243 results about "Rapid convergence" patented technology

The invention discloses an urban power grid real-time load collaborative peak regulation method based on multi-energy complementation and AI scheduling. The method comprises the following steps of multi-source data access and high-dimensional feature space construction, embedded entropy calculation and interactive network construction, domain knowledge and data-driven model fusion, hierarchical scheduling and dual-stage optimization, and real-time decision and closed-loop feedback. According to the method, the real-time performance and hierarchical scheduling thought are emphasized, and an organic closed loop is formed on the three aspects of intra-day scheduling, hour-level rolling correction and minute-level or second-level emergency response. Meanwhile, by means of a multi-stage optimizer switching mechanism, the model can complete rapid convergence of high-dimensional parameters in a short time, finer strategy fine adjustment is carried out in the later period, and the accuracy and reliability of a peak regulation scheme are guaranteed; the method can be applied to advanced power grid systems such as intelligent power grid dispatching, a multi-energy collaborative optimization platform and demand side response management, and has the characteristics of high real-time performance, strong adaptability and good expandability.

The invention relates to the technical field of adaptive filtering, discloses a variable-step robust adaptive filter and a filter network, and aims to solve the technical problem that a traditional adaptive filter cannot effectively consider convergence speed, steady-state precision and impulse noise interference resistance. A non-linear error factor is introduced, large errors caused by impulse noise are effectively suppressed through a generalized function form of the non-linear error factor, and the robustness of the filtering process is remarkably improved; meanwhile, a variable step length mechanism is provided, the theoretically optimal candidate step length is calculated on line by minimizing the mean square deviation of the next moment, and a self-adaptive target variable step length is generated through truncation and time smoothing processing, so that dynamic balance is realized between rapid convergence and low-steady-state maladjustment. According to the invention, the filter is expanded to the distributed network, each node follows a diffusion strategy of first updating and then combining, and global information sharing and performance collaborative optimization are realized by using a combined coefficient.

The invention relates to the technical field of optical communication, and discloses an optical module adaptive test and rapid convergence method based on reinforcement learning, which can dynamically generate and execute an optimal test sequence for optical modules of different types or states through an off-line trained RL intelligent agent module, avoids redundant steps in a fixed process, and improves the test efficiency. On the premise of ensuring the test coverage, the test time is obviously shortened, and meanwhile, the detection probability of potential defects is improved; through the closed-loop test of the integrated RL intelligent agent module, the test action can be adjusted in real time according to the state of a single optical module, so that the test process can quickly adapt to the individual difference and process fluctuation of the device, the online quick convergence of the test strategy is realized, and the accuracy and stability of the test are guaranteed; and through online iterative optimization of the RL agent module, new data and new changes in the production process can be continuously learned, so that a test strategy is continuously evolved, and the optimal performance is kept for a long time.

The invention relates to the technical field of wireless communication, and discloses an unloading method in an ultra-dense millimeter wave MEC network. Aiming at the problems of insufficient terminal capability, high energy consumption of an ultra-dense base station, millimeter wave coverage limitation, communication security risk and the like under the condition of sharp increase of a calculation-intensive task, the method comprises the following steps: firstly, obtaining network basic information, constructing a network architecture containing communication, calculation unloading and a security model, and establishing a multi-constraint optimization problem; based on the optimization problem, initializing a multi-strategy black-wing plinuary optimization algorithm population through logic mapping and elite selection; updating individual positions through global and local search in algorithm iteration, and screening historical optimal individuals; and finally, configuring and unloading resources according to the configuration. According to the method, the NOMA technology, a multi-step unloading framework and a hybrid communication mode are combined, the optimization efficiency is improved through a multi-strategy black-wing optimization algorithm, the local energy consumption is reduced, the communication rate is improved, the safety is guaranteed, the optimal scheme meeting time delay and safety constraints is rapidly converged, and the user experience is improved.

The invention belongs to the technical field of intelligent manufacturing and production scheduling, discloses a flexible job shop dynamic scheduling method and system considering machine aging, and designs a hierarchical environmental response strategy which can firstly evaluate the severity of environmental change. When the change is not violent, only a lightweight local optimization strategy is adopted for fine adjustment; and when the change is relatively violent, a global reconstruction strategy combining knowledge migration, reinitialization and directional repair is started. Therefore, the algorithm can intelligently allocate computing resources according to the intensity of environment change, blind global search is avoided, and the response speed and the operation efficiency of the algorithm are greatly improved. The design effectively balances the exploration and utilization capabilities of the algorithm, and maintains the diversity of the population while ensuring rapid convergence, thereby obtaining a group of Pareto optimal solution sets with good convergence and wider distribution.

The invention discloses a whole network reference unified Beidou B2b constant deviation estimation method, which comprises the following steps of: firstly, performing Beidou B2b clock error reference transformation by applying Beidou B2b service DCB information to realize constant deviation parameterization separation; secondly, adding a single-station virtual reference, and constructing a constant deviation full-rank estimation model based on a single station; then, the constant deviation information of the whole network is integrated, the reference difference of a plurality of single station solutions is eliminated, and B2b constant deviation estimation with the unified reference of the whole network is achieved; and finally, under the constant deviation constraint of the unified reference, the performance of terminal positioning convergence speed, precision and the like is improved. According to the method, constant deviation estimation based on the whole network is realized, the stability and availability of Beidou B2b constant deviation estimation are improved, the problems of to-be-estimated parameter hopping and even filtering divergence possibly caused when a user side adopts constant deviations of different references are avoided, and then rapid convergence and instantaneous horizontal decimeter-level positioning of a terminal user are guaranteed.

The invention discloses an aspheric surface processing intelligent feedback control system based on digital twinning, and the system comprises a digital twinning modeling and baseline calibration module which is used for building an aspheric surface processing digital twinning body and generating a baseline; the twinborn update and virtual-real residual calculation module is used for calculating virtual-real residual update digital twinborn bodies; the reinforcement learning candidate strategy generation module is used for generating reinforcement learning candidate strategies through the reinforcement learning strategy network; the model predictive control optimization module is used for establishing a model predictive control optimization problem and performing rolling prediction; the double-strategy fusion control module is used for executing virtual processing simulation and generating a control vector; and the virtual-real closed-loop adaptive convergence execution module is used for executing aspheric surface processing to realize closed-loop adaptive convergence control. According to the method, digital twinning is combined with reinforcement learning and model prediction control, intelligent optimization of aspheric surface machining is achieved, and the method has the advantages of being high in precision and stability and rapid in convergence.

The invention discloses a micro-grid cooperative scheduling control method, system, equipment and medium based on edge calculation, and relates to the technical field of computer platform load balancing, and the method comprises the steps: collecting power grid operation data through a micro-grid unit, carrying out the short-time power prediction of the micro-grid unit according to the power grid operation data, and obtaining a load prediction value; deploying a genetic algorithm at an edge node, and optimizing a load prediction value to obtain a scheduling strategy parameter; the scheduling strategy parameters are input into the micro-grid units, a game model is constructed to carry out multi-micro-grid cooperative game, and optimal scheduling strategy parameters are output; and performing power output scheduling through a multi-stage control mechanism. According to the method, through multi-dimensional data acquisition and sequence feature learning of micro-grid units, high-precision short-time power prediction is realized, a rapid convergence local optimization strategy is obtained, it is ensured that multiple micro-grids reach Nash equilibrium under autonomous conditions, and rapid response to sudden disturbance and effective correction of steady-state errors are realized.

The invention relates to the technical field of image recognition, and particularly provides a multi-view point cloud registration method, which comprises the following steps of: performing coarse registration on a source point cloud and a target point cloud based on multi-dimensional features by using an RANSAC (Random Sample Consensus) algorithm in a coarse registration stage to obtain a coarse registration transformation matrix; in the fine registration stage, the coarse registration transformation matrix is used as an initial value, point cloud registration is carried out in at least two resolution spaces, segmented iterative optimization is carried out by using an error loss function set in each resolution space, rapid convergence is carried out in a low-resolution space through an iterative nearest point algorithm, and the point cloud registration is realized. And local geometric alignment optimization is carried out in other resolution spaces through a generalized iterative nearest point algorithm, and fine registration of the source point cloud and the target point cloud is completed after multi-resolution space progressive optimization registration. According to the method, the robustness of feature matching of the low-overlap region is remarkably improved, the registration speed and precision are balanced, and the limitation of a traditional point cloud registration method under the low-overlap and non-ideal point cloud condition is solved.

The invention relates to the technical field of intelligent control, and discloses a multi-agent consistency detection system, which effectively improves the cooperative control capability of a multi-agent system in a dynamic environment. A real-time topology updating mechanism guarantees rapid convergence when a network structure changes, a prediction compensation module significantly reduces control delay caused by environmental interference, and a multi-dimensional health monitoring system improves the accuracy and efficiency of fault recovery. Functional decoupling is achieved through the modular design, system expansion and maintenance are facilitated, autonomous adjustment and optimization of a recovery strategy are achieved through a strategy optimization mechanism driven by reinforcement learning, the fault recovery time is shortened, and the success rate is increased. A secondary health verification mechanism ensures effectiveness of recovery operation, and secondary faults caused by continuous operation of the system in an incomplete recovery state are avoided.

The invention discloses a business processing method and device based on an ant colony algorithm, electronic equipment and a medium. The method comprises the following steps: acquiring a business service request, and analyzing the business service request to obtain a business task; determining each agent matched with the business task; wherein the intelligent agents are used for cooperative processing of business tasks; planning the cooperation path of each agent based on an ant colony algorithm to obtain a target path; and according to the cooperation sequence of the intelligent agents in the target path, executing the business task by using the intelligent agents in sequence. According to the technical scheme, efficient collaboration between agents is achieved through the improved ant colony algorithm. According to the system, a multi-dimensional pheromone matrix is adopted to record and transmit collaborative experience among departments, a knowledge verification mechanism based on swarm intelligence is established, and dynamic optimization and rapid convergence of a cross-department business process are realized.

Provided are techniques for the calibration of distillation learning from a teacher model to a student model. Specifically, the present disclosure proposes systems and methods that provide convergence with both high quality and speed. That is, example proposed systems both enable the distillation loss to be minimized at the probability mean value in the probability domain of the teacher's predictions distributions while also providing a loss that is nicely (e.g., symmetrically and / or strongly) convex around an optimum in the logit and / or probability domains (e.g., including far from the minimum) to encourage fast convergence of gradient based methods (e.g., irrespective of distance from the minimum).

The application discloses a data center micro-grid electric calculation cooperative double-layer deep reinforcement learning scheduling method and belongs to the technical field of power systems. Firstly, aiming at the high renewable energy penetration scene of the geographical distributed data center micro-grid, a double-layer multi-agent deep reinforcement learning framework is proposed. The double-layer multi-agent deep reinforcement learning framework containing a global layer and a local layer is constructed. The global layer is distributed by centralized deep reinforcement learning training to calculate the task, and the local layer is optimized by decentralized deep reinforcement learning to internally dispatch energy. The layered multi-agent double-delay deep deterministic policy gradient algorithm is combined to realize cross-layer interaction and rapid convergence. Through space-time task adjustment and energy storage cooperation, the application reduces the cost of non-renewable energy by 15%-20%, and the convergence speed is improved by 5 times compared with the prior art. Moreover, the application supports dynamic data backup, significantly improves the system reliability and economy while maintaining the service quality.

The invention discloses a low-voltage transformer area electric energy meter time synchronization method and system based on swarm intelligence, and relates to the technical field of intelligent power grid and power distribution automation. The method comprises the steps that a transformer area acquisition terminal issues reference time, and each electric energy meter collects neighborhood information through neighbor electric energy meter bidirectional message exchange; jointly estimating the clock skew and the drift rate by adopting a consistency algorithm; introducing a time-varying weight calculated according to SNR, a message success rate and historical stability to suppress the influence of an inferior link and an abnormal node; the virtual synchronization time is output through software compensation on the service side; and adjusting the synchronization period interval and the convergence step length in a linkage manner according to a double-threshold self-adaptive rule. According to the scheme, deviation and drift rate rapid convergence and stable noise suppression are realized in a time-varying and easy packet loss environment, communication energy consumption is reduced, time jump caused by hardware callback is avoided, and the method has the advantages of high precision, strong robustness and large-scale deployment.

This invention relates to the field of image recognition technology, specifically providing a multi-view point cloud registration method. In the coarse registration stage, the RANSAC algorithm is used to perform coarse registration of the source and target point clouds based on multi-dimensional features, obtaining a coarse registration transformation matrix. In the fine registration stage, the coarse registration transformation matrix is ​​used as the initial value, and point cloud registration is performed in at least two resolution spaces. Segmented iterative optimization is performed using an error loss function set for each resolution space. In the low-resolution space, an iterative nearest-point algorithm is used for rapid convergence, while in the remaining resolution spaces, a generalized iterative nearest-point algorithm is used for local geometric alignment optimization. After progressive optimization registration in multiple resolution spaces, the fine registration of the source and target point clouds is completed. This invention significantly improves the robustness of feature matching in low-overlap regions, balances registration speed and accuracy, and overcomes the limitations of traditional point cloud registration methods under low-overlap and non-ideal point cloud conditions.

The application discloses a kind of comprehensive energy system distributed optimization scheduling method and system considering privacy protection, it is related to comprehensive energy system technical field, including: with each equipment in system as node, with energy interaction between node as edge, construct the communication network topology diagram of comprehensive energy system;With total cost minimum as objective function, configure the constraint condition of system operation, build comprehensive energy system optimization scheduling model;Assume that communication network topology diagram is undirected connected graph, using the preset time distribution optimization algorithm of introducing time-varying gain, the optimization scheduling model is solved, so that comprehensive energy system converges to optimal solution from arbitrary initial state within preset time, obtain optimal optimization strategy;Wherein, in solving process, by encryption and decryption module based on deep learning adversarial network, the interaction variable between adjacent nodes is encrypted and decrypted, privacy protection is realized.The application can meet the fast convergence and privacy protection requirements of comprehensive energy system scheduling optimization solution.

The invention discloses a double-layer nonsingular terminal sliding-mode active-disturbance-rejection control method for a permanent magnet synchronous motor based on a superspiral sliding-mode observer, and the method specifically comprises the following steps: 1, building a voltage equation under an alpha-beta static coordinate system, and then building a mathematical model under a d-q rotating coordinate system through coordinate transformation; 2, designing a fractional order linear super-spiral sliding mode observer; step 3, designing a double-layer nonsingular terminal sliding mode active-disturbance-rejection controller according to the mathematical model under the d-q rotating coordinate system obtained in the step 1; and step 4, making a difference between the mechanical angular velocity given value omega ref and the mechanical angular velocity observed value, and obtaining a control signal of a PMSM current loop through a double-layer nonsingular terminal sliding mode active-disturbance-rejection controller according to the obtained error. According to the method, global rapid convergence of the speed tracking error can be ensured, the anti-interference capability of the system is improved to a certain extent, and buffeting is reduced.

The invention discloses an SOC prototype verification logic division method and system based on resource constraints. The method comprises the steps that a to-be-verified SOC prototype is split into modules meeting the resource constraints of an FPGA; the split modules are tried to be combined in pairs to form modules meeting the resource constraint of the FPGA; 1 is added to the loop variable i of the finally obtained module according to the sequence of the number of the external connecting lines from large to small, and the FPGAs are distributed in the multi-FPGA verification platform; integrating the modules distributed with the FPGAs to generate a netlist file; and converting the netlist file into a bit stream file and downloading the bit stream file into the distributed FPGA. According to the method, the process of mapping SOC design to multiple FPGAs is achieved, logic division is rapidly achieved for the design, rapid time sequence convergence is achieved through reasonable arrangement of resources, the FPGA chip resource utilization rate is increased, and time sequence optimization iteration of a net list after integration is reduced.

The invention discloses a multi-network multi-granularity resource optimization method based on federal deep reinforcement learning, and belongs to the technical field of edge computing and artificial intelligence, and the method comprises the steps: a mobile device senses a local resource state, and maps the local resource state into a structured state vector through a multi-granularity resource mapping module; local FDRL training and global prompt vector generation are carried out; performing pruning optimization on the local FDRL training model, and realizing asynchronous uploading by the multiple heterogeneous hybrid network modules based on a heterogeneous link adaptation mechanism; performing global model aggregation, prompting distillation and joint optimization; and strategy return and iterative closed-loop restarting are carried out. According to the invention, a three-level and multi-granularity state modeling method is introduced to enhance the expression ability; an off-line track guiding mechanism is introduced to realize rapid convergence in a strategy cold start stage; generating acceleration strategy convergence by adopting experience-first sampling + prompt; the uploading overhead is reduced by pruning a local FDRL training model, and the transmission efficiency and robustness are improved by adopting a heterogeneous link adaptation mechanism.

The invention discloses a UAV positioning and joint deployment and resource allocation method based on a radar, and relates to the technical field of UAV communication, the angle of a UAV relative to the radar is measured by deploying an EMVS-MIMO radar, the distance is determined by combining pulse ranging, and the three-dimensional position of the UAV can be estimated; according to the method, UAV deployment and GU association are optimized based on a BCD method, firstly, rapid convergence of a particle swarm algorithm and global optimization capacity of a grid search algorithm are combined, and UAV deployment is optimized; then, recalculating the GU association according to the new UAV deployment; a constrained particle swarm optimization algorithm is adopted to optimize the UAV transmitting power, and a variable learning rate is introduced to dynamically adjust the inertia weight, so that particles are helped to find a globally optimal solution; based on different requirements of communication and positioning GU on the bandwidth, a greedy algorithm is adopted to allocate the bandwidth of the UAV, and the fairness between the GUs is ensured.

The invention relates to the technical field of unmanned aerial vehicle auxiliary edge computing task offloading, and particularly provides an edge computing task offloading strategy based on an improved quantum particle swarm optimization (IQP) SO (Inter Quantum Particle Swarm Optimization) algorithm, and relates to an edge computing task offloading method based on the IQP SO algorithm and an edge computing task offloading system based on the IQP SO algorithm based on the IQP SO algorithm based on the IQP SO algorithm based on the IQP SO algorithm based on the IQP SO algorithm. According to the method, an MEC unloading structure of multi-user mobile equipment (UE) supported by an unmanned aerial vehicle is established; under the structure, communication, time delay and energy consumption models are formulated to evaluate time delay and energy consumption required by the unloading task of the mobile equipment; according to the improved quantum particle swarm optimization, the unloading efficiency is improved, and the time delay and energy consumption problems of tasks are considered in the optimization process; the algorithm combines quantum characteristics, has excellent global search capability and rapid convergence characteristics, and effectively avoids the problem of global optimal solution omission caused by premature convergence when optimizing an edge unloading strategy; according to the method, the average time delay and the energy consumption of the mobile edge computing task can be remarkably reduced, and the optimization of the system is realized.

The invention discloses a robot virtual-real cooperative training decision optimization system and method based on digital twinning, and the method comprises the following steps: collecting state data and disturbance data of an entity robot in a real environment, and carrying out the preprocessing of the data to generate standardized input; joint coding and state perturbation mapping are carried out on the standardized data, and a feature vector sequence embedded in a hyperspherical manifold space is generated; inputting to a virtual twin control body based on a hypersurface neural element structure, executing disturbance direction sensitive activation, and outputting an activated state vector sequence; virtual and entity control action prediction sequences are generated respectively, an embedded space difference vector is calculated, and control body parameters are updated based on a consistency optimization criterion; after convergence, the control body executes reasoning to generate a target control action sequence, the entity robot is driven to complete action execution, and control strategy optimization is achieved. According to the method, high-precision migration and rapid convergence of a robot control strategy are realized, and the execution stability in a complex disturbance environment is improved.

The embodiment of the invention provides a multi-Kalman filtering smoothing GNSS / IMU positioning method and system, electronic equipment and a storage medium, a multi-stage Kalman filtering smoothing technology is adopted, forward filtering and reverse filtering results are fused to serve as an IMU initial value, through iteration of forward filtering and reverse filtering, the operation time is theoretically prolonged, and the positioning accuracy is improved. The IMU can be quickly converged in short-route operation, so that the precision and timeliness of a small surveying and mapping task are improved; a bidirectional filtering fusion technology is applied, a forward filtering result is used as an initial reverse filtering value, a reverse filtering result is used as an initial forward filtering value of the next time, high-precision positioning attitude output can be realized through multiple iterative solutions without depending on specific initialization maneuver, and flexibility is provided for navigation in a complex environment (such as a city).

This invention proposes a fast convergence path planning method suitable for robotic arms in complex environments. First, in collision detection, an obstacle expansion method is used to treat expanded obstacles as collision detection obstacles, preventing unnecessary potential collisions between the algorithm's execution unit and obstacles during physical experiments. Second, a better constraint sampling method is adopted to further narrow the sampling interval and sample within the restricted interval. Finally, a sampling point optimization strategy is introduced, designing a cost function and performing weight adaptation. The cost of multiple sampled candidate points is calculated, and the optimal point is selected as the final sampling point. This invention solves the problem of low path generation quality in complex environments for robotic arms.

The invention discloses a communication and calculation overlap optimization method, system and device and a storage medium, which are corresponding schemes, and in the scheme, the performance of calculating communication overlap can be improved by optimizing communication parameters on the premise of not introducing excessive overhead. Moreover, the configuration with better performance can be quickly found in a huge search space through an effective optimization method; based on the above scheme, communication parameters can be dynamically optimized, resource use can be balanced, calculation and communication overlapping performance can be improved on the premise that the calculation task and the set communication task of the NCCL library can be correctly executed, the method is suitable for complex calculation communication overlapping scenes, and rapid convergence can be realized in the distributed training process of the deep neural network model; evaluation of a plurality of clusters and models shows that rapid convergence can be achieved on deep neural network model training, and meanwhile higher performance earnings can be guaranteed compared with NCCL.

The invention relates to the technical field of three-dimensional modeling, in particular to a low-calculation-power rapid three-dimensional modeling method and system based on 3DGS, and provides the following scheme that a Gaussian field is updated in the previous period of each modeling period to serve as an initial model to execute forward rendering, rendered images are obtained, and residual information is calculated. And performing robust aggregation on the residual error by taking a preset tile as a statistical unit, constructing a projection statistical spectrum in combination with Gaussian projection contribution, performing cross-period comparison to obtain a structure change degree, and performing residual error gating to form a complexity score so as to divide a complex region and a flat region. Mapping from tiles to Gaussian objects is established according to a density scheduling state, directional splitting and encryption are carried out on a complex region, appearance and transparency are finely adjusted in a limited middle region, and absorption, combination and sparsification are carried out on a flat region, so that rapid convergence and stable modeling are realized under the condition of limited computing power; and the invalid calculation proportion is remarkably reduced while the reconstruction precision and the structural stability are kept.

The invention provides an RDMA congestion control method and system, and the method comprises the steps: enabling a light-weight sensing end to only carry out the simple request sending and rate adjustment through concentrating the core calculation and decision-making functions of congestion control at a receiving end with rich resources for unified execution; therefore, the control logic deployment dilemma caused by topology asymmetry is fundamentally overcome, the occupation of calculation and memory resources of the sensing end is remarkably reduced, the resource advantages of the receiving end are fully played, the macroscopic appointment allocation of the network bandwidth and the microcosmic dynamic adjustment and optimization based on the real-time one-way delay are realized, and the real-time performance of the network bandwidth is improved. And finally, the transmission effects of high bandwidth utilization rate, low transmission delay and rapid convergence are achieved in the edge computing platform.

The application discloses a method, device and medium for controlling a robot, and relates to the technical field of robots, and the method comprises the following steps: generating a dynamic model of the robot based on dynamic parameters of the robot; determining a trajectory tracking error between an actual position and an expected position of an end of the robot, and establishing a trajectory tracking error control model based on the dynamic model and the trajectory tracking error; determining a total sliding mode surface with a recursive structure based on the trajectory tracking error; determining a sliding mode controller based on the trajectory tracking error control model and the total sliding mode surface; and controlling joints of the robot by using the sliding mode controller, so that the actual trajectory of the end of the robot coincides with the expected trajectory. The application can realize fast convergence of the trajectory of the robot, reduce the adverse effects of system uncertainty and interference on the control accuracy of the system, has stronger robustness and anti-interference capability, and helps to improve the control accuracy of the robot.

The invention discloses an intelligent power grid task unloading method for reducing terminal power consumption, and relates to the technical field of intelligent power grids and edge computing. Aiming at the problems of dimensionality explosion, slow convergence and over-high energy consumption faced by traditional deep reinforcement learning (DRL) in large-scale smart grid terminal task unloading, the method is combined with mean field game (MFG) and deep reinforcement learning, and comprises the following steps: 1, establishing an end-side collaborative task scene and an unloading model; 2, calculating time delay and energy consumption of task transmission and execution; 3, an optimization target with the lowest system energy consumption as the target is determined; 4, constructing a large-scale task unloading optimization problem; and 5, solving by adopting a multi-agent actor commentator (MFGMAAC) algorithm guided by a mean field game. According to the method, the state-action space dimension is approximately reduced through the average field, the learning efficiency is improved by adopting a centralized training distribution execution architecture, rapid convergence of an algorithm and reduction of terminal power consumption are realized, and the resource utilization rate and operation stability of the smart grid are improved.

The invention discloses a finite position set phase-locked loop considering speed reversal and calculation burden, and a control method and device thereof. The control method comprises the following steps: acquiring alpha-beta shaft voltage and current signals of a motor, constructing an IPMSM mathematical model under an alpha-beta shaft system, and calculating to obtain a counter electromotive force vector of the motor; taking the counter electromotive force vector of the motor as input, and obtaining the rotating speed and position signals of the motor through a finite position set phase-locked loop considering speed reversal and calculation burden; and the rotating speed and the position signal of the motor are input into a motor closed-loop control system, so that position-sensorless closed-loop control of the permanent magnet synchronous motor is realized. Compared with a provided limited position set phase-locked loop, the method has the advantages that the calculation burden is smaller, the speed reversal coping capacity is lower, full-computational-domain convergence can be achieved through the provided iterative search strategy, and permanent magnet synchronous motor sensorless control with rapid convergence, low calculation burden and high precision is achieved.