Engineering Applications of Artificial Intelligence最新文献

Quadruplex-depth based multi-view stereo network with wave-shaped depth cells and Epipolar Transformer 基于波形深度单元和极极变压器的四重深度多视点立体网络

IF 8 2区计算机科学 Q1 AUTOMATION & CONTROL SYSTEMS

Engineering Applications of Artificial Intelligence

Pub Date : 2025-12-20 DOI: 10.1016/j.engappai.2025.113600

Boyang Song, Jin Xiao, Xiaoguang Hu, Baochang Zhang

Most learning-based Multi-View Stereo (MVS) methods focus on accurate depth inference to obtain precise and complete point clouds, where cost aggregation plays a crucial role as a bridge between two-dimensional (2D) images and three-dimensional (3D) representations. While achieving competitive results, conventional MVS methods often adopt cascaded architectures, which risk error propagation, or they neglect the optimization of depth geometry. To address this, we propose the Quadruplex-Depth based MVS Network (ETQ-MVSNet), built upon a progressive refinement framework. Our core innovation is the Quadruplex-Depth (QD) mechanism, which predicts four depth values per pixel and constrains them to form a novel wave-shaped depth geometry. This is complemented by an adaptive initial depth range determination strategy within the Quadruplex-Depth Refinement (QDR) process to reduce prediction deviation. By preemptively modelling the wave-shaped depth map within the prediction network to reduce interpolated depth deviation during the depth fusion phase, our method significantly enhances the overall coherence of the reconstruction pipeline and improves quality. To complement the QD mechanism, which involves double regularization due to its wave-shaped cells, we also incorporate an Epipolar Transformer (ET) for visibility-aware cost aggregation, capturing robust long-range 3D relationships along epipolar lines, and an efficient multi-scale informative feature extraction network that efficiently processes images collectively and extracts high-quality features for all pipeline modules in a single pass. These two designs not only balance the pipeline's reconstruction efficiency, enhancing practical utility, but also improve reconstruction quality in non-ideal scenes. ETQ-MVSNet not only surpasses all previous progressive refinement approaches but also achieves competitive results against state-of-the-art cascaded methods, demonstrating its effectiveness, time efficiency, generalization ability, and strong scalability. Our proposed method can be extended to reconstruct images captured by mobile phones or Unmanned Aerial Vehicles (UAVs) in various applications, including digital heritage conservation and city surveying. The code will be available at https://github.com/Boyang-Song/ETQ-MVSNet.

大多数基于学习的多视图立体（MVS）方法侧重于精确的深度推断，以获得精确和完整的点云，其中成本聚合作为二维（2D）图像和三维（3D）表示之间的桥梁起着至关重要的作用。传统的MVS方法通常采用级联结构，这可能会导致误差传播，或者忽略了深度几何的优化。为了解决这个问题，我们提出了基于四路深度的MVS网络（ETQ-MVSNet），它建立在一个渐进的改进框架之上。我们的核心创新是四重深度（QD）机制，它可以预测每个像素的四个深度值，并将它们约束成一个新的波浪形深度几何形状。在四重深度细化（QDR）过程中，采用自适应初始深度范围确定策略来减少预测偏差。该方法通过对预测网络内的波形深度图进行预先建模，减少深度融合阶段的插值深度偏差，显著增强了重建管道的整体相干性，提高了重建质量。为了补充QD机制（由于其波浪状单元而涉及双正则化），我们还结合了一个Epipolar Transformer (ET)，用于可见性感知成本聚合，捕获沿Epipolar线的强大远程3D关系，以及一个高效的多尺度信息特征提取网络，该网络可以有效地集体处理图像，并在一次通过中提取所有管道模块的高质量特征。这两种设计既平衡了管道的重建效率，增强了实际效用，又提高了非理想场景下的重建质量。ETQ-MVSNet不仅超越了以往所有的渐进式细化方法，而且取得了与最先进的级联方法相媲美的结果，证明了其有效性、时间效率、泛化能力和强大的可扩展性。我们提出的方法可以扩展到重建手机或无人机（uav）在各种应用中捕获的图像，包括数字遗产保护和城市测量。代码可在https://github.com/Boyang-Song/ETQ-MVSNet上获得。

{"title":"Quadruplex-depth based multi-view stereo network with wave-shaped depth cells and Epipolar Transformer","authors":"Boyang Song, Jin Xiao, Xiaoguang Hu, Baochang Zhang","doi":"10.1016/j.engappai.2025.113600","DOIUrl":"10.1016/j.engappai.2025.113600","url":null,"abstract":"<div><div>Most learning-based Multi-View Stereo (MVS) methods focus on accurate depth inference to obtain precise and complete point clouds, where cost aggregation plays a crucial role as a bridge between two-dimensional (2D) images and three-dimensional (3D) representations. While achieving competitive results, conventional MVS methods often adopt cascaded architectures, which risk error propagation, or they neglect the optimization of depth geometry. To address this, we propose the Quadruplex-Depth based MVS Network (ETQ-MVSNet), built upon a progressive refinement framework. Our core innovation is the Quadruplex-Depth (QD) mechanism, which predicts four depth values per pixel and constrains them to form a novel wave-shaped depth geometry. This is complemented by an adaptive initial depth range determination strategy within the Quadruplex-Depth Refinement (QDR) process to reduce prediction deviation. By preemptively modelling the wave-shaped depth map within the prediction network to reduce interpolated depth deviation during the depth fusion phase, our method significantly enhances the overall coherence of the reconstruction pipeline and improves quality. To complement the QD mechanism, which involves double regularization due to its wave-shaped cells, we also incorporate an Epipolar Transformer (ET) for visibility-aware cost aggregation, capturing robust long-range 3D relationships along epipolar lines, and an efficient multi-scale informative feature extraction network that efficiently processes images collectively and extracts high-quality features for all pipeline modules in a single pass. These two designs not only balance the pipeline's reconstruction efficiency, enhancing practical utility, but also improve reconstruction quality in non-ideal scenes. ETQ-MVSNet not only surpasses all previous progressive refinement approaches but also achieves competitive results against state-of-the-art cascaded methods, demonstrating its effectiveness, time efficiency, generalization ability, and strong scalability. Our proposed method can be extended to reconstruct images captured by mobile phones or Unmanned Aerial Vehicles (UAVs) in various applications, including digital heritage conservation and city surveying. The code will be available at <span><span>https://github.com/Boyang-Song/ETQ-MVSNet</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"166 ","pages":"Article 113600"},"PeriodicalIF":8.0,"publicationDate":"2025-12-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145799639","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Single domain generalization method based on simulation-experiment data fusion and meta-learning for rotating machinery fault diagnosis 基于仿真-实验数据融合和元学习的旋转机械故障诊断单域泛化方法

IF 8 2区计算机科学 Q1 AUTOMATION & CONTROL SYSTEMS

Engineering Applications of Artificial Intelligence

Pub Date : 2025-12-20 DOI: 10.1016/j.engappai.2025.113570

Jialu Han , Qibin Wang , Lei Yin , Xinming Xie , Chenyi Lin

Single domain generalization method can diagnose unknown domains by training only one domain, which is a highly ambitious and challenging task. However, most current methods lack explicit physical constraints during data augmentation and do not consider the positive impact of learning intrinsic similarity between samples on generalization. Therefore, a novel single domain generalization framework based on simulation-experiment data fusion and meta-learning is proposed in this paper. Firstly, a diversified Wasserstein generative adversarial network (Div-WGAN) is developed to generate more diverse samples by quantitatively evaluating and feeding back the diversity of the data, and the simulation data generated by mature dynamic models are integrated into the network to ensure the physical properties of the generated samples. Then, in order to improve the generalization capability of the model, a metric-based meta-learning approach is proposed to reveal the similarity relations between the sample pairs by the learnable relation module, and a similarity score accumulation strategy is designed in the testing phase to perform fault diagnosis by synthesizing the information of multiple samples of the same category. Lastly, through extensive recognition experiments conducted on two datasets, the proposed method demonstrates excellent diagnosis performance on single domain generalization diagnosis tasks.

单域泛化方法可以通过只训练一个域来诊断未知域，这是一项雄心勃勃且具有挑战性的任务。然而，目前大多数方法在数据增强过程中缺乏明确的物理约束，并且没有考虑学习样本之间的内在相似性对泛化的积极影响。为此，本文提出了一种基于仿真-实验数据融合和元学习的单域泛化框架。首先，开发多样化的Wasserstein生成对抗网络（Div-WGAN），通过定量评估和反馈数据的多样性来生成更多样化的样本，并将成熟的动态模型生成的仿真数据集成到网络中，以保证生成样本的物理性质。然后，为了提高模型的泛化能力，提出了一种基于度量的元学习方法，通过可学习关系模块揭示样本对之间的相似关系，并在测试阶段设计了相似度评分累积策略，通过综合同一类别的多个样本的信息进行故障诊断。最后，通过在两个数据集上进行的大量识别实验，该方法在单域泛化诊断任务上表现出优异的诊断性能。

{"title":"Single domain generalization method based on simulation-experiment data fusion and meta-learning for rotating machinery fault diagnosis","authors":"Jialu Han , Qibin Wang , Lei Yin , Xinming Xie , Chenyi Lin","doi":"10.1016/j.engappai.2025.113570","DOIUrl":"10.1016/j.engappai.2025.113570","url":null,"abstract":"<div><div>Single domain generalization method can diagnose unknown domains by training only one domain, which is a highly ambitious and challenging task. However, most current methods lack explicit physical constraints during data augmentation and do not consider the positive impact of learning intrinsic similarity between samples on generalization. Therefore, a novel single domain generalization framework based on simulation-experiment data fusion and meta-learning is proposed in this paper. Firstly, a diversified Wasserstein generative adversarial network (Div-WGAN) is developed to generate more diverse samples by quantitatively evaluating and feeding back the diversity of the data, and the simulation data generated by mature dynamic models are integrated into the network to ensure the physical properties of the generated samples. Then, in order to improve the generalization capability of the model, a metric-based meta-learning approach is proposed to reveal the similarity relations between the sample pairs by the learnable relation module, and a similarity score accumulation strategy is designed in the testing phase to perform fault diagnosis by synthesizing the information of multiple samples of the same category. Lastly, through extensive recognition experiments conducted on two datasets, the proposed method demonstrates excellent diagnosis performance on single domain generalization diagnosis tasks.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"166 ","pages":"Article 113570"},"PeriodicalIF":8.0,"publicationDate":"2025-12-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145799690","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Machine learning for resilience analysis: a review of systems under cyberattacks 弹性分析的机器学习：网络攻击下的系统回顾

IF 8 2区计算机科学 Q1 AUTOMATION & CONTROL SYSTEMS

Engineering Applications of Artificial Intelligence

Pub Date : 2025-12-20 DOI: 10.1016/j.engappai.2025.113542

Ruoqing Yin, Liz Varga

Intelligent cyber-physical systems (CPS) are increasingly exposed to sophisticated cyberattacks, yet existing machine learning (ML)-focused surveys provide only fragmented coverage of the resilience lifecycle. They often separate attack modelling from defence strategies, overlook cross-sector insights, and lack clear evaluation benchmarks, limiting practical applicability. This review offers the first lifecycle-oriented synthesis of ML-enabled CPS resilience, covering detection, defence, recovery, and adaptation. It integrates diverse ML-based attack techniques with corresponding resilience mechanisms and provides a comparative analysis across two critical CPS sectors—power and water—to highlight shared vulnerabilities and sector-specific behaviours. The review further distils key limitations in current ML approaches, including data scarcity, interpretability challenges, and limited real-world validation. Finally, it proposes five actionable research directions and resilience quantification considerations to guide future development of robust and transferable ML-based CPS resilience frameworks.

智能网络物理系统（CPS）越来越多地暴露于复杂的网络攻击中，但现有的以机器学习（ML）为重点的调查只提供了弹性生命周期的零散覆盖。他们经常将攻击建模与防御策略分开，忽视了跨部门的洞察力，并且缺乏明确的评估基准，限制了实际的适用性。这篇综述提供了第一个面向生命周期的基于ml的CPS弹性的综合，包括检测、防御、恢复和适应。它集成了各种基于ml的攻击技术和相应的弹性机制，并提供了两个关键CPS部门（电力和水）的比较分析，以突出共享漏洞和部门特定行为。这篇综述进一步总结了当前机器学习方法的主要局限性，包括数据稀缺性、可解释性挑战和有限的现实验证。最后，提出了五个可操作的研究方向和弹性量化考虑，以指导未来发展稳健和可转移的基于ml的CPS弹性框架。

{"title":"Machine learning for resilience analysis: a review of systems under cyberattacks","authors":"Ruoqing Yin, Liz Varga","doi":"10.1016/j.engappai.2025.113542","DOIUrl":"10.1016/j.engappai.2025.113542","url":null,"abstract":"<div><div>Intelligent cyber-physical systems (CPS) are increasingly exposed to sophisticated cyberattacks, yet existing machine learning (ML)-focused surveys provide only fragmented coverage of the resilience lifecycle. They often separate attack modelling from defence strategies, overlook cross-sector insights, and lack clear evaluation benchmarks, limiting practical applicability. This review offers the first lifecycle-oriented synthesis of ML-enabled CPS resilience, covering detection, defence, recovery, and adaptation. It integrates diverse ML-based attack techniques with corresponding resilience mechanisms and provides a comparative analysis across two critical CPS sectors—power and water—to highlight shared vulnerabilities and sector-specific behaviours. The review further distils key limitations in current ML approaches, including data scarcity, interpretability challenges, and limited real-world validation. Finally, it proposes five actionable research directions and resilience quantification considerations to guide future development of robust and transferable ML-based CPS resilience frameworks.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"166 ","pages":"Article 113542"},"PeriodicalIF":8.0,"publicationDate":"2025-12-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145799638","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Spectrum-based anomaly detection using channel state information and attention mechanisms for elderly health monitoring 基于信道状态信息和关注机制的频谱异常检测在老年人健康监测中的应用

IF 8 2区计算机科学 Q1 AUTOMATION & CONTROL SYSTEMS

Engineering Applications of Artificial Intelligence

Pub Date : 2025-12-20 DOI: 10.1016/j.engappai.2025.113583

Abid Hussain , Xiaoqiang Zhu , Zhang Sihai , Fujiang Lin

Detecting abnormal human activities is essential in environments such as multi resident homes and elderly care facilities, where continuous visual monitoring is limited. Early identification of unsafe events, including falls and abrupt movements, can significantly reduce health risks for older adults. However, reliable anomaly detection using wireless signals remains challenging due to the scarcity of datasets that represent diverse real world behavioral deviations. This study proposes a spectrum-aware encoder architecture for anomaly detection that integrates wireless channel state information with frequency domain sensing. The method applies wavelet-based denoising, median filtering, and feature normalization to refine the channel measurements. It then extracts spectral descriptors including power spectral density, skewness, and kurtosis to capture irregular frequency signatures associated with abnormal activities. To address class imbalance in real world data, the training pipeline incorporates the Synthetic Minority Oversampling Technique. The proposed encoder employs positional encoding and multi head self attention to model long range temporal relationships in the processed sequences, forming an Artificial Intelligence framework tailored for human activity anomaly detection. Experimental results demonstrate that the spectrum-aware encoder achieves higher precision, recall, and overall robustness compared with deep learning baselines such as convolutional neural networks, long short-term memory networks, gated recurrent units, spectral temporal Transformer and attention based methods . The encoder-only design also offers reduced memory usage and faster training relative to traditional encoder decoder architectures, highlighting its suitability for real-time deployment in resource-constrained elderly-care environments.

在多住户住宅和老年人护理设施等环境中，检测异常的人类活动至关重要，因为这些环境的连续视觉监测有限。早期发现不安全事件，包括跌倒和突然移动，可以大大减少老年人的健康风险。然而，使用无线信号进行可靠的异常检测仍然具有挑战性，因为缺乏代表各种现实世界行为偏差的数据集。本研究提出一种频谱感知的异常检测编码器架构，将无线信道状态信息与频域感知相结合。该方法采用基于小波的去噪、中值滤波和特征归一化来细化通道测量。然后提取频谱描述符，包括功率谱密度、偏度和峰度，以捕获与异常活动相关的不规则频率特征。为了解决现实世界数据中的类不平衡问题，训练管道采用了合成少数派过采样技术。该编码器采用位置编码和多头自关注对处理序列中的长时间时间关系进行建模，形成适合人类活动异常检测的人工智能框架。实验结果表明，与卷积神经网络、长短期记忆网络、门控循环单元、频谱时序转换器和基于注意力的深度学习方法相比，频谱感知编码器具有更高的精度、召回率和整体鲁棒性。与传统的编码器-解码器架构相比，仅编码器的设计还提供了更少的内存使用和更快的训练速度，突出了其在资源受限的老年护理环境中实时部署的适用性。

{"title":"Spectrum-based anomaly detection using channel state information and attention mechanisms for elderly health monitoring","authors":"Abid Hussain , Xiaoqiang Zhu , Zhang Sihai , Fujiang Lin","doi":"10.1016/j.engappai.2025.113583","DOIUrl":"10.1016/j.engappai.2025.113583","url":null,"abstract":"<div><div>Detecting abnormal human activities is essential in environments such as multi resident homes and elderly care facilities, where continuous visual monitoring is limited. Early identification of unsafe events, including falls and abrupt movements, can significantly reduce health risks for older adults. However, reliable anomaly detection using wireless signals remains challenging due to the scarcity of datasets that represent diverse real world behavioral deviations. This study proposes a spectrum-aware encoder architecture for anomaly detection that integrates wireless channel state information with frequency domain sensing. The method applies wavelet-based denoising, median filtering, and feature normalization to refine the channel measurements. It then extracts spectral descriptors including power spectral density, skewness, and kurtosis to capture irregular frequency signatures associated with abnormal activities. To address class imbalance in real world data, the training pipeline incorporates the Synthetic Minority Oversampling Technique. The proposed encoder employs positional encoding and multi head self attention to model long range temporal relationships in the processed sequences, forming an Artificial Intelligence framework tailored for human activity anomaly detection. Experimental results demonstrate that the spectrum-aware encoder achieves higher precision, recall, and overall robustness compared with deep learning baselines such as convolutional neural networks, long short-term memory networks, gated recurrent units, spectral temporal Transformer and attention based methods . The encoder-only design also offers reduced memory usage and faster training relative to traditional encoder decoder architectures, highlighting its suitability for real-time deployment in resource-constrained elderly-care environments.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"166 ","pages":"Article 113583"},"PeriodicalIF":8.0,"publicationDate":"2025-12-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145799691","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Tacit algorithmic collusion in deep reinforcement learning guided price competition among a set of fast-charging electric vehicle hubs 深度强化学习中的隐性算法合谋指导了一组快速充电电动汽车轮毂之间的价格竞争

IF 8 2区计算机科学 Q1 AUTOMATION & CONTROL SYSTEMS

Engineering Applications of Artificial Intelligence

Pub Date : 2025-12-20 DOI: 10.1016/j.engappai.2025.113567

Diwas Paudel, Tapas K. Das

Players in price competition are increasingly using learning algorithms for pricing decisions. There have been reports of concern that algorithm-guided players are learning to sustain higher than competitive prices, even without communicating with one another. We examine this concern of tacit algorithmic collusion by considering a novel engineering problem where a set of electric vehicle (EV) charging hubs, serving a common pool of price-sensitive EV owners, compete by dynamically varying their prices. In the competition, each hub aims to maximize its revenue via pricing and minimize its cost of power procurement from the day-ahead (DA) and real-time (RT) electricity markets, and the in-house battery storage systems. We develop a two-step data-driven methodology to develop hub pricing strategies. The first step obtains the DA power commitment for the hubs by solving a stochastic optimization model. The second step generates the dynamic pricing strategies for each hub by solving a competitive Markov decision process model using a multi-agent deep reinforcement learning (MADRL) algorithm. We examine the pricing strategies obtained by implementing our methodology on a sample numerical hub pricing problem. Using the profits from the pricing strategies, we calculate an index measuring the level of tacit algorithmic collusion. An index value of zero indicates no collusion (perfect competition), and one indicates full collusion (monopolistic behavior). Our numerical study yields collusion values that suggest the presence of a low to moderate level of collusion in our hub pricing game.

价格竞争的参与者越来越多地使用学习算法进行定价决策。有报道担心，算法引导的玩家正在学习维持高于竞争对手的价格，即使彼此之间没有沟通。我们通过考虑一个新的工程问题来研究这种隐性算法合谋的担忧，其中一组电动汽车（EV）充电中心，服务于对价格敏感的电动汽车车主的共同池，通过动态变化价格来竞争。在竞争中，每个中心的目标是通过定价最大化其收入，并最小化其从前一天（DA）和实时（RT）电力市场以及内部电池存储系统的电力采购成本。我们开发了一个两步数据驱动的方法来制定枢纽定价策略。第一步通过求解一个随机优化模型得到轮毂的数据中心功率承诺。第二步，利用多智能体深度强化学习（MADRL）算法求解竞争马尔可夫决策过程模型，生成每个枢纽的动态定价策略。我们检查定价策略通过实施我们的方法在一个样本数值轮毂定价问题。利用定价策略的利润，我们计算了一个衡量隐性算法合谋水平的指标。指数值为0表示无共谋（完全竞争），指数值为1表示完全共谋（垄断行为）。我们的数值研究得出的合谋值表明，在我们的枢纽定价博弈中存在低至中等水平的合谋。

{"title":"Tacit algorithmic collusion in deep reinforcement learning guided price competition among a set of fast-charging electric vehicle hubs","authors":"Diwas Paudel, Tapas K. Das","doi":"10.1016/j.engappai.2025.113567","DOIUrl":"10.1016/j.engappai.2025.113567","url":null,"abstract":"<div><div>Players in price competition are increasingly using learning algorithms for pricing decisions. There have been reports of concern that algorithm-guided players are learning to sustain higher than competitive prices, even without communicating with one another. We examine this concern of tacit algorithmic collusion by considering a novel engineering problem where a set of electric vehicle (EV) charging hubs, serving a common pool of price-sensitive EV owners, compete by dynamically varying their prices. In the competition, each hub aims to maximize its revenue via pricing and minimize its cost of power procurement from the day-ahead (DA) and real-time (RT) electricity markets, and the in-house battery storage systems. We develop a two-step data-driven methodology to develop hub pricing strategies. The first step obtains the DA power commitment for the hubs by solving a stochastic optimization model. The second step generates the dynamic pricing strategies for each hub by solving a competitive Markov decision process model using a multi-agent deep reinforcement learning (MADRL) algorithm. We examine the pricing strategies obtained by implementing our methodology on a sample numerical hub pricing problem. Using the profits from the pricing strategies, we calculate an index measuring the level of tacit algorithmic collusion. An index value of zero indicates no collusion (perfect competition), and one indicates full collusion (monopolistic behavior). Our numerical study yields collusion values that suggest the presence of a low to moderate level of collusion in our hub pricing game.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"166 ","pages":"Article 113567"},"PeriodicalIF":8.0,"publicationDate":"2025-12-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145799689","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Cloud prediction via spatiotemporal-frequency differential and attentional network 基于时空频差和注意网络的云预测

IF 8 2区计算机科学 Q1 AUTOMATION & CONTROL SYSTEMS

Engineering Applications of Artificial Intelligence

Pub Date : 2025-12-19 DOI: 10.1016/j.engappai.2025.113476

Jiabing Liu , Jianhao Sun , Haiwen Wei , Qilei Li , Junzhi Shi , Mingliang Gao

Cloud prediction is pivotal for meteorology, aviation safety, and renewable energy management. A fundamental challenge in existing deep learning approaches lies in the trade-off among prediction accuracy, computational efficiency, and long-term stability. To bridge this gap, we introduce an end-to-end encoder–decoder architecture, termed Spatiotemporal-Frequency Differential and Attentional Network (SFDANet). SFDANet constructs an encoder–decoder architecture with a unique SFFE block, which integrates spatiotemporal and frequency-domain analysis to simultaneously capture localized cloud textures and global evolutionary dynamics. Between the encoder and decoder, an innovative Multi-scale Differential Pyramid (MDP) module is built to selectively enhance high-frequency details critical for rapid cloud evolution while inherently suppressing noise. To explicitly model complex temporal dynamics, we propose a new module parallel to MDP, named Multi-order Projection Attention (MPA). This module operates by projecting input features into a set of parallel subspaces. Crucially, these subspaces are designed to be both linear and non-linear. Through this architectural design, the module is capable of simultaneously capturing predictable low-order trends and intricate high-order patterns within the data. Comprehensive experiments on the WeatherBench dataset demonstrate that SFDANet achieves superior accuracy and long-term stability, while it maintains remarkable efficiency with only 0.65M parameters. The code is available at https://github.com/liu-jiabing/SFDA

云预测对气象学、航空安全和可再生能源管理至关重要。现有深度学习方法的一个基本挑战在于预测精度、计算效率和长期稳定性之间的权衡。为了弥补这一差距，我们引入了端到端的编码器-解码器架构，称为时空-频率差分和注意网络（SFDANet）。SFDANet构建了一个具有独特SFFE块的编码器-解码器架构，该架构集成了时空和频域分析，同时捕获局部云纹理和全局演化动态。在编码器和解码器之间，构建了一个创新的多尺度差分金字塔（MDP）模块，可以选择性地增强对快速云演化至关重要的高频细节，同时固有地抑制噪声。为了明确地模拟复杂的时间动力学，我们提出了一个与MDP并行的新模块，称为多阶投影注意（MPA）。该模块通过将输入特征投影到一组平行子空间中来运行。关键是，这些子空间被设计成线性和非线性的。通过这种体系结构设计，该模块能够同时捕获数据中可预测的低阶趋势和复杂的高阶模式。在WeatherBench数据集上进行的综合实验表明，SFDANet在仅使用0.65M参数的情况下，具有较好的精度和长期稳定性。代码可在https://github.com/liu-jiabing/SFDA上获得

{"title":"Cloud prediction via spatiotemporal-frequency differential and attentional network","authors":"Jiabing Liu , Jianhao Sun , Haiwen Wei , Qilei Li , Junzhi Shi , Mingliang Gao","doi":"10.1016/j.engappai.2025.113476","DOIUrl":"10.1016/j.engappai.2025.113476","url":null,"abstract":"<div><div>Cloud prediction is pivotal for meteorology, aviation safety, and renewable energy management. A fundamental challenge in existing deep learning approaches lies in the trade-off among prediction accuracy, computational efficiency, and long-term stability. To bridge this gap, we introduce an end-to-end encoder–decoder architecture, termed Spatiotemporal-Frequency Differential and Attentional Network (SFDANet). SFDANet constructs an encoder–decoder architecture with a unique SFFE block, which integrates spatiotemporal and frequency-domain analysis to simultaneously capture localized cloud textures and global evolutionary dynamics. Between the encoder and decoder, an innovative Multi-scale Differential Pyramid (MDP) module is built to selectively enhance high-frequency details critical for rapid cloud evolution while inherently suppressing noise. To explicitly model complex temporal dynamics, we propose a new module parallel to MDP, named Multi-order Projection Attention (MPA). This module operates by projecting input features into a set of parallel subspaces. Crucially, these subspaces are designed to be both linear and non-linear. Through this architectural design, the module is capable of simultaneously capturing predictable low-order trends and intricate high-order patterns within the data. Comprehensive experiments on the WeatherBench dataset demonstrate that SFDANet achieves superior accuracy and long-term stability, while it maintains remarkable efficiency with only 0.65M parameters. The code is available at <span><span>https://github.com/liu-jiabing/SFDA</span><svg><path></path></svg></span></div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"166 ","pages":"Article 113476"},"PeriodicalIF":8.0,"publicationDate":"2025-12-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145799636","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

An innovative feature clustering paradigm based on Hypergraph cooperative graph convolutional network for hyperspectral image classification 一种创新的基于超图协同图卷积网络的特征聚类范式用于高光谱图像分类

IF 8 2区计算机科学 Q1 AUTOMATION & CONTROL SYSTEMS

Engineering Applications of Artificial Intelligence

Pub Date : 2025-12-19 DOI: 10.1016/j.engappai.2025.113597

Zhen Zhang , Lehao Huang , Yabin Hu , Qingwang Wang , Chunxue Xu , Yemao Qi , Chenxi Liu

Hyperspectral Image Classification (HSIC) constitutes a pivotal endeavor in remote sensing, facilitating high-precision delineation of Earth's surface features. Conventional deep learning approaches, however, frequently fail to account for the irregular, non-Euclidean spatial arrangement of natural features, resulting in the aggregation of extraneous or misleading information that undermines the discriminative capacity of target classes. To surmount these limitations, this study proposes an innovative feature clustering paradigm, instantiated through a Hypergraph Cooperative Graph Convolutional Network (HCoGCN). By devising a Hypergraph Action Network (HACN) and a Hypergraph Node Feature Adaptive Aggregation Module (HNFA²M), this framework adeptly clusters and integrates features from homogeneous regions within non-Euclidean domains. Further refinement is achieved through a Pixel-level Compensation Mechanism (PCM), which synergistically incorporates Euclidean-space pixel-level features to bolster classification precision. The proposed method achieves the highest classification accuracies of 95.49 %, 97.66 %, and 98.75 % on the QUH-Qingyun, QUH-Pingan, and QUH-Tangdaowan datasets, respectively, outperforming existing mainstream approaches by a significant margin. Comprehensive ablation and comparative analyses substantiate the paradigm's robustness and adaptability, underscoring its efficacy in capturing intricate spatial-spectral interrelations across Euclidean and non-Euclidean spaces. This work heralds a transformative advance in HSIC by foregrounding the potency of feature clustering as a foundational strategy.

高光谱图像分类（HSIC）是遥感领域的一项关键技术，有助于对地球表面特征进行高精度描绘。然而，传统的深度学习方法经常无法解释自然特征的不规则、非欧几里得空间排列，导致无关或误导性信息的聚集，从而破坏了目标类别的判别能力。为了克服这些限制，本研究提出了一种创新的特征聚类范式，通过Hypergraph Cooperative Graph Convolutional Network （HCoGCN）实例化。通过设计超图动作网络（HACN）和超图节点特征自适应聚合模块（HNFA2M），该框架熟练地聚集和集成了非欧几里得域内同质区域的特征。进一步的细化是通过像素级补偿机制（PCM）来实现的，该机制协同结合欧几里得空间像素级特征来提高分类精度。该方法在青海青云、青海平安和青海汤岛湾数据集上的分类准确率分别达到95.49%、97.66%和98.75%，显著优于现有主流方法。综合分析和比较分析证实了该范式的稳健性和适应性，强调了其在捕获欧几里得和非欧几里得空间中复杂的空间-光谱相互关系方面的有效性。这项工作通过突出特征聚类作为基础策略的潜力，预示着HSIC的变革性进步。

{"title":"An innovative feature clustering paradigm based on Hypergraph cooperative graph convolutional network for hyperspectral image classification","authors":"Zhen Zhang , Lehao Huang , Yabin Hu , Qingwang Wang , Chunxue Xu , Yemao Qi , Chenxi Liu","doi":"10.1016/j.engappai.2025.113597","DOIUrl":"10.1016/j.engappai.2025.113597","url":null,"abstract":"<div><div>Hyperspectral Image Classification (HSIC) constitutes a pivotal endeavor in remote sensing, facilitating high-precision delineation of Earth's surface features. Conventional deep learning approaches, however, frequently fail to account for the irregular, non-Euclidean spatial arrangement of natural features, resulting in the aggregation of extraneous or misleading information that undermines the discriminative capacity of target classes. To surmount these limitations, this study proposes an innovative feature clustering paradigm, instantiated through a Hypergraph Cooperative Graph Convolutional Network (HCoGCN). By devising a Hypergraph Action Network (HACN) and a Hypergraph Node Feature Adaptive Aggregation Module (HNFA<sup>2</sup>M), this framework adeptly clusters and integrates features from homogeneous regions within non-Euclidean domains. Further refinement is achieved through a Pixel-level Compensation Mechanism (PCM), which synergistically incorporates Euclidean-space pixel-level features to bolster classification precision. The proposed method achieves the highest classification accuracies of 95.49 %, 97.66 %, and 98.75 % on the QUH-Qingyun, QUH-Pingan, and QUH-Tangdaowan datasets, respectively, outperforming existing mainstream approaches by a significant margin. Comprehensive ablation and comparative analyses substantiate the paradigm's robustness and adaptability, underscoring its efficacy in capturing intricate spatial-spectral interrelations across Euclidean and non-Euclidean spaces. This work heralds a transformative advance in HSIC by foregrounding the potency of feature clustering as a foundational strategy.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"166 ","pages":"Article 113597"},"PeriodicalIF":8.0,"publicationDate":"2025-12-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145799640","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

An improved dynamic attention mechanism-based transformers approach for motor imagery electroencephalogram signal classification 一种改进的基于动态注意机制的运动意象脑电图信号分类方法

IF 8 2区计算机科学 Q1 AUTOMATION & CONTROL SYSTEMS

Engineering Applications of Artificial Intelligence

Pub Date : 2025-12-19 DOI: 10.1016/j.engappai.2025.113524

Uzma Nawaz , Mufti Anees-ur-Rahaman , Zubair Saeed

This paper presents a novel dynamic attention mechanism-based transformers (DAMT) model for motor imagery electroencephalogram (EEG) signal classification. The DAMT model integrates channel-wise dynamic attention (CWDA) and time-wise dynamic attention (TWDA), which allows it to focus on the most significant spatial and temporal features of EEG signals. This method efficiently addresses substantial challenges, including inter-subject variability, limited datasets, and the risk of overfitting. The DAMT model captures complicated dependencies across the time and channel domains using transformer encoders and dynamic attention. This improves the overall classification accuracy of motor imaging tasks. We evaluated the model's performance on nine subjects using two benchmark datasets from the Brain-Computer Interface (BCI) Competition IV. Additionally, we utilized motor imagery EEG data from 103 subjects in the PhysioNet Motor Movement/Imagery dataset and 38 subjects in the GigaScience dataset, which focuses exclusively on motor imagery tasks. Our model outperformed the existing art, with an average accuracy of 96 % for BCI Competition IV 2a, 98 % for BCI Competition IV 2b, 93 % for PhysioNet Motor Movement/Imagery, and 95 % for GigaScience Motor Imagery. We also conducted an ablation study to determine the relevance of each component by removing CWDA and TWDA, as well as reducing the number of transformer encoder layers. The results showed a significant decrease in performance, emphasizing the need for dynamic attention processes to preserve the accuracy of the model. These results highlight the adaptability and robustness of the DAMT model for effectively classifying motor imagery.

提出了一种新的基于动态注意机制的运动意象脑电图（EEG）信号分类模型。DAMT模型将通道动态注意（CWDA）和时间动态注意（TWDA）相结合，使其能够关注脑电信号最显著的时空特征。这种方法有效地解决了实质性的挑战，包括学科间的可变性、有限的数据集和过拟合的风险。DAMT模型使用变压器编码器和动态关注捕获跨时间和通道域的复杂依赖关系。这提高了运动成像任务的整体分类精度。我们使用来自脑机接口（BCI）竞赛IV的两个基准数据集评估了该模型在九名受试者上的性能。此外，我们利用了PhysioNet运动运动/图像数据集中的103名受试者和GigaScience数据集中的38名受试者的运动图像EEG数据，该数据集专门关注运动图像任务。我们的模型优于现有的艺术，BCI竞争IV 2a的平均准确率为96%，BCI竞争IV 2b的平均准确率为98%，PhysioNet运动/图像的平均准确率为93%，GigaScience运动图像的平均准确率为95%。我们还进行了消融研究，通过去除CWDA和TWDA以及减少变压器编码器层的数量来确定每个组件的相关性。结果显示，性能显著下降，强调需要动态注意过程来保持模型的准确性。这些结果突出了DAMT模型对运动图像有效分类的适应性和鲁棒性。

{"title":"An improved dynamic attention mechanism-based transformers approach for motor imagery electroencephalogram signal classification","authors":"Uzma Nawaz , Mufti Anees-ur-Rahaman , Zubair Saeed","doi":"10.1016/j.engappai.2025.113524","DOIUrl":"10.1016/j.engappai.2025.113524","url":null,"abstract":"<div><div>This paper presents a novel dynamic attention mechanism-based transformers (DAMT) model for motor imagery electroencephalogram (EEG) signal classification. The DAMT model integrates channel-wise dynamic attention (CWDA) and time-wise dynamic attention (TWDA), which allows it to focus on the most significant spatial and temporal features of EEG signals. This method efficiently addresses substantial challenges, including inter-subject variability, limited datasets, and the risk of overfitting. The DAMT model captures complicated dependencies across the time and channel domains using transformer encoders and dynamic attention. This improves the overall classification accuracy of motor imaging tasks. We evaluated the model's performance on nine subjects using two benchmark datasets from the Brain-Computer Interface (BCI) Competition IV. Additionally, we utilized motor imagery EEG data from 103 subjects in the PhysioNet Motor Movement/Imagery dataset and 38 subjects in the GigaScience dataset, which focuses exclusively on motor imagery tasks. Our model outperformed the existing art, with an average accuracy of 96 % for BCI Competition IV 2a, 98 % for BCI Competition IV 2b, 93 % for PhysioNet Motor Movement/Imagery, and 95 % for GigaScience Motor Imagery. We also conducted an ablation study to determine the relevance of each component by removing CWDA and TWDA, as well as reducing the number of transformer encoder layers. The results showed a significant decrease in performance, emphasizing the need for dynamic attention processes to preserve the accuracy of the model. These results highlight the adaptability and robustness of the DAMT model for effectively classifying motor imagery.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"166 ","pages":"Article 113524"},"PeriodicalIF":8.0,"publicationDate":"2025-12-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145799626","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Comparative analysis of advanced machine learning models for second-grade liquid via stimulating magnetized CoFe2O4 nanoparticles over a convective moving plate 通过在对流移动板上刺激磁化CoFe2O4纳米颗粒的二级液体先进机器学习模型的比较分析

IF 8 2区计算机科学 Q1 AUTOMATION & CONTROL SYSTEMS

Engineering Applications of Artificial Intelligence

Pub Date : 2025-12-19 DOI: 10.1016/j.engappai.2025.113551

Umair Khan , Hamza Rauf , Aurang Zaib , Abeer M. Alotaibi

Applications

Magnetic nanoparticles have gained a lot of attention in a variety of sectors due to their well-known rigorous features, which have led to numerous applications. The most frequently studied magnetic nanoparticles are cobalt ferrite (CoFe₂O₄) nanoparticles due to their elevated effectiveness, chemical durability, and adjustable electrical and magnetic properties render them appropriate for various applications, including environmental remediation, catalysis, biomedicine, etc.

Novelty

The objective of the current investigation is to explore the time-dependent radiative flow of magnetized cobalt ferrite (CoFe₂O₄) nanoparticles near a stagnation-point induced by second-grade fluid subject to moving plate with convective boundary condition. This exploration offers valuable insights to improve the efficiency of the heat transport indicating it as a fundamental factor for thermal engineering in sophisticated liquid systems.

Methodology

The similarity factors transform the partial differential equations into similarity equations of a specific form. Dual numerical solutions are calculated by employing built-in solver bvp4c. Moreover, to overcome the computational expense of the numerical solver, this study also pioneers a comparative analysis of advanced machine learning surrogates. While artificial neural networks with Levenberg-Marquardt are recognized for exceptional precision, this study addresses a critical question by quantifying to what extent Gaussian Process Regression (GPR) can match this benchmark while offering uncertainty quantification.

Results

From the physical findings, we can see from the upper branch solution that temperature goes down and velocity increases. The reason for this is that as velocity increases, the viscosity factor decreases which results in reduction in temperature. And as for the machine learning results, they show GPR has reached a remarkable R² of 1 and when predicting the Nusselt number, the predicted error is comparable to ANN-LM benchmark. Thus, validating the GPR technique for high accuracy near perfect and indicating GPR can be a great option when both precision and confidence are required in predictions.

磁性纳米颗粒由于其众所周知的严格特性，在各个领域得到了广泛的关注，这导致了许多应用。最常被研究的磁性纳米粒子是钴铁氧体（CoFe2O4）纳米粒子，因为它们具有更高的效率、化学耐久性和可调节的电和磁特性，使它们适合于各种应用，包括环境修复、催化、生物医学、本研究的目的是探索二级流体在对流边界条件下受移动板作用时磁化的钴铁氧体（CoFe2O4）纳米颗粒在停滞点附近的随时间的辐射流动。这一探索为提高热传递效率提供了有价值的见解，表明它是复杂液体系统热工程的基本因素。方法相似因子将偏微分方程转化为特定形式的相似方程。利用内置的求解器bvp4c计算了对偶数值解。此外，为了克服数值求解器的计算费用，本研究还开创性地对先进的机器学习替代品进行了比较分析。虽然Levenberg-Marquardt人工神经网络被认为具有卓越的精度，但本研究通过量化高斯过程回归（GPR）在多大程度上可以匹配该基准来解决一个关键问题，同时提供不确定性量化。结果从物理上看，上支路溶液温度下降，速度增大。这样做的原因是，随着速度的增加，粘度系数降低，导致温度降低。而对于机器学习的结果，他们显示GPR已经达到了显著的R2为1，并且在预测Nusselt数时，预测误差与ANN-LM基准相当。因此，当预测需要精度和信心时，验证GPR技术近乎完美的高精度和指示GPR可能是一个很好的选择。

{"title":"Comparative analysis of advanced machine learning models for second-grade liquid via stimulating magnetized CoFe2O4 nanoparticles over a convective moving plate","authors":"Umair Khan , Hamza Rauf , Aurang Zaib , Abeer M. Alotaibi","doi":"10.1016/j.engappai.2025.113551","DOIUrl":"10.1016/j.engappai.2025.113551","url":null,"abstract":"<div><h3>Applications</h3><div>Magnetic nanoparticles have gained a lot of attention in a variety of sectors due to their well-known rigorous features, which have led to numerous applications. The most frequently studied magnetic nanoparticles are cobalt ferrite (CoFe<sub>2</sub>O<sub>4</sub>) nanoparticles due to their elevated effectiveness, chemical durability, and adjustable electrical and magnetic properties render them appropriate for various applications, including environmental remediation, catalysis, biomedicine, etc.</div></div><div><h3>Novelty</h3><div>The objective of the current investigation is to explore the time-dependent radiative flow of magnetized cobalt ferrite (CoFe<sub>2</sub>O<sub>4</sub>) nanoparticles near a stagnation-point induced by second-grade fluid subject to moving plate with convective boundary condition. This exploration offers valuable insights to improve the efficiency of the heat transport indicating it as a fundamental factor for thermal engineering in sophisticated liquid systems.</div></div><div><h3>Methodology</h3><div>The similarity factors transform the partial differential equations into similarity equations of a specific form. Dual numerical solutions are calculated by employing built-in solver bvp4c. Moreover, to overcome the computational expense of the numerical solver, this study also pioneers a comparative analysis of advanced machine learning surrogates. While artificial neural networks with Levenberg-Marquardt are recognized for exceptional precision, this study addresses a critical question by quantifying to what extent Gaussian Process Regression (GPR) can match this benchmark while offering uncertainty quantification.</div></div><div><h3>Results</h3><div>From the physical findings, we can see from the upper branch solution that temperature goes down and velocity increases. The reason for this is that as velocity increases, the viscosity factor decreases which results in reduction in temperature. And as for the machine learning results, they show GPR has reached a remarkable R<sup>2</sup> of 1 and when predicting the Nusselt number, the predicted error is comparable to ANN-LM benchmark. Thus, validating the GPR technique for high accuracy near perfect and indicating GPR can be a great option when both precision and confidence are required in predictions.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"166 ","pages":"Article 113551"},"PeriodicalIF":8.0,"publicationDate":"2025-12-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145799634","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Improved self-attention multi-model fusion for load forecasting in regional integrated energy systems 区域综合能源系统负荷预测的改进自关注多模型融合

IF 8 2区计算机科学 Q1 AUTOMATION & CONTROL SYSTEMS

Engineering Applications of Artificial Intelligence

Pub Date : 2025-12-19 DOI: 10.1016/j.engappai.2025.113557

Jian Shi , Jiashen Teh , Bader Alharbi

The electric, cooling, and heating loads of the Regional Integrated Energy System (RIES) exhibit high randomness, volatility, and complex interdependencies, making accurate forecasting challenging. To address this, a novel RIES prediction model is developed by integrating Complementary Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), Sample Entropy (SE), Variational Mode Decomposition (VMD), Multiple Linear Regression (MLR), and Bi-directional Temporal Convolutional Network-Transformer (BiTCN-TR). CEEMDAN is a signal decomposition technique that addresses non-stationarity by injecting adaptive noise. The model replaces traditional word embedding and position encoding with Bi-directional Temporal Convolutional Network (BiTCN), enhancing the extraction of high-dimensional feature data and long-term temporal dependencies. CEEMDAN decomposes the load series, followed by SE for identifying the most complex modes and VMD for secondary decomposition to reduce non-stationarity. Modal components are classified as high or low-frequency using zero-crossing rates, with BiTCN-TR predicting high-frequency components and MLR handling low-frequency components. Results demonstrate the model's ability to improve forecasting accuracy and parameter optimization, contributing to stable and reliable energy system management.

区域综合能源系统（RIES）的电、冷、热负荷表现出高度的随机性、波动性和复杂的相互依赖性，使得准确预测具有挑战性。为了解决这一问题，通过集成自适应噪声互补集成经验模态分解（CEEMDAN）、样本熵（SE）、变分模态分解（VMD）、多元线性回归（MLR）和双向时间卷积网络变压器（BiTCN-TR），开发了一种新的RIES预测模型。CEEMDAN是一种通过注入自适应噪声来解决非平稳性的信号分解技术。该模型用双向时间卷积网络（BiTCN）取代传统的词嵌入和位置编码，增强了高维特征数据和长期时间依赖关系的提取能力。CEEMDAN对荷载序列进行分解，然后通过SE识别最复杂的模态，VMD进行二次分解，减少非平稳性。使用过零率将模态分量分类为高频或低频，BiTCN-TR预测高频分量，MLR处理低频分量。结果表明，该模型能够提高预测精度和参数优化，有助于稳定可靠的能源系统管理。

{"title":"Improved self-attention multi-model fusion for load forecasting in regional integrated energy systems","authors":"Jian Shi , Jiashen Teh , Bader Alharbi","doi":"10.1016/j.engappai.2025.113557","DOIUrl":"10.1016/j.engappai.2025.113557","url":null,"abstract":"<div><div>The electric, cooling, and heating loads of the Regional Integrated Energy System (RIES) exhibit high randomness, volatility, and complex interdependencies, making accurate forecasting challenging. To address this, a novel RIES prediction model is developed by integrating Complementary Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), Sample Entropy (SE), Variational Mode Decomposition (VMD), Multiple Linear Regression (MLR), and Bi-directional Temporal Convolutional Network-Transformer (BiTCN-TR). CEEMDAN is a signal decomposition technique that addresses non-stationarity by injecting adaptive noise. The model replaces traditional word embedding and position encoding with Bi-directional Temporal Convolutional Network (BiTCN), enhancing the extraction of high-dimensional feature data and long-term temporal dependencies. CEEMDAN decomposes the load series, followed by SE for identifying the most complex modes and VMD for secondary decomposition to reduce non-stationarity. Modal components are classified as high or low-frequency using zero-crossing rates, with BiTCN-TR predicting high-frequency components and MLR handling low-frequency components. Results demonstrate the model's ability to improve forecasting accuracy and parameter optimization, contributing to stable and reliable energy system management.</div></div>","PeriodicalId":50523,"journal":{"name":"Engineering Applications of Artificial Intelligence","volume":"166 ","pages":"Article 113557"},"PeriodicalIF":8.0,"publicationDate":"2025-12-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145799635","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0