A distributed closed-loop predictive-scheduling method and system for unit commitment of renewable energy power system

Abstract

The application discloses a kind of distributed closed-loop prediction-scheduling methods and systems for renewable energy power system unit combination, belong to power system dispatching field.The method is respectively in power system operation subject side and multiple renewable energy prediction subject side acquisition and local save system operation data, prediction characteristic data and actual power data;Each prediction subject first local training integrated learning prediction model, then carry out distributed closed-loop iteration through prediction copy consistency.In iteration, operation subject locally updates operation subject side prediction copy, unit combination variable, rescheduling variable, consensus prediction variable and penalty parameter, each prediction subject parallelly updates renewable energy prediction side prediction copy and meta-learner parameter, only exchange consensus prediction variable, penalty parameter and regularization term value between subjects.After convergence, each prediction subject generates renewable energy power prediction value of dispatching cost orientation, and operation subject calculates unit combination result accordingly and issues control instruction.

Description

Technical Field

[0001] This invention belongs to the technical field of power system optimization and control, and mainly relates to a distributed closed-loop prediction-dispatch method and system for unit combination in power systems containing renewable energy. Background Technology

[0002] As the penetration rate of renewable energy sources such as wind and solar power in the power system continues to increase, the reliance of unit combination problems on renewable energy power forecasting results has significantly strengthened. Traditional open-loop forecasting-dispatch processes typically train forecasting models using statistical indicators such as mean squared error and mean absolute error, and then input the forecast results into the unit combination model. This process makes the forecasting model primarily pursue statistical accuracy, but it cannot directly perceive the economic consequences of forecasting errors in unit start-up and shutdown, reserve configuration, wind and solar curtailment, load shedding, and rescheduling costs. Therefore, even small statistical errors can still lead to high operating costs in actual dispatching.

[0003] Closed-loop forecasting-scheduling methods feed back the downstream unit combination or rescheduling costs to the upstream forecasting model, enabling the model to learn scheduling cost-oriented forecasting patterns and thus obtain scheduling cost-oriented renewable energy power forecasts. However, existing closed-loop forecasting-scheduling methods often employ centralized modeling and solving, requiring the historical meteorological characteristics, actual power generation, forecasting model structure, and forecasting model parameters of the renewable energy forecasting entity to be centralized at the power system operator's side, or requiring the power system operator's network constraints, unit operation constraints, cost coefficients, and scheduling models to be exposed to the forecasting entity. These data and models typically belong to different operators, possessing commercial sensitivity and operational safety attributes, making direct centralized sharing difficult in practical engineering. Furthermore, the closed-loop forecasting-scheduling problem inherently involves a nested relationship between upper-level forecasting model parameter optimization and lower-level unit combination mixed-integer optimization, belonging to a two-level mixed-integer programming problem. Since the lower-level unit combination problem includes binary start-stop variables, its optimal solution mapping typically exhibits discontinuous, non-smooth, and ensemble-valued characteristics, making convergence difficult for general distributed optimization algorithms.

[0004] Therefore, a distributed closed-loop forecasting-dispatch method is needed for the combination of generating units in power systems containing renewable energy. This method would enable power system operators and multiple renewable energy forecasting entities to collaboratively solve the closed-loop forecasting-dispatch model according to a defined time sequence. This would allow them to obtain renewable energy power forecasting results and generating unit combination control instructions that are oriented towards dispatching costs while protecting the local privacy data and models of each entity. Summary of the Invention

[0005] The purpose of this invention is to provide a distributed closed-loop prediction-scheduling method and system for power systems with renewable energy units, so as to solve the problem that closed-loop prediction-scheduling in the prior art requires centralized sharing of private data and models of multiple entities.

[0006] To achieve the above objectives, the present invention provides the following technical solution.

[0007] Firstly, a distributed closed-loop prediction-dispatch method for unit combinations in power systems containing renewable energy is provided, including the following steps:

[0008] S1. Set the training set time window, test set time window, and time resolution. The training set time window corresponds to historical moments and is used to construct historical samples and train the closed-loop prediction-dispatch model. The test set time window corresponds to future dispatch moments and is used to generate renewable energy power predictions and unit combination results for future dispatch moments. The time resolution can be at the hourly level or a finer granularity, specifically determined by the power system dispatch requirements.

[0009] S2. The power system operator and multiple renewable energy forecasting entities collect data and store it locally. The power system operator collects system operating parameters within the training set time window, including line topology, line power flow distribution factor, line transmission capacity, node load, unit output upper and lower limits, ramp limits, start-up and shutdown costs, fuel costs, reserve demand, and reschedule costs, and stores these system operating parameters locally. Renewable energy forecasting entities... Feature data for renewable energy forecasting within the collection time window and actual power generation data of renewable energy This includes local private meteorological data, relevant parameters of renewable energy units, and actual power generation, forming a local training sample. ,in, For the first A training sample set of renewable energy entities. The samples in the training sample set are numbered, and the samples are stored locally in the corresponding prediction subject.

[0010] S3. Each renewable energy forecasting entity trains its basic forecasting model locally and generates historical baseline forecast outputs within the training set time window. Specifically, each renewable energy forecasting entity constructs a first-layer basic forecasting model and a second-layer meta-learner locally. The model type, parameter settings, and training process of the first-layer basic forecasting model are determined and saved locally by the corresponding renewable energy forecasting entity. The second-layer meta-learner is used to combine the outputs of the first-layer basic forecasting model. Each renewable energy forecasting entity trains its basic forecasting model based on locally saved data, generates basic forecast outputs within the training set time window, and obtains reference parameters for the meta-learner. Specifically, for the renewable energy forecasting entity... Its local base prediction model set is denoted as , No. The basic prediction models are: The historical baseline prediction output is:

[0011]

[0012] Each renewable energy forecasting entity can use neural networks, tree models, linear models, or other forecasting models as its base forecasting model. The structure, parameters, and training samples of the base forecasting model are all stored locally. Each forecasting entity can also train an initial meta-learner based on statistical error to obtain meta-learner reference parameters. .

[0013] S4. Execute distributed closed-loop prediction-scheduling iteration. This step is the core of the invention, including outer-layer reconstruction and decomposition iteration and inner-layer overlapping block coordinate descent-alternating direction multiplier method iteration. The outer-layer iteration is responsible for transforming the bi-layer mixed integer programming into a sequence of master problems and subproblems, and gradually adding optimality cuts; the inner-layer iteration is responsible for distributedly solving the penalized relaxation master problem generated in each round of outer-layer iteration without centrally sharing the private data and models of each entity. Specifically, the outer-layer reconstruction and decomposition iteration includes establishing a closed-loop prediction-scheduling bi-layer mixed integer programming model, solving the penalized relaxation master problem using the inner-layer overlapping block coordinate descent-alternating direction multiplier method, solving the first subproblem locally by the power system operator, solving the second subproblem locally by the power system operator, and executing iteration return or termination according to the inner-layer convergence condition and the outer-layer convergence condition; in the inner-layer overlapping block coordinate descent-alternating direction multiplier method iteration, the power system operator and multiple renewable energy prediction entities collaboratively solve the penalized relaxation master problem. Step S4 specifically includes steps S41 to S44.

[0014] S41. Establish a closed-loop prediction-scheduling two-layer hybrid integer programming model. The model, with the difference between the scheduling operation cost and the baseline optimal cost, and the meta-learner regularization term as the objective, is expressed as:

[0015]

[0016] in, For the training sample set, This is the set of meta-learner parameters for each renewable energy forecasting subject. For training samples Optimal values ​​of continuous variables for unit combination. For training samples Optimal values ​​for start-up and shutdown variables of the unit combination. For continuous decision variables in the unit combination model, For binary decision variables in the unit combination model, For training samples The rescheduled variable under, The cost coefficient vector is a continuous variable in the unit combination objective function. , For the fuel cost coefficient vector, This is a vector of expected load shedding and wind curtailment penalty cost coefficients. This is a vector of unit start-up and shutdown cost coefficients. This is a vector of rescheduling cost coefficients, which includes the upward and downward adjustment of ancillary service costs, the actual load shedding, and the cost of wind curtailment penalties. For training samples The corresponding pre-calculated baseline optimal operating cost, These are the regularization coefficients for the meta-learner parameters. For the meta-learner parameter regularization term, For training samples The corresponding renewable energy power forecast, Indicates sample The following is a prediction feature matrix composed of the outputs of various basic prediction models, representing the actual power generation of renewable energy. Based on actual renewable energy power The rescheduling feasible domain, For predicted power The two feasible domains of unit combinations are both stored locally by the power system operator.

[0017] S42. The penalized relaxation master problem is solved using the overlapping block coordinate descent-alternating direction multiplier method. Specifically, it is transformed into the penalized relaxation master problem. The first subproblem SP1 and the second subproblem SP2; in the In the outer-layer reconstruction and decomposition iteration, the power system operator initializes the outer-layer iteration count. The lower realm Upper Realm and outer convergence threshold and in the wheel( Constructing the penalized relaxation master problem in the outer layer reconstruction and decomposition iteration :

[0018]

[0019] in, For the first The optimal cut corresponds to the fixed unit start-up and shutdown variables. and These are all coefficient matrices representing the constraints of the unit combination problem. For the bilinear term in the relaxation optimality cut, These are the dual variables corresponding to the unit combination constraints. The right-hand side of the unit combination constraint is determined by the predicted power parameterization. For use in relaxing bilinear terms Auxiliary variables, This is the penalty coefficient.

[0020] The penalty relaxation master problem is solved iteratively using the overlapping block coordinate descent-alternating direction multiplier method described in steps S421 to S428.

[0021] S421, Set the overlapping block coordinate descent variable block. Let:

[0022]

[0023] in, This mainly corresponds to the prediction-side variables caused by the prediction power and the meta-learner parameters. Includes unit start-up and shutdown variables and dual variables. It includes continuous variables for unit combination and rescheduling variables.

[0024] S422, Solution for Power System Operators - A fixed problem. In the... During the sub-overlapping block coordinate descent iteration, the power system operating entity is fixed. And solve it locally:

[0025]

[0026] This step involves unit start-up and shutdown variables, dual variables, network constraints, and rescheduling constraints. It is completed solely by the power system operator based on a local model and does not require communication with renewable energy forecasting entities.

[0027] S423, Construction - Fixed issues. The main operating entity of the power system is fixed. The solution is obtained by the power system operator and various renewable energy forecasting entities through a distributed solution using the alternating direction multiplier method. - Fixed issue, introducing operational entity to predict replicas Predicting subject and predicting copy and consensus predictor variables and define:

[0028]

[0029] but - Fixed problems can be written as:

[0030]

[0031] Among them, represents the set of local constraints for the power system operating entity. For the local forecasting constraint set of renewable energy forecasting subjects, Maintained by the power system operator. Maintained by the renewable energy forecasting entity, the two are coupled only through the consistency constraint of the predicted power replica. For the objective function on the operating entity side, To predict the objective function on the subject side. and The expressions are as follows:

[0032]

[0033]

[0034] in, Main body for renewable energy forecasting The meta-learner parameter vector satisfies This is used to combine the outputs of multiple base prediction models on the subject's local side into a prediction copy for the subject:

[0035]

[0036] S424. Initialize the alternating direction multiplier method iteration. Power system operators set or inherit consensus prediction variables. Dual variables on the operating entity side Predicting the dual variables on the subject side and penalty parameters .like Then, the consensus predictor variables and dual variables obtained from the previous round of overlapping block coordinate descent iteration are used for hot start.

[0037] S425, the power system operator and various renewable energy forecasting entities solve the local subproblem in parallel. In the... In the alternating direction multiplier iteration, the power system operator solves the operation sub-problem locally:

[0038]

[0039] Meanwhile, renewable energy forecasting entities Parallel solution based on local base prediction output and local prediction constraints:

[0040]

[0041] in, , , As the dual variable on the operating entity side, To predict the dual variables on the subject side, For the first Penalty parameters in the alternating direction multiplier method iteration;

[0042] S426. Information exchange using the alternating direction multiplier method. Power system operators communicate with renewable energy forecasting entities. Send consensus predictor variables and penalty parameters:

[0043]

[0044] Renewable energy forecasting entities Returns the predicted copy and regularization term values ​​of the predicted subject:

[0045]

[0046] in, To send to renewable energy forecasting entities consensus predictor variables, For penalty parameters, For predicting the subject, predict the copy. This is the value of the regularization term.

[0047] In addition to the variables mentioned above, renewable energy forecasting entities will not send local characteristic data, actual power generation data, basic forecasting model structure, basic forecasting model parameters, and meta-learner parameters; power system operators will not send system operating parameters, cost coefficients, network constraints, unit operating constraints, rescheduling constraints, and local decision variables.

[0048] S427. Power system operators update consensus forecast variables, and power system operators and renewable energy forecasters respectively update local dual variables:

[0049]

[0050]

[0051]

[0052] S428. Determine if the inner loop of the overlapping block coordinate descent-alternating direction multiplier method has converged. The convergence conditions include the convergence condition of the alternating direction multiplier iteration, the convergence condition of the overlapping block coordinate descent iteration, and the convergence condition of the outer reconstruction and decomposition iteration. For the original residual of the alternating direction multiplier iteration... , and dual residuals for:

[0053]

[0054]

[0055] If the following conditions are met:

[0056]

[0057] Then it is determined that the alternating direction multiplier iteration is convergent, where and These are the original feasibility threshold and the dual feasibility threshold, respectively; otherwise, the penalty parameter is updated according to the following rule, letting... Then return to step S425, where the penalty parameters are updated according to the relative magnitudes of the original residual and the dual residual:

[0058]

[0059] in, and These are the coefficients for increasing and decreasing the penalty parameter, respectively. This is the coefficient for judging the relative magnitude of the original residual and the dual residual.

[0060] After the alternating direction multiplier iteration converges, if the objective function value of the penalized relaxation master problem in two consecutive overlapping block coordinate descent iterations satisfies:

[0061]

[0062] If the overlapping block coordinate descent-alternating direction multiplier method for solving the penalized relaxation master problem converges, then the current solution to the penalized relaxation master problem is output. ,in Set the threshold for convergence of the overlapping block coordinate descent iteration; otherwise, let Then return to step S422.

[0063] S43. The power system operator solves the first subproblem, SP1. After the penalized relaxation main problem converges, the power system operator obtains the current predicted power. Heyuan learner parameters And solve the first subproblem SP1 locally:

[0064]

[0065] This step only involves the unit combination model and constraints of the power system operator and does not require interaction with renewable energy forecasting entities.

[0066] S44. The power system operator solves the second subproblem SP2 and performs an outer-layer convergence check. The power system operator solves the problem locally:

[0067]

[0068] This step is used to evaluate the unit combination and reschedule costs under the current predicted power and to generate an optimality cut for the next round of the outer master problem. The power system operator updates the lower bound based on the objective function value of the penalized relaxation master problem. And update based on the objective function value of the second subproblem SP2 If the following conditions are met:

[0069]

[0070] Then it is determined that the outer layer reconstruction and decomposition iterations have converged, where, Step S4 ends when the outer layer reconstruction and decomposition iteration convergence threshold is reached; otherwise, the power system operator generates a new optimality cut based on the solution of SP2, and adds the optimality cut to the next round of penalized relaxation principal problem, letting... Then return to step S42 to continue execution.

[0071] S5. Each renewable energy forecasting entity constructs an ensemble learning forecasting model based on the converged meta-learner parameters and its local basic forecasting model. Renewable Energy Forecasting Entity Based on the local base prediction model and the converged meta-learner parameters The prediction form of the linear meta-learner in constructing an ensemble learning prediction model is as follows:

[0072]

[0073] in, , For the main body of renewable energy forecasting, Main body for renewable energy forecasting The set of basic prediction models, For the first The local prediction output of each basic prediction model For the intercept term, For the first The meta-learner weights corresponding to each basic prediction model, wherein the meta-learner parameters are: The meta-learner parameter regularization term only applies to the linear meta-learner parameters. And it does not affect the basic prediction model. When L2 regularization is used:

[0074]

[0075] When L1 regularization is used:

[0076]

[0077] in, Main body for renewable energy forecasting Based on the meta-learner reference parameters obtained in step S3;

[0078] The regularization coefficient can be determined based on the running cost of the validation set:

[0079]

[0080] in, This is the candidate set of regularization coefficients. To use regularization coefficients The unit combination and rescheduling operation costs are then obtained on the validation set.

[0081] S6. Renewable energy forecasting entities predict renewable energy power within the test set time window. For the test set samples, each renewable energy forecasting entity uses the ensemble learning prediction model, that is, first uses the local basic prediction model to generate basic prediction output, and then combines it through the converged linear meta-learner to obtain the dispatch cost-oriented renewable energy power prediction value, and sends the prediction value to the power system operator.

[0082] S7. Power system operator's computer group combination results. The power system operator inputs the dispatch cost-oriented renewable energy power forecast value obtained in step S6 into the unit combination model, solves the unit combination problem at future dispatch times locally, and obtains the unit start-up, shutdown, output, reserve and related dispatch plans.

[0083] S8. The power system operator issues control instructions to the power system generating units. The power system operator converts the generating unit combination results obtained in step S7 into control instructions and issues them to the corresponding power system generating units according to the dispatch execution process.

[0084] Secondly, a distributed closed-loop forecasting-dispatch system for power systems with renewable energy units is provided, including:

[0085] The local data acquisition module is configured to collect system operating parameters on the side of the power system operator and to collect forecast characteristic data and actual renewable energy power generation data on multiple renewable energy forecasting entities.

[0086] The basic forecast training module is configured to enable each renewable energy forecasting entity to train its basic forecasting model locally and generate historical baseline forecast outputs.

[0087] A distributed closed-loop iteration module is configured to perform distributed closed-loop prediction-scheduling iteration. The distributed closed-loop iteration module includes an outer reconstruction and decomposition module and an inner distributed solution module. The outer reconstruction and decomposition module is configured to establish a closed-loop prediction-scheduling two-layer hybrid integer programming model, construct a penalized relaxation master problem, solve the first subproblem, and solve the second subproblem. The inner distributed solution module is configured to perform overlapping block coordinate descent-alternating direction multiplier method iteration to collaboratively solve the penalized relaxation master problem.

[0088] The information exchange module is configured to transmit consensus forecast variables, penalty parameters, forecast copies of forecasts from forecast entities, and regularization terms between the power system operator and each renewable energy forecast entity.

[0089] The integrated model building module is configured to enable each renewable energy forecasting entity to build an integrated learning forecasting model based on the converged meta-learner parameters and the local basic forecasting model, respectively.

[0090] The test set prediction module is configured to generate scheduling cost-oriented renewable energy power prediction values ​​within the test set time window based on the ensemble learning prediction model.

[0091] The control instruction generation module is configured to enable the power system operator to issue control instructions based on the computer group combination results of the dispatch cost-oriented renewable energy power forecast.

[0092] Compared with existing technologies, this invention discloses a distributed closed-loop prediction-dispatch method and system for power systems with renewable energy units. Compared with traditional open-loop prediction-dispatch methods and centralized closed-loop prediction-optimization methods, it has significant advantages in terms of information disclosure control, dispatch economy and engineering feasibility.

[0093] Regarding information disclosure control, this invention does not require power system operators to centrally acquire the private meteorological characteristics, actual power generation data, basic prediction model structure, basic prediction model parameters, and meta-learner parameters of each renewable energy forecasting entity. Nor does it require renewable energy forecasting entities to acquire the system operating parameters, cost coefficients, network constraints, unit operating constraints, rescheduling constraints, and local decision variables of the power system operators. Each entity only exchanges consensus forecasting variables, penalty parameters, forecasting entity copies, and regularization term values ​​during the iteration process, thereby significantly reducing the scope of information disclosure across entity data and models. This allows closed-loop prediction-scheduling training to be completed under conditions of local data storage, local modeling, and local computation across multiple entities.

[0094] In terms of economic benefits, this invention, without centrally sharing private data and models, still connects the training process of the renewable energy prediction model with actual operational decisions such as downstream unit combination and rescheduling through a closed-loop prediction-scheduling mechanism. This allows the meta-learner parameters to perceive the impact of prediction errors on start-up and shutdown costs, fuel costs, reserve configuration, wind and solar curtailment, load reduction, and rescheduling costs. The resulting scheduling cost-oriented prediction values ​​can avoid error types that have a significant impact on system operating costs. Compared with open-loop prediction methods that only target statistical errors, it can effectively reduce actual operating costs. At the same time, its economic improvement effect is close to that of centralized closed-loop prediction-optimization methods, maintaining high scheduling economy while reducing information disclosure.

[0095] In terms of algorithm implementation, this invention decomposes the closed-loop prediction-scheduling bi-level mixed integer programming problem into an outer-layer reconstruction and decomposition iteration and an inner-layer overlapping block coordinate descent-alternating direction multiplier method iteration. The outer-layer iteration updates the cutting plane and upper and lower bounds stepwise through the penalized relaxation master problem, the first subproblem, and the second subproblem; the inner-layer iteration achieves distributed collaborative solution of the master problem through operator prediction replicas, prediction entity prediction replicas, and consensus prediction variables. This nested solution structure clarifies the computational tasks, interaction variables, and convergence judgment logic undertaken by the power system operator and the renewable energy prediction entity, improving the solvability and deployability of the complex bi-level mixed integer programming problem in actual power systems.

[0096] Regarding model compatibility, this invention allows each renewable energy forecasting entity to independently select and train its local base forecasting model. The base forecasting model can employ neural networks, tree models, linear models, or other forecasting models; only the second-layer linear meta-learner and its forecasting copy participate in the closed-loop optimization. Therefore, this invention not only preserves the existing forecasting model systems and private model assets of each forecasting entity, but also achieves near-centralized closed-loop forecasting-optimization economic improvements through distributed closed-loop optimization, combining model compatibility, cross-entity collaborative capabilities, and engineering practical value. Attached Figure Description

[0097] Figure 1 This is a diagram illustrating the overall structure of the distributed closed-loop prediction-scheduling method of the present invention, showing the collaborative relationship between the power system operator and the renewable energy prediction entity, as well as the corresponding theoretical model structure.

[0098] Figure 2 The flowcharts for the outer-layer reconstruction and decomposition (R&D) iteration and the inner-layer overlapping block coordinate descent-alternating direction multiplier method (OBCD-ADMM) iteration of this invention are shown in (a) the outer-layer reconstruction and decomposition iteration flowchart, illustrating the main problem, the first subproblem, the second subproblem, the generation of optimality cuts, and the outer-layer convergence judgment process; (b) the inner-layer OBCD-ADMM iteration flowchart, illustrating the solution of the power system operation entity. - Fixed problems, collaborative solutions involving operating entities and forecasting entities - The process of fixing the problem, updating global and dual variables, and determining convergence.

[0099] Figure 3 The figures show a comparison of the scheduling results and renewable energy power prediction results of the traditional open-loop forecasting-schedule method and the method of the present invention on a typical scheduling day in the embodiments of the present invention. (a) shows the unit output, load shedding, rescheduling and renewable energy power prediction results under the traditional open-loop forecasting-schedule method; (b) shows the unit output, load shedding, rescheduling and renewable energy power prediction results under the distributed closed-loop forecasting-schedule method of the present invention.

[0100] Figure 4 The diagram shows the iterative convergence curves of the distributed closed-loop prediction-scheduling algorithm described in this embodiment of the invention. (a) shows the changes in the upper bound, the objective value of the penalized relaxation master problem, and the optimal gap during the outer R&D iteration process; (b) shows a typical inner OBCD iteration process. -Fixed problem target value, - The curves showing the change between the target value and the target value gap in the fixed problem; (c) The curves showing the change between the original residual, the dual residual and their threshold during the ADMM iteration process; (d) The curves showing the change between the L2 norm of the operator's predicted replica and the predictor's predicted replica during the ADMM iteration process.

[0101] Figure 5 The figures show a comparison of the cost reduction rate and solution time of the present invention's method (DCPO) and the centralized closed-loop prediction-scheduling method (CPO) in small-scale and large-scale systems in the test examples of this invention. (a) Comparison of the cost reduction rate of the two methods in small-scale systems; (b) Comparison of the solution time of the two methods in small-scale systems; (c) Comparison of the cost reduction rate of the two methods in large-scale systems; (d) Comparison of the solution time of the two methods in large-scale systems. Detailed Implementation

[0102] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments and test examples are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. This embodiment uses wind power as a renewable energy source for illustration, but other renewable energy power prediction and unit combination scheduling scenarios with predictive uncertainty, such as photovoltaic and hydropower, can also be implemented by reference.

[0103] Example 1

[0104] This embodiment uses an improved IEEE 9-node wind power system as an example to illustrate the specific implementation process of the method of the present invention. The system includes 9 nodes, 9 branches, 3 thermal power units, 1 wind farm, and 3 loads. The power system operator is denoted as ISO, and the renewable energy forecasting operator is denoted as WF. In this embodiment, WF represents one wind farm; when multiple wind farms exist, each WF executes in parallel in the same manner. Simulation calculations are performed in the MATLAB 2022b environment, the optimization modeling tool is YALMIP, and the solver is Gurobi 12.0.1.

[0105] like Figure 1 As shown, this invention includes two types of entities at the implementation level: power system operators and renewable energy forecasting entities. ISO locally stores the unit combination model, network constraints, cost parameters, and rescheduling model; WF locally stores forecasting characteristics, actual wind power, the basic forecasting model, and its parameters. During the distributed closed-loop forecasting-scheduling iteration process, both parties only exchange variables related to forecast replica consistency, without centrally sharing their respective private data and models.

[0106] Step S1 involves setting the training set time window, the test set time window, and the time resolution. In this embodiment, the training set time window is 7 days, the test set time window is 1 day, and the time resolution is 1 hour; therefore, each scheduling day contains 24 time slots. The training sample set is denoted as... For training samples WF locally stores meteorological feature vectors and actual wind power .

[0107] In step S2, the power system operator and multiple renewable energy forecasting entities collect and store data locally. Specifically, the ISO collects and stores locally parameters such as line topology, line power flow distribution factor, line transmission capacity, node load, turbine output upper and lower limits, ramp limits, start-up and shutdown costs, fuel costs, reserve requirements, and reschedule costs. The WF collects and stores locally meteorological characteristic data, wind turbine-related parameters, and actual wind power data used for wind power forecasting. This data is stored locally by its respective entity and is not centrally transmitted between the ISO and WF.

[0108] In step S3, each renewable energy forecasting entity trains its basic forecasting model locally and generates historical baseline forecast outputs. In this embodiment, WF constructs a first-layer basic forecasting model and a second-layer meta-learner locally. The first-layer basic forecasting model includes a multilayer perceptron, convolutional neural network, long short-term memory network, extreme gradient boosting, random forest, and support vector machine. For the... The historical baseline prediction output of the basic prediction model is as follows:

[0109]

[0110] The second-layer meta-learner is used to combine the outputs of the first-layer basic prediction model, and its form is as follows:

[0111]

[0112] in, These are the meta-learner parameters. WF first trains an initial meta-learner with the objective of minimizing the mean squared error, thus obtaining the meta-learner reference parameters. This is used for subsequent regularization term calculations.

[0113] Step S4 involves executing the distributed closed-loop prediction-scheduling iteration. This step is the core of this embodiment, including the outer-layer reconstruction and decomposition iteration and the inner-layer overlapping block coordinate descent-alternating direction multiplier method iteration. The outer-layer iteration is responsible for transforming the closed-loop prediction-scheduling two-layer mixed integer programming into a penalized relaxation master problem, a first subproblem SP1, and a second subproblem SP2. The inner-layer iteration is responsible for solving the penalized relaxation master problem generated in each round of the outer-layer iteration in a distributed manner, without centrally sharing ISO and WF private data. The algorithm structure is as follows: Figure 2 As shown.

[0114] Execute step S41: Establish a closed-loop prediction-scheduling two-layer hybrid integer programming model. The model, with the difference between the scheduling operation cost and the baseline optimal cost, and the meta-learner regularization term as its objective, is expressed as:

[0115]

[0116] in, For unit combination continuous variables, For unit start-up and shutdown variables, For rescheduling variables, For day-ahead unit combination feasible domains maintained locally by ISO, This is a rescheduling feasible domain maintained locally by ISO.

[0117] Execute step S42, using the overlapping block coordinate descent-alternating direction multiplier method to solve the penalized relaxation master problem. Initialize the outer iteration count, cut plane set, lower bound, upper bound, and outer convergence threshold using ISO. In the... In the outer-layer iteration, an auxiliary variable is introduced to handle the bilinear term in the optimality cut. Sum of penalty coefficients Constructing the penalized relaxation principal problem:

[0118]

[0119] The penalized relaxation master problem includes prediction constraints, unit combination constraints, rescheduling constraints, and an optimality cut generated based on historical unit start-up and shutdown solutions. Let:

[0120]

[0121] In the In the sub-overlapping block coordinate descent iteration, ISO is first fixed. Local solution - Fixed problem:

[0122]

[0123] This problem involves unit start-up and shutdown variables, dual variables, network constraints, operational constraints, and rescheduling constraints, and is solved solely by ISO based on a local model.

[0124] Subsequently, ISO fixed The solution is obtained by ISO and WF using the alternating direction multiplier method in a distributed manner. - Fixed problem:

[0125]

[0126] To achieve distributed solving, ISO-side predictive replicas are introduced. WF-side predicted replica and consensus predictor variables And set consistency constraints:

[0127]

[0128] In the In the alternating direction multiplier iteration, ISO solves the operational side problem locally:

[0129]

[0130] WF local solution to the prediction side subproblem:

[0131]

[0132] in, , ISO sends consensus predictor variables and penalty parameters to WF. WF returns the predicted principal copy and regularization term value to ISO. In addition to the variables mentioned above, WF does not send local meteorological characteristics, actual wind power, basic forecast model structure, basic forecast model parameters, and meta-learner parameters; ISO does not send system operating parameters, cost coefficients, network constraints, unit operating constraints, rescheduling constraints, and local decision variables.

[0133] After solving the local subproblem, ISO updates the consensus predictor variables:

[0134]

[0135] ISO and WF update their local dual variables respectively:

[0136]

[0137]

[0138] When the original residual and dual residual of the alternating direction multiplier iteration meet the preset threshold, the alternating direction multiplier iteration is determined to be converged; when the change in the objective function value of the penalized relaxation master problem of two consecutive overlapping block coordinate descent iterations meets the preset threshold, the penalized relaxation master problem is determined to be converged.

[0139] Execute step S43: The power system operator solves the first subproblem SP1. After the penalized relaxation master problem converges, ISO obtains the current predicted power. Heyuan learner parameters And solve it locally:

[0140]

[0141] Execute step S44: The power system operator solves the second subproblem SP2 and performs an outer-layer convergence check. ISO solves locally:

[0142]

[0143] ISO updates the lower bound of the outer layer based on the objective function value of the penalized relaxation master problem, and updates the upper bound of the outer layer based on the objective function value of SP2. If the difference between the upper and lower bounds of the outer layer satisfies the preset convergence condition, then step S4 ends; otherwise, ISO generates a new optimality cut based on the unit start-up and shutdown solution obtained from SP2, adds the optimality cut to the next round of penalized relaxation master problem, and returns to step S42.

[0144] In step S5, each renewable energy forecasting entity constructs an ensemble learning forecasting model based on the converged meta-learner parameters and its local base forecasting model. In this embodiment, WF obtains:

[0145]

[0146] In step S6, WF uses the ensemble learning prediction model to predict the wind power within the test set time window, obtains the wind power prediction value oriented towards scheduling cost, and sends the prediction value to ISO.

[0147] Step S7 involves ISO combining the computer group results based on the aforementioned scheduling cost-oriented wind power forecast values. Specifically, ISO will test the daily forecast power. Input the day-ahead unit combination model and solve for the unit start-up and shutdown status, unit output, reserve configuration and related scheduling plan.

[0148] In step S8, ISO converts the unit combination result into control commands and issues the control commands to the power system units.

[0149] This embodiment selects one typical scheduling day as the test set and the previous 7 days as the training set to compare the traditional open-loop prediction-scheduling method with the method of this invention. The results are as follows: Figure 3 As shown, traditional open-loop prediction-schedule methods train prediction models with the goal of minimizing statistical error, resulting in significant overestimation of wind power predictions on the scheduled dispatch day. Because the day-ahead unit combination model assumes higher available wind power, the system's scheduled conventional unit operating capacity and reserve capacity are insufficient during the day-ahead phase, leading to substantial upward adjustments and reschedules during the real-time phase, resulting in load reduction. In contrast, the method of this invention uses closed-loop prediction-schedule training, enabling the meta-learner parameters to perceive the impact of prediction errors on unit start-up and shutdown, reserve configuration, and reschedule costs, ultimately generating slightly conservative but more economical wind power predictions.

[0150] From a cost structure perspective, the traditional open-loop forecasting-dispatch method has a baseline fuel cost of RMB 12,693.48, an no-load cost of RMB 2,244.00, a start-stop cost of RMB 500.00, an upward rescheduling cost of RMB 3,321.51, and a downward rescheduling cost of RMB 25.79, resulting in a total operating cost of RMB 18,784.78 and a load reduction of 6.98 MW. The method of this invention has a baseline fuel cost of RMB 13,716.55, an no-load cost of RMB 2,319.00, a start-stop cost of RMB 800.00, an upward rescheduling cost of RMB 91.05, and a downward rescheduling cost of RMB 109.40, resulting in a total operating cost of RMB 17,035.99 and no load reduction. Although the method of this invention increases fuel costs, no-load costs, and start-stop costs during the day-ahead phase, it significantly reduces real-time upward rescheduling costs and eliminates load reduction, thus reducing the total operating cost by 9.31%.

[0151] From the perspective of forecasting indicators, the traditional open-loop forecasting-scheduling method has a mean squared error of 461.87, a mean absolute error of 17.197, and a mean absolute percentage error of 22.691%; while the method of this invention has a mean squared error of 651.06, a mean absolute error of 23.107, and a mean absolute percentage error of 24.917%. This indicates that the method of this invention does not simply pursue statistical forecasting accuracy, but rather forecasting effectiveness in the context of scheduling costs. Furthermore, the average over-prediction error of the traditional method is 27.148%, while the average over-prediction error of the method of this invention is reduced to 6.6476%; the average under-prediction error of the traditional method is 13.777%, while the average under-prediction error of the method of this invention is 26.578%. In scenarios involving wind turbine units, over-prediction of wind power is more likely to cause insufficient operation of conventional units, insufficient reserve capacity, and real-time load reduction, while moderate under-prediction usually only increases the day-ahead operation capacity and downward adjustment space. The method of this invention learns from this asymmetric economic consequence, thus proactively reducing the high-cost over-prediction error.

[0152] The calculation methods for underprediction error (MUPE) and overprediction error (MOPE) are as follows:

[0153]

[0154]

[0155] like Figure 4As shown, the method of the present invention can converge stably in this embodiment. The outer layer reconstruction and decomposition iteration can tighten the upper and lower bounds in fewer rounds; the objective function value of the inner layer overlapping block coordinate descent iteration quickly stabilizes; in the alternating direction multiplier method iteration, the original residuals show an overall decreasing trend, and the difference between the ISO side prediction replica and the WF side prediction replica gradually decreases. Since the subsequent round of overlapping block coordinate descent iteration can be warm-started using the consensus prediction variables and dual variables of the previous round of alternating direction multiplier method, the number of subsequent alternating direction method iterations is significantly reduced. This result shows that the present invention can not only achieve closed-loop prediction-scheduling training across subjects, but also obtain a stable distributed solution process under limited communication volume.

[0156] Test case

[0157] To further verify the economy and computational efficiency of the method of the present invention under different system scales, this test case uses an improved IEEE 9-node system and an improved IEEE 118-node system for testing. The improved IEEE 118-node system includes 118 nodes, 186 branches, 45 thermal power units, 5 wind power units, and 91 loads. The training set and test set are divided in the same way as in Example 1.

[0158] This test case sets the traditional open-loop prediction-scheduling method as the baseline and the centralized closed-loop prediction-scheduling method as the comparison method. The traditional open-loop prediction-scheduling method first trains the prediction model with statistical error as the objective, and then inputs the prediction results into the unit combination model; the centralized closed-loop prediction-scheduling method can feed the scheduling cost back to the prediction training, but requires centralized access to the prediction main data, prediction model, and ISO-side unit combination and rescheduling model; the method of this invention completes the closed-loop prediction-scheduling training without centrally sharing the above-mentioned private data and models.

[0159] like Figure 5 As shown, during 31 test days of the improved IEEE 9-node system, the centralized closed-loop prediction-scheduling method achieved an average cost reduction of 4.15% compared to the traditional open-loop prediction-scheduling baseline, while the method of this invention achieved an average cost reduction of 3.72%, a difference of 0.43 percentage points. This result indicates that the method of this invention, due to its distributed solution and limited information interaction, has a small economic advantage compared to the centralized closed-loop prediction-scheduling method; however, this advantage is compensated by the fact that the prediction subject data, basic prediction model, and ISO-side scheduling model are not centrally disclosed. In terms of computation time, the average solution time of the method of this invention is 194 seconds, while the average solution time of the centralized closed-loop prediction-scheduling method is 320 seconds, demonstrating that the method of this invention has higher solution efficiency and more stable computational performance.

[0160] In the improved IEEE 118-node system, the average cost reduction rate of the centralized closed-loop predictive-schedule method was 2.82%, while the average cost reduction rate of the method of this invention was 2.81%, a difference of only 0.01 percentage points. Compared with small-scale systems, the economic gap in large-scale systems did not widen but rather narrowed further, indicating that the distributed reconfiguration method of this invention does not significantly lose scheduling economy as the system size increases. This result shows that the method of this invention can maintain decision quality close to that of the centralized closed-loop predictive-schedule method in large-scale power systems.

[0161] In terms of computational efficiency, in the improved IEEE 118-node system, the average solution time of the method of this invention is 2446 seconds, while the average solution time of the centralized closed-loop prediction-scheduling method is 4443 seconds. The average solution time of the method of this invention is approximately half that of the centralized method. Meanwhile, the solution time of the centralized closed-loop prediction-scheduling method fluctuates significantly, exceeding 15000 seconds on some test days; the solution time of the method of this invention is more concentrated, indicating better computational stability in response to changes in test day data. For real-world day-ahead scheduling scenarios, stable and predictable computation time is just as important as average economic efficiency; therefore, this result further demonstrates the engineering deployability of this invention.

[0162] In summary, this test case demonstrates that the method of this invention can achieve significant reductions in information disclosure and improvements in computational efficiency in small-scale systems with relatively small economic losses. In large-scale systems, the method of this invention maintains almost the same cost reduction effect as centralized closed-loop prediction-scheduling methods, while significantly shortening the average solution time and reducing solution time fluctuations. Therefore, this invention is applicable to combined scheduling scenarios involving renewable energy units, where power system operators and multiple renewable energy forecasting entities participate, but cannot centrally share private data and models.

Claims

1. A distributed closed-loop prediction-dispatch method for unit combinations in power systems containing renewable energy, characterized in that, Includes the following steps: S1. Set the training set time window, the test set time window, and the time resolution. The training set time window corresponds to historical moments, and the test set time window corresponds to future scheduling moments. S2. The power system operator and multiple renewable energy forecasting entities collect data and store it locally. The power system operator collects system operating parameters within the training set time window, including line topology, line power flow distribution factor, line transmission capacity, node load, unit output upper and lower limits, ramping limits, start-up and shutdown costs, fuel costs, reserve requirements, and rescheduling costs. Each renewable energy forecasting entity collects characteristic data for renewable energy forecasting and actual renewable energy power generation data within the training set time window. The characteristic data includes private meteorological data and relevant parameters of renewable energy units. S3. Each renewable energy forecasting entity trains its basic forecasting model locally and generates historical baseline forecasting outputs. Each renewable energy forecasting entity constructs a first-layer basic forecasting model and a second-layer meta-learner locally. The model type, parameter settings, and training process of the first-layer basic forecasting model are determined and saved locally by the corresponding renewable energy forecasting entity. The second-layer meta-learner is used to combine the outputs of the first-layer basic forecasting model. Each renewable energy forecasting entity trains its basic forecasting model based on locally saved data, generates basic forecasting outputs within the training set time window, and obtains reference parameters for the meta-learner. S4. Execute distributed closed-loop prediction-scheduling iteration, wherein the distributed closed-loop prediction-scheduling iteration includes outer-layer reconstruction and decomposition iteration and inner-layer overlapping block coordinate descent-alternating direction multiplier method iteration; the outer-layer reconstruction and decomposition iteration includes establishing a closed-loop prediction-scheduling two-layer mixed integer programming model, solving the penalized relaxation master problem using the inner-layer overlapping block coordinate descent-alternating direction multiplier method, solving the first subproblem locally by the power system operator, solving the second subproblem locally by the power system operator, and performing iteration return or termination according to the inner-layer convergence condition and the outer-layer convergence condition; in the inner-layer overlapping block coordinate descent-alternating direction multiplier method iteration, the power system operator and multiple renewable energy prediction entities collaboratively solve the penalized relaxation master problem; S5. Each renewable energy forecasting entity constructs an integrated learning forecasting model based on the converged meta-learner parameters and the local basic forecasting model. S6. Each renewable energy forecasting entity uses the integrated learning forecasting model to predict the renewable energy power within the test set time window, and obtains the renewable energy power forecast value guided by scheduling cost. S7. The combined results of computer groups based on the dispatch cost-oriented renewable energy power forecast values ​​of the power system operator; S8. The power system operator converts the unit combination results into control instructions and issues the control instructions to the power system units.

2. The distributed closed-loop prediction-dispatch method for unit combination in a power system containing renewable energy, as described in claim 1, is characterized in that... The closed-loop prediction-scheduling two-level hybrid integer programming model established in step S4 is as follows: in, For the training sample set, The sample number in the training sample set. This is the set of meta-learner parameters for each renewable energy forecasting subject. For training samples Optimal values ​​of continuous variables for unit combination. For training samples Optimal values ​​for start-up and shutdown variables of the unit combination. For continuous decision variables in the unit combination model, For binary decision variables in the unit combination model, For training samples The rescheduled variable under, The cost coefficient vector is a continuous variable in the unit combination objective function. , For the fuel cost coefficient vector, This is a vector of expected load shedding and wind curtailment penalty cost coefficients. This is a vector of unit start-up and shutdown cost coefficients. This is a vector of rescheduling cost coefficients, which includes the upward and downward adjustment of ancillary service costs, the actual load shedding, and the cost of wind curtailment penalties. For training samples The corresponding pre-calculated baseline optimal operating cost, These are the regularization coefficients for the meta-learner parameters. For the meta-learner parameter regularization term, For training samples The corresponding renewable energy power forecast, Indicates sample The following is a prediction feature matrix composed of the outputs of each basic prediction model. The actual power generation capacity of renewable energy. Based on actual renewable energy power The rescheduling feasible domain, For predicted power The feasible domain for unit combinations.

3. The distributed closed-loop prediction-dispatch method for unit combination in a power system containing renewable energy, as described in claim 2, is characterized in that... Step S4, the outer-layer reconstruction and decomposition iteration, transforms the closed-loop prediction-scheduling two-layer hybrid integer programming model into a penalized relaxation master problem. First subproblem SP1 and second subproblem SP2; power system operator initializes outer iteration count. Lower Boundary Upper Realm and outer convergence threshold and in the wheel( Constructing the penalized relaxation master problem in the outer layer reconstruction and decomposition iteration ; in, For the first The optimal cut corresponds to the fixed unit start-up and shutdown variables. and These are all coefficient matrices representing the constraints of the unit combination problem. For the bilinear term in the relaxation optimal cut, These are the dual variables corresponding to the unit combination constraints. The right-hand side of the unit combination constraint is determined by the predicted power parameterization. For use in relaxing bilinear terms Auxiliary variables, This is the penalty coefficient.

4. The distributed closed-loop prediction-dispatch method for unit combination in a power system containing renewable energy, as described in claim 3, is characterized in that... The main problem of penalty relaxation Solved using the inner overlapping block coordinate descent-alternating direction multiplier method; let... , , ,in, This mainly corresponds to the prediction-side variables caused by the prediction power and the meta-learner parameters. Includes unit start-up and shutdown variables and dual variables. Includes continuous variables of unit combination and rescheduling variables; in the... During the sub-overlapping block coordinate descent iteration, the power system operating entity is fixed. And solve locally - Fixed problem: Then, the power system operator is fixed. The solution is obtained by the power system operator and various renewable energy forecasting entities through a distributed solution using the alternating direction multiplier method. - Fixed problem; the aforementioned - Introducing operator-predicted copies in fixed-problem scenarios Predicting subject and predicting copy and consensus predictor variables And assume: Then the -The fixed problem is: in, For the local constraint set of the power system operator, For the local forecasting constraint set of renewable energy forecasting subjects, Maintained by the power system operator. Maintained by renewable energy forecasting entities, For the objective function on the operating entity side, To predict the objective function on the subject side; and The expressions are as follows: in, Main body for renewable energy forecasting The meta-learner parameter vector satisfies This is used to combine the outputs of multiple base prediction models on the subject's local side into a prediction copy for the subject: 。 5. A distributed closed-loop prediction-dispatch method for unit combination in a power system containing renewable energy, as described in claim 4, is characterized in that... In the In the alternating direction multiplier iteration, the power system operator solves the operation sub-problem locally: Each renewable energy forecasting entity solves the forecasting side problem in parallel: in, , , As the dual variable on the operating entity side, To predict the dual variables on the subject side, For the first The penalty parameter in the alternating direction multiplier method iteration; after solving the operation side problem and the prediction side problem, the power system operator updates the consensus prediction variables: The power system operators and each renewable energy forecasting entity update their local dual variables respectively: 。 6. A distributed closed-loop prediction-dispatch method for unit combination in a power system containing renewable energy, as described in claim 5, is characterized in that... The information exchange between the power system operator and the renewable energy forecasting entity is as follows: In the... In the alternating direction multiplier iteration, the power system operator moves towards the renewable energy forecasting entity. send: Renewable energy forecasting entities Return to the power system operator: The information exchange does not include local characteristic data, actual power generation data, basic prediction model structure, basic prediction model parameters and meta-learner parameters of the renewable energy prediction subject, nor does it include system operation parameters, cost coefficients, network constraints, unit operation constraints, rescheduling constraints and local decision variables of the power system operator.

7. A distributed closed-loop prediction-dispatch method for unit combination in a power system containing renewable energy, as described in claim 6, is characterized in that... The convergence conditions in step S4 include the convergence conditions for alternating direction multiplier iteration, overlapping block coordinate descent iteration, and outer layer reconstruction and decomposition iteration; among them, the original residual of the alternating direction multiplier iteration... , and dual residuals for: When the following conditions are met: When the alternating direction multiplier iteration is converged, it is determined that the iteration is complete. and These are the original feasibility threshold and the dual feasibility threshold, respectively; when not satisfied, update... And continue to solve - Fixed problem; penalty parameters are updated according to the relative magnitudes of the original residual and the dual residual: in, and These are the coefficients for increasing and decreasing the penalty parameter, respectively. The coefficient for determining the relative magnitude of the original residual and the dual residual; When the punishment is relaxed, the main problem is... The objective function value changes in two consecutive overlapping block coordinate descent iterations as follows: At that time, determine if the overlapping block coordinate descent iteration has converged, and output the current solution to the penalized relaxation master problem. and ,in Set the convergence threshold for the overlapping block coordinate descent iteration; if the threshold is not met, update... And continue solving the penalized relaxation master problem; after the penalized relaxation master problem converges, the power system operator locally solves the first subproblem SP1: And solve the second subproblem SP2 locally: The power system operator updates the lower bound based on the objective function value of the penalized relaxation principal problem. Update the upper bound based on the objective function value of the second subproblem SP2. And generate the optimality cut for the next round of penalized relaxation principal problem based on the solution of SP2; when: At that time, it is determined that the outer layer reconstruction and decomposition iterations have converged, where, Set the outer layer reconstruction and decomposition iteration convergence threshold; if the threshold is not met, update the outer layer iteration count and continue executing step S4.

8. A distributed closed-loop prediction-dispatch method for unit combinations in power systems containing renewable energy, as described in claim 1, is characterized in that... In step S5, each renewable energy forecasting subject uses a linear meta-learner to combine its local base forecast output to obtain a forecasting copy of the subject: in, , For the main body of renewable energy forecasting, Main body for renewable energy forecasting The set of basic prediction models, For the first The local prediction output of each basic prediction model For the intercept term, For the first The meta-learner weights corresponding to each basic prediction model, wherein the meta-learner parameters are: The meta-learner parameter regularization term only applies to the linear meta-learner parameters. And it does not affect the basic prediction model. When using L2 regularization: When using L1 regularization: in, Main body for renewable energy forecasting The meta-learner reference parameters obtained in step S3.

9. A distributed closed-loop predictive-dispatch system comprising a combination of renewable energy power system units, used to implement the method described in any one of claims 1-8, characterized in that, include: The local data acquisition module is configured to collect system operating parameters on the side of the power system operator and to collect forecast characteristic data and actual renewable energy power generation data on multiple renewable energy forecasting entities. The basic forecast training module is configured to enable each renewable energy forecasting entity to train its basic forecasting model locally and generate historical baseline forecast outputs. A distributed closed-loop iteration module is configured to perform distributed closed-loop prediction-scheduling iteration. The distributed closed-loop iteration module includes an outer reconstruction and decomposition module and an inner distributed solution module. The outer reconstruction and decomposition module is configured to establish a closed-loop prediction-scheduling two-layer hybrid integer programming model, construct a penalized relaxation master problem, solve the first subproblem, and solve the second subproblem. The inner distributed solution module is configured to perform overlapping block coordinate descent-alternating direction multiplier method iteration to collaboratively solve the penalized relaxation master problem. The information exchange module is configured to transmit consensus forecast variables, penalty parameters, forecast copies of forecasts from forecast entities, and regularization terms between the power system operator and each renewable energy forecast entity. The integrated model building module is configured to enable each renewable energy forecasting entity to build an integrated learning forecasting model based on the converged meta-learner parameters and the local basic forecasting model, respectively. The test set prediction module is configured to generate scheduling cost-oriented renewable energy power prediction values ​​within the test set time window based on the integrated learning prediction model. The control instruction generation module is configured to enable the power system operator to issue control instructions based on the combined results of the computer group of the dispatch cost-oriented renewable energy power forecast value.

Images