A rolling optimization dispatch method, system, device and medium of a power system

Abstract

The application discloses a rolling optimization dispatching method, system, device and medium of a power system, relates to the field of power system dispatching operation, and comprises the following steps: constructing an operation state vector according to operation data of a power system to be dispatched at a current rolling optimization moment, and inputting a pre-trained operation boundary prediction model to obtain a dynamic operation boundary parameter set; wherein the operation boundary prediction model is constructed based on deep reinforcement learning and combined training of multiple agents; and a multi-scale dispatching model is constructed according to the dynamic operation boundary parameter set, with the objective of minimizing the operation cost of the power system within a preset rolling window; wherein the rolling window takes the current rolling optimization moment as the starting boundary and comprises multiple cross sections of different time scales; a dispatching decision is obtained by solving the multi-scale dispatching model, and dispatching is performed according to the dispatching decision, thereby completing single-round rolling optimization dispatching. Through implementation of the application, the accuracy of the dispatching decision of the power system can be improved.

Description

Technical Field

[0001] This application relates to the field of power system dispatching and operation, and in particular to a rolling optimization clearing method, system, equipment and medium for power systems. Background Technology

[0002] In the operation of power systems, clearing, as the equilibrium state of supply and demand in the electricity market, aims to achieve optimal economic allocation of resources while meeting system security constraints. Existing traditional power system clearing methods often employ optimization models based on mathematical programming, such as constructing linear or mixed-integer programming models with the objective of minimizing operating costs, and considering constraints such as power balance, network transmission security, and unit physical operation to assist in market clearing.

[0003] However, traditional power system clearing methods typically employ separate segmented optimization for each time point. The models at each stage are relatively independent with fixed parameters, leading to actual clearing decisions that are often not globally optimal. Furthermore, the models lack the ability to perceive and adjust to complex dynamic environments, causing model parameters to fail to accurately reflect the operating characteristics of the power system, thus compromising the accuracy of actual clearing decisions. Therefore, improving the accuracy of power system clearing decisions remains a pressing issue that needs to be addressed by existing technologies. Summary of the Invention

[0004] This application provides a rolling optimization clearing method, system, equipment, and medium for power systems to address the technical problem of insufficient accuracy in existing power system clearing decisions.

[0005] According to a first aspect of the embodiments of this application, a rolling optimization clearing method for a power system is provided, comprising: Based on the operating data of the power system to be scheduled at the current rolling optimization moment, an operating state vector of the power system is constructed, and the operating state vector is input into a pre-trained operating boundary prediction model to dynamically predict the operating boundary of the power system, thereby obtaining the dynamic operating boundary parameter set of the power system; wherein, the operating boundary prediction model is constructed based on deep reinforcement learning and is constructed by combining and training multiple agents. Based on the dynamic operating boundary parameter set, a multi-scale clearing model is constructed with the objective of minimizing the operating cost of the power system within a preset rolling window; wherein the rolling window takes the current rolling optimization moment as the starting boundary; the rolling window includes multiple sections with different time scales; Solve the multi-scale clearing model to obtain the clearing decision of the power system, and schedule the power system according to the clearing decision to complete the single-round rolling optimization clearing of the power system.

[0006] This application first constructs an operating state vector based on the operating data of the power system at the current rolling optimization moment, and then inputs an operating boundary prediction model obtained by combining and training multiple agents based on deep reinforcement learning to dynamically predict the operating boundary and obtain a dynamic operating boundary parameter set. Next, a multi-scale clearing model is constructed with the objective of minimizing the operating cost of the power system within a rolling window including multiple time scales. The clearing decision is solved and scheduled to complete the single-round rolling optimization clearing. By combining multiple agents to construct the operating boundary prediction model and training it based on deep reinforcement learning, different agents can be driven to optimize the power system... By sensing data from different dimensions within the system, the comprehensiveness of data dimensions is improved, thereby enhancing the model's ability to predict the operating boundary of the power system and increasing the accuracy of the operating boundary prediction. This, in turn, improves the accuracy of subsequent clearing decisions through more accurate operating boundaries. By constructing a rolling window that includes multiple sections with different time scales and building a multi-scale clearing model with the goal of minimizing the operating cost of the power system within this rolling window, the clearing decision can be considered and integrated at different time scales simultaneously. This avoids the problem of non-global optimal clearing decisions caused by separate segmentation optimization for each time node, as seen in existing models, thus improving the accuracy of clearing decisions.

[0007] In some embodiments of this application, the running boundary prediction model is constructed based on deep reinforcement learning by combining and training multiple agents, specifically including: With the objective of minimizing the operating cost of the power system within the rolling window, the multi-scale clearing model is simplified to obtain a simplified clearing model. Based on the historical operating data of the power system, the multiple agents are iteratively trained. In each iteration, the multiple agents determine the boundary parameter set corresponding to the input state of the current round. The simplified clearing model, combined with the historical operating data, is used to simulate the boundary parameter set to obtain the input state of the next round and the execution reward of the current round. The iteration is completed when the execution reward converges to the optimal value. The multiple agents obtained after the last iteration are combined to construct a running boundary prediction model.

[0008] This application first simplifies the multi-scale clearing model to obtain a simplified clearing model with the goal of minimizing the operating cost of the power system within the rolling window. Then, iteratively trains multiple agents based on historical operating data, and combines the multiple agents obtained after the last iteration to construct an operating boundary prediction model. By constructing the simplified clearing model, a lightweight simulation environment that is close to the effect of the full environment can be provided for agent training, thereby improving training efficiency. By training and combining multiple agents based on deep reinforcement learning, different agents can be driven to perceive data of different dimensions in the power system, improving the comprehensiveness of data dimensions, thereby improving the model's ability to predict the operating boundary of the power system and improving the accuracy of the operating boundary prediction. In this way, the accuracy of subsequent clearing decisions can be improved through more accurate operating boundaries.

[0009] In some embodiments of this application, the plurality of agents includes a primary policy agent, a network evaluation agent, and a backup evaluation agent; the step of determining the boundary parameter set corresponding to the input state of the current round through the plurality of agents specifically includes: The network evaluation agent performs forward propagation prediction on the input state of the current round to determine the network risk assessment index of the current round, and the backup evaluation agent performs forward propagation prediction on the input state of the current round to determine the backup risk assessment index of the current round. By integrating the input state of the current round, network risk assessment indicators, and backup risk assessment indicators, a fused state feature is obtained. Then, through the main policy agent, the fused state feature is forward propagated to predict the boundary parameter set of the current round.

[0010] This application sequentially uses a network evaluation agent and a backup evaluation agent to perform forward propagation prediction on the input state of the current round, obtaining the network risk assessment index and the backup risk assessment index for the current round. Then, it fuses the input state, network risk assessment index, and backup risk assessment index of the current round to obtain fused state features. Finally, it uses a master policy agent to perform forward propagation prediction to obtain the boundary parameter set. This allows the network evaluation agent and the backup evaluation agent to perceive data of different dimensions, improving the comprehensiveness of data dimensions. Furthermore, the master policy agent's fusion processing of the input state and perceived data further improves the matching degree between the output data and the actual current input, thereby improving the model's ability to predict the power system operating boundary and increasing the accuracy of the operating boundary prediction.

[0011] In some embodiments of this application, the step of simulating the boundary parameter set using the simplified clearing model and historical operating data to obtain the input state for the next round and the execution reward for the current round specifically includes: The simplified clearing model is used to calculate the boundary parameter set, determine the simulation scheduling plan, and perform simulation based on the simulation scheduling plan and the historical running data to obtain the simulation running data and the input state for the next round. Based on the simulated operation data, the basic reward, time scale penalty, and collaborative reward for the current round are calculated sequentially and aggregated to obtain the execution reward for the current round.

[0012] This application first calculates the boundary parameter set using a simplified clearing model to determine the simulated scheduling plan. It then combines historical running data with simulation to obtain simulated running data and the input state for the next round. Next, using the simulated running data, it sequentially calculates the basic reward, time-scale penalty, and collaborative reward for the current round, and aggregates them to obtain the execution reward for the current round. By providing a simplified clearing model for the calculation environment of the boundary parameter set and a simulation environment for historical running data, the simulation efficiency can be improved with a simplified and lightweight environment, thereby improving the training efficiency of each round. By sequentially calculating the basic reward, time-scale penalty, and collaborative reward and aggregating them into the execution reward, the comprehensiveness of the reward function types in deep reinforcement learning can be improved, thus enhancing the objectivity of reward evaluation and improving the accuracy of the running boundary prediction model obtained after training, thereby improving the prediction accuracy of subsequent running boundaries.

[0013] In some embodiments of this application, the plurality of intelligent agents includes a primary policy intelligent agent, a network evaluation intelligent agent, and a backup evaluation intelligent agent; the step of inputting the operating state vector into a pre-trained operating boundary prediction model to dynamically predict the operating boundary of the power system and obtain the dynamic operating boundary parameter set of the power system specifically includes: The network evaluation agent performs forward propagation prediction on the running state vector to obtain network risk assessment indicators, and the backup evaluation agent performs forward propagation prediction on the running state vector to obtain backup risk assessment indicators. By fusing the operating state vector, the network risk assessment index, and the backup risk assessment index, a fused state feature is obtained. The fused state feature is then forward-propagated and predicted by the main policy agent to obtain a dynamic operating boundary parameter set.

[0014] This application first uses a network assessment agent and a backup assessment agent to perform forward propagation prediction on the operating state vector to obtain network risk assessment indicators and backup risk assessment indicators. Then, the main policy agent performs fusion processing and prediction on the operating state vector, network risk assessment indicators, and backup risk assessment indicators to obtain a dynamic operating boundary parameter set. This allows for the perception of data from different dimensions in the power system, improving the comprehensiveness of data dimensions, thereby enhancing the model's ability to predict the operating boundary of the power system and improving the accuracy of the operating boundary prediction. In turn, a more accurate operating boundary improves the accuracy of subsequent clearing decisions.

[0015] In some embodiments of this application, the dynamic operating boundary parameter set includes a multi-objective composite weight vector, a dynamic correction factor for reserve capacity demand, and a dynamic buffer for cross-sectional power limits; the step of constructing a multi-scale clearing model based on the dynamic operating boundary parameter set, with the objective of minimizing the operating cost of the power system within a preset rolling window, specifically includes: Based on the multi-objective synthesized weight vector, a clearing objective function is constructed with the goal of minimizing the operating cost of the power system within a preset rolling window; Based on the power system design, and in conjunction with the dynamic correction factor for reserve capacity demand and the dynamic buffer of the cross-sectional power limit, multiple associated constraints are set for the clearing objective function; wherein, the multiple associated constraints include power balance constraints, network power flow security constraints, physical operation limit constraints, and system service coupling constraints. A multi-scale clearing model is constructed based on the clearing objective function and the multiple associated constraints.

[0016] This application first constructs a clearing objective function based on a multi-objective composite weight vector, with the goal of minimizing the operating cost of the power system within a preset rolling window. Then, based on the system design, it sets multiple related constraints, including power balance constraints, network flow security constraints, physical operation limit constraints, and system service coupling constraints, by combining the dynamic correction factor for reserve capacity demand and the dynamic buffer amount of cross-sectional power limit. This results in the construction of a multi-scale clearing model, which can construct a multi-scale clearing model that is more suitable for the actual current task, thereby improving the accuracy of subsequent clearing decisions based on the multi-scale clearing model.

[0017] In some embodiments of this application, the step of constructing a clearing objective function based on the multi-objective synthesized weight vector, with the objective of minimizing the operating cost of the power system within a preset rolling window, specifically includes: Based on the system design of the power system, the total operating cost of the power system is decomposed to determine various types of sub-operating costs. Based on the system design, each type of sub-operating cost is modeled to obtain various sub-operating cost functions of the power system. With the goal of minimizing the operating cost of the power system within a preset rolling window, a clearing objective function is constructed by assigning corresponding dynamic weights to each sub-operating cost function through the multi-objective composite weight vector.

[0018] This application first decomposes the overall operating cost based on the system design to determine various types of sub-operating costs and models them to obtain multiple sub-operating cost functions. This improves the comprehensiveness of the sub-operating cost functions and provides a foundation for the subsequent construction of the clearing function. Then, with the goal of minimizing the operating cost of the power system within a preset rolling window, a clearing function is constructed by dynamically assigning weights to each sub-operating cost function through a multi-objective composite weight vector. This multi-objective composite vector weight can express the emphasis on different sub-operating costs, thereby constructing a clearing function that is more suitable for the current task. This improves the accuracy when constructing a multi-scale clearing model and solving the clearing decision using the clearing function.

[0019] According to a second aspect of the embodiments of this application, a rolling optimization clearing system for a power system is provided, including an operating boundary prediction module, a clearing model construction module, and a clearing solution scheduling module; The operation boundary prediction module is used to construct the operation state vector of the power system based on the operation data of the power system to be scheduled at the current rolling optimization time, and input the operation state vector into a pre-trained operation boundary prediction model to dynamically predict the operation boundary of the power system and obtain the dynamic operation boundary parameter set of the power system; wherein, the operation boundary prediction model is constructed based on deep reinforcement learning and combined training of multiple agents. The clearing model construction module is used to construct a multi-scale clearing model based on the dynamic operating boundary parameter set, with the objective of minimizing the operating cost of the power system within a preset rolling window; wherein the rolling window takes the current rolling optimization time as the starting boundary; the rolling window includes multiple sections with different time scales; The clearing solution scheduling module is used to solve the multi-scale clearing model, obtain the clearing decision of the power system, and schedule the power system according to the clearing decision to complete the single-round rolling optimization clearing of the power system.

[0020] In some embodiments of this application, the running boundary prediction model is constructed based on deep reinforcement learning by combining and training multiple agents, specifically including: With the objective of minimizing the operating cost of the power system within the rolling window, the multi-scale clearing model is simplified to obtain a simplified clearing model. Based on the historical operating data of the power system, the multiple agents are iteratively trained. In each iteration, the multiple agents determine the boundary parameter set corresponding to the input state of the current round. The simplified clearing model, combined with the historical operating data, is used to simulate the boundary parameter set to obtain the input state of the next round and the execution reward of the current round. The iteration is completed when the execution reward converges to the optimal value. The multiple agents obtained after the last iteration are combined to construct a running boundary prediction model.

[0021] In some embodiments of this application, the plurality of agents includes a primary policy agent, a network evaluation agent, and a backup evaluation agent; the step of determining the boundary parameter set corresponding to the input state of the current round through the plurality of agents specifically includes: The network evaluation agent performs forward propagation prediction on the input state of the current round to determine the network risk assessment index of the current round, and the backup evaluation agent performs forward propagation prediction on the input state of the current round to determine the backup risk assessment index of the current round. By integrating the input state of the current round, network risk assessment indicators, and backup risk assessment indicators, a fused state feature is obtained. Then, through the main policy agent, the fused state feature is forward propagated to predict the boundary parameter set of the current round.

[0022] In some embodiments of this application, the step of simulating the boundary parameter set using the simplified clearing model and historical operating data to obtain the input state for the next round and the execution reward for the current round specifically includes: The simplified clearing model is used to calculate the boundary parameter set, determine the simulation scheduling plan, and perform simulation based on the simulation scheduling plan and the historical running data to obtain the simulation running data and the input state for the next round. Based on the simulated operation data, the basic reward, time scale penalty, and collaborative reward for the current round are calculated sequentially and aggregated to obtain the execution reward for the current round.

[0023] In some embodiments of this application, the plurality of intelligent agents includes a main policy intelligent agent, a network evaluation intelligent agent, and a backup evaluation intelligent agent; the operation boundary prediction module includes an evaluation index prediction unit and an operation boundary prediction unit; The evaluation index prediction unit is used to perform forward propagation prediction on the operating state vector through the network evaluation agent to obtain network risk evaluation index, and to perform forward propagation prediction on the operating state vector through the backup evaluation agent to obtain backup risk evaluation index. The operation boundary prediction unit is used to fuse the operation state vector, the network risk assessment index, and the backup risk assessment index to obtain fused state features, and then use the main policy agent to perform forward propagation prediction on the fused state features to obtain a dynamic operation boundary parameter set.

[0024] In some embodiments of this application, the dynamic operating boundary parameter set includes a multi-objective composite weight vector, a dynamic correction factor for reserve capacity demand, and a dynamic buffer amount for cross-sectional power limits; the clearing model construction module includes an objective function construction unit, an associated constraint setting unit, and a clearing model construction unit. The objective function construction unit is used to construct a clearing objective function based on the multi-objective synthesized weight vector, with the objective of minimizing the operating cost of the power system within a preset rolling window; The associated constraint setting unit is used to set multiple associated constraints of the clearing objective function according to the system design of the power system, combined with the dynamic correction factor of the reserve capacity demand and the dynamic buffer amount of the cross-sectional power limit; wherein, the multiple associated constraints include power balance constraints, network power flow security constraints, physical operation limit constraints and system service coupling constraints. The clearing model construction unit is used to construct a multi-scale clearing model based on the clearing objective function and the multiple associated constraints.

[0025] In some embodiments of this application, the objective function construction unit includes a cost modeling subunit and an objective construction subunit; The cost modeling subunit is used to decompose the full-process operation cost of the power system according to the system design of the power system, determine the various types of sub-operation costs of the power system, and model each type of sub-operation cost according to the system design to obtain the various sub-operation cost functions of the power system. The target construction subunit is used to construct a clearing target function by assigning corresponding dynamic weights to each sub-operation cost function through the multi-objective composite weight vector, with the goal of minimizing the operating cost of the power system within a preset rolling window.

[0026] This application first constructs an operating state vector based on the operating data of the power system at the current rolling optimization moment, and then inputs an operating boundary prediction model obtained by combining and training multiple agents based on deep reinforcement learning to dynamically predict the operating boundary and obtain a dynamic operating boundary parameter set. Next, a multi-scale clearing model is constructed with the objective of minimizing the operating cost of the power system within a rolling window including multiple time scales. The clearing decision is solved and scheduled to complete the single-round rolling optimization clearing. By combining multiple agents to construct the operating boundary prediction model and training it based on deep reinforcement learning, different agents can be driven to optimize the power system... By sensing data from different dimensions within the system, the comprehensiveness of data dimensions is improved, thereby enhancing the model's ability to predict the operating boundary of the power system and increasing the accuracy of the operating boundary prediction. This, in turn, improves the accuracy of subsequent clearing decisions through more accurate operating boundaries. By constructing a rolling window that includes multiple sections with different time scales and building a multi-scale clearing model with the goal of minimizing the operating cost of the power system within this rolling window, the clearing decision can be considered and integrated at different time scales simultaneously. This avoids the problem of non-global optimal clearing decisions caused by separate segmentation optimization for each time node, as seen in existing models, thus improving the accuracy of clearing decisions.

[0027] According to a third aspect of the embodiments of this application, a computer device is provided, comprising: a processor; a memory; and a computer program stored in the memory and configured to be executed by the processor; wherein the processor executes the computer program to implement a rolling optimization clearing method for a power system as described in this application.

[0028] According to a fourth aspect of the embodiments of this application, a computer-readable storage medium is provided, the computer-readable storage medium storing a plurality of instructions adapted for loading by a processor to execute a rolling optimization clearing method for a power system as described in this application. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating a rolling optimization clearing method for a power system according to certain embodiments of this application. Figure 2 This is a module structure diagram of a rolling optimization clearing system for a power system, as shown in some embodiments of this application. Detailed Implementation

[0030] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below in conjunction with the accompanying drawings are exemplary and are only used to explain some embodiments of this application, and should not be construed as limiting the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments shown in this application without inventive effort are within the protection scope of this application.

[0031] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, unless otherwise explicitly specified, "a plurality of" or "several" means two or more.

[0032] In existing traditional power system clearing methods, many optimization models based on mathematical programming are used to achieve clearing. For example, linear or mixed integer programming models are constructed with the goal of minimizing operating costs, and constraints such as power balance, network transmission security, and unit physical operation are considered to assist market clearing. However, these methods have the following drawbacks: (1) Each time node is optimized separately, and the models at each stage are relatively independent with fixed parameters, which often results in the actual clearing decision not being globally optimal; (2) The model lacks the ability to perceive and adjust to complex dynamic environments, which means that the model parameters cannot accurately reflect the operating characteristics of the power system, making it difficult to guarantee the accuracy of the actual clearing decision. Therefore, how to improve the accuracy of clearing decisions in the power system remains a problem that needs to be solved by existing technologies.

[0033] Based on the above technical background, please refer to Figure 1 This application provides a rolling optimization clearing method for a power system, including steps S101 to S103, each step of which is as follows: Step S101: Based on the operating data of the power system to be scheduled at the current rolling optimization moment, construct the operating state vector of the power system, and input the operating state vector into the pre-trained operating boundary prediction model to dynamically predict the operating boundary of the power system, thereby obtaining the dynamic operating boundary parameter set of the power system; wherein, the operating boundary prediction model is constructed based on deep reinforcement learning and combined training of multiple agents.

[0034] Specifically, the operational data includes, but is not limited to: system load forecast curves, specifically time-series curves formed by forecasting the total power demand of the power system within a specific future time period; new energy power generation forecast curves, specifically time-series curves formed by forecasting the power generation of renewable energy sources such as wind farms and photovoltaic power plants within a specific future time period; power grid topology information, specifically model data describing the connection relationships of various electrical devices in the power system; line transmission limits, specifically the maximum active power limit allowed for long-term operation of each transmission line and transformer under safe operation conditions; and physical operating parameters reported by each controllable generation unit and regulation unit, including power generation energy consumption characteristics, active and reactive power regulation ranges, ramp rate, and reserve capacity response performance. More specifically, when constructing the operating state vector of the power system, the collected operational data is first cleaned to remove noise data, and then vectorized to construct the operating state vector. .

[0035] In some embodiments of this application, the running boundary prediction model is constructed based on deep reinforcement learning by combining and training multiple agents, specifically including: With the objective of minimizing the operating cost of the power system within the rolling window, the multi-scale clearing model is simplified to obtain a simplified clearing model. Based on the historical operating data of the power system, the multiple agents are iteratively trained. In each iteration, the multiple agents determine the boundary parameter set corresponding to the input state of the current round. The simplified clearing model, combined with the historical operating data, is used to simulate the boundary parameter set to obtain the input state of the next round and the execution reward of the current round. The iteration is completed when the execution reward converges to the optimal value. The multiple agents obtained after the last iteration are combined to construct a running boundary prediction model.

[0036] Specifically, the plurality of intelligent agents includes a primary policy agent, a network evaluation agent, and a backup evaluation agent; the primary policy agent, the network evaluation agent, and the backup evaluation agent are all deep neural networks; the network evaluation agent is used to assess network congestion risk on the running state vector and output network risk assessment indicators. The standby assessment agent is used to assess the standby capability of the operating state vector and output standby risk assessment indicators. .

[0037] More specifically, the network evaluates the agent, including: an input layer for processing the input running state vector. The filtering is performed on feature sub-vectors related to network congestion risk; three hidden layers are set sequentially, with 256, 128 and 64 neurons in each layer, respectively, and activated based on the ReLU function; the output layer is a fully connected layer with m neurons, where m is the number of sections with different time scales in the scrolling window, activated based on the Sigmoid function, and used to output the network risk assessment index to the feature fusion layer of the main policy agent; the network risk assessment index is in vector form.

[0038] More specifically, the alternative evaluation agent includes: an input layer for processing the input running state vector. The filtering is performed using feature sub-vectors related to the system's backup capability; three hidden layers are set sequentially, with 128, 64 and 32 neurons in each layer, respectively, and activated based on the ReLU function; the output layer is a fully connected layer with one neuron, activated based on the Sigmoid function, used to output the backup risk assessment index to the feature fusion layer of the main policy agent; the backup risk assessment index is in scalar form.

[0039] More specifically, the main policy agent includes: an input layer for receiving a running state vector. The system consists of: a first feature extraction hidden layer, which extracts features from the running state vector and outputs a 256-dimensional feature vector; a feature fusion layer, which fuses the 256-dimensional running state vector, an m-dimensional network risk assessment index, and a 1-dimensional backup risk assessment index to obtain a (256+m+1)-dimensional fused state feature; a second feature extraction hidden layer, which performs feature transformation on the fused state feature to obtain a 128-dimensional intermediate feature vector; a decision hidden layer, which performs feature transformation on the 128-dimensional intermediate feature vector to obtain a 64-dimensional intermediate feature vector; and an output layer, which includes three fully connected layers, used for branch decisions to obtain the corresponding multi-objective composite weight vector. , Reserve capacity demand dynamic adjustment factor Dynamic buffer amount of cross-sectional power limit The output is integrated into a dynamic runtime boundary parameter set. .

[0040] Specifically, in each iteration of the training of the multiple agents, the input state of the current iteration is first converted into a running state vector. The data is passed to the network evaluation agent and the backup evaluation agent for forward propagation prediction, and outputs network risk assessment indicators. and backup risk assessment indicators Then run state vector Network risk assessment indicators and backup risk assessment indicators The parameters are jointly passed to the main policy agent for forward propagation prediction, and the output is a dynamic set of operational boundary parameters. And based on the dynamic running boundary parameter set A simulated scheduling plan is obtained by simplifying the clearing model calculation, and the execution of the simulated scheduling plan under uncertainty is simulated to obtain the input state for the next round. and the execution reward for the current round Finally, at the end of the current iteration, the interaction experience tuples will be... Store the data in the experience replay buffer to calculate the total loss for multiple agents and update the network parameters for each agent.

[0041] This application first simplifies the multi-scale clearing model to obtain a simplified clearing model with the goal of minimizing the operating cost of the power system within the rolling window. Then, iteratively trains multiple agents based on historical operating data, and combines the multiple agents obtained after the last iteration to construct an operating boundary prediction model. By constructing the simplified clearing model, a lightweight simulation environment that is close to the effect of the full environment can be provided for agent training, thereby improving training efficiency. By training and combining multiple agents based on deep reinforcement learning, different agents can be driven to perceive data of different dimensions in the power system, improving the comprehensiveness of data dimensions, thereby improving the model's ability to predict the operating boundary of the power system and improving the accuracy of the operating boundary prediction. In this way, the accuracy of subsequent clearing decisions can be improved through more accurate operating boundaries.

[0042] In some embodiments of this application, determining the boundary parameter set corresponding to the input state of the current round through the plurality of agents specifically includes: The network evaluation agent performs forward propagation prediction on the input state of the current round to determine the network risk assessment index of the current round, and the backup evaluation agent performs forward propagation prediction on the input state of the current round to determine the backup risk assessment index of the current round. By integrating the input state of the current round, network risk assessment indicators, and backup risk assessment indicators, a fused state feature is obtained. Then, through the main policy agent, the fused state feature is forward propagated to predict the boundary parameter set of the current round.

[0043] This application sequentially uses a network evaluation agent and a backup evaluation agent to perform forward propagation prediction on the input state of the current round, obtaining the network risk assessment index and the backup risk assessment index for the current round. Then, it fuses the input state, network risk assessment index, and backup risk assessment index of the current round to obtain fused state features. Finally, it uses a master policy agent to perform forward propagation prediction to obtain the boundary parameter set. This allows the network evaluation agent and the backup evaluation agent to perceive data of different dimensions, improving the comprehensiveness of data dimensions. Furthermore, the master policy agent's fusion processing of the input state and perceived data further improves the matching degree between the output data and the actual current input, thereby improving the model's ability to predict the power system operating boundary and increasing the accuracy of the operating boundary prediction.

[0044] In some embodiments of this application, the step of simulating the boundary parameter set using the simplified clearing model and historical operating data to obtain the input state for the next round and the execution reward for the current round specifically includes: The simplified clearing model is used to calculate the boundary parameter set, determine the simulation scheduling plan, and perform simulation based on the simulation scheduling plan and the historical running data to obtain the simulation running data and the input state for the next round. Based on the simulated operation data, the basic reward, time scale penalty, and collaborative reward for the current round are calculated sequentially and aggregated to obtain the execution reward for the current round.

[0045] For example, assuming the multi-scale clearing model is a mixed-integer linear programming problem involving N nodes, G generators, L lines, and T time periods, the corresponding simplified clearing model can be constructed as follows: (1) Time period: Select Three time periods, among which (1) Represent the current time period, the midpoint and the end point of the rolling window respectively; (2) Network: Use the DC power flow model based on the power transmission distribution factor and cluster the network nodes into S regions (S < N). Only retain the inter-regional tie lines as key sections. The key sections are multiple sections with different time scales within the rolling window; (3) Units: Aggregate the same type of units in the same region into an equivalent unit. Its total capacity and total reserve capacity are the sum of the units. The cost curve is taken as a weighted average; (4) Constraints: Only retain the active power balance constraints for the three time periods. The line power flow constraints are only applied to the key tie lines between regions. The unit constraints retain the upper and lower limits of the output after aggregation and the total ramping capacity constraints between the three time periods. The reserve requirement is simplified to a total capacity constraint and does not distinguish between upper and lower limits; (5) Solution: Call the linear programming solver and set the optimality tolerance to 1e-3 (higher than the original multi-scale clearing model's 1e-6). The maximum solution time is set to 5 seconds.

[0046] Specifically, the execution reward for the current round Specifically, it can be expressed as: ; in, These are basic rewards, time-scale penalties, and collaborative rewards; The basic reward This is used to correlate the total operating cost corresponding to the simulated scheduling plan, and it is negatively correlated with the total operating cost, specifically: ; ; in, The scaling factor for the base reward is preferably set to [0.001, 0.01]. The total system operating cost is obtained by solving the simplified clearing model based on the boundary parameter set; Indicates the current moment. This is the sum of all time segments within the scrolling window; This represents the set of conventional generating units and the set of rapid adjustment units in the simplified clearing model; These are the unit active power, increased reserve capacity, decreased reserve capacity, and AGC regulation capacity output from the simplified clearing model solution, respectively. To simplify the clearing model of the units The cost of electricity generation; The reserve capacity reserve cost factor for providing the unit reserve capacity required for the unit to be adjusted upward or downward reflects the additional losses incurred in maintaining the reserve status; A cost factor is reserved for the adjustment capacity per unit AGC capacity of the rapid adjustment unit, reflecting the equipment wear and tear caused by frequent adjustments; The time scale penalty This is used to characterize the penalty caused by improper dynamic parameter settings, resulting in short-term safety exceeding limits or long-term insufficient reserve. Specifically: ; in, The line power flow violation penalty coefficient is preferably set to [0.1, 1.0]. To simplify the set of critical paths in the clearing model; To simulate the line in the scheduling plan At any moment The actual effective current; For the line The static transmission limit; The output of the main policy agent during training is used to simplify the clearing model, corresponding to the line. At any moment Dynamic buffer amount of cross-sectional power limit; The preferred value for the reserve shortage penalty coefficient is [0.5, 2.0]. To simulate the timing in the scheduling plan The total amount of reserve demand actually triggered by the time of uncertainty; For dynamic parameter settings, at time The total amount of standby capacity that can actually be called upon at that time; The collaborative reward This is used to encourage the optimal allocation of resources among energy and ancillary service providers, specifically: ; in, The scaling factor for collaborative rewards is preferably set to [0.01, 0.1]. These represent the quantitative importance coefficients corresponding to the invocation of reserve adjustment, reserve reduction, and AGC adjustment service, respectively, and are set based on historical operating data; These are the time points in the simulated scheduling plan. The actual increase and decrease in reserve capacity that are called up; To simulate the timing in the scheduling plan The actual AGC adjustment capacity that is invoked; For a moment The costs incurred to reserve spare capacity; For a moment The reserved costs incurred to accommodate AGC adjustment capacity; To reserve positive numbers, in order to prevent the denominator from being 0.

[0047] This application first calculates the boundary parameter set using a simplified clearing model to determine the simulated scheduling plan. It then combines historical running data with simulation to obtain simulated running data and the input state for the next round. Next, using the simulated running data, it sequentially calculates the basic reward, time-scale penalty, and collaborative reward for the current round, and aggregates them to obtain the execution reward for the current round. By providing a simplified clearing model for the calculation environment of the boundary parameter set and a simulation environment for historical running data, the simulation efficiency can be improved with a simplified and lightweight environment, thereby improving the training efficiency of each round. By sequentially calculating the basic reward, time-scale penalty, and collaborative reward and aggregating them into the execution reward, the comprehensiveness of the reward function types in deep reinforcement learning can be improved, thus enhancing the objectivity of reward evaluation and improving the accuracy of the running boundary prediction model obtained after training, thereby improving the prediction accuracy of subsequent running boundaries.

[0048] In some embodiments of this application, the step of inputting the operating state vector into a pre-trained operating boundary prediction model to dynamically predict the operating boundary of the power system and obtain the dynamic operating boundary parameter set of the power system specifically includes: The network evaluation agent performs forward propagation prediction on the running state vector to obtain network risk assessment indicators, and the backup evaluation agent performs forward propagation prediction on the running state vector to obtain backup risk assessment indicators. By fusing the operating state vector, the network risk assessment index, and the backup risk assessment index, a fused state feature is obtained. The fused state feature is then forward-propagated and predicted by the main policy agent to obtain a dynamic operating boundary parameter set.

[0049] This application first uses a network assessment agent and a backup assessment agent to perform forward propagation prediction on the operating state vector to obtain network risk assessment indicators and backup risk assessment indicators. Then, the main policy agent performs fusion processing and prediction on the operating state vector, network risk assessment indicators, and backup risk assessment indicators to obtain a dynamic operating boundary parameter set. This allows for the perception of data from different dimensions in the power system, improving the comprehensiveness of data dimensions, thereby enhancing the model's ability to predict the operating boundary of the power system and improving the accuracy of the operating boundary prediction. In turn, a more accurate operating boundary improves the accuracy of subsequent clearing decisions.

[0050] Step S102: Based on the dynamic operating boundary parameter set, construct a multi-scale clearing model with the objective of minimizing the operating cost of the power system within a preset rolling window; wherein, the rolling window takes the current rolling optimization time as the starting boundary; the rolling window includes multiple sections with different time scales.

[0051] Specifically, the scrolling window includes A series of continuous cross-sections with different time scales (or time resolutions) are denoted as . Used to represent multi-scale time characteristics, such as to Corresponding to the high-frequency period of the next 15 minutes, to This corresponds to a low-frequency period in the next few hours.

[0052] Specifically, in the task requirements of this application, the variables to be decided include: conventional power generation units. Time cross-section Active power generation New energy power stations Time cross-section Planned power generation Conventional power generation unit Time cross-section The system provides the ability to increase and decrease reserve capacity. Quick adjustment unit Time cross-section Reserved automatic power generation control and regulation capacity ;node Time cross-section voltage phase angle .

[0053] In some embodiments of this application, the dynamic operating boundary parameter set includes a multi-objective composite weight vector, a dynamic correction factor for reserve capacity demand, and a dynamic buffer for cross-sectional power limits; the step of constructing a multi-scale clearing model based on the dynamic operating boundary parameter set, with the objective of minimizing the operating cost of the power system within a preset rolling window, specifically includes: Based on the multi-objective synthesized weight vector, a clearing objective function is constructed with the goal of minimizing the operating cost of the power system within a preset rolling window; Based on the power system design, and in conjunction with the dynamic correction factor for reserve capacity demand and the dynamic buffer of the cross-sectional power limit, multiple associated constraints are set for the clearing objective function; wherein, the multiple associated constraints include power balance constraints, network power flow security constraints, physical operation limit constraints, and system service coupling constraints. A multi-scale clearing model is constructed based on the clearing objective function and the multiple associated constraints.

[0054] Specifically, the clearing objective function is as follows: ; in, This is a multi-objective composite weight vector, where each weight component satisfies... and ; These are the sets of all conventional power generation units and the sets of all fast regulation units with AGC capabilities, respectively. For power generation unit Time cross-section contribution; For power generation unit Time cross-section Output The power generation cost function is a quadratic convex function that reflects the cost of the physical quantities consumed by the unit in converting primary energy into electrical energy. For power generation unit Provides reserve capacity reservation cost coefficients corresponding to unit increases and decreases in reserve capacity, reflecting the additional operating losses or opportunity costs incurred by the unit in keeping some capacity available for use; For rapid adjustment unit The reserved unit AGC adjustment capacity reserve cost coefficient reflects the equipment wear and control costs incurred by the rapid adjustment unit in providing rapid adjustment services.

[0055] Specifically, the plurality of association constraints include: The power balance constraints include node active power balance constraints and node reactive power balance constraints; the node active power balance constraints apply to each node. and each time section Specifically: ; The reactive power balance constraint at the nodes is specifically as follows: ; in, These represent connections on the nodes. A collection of conventional generating units and new energy power stations; For nodes Time cross-section The load forecast value; For nodes The collection of all directly connected power transmission lines; For the line susceptance; The lines are respectively end nodes or end node The voltage phase angle; Conventional power generation unit Time cross-section The reactive power generation capacity; For nodes Time cross-section The reactive power of parallel capacitors or reactors; For nodes Time cross-section The reactive power of the load; The lines are respectively Conductivity, susceptance, and susceptance to ground; They are nodes and nodes Time cross-section The voltage amplitude; For nodes and nodes Time cross-section The voltage phase angle difference; The network power flow security constraint is approximated by DC power flow, specifically as follows: ; in, This is a pre-defined set of key power transmission sections; For the line The static active power transmission limit; To correspond to the line Time cross-section The dynamic buffer value of the cross-sectional power limit is non-negative, and the preferred range is [value missing]. ; For the line susceptance; The physical operational limit constraints include upper and lower limits for power generation unit output and upper and lower limits for unit ramp rate; the upper and lower limits for power generation unit output are specifically as follows: ; The upper and lower limits of the unit's ramp rate constraints are as follows: ; in, For the unit Minimum and maximum output; For the unit The maximum upward and downward climbing rates; Time section and The time interval between; The system service coupling constraints include capacity coupling constraints, total reserve requirement coupling constraints, and AGC regulation capacity coupling constraints; the capacity coupling constraints apply to units providing reserve, specifically: ; ; The total reserve requirement coupling constraint is specifically as follows: ; ; The AGC adjustment capacity coupling constraint is specifically as follows: ; ; in, It is obtained based on the dynamic adjustment factor of reserve capacity demand, and the acquisition methods include decomposition or direct application. These are the system's upward and downward adjustments to the reserve baseline demand, respectively, based on pre-set statistical data of load and renewable energy forecasting errors. For unit Maximum available AGC adjustment capacity; The AGC regulation capacity reference value required to maintain system frequency stability is related to the system size and the control requirements for frequency deviation; This is the dynamic adjustment coefficient for AGC capacity demand.

[0056] This application first constructs a clearing objective function based on a multi-objective composite weight vector, with the goal of minimizing the operating cost of the power system within a preset rolling window. Then, based on the system design, it sets multiple related constraints, including power balance constraints, network flow security constraints, physical operation limit constraints, and system service coupling constraints, by combining the dynamic correction factor for reserve capacity demand and the dynamic buffer amount of cross-sectional power limit. This results in the construction of a multi-scale clearing model, which can construct a multi-scale clearing model that is more suitable for the actual current task, thereby improving the accuracy of subsequent clearing decisions based on the multi-scale clearing model.

[0057] In some embodiments of this application, the step of constructing a clearing objective function based on the multi-objective synthesized weight vector, with the objective of minimizing the operating cost of the power system within a preset rolling window, specifically includes: Based on the system design of the power system, the total operating cost of the power system is decomposed to determine various types of sub-operating costs. Based on the system design, each type of sub-operating cost is modeled to obtain various sub-operating cost functions of the power system. With the goal of minimizing the operating cost of the power system within a preset rolling window, a clearing objective function is constructed by assigning corresponding dynamic weights to each sub-operating cost function through the multi-objective composite weight vector.

[0058] This application first decomposes the overall operating cost based on the system design to determine various types of sub-operating costs and models them to obtain multiple sub-operating cost functions. This improves the comprehensiveness of the sub-operating cost functions and provides a foundation for the subsequent construction of the clearing function. Then, with the goal of minimizing the operating cost of the power system within a preset rolling window, a clearing function is constructed by dynamically assigning weights to each sub-operating cost function through a multi-objective composite weight vector. This multi-objective composite vector weight can express the emphasis on different sub-operating costs, thereby constructing a clearing function that is more suitable for the current task. This improves the accuracy when constructing a multi-scale clearing model and solving the clearing decision using the clearing function.

[0059] Step S103: Solve the multi-scale clearing model to obtain the clearing decision of the power system, and schedule the power system according to the clearing decision to complete the single-round rolling optimization clearing of the power system.

[0060] Specifically, the clearing decision is characterized by a pre-scheduling scheme across multiple future time scales, including: the power generation plan curve of each controllable unit, i.e., the time-series curve formed by the planned active power output values ​​of each conventional generator unit and each new energy power station at various future time segments; the reserve capacity reservation status, i.e., the reserved upward and downward reserve capacity values ​​of each unit that can provide reserve services at various future time periods; and the adjustment service call plan, i.e., the automatic power generation control adjustment capacity size and direction allocated to units with rapid adjustment capabilities at various future time periods. Specifically, when scheduling according to the clearing decision, the pre-scheduling scheme corresponding to the next immediate execution cycle is extracted and converted into corresponding deterministic control instructions, which are then issued to each corresponding controllable unit for execution through the scheduling automation system, thereby completing the rolling optimization clearing at the current rolling optimization time (corresponding to one optimization round).

[0061] Compared to existing technologies, this application first constructs an operating state vector based on the operating data of the power system at the current rolling optimization moment. Then, it inputs an operating boundary prediction model obtained by combining and training multiple agents using deep reinforcement learning to dynamically predict the operating boundary, thus obtaining a dynamic operating boundary parameter set. Next, it constructs a multi-scale clearing model with the objective of minimizing the operating cost of the power system within a rolling window encompassing multiple time scales. The model solves for the clearing decision and performs scheduling to complete the single-round rolling optimization clearing. By combining multiple agents to construct the operating boundary prediction model and training it using deep reinforcement learning, it can drive different agents. By sensing data from different dimensions within the power system, the comprehensiveness of data dimensions is improved, thereby enhancing the model's ability to predict the power system's operating boundaries and increasing the accuracy of these predictions. This, in turn, improves the accuracy of subsequent clearing decisions through more accurate operating boundaries. Furthermore, by constructing a rolling window encompassing multiple time scales and aiming to minimize the power system's operating costs within this rolling window, a multi-scale clearing model can be built. This model can simultaneously consider and integrate clearing decisions at different time scales, avoiding the non-globally optimal problem caused by separate segmentation optimization for each time node, as seen in existing methods, thus improving the accuracy of clearing decisions.

[0062] For a method corresponding to the one described above, please refer to [link to relevant documentation]. Figure 2 The present application provides a rolling optimization clearing system for a power system, including an operating boundary prediction module 210, a clearing model construction module 220, and a clearing solution scheduling module 230. The operation boundary prediction module 210 is used to construct the operation state vector of the power system based on the operation data of the power system to be scheduled at the current rolling optimization time, and input the operation state vector into a pre-trained operation boundary prediction model to dynamically predict the operation boundary of the power system and obtain the dynamic operation boundary parameter set of the power system; wherein, the operation boundary prediction model is constructed based on deep reinforcement learning and combined training of multiple agents. The clearing model construction module 220 is used to construct a multi-scale clearing model based on the dynamic operating boundary parameter set, with the objective of minimizing the operating cost of the power system within a preset rolling window; wherein the rolling window takes the current rolling optimization time as the starting boundary; the rolling window includes multiple sections with different time scales; The clearing solution scheduling module 230 is used to solve the multi-scale clearing model, obtain the clearing decision of the power system, and schedule the power system according to the clearing decision to complete the single-round rolling optimization clearing of the power system.

[0063] In some embodiments of this application, the running boundary prediction model is constructed based on deep reinforcement learning by combining and training multiple agents, specifically including: With the objective of minimizing the operating cost of the power system within the rolling window, the multi-scale clearing model is simplified to obtain a simplified clearing model. Based on the historical operating data of the power system, the multiple agents are iteratively trained. In each iteration, the multiple agents determine the boundary parameter set corresponding to the input state of the current round. The simplified clearing model, combined with the historical operating data, is used to simulate the boundary parameter set to obtain the input state of the next round and the execution reward of the current round. The iteration is completed when the execution reward converges to the optimal value. The multiple agents obtained after the last iteration are combined to construct a running boundary prediction model.

[0064] In some embodiments of this application, the plurality of agents includes a primary policy agent, a network evaluation agent, and a backup evaluation agent; the step of determining the boundary parameter set corresponding to the input state of the current round through the plurality of agents specifically includes: The network evaluation agent performs forward propagation prediction on the input state of the current round to determine the network risk assessment index of the current round, and the backup evaluation agent performs forward propagation prediction on the input state of the current round to determine the backup risk assessment index of the current round. By integrating the input state of the current round, network risk assessment indicators, and backup risk assessment indicators, a fused state feature is obtained. Then, through the main policy agent, the fused state feature is forward propagated to predict the boundary parameter set of the current round.

[0065] In some embodiments of this application, the step of simulating the boundary parameter set using the simplified clearing model and historical operating data to obtain the input state for the next round and the execution reward for the current round specifically includes: The simplified clearing model is used to calculate the boundary parameter set, determine the simulation scheduling plan, and perform simulation based on the simulation scheduling plan and the historical running data to obtain the simulation running data and the input state for the next round. Based on the simulated operation data, the basic reward, time scale penalty, and collaborative reward for the current round are calculated sequentially and aggregated to obtain the execution reward for the current round.

[0066] In some embodiments of this application, the plurality of intelligent agents include a main policy intelligent agent, a network evaluation intelligent agent, and a backup evaluation intelligent agent; the operation boundary prediction module 210 includes an evaluation index prediction unit and an operation boundary prediction unit; The evaluation index prediction unit is used to perform forward propagation prediction on the operating state vector through the network evaluation agent to obtain network risk evaluation index, and to perform forward propagation prediction on the operating state vector through the backup evaluation agent to obtain backup risk evaluation index. The operation boundary prediction unit is used to fuse the operation state vector, the network risk assessment index, and the backup risk assessment index to obtain fused state features, and then use the main policy agent to perform forward propagation prediction on the fused state features to obtain a dynamic operation boundary parameter set.

[0067] In some embodiments of this application, the dynamic operating boundary parameter set includes a multi-objective composite weight vector, a dynamic correction factor for reserve capacity demand, and a dynamic buffer amount for cross-sectional power limits; the clearing model construction module 220 includes an objective function construction unit, an associated constraint setting unit, and a clearing model construction unit. The objective function construction unit is used to construct a clearing objective function based on the multi-objective synthesized weight vector, with the objective of minimizing the operating cost of the power system within a preset rolling window; The associated constraint setting unit is used to set multiple associated constraints of the clearing objective function according to the system design of the power system, combined with the dynamic correction factor of the reserve capacity demand and the dynamic buffer amount of the cross-sectional power limit; wherein, the multiple associated constraints include power balance constraints, network power flow security constraints, physical operation limit constraints and system service coupling constraints. The clearing model construction unit is used to construct a multi-scale clearing model based on the clearing objective function and the multiple associated constraints.

[0068] In some embodiments of this application, the objective function construction unit includes a cost modeling subunit and an objective construction subunit; The cost modeling subunit is used to decompose the full-process operation cost of the power system according to the system design of the power system, determine the various types of sub-operation costs of the power system, and model each type of sub-operation cost according to the system design to obtain the various sub-operation cost functions of the power system. The target construction subunit is used to construct a clearing target function by assigning corresponding dynamic weights to each sub-operation cost function through the multi-objective composite weight vector, with the goal of minimizing the operating cost of the power system within a preset rolling window.

[0069] This application first constructs an operating state vector based on the operating data of the power system at the current rolling optimization moment, and then inputs an operating boundary prediction model obtained by combining and training multiple agents based on deep reinforcement learning to dynamically predict the operating boundary and obtain a dynamic operating boundary parameter set. Next, a multi-scale clearing model is constructed with the objective of minimizing the operating cost of the power system within a rolling window including multiple time scales. The clearing decision is solved and scheduled to complete the single-round rolling optimization clearing. By combining multiple agents to construct the operating boundary prediction model and training it based on deep reinforcement learning, different agents can be driven to optimize the power system... By sensing data from different dimensions within the system, the comprehensiveness of data dimensions is improved, thereby enhancing the model's ability to predict the operating boundary of the power system and increasing the accuracy of the operating boundary prediction. This, in turn, improves the accuracy of subsequent clearing decisions through more accurate operating boundaries. By constructing a rolling window that includes multiple sections with different time scales and building a multi-scale clearing model with the goal of minimizing the operating cost of the power system within this rolling window, the clearing decision can be considered and integrated at different time scales simultaneously. This avoids the problem of non-global optimal clearing decisions caused by separate segmentation optimization for each time node, as seen in existing models, thus improving the accuracy of clearing decisions.

[0070] It should be understood that the system provided in this application is corresponding to the aforementioned method. The rolling optimization clearing system for a power system provided in this application can implement the rolling optimization clearing method for a power system provided in any of the embodiments of this application.

[0071] Adaptively, embodiments of this application also provide a computer device and a computer-readable storage medium.

[0072] The computer device includes: a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor; The processor executes the computer program to implement a rolling optimization clearing method for a power system according to this application.

[0073] The computer-readable storage medium stores multiple instructions adapted for loading by a processor to execute a rolling optimization clearing method for a power system according to this application.

[0074] The above description represents some embodiments of this application, providing a further detailed explanation of the purpose, technical solution, and beneficial effects of this application. It should be understood that the above-described embodiments of this application should not be construed as limiting this application. In particular, any changes, modifications, equivalent substitutions, and variations made by those skilled in the art within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A rolling optimization clearing method for a power system, characterized in that, include: Based on the operating data of the power system to be scheduled at the current rolling optimization moment, an operating state vector of the power system is constructed, and the operating state vector is input into a pre-trained operating boundary prediction model to dynamically predict the operating boundary of the power system, thereby obtaining the dynamic operating boundary parameter set of the power system; wherein, the operating boundary prediction model is constructed based on deep reinforcement learning and is constructed by combining and training multiple agents. Based on the dynamic operating boundary parameter set, a multi-scale clearing model is constructed with the objective of minimizing the operating cost of the power system within a preset rolling window; wherein, the rolling window takes the current rolling optimization moment as the starting boundary; the rolling window includes multiple sections with different time scales; Solve the multi-scale clearing model to obtain the clearing decision of the power system, and schedule the power system according to the clearing decision to complete the single-round rolling optimization clearing of the power system.

2. The rolling optimization clearing method for a power system according to claim 1, characterized in that, The operational boundary prediction model is constructed based on deep reinforcement learning, through combined training of multiple agents, and specifically includes: With the objective of minimizing the operating cost of the power system within the rolling window, the multi-scale clearing model is simplified to obtain a simplified clearing model. Based on the historical operating data of the power system, the multiple agents are iteratively trained. In each iteration, the multiple agents determine the boundary parameter set corresponding to the input state of the current round. The simplified clearing model, combined with the historical operating data, is used to simulate the boundary parameter set to obtain the input state of the next round and the execution reward of the current round. The iteration is completed when the execution reward converges to the optimal value. The multiple agents obtained after the last iteration are combined to construct a running boundary prediction model.

3. The rolling optimization clearing method for a power system according to claim 2, characterized in that, The plurality of agents includes a primary policy agent, a network evaluation agent, and a backup evaluation agent; the step of determining the boundary parameter set corresponding to the input state of the current round through the plurality of agents specifically includes: The network evaluation agent performs forward propagation prediction on the input state of the current round to determine the network risk assessment index of the current round, and the backup evaluation agent performs forward propagation prediction on the input state of the current round to determine the backup risk assessment index of the current round. By integrating the input state of the current round, network risk assessment indicators, and backup risk assessment indicators, a fused state feature is obtained. Then, through the main policy agent, the fused state feature is forward propagated to predict the boundary parameter set of the current round.

4. The rolling optimization clearing method for a power system according to claim 2, characterized in that, The process involves simulating the boundary parameter set using the simplified clearing model and historical operational data to obtain the input state for the next round and the execution reward for the current round. Specifically, this includes: The simplified clearing model is used to calculate the boundary parameter set, determine the simulation scheduling plan, and perform simulation based on the simulation scheduling plan and the historical running data to obtain the simulation running data and the input state for the next round. Based on the simulated operation data, the basic reward, time scale penalty, and collaborative reward for the current round are calculated sequentially and aggregated to obtain the execution reward for the current round.

5. The rolling optimization clearing method for a power system according to claim 1, characterized in that, The plurality of intelligent agents includes a primary policy intelligent agent, a network evaluation intelligent agent, and a backup evaluation intelligent agent; the step of inputting the operating state vector into a pre-trained operating boundary prediction model to dynamically predict the operating boundary of the power system, thereby obtaining the dynamic operating boundary parameter set of the power system, specifically includes: The network evaluation agent performs forward propagation prediction on the running state vector to obtain network risk assessment indicators, and the backup evaluation agent performs forward propagation prediction on the running state vector to obtain backup risk assessment indicators. By fusing the operating state vector, the network risk assessment index, and the backup risk assessment index, a fused state feature is obtained. The fused state feature is then forward-propagated and predicted by the main policy agent to obtain a dynamic operating boundary parameter set.

6. The rolling optimization clearing method for a power system according to claim 1, characterized in that, The dynamic operating boundary parameter set includes a multi-objective composite weight vector, a dynamic correction factor for reserve capacity demand, and a dynamic buffer for cross-sectional power limits; based on the dynamic operating boundary parameter set, a multi-scale clearing model is constructed with the objective of minimizing the operating cost of the power system within a preset rolling window, specifically including: Based on the multi-objective synthesized weight vector, a clearing objective function is constructed with the goal of minimizing the operating cost of the power system within a preset rolling window; Based on the power system design, and in conjunction with the dynamic correction factor for reserve capacity demand and the dynamic buffer of the cross-sectional power limit, multiple associated constraints are set for the clearing objective function; wherein, the multiple associated constraints include power balance constraints, network power flow security constraints, physical operation limit constraints, and system service coupling constraints. A multi-scale clearing model is constructed based on the clearing objective function and the multiple associated constraints.

7. A rolling optimization clearing method for a power system according to claim 6, characterized in that, The step of constructing a clearing objective function based on the multi-objective synthesized weight vector, with the objective of minimizing the operating cost of the power system within a preset rolling window, specifically includes: Based on the system design of the power system, the total operating cost of the power system is decomposed to determine various types of sub-operating costs. Based on the system design, each type of sub-operating cost is modeled to obtain various sub-operating cost functions of the power system. With the goal of minimizing the operating cost of the power system within a preset rolling window, a clearing objective function is constructed by assigning corresponding dynamic weights to each sub-operating cost function through the multi-objective composite weight vector.

8. A rolling optimization clearing system for a power system, characterized in that, It includes a boundary prediction module, a clearing model construction module, and a clearing solution scheduling module; The operation boundary prediction module is used to construct the operation state vector of the power system based on the operation data of the power system to be scheduled at the current rolling optimization time, and input the operation state vector into a pre-trained operation boundary prediction model to dynamically predict the operation boundary of the power system and obtain the dynamic operation boundary parameter set of the power system; wherein, the operation boundary prediction model is constructed based on deep reinforcement learning and combined training of multiple agents. The clearing model construction module is used to construct a multi-scale clearing model based on the dynamic operating boundary parameter set, with the objective of minimizing the operating cost of the power system within a preset rolling window; wherein the rolling window takes the current rolling optimization time as the starting boundary; the rolling window includes multiple sections with different time scales; The clearing solution scheduling module is used to solve the multi-scale clearing model, obtain the clearing decision of the power system, and schedule the power system according to the clearing decision to complete the single-round rolling optimization clearing of the power system.

9. A computer device, characterized in that, include: processor; Memory; A computer program stored in the memory and configured to be executed by the processor; When the processor executes the computer program, it implements a rolling optimization clearing method for a power system as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of instructions adapted for loading by a processor to execute a rolling optimization clearing method for a power system as described in any one of claims 1 to 7.

Images