Abstract:Aiming at the problem of insufficient multi-unit synergetic control under hydraulic coupling in the single penstock multi-machine system,this paper proposes a multi-unit synergetic calming control strategy based on synergetic control theory for establishing multi-unit linkage.First,the dynamic head-coupling change term of the shared pipeline is reconsidered,and the macro-variables containing the state variables of multiple units are constructed.The equilibrium point is preset by taking the output deviation of the main state variables of multi-unit as the control core.Then,the macro-variables are substituted into the control manifold to derive a new form of coupled head term,which transforms the original linear model into a nonlinear model to track the unit changes.By comparing the state outputs in the original form,it is verified that the proposed strategy performs well in reducing the output oscillations of the unit and enhancing the stability of the system's isolated network operation,which provides a feasible solution for the treatment of hydraulic coupling.

Abstract:An integrated energy system(IES)that incorporates distributed energy resources such as photovoltaics,ener-gy storage,and gas turbines has the potential to provide a multi-energy coordinated and complementary energy utilization form,which can play an important role in participating in grid demand response.To effectively respond to grid peak regulation demands,this paper proposes an optimization method for IES intraday scheduling considering multi-energy complementarity and internal user response as the energy management means of IES.Firstly,based on the multi-energy coupling operation architecture of IES,the response characteristics of internal users are analyzed.The electric load demand of internal users is changed by subsidy price and load reduction respectively,and then the energy management strategy optimization model of IES participating in power grid demand response under photovoltaic output and load uncertainty is constructed.Secondly,the deep reinforcement learning algorithm based on TD3 is used to solve the IES energy man-agement strategy.Finally,the case study shows that the proposed energy management strategy optimization model and strategy optimization method can reasonably formulate the energy conversion control and demand response scheme within the system to fully tap the response potential of the system and effectively achieve the peak regulation demand response goal of the grid.

Abstract:Hypersonic flight vehicle is a multi-input multi-output system with strong nonlinearity and strong coupling characteristics.In addition,when the elevator is used as the only control surface,the coupling term between the elevator and lift force in the system model leads to the non-minimum phase behavior of the hypersonic flight vehicle system,which poses certain challenge to its tracking control problem.This paper investigates the tracking control problem of hypersonic flight vehicle based on discrete-time output regulation theory,where the elevator is the only control surface.The tracking control problem of hypersonic flight vehicle is firstly formulated as an approximate discrete-time output regulation problem.Since it is difficult to obtain the exact solution of the discrete regulator equations for the hypersonic flight vehicle,the approximate solution of the discrete regulator equations is obtained by neural network method and then the discrete-time neural network controller is designed to achieve tracking control of hypersonic flight vehicle in this paper.The simulation results show that the proposed control algorithm can lead to satisfactory tracking performance.

Abstract:Currently,studies on the bifurcation dynamics of neural networks mainly focus on the binary interactions between neurons,while higher-order interactions between neurons in the form of groups and clusters are common in neural networks.However,the effect of higher-order interactions on the dynamics of neural networks is not well understood.The study of neural networks with higher-order interactions can further explore the higher-order properties and dynamics of real neural networks.In this paper,we propose a class of fractional-order time-delayed neural network with ring-star structure under higher-order interactions.The time delay is chosen as the bifurcation parameter,and the stability of the system and the sufficient condition for Hopf bifurcation are given,which reveals the mechanism of the higher-order coupling coefficient,the self-feedback coefficient and the fractional-order on the system dynamics.

Abstract:To enhance the security of industrial cyber-physical systems(ICPS)simultaneously subject to false data injection(FDI)attacks on both the actuator channel and sensor channel,this paper proposes a secure control strategy based on reconstructing FDI attack signals and ICSP states.Firstly,an augmented state comprising the system state and sensor attack signals is constructed to establish the augmented system.Secondly,based on the established augmented system,an adjustable proportional-integral observer is designed to reconstruct the system state,as well as sensor and actuator FDI attack signals.Then,a feedback controller is designed using the reconstructed system state and actuator FDI attack signals.Furthermore,the Lyapunov functions and the finite frequency domain H∞ are used to derive the conditions required by the system to meet stability and robustness.Finally,using the linearized longitudinal dynamics of a vertical take-off and landing aircraft as a case study,the simulation results demonstrate that the proposed control strategy can effectively defend against FDI attacks on ICPS while maintaining system stability.

Abstract:The spacing policy used in collaborative control of vehicle platoon is one of the key factors to maintain inner-vehicle stability and string stability of vehicle platoon system.Most of the existing research on mixed vehicle platoon control adopts the constant spacing policy,which is difficult to apply to the complex road driving environment.Therefore,the cooperative control method of mixed vehicle platoon based on variable spacing policy and the stability problem of mixed vehicle platoon are studied.Firstly,for the mixed vehicle platoon composed of human-driven vehicles(HDV)and connected autonomous vehicles(CAV),a quadratic variable spacing policy is designed,and the mixed vehicle platoon system model based on the variable spacing policy is constructed.Secondly,a cooperative controller of mixed vehicle platoon based on multi-agent consistency is proposed,and the head-to-tail transfer function of mixed vehicle platoon for a bidirectional multi-vehicle leader-follower topology,which includes multiple connected autonomous vehicles,is derived.Finally,numerical simulation experiments are designed to verify the effectiveness of the proposed controller,and the effects of driver response time delay,vehicle communication time delay,number of connected autonomous vehicles,information topologies,different types of vehicle order and control gains on the stability of the mixed vehicle platoon are discussed.

Abstract:This paper investigates the bipartite consensus problem for first-order multi-agent systems with the detail-balanced communication topology.Different from the common dynamic event-triggered control methods in the literature,a new dynamic event-triggered control strategy combined with the reset mechanism is proposed,in which the external dynamic variable in the trigger condition threshold can be adjusted according to the preset reset condition.If the local disagreement state error reaches the preset reset condition,the external dynamic variable will be reset to its initial value,to avoid the frequent triggering phenomena when the system is close to the consensus point,and further reduce the communi-cation burden of the system while ensuring the desired control performance.The proposed reset event-triggering condition in this paper only depends on the local states of the agents and does not require any global information.Moreover,the al-gebraic graph theory and the Lyapunov stability theory are applied to prove the practical bipartite consensus of the system.In addition,a theoretical analysis of Zeno-free is given.Finally,the simulation verifies the effectiveness of the proposed method.

Abstract:Under the background of the dual carbon goals,the development of the manufacturing industry faces both challenges and opportunities.In response to national policies,vigorously reducing carbon emissions,this paper focuses on the green scheduling problem of distributed hybrid flow shops.For factory workshops with the same processing capabilities,this paper constructs a green scheduling problem model for distributed homogeneous hybrid flow-shops.Combining the characteristics of actual factories,a formula for calculating carbon emissions during processing is provided.An improved NSGA-Ⅱ algorithm is proposed,including a hybrid initialization strategy,an update strategy,and a carbon reduction strategy to enhance the algorithm's performance.In the experimental validation of the algorithm,ablation studies were designed to verify the effectiveness of the proposed strategies.Additionally,comparative experiments with various advanced multi-objective optimization algorithms validate the effectiveness of the improved algorithm in solving this problem.

Abstract:This paper proposes a dynamic sensor-weapon-target assignment(SWTA)method based on deep reinforce-ment learning(DRL)and graph neural network(GNN),aimed at addressing the complex and dynamic decision-making requirements on modem battlefields.Traditional static methods are inefficient and lack adaptability in real-time chang-ing battlefield environments.To tackle this issue,DRL is combined with GNN to build an intelligent decision-making framework.This framework leverages environmental interaction and deep learning to optimize decision-making strategies,thereby improving resource allocation efficiency and decision accuracy.Guided by the OODA loop theory,the framework uses GNN to capture the relationships between weapons,targets,and sensors in the battlefield,quickly generating assign-ment solutions.The DRL component then optimizes these strategies,enabling resource allocation optimization in dynamic environments.The optimization process takes into account operational effectiveness,resource consumption,and the pro-tection of key locations.Experiments demonstrate that this method performs excellently in various scenarios,significantly enhancing resource utilization and operational outcomes.

Abstract:A networked system refers to a system composed of multiple subsystems with the capability of interaction and task execution,connected through a network.Due to the properties of high dimension,multiple constraints,non-convexity,and nonlinearity in various application scenarios,the analysis and control of networked systems have received widespread attention from different communities in this century.To address uncertainties in dynamic environments and complex sys-tems,end-to-end methods like reinforcement learning have been introduced to learn control policies of networked systems.However,the high-dimensional nature of networked systems poses significant challenges to the learning efficiency.In fa-ct,many studies have found that the performance of networked systems is often closely related to the network structures.By approaching the problem from a graph perspective,complex optimization and control problems can be transformed into simple combinatorial optimization problems,enabling the scalability of the method to practical large-scale networks.In light of this,this paper systematically reviews learning-based control methods for networked systems from the graph perspective.By examining optimal control problems formulated by linear quadratic regulation and Markov games,respec-tively,it highlights the critical role of graph structures in learning control policies for networked systems.Additionally,some challenges in this field are outlined and a prospective outlook on future directions is provided.

Abstract:To address the current issues of slow operating speeds and limited working ranges of autonomous marine vehicles,this paper proposes a new high-speed single-strut unmanned hydrofoil vessel:Sea-U-Foil.Unlike traditional autonomous marine vehicles,Sea-U-Foil features three motion modes,making it suitable for various marine tasks.This paper focuses on an in-depth study of the foilborne mode.The foilborne mode of Sea-U-Foil draws on the design principles of fixed-wing aircraft,incorporating the hydrofoil as its critical component.When the hydrofoil moves forward in water,it generates upward lift,which elevates the hull above the water surface,thereby reducing the contact area between the hull and the water.The decrease in contact area further reduces the drag experienced by Sea-U-Foil.As a result of the significant reduction in drag,Sea-U-Foil achieves faster operating speeds,lower energy consumption,and a broader working range.Additionally,according to Bernoulli's principle,the lift generated by the hydrofoil is proportional to the square of the speed.As speed increases,lift also increases,which enhances the payload capacity of Sea-U-Foil.In this paper,a detailed introduction is presented to the working principle,design,and control of Sea-U-Foil,and validates the effectiveness of our design through experiments,demonstrating the superior performance of Sea-U-Foil compared to traditional monohull and catamaran vessels.A related video can be found here:

Abstract:User interaction-based evolutionary optimization can effectively improve the performance of personalized recommendation.However,existing studies have overlooked the deviation between the encoded individuals and the de-coded candidates,often resulting in a significant deviation in the search direction and low search efficiency.Moreover,the quantitative representation of user interaction evaluation is also a major challenge.To address this,this paper proposes a personalized differential evolution algorithm that integrates Kalman filter deviation estimation and surrogate models.First-ly,a deep belief network trained with user evaluation and product attributes is constructed to achieve quantitative evaluation of user interactions.Then,a Kalman filter estimator is designed to track the deviation between genotypes and phenotypes during the evolution process,and a differential evolution operator is designed based on this deviation to change the popula-tion distribution and guide the search direction.Finally,this algorithm is applied to the Amazon personalized search dataset to verify its effectiveness.

Abstract:Online diverse content generation is one of the emergent research directions in the field of procedural content generation in recent years.It can not only meet users' different preferences and enhance user experience,but also provide a large amount of scenarios and problems for training and testing artificial intelligence algorithms.Recent research proposed online diverse content generation methods based on negatively correlated ensemble reinforcement learning,such methods can not effectively meet the specific preferences of different users.Furthermore,training and deploying individual learning models requires significant computational resources.To address those two issues,this paper proposes an online content generation approach based on multi-objective ensemble pruning,built upon the negatively correlated ensemble reinforce-ment learning framework.This approach searches for the weights for integrating individual learning models through an efficient multi-objective optimization algorithm,so that the obtained ensemble model can not only effectively match user preferences,but also offer a Pareto set that exhibits a tradeoff between model performance and computational resource consumption.This approach matches user preferences by adjusting the weights of individual learning models instead of retraining models,thereby reducing the computational resource consumption.

Abstract:In recent years,multi-agent reinforcement learning(MARL)has gained significant attention due to its poten-tial in solving complex real-world problems.The centralized training with decentralized execution(CTDE)framework has been widely adopted in complex multi-agent systems.CTDE alleviates the non-stationarity problem in MARL by utilizing centralized training,but it also introduces new challenges,particularly in handling the information of all agents during training,especially the exponentially growing joint action space as the number of agents increases.This paper provides a selective review of the algorithms and developments in cooperative multi-agent reinforcement learning within the CTDE framework,focusing on two main approaches:Value function factorization methods and policy-based methods.This pa-per summarizes the performance of these algorithms in addressing the complexity of joint action spaces,non-stationarity,and estimation errors.It also proposes new perspectives and ideas,aiming to provide researchers in the field with deeper insights and guidance for future studies.

Abstract:By using cross-task knowledge transfer,multi-task optimization can achieve better convergence performance than traditional single-task optimization.However,in multi-task optimization,the deviation of the search space and opti-mization scenarios,as well as noise that may interfere with knowledge transfer,can lead to a decrease in the efficiency of effective knowledge transfer and even negative transfer.An multi-task optimization algorithm based on partial least squares(PLS)subspace alignment and reuse population mechanism(PR-MTEA)is proposed to solve the problem.Firstly,by intro-ducing the PLS subspace projection strategy,the high-dimensional task search space is transformed into a low-dimensional space and specific low-dimensional subspaces are established for each task's population.Secondly,real-time adjustment of the Bregman divergence of the subspace is used to obtain an alignment matrix and achieve cross-task knowledge transfer.Finally,a population reuse mechanism based on the Residual structure is designed to avoid negative transfer and getting stuck in local optima,as well as to improve the convergence of the algorithm.Comparative experimental results with four other advanced multi-task algorithms show that PR-MTEA has better convergence performance and faster search ability.In addition,sensor coverage problem is conducted to test and analyze the feasibility and applicability of the improved algorithm.

Abstract:Based on the multi-variable coupling characteristics of the thermal dynamics and energy consumption of the HVAC system and the problem of insufficient data prediction accuracy,this paper proposes a fusion prediction model integrating time convolution-long short-term memory-attention mechanism(TCN-LSTM-Attention).Firstly,in order to better capture the short-term and long-term dependencies in the building operation data,a TCN-LSTM-Attention building thermal dynamics and energy consumption prediction model is established to predict HVAC energy consumption,indoor temperature,and PMV.The improved particle swarm optimization(IPSO)algorithm is used to optimize the hyperparam-eters of the prediction model,reduce the prediction error of the model,and analyze the model's approximation ability.Secondly,the EnergyPlus is used to build a building simulation model for verification.The prediction model is verified by using the operation data of an office building in Hebei Province.The experiments show that this model has better prediction accuracy and prediction stability compared with the comparison algorithms,and the generalization of the algorithm when the building envelope parameters change is verified.

Abstract:Differential games offer a powerful framework for modeling and analyzing the decision-making problems in multi-agent systems under competitive environments,with extensive application prospects in fields like economics and industry.However,as the problem modeling approaches real-world scenarios,obtaining theoretical solutions becomes increasingly challenging.In recent years,with the rapid advancement of artificial intelligence,multi-agent deep rein-forcement learning has achieved significant breakthroughs,providing promising alternatives for addressing the challenges encountered in differential games.This paper firstly provides a comprehensive review of the fundamental principles and lat-est development of both fields.The paper further details the reinforcement learning methods applied to differential games,categorizing them based on the problem formulations and deep-learning network architectures,and illustrates how recent researches overcome the challenges faced by traditional methods.Finally,the paper summarizes the current progress in this interdisciplinary research area and suggests potential future directions.

Abstract:This study addresses the dynamic cascading workshop energy-efficiency scheduling problem(DCSESP)that consists of distributed flow shops(DFS),hybrid flow shops(HFS),and the transportation stage between them.The objective is to minimize the total tardiness and total energy consumption.To achieve this,a mixed-integer linear programming(MILP)model is developed,and a graph-based deep reinforcement learning(GDRL)algorithm is proposed.First,a heterogeneous graph model is designed for the DCSESP,and a three-stage node embedding method is introduced to capture the real-time shop floor state features.Second,based on these state features,the algorithm directly selects jobs and operations for both stages of the cascaded dual-shop system.Finally,a multilayer perceptron(MLP)and graph attention network(GAT)are integrated into the proximal policy optimization(PPO)algorithm with an actor-critic framework to facilitate learning and decision-making for rapid joint scheduling.The experimental results demonstrated that the proposed GDRL algorithm outperformed the three state-of-the-art scheduling methods in solving the DCSESP,particularly in complex scheduling scenarios,where it achieved higher optimization performance and robustness.

Abstract:Traditional multi-objective optimization algorithms often suffer from rigid reference point distribution,weak environmental adaptability,and population diversity degradation,leading to imbalanced solution set distribution and low convergence efficiency.This paper proposes multi-objective optimization with adaptive reference-point updates and pop-ulation prediction.Firstly,an elite gene-guided reproductive crossover operator is designed to enhance global search and diversity through a triple mechanism of interference,exchange,and inheritance.Secondly,a population prediction inte-grates regularized regression with boundary perturbation to forecast new solutions,achieving dynamic fusion of historical information and new populations via error correction.Thirdly,an adaptive reference point update strategy dynamically eliminates invalid points and generates new ones to improve coverage in high-dimensional objective spaces.Finally,a complete algorithmic framework is established based on these strategies.Experimental results demonstrate the algorithm's superior performance on benchmark test problems and a real-world aluminum electrolysis process parameter optimization case.

Abstract:Due to the random arrival of orders during production,companies may face challenges that result in the execution of a scheduling plan that is not optimal.The real-time dynamic arrival of orders is a critical issue.To address the dynamic multiplicity flexible job-shop scheduling problem(DMFJSP),which involves dynamic arrivals of job orders and multiple job types,a multi-policy dueling double deep Q-network(MPD3QN)solution is proposed.Firstly,to reduce the complexity of DMFJSP,a simplified fluid relaxation model is introduced and a multi-criteria selection strategy is developed based on this model to aid production scheduling decisions.Secondly,a Markov decision process(MDP)framework is constructed by extracting 19 state features related to jobs and machines,and 20 composite rules are designed to form the action space.Then,the MPD3QN algorithm is formulated through the integration of prioritised experience replay,a soft update mechanism,and an adaptive action selection strategy.Finally,the proposed method is evaluated through 81 test instances,and its performance is compared against three existing deep reinforcement learning scheduling approaches.Simulation results confirm its superior performance in scheduling efficiency and robustness.