Multi-element computing power integrated scheduling method, system and equipment for smart power grid

Abstract

The invention relates to the technical field of intelligent power grids, in particular to a multivariate computing power integrated scheduling method, system and equipment for an intelligent power grid, and the method comprises the steps: carrying out the priority division of a power data set according to a preset business priority label, and obtaining the power feature data of different priorities; performing redundancy elimination processing on the power feature data with different priorities to obtain redundancy-eliminated power feature data; performing secondary classification on the redundancy-removed power feature data to obtain demand feature data of different demand types; performing learning optimization on the demand feature data of different demand types to obtain an optimal power distribution strategy; and based on the optimal power distribution strategy, performing dynamic heterogeneous cooperative scheduling on the multivariate computing power resources by using a computing power chip scheduling algorithm. According to the method, multiple computing power resources are integrated, dynamic scheduling of the computing power resources is carried out on the chip according to the optimal power distribution strategy, the utilization efficiency of the computing power resources is remarkably improved, and therefore the maximization of the use efficiency of the chip is achieved.

Description

Technical Field

[0001] The present invention relates to the field of smart grid technology, and in particular to a multi-computing power integrated scheduling method, system and equipment for smart grids. Background Art

[0002] With the continuous development of power systems and the increasing demand for intelligence, traditional power grids have gradually revealed their limitations in responding to the complex needs of modern society, especially in terms of stability, reliability and efficiency. With the continuous development and application of digital technology, smart grids have emerged as a new stage of power grid development. Smart grids aim to achieve intelligent management and efficient operation of power grids, which requires real-time monitoring of various parameters in the power grid, such as voltage, current and power. Rapid analysis and processing of these data are the key to achieving precise control and optimized scheduling.

[0003] However, the smart grid has an extremely large demand for data processing, which poses an unprecedented challenge to the computing power of grid control chips. Traditional grid control chips perform poorly when processing massive amounts of data and cannot meet the high requirements of smart grids for real-time, accuracy and reliability. Especially in terms of advanced functions such as fault diagnosis, reactive power compensation and distributed energy access, the demand for high-computing power grid control chips is becoming more urgent. At the same time, the widespread application of digital technology has made the connection between the grid and digital devices and systems closer. These devices and systems generate a large amount of data. In order to realize the digital management of the grid, a chip that can seamlessly connect with digital devices and has powerful data processing capabilities is needed. This chip not only needs to be able to monitor the operating status of the grid in real time and quickly identify potential faults, but also needs to be able to evaluate and predict the reliability of the grid, providing a scientific basis for the maintenance and upgrade of the grid.

[0004] In summary, the existing power grid control chips have obvious deficiencies in computing power integration and cannot simultaneously meet the needs of diversified computing power and multi-business integration. Therefore, it is urgent to develop a chip computing power integration method that can efficiently integrate diversified computing power and flexibly schedule according to different business needs. This has become the key to solving the current bottleneck in power grid development. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a multi-computing power integrated scheduling method, system and equipment for smart grid.

[0006] In a first aspect, the present invention provides a multi-computing power integrated scheduling method for a smart grid, the method comprising the following steps:

[0007] Collecting power data sets of different business scenarios of the smart grid, and prioritizing the power data sets according to preset business priority tags to obtain power feature data of different priorities;

[0008] Using the pre-built redundant detection fusion model to perform redundancy processing on the power feature data of different priorities, and obtain the corresponding redundancy-free power feature data;

[0009] Secondarily classifying the de-redundant power characteristic data according to power demand characteristics to obtain demand characteristic data of different demand types;

[0010] Learning and optimizing demand characteristic data of different demand types to obtain an optimal power allocation strategy; the demand types include real-time demand types, safety demand types, and basic demand types;

[0011] Based on the optimal power allocation strategy, the computing chip scheduling algorithm is used to perform dynamic heterogeneous collaborative scheduling of multiple computing resources.

[0012] In a further embodiment, the redundant detection fusion model is constructed by fusing a genetic algorithm and a particle swarm algorithm, and the steps of performing redundancy processing on power feature data of different priorities using the pre-constructed redundant detection fusion model to obtain corresponding de-redundant power feature data include:

[0013] An initial particle group including a plurality of initial particles is randomly generated; each initial particle represents power characteristic data;

[0014] According to the relative positions and quantity distributions of the power characteristic data of different priorities, a fitness function is constructed; the fitness function is used to evaluate the redundancy of the current particles;

[0015] Use the fitness function to preliminarily evaluate the initial particle swarm, calculate the fitness value of each initial particle, and select high-quality solutions based on the fitness value of each initial particle;

[0016] Selecting a number of pairs of parent particles from the high-quality solutions, and performing crossover operations on the parent particles using a genetic algorithm according to extreme values ​​of the power characteristic data to generate new child particles;

[0017] The redundancy of the new generation particles is evaluated twice according to the fitness function, the fitness value of each new generation particle is calculated, and the redundant data points are determined according to the fitness value of each new generation particle;

[0018] Deleting the redundant data points from the new offspring particles or replacing them with non-redundant data points to generate an offspring individual population after the redundant data are removed;

[0019] A mutation operation is performed on the offspring individual population after the redundant data is removed, and the particle group is iteratively updated until a maximum number of iterations is reached, and the redundant power characteristic data is output.

[0020] In a further embodiment, the step of selecting a high-quality solution according to the fitness value of each initial particle comprises:

[0021] According to the fitness value of each initial particle, the initial particles with redundancy lower than the preset redundancy threshold are selected as high-quality solutions.

[0022] In a further embodiment, when the demand characteristic data is de-redundant power characteristic data of a real-time demand type, the step of learning and optimizing the demand characteristic data of different demand types to obtain an optimal power allocation strategy includes:

[0023] According to the demand characteristic data, obtaining local load consumption data of each distributed generator;

[0024] According to the local load consumption data of each distributed generator, the load increment of each distributed generator is calculated by using the first-order discrete time dynamic average consistency algorithm;

[0025] The state of each distributed generator at different times is updated according to the load increment of each distributed generator, and the average grid power of all distributed generators is obtained through iterative calculation;

[0026] Inputting the average grid power into a pre-trained deep neural network model to obtain an optimal power solution for each distributed generator;

[0027] The local power supply and demand deviation of each distributed generator is calculated according to the optimal power solution of each distributed generator, the local power output and the local real-time load demand;

[0028] Utilizing a supply-demand consistency algorithm to optimize and update the local power supply-demand deviation to obtain a global power supply-demand deviation;

[0029] The optimal power solution of each distributed generator is adjusted according to the global power supply and demand deviation until a balance is reached, and an optimal power allocation strategy is output.

[0030] In a further embodiment, when the demand characteristic data is de-redundant power characteristic data of a safety demand type, the step of learning and optimizing the demand characteristic data of different demand types to obtain an optimal power allocation strategy includes:

[0031] Using a multi-layer perceptron to learn and predict a set of potential state points from the demand feature data;

[0032] Select one or more initial state variables from the potential state point set as initial state candidate points of the Gauss-Newton iterative algorithm;

[0033] Sorting the initial state candidate points, and selecting the initial state candidate point with the highest sorting as the hot start point of the Gauss-Newton iterative algorithm;

[0034] The number of iterations is set, and in each iteration of the Gauss-Newton iterative algorithm, the difference between the current state point and the target state point is analyzed according to the Jacobian matrix of the hot start point to obtain an error vector;

[0035] Construct the Hessian matrix according to the Jacobian matrix, solve the incremental equation according to the Hessian matrix and the error vector, and obtain the state update;

[0036] The hot start point is updated according to the state update amount and iteratively optimized until the maximum number of iterations is reached, the final hot start point is output as the optimal state point, and the optimal power allocation strategy is determined according to the optimal state point.

[0037] In a further embodiment, when the demand characteristic data is de-redundant power characteristic data of a basic demand type, the step of learning and optimizing the demand characteristic data of different demand types to obtain an optimal power allocation strategy includes:

[0038] Identifying uncertainty factors in power distribution according to the demand characteristic data, and establishing a mapping relationship between the uncertainty factors and power distribution constraints;

[0039] According to the mapping relationship, the power distribution constraint is learned and identified by using support vector machine learning to obtain a probability distribution of the power distribution constraint;

[0040] Determine active constraints under the current power distribution state according to the probability distribution of power distribution constraints to form an active constraint set;

[0041] The gradient descent method is used to iteratively calculate the active constraint set under the current power distribution state to solve the optimal power allocation strategy under nonlinear constraints.

[0042] In a further implementation scheme, the step of dynamically scheduling heterogeneous collaborative computing resources using a computing chip scheduling algorithm based on an optimal power allocation strategy includes:

[0043] In response to the unit matrix test task instruction, the unit matrix test task is respectively run on the algorithm chip of each computing node to obtain the task test time consumption of the algorithm chip of each computing node; the algorithm chip includes at least a central processing unit and a graphics processing unit;

[0044] According to the task test time of each computing node algorithm chip, the cluster computing capacity evaluation index is obtained;

[0045] Determine the computing tasks to be scheduled, and divide the computing tasks to be scheduled into fine-grained units based on the cluster computing capacity evaluation index using the unit matrix as a unit to obtain several minimum task units;

[0046] According to the optimal power allocation strategy, the optimal computing resources are dynamically scheduled for each minimum task unit in the cluster.

[0047] In a further implementation scheme, the cluster computing capability evaluation index includes the computing capability ratio of each algorithm chip in the node, the comprehensive computing capability ratio of each computing node, and the comprehensive computing capability of the cluster per unit time.

[0048] In a second aspect, the present invention provides a multi-computing power integrated dispatching system for smart grids, the system comprising:

[0049] A data acquisition module is used to collect power data sets of different business scenarios of the smart grid, and prioritize the power data sets according to preset business priority tags to obtain power feature data of different priorities;

[0050] A redundancy detection module, used to perform redundancy removal processing on power feature data of different priorities using a pre-built redundancy detection fusion model to obtain corresponding redundancy-free power feature data;

[0051] A demand classification module, used for secondary classification of the de-redundant power characteristic data according to power demand characteristics to obtain demand characteristic data of different demand types;

[0052] A power allocation module is used to learn and optimize the demand characteristic data of different demand types to obtain an optimal power allocation strategy; the demand types include real-time demand types, safety demand types and basic demand types;

[0053] The dynamic scheduling module is used to dynamically coordinate heterogeneous scheduling of multiple computing resources based on the optimal power allocation strategy and the computing chip scheduling algorithm.

[0054] In a third aspect, the present invention further provides a computer device, comprising a processor and a memory, wherein the processor is connected to the memory, the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory so that the computer device performs the steps of implementing the above method.

[0055] The present invention provides a multi-computing power integrated scheduling method, system and equipment for smart grids. The method collects power data sets of different business scenarios of smart grids, and prioritizes the power data sets according to preset business priority labels to obtain power feature data of different priorities; uses a pre-built redundant detection fusion model to perform redundancy processing on power feature data of different priorities to obtain corresponding de-redundant power feature data; performs secondary classification on de-redundant power feature data according to power demand characteristics to obtain demand feature data of different demand types; performs learning optimization on demand feature data of different demand types to obtain an optimal power allocation strategy; based on the optimal power allocation strategy, uses a computing power chip scheduling algorithm to perform dynamic heterogeneous collaborative scheduling on multi-computing power resources. Compared with the prior art, the method uses a redundant detection fusion model to perform redundancy processing on power feature data, and schedules chips according to the optimal power allocation strategy. The allocation of computing power resources is optimized through reasonable computing power allocation and scheduling strategies, and the chip utilization efficiency is maximized, providing strong support for the development of smart grids. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a schematic diagram of a process flow of a multi-computing power integrated scheduling method for a smart grid provided by an embodiment of the present invention;

[0057] Figure 2 It is a schematic diagram of the multi-computing power integrated scheduling process provided by an embodiment of the present invention;

[0058] Figure 3 is a redundant data processing flow chart provided by an embodiment of the present invention;

[0059] Figure 4 It is a schematic diagram of the algorithm chip scheduling process provided by an embodiment of the present invention;

[0060] Figure 5 is a schematic diagram of an algorithm core particle module provided by an embodiment of the present invention;

[0061] Figure 6 It is a schematic diagram of the integrated packaging of algorithm core particles provided by an embodiment of the present invention;

[0062] Figure 7 It is a block diagram of a multi-computing power integrated dispatching system for smart grids provided by an embodiment of the present invention;

[0063] Figure 8 It is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0064] The following specifically illustrates the implementation mode of the present invention in conjunction with the accompanying drawings. The embodiments are provided for illustrative purposes only and cannot be understood as limiting the present invention. The accompanying drawings are provided for reference and illustration only and do not constitute a limitation on the scope of patent protection of the present invention, because many changes can be made to the present invention without departing from the spirit and scope of the present invention.

[0065] refer to Figure 1 The embodiment of the present invention provides a multi-computing power integrated scheduling method for smart grid, which is applied to a multi-chip integrated packaging structure, such as Figure 1 As shown, the multi-computing power integrated scheduling method includes the following steps:

[0066] S1. Collect power data sets of different business scenarios of the smart grid, and prioritize the power data sets according to preset business priority tags to obtain power feature data of different priorities.

[0067] This embodiment takes the multi-computing power and multi-service fusion control chip in the smart grid as an example. The chip is deployed in the data center of the smart grid. It is mainly responsible for processing power data from different business fields. The power data includes various grid attributes, such as voltage, current, power, etc. Figure 2 As shown, this embodiment allocates and schedules computing power according to the priority and demand characteristics of the data. All input power data will be assigned two labels, namely, a priority label and a demand characteristic label, to obtain a power data set containing a priority label and a demand characteristic label. For example, this embodiment divides the data into high priority and low priority according to the importance of power facilities and data characteristics. Among them, high-priority data may include power consumption data of important public service facilities and infrastructure such as hospitals, fire protection facilities, key facilities of government departments, transportation hubs, communication facilities, water supply and drainage facilities. These data have a large fluctuation range and are not easy to capture features, so they are directly processed without training and modeling; low-priority data includes power consumption data of normal residents and enterprises. These data are regular and can be trained through algorithms to improve the efficiency of computing power chip scheduling.

[0068] S2. Use the pre-built redundant detection fusion model to perform redundancy removal processing on the power feature data of different priorities to obtain corresponding redundancy removal power feature data.

[0069] In this embodiment, the redundant detection fusion model is constructed by fusing genetic algorithm and particle swarm algorithm. The steps of performing redundancy processing on power feature data of different priorities using the pre-constructed redundant detection fusion model to obtain corresponding de-redundant power feature data include:

[0070] An initial particle group including a plurality of initial particles is randomly generated; each initial particle represents power characteristic data;

[0071] According to the relative positions and quantity distributions of the power characteristic data of different priorities, a fitness function is constructed; the fitness function is used to evaluate the redundancy of the current particles;

[0072] Use the fitness function to preliminarily evaluate the initial particle swarm, calculate the fitness value of each initial particle, and select high-quality solutions based on the fitness value of each initial particle;

[0073] Selecting a number of pairs of parent particles from the high-quality solutions, and performing crossover operations on the parent particles using a genetic algorithm according to extreme values ​​of the power characteristic data to generate new child particles;

[0074] The redundancy of the new generation particles is evaluated twice according to the fitness function, the fitness value of each new generation particle is calculated, and the redundant data points are determined according to the fitness value of each new generation particle;

[0075] Deleting the redundant data points from the new offspring particles or replacing them with non-redundant data points to generate an offspring individual population after the redundant data are removed;

[0076] A mutation operation is performed on the offspring individual population after the redundant data is removed, and the particle group is iteratively updated until a maximum number of iterations is reached, and the redundant power characteristic data is output.

[0077] In the field of smart grid, with the continuous increase of power data, redundant data not only occupies database space, but also affects the speed and efficiency of data processing. Therefore, this embodiment is based on a redundant detection fusion model that integrates genetic algorithm and particle swarm algorithm to achieve effective de-redundancy processing of power distribution data of different priorities and improve the detection accuracy of redundant data. Figure 3 As shown, the specific implementation process is:

[0078] The position and speed of the particle group are randomly initialized, and each particle represents a power characteristic number. In this embodiment, an initial particle group including multiple initial particles is obtained by random initialization, and then a fitness function is constructed according to the relative position and quantity distribution between power characteristic data of different priorities. The fitness function is used to evaluate the redundancy of each particle (i.e., each power characteristic data). The redundancy of the initial particle group is preliminarily evaluated using the fitness function, and the fitness value of each particle is calculated. According to the fitness value, particles with lower redundancy are screened out as high-quality solutions. In this embodiment, initial particles with redundancy lower than a preset redundancy threshold can be screened out as high-quality solutions according to the fitness value of each initial particle.

[0079] Several pairs of parent particles are selected from the selected high-quality solutions. According to the extreme values ​​in the data group (for example, the highest and lowest particles can be selected as parents according to the ranking of fitness values), a genetic algorithm is used to perform a crossover operation on the parent particles. The crossover operation generates new child particles by exchanging some features of the parent particles. The specific mode of the crossover operation can be single-point crossover, multi-point crossover, etc. The new child particles obtained after the crossover operation are subjected to a secondary determination of redundant data. Specifically, the fitness value of each new child particle is calculated according to the fitness function, and the redundant data points are determined according to the fitness value. If there are redundant data points, It deletes or replaces other non-redundant data points from the new offspring particles using the nearest neighbor replacement method to generate a population of offspring individuals after removing redundant data, and performs a mutation operation on the offspring individual population after removing redundant data. The specific method of the mutation operation can be to randomly change certain properties of the particles (such as the location of the cluster center, the distribution of data points, etc.) to increase the diversity of the population and improve the possibility of finding the global optimal solution. Repeat the above steps (i.e., crossover operation, secondary determination and removal of redundant data, mutation operation), and iteratively update the particle group. At the same time, it is determined whether the current number of iterations reaches the preset maximum number of iterations. If not, the iteration process continues; if it is reached, the iteration process is terminated and the redundant power characteristic data is output.

[0080] In the redundant data processing process, the classified power feature data contains repeated or similar grid attribute values. The grid attributes represent the basic characteristics of the grid operation, such as voltage, current, power, etc. The differences between these attributes can reflect the redundancy of the data. In this embodiment, the grid attributes are extracted from the classified power feature data, and a fitness function is constructed based on the extracted grid attributes. By calculating the distance between any two grid attributes, that is, the fitness value, redundant data can be identified, that is, those data points that are very close or repeated in attribute characteristics. Therefore, in this embodiment, the distance between grid attributes is used to measure redundancy. Specifically, in this embodiment, a fitness function is constructed based on the grid attribute characteristics of the power feature data. The fitness function is used to measure the distance between any two grid attributes, that is, the similarity or difference between them. In this embodiment, the fitness function f adopts the following formula:

[0081]

[0082] In the formula, d(x i ,x j ) are any two grid attributes x i With x j The distance between them is the fitness value; For x i Completely transformed into x j The distance (or difference) of For xj Completely transformed into x i Distances (or differences) can be obtained by calculating the difference between attributes, Euclidean distance or other suitable measurement methods. Those skilled in the art can quantify the degree of redundancy by calculating indicators such as the degree of repetition, distance measurement or clustering density between data points according to the specific implementation situation.

[0083] This embodiment achieves high-precision detection and removal of redundant data in the smart grid by integrating the genetic algorithm and the standard particle swarm algorithm. The genetic algorithm searches for the optimal solution by simulating natural selection and genetic mechanisms, while the particle swarm algorithm searches for the global optimal solution by simulating the foraging behavior of bird flocks. In this embodiment, operations such as crossover and mutation of the genetic algorithm are combined with search mechanisms such as speed update and position update of the particle swarm algorithm to form a fusion algorithm. The fusion algorithm is used to search the grid attributes of the power characteristic data, find and mark the redundant data, so as to improve the detection accuracy of the redundant data. According to the calculation result of the fitness function, the grid attributes with high similarity are marked as redundant data. According to the marking result of the redundant data, the redundant data is deleted. At the same time, in order to ensure the accuracy of data removal, the results of automatic removal are verified by manual verification. For possible misjudgment, manual confirmation and adjustment can be performed.

[0084] Through the above steps, the smart grid in this embodiment can realize accurate detection and efficient removal of redundant data. The redundant data processing scheme based on the fusion of genetic algorithm and standard particle swarm algorithm not only improves the detection accuracy of redundant data, but also ensures the accuracy of data removal through the combination of automatic and manual methods, which not only saves database space, but also improves the efficiency of data processing, and provides strong support for the intelligent management and efficient operation of smart grids. In summary, through classification and priority sorting, the data in this embodiment can enter the processing flow in an orderly manner according to its importance and priority, which significantly reduces the data processing waiting time. At the same time, the space occupancy and subsequent processing pressure are reduced through redundant data processing, making the entire data processing system more efficient. This orderly processing method not only improves the speed of data processing, but also reduces the overall burden of the system.

[0085] S3. Perform secondary classification on the de-redundant power characteristic data according to the power demand characteristics to obtain demand characteristic data of different demand types.

[0086] Different power grid services have different characteristics in their demands for computing power, such as real-time demands, safety demands, and basic demands. The scheduling method needs to be able to identify these demands and make reasonable allocation of computing power based on their characteristics. For demand characteristic labels, this embodiment can further classify the data into de-redundant power characteristic data containing demand characteristic labels based on the real-time, safety, and basic demands of the electricity consumption data, and obtain demand characteristic data of different demand types. The demand types include real-time demand types, safety demand types, and basic demand types. For example, real-time demand data may include electricity consumption data of financial trading companies, medical industry companies, etc., which have high real-time requirements; safety demand data may include electricity consumption data of heavy industrial enterprises such as construction sites and the steel industry. The electricity consumption data has a small variation range, but has high stability requirements; basic demand data may include electricity consumption data of residential buildings, schools, and commercial stores, which have no special requirements for real-time and stability.

[0087] S4. Perform learning and optimization on demand characteristic data of different demand types to obtain an optimal power allocation strategy; the demand types include real-time demand types, safety demand types and basic demand types.

[0088] In this embodiment, when the demand characteristic data is redundancy-free power characteristic data of a real-time demand type, the step of learning and optimizing the demand characteristic data of different demand types to obtain an optimal power allocation strategy includes:

[0089] According to the demand characteristic data, obtaining local load consumption data of each distributed generator;

[0090] According to the local load consumption data of each distributed generator, the load increment of each distributed generator is calculated by using the first-order discrete time dynamic average consistency algorithm;

[0091] The state of each distributed generator at different times is updated according to the load increment of each distributed generator, and the average grid power of all distributed generators is obtained through iterative calculation;

[0092] Inputting the average grid power into a pre-trained deep neural network model to obtain an optimal power solution for each distributed generator;

[0093] The local power supply and demand deviation of each distributed generator is calculated according to the optimal power solution of each distributed generator, the local power output and the local real-time load demand;

[0094] Utilizing a supply-demand consistency algorithm to optimize and update the local power supply-demand deviation to obtain a global power supply-demand deviation;

[0095] The optimal power solution of each distributed generator is adjusted according to the global power supply and demand deviation until a balance is reached, and an optimal power allocation strategy is output.

[0096] When the demand characteristic data is a real-time demand type de-redundant power characteristic data, this embodiment adopts the learning optimal solution algorithm to obtain the optimal power allocation scheme. Specifically, this embodiment uses the local load consumption data r of each distributed generator to obtain the optimal power allocation scheme. i (t), where i represents the number of the distributed generator and t represents time. The load increment Δr of each distributed generator is calculated by the first-order discrete-time dynamic average consistency (DAC) algorithm. i (t) = r i (t)-r i (th), where h is the time discretization unit. Then, the state x of each distributed generator at time t is updated according to the DAC protocol. i (t), that is:

[0097]

[0098] In the formula, a ij is the interaction gain between distributed generators.

[0099] In this embodiment, the state of each distributed generator at different times is updated according to the load increment of each distributed generator, and the average grid power of all distributed generators is obtained by iterative calculation, and its mathematical expression is:

[0100]

[0101] The average grid power is used as input data and input into the trained deep neural network (DNN) model. The DNN model outputs the optimal power solution for each distributed generator through calculations of the internal neural network layer based on the input data. i and local real-time load demand L i Perform optimization processing of the supply and demand consistency algorithm (BGD) to calculate the local power supply and demand deviation Δ of each distributed generator i [0] = L i -O i Then, the local power supply and demand deviation is updated according to the BGD algorithm to obtain the global power supply and demand deviation, that is:

[0102]

[0103] Where ζ is the step size parameter; Δ i [n] is the local power supply and demand deviation in the i-th distributed generator bus; n is the distributed generator index.

[0104] In this embodiment, the local power supply and demand deviation is iteratively updated to obtain the global power supply and demand deviation, and its mathematical expression is:

[0105]

[0106] Finally, the optimal power solution of each distributed generator is adjusted according to the global power supply and demand deviation to obtain the final power output P i =O i +δ, and limit it between Pimin and Pimax, generate the optimal power allocation strategy according to the final power output Pi, and distribute and control power to the power system. Through the above steps, this embodiment realizes the iterative optimization of the demand characteristic data to obtain the optimal power allocation strategy. This method combines the advantages of the DAC algorithm, DNN prediction and BGD algorithm, and has the characteristics of strong real-time performance and good optimality.

[0107] In this embodiment, when the demand characteristic data is de-redundant power characteristic data of a safety demand type, the step of learning and optimizing the demand characteristic data of different demand types to obtain an optimal power allocation strategy includes:

[0108] Using a multi-layer perceptron to learn and predict a set of potential state points from the demand feature data;

[0109] Select one or more initial state variables from the potential state point set as initial state candidate points of the Gauss-Newton iterative algorithm;

[0110] Sorting the initial state candidate points, and selecting the initial state candidate point with the highest sorting as the hot start point of the Gauss-Newton iterative algorithm;

[0111] The number of iterations is set, and in each iteration of the Gauss-Newton iterative algorithm, the difference between the current state point and the target state point is analyzed according to the Jacobian matrix of the hot start point to obtain an error vector;

[0112] Construct the Hessian matrix according to the Jacobian matrix, solve the incremental equation according to the Hessian matrix and the error vector, and obtain the state update;

[0113] The hot start point is updated according to the state update amount and iteratively optimized until the maximum number of iterations is reached, the final hot start point is output as the optimal state point, and the optimal power allocation strategy is determined according to the optimal state point.

[0114] When the demand characteristic data is redundant power characteristic data of the safety demand type, the present embodiment adopts a learning hot start algorithm to obtain an optimal power allocation scheme. Specifically, the present embodiment adopts a multilayer perceptron as a neural network (NN) of the learning hot start algorithm, and uses the demand characteristic data as the input of the multilayer perceptron (MLP). The multilayer perceptron learns the mapping relationship between data of different demand types and power allocation strategies, and uses the multilayer perceptron to learn and predict points near the potential state to form a potential state point set. One or more initial state variables are selected from the potential state point set as initial state candidate points of the Gauss-Newton iterative algorithm. The initial state candidate points are sorted according to factors such as data priority and importance, and the initial state point with the highest sorting is selected as the hot start point of the Gauss-Newton iterative algorithm. The maximum number of iterations is set according to the hot start point to prevent the iterative process from falling into an infinite loop. In each iteration of the Gauss-Newton iterative algorithm, the Jacobian matrix J(x k ), where x k is the hot start point for the current iteration number k; the error vector f(x k ), the error vector reflects the difference between the current state point and the target state point, and the Hessian matrix is ​​constructed by multiplying the transpose of the Jacobian matrix with itself. The incremental equation is solved according to the Hessian matrix and the error vector to obtain the state update amount, which represents the direction of movement from the current state point to a state point closer to the target state point. The hot start point is iteratively updated using the state update amount to obtain a new hot start point, and the above steps are repeated for iterative optimization until the maximum number of iterations is reached or the preset convergence conditions are met. During the iteration process, the hot start point will gradually approach the optimal solution. When the iteration process ends, the final hot start point is output as the optimal state point, and the optimal power allocation value of each node is determined according to the optimal state point, thereby obtaining the optimal power allocation strategy.

[0115] In this embodiment, when the demand characteristic data is de-redundant power characteristic data of a basic demand type, the step of learning and optimizing the demand characteristic data of different demand types to obtain an optimal power allocation strategy includes:

[0116] Identifying uncertainty factors in power distribution according to the demand characteristic data, and establishing a mapping relationship between the uncertainty factors and power distribution constraints;

[0117] According to the mapping relationship, the power distribution constraint is learned and identified by using support vector machine learning to obtain a probability distribution of the power distribution constraint;

[0118] Determine active constraints under the current power distribution state according to the probability distribution of power distribution constraints to form an active constraint set;

[0119] The gradient descent method is used to iteratively calculate the active constraint set under the current power distribution state to solve the optimal power allocation strategy under nonlinear constraints.

[0120] When the demand characteristic data is the redundant power characteristic data of the basic demand type, this embodiment adopts a learning constraint algorithm to obtain the optimal power distribution plan. Specifically, the uncertainty factors in power distribution, such as power fluctuation, are identified according to the processed demand characteristic data, and the mapping relationship between these uncertainty factors and power distribution constraints is established through statistical analysis, data mining and other methods. The power distribution constraints are learned and identified using a support vector machine (SVM). By training the SVM model, different constraints in power distribution are identified and classified, such as voltage stability constraints, power balance constraints, line capacity constraints, or physical constraints, safety constraints, and priority requirements of the power grid are identified. According to the learning results of the SVM, The probability distribution of power distribution constraints is calculated, and the active constraints under the current power distribution state are analyzed according to the probability distribution of power distribution constraints. Active constraints refer to the constraints that play a major role in the current power distribution. The identified active constraints constitute an active constraint set, that is, a set of constraints that affect the optimal solution. The active constraint set under the current power distribution state is iteratively calculated using the gradient descent method. By continuously adjusting the power distribution plan, the power distribution reaches the optimal state while satisfying all active constraints. In the iterative calculation process, the nonlinear constraints of power distribution are considered, such as nonlinear load characteristics, nonlinear losses of transmission lines, etc. The optimal power distribution strategy is obtained by solving the optimal solution under the nonlinear constraints.

[0121] In summary, this embodiment allocates the most suitable learning optimization algorithm to data of different demand types, which can ensure that each type of data can be processed in the best way, thereby improving the accuracy of computing resource scheduling and ensuring the stability and efficiency of smart grid operation.

[0122] S5. Based on the optimal power allocation strategy, the computing chip scheduling algorithm is used to dynamically coordinate the heterogeneous scheduling of multiple computing resources.

[0123] In this embodiment, the steps of dynamically scheduling heterogeneous collaborative computing resources using a computing chip scheduling algorithm based on an optimal power allocation strategy include:

[0124] In response to the unit matrix test task instruction, the unit matrix test task is respectively run on the algorithm chip of each computing node to obtain the task test time consumption of the algorithm chip of each computing node; the algorithm chip includes at least a central processing unit and a graphics processing unit;

[0125] According to the task test time of the algorithm chip of each computing node, a cluster computing capacity evaluation index is obtained; the cluster computing capacity evaluation index includes the computing capacity ratio of each algorithm chip in the node, the comprehensive computing capacity ratio of each computing node and the comprehensive computing capacity of the cluster per unit time;

[0126] Determine the computing tasks to be scheduled, and divide the computing tasks to be scheduled into fine-grained units based on the cluster computing capacity evaluation index using the unit matrix as a unit to obtain several minimum task units;

[0127] According to the optimal power allocation strategy, the optimal computing resources are dynamically scheduled for each minimum task unit in the cluster.

[0128] Specifically, Figure 4 As shown, in this embodiment, the unit matrix test task is respectively run on the algorithm chip (including CPU and GPU) of each computing node, and the time consumed by the algorithm chip of each computing node to complete the unit matrix test task is collected and recorded to obtain the task test time consumption, wherein the algorithm chip includes CPU and GPU; according to the task test time consumption of the algorithm chip of each computing node, the computing power ratio of each algorithm chip in the computing node is obtained, which can be achieved by comparing the time taken by different chips to complete the same task; then, the computing power ratios of all algorithm chips in each computing node are integrated to obtain the comprehensive computing power ratio of each computing node, which can be obtained by weighted averaging the computing power ratios of all algorithm chips in the node or other suitable calculation methods; then, the comprehensive computing power ratios of each computing node are summarized to obtain the comprehensive computing power of the cluster per unit time, for example, the comprehensive computing power ratio of each computing node can be multiplied by the corresponding number of nodes to obtain the cluster unit time. The computing tasks to be scheduled are determined based on the comprehensive computing power within the time. According to the cluster computing power evaluation index, the computing tasks to be scheduled are finely divided with the unit matrix as the unit. This means splitting the large task into multiple smaller and easy-to-manage task units. Each task unit can be regarded as a minimum task unit. According to the optimal power allocation strategy and the node computing power evaluation information, the optimal computing power resources are dynamically scheduled for each minimum task unit in the cluster. The computing power resources include not only computing nodes, but also specific algorithm chips (such as CPU or GPU), memory, storage and other resources in the nodes. For example, this embodiment can allocate nodes with sufficient computing power to tasks based on the comprehensive computing power ratio of each computing node and the current power allocation situation, while considering the impact of power allocation on node performance, to ensure that the computing power of the node can be fully utilized when the task is allocated, while maintaining low energy consumption. After the task is scheduled to the corresponding computing power resource, the computing task is started.

[0129] At the same time, this embodiment establishes a data cache queue and a data processing queue on each computing node to ensure the concurrency of data transmission and data processing, avoid idle node resources, perform dynamic task scheduling according to the determined processing capacity of each algorithm chip, complete node-level collaborative scheduling, monitor the power consumption of the node in real time, and dynamically adjust task allocation as needed to ensure that it does not exceed the preset power threshold, thereby realizing dynamic computing power resource scheduling based on the optimal power allocation strategy. In summary, this embodiment schedules chips according to the size and type of data to obtain a reasonable computing power allocation. For example, for high-priority data or real-time demand data, more computing power resources are allocated to ensure fast processing; for low-priority data or basic demand data, fewer computing power resources are allocated to reduce costs, so that high-priority data and real-time demand data can be quickly processed, providing strong support for the intelligent management and efficient operation of smart grids.

[0130] In the data processing flow provided in this embodiment, in order to reduce the computational burden caused by repeated calls to redundant data removal algorithms and learning optimization algorithms and to improve the communication efficiency within the chip, Figure 5 , Figure 6 As shown, this embodiment can modularize these algorithm functions and design them into independent small chiplets to form a multi-chiplet integrated packaging structure. Specifically, in the multi-chiplet integrated packaging structure, this embodiment repeatedly calls the redundant data removal algorithm and the learning optimization algorithm. In order to improve the communication efficiency within the chip, this embodiment designs these algorithms into independent small chiplets. These small chiplets can be modules with specific functions such as data processing and algorithm optimization. The chiplets are integrated by designing efficient communication interfaces and protocols to form a complete processing unit. This embodiment can consider using high-speed serial interfaces (such as PCIe, CXL, etc.) or dedicated on-chip networks (NoC) to realize data transmission between chiplets to reduce communication delays and power consumption.

[0131] In order to further improve the integration and performance of the system, this embodiment considers using 3D FO (Fan-Out) multi-chip integrated packaging technology. According to the size and layout requirements of the algorithm core particles, the designed algorithm core particles are embedded in resin, silicon substrate or other carriers, reconstructed into a standard-sized wafer, and the interconnection between the chips is realized through the rewiring (RDL) method. In which, the metal copper pillar is used as a vertical interconnection structure to connect the core particles of different layers to realize the vertical interconnection between the core particles of different layers. The chip with the bump is flipped to the lower chip and connected to the rewiring layer of the lower chip through the bump. At the same time, the metal copper pillar is used to realize the electrical connection between the upper and lower core particles. In the interconnection technology of the core particles, this embodiment can use the through silicon via (TSV) technology to form a vertical direction by filling the silicon hole with conductive materials such as tungsten, copper, polysilicon, etc. The electrical interconnection on the substrate effectively reduces the size of the package and improves the electrical performance. TSV technology is one of the key technologies for three-dimensional integration. It allows high-density integration and low-power operation, and is very suitable for high-performance computing and portable electronic products. Copper pillar bump technology is used to realize chip substrate interconnection of integrated circuit packaging. It has superior electrical performance, thermal performance and reliability, and can meet RoHS requirements. This technology can reduce the number of substrate layers and reduce the overall packaging cost. Compared with wire bonding, its overall packaging cost can be saved by about 20%. The flip chip (FC) packaging process mainly includes steps such as bump metallization, bump manufacturing, wafer assembly and bottom filling. This process can achieve high-density chip interconnection and has the best thermal and electrical performance and the highest reliability.

[0132] The embodiment of the present invention provides a multi-computing power integrated scheduling method for smart grids. The method collects power data sets of different business scenarios of the smart grid, and prioritizes the power data sets according to preset business priority labels to obtain power feature data of different priorities; uses a pre-built redundant detection fusion model to perform redundancy processing on the power feature data of different priorities to obtain corresponding de-redundant power feature data; performs secondary classification on the de-redundant power feature data according to power demand characteristics to obtain demand feature data of different demand types; performs learning optimization on the demand feature data of different demand types to obtain an optimal power allocation strategy; the demand types include real-time demand types, security demand types and basic demand types; based on the optimal power allocation strategy, uses a computing power chip scheduling algorithm to perform dynamic heterogeneous collaborative scheduling of multi-computing power resources. Compared with the existing technology, this method uses a redundant detection fusion model to perform redundancy processing on power feature data, and integrates multiple computing resources to dynamically schedule computing resources for chips according to the optimal power allocation strategy, ensuring that computing resources are reasonably allocated and significantly improving the utilization efficiency of computing resources, thereby maximizing the efficiency of chip use. At the same time, it can be flexibly adjusted and optimized according to the needs of different businesses, which can meet the high requirements of smart grids for real-time and accuracy and improve the intelligent management level of smart grids.

[0133] It should be noted that the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0134] In one embodiment, Figure 7 As shown, an embodiment of the present invention provides a multi-computing power integrated dispatching system for a smart grid, the system comprising:

[0135] The data acquisition module 101 is used to collect power data sets of different business scenarios of the smart grid, and prioritize the power data sets according to preset business priority tags to obtain power feature data of different priorities;

[0136] The redundancy detection module 102 is used to perform redundancy removal processing on power feature data of different priorities using a pre-built redundancy detection fusion model to obtain corresponding redundancy-removed power feature data;

[0137] A demand classification module 103, used for performing secondary classification on the de-redundant power characteristic data according to power demand characteristics to obtain demand characteristic data of different demand types;

[0138] The power allocation module 104 is used to learn and optimize the demand characteristic data of different demand types to obtain an optimal power allocation strategy; the demand types include real-time demand types, safety demand types and basic demand types;

[0139] The dynamic scheduling module 105 is used to perform dynamic heterogeneous collaborative scheduling of multiple computing resources based on the optimal power allocation strategy and using the computing chip scheduling algorithm.

[0140] For the specific definition of a multi-computing power integrated scheduling system for smart grids, please refer to the above-mentioned definition of a multi-computing power integrated scheduling method for smart grids, which will not be repeated here. A person of ordinary skill in the art will appreciate that the various modules and steps described in conjunction with the embodiments disclosed in this application can be implemented in hardware, software, or a combination of both. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0141] The embodiment of the present invention provides a multi-computing power integrated scheduling system for smart grids. The system collects power data sets of different business scenarios of smart grids through a data acquisition module, and prioritizes the power data sets according to preset business priority labels to obtain power feature data of different priorities; the redundancy detection module uses a pre-built redundancy detection fusion model to perform redundancy processing on power feature data of different priorities to obtain corresponding de-redundant power feature data; the demand classification module performs secondary classification on the de-redundant power feature data according to power demand characteristics to obtain demand feature data of different demand types; the power allocation module performs learning and optimization on the demand feature data of different demand types to obtain an optimal power allocation strategy; the demand types include real-time demand types, safety demand types and basic demand types; the dynamic scheduling module uses a computing power chip scheduling algorithm based on the optimal power allocation strategy to perform dynamic heterogeneous collaborative scheduling of multi-computing power resources. Compared with the existing technology, this system uses a redundant detection fusion model to perform redundancy processing on power feature data, and integrates multiple computing resources to dynamically schedule computing resources for chips according to the optimal power allocation strategy, ensuring that computing resources are reasonably allocated and significantly improving the utilization efficiency of computing resources, thereby maximizing the efficiency of chip use. At the same time, it can be flexibly adjusted and optimized according to the needs of different businesses, which can meet the high requirements of smart grids for real-time and accuracy and improve the intelligent management level of smart grids.

[0142] Figure 8A computer device provided by an embodiment of the present invention includes a memory, a processor and a transceiver, which are connected via a bus; the memory is used to store a set of computer program instructions and data, and can transmit the stored data to the processor, and the processor can execute the program instructions stored in the memory to perform the steps of the above method.

[0143] The memory may include a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories; the processor may be a central processing unit, a microprocessor, an application-specific integrated circuit, a programmable logic device, or a combination thereof. By way of example but not limitation, the programmable logic device may be a complex programmable logic device, a field programmable gate array, a general purpose array logic, or any combination thereof.

[0144] Additionally, the memory may be a physically separate unit or may be integrated with the processor.

[0145] It can be understood by those skilled in the art that Figure 8 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have the same component arrangement.

[0146] The above-mentioned embodiments only express several preferred implementation modes of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in the technical field, several improvements and substitutions can be made without departing from the technical principles of the present invention, and these improvements and substitutions should also be regarded as the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be based on the protection scope of the claims.

Claims

1. A multi-computing power integrated scheduling method for smart grid, characterized in that: The multi-computing power integrated scheduling method comprises the following steps: Collecting power data sets of different business scenarios of the smart grid, and prioritizing the power data sets according to preset business priority tags to obtain power feature data of different priorities; Using the pre-built redundant detection fusion model to perform redundancy processing on the power feature data of different priorities, and obtain the corresponding redundancy-free power feature data; Secondarily classifying the de-redundant power characteristic data according to power demand characteristics to obtain demand characteristic data of different demand types; Learning and optimizing demand characteristic data of different demand types to obtain an optimal power allocation strategy; the demand types include real-time demand types, safety demand types, and basic demand types; Based on the optimal power allocation strategy, the computing chip scheduling algorithm is used to perform dynamic heterogeneous collaborative scheduling of multiple computing resources.

2. A multi-computing power integrated scheduling method for smart grid according to claim 1, characterized in that: The redundant detection fusion model is constructed by fusing a genetic algorithm and a particle swarm algorithm. The steps of performing redundancy processing on power feature data of different priorities using the pre-constructed redundant detection fusion model to obtain corresponding de-redundant power feature data include: An initial particle group including a plurality of initial particles is randomly generated; each initial particle represents power characteristic data; According to the relative positions and quantity distributions of the power characteristic data of different priorities, a fitness function is constructed; the fitness function is used to evaluate the redundancy of the current particles; Use the fitness function to preliminarily evaluate the initial particle swarm, calculate the fitness value of each initial particle, and select high-quality solutions based on the fitness value of each initial particle; Selecting a number of pairs of parent particles from the high-quality solutions, and performing a crossover operation on the parent particles using a genetic algorithm according to the extreme values ​​of the power characteristic data to generate new child particles; The redundancy of the new generation particles is evaluated twice according to the fitness function, the fitness value of each new generation particle is calculated, and the redundant data points are determined according to the fitness value of each new generation particle; Deleting the redundant data points from the new offspring particles or replacing them with non-redundant data points to generate an offspring individual population after the redundant data are removed; A mutation operation is performed on the offspring individual population after the redundant data is removed, and the particle group is iteratively updated until a maximum number of iterations is reached, and the redundant power characteristic data is output.

3. A multi-computing power integrated scheduling method for smart grid as claimed in claim 2, characterized in that: The step of selecting a high-quality solution according to the fitness value of each initial particle comprises: According to the fitness value of each initial particle, the initial particles with redundancy lower than the preset redundancy threshold are selected as high-quality solutions.

4. A multi-computing power integrated scheduling method for smart grid according to claim 1, characterized in that: When the demand characteristic data is redundancy-free power characteristic data of a real-time demand type, the step of learning and optimizing the demand characteristic data of different demand types to obtain an optimal power allocation strategy includes: According to the demand characteristic data, obtaining local load consumption data of each distributed generator; According to the local load consumption data of each distributed generator, the load increment of each distributed generator is calculated by using the first-order discrete time dynamic average consistency algorithm; The state of each distributed generator at different times is updated according to the load increment of each distributed generator, and the average grid power of all distributed generators is obtained through iterative calculation; Inputting the average grid power into a pre-trained deep neural network model to obtain an optimal power solution for each distributed generator; The local power supply and demand deviation of each distributed generator is calculated according to the optimal power solution of each distributed generator, the local power output and the local real-time load demand; Utilizing a supply-demand consistency algorithm to optimize and update the local power supply-demand deviation to obtain a global power supply-demand deviation; The optimal power solution of each distributed generator is adjusted according to the global power supply and demand deviation until a balance is reached, and an optimal power allocation strategy is output.

5. The method for integrated scheduling of multiple computing power for smart grid according to claim 1, characterized in that: When the demand characteristic data is redundant power characteristic data of a safety demand type, the step of learning and optimizing the demand characteristic data of different demand types to obtain an optimal power allocation strategy includes: Using a multi-layer perceptron to learn and predict a set of potential state points from the demand feature data; Select one or more initial state variables from the potential state point set as initial state candidate points of the Gauss-Newton iterative algorithm; Sorting the initial state candidate points, and selecting the initial state candidate point with the highest sorting as the hot start point of the Gauss-Newton iterative algorithm; The number of iterations is set, and in each iteration of the Gauss-Newton iterative algorithm, the difference between the current state point and the target state point is analyzed according to the Jacobian matrix of the hot start point to obtain an error vector; Construct the Hessian matrix according to the Jacobian matrix, solve the incremental equation according to the Hessian matrix and the error vector, and obtain the state update; The hot start point is updated according to the state update amount and iteratively optimized until the maximum number of iterations is reached, the final hot start point is output as the optimal state point, and the optimal power allocation strategy is determined according to the optimal state point.

6. A multi-computing power integrated scheduling method for smart grid according to claim 1, characterized in that: When the demand characteristic data is de-redundant power characteristic data of a basic demand type, the step of learning and optimizing the demand characteristic data of different demand types to obtain an optimal power allocation strategy includes: Identifying uncertainty factors in power distribution according to the demand characteristic data, and establishing a mapping relationship between the uncertainty factors and power distribution constraints; According to the mapping relationship, the power distribution constraint is learned and identified by using support vector machine learning to obtain a probability distribution of the power distribution constraint; Determine active constraints under the current power distribution state according to the probability distribution of power distribution constraints to form an active constraint set; The gradient descent method is used to iteratively calculate the active constraint set under the current power distribution state to solve the optimal power allocation strategy under nonlinear constraints.

7. A multi-computing power integrated scheduling method for smart grid according to claim 1, characterized in that: The steps of dynamically scheduling heterogeneous collaborative computing resources using a computing chip scheduling algorithm based on an optimal power allocation strategy include: In response to the unit matrix test task instruction, the unit matrix test task is respectively run on the algorithm chip of each computing node to obtain the task test time consumption of the algorithm chip of each computing node; the algorithm chip includes at least a central processing unit and a graphics processing unit; According to the task test time of each computing node algorithm chip, the cluster computing capacity evaluation index is obtained; Determine the computing tasks to be scheduled, and divide the computing tasks to be scheduled into fine-grained units based on the cluster computing capacity evaluation index using the unit matrix as a unit to obtain several minimum task units; According to the optimal power allocation strategy, the optimal computing resources are dynamically scheduled for each minimum task unit in the cluster.

8. A multi-computing power integrated scheduling method for smart grid according to claim 7, characterized in that: The cluster computing capability evaluation index includes the computing capability ratio of each algorithm chip in the node, the comprehensive computing capability ratio of each computing node, and the comprehensive computing capability of the cluster per unit time.

9. A multi-computing power integrated dispatching system for smart grid, characterized in that: The system comprises: A data acquisition module is used to collect power data sets of different business scenarios of the smart grid, and prioritize the power data sets according to preset business priority tags to obtain power feature data of different priorities; A redundancy detection module, used to perform redundancy removal processing on power feature data of different priorities using a pre-built redundancy detection fusion model to obtain corresponding redundancy-free power feature data; A demand classification module, used for secondary classification of the de-redundant power characteristic data according to power demand characteristics to obtain demand characteristic data of different demand types; A power allocation module is used to learn and optimize the demand characteristic data of different demand types to obtain an optimal power allocation strategy; the demand types include real-time demand types, safety demand types and basic demand types; The dynamic scheduling module is used to dynamically coordinate heterogeneous scheduling of multiple computing resources based on the optimal power allocation strategy and the computing chip scheduling algorithm.

10. A computer device, characterized in that: The computer device comprises a processor and a memory, wherein the processor is connected to the memory, the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the computer device executes the method according to any one of claims 1 to 8.

Images