Intelligent regulation and balance optimization method and system for new energy cluster grid connection

Abstract

The application provides a new energy cluster grid-connected intelligent regulation and balance optimization method and system, and belongs to the field of power system intelligent dispatching and multi-agent collaborative control. The method comprises the following steps: obtaining new energy cluster and power grid dispatching data to construct global state and local observation; establishing a hierarchical time sequence system evolution model, generating intention control quantity based on local observation through an Actor network, and converting the intention control quantity into an execution control quantity through tensor interval projection correction; performing dispatching allocation according to the execution control quantity to obtain grid-connected power and power distribution; constructing a gradient constraint correction loss based on the deviation degree of the intention control quantity, and constructing a power balance regulation and control loss based on the power concentration degree; updating Actor-Critic network parameters in a centralized training and decentralized execution framework to output a collaborative control strategy. The application synchronously improves operation data compliance, power grid dispatching stability and grid-connected consumption balance, and is suitable for new energy multi-cluster distributed grid-connected dispatching scenes.

Description

Technical Field

[0001] This invention belongs to the field of intelligent dispatching and multi-agent cooperative control technology of power systems, and particularly relates to an intelligent control and balance optimization method and system for new energy cluster grid connection. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] With the continuous advancement of the construction of new power systems, wind power, photovoltaic and other new energy power generation units are being connected to the grid on a large scale. The main body of new energy has transformed from traditional passive power generation equipment to intelligent and controllable clusters with autonomous regulation capabilities. In the real-time dispatch scenario of the power system, each new energy cluster needs to dynamically optimize the grid-connected power allocation strategy based on its own adjustable output, the overall grid load demand, compliance operation boundaries, and the behavior of neighboring clusters. Affected by meteorological conditions, wind and solar power output exhibits strong fluctuations and temporal uncertainties, forming complex dynamic coupling relationships between clusters. Although wind and solar resources have spatiotemporal complementary characteristics, and multi-cluster collaboration can mitigate power fluctuations, each cluster has independent operating constraints and differentiated objectives, which cannot be solved by traditional single-entity static optimization. Therefore, there is an urgent need for refined regulation methods that involve dynamic interaction and temporal evolution among multiple entities.

[0004] To ensure the safe operation of the system, the existing dispatching system has established standardized data compliance boundaries to suppress abnormal control outputs. However, in the absence of a robust dynamic boundary correction mechanism, renewable energy sources are prone to continuously generating data exceeding limits, causing grid-connected power to deviate from the allowable range. Long-term operation of this can lead to frequent adjustments to dispatching instructions, increased curtailment rates, and excessive power concentration, severely impacting system stability. Furthermore, renewable energy grid connection exhibits typical Markov time-series evolution characteristics; current control actions influence subsequent cycles through historical states and power distribution. Traditional static optimization cannot characterize this dynamic feedback, making it difficult to meet the control needs of highly volatile power grids.

[0005] Existing control methods can be mainly divided into three categories. Rule-based control generates instructions based on fixed thresholds, which is simple to implement but cannot adapt to random fluctuations and multi-cluster coupling, resulting in poor adaptability. Static modeling can quantify the power interaction between clusters, but it severs the temporal correlation between time periods, ignores the transmission effect of historical policies on long-term stability, and cannot match the dynamic evolution characteristics. Multi-agent reinforcement learning has dynamic adaptive capabilities, but existing schemes only perform post-cutoff correction at the output stage, failing to deeply embed data compliance constraints, collaborative absorption incentives, and power balance control into the entire training process. This leads to the continuous generation of out-of-bounds control variables during iteration, resulting in insufficient absorption stability and policy robustness. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, this invention provides an intelligent regulation and equalization optimization method and system for grid connection of new energy clusters. It aims to solve the problem of the integration of data compliance constraints, multi-cluster collaborative absorption, time-series dynamic feedback and power equalization control in the existing technology, and overcome the technical defects of poor regulation data compliance, uneven absorption of new energy, weak grid operation stability and poor convergence effect of intelligent strategy.

[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides an intelligent control and balance optimization method for grid connection of new energy clusters; Intelligent control and equilibrium optimization methods for grid connection of new energy clusters include: Acquire data on the collection of new energy clusters and power grid dispatching, and construct a global system status and local observations of each new energy cluster; Based on the upper and lower limits of operating data, upper-level coordination variables, and lower-level cluster action space, a hierarchical time-series power system state evolution model is established to provide a state transition framework for multi-cluster collaborative decision-making. Each new energy cluster outputs an intention control quantity through the Actor strategy network based on local observations, and projects the intention control quantity into an execution control quantity that satisfies the upper and lower limits of the running data and the reporting ratio constraints through tensor interval projection correction. Based on the joint execution control quantities of multiple new energy clusters, grid load demand, available output and upper-level coordination variables, the scheduling and allocation calculation is completed to obtain grid-connected power, grid-connected power, power distribution and real-time control performance indicators of each cluster; The gradient constraint correction loss is constructed based on the degree to which the intended control quantity deviates from the legal range, and the power balance regulation loss is constructed based on the grid-connected power concentration. Under the framework of centralized training and decentralized execution, the parameters of the Actor network and Critic network are updated, and the grid-connected collaborative control strategy of each new energy cluster is output.

[0008] As a further technical solution, data on the collection of new energy clusters and power grid dispatch are acquired to construct a global system state and local observations of each new energy cluster, including: Obtain the set of new energy clusters participating in power grid dispatch control and the set of discrete dispatch periods; To construct the global system state, obtain the available power output, grid load demand, lower and upper limits of operating data for each renewable energy cluster during each scheduling period, as well as the historical grid-connected power, historical power distribution share, and upper-level scheduling coordination variables from the previous period.

[0009] in, Indicates the global state of the system. Indicates the time period of each new energy cluster Available output Indicates the power grid load demand. Indicates time period The set of lower limit boundaries for the grid-connected power of each new energy cluster. Represents the corresponding upper bound set, and for any cluster All meet , This indicates the grid-connected electricity volume in the previous time period. This indicates the power distribution in the previous time period. This represents the upper-level coordination variables from the previous time period; And construct local observations for each new energy cluster:

[0010] in, Indicates local observation. Indicates the i-th cluster during the time period Available output; This represents the grid-connected power consumption of the i-th cluster in the previous time period; This represents the power distribution of the i-th cluster in the previous time period.

[0011] As a further technical solution, based on the upper and lower limits of operational data, upper-level coordination variables, and lower-level cluster action space, a hierarchical time-series power system state evolution model is established to provide a state transition framework for multi-cluster collaborative decision-making, including: The control process of grid connection of multiple new energy clusters is modeled as a hierarchical time-series system evolution model consisting of an exogenous system constraint layer, an upper-level scheduling and coordination layer, and a lower-level new energy cluster. Establish state transition relationships, where the system state for the next time period is determined by the current state, upper-level coordination variables, and lower-level joint execution control variables.

[0012] As a further technical solution, each new energy cluster outputs an intention control quantity through an Actor policy network based on local observations, and projects the intention control quantity into an execution control quantity that satisfies the upper and lower limits of the operating data and the reporting ratio constraints through tensor interval projection correction. Specifically, this includes: Each new energy cluster inputs its local observations into the Actor policy network, which maps and generates intention control quantities, including intention grid-connected power and intention reporting ratio. Tensor interval projection correction is performed on the intent control quantity. By taking the maximum lower limit and the minimum upper limit of the interval projection operation, the actual execution control quantity and execution reporting ratio participating in the scheduling allocation are obtained. The grid-connected power reported by each cluster is calculated based on the product of the reported ratio and the available output. The reported grid-connected power is then compared with the grid-connected power command obtained after projection correction to obtain the candidate grid-connected power. The execution of grid-connected power commands, execution reporting ratios, reported grid-connected power, and candidate grid-connected power from each new energy cluster are aggregated to form a joint schedule, which is used for subsequent scheduling and allocation calculations.

[0013] As a further technical solution, scheduling and allocation calculations are performed based on the joint execution control quantities of multiple new energy clusters, grid load demand, available output, and upper-level coordination variables. This yields grid-connected power, grid-connected power, power distribution, and real-time control performance indicators for each cluster, including: Receive the joint execution control input from all new energy clusters, determine the grid connection priority based on the grid connection power, and sort the clusters accordingly; For each new energy cluster, grid connection control performance indicators are determined based on grid-connected power and grid-connected electricity to reflect the cluster's contribution to the grid supply and demand balance. The system operating loss is calculated based on the product of the unit grid-connected power loss coefficient and the grid-connected power of each cluster; the collaborative benefit allocation coefficient is calculated based on the contribution weight of each cluster in grid-connected collaboration; and the data boundary violation penalty term is constructed based on the intention control quantity to define the data boundary violation degree. Based on grid-connected control performance indicators, system operating losses, grid-connected collaborative benefits, data out-of-bounds penalty items, and power balance regulation items, the stage control performance indicators of each cluster are obtained.

[0014] As a further technical solution, the gradient constraint correction loss is:

[0015] in, This is the gradient constraint correction loss; Represents a cluster During the period The degree of violation of the intent to control quantity; For training batches; The power equalization control loss is:

[0016] in, For power equalization control losses; For the first A new energy cluster during the period The squared deviation of the actual grid-connected power share from the average power share of the same period.

[0017] As a further technical solution, the parameters of the Actor network and Critic network are updated under the framework of centralized training and distributed execution, and the grid-connected collaborative control strategies of each new energy cluster are output, including: A centralized training and distributed execution framework is adopted. During the training phase, the Critic network uses the global state to evaluate the value of each cluster and assists in the joint optimization of multiple clusters. During the execution phase, each cluster generates control variables independently based only on its own local observations, which meets the requirements of distributed control.

[0018] During training, the temporal difference residuals are calculated based on control reward and state value estimation, and the advantage function is calculated using generalized advantage estimation. The Critic network is updated based on the mean square error between the system reward objective and the value estimate; the Actor network is based on the PPO pruning objective and integrates the entropy regularization term, gradient constraint correction loss and power balance regulation loss to form a comprehensive objective for policy update. By iterating until the maximum number of training rounds is reached or the convergence condition is met, the collaborative control strategy of each cluster after training is output.

[0019] The second aspect of this invention provides an intelligent control and balance optimization system for grid connection of new energy clusters.

[0020] Intelligent control and balance optimization system for grid connection of new energy clusters includes: The data input module is configured to: acquire the collection of new energy clusters and power grid dispatch data, and construct the global status of the system and the local observations of each new energy cluster; The evolution modeling module is configured to: establish a hierarchical time-series power system state evolution model based on the upper and lower limits of the operating data, upper-level coordination variables, and lower-level cluster action space, providing a state transition framework for multi-cluster collaborative decision-making; The tensor projection correction module is configured such that: each new energy cluster outputs an intention control quantity through the Actor strategy network based on local observations, and projects the intention control quantity into an execution control quantity that satisfies the upper and lower limits of the running data and the reporting ratio constraints through tensor interval projection correction; The scheduling and allocation calculation module is configured to perform scheduling and allocation calculations based on the joint execution control quantities of multiple new energy clusters, grid load demand, available output and upper-level coordination variables, and obtain grid-connected power, grid-connected power, power distribution and real-time control performance indicators of each cluster. The gradient constraint and power equalization module is configured to: construct gradient constraint correction loss based on the degree to which the intended control quantity deviates from the legal range, and construct power equalization regulation loss based on the grid-connected power concentration. The collaborative control training and strategy output module is configured to update the parameters of the Actor network and Critic network under the framework of centralized training and decentralized execution, and output the grid-connected collaborative control strategy for each new energy cluster.

[0021] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of the intelligent control and equalization optimization method for grid connection of new energy clusters as described in the first aspect of the present invention.

[0022] The fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the intelligent regulation and equalization optimization method for grid connection of new energy clusters as described in the first aspect of the present invention.

[0023] The above one or more technical solutions have the following beneficial effects: (1) This invention constructs a hierarchical time-series power system state evolution model, which simultaneously incorporates the system constraint layer, the upper-level scheduling and coordination unit and the lower-level multiple new energy clusters. It realizes the coupled modeling of four mechanisms: control data compliance boundary constraints, multi-cluster grid connection collaborative absorption incentives, multi-subject power collaborative interaction, and Markov time-series system state feedback. It breaks the limitations of traditional models that can only depict power interaction or static time series, and fully depicts the objective physical laws of dynamic power fluctuations and random evolution of new energy output in the power grid. It greatly improves the completeness and accuracy of the time-series state characterization of high-fluctuation new energy power grids.

[0024] (2) This invention distinguishes between the control quantity to be corrected in the original output of the neural network and the compliant execution control quantity in the access scheduling system. It uses tensor interval projection correction to perform interval constraint processing on the network output data to ensure that the output control parameters input to the scheduling platform meet the upper and lower limits of the system data throughout the process. At the same time, it embeds the gradient constraint correction mechanism into the back propagation training process of the neural network to suppress the output of out-of-limit data from the root of weight iteration, improves the convergence stability of the intelligent adjustment strategy, and solves the inherent technical defects of existing algorithms that can only achieve surface compliance by post-truncation and training nodes continuously generate abnormal out-of-limit data.

[0025] (3) The present invention embeds a grid-connected collaborative regulation reward term in the multi-agent reward function, making full use of the spatiotemporal power complementarity of wind power, photovoltaic and wind-solar hybrid units, guiding each new energy cluster to coordinate and regulate grid-connected power output, smoothing overall power fluctuations across the entire domain, improving the level of new energy consumption from the perspective of regulation mechanism, and enhancing the balance of long-term grid operation.

[0026] (4) This invention introduces a multi-entity grid-connected power equalization control mechanism, constructs an equalization control loss term based on the power concentration index, and continuously optimizes the uniformity of power distribution among multiple clusters during the strategy iteration optimization process, avoiding excessive concentration of grid-connected power in a few new energy clusters during long-term scheduling; effectively reducing the risk of local grid power oscillation, so that the control strategy learned by the agent can simultaneously take into account the stability of new energy consumption, grid connection security and the power equalization operation requirements of the whole network.

[0027] (5) The present invention adopts a constrained MAPPO collaborative control framework with centralized training and decentralized execution. During the training phase, multi-agent joint optimization is carried out based on the global power grid operation status, which greatly improves the convergence effect and operating condition robustness of the multi-cluster collaborative strategy. During the online control phase, each new energy cluster can independently complete the grid-connected power output control by relying only on local observation information, which is suitable for new power system distributed intelligent dispatch application scenarios.

[0028] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0029] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0030] Figure 1 This is a flowchart of the method in the first embodiment.

[0031] Figure 2 This is a diagram of the hierarchical time-series power system state evolution control architecture for the first embodiment.

[0032] Figure 3 This is a centralized training-distributed execution framework for the constrained MAPPO power grid cooperative control algorithm in the first embodiment.

[0033] Figure 4 This is a system structure diagram of the second embodiment. Detailed Implementation

[0034] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0035] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0036] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0037] Example 1 This embodiment discloses an intelligent regulation and equalization optimization method for grid connection of new energy clusters. It constructs a hierarchical time-series state evolution model to characterize the dynamic coupling relationship of multiple clusters, uses tensor interval projection correction to ensure the compliance of scheduling data, introduces gradient constraint correction to suppress data out-of-bounds from the training root, and combines power equalization regulation and collaborative absorption reward to achieve complementary output of multiple clusters and power equalization optimization of the whole network.

[0038] like Figure 1 As shown, the intelligent control and equilibrium optimization method for grid connection of new energy clusters includes: Step S1: Obtain the set of new energy clusters and power grid dispatch data to construct the global state of the system and the local observation of each new energy cluster.

[0039] Obtain the set of new energy clusters participating in power grid dispatch and control:

[0040] in, Represents a collection of new energy clusters. This indicates the number of new energy clusters. Each new energy cluster represents a wind power cluster, a photovoltaic cluster, a wind-solar hybrid cluster, or a regional new energy aggregation unit. The system operates according to discrete scheduling time periods, and the set of scheduling time periods is as follows:

[0041] in, Represents the set of scheduling periods. This indicates the total number of scheduling periods.

[0042] To construct the global system state, obtain the available power output, grid load demand, lower and upper limits of operating data for each renewable energy cluster during each scheduling period, as well as the historical grid-connected power, historical power distribution share, and upper-level scheduling coordination variables from the previous period.

[0043] in, Indicates the global state of the system. Indicates the time period of each new energy cluster Available output Indicates the power grid load demand. Indicates time period The set of lower limit boundaries for the grid-connected power of each new energy cluster. Represents the corresponding upper bound set, and for any cluster All meet , This indicates the grid-connected electricity volume in the previous time period. This indicates the power distribution in the previous time period. This represents the upper-level coordination variables of the previous period, including the grid-connected coordination benefit allocation coefficient and the system power balance adjustment weight, etc.

[0044] And construct local observations for each new energy cluster:

[0045] in, Indicates local observation. Indicates the i-th cluster during the time period Available output; This represents the grid-connected power consumption of the i-th cluster in the previous time period; This represents the power distribution of the i-th cluster in the previous time period.

[0046] This local observation includes the cluster's own available output, grid load demand, data compliance boundaries, its own historical grid connection performance, and publicly available coordination information, but does not include the complete policy parameters and local private information of other clusters, in order to meet the requirements of localized independent decision-making in distributed grid connection control.

[0047] Step S2: Based on the upper and lower limits of the operating data, the upper-level coordination variables, and the lower-level cluster action space, establish a hierarchical time-series power system state evolution model to provide a state transition framework for multi-cluster collaborative decision-making.

[0048] S2.1, as Figure 2 As shown, the grid connection control process of multiple new energy clusters is modeled as a hierarchical time-series system evolution model consisting of an exogenous system constraint layer, an upper-level scheduling and coordination layer, and a lower-level new energy cluster.

[0049] The exogenous system constraint layer is used to define the lower limit boundary of the operating data of each new energy cluster within a given scheduling period t. and upper limit boundary This serves as the legal physical range for the control output of the entire system, preventing scheduling instructions from becoming invalid due to out-of-bounds input data.

[0050] The upper-level scheduling and coordination layer is used to set variables for grid-connected coordination, power allocation, and system balance adjustment. Coordinated decision-making is carried out with the overall optimization objectives of system operation stability, balance of new energy consumption, curtailment rate and power concentration.

[0051] The lower-level new energy clusters are used to, under the constraints of the data compliance boundaries given by the exogenous system constraint layer and the coordination variables output by the upper-level scheduling and coordination layer, based on the local observation information of each cluster. The grid-connected power and reporting ratio are independently selected, and distributed decision-making is carried out with the optimization goal of achieving optimal long-term control performance.

[0052] S2.2, construct the state transition dependency relationship between adjacent scheduling periods, so that the system state of the next period is jointly determined by the system state of the current period, the upper-level coordination variables, and the joint execution control variables of all lower-level new energy clusters:

[0053] in, Indicates the lower-level joint execution control quantity. This represents the state transition probability kernel. This relationship is used to characterize the exogenous changes in renewable energy output and grid load, as well as the endogenous feedback of dispatch allocation results on historical grid-connected power and power distribution, reflecting the natural physical laws of dynamic temporal fluctuations in the power grid.

[0054] S2.3 The upper-level scheduling and coordination layer takes system operation stability, renewable energy consumption balance, curtailment rate and power concentration as comprehensive objectives, and adjusts the grid connection coordination benefit allocation and system power balance weight; the lower-level renewable energy cluster takes optimal long-term control performance as its objective, and carries out compliance control under data boundary and scheduling allocation constraints.

[0055] In step S3, each new energy cluster outputs an intention control quantity through the Actor strategy network based on local observations, and projects the intention control quantity into an execution control quantity that satisfies the upper and lower limits of the running data and the reporting ratio constraints through tensor interval projection correction.

[0056] S3.1, during training and execution, the first A new energy cluster based on local observation The intention control quantity is output through the Actor policy network:

[0057] in, Indicates the intended grid-connected power. This indicates the proportion of intent reported. The intent control quantity reflects the raw output of the policy network and may have a tendency to exceed the upper and lower limits of the operational data. The intent control quantity is the raw output of the policy network without any compliance correction processing, directly reflecting the policy network's true control tendency under the current local observation, and may exceed the lower limit of the operational data. and upper limit boundary The tendency to overstep boundaries.

[0058] S3.2, Perform neural network tensor interval projection correction on the intended grid-connected power. By performing interval projection operations with the lower limit at the maximum and the upper limit at the minimum, the intended grid-connected power is compressed to the lower limit boundary of the operating data. and upper limit boundary Within the defined legal range, the actual execution control quantity entering the scheduling allocation is obtained. :

[0059] This correction technique ensures that the executed control quantity is always within the legal range of system data, avoiding invalid scheduling instructions caused by out-of-bounds input data to the scheduling system. The meaning of this projection correction calculation is: if the intended grid-connected power is lower than the lower limit boundary... Then it will be upgraded to If the intended grid-connected power exceeds the upper limit boundary. Then reduce it to If the intended grid-connected power is within the legal range, it remains unchanged. Through the above processing, it is ensured that the executed grid-connected power is always within the legal range of system data, avoiding invalid scheduling instructions caused by out-of-bounds input data to the scheduling system.

[0060] S3.3 performs tensor interval projection correction on the intention reporting ratio. By using interval projection operations with the lower limit at its maximum and the upper limit at its minimum, the intention reporting ratio is constrained to a legal ratio range of 0 to 1, thus obtaining the execution reporting ratio. :

[0061] The projection correction ensures that the reporting ratio is strictly within the range of [0,1], which conforms to the actual physical meaning of the reporting ratio.

[0062] S3.4, Calculate the first [unit / item] based on the execution reporting ratio and the available output of the corresponding new energy cluster. A new energy cluster during the period Reported grid-connected power:

[0063] in, Indicates the first The upper limit of the power that can participate in scheduling for each new energy cluster is determined based on the current available output and the execution reporting ratio; Indicates the reporting ratio; Indicates the first A new energy cluster during the period Available output. The grid-connected power command obtained in step S3.2 Comparison to obtain candidate grid-connected power :

[0064] S3.5, the executed grid-connected power command, executed reporting ratio, reported grid-connected power, and candidate grid-connected power of all new energy clusters are aggregated to form a joint scheduling input and sent to the scheduling and allocation module. In step S4, the grid-connected power command is executed. Used to determine priority, candidate grid-connected power Used to limit the upper limit of the actual grid-connected power.

[0065] In the above technical solution, the intent control variable retains the original output features of the policy network and is used to calculate the data out-of-bounds penalty in subsequent gradient constraint correction training; the execution control variable, after projection correction, is used for actual scheduling and allocation calculation. The two types of control variables function independently and jointly participate in the iterative optimization process of the policy network: the intent control variable provides the policy network with unconstrained gradient update directions, while the execution control variable ensures the compliance of the system input data. Together, they achieve the dual technical effect of "correcting out-of-bounds tendencies at the root during the training phase and ensuring data compliance during the execution phase."

[0066] Step S4: Based on the joint execution control quantities of multiple new energy clusters, grid load demand, available output, and upper-level coordination variables, the scheduling and allocation calculation is completed to obtain grid-connected power, grid-connected power, power distribution, and real-time control performance indicators of each cluster.

[0067] S4.1, the scheduling and allocation module receives the joint scheduling input generated in step S3. Among them, For step S3.2, the power intended for grid connection The grid-connected power command obtained after tensor interval projection correction is used to characterize the first... The scheduling requirements of each new energy cluster are determined and the sorting priority is determined. This does not represent the actual grid-connected power after dispatching is completed. The actual grid-connected power is denoted as... The scheduling and allocation process includes: (1) Execute grid-connected power command Reported grid-connected power Determine the candidate grid-connected power for each new energy cluster:

[0068] (2) Execute grid-connected power command Sort the new energy clusters from largest to smallest to obtain the sorting mapping. ,in Indicates time period Sort the first A new energy cluster.

[0069] (3) In terms of grid load demand Under the constraints, the actual grid-connected power of each new energy cluster is determined sequentially according to the ranking results:

[0070] in, This is used to ensure that the remaining acceptable power is not less than zero.

[0071] (4) Based on the actual grid-connected power, the grid-connected power, clustered aggregated grid-connected power, and power distribution share are formed:

[0072]

[0073]

[0074] in, Indicates the duration of a single scheduling period; Indicates the first A new energy cluster during the period The actual grid-connected power generation; Indicates time period The aggregated grid-connected power of all new energy clusters; Indicates the first The actual grid-connected power of each new energy cluster accounts for the proportion of the aggregated grid-connected power.

[0075] S4.2, for the first A new energy cluster, grid connection control performance indicators Determined by the unified grid-connected power and grid-connected electricity, it reflects the contribution of the cluster to the grid's supply and demand balance. The larger the grid-connected electricity and the higher the unified grid-connected power across the entire grid, the better the grid-connected control performance of the cluster.

[0076] in, Indicates the grid-connected power during a given time period; Indicates the first A new energy cluster during the period The grid-connected power generation.

[0077] S4.3, System Operating Costs Determined by the unit loss factor and the grid-connected power consumption, it reflects the physical losses during grid-connected power transmission.

[0078] in, Indicates the first The unit grid-connected power loss coefficient of each cluster is determined by a combination of the grid connection distance, transmission line parameters, and operating conditions of each new energy cluster.

[0079] S4.4, Allocate grid-connection synergy benefits based on the contribution of each cluster to system absorption, and calculate the synergy benefit allocation coefficient. This is used to incentivize complementary outputs from multiple clusters and smooth out overall grid connection fluctuations.

[0080] in, Indicates the first The contribution weight of each cluster in grid-connected coordination can be determined by a combination of incremental grid-connected power, output smoothing contribution, and the degree of reporting participation. Indicates the first A new energy cluster during the period The contribution weight of grid connection collaboration To sum the subscripts; This indicates the total number of new energy clusters participating in the current scheduling; This indicates that all new energy clusters are in the same time period. The sum of contribution weights; To prevent extremely small constants with a denominator of zero.

[0081] S4.5 defines the degree of data boundary breach based on the intention control variable rather than the execution control variable. Specifically, the intention grid-connected power is calculated separately. Below the lower limit of the running data boundary The degree of downward deviation And the intended grid-connected power exceeds the upper limit of the operating data boundary. degree of deviation :

[0082] Accordingly, the data out-of-bounds penalty term is defined as follows:

[0083] in, Indicates the first A new energy cluster during the period The degree of violation of the intended grid connection power. >0 represents the data out-of-bounds penalty coefficient. This setting avoids the problem that the out-of-bounds penalty is always zero after the control variable is projected through hard constraints, ensuring that gradient constraint correction training techniques can effectively apply to policy network weight updates.

[0084] S4.6, based on grid-connected control performance indicators, system operating losses, grid-connected collaborative benefits, data out-of-bounds penalty terms, and power balance regulation terms, obtains the stage control performance indicators for each cluster. :

[0085] in, This represents the total benefit generated by the grid connection and coordination mechanism. The data out-of-bounds penalty coefficient, This is the system equilibrium adjustment coefficient. Indicates the first A new energy cluster during the period The share of grid-connected power, This represents the average power distribution.

[0086] Step S5: Construct gradient constraint correction loss based on the degree to which the intended control quantity deviates from the legal range, and construct power balance regulation loss based on the grid-connected power concentration.

[0087] S5.1, in training batches Calculate the average gradient constraint correction loss This is used to encourage the policy network to reduce out-of-bounds data output.

[0088] in, Represents a cluster During the period The degree of violation of the intention control quantity is defined as the sum of the squares of the intention control quantity being below the lower boundary and above the upper boundary. This loss is applied to the Actor update process, providing additional gradient constraint signals during the backpropagation phase of the neural network. This guides the policy network to internalize the data compliance boundary during the learning phase, reducing the dependence on the execution layer projection correction and improving the stability of the control output from the root.

[0089] S5.2, based on the grid-connected power share of each new energy cluster in time period t. Calculate the grid-connected power concentration index to measure whether grid-connected capacity is excessively concentrated in a few clusters:

[0090] in, This indicates the concentration of grid-connected power in time period t. The higher the power concentration, the greater the risk of local power fluctuations in the power grid. The lower the concentration, the more evenly the grid-connected power is distributed among the clusters, and the more stable the overall operation of the power grid.

[0091] S5.3 further constructs grid-connected power balancing control loss This is used to provide additional constraint signals during strategy updates, suppress concentrated fluctuations in the output of new energy clusters, and reduce the risk of local power oscillations in the power grid.

[0092] in, For power equalization control losses; For the first A new energy cluster during the period The squared deviation of the power share relative to the average power share during the same period; this item does not require all clusters to have completely equal grid-connected power, but is used to prevent a few clusters from occupying an excessively high grid-connected share for a long time, and to improve the overall supply and demand balance of the network.

[0093] The gradient constraint correction loss and power equalization control loss Together, they participate in the policy update process of the Actor network, working synergistically with the PPO policy optimization objective and entropy regularization term in subsequent steps to form a comprehensive objective function. This achieves integrated optimization that combines improved control performance with data compliance and power balance constraints. Specifically, the gradient constraint correction loss constrains the policy vertically, while the power balance regulation loss constrains the policy horizontally. The two work together to improve the overall performance of multi-cluster grid-connected regulation.

[0094] Step S6: Update the parameters of the Actor network and Critic network under the centralized training and decentralized execution framework, and output the grid-connected collaborative control strategy for each new energy cluster.

[0095] S6.1 adopts a centralized training-distributed execution framework, such as... Figure 3 As shown, during the training phase, the Actor network of each new energy cluster i generates the intended control quantity based on its own local observations, while the Critic network uses the global state of the system to evaluate the long-term control performance of each cluster. During training, the Critic network can acquire the state information and control actions of all clusters to estimate the expected cumulative reward of cluster i in the global state, thus providing the Actor network with a more accurate policy gradient direction. This effectively alleviates the non-stationarity problem in multi-agent environments and significantly improves the convergence effect and robustness of multi-cluster cooperative strategies.

[0096] During the execution phase, each new energy cluster relies solely on its own local observations to independently input the Actor network to generate control quantities. After tensor interval projection correction, these control quantities are converted into compliant execution control quantities and participate in scheduling and allocation. During execution, there is no need to obtain policy parameters, private observation information, or the global system state from other clusters, fully meeting the requirements of distributed grid-connected control for localized independent decision-making.

[0097] S6.2 Calculate the TD residual based on control rewards and state value estimates. :

[0098] in, Indicates the first A new energy cluster during the period Single-step timing difference residual; This indicates that the cluster is in the time period The immediate control reward gained; This represents the discount factor for future returns, and ; and These represent the system's global state for the current time period and the next time period, respectively. and The parameters are respectively The The Critic network provides value estimates for the current global state and the next global state; Indicates the first The trainable parameters of a Critic network. When When in a terminated state, let .like This indicates that the actual reward generated by the current action is higher than the original estimate of the Critic network; if This indicates that the actual return generated by the current action is lower than the original estimate.

[0099] Based on the generalized advantage estimation (GAE), the TD residuals of multiple future time periods are discounted and weighted to accumulate, and the advantage function is calculated. :

[0100] in, This represents the GAE parameter, used to make a trade-off between estimation bias and variance. Indicates relative to the current time period The future step size offset; This indicates the end time period of the current training trajectory.

[0101] This advantage estimate applies a discounted weighted accumulation of TD residuals over multiple future periods, which helps improve the stability of multi-agent policy updates.

[0102] S6.3, the Critic network uses the mean square error between the system reward objective and the value estimate for updates, improving the accuracy of value assessment of the power grid dynamic state. :

[0103] in, This represents the critic's estimate of the value of the current global state; For Critic network parameters, The goal is to achieve system returns. By minimizing this mean square error, the Critic network progressively learns to accurately assess the long-term expected returns under different global states, providing a more reliable value benchmark for the Actor network.

[0104] In S6.4, the Actor network uses the PPO pruning objective as a basis and integrates the entropy regularization term, gradient constraint correction loss, and power balance regulation loss to form a comprehensive objective for policy update.

[0105] First, regarding the first For each cluster, define the probability ratio between the old and new strategies. :

[0106] in, Indicates the current strategy. This indicates the old strategy before the update. Represents a cluster Local observations Indicates the first A new energy cluster based on local observation during time period t The intention control quantity is sampled and recorded by the old Actor policy before the update, including the intention grid connection power and the intention reporting ratio, and has not yet undergone tensor interval projection correction. This probability characterizes the change in the selection probability of the same action before and after the policy update.

[0107] PPO trimming target The construction is as follows:

[0108] in, This represents the PPO pruning factor. This pruning operation limits the policy update magnitude, preventing policy collapse caused by excessively large single-step updates, thereby reducing fluctuations during the training process of the grid-connected control policy and improving learning stability.

[0109] Meanwhile, to prevent the strategy from prematurely converging to a single grid-connected control mode, an entropy regularization term is introduced:

[0110] in, This represents the policy distribution entropy. The entropy regularization term encourages the agent to maintain the necessary exploration capabilities, reducing the risk of getting trapped in local optima during training.

[0111] S6.5 combines the PPO pruning objective, entropy regularization term, gradient constraint correction loss, and power equalization regulation loss into a constrained Actor objective:

[0112] in, For entropy regularization weights, The loss weights are corrected for gradient constraints. The weights for power balancing are used to adjust the loss. This objective function incorporates four types of learning signals: control performance improvement, exploration-maintain, data compliance, and system power balancing, achieving multi-dimensional collaborative optimization.

[0113] The Actor network maximizes this comprehensive objective function. The strategy is updated. Specifically, the gradient ascent method is used to update the network parameters along the gradient direction of the objective function. During backpropagation, the gradient constraint correction loss generates a penalty gradient, which drives the policy network weights to be updated in the direction of reducing out-of-bounds output; the power equalization adjustment loss generates an equalization gradient, which drives the output distribution of each cluster to tend to be uniform.

[0114] S6.6, repeat steps S3 (generating the intentional control quantity and correcting the tensor interval projection), S4 (scheduling allocation and calculating control performance), S5 (constructing the loss function), and steps S6.1 to S6.5 (updating the Critic network and Actor network) until the maximum number of training rounds is reached or the policy convergence condition is met. Once the convergence condition is met, save the trained Actor policy parameters for each new energy cluster as the grid-connected collaborative control strategy.

[0115] In step S6.7, during the actual scheduling execution phase, the Actor policy networks trained in step S6.6 are deployed to the local controllers of the corresponding renewable energy clusters. Each renewable energy cluster generates its intended grid-connected power and intended reporting ratio based solely on its own local observations input to the Actor network. After tensor interval projection correction in step S3, an execution control quantity that satisfies the upper and lower limits of the operating data is formed. The execution control quantity is input into the scheduling allocation module in step S4 to participate in the overall network scheduling allocation calculation. The scheduling allocation results are then fed back as the historical grid-connected power and power distribution in the next time period, enabling the system to continuously adapt to renewable energy output fluctuations, grid load changes, and multi-cluster control strategy interactions during multi-time period scheduling, ensuring the long-term operational stability of the power grid.

[0116] The technical solution provided in this embodiment achieves efficient convergence of multi-cluster collaborative strategies through global information fusion in the centralized training phase, and satisfies the engineering constraints of communication bandwidth and privacy protection in actual power grid distributed control through local observation and independent decision-making in the decentralized execution phase, thus achieving an organic unity between training performance and deployment feasibility.

[0117] Example 2 This embodiment discloses an intelligent control and balance optimization system for grid connection of new energy clusters; like Figure 4 As shown, the intelligent control and balance optimization system for grid connection of new energy clusters includes: The data input module is configured to: acquire the collection of new energy clusters and power grid dispatch data, and construct the global status of the system and the local observations of each new energy cluster; The evolution modeling module is configured to: establish a hierarchical time-series power system state evolution model based on the upper and lower limits of the operating data, upper-level coordination variables, and lower-level cluster action space, providing a state transition framework for multi-cluster collaborative decision-making; The tensor projection correction module is configured such that: each new energy cluster outputs an intention control quantity through the Actor strategy network based on local observations, and projects the intention control quantity into an execution control quantity that satisfies the upper and lower limits of the running data and the reporting ratio constraints through tensor interval projection correction; The scheduling and allocation calculation module is configured to perform scheduling and allocation calculations based on the joint execution control quantities of multiple new energy clusters, grid load demand, available output and upper-level coordination variables, and obtain grid-connected power, grid-connected power, power distribution and real-time control performance indicators of each cluster. The gradient constraint and power equalization module is configured to: construct gradient constraint correction loss based on the degree to which the intended control quantity deviates from the legal range, and construct power equalization regulation loss based on the grid-connected power concentration. The collaborative control training and strategy output module is configured to update the parameters of the Actor network and Critic network under the framework of centralized training and decentralized execution, and output the grid-connected collaborative control strategy for each new energy cluster.

[0118] Example 3 The purpose of this embodiment is to provide a computer-readable storage medium.

[0119] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the intelligent control and balance optimization method for grid connection of new energy clusters as described in Embodiment 1.

[0120] Example 4 The purpose of this embodiment is to provide an electronic device.

[0121] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the intelligent control and equalization optimization method for grid connection of new energy clusters as described in Embodiment 1.

[0122] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0123] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0124] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for intelligent regulation and equilibrium optimization for grid connection of new energy clusters, characterized in that, include: Acquire data on the collection of new energy clusters and power grid dispatching, and construct a global system status and local observations of each new energy cluster; Based on the upper and lower limits of operating data, upper-level coordination variables, and lower-level cluster action space, a hierarchical time-series power system state evolution model is established to provide a state transition framework for multi-cluster collaborative decision-making. Each new energy cluster outputs an intention control quantity through the Actor strategy network based on local observations, and projects the intention control quantity into an execution control quantity that satisfies the upper and lower limits of the running data and the reporting ratio constraints through tensor interval projection correction. Based on the joint execution control quantities of multiple new energy clusters, grid load demand, available output and upper-level coordination variables, the scheduling and allocation calculation is completed to obtain grid-connected power, grid-connected power, power distribution and real-time control performance indicators of each cluster; The gradient constraint correction loss is constructed based on the degree to which the intended control quantity deviates from the legal range, and the power balance regulation loss is constructed based on the grid-connected power concentration. Under the framework of centralized training and decentralized execution, the parameters of the Actor network and Critic network are updated, and the grid-connected collaborative control strategy of each new energy cluster is output.

2. The intelligent regulation and equilibrium optimization method for new energy cluster grid connection as described in claim 1, characterized in that, Acquire data from the collection of new energy clusters and power grid dispatching to construct a global system state and local observations of each new energy cluster, including: Obtain the set of new energy clusters participating in power grid dispatch control and the set of discrete dispatch periods; To construct the global system state, obtain the available power output, grid load demand, lower and upper limits of operating data for each renewable energy cluster during each scheduling period, as well as the historical grid-connected power, historical power distribution share, and upper-level scheduling coordination variables from the previous period. in, Indicates the global state of the system. Indicates the time period of each new energy cluster Available output Indicates the power grid load demand. Indicates time period The set of lower limit boundaries for the grid-connected power of each new energy cluster. Represents the corresponding upper bound set, and for any cluster All meet , This indicates the grid-connected electricity volume in the previous time period. This indicates the power distribution in the previous time period. This represents the upper-level coordination variables from the previous time period; And construct local observations for each new energy cluster: in, Indicates local observation. Indicates the i-th cluster during the time period Available output; This represents the grid-connected power consumption of the i-th cluster in the previous time period; This represents the power distribution of the i-th cluster in the previous time period.

3. The intelligent regulation and equilibrium optimization method for new energy cluster grid connection as described in claim 1, characterized in that, Based on the upper and lower limits of operational data, upper-level coordination variables, and lower-level cluster action space, a hierarchical time-series power system state evolution model is established to provide a state transition framework for multi-cluster collaborative decision-making, including: The control process of grid connection of multiple new energy clusters is modeled as a hierarchical time-series system evolution model consisting of an exogenous system constraint layer, an upper-level scheduling and coordination layer, and a lower-level new energy cluster. Establish state transition relationships, where the system state for the next time period is determined by the current state, upper-level coordination variables, and lower-level joint execution control variables.

4. The intelligent regulation and equilibrium optimization method for new energy cluster grid connection as described in claim 1, characterized in that, Each new energy cluster outputs an intention control quantity through an Actor policy network based on local observations, and projects this intention control quantity into an execution control quantity that satisfies the upper and lower limits of the operating data and the reporting ratio constraints through tensor interval projection correction. Specifically, this includes: Each new energy cluster inputs its local observations into the Actor policy network, which maps and generates intention control quantities, including intention grid-connected power and intention reporting ratio. Tensor interval projection correction is performed on the intent control quantity. By taking the maximum lower limit and the minimum upper limit of the interval projection operation, the actual execution control quantity and execution reporting ratio participating in the scheduling allocation are obtained. The grid-connected power reported by each cluster is calculated based on the product of the reported ratio and the available output. The reported grid-connected power is then compared with the grid-connected power command obtained after projection correction to obtain the candidate grid-connected power. The execution of grid-connected power commands, execution reporting ratios, reported grid-connected power, and candidate grid-connected power from each new energy cluster are aggregated to form a joint schedule, which is used for subsequent scheduling and allocation calculations.

5. The intelligent regulation and equilibrium optimization method for new energy cluster grid connection as described in claim 1, characterized in that, Based on the joint execution control variables of multiple new energy clusters, grid load demand, available output, and upper-level coordination variables, scheduling and allocation calculations are completed to obtain grid-connected power, grid-connected power, power distribution, and real-time control performance indicators of each cluster, including: Receive the joint execution control input from all new energy clusters, determine the grid connection priority based on the grid connection power, and sort the clusters accordingly; For each new energy cluster, grid connection control performance indicators are determined based on grid-connected power and grid-connected electricity to reflect the cluster's contribution to the grid supply and demand balance. The system operating loss is calculated based on the product of the unit grid-connected power loss coefficient and the grid-connected power of each cluster; the collaborative benefit allocation coefficient is calculated based on the contribution weight of each cluster in grid-connected collaboration; and the data boundary violation penalty term is constructed based on the intention control quantity to define the data boundary violation degree. Based on grid-connected control performance indicators, system operating losses, grid-connected collaborative benefits, data out-of-bounds penalty items, and power balance regulation items, the stage control performance indicators of each cluster are obtained.

6. The intelligent regulation and equilibrium optimization method for new energy cluster grid connection as described in claim 1, characterized in that, The gradient constraint correction loss is: in, This is the gradient constraint correction loss; Represents a cluster During the period The degree of violation of the intent to control quantity; For training batches; The power equalization control loss is: in, For power equalization control losses; For the first A new energy cluster during the period The squared deviation of the actual grid-connected power share from the average power share of the same period.

7. The intelligent regulation and equilibrium optimization method for new energy cluster grid connection as described in claim 1, characterized in that, Under a centralized training and distributed execution framework, the parameters of the Actor and Critic networks are updated, and the grid-connected collaborative control strategies for each new energy cluster are output, including: A centralized training and distributed execution framework is adopted. During the training phase, the Critic network uses the global state to evaluate the value of each cluster and assists in the joint optimization of multiple clusters. During the execution phase, each cluster generates control variables independently based only on its own local observations, which meets the requirements of distributed control. During training, the temporal difference residuals are calculated based on control reward and state value estimation, and the advantage function is calculated using generalized advantage estimation. The Critic network is updated based on the mean square error between the system reward objective and the value estimate; the Actor network is based on the PPO pruning objective and integrates the entropy regularization term, gradient constraint correction loss and power balance regulation loss to form a comprehensive objective for policy update. By iterating until the maximum number of training rounds is reached or the convergence condition is met, the collaborative control strategy of each cluster after training is output.

8. An intelligent control and balance optimization system for grid connection of new energy clusters, characterized in that, include: The data input module is configured to: acquire the collection of new energy clusters and power grid dispatch data, and construct the global status of the system and the local observations of each new energy cluster; The evolution modeling module is configured to: establish a hierarchical time-series power system state evolution model based on the upper and lower limits of the operating data, upper-level coordination variables, and lower-level cluster action space, providing a state transition framework for multi-cluster collaborative decision-making; The tensor projection correction module is configured such that: each new energy cluster outputs an intention control quantity through the Actor strategy network based on local observations, and projects the intention control quantity into an execution control quantity that satisfies the upper and lower limits of the running data and the reporting ratio constraints through tensor interval projection correction; The scheduling and allocation calculation module is configured to perform scheduling and allocation calculations based on the joint execution control quantities of multiple new energy clusters, grid load demand, available output and upper-level coordination variables, and obtain grid-connected power, grid-connected power, power distribution and real-time control performance indicators of each cluster. The gradient constraint and power equalization module is configured to: construct gradient constraint correction loss based on the degree to which the intended control quantity deviates from the legal range, and construct power equalization regulation loss based on the grid-connected power concentration. The collaborative control training and strategy output module is configured to update the parameters of the Actor network and Critic network under the framework of centralized training and decentralized execution, and output the grid-connected collaborative control strategy for each new energy cluster.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When executed by the processor, the program implements the steps in the intelligent regulation and balance optimization method for grid connection of new energy clusters as described in any one of claims 1-7.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the intelligent regulation and balance optimization method for grid connection of new energy clusters as described in any one of claims 1-7.

Images