Forecasting road network information based on spatio-temporal features has emerged as a key research direction in intelligent transportation. Despite numerous studies examining the intrinsic connection between complex spatial structures and temporal variations, achieving substantial improvements in forecasting remains challenging, particularly for long-term forecasting and handling abrupt changes. Extant research extensively leverages predefined graph structures or adaptive graphs to represent spatial relationships, but often overlooks the spatial dependence in node temporal information and the impact of road outliers. In response, this paper proposes a multiple spatio-temporal fusion network model (McFasten). Firstly, the spatio-temporal characteristics of the road network are analyzed using a multi-scale attention mechanism. Next, temporal information is decomposed, and anomalies are processed using Wavelet Convolution, and the spatial relationships inherent in the temporal sequences are captured by an improved Convolutional Block Attention Module (CBAM). Furthermore, the adjacency matrix is dynamically adjusted via a parameter matrix, enhancing the model’s capability for learning long-term implicit relationships. Finally, Efficient Multi-Scale Attention (EMA) is integrated into the output forecasting, achieving cross-dimensional feature integration. On real-world road network datasets, our proposed model outperforms the state-of-the-art baseline by up to 16

Abstract Photovoltaic (PV) power forecasting plays a vital role in ensuring the secure and stable operation of power systems. To address the strong nonlinearity and complex local and global temporal correlations of PV power caused by environmental factors, this study proposes an ultra-short-term PV power forecasting framework based on CEEMDAN-VMD and BiTGRU-ACGN. First, complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) is employed to perform the initial decomposition of the original PV sequence. Subsequently, sample entropy (SE) and K-means clustering are combined to classify the decomposed components into different frequency-domain subsets. To further eliminate noise interference and enhance multi-scale feature extraction capability, variational mode decomposition (VMD) is introduced to conduct a secondary refined decomposition on the high-frequency subsequences.Based on the reconstructed component sequences, a parallel dual-branch forecasting network is constructed. The first branch integrates bidirectional dilated causal convolution with a bidirectional gated recurrent unit (BiTGRU) to collaboratively capture local dynamic features. The second branch adopts an attention-based convolutional gating network (ACGN), which utilizes positional encoding, a multi-head self-attention mechanism, and a gated convolutional feed-forward neural network to model global temporal dependencies. Through the collaborative fusion of dual-branch features, the proposed framework simultaneously strengthens the modeling capability for both local details and global patterns. Experiments are conducted using datasets collected from a photovoltaic power station in Northwest China and the Alice Springs PV dataset in Australia. The results demonstrate that the proposed model achieves satisfactory forecasting performance across different seasons. In particular, during summer, when random fluctuations are most severe, R² increases by 1.97%–26.49%, while MAE and RMSE decrease by 43.23%–51.16% and 47.15%–81.42%, respectively. These experimental results verify that the proposed method possesses excellent forecasting capability and strong generalization performance, thereby providing effective support for optimal smart-grid scheduling and practical operation and maintenance tasks.

This paper studies the leader-follower consensus problem of multi-agent systems (MASs) with matched disturbances under Markov-switching communication topology, where a distributed adaptive integral sliding mode (AISM) control strategy is proposed. By introducing an integral sliding surface, the strategy effectively mitigates the impact of matched disturbances on system dynamics. Meanwhile, an adaptive sliding mode gain is designed to compensate for disturbances, avoiding the excessive conservatism or high control energy consumption associated with fixed-gain sliding mode control. In addition, to address the random switching effects introduced by Markov-switching topology, a Lyapunov function is constructed based on infinitesimal generator theory, proving the stability of the MAS. Then, the matrix equation and other conditions are used to ensure that MAS can achieve mean square consensus. Finally, the effectiveness and robustness of the designed control strategy are verified through the numerical simulations.

Lag formation control is a fundamental cooperative trajectory-following mechanism for multi connected autonomous vehicle systems (MCAVSs), where the follower vehicles intentionally behind that of a designated leader while maintaining a stable platoon formation despite tracking delays. This study addresses the lag formation problem in MCAVSs through a distributed event-triggered control scheme applicable to both directed and undirected communication works. This event-based communication framework prevents continuous information exchange among follower vehicles, thereby reducing network load and computational demand. on Lyapunov stability analysis, a set of sufficient conditions is established to guarantee the MCAVSs achieve the desired lag formation under the proposed protocol. Notably, behavior is rigorously excluded by demonstrating that successive triggering intervals have positive lower bound, ensuring physically realizable control updates. Finally, two numerical simulation studies verify the effectiveness, communication efficiency, and delay robustness the proposed control method. This study presents a novel integration of lag formation distributed event-triggering mechanisms tailored for real-world MCAVS scenarios in contrast existing event-triggered and lag consensus approaches, enabling both theoretical stability practical implementability.

Precise photovoltaic (PV) power forecasting facilitates optimal power grid scheduling, managing energy, and maintaining grid stability. To achieve precise PV forecasting, this paper proposed a spatio-temporal feature fusion ultra-short-term PV power forecasting approach based on ADARO-PTSFENet. First, the spatial feature extraction module is constructed by integrating the depth perception module and global perception module proposed in this paper, which are connected in parallel with bidirectional long short-term memory (BiLSTM) to form a parallel temporal-spatial feature extraction network (PTSFENet) model. Subsequently, to address the limitations of the artificial rabbit optimization (ARO), a new adaptive weighting factor was proposed to improve its search process. Then, to improve population diversity, the Tent chaos mapping technique was employed, and the adaptive dynamic artificial rabbit optimization (ADARO) was proposed. In comparison with other benchmark algorithms, ADARO demonstrates superior performance. Finally, ADARO was integrated to optimize the hyperparameters of the PTSFENet, thereby forming the novel ADARO-PTSFENet PV power forecasting approach. The experimental results show that, compared with the other seven baseline models, the ADARO-PTSFENet model has excellent predictive performance while having a lower number of parameters. The root mean square error (RMSE) on the first dataset is reduced by 19.18% to 43.05%. The mean absolute error (MAE) decreased by 23.82% to 47.57%, the RMSE on the second dataset decreased by 5.52% to 36.01%, and the MAE decreased by 17.17% to 47.65%. In the one-year datasets, compared with other comparison models, the time generalization performance and predictive ability of the proposed model are the best, demonstrating that the model proposed in this paper has excellent predictive ability.

Traffic congestion is getting increasingly severe, making accurate traffic flow prediction crucial for relieving traffic pressure. Existing methods have problems in extracting multidimensional features from time series, particularly in capturing long-range temporal dependencies, effectively mining frequency-domain characteristics (e.g., periodic patterns), and modeling complex inter-correlations among multivariate data, all of which limit prediction accuracy. Therefore, this study proposes an integrated model called DSTIT-TCN to address this issue. Firstly, efficient multi-dimensional feature extraction of time series is accomplished through the integration of Discrete Cosine Transform (DCT), Squeeze-and-Excitation Networks (SENet), and Temporal Convolutional Network (TCN). Furthermore, the optimized features are subsequently fed into an iTransformer network to capture multivariate inter-correlations. Finally, the parallel connection with TCN enhances temporal dependency modeling capabilities and synergistically improves the prediction performance. To validate the proposed model’s efficacy, we used the UK motorway dataset for evaluation. Experimental results demonstrate substantial performance gains, with our model achieving improvements across all key metrics: R² increased by 0.0101

With the increasing duration of operation, factors such as process parameters, operating conditions, and environmental variables in industrial processes gradually change, leading to variations in the correlations among process variables. The complexity of process data, characterized by high-dimensional redundancy, nonlinear correlations, noise interference, and slow-varying fault characteristics, poses significant challenges for unsupervised fault detection. Therefore, this article proposes an industrial process monitoring method based on fusion of multi-scale sparse autoencoder and dual-branch slow feature architecture (MSAE-DSFA). In the first stage, the original data is mapped to different dimensions through a multi-scale whitening-sparse encoder, capturing features from local details to global structures. This approach addresses the limitations of single-scale representation capabilities and eliminates issues related to process feature redundancy and decorrelation. Simultaneously, in the second stage, a primary-secondary slow feature extraction method is designed to capture the temporal stability characteristics of latent features from the previous stage, further mitigating environmental noise and other factors while extracting slowly changing latent representations to ensure sensitivity to progressive anomalies. Finally, based on the latent space and reconstruction error, statistical measures are calculated. The features extracted by the MSAE-DSFA method contain crucial information about the operational process, with reconstructed data demonstrating closer alignment to actual industrial processes. Experimental validation is conducted on two industrial cases.

Multi-agent reinforcement learning (MARL) frameworks offer significant potential for vehicle-road collaborative systems, promising substantial reductions in travel times and enhanced network throughput. While existing MARL approaches excel at multi-intersection signal control, they often lack guarantees regarding the optimality of their joint learning policies. This limitation hinders both their scalability and effectiveness in large-scale network-level signal control. Key challenges involve escalating system complexity, diminishing learning efficiency, heightened difficulty in parameter optimization, and the persistent scalability-performance trade-off. To address these issues, we propose a hierarchically structured MARL model with spatially partitioned zones. This framework first decomposes the traffic network, then implements a tiered agent architecture featuring coordinated agent interactions within each level, thus enhancing its potential for scalable traffic signal control. Simulations on 6 6 and Manhattan networks demonstrate that the proposed method significantly outperforms single-layer baselines. Notably, under near-saturated conditions in the large-scale Manhattan scenario, it effectively mitigates the performance degradation typical of scale expansion, exhibiting robust scalability. Furthermore, the introduced “soft relative priority” mechanism secures transit advantages for emergency vehicles (EMVs) while minimizing disruptions to general traffic, thereby balancing emergency priority with overall network efficiency.

In intelligent transport systems (ITS), accurate and real-time traffic flow prediction is essential to reduce congestion and manage transport. Numerous existing studies have used hybrid models to extract various intrinsic patterns of traffic flow data, but their noise is not treated when extracting spatial features. Different from the existing studies, this document proposes a new hybrid model called "MFMA-Informer (Multiscale Fusion and Multiscale Attention-Informer)," which aims to reduce noise interference while deeply mining the intrinsic model of spatiotemporal information. The model consists of two parts: feature extraction and feature fusion. First, in the spatial feature extractor model, the MFMA model is proposed, which uses a multiscale temporal convolutional network (TCN) and long short-term memory (LSTM) to obtain multiscale spatial features and reduces noise by computing attention at different scales. Second, in the temporal properties model, the longtime dependency of traffic flow data has been fully achieved because the Informer model has the advantage of encoding temporal data. In addition, in the feature fusion module, interdependencies are constructed on the extracted spatiotemporal features through cross-attention to further enhance the consistency of spatiotemporal feature expression. Finally, to prove the effectiveness of the framework, numerous trials were performed on four different real-world datasets, which showed that MFMA-Informer had excellent forecasting and generalization capabilities.

The accelerating urbanization has made traffic congestion a major bottleneck in urban mobility. Traditional traffic signal control methods often prove inadequate in complex traffic environments. Reinforcement Learning (RL)-based Traffic Signal Control (TSC) methods have recently gained prominence for their adaptive decision-making capabilities in such dynamic settings. As the environmental state is crucial for control performance, most existing studies heuristically design state features using metrics like queue lengths, often resulting in sensitive and unstable models. Addressing this limitation, this paper proposes a Bidirectional Pressure (BDP)-enhanced state representation method, explicitly incorporating traffic pressure from both upstream and downstream lanes into state modeling. Integrated with state-of-the-art RL frameworks, our method demonstrates at least 16.15% improvement in average travel time across synthetic and real-world datasets compared to existing approaches.

In high-dynamic scenarios, moving objects corrupt feature matches in Visual-Inertial Odometry (VIO), invalidating epipolar constraints and causing state estimation divergence. When dynamic features dominate tracked points, standard RANSAC fits dynamic rather than camera ego-motion. This paper proposes a robust VIO framework for CPU-only platforms combining a Dual-Screening strategy with an Event-Triggered Mechanism. A Depth-Motion Geometric Consistency model removes dynamic candidates from the RANSAC input before they reach the backend. The Event-Triggered Mechanism monitors scene dynamics via the optical flow tracking rate and activates depth verification only when instability is detected, bypassing overhead during stable operation. On the OpenLORIS-Scene dataset, the proposed method reduces Absolute Trajectory Error (ATE) by 36.8% relative to VINS-Fusion in pedestrian-dense sequences at $\mathbf{4 9 ~ H z}$ on a standard CPU.

This paper investigates the leader-follower consensus problem for high-order multi-input-multi-output (MIMO) linear multi-agent systems (MASs) subject to unknown matched disturbances. First, a continuous-time adaptive integral sliding mode (AISM) control strategy is proposed under a directed topology to enhance robustness. To alleviate the communication burden while maintaining stability, an event-triggered adaptive integral sliding mode (ET-AISM) control strategy is further developed. Distinguishing itself from existing literature, this work incorporates a dynamic event-triggered mechanism (DETM) featuring internal dynamic variables. Based on Lyapunov stability theory, the robust stability of the closed-loop system is established, and the existence of a practical sliding motion (PSM) is verified. Moreover, the exclusion of the Zeno phenomenon is rigorously proven. Finally, the effectiveness of the proposed strategy is verified through the simulations on a theory system and damped mass-spring system.

This paper investigates the distributed recursive filtering problem for two-dimensional (2-D) systems over sensor networks with an adaptive event-triggered scheme (ETS) and a binary encoding scheme. To alleviate the computational and communication burden, an adaptive ETS with bidirectional evolutionary properties is proposed. In addition, a binary encoding model accounting for two-dimensional time instants is developed to improve data security and facilitate the transmission of data. Within this framework, a distributed recursive filter structure that relies on the topology of the sensor networks is presented, incorporating the adaptive ETS and binary encoding scenarios. Subsequently, the 2-D mathematical induction approach is employed to derive the upper bound of the filtering error covariance (FEC), and the filter gains with satisfactory performance are obtained by minimizing the trace of the upper bound. Moreover, the boundedness criterion of FEC is developed, and the effect of triggering condition on the FEC is analyzed. Finally, the validity of the proposed distributed recursive filtering algorithm is demonstrated through the case of the heat-exchange process.

In response to the current severe traffic congestion issues, highly reliable traffic flow prediction serves as a fundamental prerequisite for optimizing municipal road networks and mitigating systemic vehicular congestion. Aiming to elevate the precision of short-term traffic flow prediction, this paper first addresses the low precision of the Dung Beetle Optimizer (DBO) algorithm by introducing an exponential adaptive weight in the way of position update for the ball-rolling dung beetle, along with incorporating a Cauchy-Gaussian mutation strategy. We propose the Multi-strategy improved Dung Beetle Optimizer (MDBO), which is validated using eight benchmark test functions, demonstrating that MDBO outperforms common optimization algorithms in solution accuracy. Secondly, we adopt a combined prediction model, Traffic Flow Temporal-Spatio Network (TFTSNet), which constructs spatial feature modules and temporal feature modules in parallel fusion. Finally, we achieve short-term traffic flow prediction by optimizing the TFTSNet combined prediction model using MDBO. The experiment evaluated model performance using publicly available traffic flow datasets. The results demonstrate that, compared to other state-of-the-art models, the proposed joint prediction model based on MDBO-optimized TFTSNet achieves substantial enhancements in both prediction precision and generalization capability. Root mean square error (RMSE) decreased by 8.7-35.7%, mean absolute error (MAE) decreased by 6.6-40.0%, and R2 reached 0.975, showcasing robust predictive capabilities and engineering reference value.

This paper investigates the problem of priority access signal control for emergency vehicles (EMVs) in urban road networks. Based on the characteristics of traffic road network, a cooperative multi-agent reinforcement learning model leveraging the store-and-forward paradigm is proposed, as a way to improve emergency vehicle priority and alleviate urban traffic pressure. Initially, we establish cooperative relationships among agents at the network modeling level, modeling the traffic flow between pairs of adjacent agents as a co-optimization objective, which achieves parameter decoupling among algorithms. Subsequently, through regional partitioning, we design a spatiotemporal information extraction algorithm, termed store-and-forward graph attention mechanism for multi-agent proximal policy optimization (SF-GMPPO), to control information exchange among agents, thereby learning optimal traffic signal control policies. Finally, the proposed algorithm model is tested in both synthetic and real-world network environments. Experimental results demonstrate that the proposed algorithm surpasses similar algorithms in terms of training convergence and control performance. Specifically, in the 5x5 synthetic road network, the average speed of EMVs increased by 2.463 meters per second (m/s) compared to social vehicles. In the Lanzhou Chengguan road network with 128 intersections, the proposed algorithm achieved high traffic control efficiency, resulting in the speed of social vehicles approaching that of EMVs, with EMVs operating at an average speed of 13.061 m/s and social vehicles at 13.032 m/s under efficient traffic flow management. These results highlight the effectiveness of the proposed approach in enhancing EMV priority while maintaining overall traffic efficiency.

In this paper, a strategy for safe driving of autonomous vehicles in complex environments with obstacles is proposed, which combines a double-layer model predictive control (DLMPC) with an event-triggered mechanism. At first, a safe collision distance model is established in the upper path planning layer, utilizing a vehicle point mass model and based on the distance relationship between the vehicle and obstacles. The design of the planning layer improves the safety of the vehicle. After solving the optimization problem, the optimal input variable is derived and then transmitted to the dynamic control layer for accurate tracking. It is followed by designing the triggered condition according to the stability and feasibility of the system, which reduces the computation of the controller and improves the real-time performance of the system. Finally, the simulation results demonstrate the efficacy of the proposed control strategy in addressing the obstacle avoidance challenge faced by autonomous vehicles in complex road conditions. Compared with traditional model predictive control (MPC), this strategy improves the precision of tracking control and reduces the computational complexity.

This paper investigates the bounded filtering problem for two-dimensional (2-D) discrete systems with encryption-decryption mechanism (EDM) and dynamic event-triggered mechanism (ETM). Firstly, considering the potential information leakage, a novel EDM is designed for 2-D systems based on the quantization-based encoding-decoding mechanism. In addition, the dynamic ETM with bidirectional evolutionary characteristics is proposed to alleviate the computational and communication burdens. In such a framework, the filtering error systems (FESs) at the eavesdropper side and the user side are obtained, respectively. Subsequently, the encryption parameters are devised such that the filtering error at the eavesdropper side diverges. Additionally, the boundedness criteria are established to ensure that the client-side FESs are bounded in terms of Lyapunov stability analysis approach. Finally, it is demonstrated that the proposed bounded filtering algorithm is valid for 2-D systems in several types of industrial environments.

Cyber-physical systems (CPSs) play a vital role in modern industry due to their scalability, resource-sharing capabilities, and cost-effectiveness. However, their reliance on network communication makes them vulnerable to false data injection (FDI) attacks, which can significantly degrade system performance. This paper proposes a self-triggered nonlinear model predictive control (STMPC) strategy to mitigate the effects of FDI attacks. The proposed STMPC method optimizes a cost function while ensuring control input constraints and incorporates a sample-and-hold strategy to reduce communication overhead. A key feature of this approach is its ability to detect and counteract FDI attacks by reconstructing control signals using historical data, thereby enhancing system resilience. Simulation results with an Autonomous Underwater Vehicle (AUV) demonstrate the effectiveness of the proposed method in trajectory tracking and stability maintenance under attack conditions.

While the introduction of networking has increased the efficiency of Industrial Cyber-Physical Systems (ICPS), it has also lowered the cost for attackers, significantly increasing security risks. Current research on ICPS attack detection focuses on deep learning methods. However, the dependence on large labeled datasets often hinders these systems from adapting quickly to the dynamic changes and real-time demands of the ICPS environment. To address these issues, we present an attack detection method based on improved meta pseudo label (ADIMPL). ADIMPL innovatively combines two-layer network traffic feature extraction with the compact SqueezeNet deep neural network, achieving high performance with a minimal number of labeled samples. Additionally, the method dynamically adapts to changing attack patterns, significantly increasing detection accuracy while enhancing the robustness and real-time processing capabilities of the detection system. Extensive experiments on real-world industrial CPS datasets (CIC-IDS2017, CIC-IDS2018, and the CIC-Attack Dataset 2023) demonstrate that ADIMPL can effectively, robustly, and in real-time detect network attacks against industrial CPS. Notably, ADIMPL achieves a detection accuracy of 99.13

This paper presents a Multi-Dimensional Network model aimed at analyzing mutually coupled membrane fouling in membrane bioreactors (MBR), addressing the limitations of existing diagnostic methods. First, fouling data were processed through wavelet transformation, using time-domain information as the network input. Second, initial features were extracted via convolution, with an inception structure applied for multidimensional feature extraction. Third, horizontal dense blocks were utilized to reduce network complexity while preserving key information. An attention mechanism was then incorporated to enhance model expressiveness, generalization, and diagnostic stability. Comparative testing revealed strong overall performance in diagnosing membrane fouling in hollow fiber membrane systems, achieving an average accuracy of 99.44%, which increased to 99.72% under noisy conditions. This underscores the model's effectiveness in filtering contaminants in complex environments while maintaining excellent noise resistance and stability. Ultimately, the model significantly reduces operational and maintenance costs of wastewater treatment, providing a valuable theoretical foundation for advancing membrane water treatment technologies.