the-art methods including autoregressive integrated moving average model (ARIMA), wavelet-based artificial neural network (WANN) and long-short term memory model (LSTM), because of its ability to decrease the affections of abnormal value and data randomness of a single water station to some extent. The results are verified on an experiment dataset that including 30-year observed data of water level at several collection stations in the Yangtze River. For forecasting the 8-o'clock water levels of future 5 days, accuracy of the CNN-GRU model is better than that of ARIMA, WANN and LSTM models with three evaluation factors including Nash-Sutcliffe efficiency coefficient (NSE), average relative error (MRE) and root mean square error (RMSE).