A source-grid-load-storage collaborative scheduling method based on divergence regularized distribution robust optimization

Abstract

This invention discloses a source-grid-load-storage coordinated scheduling method based on divergence regularization and bibliometric optimization, belonging to the field of power system optimization. Addressing the problems of random output fluctuations, distribution shifts, insufficient robustness of traditional methods, and inefficiency in solving problems under high-proportion wind and solar grid integration, this invention first constructs a continuously differentiable comprehensive loss function containing multiple uncertainties and soft and hard constraints. Then, it uses a generalized Sinkhorn distance with χ²-divergence regularization and Gaussian reference measure to construct a fuzzy set, equivalently transforming the original Min-Max problem into a nested two-layer continuous dual model. Finally, it employs a nested stochastic gradient descent algorithm, with the inner layer estimating the dual multipliers and the outer layer updating the scheduling strategy and projecting it onto the physical feasible region, achieving a rapid solution. This invention overcomes the limitations of historical data support sets, can prevent risks from unknown extreme weather events, avoids the dimensionality curse, is compatible with first-order AI algorithms, and balances grid operation safety and economy, making it suitable for robust scheduling of large-scale, high-proportion renewable energy power systems.

Description

Technical Field

[0001] This invention belongs to the field of power system optimization, specifically relating to a source-grid-load-storage coordinated scheduling method based on divergence regularized split-Brow bar optimization. Background Technology

[0002] The penetration rate of renewable energy sources such as wind power and photovoltaics in new power systems is constantly rising, gradually forming a system with a "high proportion of wind, solar and energy storage" grid connection. However, the output of new energy sources has strong randomness, intermittency and volatility, which brings unprecedented challenges to power system dispatch: (1) Wind and solar power output is significantly affected by meteorological conditions. Traditional deterministic dispatch methods based on point prediction often fail in actual operation, resulting in increased cost of timely adjustment and increased wind and solar curtailment rate. (2) The statistical characteristics of wind and solar power output prediction error are not fixed. Seasonal changes, extreme weather, equipment aging and other factors cause the error distribution to shift. Traditional stochastic optimization methods assume that the error distribution is known and fixed, and are not robust enough in the case of distribution shift.

[0003] To address the risks posed by prediction errors, the sub-Brow bar optimization method has been introduced into the field of power system dispatching. This technique constructs a fuzzy set containing probability distributions of various adverse scenarios and seeks the optimal dispatching strategy under the worst probability distribution. However, existing sub-Brow bar optimization methods still have the following technical bottlenecks when facing large-scale power system dispatching scenarios: (1) There is a "blind spot" in the prevention of extreme weather events that have not been seen in history: traditional methods based on... - The divergence-based Brull bar optimization method strictly requires, mathematically, that the predicted distribution and the actual extreme distribution must have the same probability support set. This means that if a certain type of extreme weather has never occurred in historical data, this method cannot include it in the worst-case scenario assessment, which poses a great safety hazard. (2) Large-scale power grid solutions face the "curse of dimensionality" and computational bottleneck: Although the Brull bar optimization method based on Wasserstein distance can break the support set restriction, it must be transformed into a huge semidefinite programming problem when solving, or rely on the calculation of the second-order Hessian matrix. When facing a real power grid with many nodes, it faces a serious "curse of dimensionality". (3) Mixed integer hard constraints hinder the application of first-order efficient AI algorithms: In the traditional power grid dispatch optimal flow model that includes energy storage, discrete hard constraints such as mutual exclusion of energy storage charging and discharging are inevitably included. These discontinuous problems make it impossible to directly apply the modern first-order gradient descent algorithm, which has great computational advantages, further aggravating the difficulty of solving the problem.

[0004] In summary, a novel technical solution is urgently needed in the field of high-proportion wind, solar, and energy storage joint dispatch. This solution must not only overcome the limitations of historical data and accommodate unprecedented extreme wind and solar power output scenarios, but also efficiently handle the non-convex physical constraints of energy storage and other equipment. More importantly, it must completely break free from the constraints of second-order matrix calculations and achieve rapid and stable solutions for large-scale, high-dimensional power grid dispatch models, thereby achieving the optimal balance between power grid security and economy under extreme uncertainty. Summary of the Invention

[0005] To address the above shortcomings, this invention discloses a source-network-load-storage collaborative scheduling method based on divergence regularization and robust optimization, which solves the problems of traditional scheduling methods being unable to prevent unknown extreme risks due to support set limitations, and the low computational efficiency of large-scale robust optimization.

[0006] To achieve the above objectives, the technical solution adopted by this invention is as follows: A source-grid-load-storage collaborative scheduling method based on divergence regularization split-bar optimization, characterized by comprising the following steps:

[0007] S1. Construct a power system operation model considering multiple uncertainties and transform it into a comprehensive loss function suitable for continuous gradient optimization: Collect power system network topology and historical data, define the set of uncertain random scenario variables and the global scheduling decision variable matrix of the power system; construct a comprehensive loss function that is highly coupled with the physical decision variables and the random scenario of new energy, including the operating cost model of conventional generator units and energy storage systems, the global power balance penalty function, the line active power flow limit violation penalty function, the unit ramping limit violation penalty function, the wind and solar curtailment penalty cost function, the energy storage SOC upper and lower limit violation and initial and final state limitation penalty function, and the energy storage state mutual exclusion penalty function; define the insurmountable physical limitations of equipment output and available power of new energy as the hard constraint boundary of the decision variables, and finally transform the source-grid-load-storage coordinated scheduling problem into a continuous nonlinear optimization model with hard constraint boundaries;

[0008] S2. Construct and equivalently transform a distributed Sinkhorn bar collaborative scheduling model based on generalized Sinkhorn distance: To overcome the prediction blind spot caused by the actual meteorological distribution offset, the comprehensive loss function constructed in step S1 is used as the optimization objective, and a model with... - Construct an uncertain fuzzy set using divergence regularization and generalized Sinkhorn distance of Gaussian reference measure, and establish the original Min-Max sub-Bruker scheduling problem; through the inverse cumulative distribution function and Lagrange duality theory, the original problem is equivalently transformed into a nested bilayer continuous dual model with inner dual multipliers and outer scheduling strategy as decision variables, so as to achieve efficient solution in large-scale scenarios.

[0009] S3. Efficient solution of power grid dispatch strategy based on nested stochastic gradient descent: The inner estimation module and the outer update module are executed alternately. The inner estimation module fixes the current dispatch strategy, applies noise to generate a disturbance scenario around the historical wind nominal sample, and approximates the optimal dual multiplier under the current sample through first-order gradient descent. The outer update module substitutes the optimal dual multiplier into the weight evaluation formula, calculates the outer weighted gradient of the comprehensive loss function with respect to the power system decision variables, and ensures that the output of generators and energy storage does not exceed the limit through physical boundary projection, thus completing the update of decision variables. When the iterative convergence condition is met, the optimization stops and the optimal power grid dispatch strategy is output to achieve risk-resistant and robust dispatch of a high proportion of new energy power system.

[0010] Furthermore, in step S1, the set of uncertain random scenario variables and the matrix of decision variables are defined as follows:

[0011] First, define the sources of uncertainty as shown in formulas (1)-(3):

[0012]

[0013]

[0014]

[0015] In the formula: The scheduling period; and They are respectively The vector between the theoretical available power output limit of a wind farm, which is constantly affected by weather, and the theoretical available power output limit of a photovoltaic power station; For the first The theoretical upper limit of usable output of a wind farm affected by weather conditions; For the first The theoretical maximum usable output of a photovoltaic power station affected by weather conditions; subscript Total number of wind farms; subscript This represents the total number of photovoltaic power stations; for The load demand vector of network nodes at any given time. for The Middle Active load demand vector of each network node; subscript This represents the total number of nodes in the power system.

[0016] Then, define the set of uncertain random scenario variables representing extreme weather and user behavior at that moment. As shown in formula (4):

[0017]

[0018] Next, a discrete empirical nominal probability distribution is constructed based on historical data, as shown in formula (5):

[0019]

[0020] In the formula: For the first One historical observation sample; This represents the total number of days for which historical data was collected. To focus on the sample The Dirac measure function at the location; It is a discrete empirical nominal probability distribution constructed based on historical data;

[0021] For any time The scheduling instructions for each controlled device are shown in formulas (6)-(10):

[0022]

[0023]

[0024]

[0025]

[0026]

[0027] In the formula: For conventional generator sets at time The active power output vector; subscript This represents the total number of conventional thermal power units. , For energy storage systems at time The charging power and discharging power vector; subscript This represents the total number of energy storage systems. For wind farms at all times The actual grid-connected power output vector; For photovoltaic power stations at all times The actual grid-connected power output vector; For the first The conventional generator set at the time The instructions for generating active power; , The first An energy storage system at any time Charging power and discharging power commands; For the first Each wind farm at any time The actual grid-connected power consumption; For the first A photovoltaic power station at time The actual grid-connected power consumption;

[0028] Finally, the global dispatch decision variable matrix of the power system was obtained. The definition is shown in formula (11):

[0029] .

[0030] Furthermore, in step S1, a comprehensive loss function highly coupled with the physical decision variables and the new energy stochastic scenario is constructed, specifically including:

[0031] The operating cost model for conventional generator sets is constructed as shown in formula (12):

[0032]

[0033] In the formula: This represents the total operating cost of the generator sets in the system. , , For the first Cost coefficient of a generator set;

[0034] A cost model for the energy storage system is constructed, and the depreciation cost of energy storage operation during the system scheduling cycle is calculated, as shown in formula (13):

[0035]

[0036] In the formula: This represents the total operating cost of the energy storage system within the system. This is the lifetime loss factor per unit throughput power of the energy storage system;

[0037] The cost function for the curtailment penalty is defined as shown in formula (14):

[0038]

[0039] In the formula, The cost of penalizing the abandonment of wind and solar power; Weighting for wind curtailment power penalty; The penalty weight for wasted light power;

[0040] The net injected power at each node is as shown in formula (15):

[0041]

[0042] In the formula: For each network node exist Net injection power at any given time; , , and The elements of the 0-1 correlation matrix indicate the number of elements. generator, the first The first wind farm, The first photovoltaic power station and the first Is the energy storage system connected to the node? If connected, the value is 1; otherwise, it is 0.

[0043] A global power balance penalty function based on the L2 norm is introduced as shown in Equation (16):

[0044]

[0045] In the formula: For each network node exist Net injection power at any given time; This is the global power balance penalty function; For power balance penalty weights;

[0046] The line active power flow over-limit penalty function based on the power transfer distribution factor matrix is ​​shown in formulas (17)-(18):

[0047]

[0048]

[0049] In the formula: For the first The transmission line is in Trend value for a given time period; For PTDF matrix elements, representing nodes Injecting unit power to the line Sensitivity to current shifts; This represents the total number of lines; For the active power flow exceeding the limit penalty function of the line; For the line The maximum thermal stability transfer limit; The penalty weight for exceeding the limits of the trend;

[0050] The unit ramp-up penalty function is constructed as shown in formula (19):

[0051]

[0052] In the formula: This is the function for penalizing over-limit ramping of the generator unit. For the unit Maximum permissible climbing rate limit; Weighting for penalties for exceeding the climbing limit;

[0053] The penalty functions for exceeding the upper and lower limits of the energy storage SOC and the initial and final state constraints are constructed as shown in formulas (20)-(22):

[0054]

[0055]

[0056]

[0057] In the formula: For energy storage At any moment The energy storage state of charge; , These represent the charging and discharging efficiencies of energy storage, respectively. For energy storage Rated capacity; The scheduling time step; , These are the upper and lower safety limits for preventing battery overcharging and over-discharging; For the penalty function for exceeding the limit of the energy storage state of charge; , Energy storage at the beginning and end of the scheduling process The state of charge; , For penalty weighting; The initial and final state constraints of the energy storage are defined by a penalty function.

[0058] The mutual exclusion penalty function for energy storage charging and discharging is shown in formula (23):

[0059]

[0060] In the formula: The mutual exclusion penalty function for energy storage charging and discharging; Mutually exclusive penalty weights;

[0061] By adding the power generation and operation costs to various physical penalty functions, a comprehensive loss function is constructed. As shown in formula (24):

[0062]

[0063] In the formula, To cover scheduling cycles All time periods The global dispatch decision variables of the power system; To cover scheduling cycles All time periods Real-world, unknown, random scenario variables;

[0064] Meanwhile, power system global dispatch decision variables The equipment must satisfy both the hard constraints of the absolute physical boundary and the hard constraints of the available output power in the local weather, as shown in formulas (25)-(29):

[0065]

[0066]

[0067]

[0068]

[0069]

[0070] In the formula: , The first Minimum and maximum technical output of a conventional generator set; , The first The maximum charging and discharging power of an energy storage system; For wind power Theoretical maximum output; For photovoltaic Theoretical maximum output;

[0071] Furthermore, these hard constraints constitute the global scheduling decision variables of the power system. feasible domain Then, the global dispatch decision variables of the power system The model is transformed into a continuous nonlinear optimization model with hard-constrained boundaries, as shown in equation (30):

[0072]

[0073] In the formula: and These are the global dispatch decision variables of the power system. The absolute physical lower limit and absolute physical upper limit of various power levels.

[0074] Furthermore, in step S2, a penalized Min-Max sub-bar optimization framework is introduced, and the original problem model for minimizing the expected comprehensive loss of the system under the worst distribution is established as shown in formula (31):

[0075]

[0076] In the formula: For the global dispatch decision variables of the power system, and satisfy the physical boundary feasible region. ; For real, unknown, random scenario variables; To obtain from the empirical nominal probability distribution Historical nominal scene samples obtained from sampling; In the potential worst probability distribution Below, the comprehensive loss function The mathematical expectation; To quantify the empirical nominal probability distribution With the potential worst probability distribution The generalized Sinkhorn distance between the differences; This is the divergence regularization coefficient, used to control the computational smoothness during the joint probability measure shift process. ; This is the robust penalty regularization coefficient, used to adjust the conservatism of the scheduling strategy in the face of extreme environmental fluctuations. ;

[0077] Build with - The expression for constructing an uncertain fuzzy set using divergence regularization and the generalized Sinkhorn distance of the Gaussian reference measure is shown in formula (32):

[0078]

[0079] In the formula: Empirical nominal probability distribution and the potential worst probability distribution The joint probability distribution mapping; The complete set of the joint distribution that satisfies this marginal distribution condition; To measure variables in real-world, unknown, random scenarios Deviation from historical nominal scene samples The cost metric function for degree is based on the L2 norm Euclidean distance. ; A Gaussian reference measure independent of empirical distribution. , It is the identity matrix. The standard deviation of the Gaussian reference metric noise is set to control the boundary range of the exploration of extremely harsh scenarios; To measure the joint distribution mapping Compared with the benchmark measurement Differences - Divergence; This invention employs here... -Difference - Divergence is used to ensure the second-order smoothness and computational stability of subsequent algorithms; its generating function is defined as follows: ,in Let be the non-negative real number independent variable in the domain of the mathematical mapping.

[0080] Furthermore, in step S2, the expression for reconstructing the original maximization problem containing divergence constraints into a one-dimensional form is shown in equation (33):

[0081]

[0082] In the formula: Given a specific historical nominal scene sample and current power system global dispatch decision variables Under the condition of worst probability distribution, the conditional expected loss function faced by the power system; It is a density ratio function; Scalar random variable The inverse cumulative distribution function; To satisfy the function space where the probability density integral is 1;

[0083] By introducing the Lagrange dual multiplier and using the strong duality theorem, the maximization problem is equivalently transformed into an unconstrained minimization problem, and finally a nested bi-level dual scheduling model is obtained, as shown in formulas (34)-(35):

[0084]

[0085]

[0086] In the formula: As an inner-layer decision variable, it physically represents a sample of a historical nominal scenario. The threshold for economic penalties triggered by the power system's response to current extreme weather scenarios; Indicates a Gaussian reference measure independent of empirical distribution. The expected value is to be obtained. for - Divergence generation function The conjugate dual function of is mathematically defined as . ,in The independent variable introduced to describe the conjugate dual mapping relationship; the one targeted at... - Divergence, whose exact analytical expression for its conjugate dual function is shown in equation (36):

[0087] .

[0088] Furthermore, in step S3, the inner estimation module fixes the current global scheduling decision variables of the power system. Noise is applied around historical scenic samples to generate a perturbation scenario. The optimal dual multiplier for the current sample is obtained by approximating it using first-order gradient descent.

[0089] Around the central sample, based on Gaussian white noise The generated perturbation scenario is shown in formula (37):

[0090]

[0091] In the formula, This refers to the randomly selected nominal historical sample number; For the first batch A historical center observation sample; This refers to the sequence number of the random perturbation scene generated based on the Gaussian reference measure; This represents the total number of perturbation scenarios generated for a single historical sample. To surround the central sample The generated first A random disturbance scenario variable; This is the corresponding Gaussian random noise vector;

[0092] To quickly approximate the penalty weight under the worst-case scenario, the current strategy is fixed. Define intermediate state variables Used to quantify the current strategy in a perturbation scenario. Physical loss deviation; initialization of inner layer variables Calculate the current The intermediate mapping variables are shown in formula (38):

[0093]

[0094] In the formula, Estimate the number of iterations for the dual multipliers in the inner algorithm; This is the maximum number of iterations set for the inner layer; For the inner layer In the next iteration, intermediate mapping state variables are used to quantify the deviation of the system's physical loss. For the inner layer The dual multiplier value of the step;

[0095] use -Analytical expression of the derivative of the divergence conjugate function Calculate the inner objective function with respect to The stochastic gradient approximation value is shown in equation (39):

[0096]

[0097] In the formula, For the inner dual objective function with respect to The stochastic gradient approximation estimate;

[0098] The multipliers are updated using gradient descent as shown in equation (40):

[0099]

[0100] In the formula, A constant learning rate for optimizing the inner dual variables;

[0101] The outer update module substitutes the optimal dual multiplier into the weight evaluation formula to calculate the outer weighted gradient of the comprehensive loss function with respect to the power system decision variables, which is obtained using the inner layer calculation. Calculate the comprehensive penalty loss with respect to the physical scheduling decision variables. Outer robust weighted gradient For each sample in the batch and the perturbation scene it generates Substitute back the optimal multiplier The dynamic weights of the worst-case distribution shift in the current scene are calculated as shown in formula (41):

[0102]

[0103] In the formula, This represents the number of outer iterations. The intermediate mapping state variables used to quantify the deviation of the system's physical loss when the iteration reaches the optimal point;

[0104] The dynamic weight expression, which is spontaneously assigned based on the severity of the penalty for each disturbance scenario, is shown in formula (42):

[0105]

[0106] The stochastic gradient estimator that approximates the expected risk of the entire network scheduling is constructed as shown in Equation (43):

[0107]

[0108] In the formula, For the first In the outer layer iteration, the global weighted stochastic gradient estimator of the outer layer policy is comprehensively considered after taking into account both batch samples and perturbation scenarios; This is the maximum number of iterations for the outermost layer. The sample size for a single iteration;

[0109] By using physical boundary projection to ensure that generator and energy storage commands do not exceed limits, the decision variables are updated as follows:

[0110] The outer layer stochastic gradient is subjected to extreme value truncation as shown in Equation (44):

[0111]

[0112] In the formula, The maximum allowable gradient clipping threshold set to prevent gradient explosion;

[0113] Introducing momentum factor and decaying learning rate Accumulate historical gradient directions and calculate intermediate scheduling instructions as shown in formulas (45)-(46):

[0114]

[0115]

[0116] In the formula, This is the momentum factor in the momentum gradient descent method, used to accumulate historical gradient directions to accelerate convergence. For the first The momentum-velocity vector of the outermost layer of the wheel during iteration; For the outermost layer The decaying learning rate of each iteration; This is an intermediate temporary scheduling instruction that has been updated by momentum gradient but has not yet been projected onto the physical boundary.

[0117] Furthermore, the physical boundary projection operator is invoked to forcibly truncate the temporary decision instructions and map them back to the physical security domain of each device, as shown in formula (47):

[0118]

[0119] And the updated As the starting point for the next iteration.

[0120] Furthermore, in step S3, when the iterative convergence condition is met, the optimal source-network-load-storage scheduling strategy is output: when the iteration number... Reaching the set maximum number of outer iterations If the proxy loss value is less than the preset threshold for multiple consecutive rounds, optimization stops; the final robust optimal scheduling strategy matrix is ​​then obtained. The scheduling plan is broken down into scheduling plans within the scheduling cycle according to the time series; thus, robust control against risks is truly achieved for the new power system with a high proportion of new energy sources. This not only effectively avoids the risk of grid frequency collapse caused by extreme weather, but also significantly reduces the operation and maintenance economic costs brought about by conventional deterministic scheduling.

[0121] This invention also provides a source-grid-load-storage collaborative scheduling system based on divergence regularized split-bar optimization, used to implement the source-grid-load-storage collaborative scheduling method based on divergence regularized split-bar optimization as described above, including:

[0122] The modeling module is used to collect power system data, define uncertain random variables, and construct a continuous and differentiable comprehensive loss function and a feasible region with physical hard constraints on decision variables.

[0123] The robust optimization module is used to construct fuzzy sets based on the generalized Sinkhorn distance, establish the Min-Max scheduling primal problem and transform its duality into a nested bi-level continuous dual problem;

[0124] The solver module is used to perform inner dual multiplier estimation and outer scheduling policy update using a nested stochastic gradient descent algorithm.

[0125] The scheduling execution module is used to output and distribute the optimal power grid scheduling strategy when the convergence conditions are met.

[0126] The present invention provides an electronic device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements a source-network-load-storage collaborative scheduling method based on divergence regularization split-bar optimization as described above.

[0127] The present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the program is executed by a processor, it implements the source-grid-load-storage collaborative scheduling method based on divergence regularization split-bar optimization as described above.

[0128] The present invention has the following advantages and beneficial effects:

[0129] (1) Overcoming the limitations of probabilistic support sets and improving robustness against unseen meteorological disasters: This invention innovatively introduces a generalized Sinkhorn distance to construct a fuzzy set. By superimposing Gaussian reference measures around historical nominal samples, the rigid limitations of support sets are broken. This gives the model the ability to generate severe boundary scenarios in historical data gaps. Even if the power grid encounters rare extreme weather events, the generated robust scheduling instructions can still spontaneously reserve sufficient backup capacity for the system, safeguarding the bottom line of power grid safety.

[0130] (2) Breaking the "curse of dimensionality" of large-scale scheduling and improving computational efficiency: This invention establishes a sub-Bruker optimization model based on generalized Sinkhorn distance, reduces dimensionality through Lagrange duality, and adopts the Nested Stochastic Gradient Descent (SGD) algorithm based on first-order gradient. It does not require storing a large high-dimensional Jacobian or Hessian matrix in memory, which greatly improves the scheduling timeliness of large distribution networks with massive nodes.

[0131] (3) Achieving continuous modeling of physical soft and hard constraints, aligning with efficient artificial intelligence algorithms: This invention transforms complex physical coupling constraints into a continuously differentiable comprehensive loss function through a weighted quadratic penalty function; simultaneously, it performs hard constraint truncation in conjunction with the physical boundary projection operator. This achieves efficient convergence while eliminating dangerous instructions that violate physical constraints.

[0132] (4) Achieving a dynamic balance between the economic efficiency and safety of the new power system: The comprehensive loss function constructed in this invention fully covers the cost of power generation fuel, the depreciation cost of energy storage life, and various physical limit violations. Under normal weather conditions, the system model can balance the cost of wind and solar power integration with the cost of equipment operation and maintenance, avoiding resource waste caused by excessive conservatism; under extreme weather warnings, the system can spontaneously increase the system's risk resistance weight by utilizing the generalized Sinkhorn fuzzy set, effectively avoiding the remedial costs caused by frequent limit violations in traditional deterministic dispatching, and realizing the global optimal operation of the new power system with a high proportion of wind, solar and energy storage participation under uncertain environments.

[0133] Therefore, this invention breaks through the limitations of traditional distributed bar optimization methods on the data support set, effectively avoids the risk of equipment exceeding limits under extreme disasters, achieves effective defense against unknown extreme weather events, and ensures the safety and stability of the power grid; it significantly reduces the computational complexity and memory consumption of large-scale power grid dispatch, thereby improving computational efficiency; and by comprehensively balancing economy and robustness, it achieves efficient utilization of wind and solar resources and reduces power grid operating costs, thus promoting the transformation of the energy structure. Attached Figure Description

[0134] Figure 1 This is a flowchart of a source-grid-load-storage collaborative scheduling method based on divergence regularization split-bar optimization according to the present invention.

[0135] Figure 2 A flowchart of an efficient algorithm for solving power grid scheduling strategies based on nested stochastic gradient descent is provided in this embodiment of the invention.

[0136] Figure 3 This is a diagram showing the convergence curve of the comprehensive loss of the generalized Sinkhorn split-bar scheduling model based on the nested stochastic gradient descent algorithm in an embodiment of the present invention.

[0137] Figure 4 The scheduling results are shown in the figure using the Brussels bar optimization method based on Wasserstein distance.

[0138] Figure 5 The diagram shows the optimized scheduling results achieved using the method proposed in this patent.

[0139] Figure 6 A comparison chart showing the power deficit in power grids after encountering extreme weather events, using a Wasserstein distance-based bibliometric optimization method versus the method proposed in this patent to address power grid deficits in scenarios unseen in the past. Detailed Implementation

[0141] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0142] Example 1

[0143] like Figure 1 As shown, this embodiment of the invention provides a complete execution flow of a source-network-load-storage collaborative scheduling method based on divergence regularization split-bar optimization, specifically including the following steps:

[0144] S1. Construct a power system operation model considering multiple uncertainties and transform it into a comprehensive loss function suitable for continuous gradient optimization: Collect power system network topology and historical data, define the set of uncertain random scenario variables and the matrix of global dispatch decision variables for the power system; construct a comprehensive loss function that is highly coupled with the physical decision variables and the random scenarios of new energy sources, including the operating cost model of conventional generating units and energy storage systems, the global power balance penalty function, the line active power flow exceeding limit penalty function, the unit ramping exceeding limit penalty function, the wind and solar curtailment penalty cost function, the energy storage SOC upper and lower limit exceeding limit and the initial and final state limit penalty function, and the energy storage state mutual exclusion penalty function; define the insurmountable physical limitations of equipment output and available power of new energy sources as the hard constraint boundary of the decision variables, and finally transform the source-grid-load-storage coordinated dispatch problem into a continuous nonlinear optimization model with hard constraint boundaries; the specific method is as follows:

[0145] Collect network topology and historical data, and define uncertain random variables:

[0146] Define the sources of uncertainty as shown in formulas (1)-(3):

[0147]

[0148]

[0149]

[0150] In the formula: Scheduling period (unit: hours); and They are respectively The vector between the theoretical available power output limit of a wind farm, which is constantly affected by weather, and the theoretical available power output limit of a photovoltaic power station; For the first The theoretical upper limit of usable output of a wind farm affected by weather conditions; For the first The theoretical maximum usable output of a photovoltaic power station affected by weather conditions; subscript Total number of wind farms; subscript This represents the total number of photovoltaic power stations; for The load demand vector of network nodes at any given time. for The Middle Active load demand vector of each network node; subscript This represents the total number of nodes in the power system.

[0151] Then, define the set of uncertain random scenario variables representing extreme weather and user behavior at that moment. As shown in formula (4):

[0152]

[0153] Next, a discrete empirical nominal probability distribution is constructed based on historical data, as shown in formula (5):

[0154]

[0155] In the formula: For the first One historical observation sample; This represents the total number of days for which historical data was collected. To focus on the sample The Dirac measure function at the location; It is a discrete empirical nominal probability distribution constructed based on historical data.

[0156] For any time The scheduling instructions for each controlled device are shown in formulas (6)-(10):

[0157]

[0158]

[0159]

[0160]

[0161]

[0162] In the formula: For conventional generator sets at time The active power output vector; subscript This represents the total number of conventional thermal power units. , For energy storage systems at time The charging power and discharging power vector; subscript This represents the total number of energy storage systems. For wind farms at all times The actual grid-connected power output vector; For photovoltaic power stations at all times The actual grid-connected power output vector; For the first The conventional generator set at the time The instructions for generating active power; , The first An energy storage system at any time Charging power and discharging power commands; For the first Each wind farm at any time The actual grid-connected power consumption; For the first A photovoltaic power station at time The actual grid-connected power consumption;

[0163] Finally, the global dispatch decision variable matrix of the power system was obtained. The definition is shown in formula (11):

[0164]

[0165] S2. Construct and equivalently transform a distributed Sinkhorn bar collaborative scheduling model based on generalized Sinkhorn distance: To overcome the prediction blind spot caused by the actual meteorological distribution offset, the comprehensive loss function constructed in step S1 is used as the optimization objective, and a model with... - Divergence regularization and the generalized Sinkhorn distance of Gaussian reference measure are used to construct an uncertain fuzzy set, establishing the Min-Max sub-Bruker scheduling primal problem. Through the inverse cumulative distribution function and Lagrange duality theory, the primal problem is equivalently transformed into a nested bi-layer continuous dual model with inner dual multipliers and outer scheduling policy as decision variables, achieving efficient solutions in large-scale scenarios. The specific method is as follows:

[0166] The conventional operating cost model of the generator set is constructed as shown in formula (12), and the conventional power generation fuel cost within the system scheduling cycle is calculated:

[0167]

[0168] In the formula: This represents the total operating cost of the generator sets in the system. , , For the first Cost coefficient of a generator set.

[0169] A cost model for the energy storage system is constructed, and the depreciation cost of energy storage operation during the system scheduling cycle is calculated, as shown in formula (13):

[0170]

[0171] In the formula: This represents the total operating cost of the energy storage system within the system. This is the lifetime loss factor per unit throughput power of the energy storage system.

[0172] To maximize the consumption of new energy sources and quantify the loss of unused wind and solar resources, the penalty cost function for wind and solar curtailment is defined as shown in formula (14):

[0173]

[0174] In the formula, The cost of penalizing the abandonment of wind and solar power; Weighting for wind curtailment power penalty; The penalty weight for abandoned light power.

[0175] The net injected power at each node is as shown in formula (15):

[0176]

[0177] In the formula: For network nodes exist Net injection power at any given time; , , and The elements of the 0-1 correlation matrix indicate the number of elements. generator, the first The first wind farm, The first photovoltaic power station and the first Is the energy storage system connected to the node? If the condition is met, the value is 1; otherwise, it is 0.

[0178] A global power balance penalty function based on the L2 norm is introduced as shown in Equation (16):

[0179]

[0180] In the formula: For network nodes exist Net injection power at any given time; This is the global power balance penalty function; The power balance penalty weight.

[0181] Using the power transfer distribution factor matrix (PTDF) in the optimal DC power flow, the line active power flow limit violation inequality constraint is transformed into a one-sided continuously differentiable quadratic penalty function, and the line active power flow limit violation penalty function is obtained as shown in formulas (17)-(18):

[0182]

[0183]

[0184] In the formula: For the first The transmission line is in Trend value for a given time period; For PTDF matrix elements, representing nodes Injecting unit power to the line Sensitivity to current shifts; This represents the total number of lines; For the active power flow exceeding the limit penalty function of the line; For the line The maximum thermal stability transfer limit; The weight of penalties is imposed for exceeding the limits of trendsetting.

[0185] The unit ramp-up penalty function is constructed as shown in formula (19):

[0186]

[0187] In the formula: For the unit's ramp-up penalty function; For the unit Maximum permissible climbing rate limit; The penalty weight for exceeding the limit while climbing.

[0188] The penalty functions for exceeding the upper and lower limits of the energy storage SOC and the initial and final state constraints are constructed as shown in formulas (20)-(22):

[0189]

[0190]

[0191]

[0192] In the formula: For energy storage At any moment The energy storage state of charge; , These represent the charging and discharging efficiencies of energy storage, respectively. For energy storage Rated capacity; The scheduling time step; , These are the upper and lower safety limits for preventing battery overcharging and over-discharging; For the penalty function for exceeding the limit of the energy storage state of charge; , Energy storage at the beginning and end of the scheduling process The state of charge; , For penalty weighting; The initial and final state constraints of the energy storage are defined by a penalty function.

[0193] The mutual exclusion penalty function for energy storage charging and discharging is shown in formula (23):

[0194]

[0195] In the formula: The mutual exclusion penalty function for energy storage charging and discharging; These are mutually exclusive penalty weights.

[0196] The above power generation and operation costs are added to the various physical penalty functions to construct a comprehensive loss function. As shown in formula (24):

[0197]

[0198] In the formula: To cover scheduling cycles All time periods The global dispatch decision variables of the power system; To cover scheduling cycles All time periods Real-world, unknown, random scenario variables;

[0199] Meanwhile, power system dispatch decision variables The equipment must satisfy both the hard constraints of the absolute physical boundary and the hard constraints of the available output power in the local weather, as shown in formulas (25)-(29):

[0200]

[0201]

[0202]

[0203]

[0204]

[0205] In the formula: , The first Minimum and maximum technical output of a conventional generator set; , The first The maximum charging and discharging power of an energy storage system; For wind power Theoretical maximum output; For photovoltaic Theoretical maximum output;

[0206] The aforementioned hard constraints constitute the decision variables. feasible domain Unlike complex soft constraints that are coupled across variables, such as node power balance and network flow, absolute hard constraints define the objective physical limits of device output, which cannot be violated at any optimization moment to prevent unreasonable scheduling instructions.

[0207] Furthermore, these hard constraints constitute the global scheduling decision variables of the power system. feasible domain Then, the global dispatch decision variables of the power system The model is transformed into a continuous nonlinear optimization model with hard-constrained boundaries, as shown in equation (30):

[0208]

[0209] In the formula: and These are the global dispatch decision variables of the power system. The absolute physical lower limit and absolute physical upper limit of various power levels.

[0210] Construct a model for the original Brussels bar scheduling problem based on generalized Sinkhorn distance:

[0211] Considering the potential worst-case probability distribution of the actual occurrence It often deviates from the empirical nominal probability distribution based on historical data statistics. (Such as extreme weather disasters never seen in history). A penalized Min-Max sub-bar optimization framework is introduced to establish the original problem model that minimizes the expected comprehensive loss of the system under the worst distribution, as shown in formula (31):

[0212]

[0213] In the formula: For the global dispatch decision variables of the power system, and satisfy the physical boundary feasible region. ; For real, unknown, random scenario variables; To obtain from the empirical nominal probability distribution Historical nominal scene samples obtained from sampling; In the potential worst probability distribution Below, the comprehensive loss function The mathematical expectation; To quantify the empirical nominal probability distribution With the potential worst probability distribution The generalized Sinkhorn distance between the differences; This is the divergence regularization coefficient, used to control the computational smoothness during the joint probability measure shift process. ; This is the robust penalty regularization coefficient, used to adjust the conservatism of the scheduling strategy in the face of extreme environmental fluctuations. ;

[0214] Unlike traditional divergence measures that require support sets to have the same probability, constructing a divergence measure with... - The expression for constructing an uncertain fuzzy set using divergence regularization and the generalized Sinkhorn distance of the Gaussian reference measure is shown in formula (32):

[0215]

[0216] In the formula: Empirical nominal probability distribution and the potential worst probability distribution The joint probability distribution mapping; The complete set of the joint distribution that satisfies this marginal distribution condition; To measure variables in real-world, unknown, random scenarios Deviation from historical nominal scene samples The cost metric function for degree is based on the L2 norm Euclidean distance. ; A Gaussian reference measure independent of empirical distribution. , It is the identity matrix. The standard deviation of the Gaussian reference metric noise is set to control the boundary range of the exploration of extremely harsh scenarios; To measure the joint distribution mapping Compared with the benchmark measurement Differences - Divergence. To ensure the second-order smoothness and computational stability of subsequent algorithms, this embodiment adopts the above-mentioned... -Difference - Divergence, whose generating function is defined as ,in, Let be the non-negative real number independent variable in the domain of the mathematical mapping.

[0217] Derivation based on the dual theory of the inverse cumulative distribution function:

[0218] In the original question This involves optimizing infinite-dimensional probability measures, falling into the NP-hard category. This invention performs dimensionality reduction mapping of the data space through a system, utilizing the inverse cumulative distribution function sampling technique from probability theory to transfer the optimization of the measure space to the integral variable space: decomposing the joint distribution into conditional distributions. Based on the principle of commutativity, the original problem is transformed into a problem for each historical scenario. Find the worst-case distribution. The expression for reconstructing the original maximization problem with divergence constraints into a one-dimensional form is shown in equation (33):

[0219]

[0220] In the formula: Given a specific historical nominal scene sample and current power system global dispatch decision variables Under the condition of worst probability distribution, the conditional expected loss function faced by the power system; It is a density ratio function; Scalar random variable The inverse cumulative distribution function; To satisfy the function space where the probability density integral is 1.

[0221] Construct a nested bi-level minimization (Min-Min) dual problem model that can be computed by a first-order algorithm:

[0222] By introducing the Lagrange dual multiplier and using the strong duality theorem, the above maximization problem is equivalently transformed into an unconstrained minimization problem, and finally a nested bi-level dual scheduling model is obtained, as shown in formulas (34)-(35):

[0223]

[0224]

[0225] In the formula: As inner-layer decision variables, they physically represent specific historical samples. The economic penalty threshold (shadow price) triggered by the power system's response to current extreme weather scenarios. Indicates the nominal distribution based on historical experience The expected value is to be obtained. for - Divergence generation function The conjugate dual function of is mathematically defined as . ,in Independent variables introduced to describe the conjugate dual mapping relationship; for the purposes of this invention - Divergence, whose exact analytical expression for its conjugate dual function is shown in equation (36):

[0226]

[0227] S3. Efficient solution of power grid dispatch strategy based on nested stochastic gradient descent: The inner estimation module and the outer update module are executed alternately. The inner estimation module fixes the current dispatch strategy, applies noise to generate a disturbance scenario around the historical wind nominal sample, and approximates the optimal dual multiplier under the current sample through first-order gradient descent. The outer update module substitutes the optimal dual multiplier into the weight evaluation formula, calculates the outer weighted gradient of the comprehensive loss function with respect to the power system decision variables, and ensures that the output of generators and energy storage does not exceed the limit through physical boundary projection, thus completing the update of decision variables. When the iterative convergence condition is met, the optimization stops and the optimal power grid dispatch strategy is output to achieve risk-resistant and robust dispatch of a high proportion of new energy power system. Figure 2 As shown, the specific steps include:

[0228] (1) Algorithm initialization and parameter configuration

[0229] Initialize power system physical decision variables (Including the initial planned output of each generator set and the reference power for energy storage charging and discharging), and initialize the number of outer iterations. Sample size per iteration Number of inner iterations Number of inner layer perturbation samples Initialize the historical sample set of wind and solar loads. Set the initial outer layer learning rate. and inner constant learning rate .

[0230] (2) Batch random sampling and inner dual multiplier estimation

[0231] The inner estimation module fixes the current global dispatch decision variables of the power system. Noise is applied around historical scenic nominal samples to generate perturbation scenarios. The optimal dual multiplier for the current sample is obtained by approximation using first-order gradient descent. Around the central sample, based on a reference measure (such as Gaussian white noise)... The generated perturbation scenario is shown in formula (37):

[0232]

[0233] In the formula: This refers to the randomly selected nominal historical sample number; For the first batch A historical center observation sample; This refers to the sequence number of the random perturbation scene generated based on the Gaussian reference measure; This represents the total number of perturbation scenarios generated for a single historical sample. To surround the central sample The generated first A random disturbance scenario variable; This is the corresponding Gaussian random noise vector.

[0234] To quickly approximate the penalty weight under the worst-case scenario, the current strategy is fixed. Define intermediate state variables Used to quantify the current strategy in a perturbation scenario. Physical loss deviation. Initialize inner layer variables. Calculate the current The intermediate mapping variables are shown in formula (38):

[0235]

[0236] In the formula, Estimate the number of iterations for the dual multipliers in the inner algorithm; This is the maximum number of iterations set for the inner layer; For the inner layer In the next iteration, intermediate mapping state variables are used to quantify the deviation of the system's physical loss. For the inner layer The dual multiplier value of the step.

[0237] use -Analytical expression of the derivative of the divergence conjugate function Calculate the inner objective function with respect to The stochastic gradient approximation value is shown in equation (39):

[0238]

[0239] In the formula: For the inner dual objective function with respect to The stochastic gradient approximation estimate.

[0240] The multipliers are updated using gradient descent as shown in equation (40):

[0241]

[0242] In the formula: A constant learning rate for optimizing the inner dual variables.

[0243] (3) Gradient estimation and physical projection update of outer scheduling policy

[0244] Obtained using inner layer calculations Calculate the comprehensive penalty loss with respect to the physical scheduling decision variables. Outer robust weighted gradient .

[0245] For each sample in the batch and the perturbation scene it generates Substitute back the optimal multiplier The dynamic weights of the worst-case distribution shift in the current scene are calculated as shown in formula (41):

[0246]

[0247] In the formula: An intermediate mapping state used to quantify the deviation of the physical loss of the system when the iteration reaches the optimum; This represents the number of outer iterations.

[0248] The dynamic weight expression, which is spontaneously assigned based on the severity of the penalty for each disturbance scenario, is shown in formula (42):

[0249]

[0250] In the formula: These are the dynamic weights that the algorithm automatically assigns based on the severity of the penalty for each perturbation scenario.

[0251] The stochastic gradient estimator that approximates the expected risk of the entire network scheduling is constructed as shown in Equation (43):

[0252]

[0253] In the formula: For the first In the outer layer iteration, the global weighted stochastic gradient estimator of the outer layer policy is comprehensively considered after taking into account both batch samples and perturbation scenarios; This is the maximum number of outer iteration rounds set; This is the sample size for a single iteration.

[0254] To eliminate the severe oscillations in dispatch commands caused by strong physical penalty terms near the minimum valley, this invention employs a momentum mechanism and a gradient-trunculated projective gradient descent method to update the power system control variables. The outer-layer stochastic gradient is truncated at extreme values ​​as shown in formula (44):

[0255]

[0256] In the formula, The maximum allowable gradient clipping threshold set to prevent gradient explosion.

[0257] Introducing momentum factor and decaying learning rate Accumulate historical gradient directions and calculate intermediate scheduling instructions as shown in formulas (45)-(46):

[0258]

[0259]

[0260] In the formula: The momentum factor (usually around 0.9) in the momentum gradient descent method is used to accumulate historical gradient directions to accelerate convergence; For the first The momentum-velocity vector of the outermost layer of the wheel during iteration; For the outermost layer The decaying learning rate of each iteration; This is an intermediate temporary scheduling instruction that has been updated by momentum gradient but has not yet been projected onto the physical boundary.

[0261] To prevent the algorithm from outputting control commands that violate safety regulations (such as requiring the generator to output negative power), the system calls the physical boundary projection operator to forcibly truncate the temporary decision commands and map them back to the physical safety domain of each device, as shown in formula (47):

[0262]

[0263] And the updated As the starting point for the next iteration.

[0264] When the number of iterations Reaching the set maximum number of convergence rounds Optimization stops when the proxy loss value is less than a preset threshold for multiple consecutive rounds. The final robust optimal scheduling strategy matrix is ​​then obtained. By breaking down the time series into 24-hour day-ahead dispatch plans, we can truly achieve robust control against risks in new power systems with a high proportion of renewable energy participation. This not only effectively avoids the risk of grid frequency collapse caused by extreme weather, but also significantly reduces the operation and maintenance costs brought about by conventional deterministic dispatch.

[0265] Example 2

[0266] To verify the technical advantages of the scheduling method proposed in this invention in improving the absorption of new energy sources and the economic efficiency of the system, such as... Figure 3 and Figure 4 As shown, the power allocation strategies for the same power grid under the same meteorological and load nominal conditions are compared for the previous 24 hours. The positive bars above the horizontal axis represent the total power supply of the system (including the actual grid-connected capacity of conventional thermal power units, wind power, and photovoltaic power, as well as energy storage discharge), while the negative bars below the horizontal axis represent the total power demand of the system (including the original electrical load and energy storage charging). Compared with the traditional Wasserstein distance-based split-blob optimization method, the generalized Sinkhorn split-blob optimization method reduces the output ratio of thermal power units during the midday peak photovoltaic power generation period, freeing up space for photovoltaic power to reach full grid connection; during the nighttime peak wind power generation period, it also prioritizes wind power consumption. In conjunction with the peak output of new energy sources, the algorithm of this invention enables the energy storage system to perform deep and concentrated charging during midday and early morning, efficiently storing excess energy and discharging it during the evening peak load period. By introducing a generalized Sinkhorn distance and divergence regularization mechanism, this invention effectively filters out extreme weather scenarios with extremely low probability and high defense costs in traditional robust optimization, avoiding overly conservative approaches. The algorithm's internal dynamic weight evaluation mechanism enables it to intelligently weigh the "marginal fuel cost of deep peak shaving of thermal power" against the "penalty cost of curtailment of new energy power" while ensuring the safety of the bottom line.

[0267] like Figure 5 As shown, the dynamic trajectory of the number of outer iterations and the overall system loss during the optimization process is illustrated, and the algorithm can achieve stable convergence within a hundred-step iteration range.

[0268] Example 3

[0269] like Figure 6 As shown, to verify the technical effectiveness of the method of the present invention in overcoming the limitations of probability distribution support sets and preventing risks from unknown weather scenarios, this embodiment artificially constructs an extremely severe weather period that has not been recorded in the historical dataset (such as...). Figure 6(The gray highlighted area is shown from 13:00 to 18:00). During this period, a sudden extreme weather event is simulated, causing a severe decrease in photovoltaic and wind power output. As shown in the bar chart, the traditional Wasserstein distance-based distributed bar optimization method shows a continuous increase in grid power deficit when facing extreme weather, exceeding that of the method proposed in this patent. This confirms that traditional methods, due to their strict reliance on known historical experience distributions, cannot predict and prevent extreme scenarios in historical blind spots, resulting in insufficient reserve capacity in their day-ahead scheduling plans, necessitating power curtailment measures during disasters. The fundamental reason for the superior performance mentioned above lies in the innovative introduction of an independent Gaussian reference measure when constructing the uncertain fuzzy set. This mechanism breaks the rigid dependence of traditional algorithms on historical support sets, giving the model the ability to explore the severe boundaries of historical data gaps.

Claims

1. A source-grid-load-storage collaborative scheduling method based on divergence regularized split-bar optimization, characterized in that, Includes the following steps: S1. Construct a power system operation model considering multiple uncertainties and transform it into a comprehensive loss function suitable for continuous gradient optimization: Collect power system network topology and historical data, define the set of uncertain random scenario variables and the global scheduling decision variable matrix of the power system; construct a comprehensive loss function that is highly coupled with the physical decision variables and the random scenario of new energy, including the operating cost model of conventional generator units and energy storage systems, the global power balance penalty function, the line active power flow limit violation penalty function, the unit ramping limit violation penalty function, the wind and solar curtailment penalty cost function, the energy storage SOC upper and lower limit violation and initial and final state limitation penalty function, and the energy storage state mutual exclusion penalty function; define the insurmountable physical limitations of equipment output and available power of new energy as the hard constraint boundary of the decision variables, and finally transform the source-grid-load-storage coordinated scheduling problem into a continuous nonlinear optimization model with hard constraint boundaries; S2. Construct and equivalently transform a distributed Sinkhorn bar collaborative scheduling model based on generalized Sinkhorn distance: To overcome the prediction blind spot caused by the actual meteorological distribution offset, the comprehensive loss function constructed in step S1 is used as the optimization objective, and a model with... - Construct an uncertain fuzzy set using divergence regularization and generalized Sinkhorn distance of Gaussian reference measure, and establish the original Min-Max sub-Bruker scheduling problem; through the inverse cumulative distribution function and Lagrange duality theory, the original problem is equivalently transformed into a nested bilayer continuous dual model with inner dual multipliers and outer scheduling strategy as decision variables, so as to achieve efficient solution in large-scale scenarios; S3. Efficient solution of power grid dispatch strategy based on nested stochastic gradient descent: The inner estimation module and the outer update module are executed alternately. The inner estimation module fixes the current dispatch strategy, applies noise to generate a disturbance scenario around the historical wind nominal sample, and approximates the optimal dual multiplier under the current sample through first-order gradient descent. The outer update module substitutes the optimal dual multiplier into the weight evaluation formula, calculates the outer weighted gradient of the comprehensive loss function with respect to the power system decision variables, and ensures that the output of generators and energy storage does not exceed the limit through physical boundary projection, thus completing the update of decision variables. When the iterative convergence condition is met, the optimization stops and the optimal power grid dispatch strategy is output to achieve risk-resistant and robust dispatch of a high proportion of new energy power system.

2. The source-grid-load-storage collaborative scheduling method based on divergence regularization split-bar optimization according to claim 1, characterized in that, In step S1, the set of uncertain random scenario variables and the matrix of decision variables are defined as follows: First, define the sources of uncertainty as shown in formulas (1)-(3): In the formula: The scheduling period; and They are respectively The vector between the theoretical available power output limit of a wind farm, which is constantly affected by weather, and the theoretical available power output limit of a photovoltaic power station; For the first The theoretical upper limit of usable output of a wind farm affected by weather conditions; For the first The theoretical upper limit of usable output of a photovoltaic power station affected by weather conditions; subscript Total number of wind farms; subscript This represents the total number of photovoltaic power stations; for The load demand vector of network nodes at any given time. for The Middle Active load demand vector of each network node; subscript This represents the total number of nodes in the power system. Then, define the set of uncertain random scenario variables representing extreme weather and user behavior at that moment. As shown in formula (4): Next, a discrete empirical nominal probability distribution is constructed based on historical data, as shown in formula (5): In the formula: For the first One historical observation sample; This represents the total number of days for which historical data was collected. To focus on the sample The Dirac measure function at the location; To construct discrete empirical nominal probability distributions based on historical data; For any time The scheduling instructions for each controlled device are shown in formulas (6)-(10): In the formula: For conventional generator sets at time The active power output vector; subscript This represents the total number of conventional thermal power units. , For energy storage systems at time The charging power and discharging power vector; subscript This represents the total number of energy storage systems. For wind farms at all times The actual grid-connected power output vector; For photovoltaic power stations at all times The actual grid-connected power output vector; For the first The conventional generator set at the time The instructions for generating active power; , The first An energy storage system at any time Charging power and discharging power commands; For the first Each wind farm at any time The actual grid-connected power consumption; For the first A photovoltaic power station at time The actual grid-connected power consumption; Finally, the global dispatch decision variable matrix of the power system was obtained. The definition is shown in formula (11): 。 3. The source-grid-load-storage collaborative scheduling method based on divergence regularization split-bar optimization according to claim 2, characterized in that, In step S1, a comprehensive loss function that is highly coupled with the physical decision variables and the stochastic scenario of new energy is constructed, specifically including: The operating cost model for conventional generator sets is constructed as shown in formula (12): In the formula: This represents the total operating cost of the generator sets in the system. , , For the first Cost coefficient of a generator set; A cost model for the energy storage system is constructed, and the depreciation cost of energy storage operation during the system scheduling cycle is calculated, as shown in formula (13): In the formula: This represents the total operating cost of the energy storage system within the system. This is the lifetime loss factor per unit throughput power of the energy storage system; The cost function for the curtailment penalty is defined as shown in formula (14): In the formula, The cost of penalizing the abandonment of wind and solar power; Weighting for wind curtailment power penalty; The penalty weight for wasted light power; The net injected power at each node is as shown in formula (15): In the formula: For each network node exist Net injection power at any given time; , , and The elements of the 0-1 correlation matrix indicate the number of elements. generator, the first The first wind farm, The first photovoltaic power station and the first Is the energy storage system connected to the node? If connected, the value is 1; otherwise, it is 0. A global power balance penalty function based on the L2 norm is introduced as shown in Equation (16): In the formula: For each network node exist Net injection power at any given time; This is the global power balance penalty function; For power balance penalty weights; The line active power flow over-limit penalty function based on the power transfer distribution factor matrix is ​​shown in formulas (17)-(18): In the formula: For the first The transmission line is in Trend value for a given time period; For PTDF matrix elements, representing nodes Injecting unit power to the line Sensitivity to current shifts; This represents the total number of lines; For the active power flow exceeding the limit penalty function of the line; For the line The maximum thermal stability transfer limit; The penalty weight for exceeding the limits of the trend; The unit ramp-up penalty function is constructed as shown in formula (19): In the formula: This is the function for penalizing over-limit ramping of the generator unit. For the unit Maximum permissible climbing rate limit; Weighting for penalties for exceeding the climbing limit; The penalty functions for exceeding the upper and lower limits of the energy storage SOC and the initial and final state constraints are constructed as shown in formulas (20)-(22): In the formula: For energy storage At any moment The energy storage state of charge; , These represent the charging and discharging efficiencies of energy storage, respectively. For energy storage Rated capacity; The scheduling time step; , These are the upper and lower safety limits for preventing battery overcharging and over-discharging; For the penalty function for exceeding the limit of the energy storage state of charge; , Energy storage at the beginning and end of the scheduling process The state of charge; , For penalty weighting; The initial and final state constraints of the energy storage are defined by a penalty function. The mutual exclusion penalty function for energy storage charging and discharging is shown in formula (23): In the formula: The mutual exclusion penalty function for energy storage charging and discharging; Mutually exclusive penalty weights; By adding the power generation and operation costs to various physical penalty functions, a comprehensive loss function is constructed. As shown in formula (24): In the formula, To cover scheduling cycles All time periods The global dispatch decision variables of the power system; To cover scheduling cycles All time periods Real-world, unknown, random scenario variables; Meanwhile, power system global dispatch decision variables The equipment must satisfy both the hard constraints of the absolute physical boundary and the hard constraints of the available output power in the local weather, as shown in formulas (25)-(29): In the formula: , The first Minimum and maximum technical output of a conventional generator set; , The first The maximum charging and discharging power of an energy storage system; For wind power Theoretical maximum output; For photovoltaic Theoretical maximum output; Furthermore, these hard constraints constitute the global scheduling decision variables of the power system. feasible domain Then, the global dispatch decision variables of the power system The model is transformed into a continuous nonlinear optimization model with hard-constrained boundaries, as shown in equation (30): In the formula: and These are the global dispatch decision variables of the power system. The absolute physical lower limit and absolute physical upper limit of various power levels.

4. The source-grid-load-storage collaborative scheduling method based on divergence regularization split-bar optimization according to claim 3, characterized in that, In step S2, a penalized Min-Max sub-bar optimization framework is introduced, and the original problem model for minimizing the expected comprehensive loss of the system under the worst distribution is established as shown in formula (31): In the formula: For the global dispatch decision variables of the power system, and satisfy the physical boundary feasible region. ; For real, unknown, random scenario variables; To obtain from the empirical nominal probability distribution Historical nominal scene samples obtained from sampling; In the potential worst probability distribution Below, the comprehensive loss function The mathematical expectation; To quantify the empirical nominal probability distribution With the potential worst probability distribution The generalized Sinkhorn distance between the differences; This is the divergence regularization coefficient, used to control the computational smoothness during the joint probability measure shift process. ; This is the robust penalty regularization coefficient, used to adjust the conservatism of the scheduling strategy in the face of extreme environmental fluctuations. ; Build with - The expression for constructing an uncertain fuzzy set using divergence regularization and the generalized Sinkhorn distance of the Gaussian reference measure is shown in formula (32): In the formula: Empirical nominal probability distribution and the potential worst probability distribution The joint probability distribution mapping; The complete set of the joint distribution that satisfies this marginal distribution condition; To measure variables in real-world, unknown, random scenarios Deviation from historical nominal scene samples The cost metric function for degree is based on the L2 norm Euclidean distance. ; A Gaussian reference measure independent of empirical distribution. , It is the identity matrix. The standard deviation of the Gaussian reference metric noise is set to control the boundary range of the exploration of extremely harsh scenarios; To measure the joint distribution mapping Compared with the benchmark measurement Differences - Divergence.

5. The source-grid-load-storage collaborative scheduling method based on divergence regularization split-bar optimization according to claim 4, characterized in that, In step S2, the expression for reconstructing the original maximization problem containing divergence constraints into a one-dimensional form is shown in equation (33): In the formula: Given a specific historical nominal scene sample and current power system global dispatch decision variables Under the condition of worst probability distribution, the conditional expected loss function faced by the power system; It is a density ratio function; Scalar random variable The inverse cumulative distribution function; To satisfy the function space where the probability density integral is 1; By introducing the Lagrange dual multiplier and using the strong duality theorem, the maximization problem is equivalently transformed into an unconstrained minimization problem, and finally a nested bi-level dual scheduling model is obtained, as shown in formulas (34)-(35): In the formula: As an inner-layer decision variable, it physically represents a sample of a historical nominal scenario. The threshold for economic penalties triggered by the power system's response to current extreme weather scenarios; Indicates a Gaussian reference measure independent of empirical distribution. The expected value is to be obtained. for - The conjugate dual function of the divergence generating function, for - Divergence, whose exact analytical expression for its conjugate dual function is shown in equation (36): 。 6. The source-grid-load-storage collaborative scheduling method based on divergence regularization split-bar optimization according to claim 5, characterized in that, In step S3, the inner estimation module fixes the current global scheduling decision variables of the power system. Noise is applied around historical scenic samples to generate a perturbation scenario. The optimal dual multiplier for the current sample is obtained by approximating it using first-order gradient descent. Around the central sample Based on Gaussian white noise The generated perturbation scenario is shown in formula (37): In the formula, This refers to the randomly selected nominal historical sample number; For the first batch A historical center observation sample; This refers to the sequence number of the random perturbation scene generated based on the Gaussian reference measure; This represents the total number of perturbation scenarios generated for a single historical sample. To surround the central sample The generated first A random disturbance scenario variable; This is the corresponding Gaussian random noise vector; To quickly approximate the penalty weight under the worst-case scenario, the current strategy is fixed. Define intermediate state variables Used to quantify the current strategy in a perturbation scenario. Physical loss deviation; initialization of inner layer variables Calculate the current The intermediate mapping variables are shown in formula (38): In the formula, Estimate the number of iterations for the dual multipliers in the inner algorithm; This is the maximum number of iterations set for the inner layer; For the inner layer In the next iteration, intermediate mapping state variables are used to quantify the deviation of the system's physical loss. For the inner layer The dual multiplier value of the step; use - The analytical expression of the derivative of the divergence conjugate function is used to calculate the inner objective function with respect to... The stochastic gradient approximation value is shown in equation (39): In the formula, For the inner dual objective function with respect to The stochastic gradient approximation estimate; The multipliers are updated using gradient descent as shown in equation (40): In the formula, A constant learning rate for optimizing the inner dual variables; The outer update module substitutes the optimal dual multiplier into the weight evaluation formula to calculate the outer weighted gradient of the comprehensive loss function with respect to the power system decision variables, which is obtained using the inner layer calculation. Calculate the comprehensive penalty loss with respect to the physical scheduling decision variables. Outer robust weighted gradient For each sample in the batch and the perturbation scene it generates Substitute back the optimal multiplier The dynamic weights of the worst-case distribution shift in the current scene are calculated as shown in formula (41): In the formula, This represents the number of outer iterations. The intermediate mapping state variables used to quantify the deviation of the system's physical loss when the iteration reaches the optimal point; The dynamic weight expression, which is spontaneously assigned based on the severity of the penalty for each disturbance scenario, is shown in formula (42): The stochastic gradient estimator that approximates the expected risk of the entire network scheduling is constructed as shown in Equation (43): In the formula, For the first In the outer layer iteration, the global weighted stochastic gradient estimator of the outer layer policy is comprehensively considered after taking into account both batch samples and perturbation scenarios; This is the maximum number of iterations for the outermost layer. The sample size for a single iteration; By using physical boundary projection to ensure that generator and energy storage commands do not exceed limits, the decision variables are updated as follows: The outer layer stochastic gradient is subjected to extreme value truncation as shown in Equation (44): In the formula, The maximum allowable gradient clipping threshold set to prevent gradient explosion; Introducing momentum factor and decaying learning rate Accumulate historical gradient directions and calculate intermediate scheduling instructions as shown in formulas (45)-(46): In the formula, This is the momentum factor in the momentum gradient descent method, used to accumulate historical gradient directions to accelerate convergence. For the first The momentum-velocity vector of the outermost layer of the wheel during iteration; For the outermost layer The decaying learning rate of each iteration; This is an intermediate temporary scheduling instruction that has been updated by momentum gradient but has not yet been projected onto the physical boundary. Furthermore, the physical boundary projection operator is invoked to forcibly truncate the temporary decision instructions and map them back to the physical security domain of each device, as shown in formula (47): And the updated As the starting point for the next iteration.

7. The source-grid-load-storage collaborative scheduling method based on divergence regularization split-bar optimization according to claim 6, characterized in that, In step S3, when the iterative convergence condition is met, the optimal source-grid-load-storage scheduling strategy is output: when the number of iterations... Reaching the set maximum number of outer iterations If the proxy loss value is less than the preset threshold for multiple consecutive rounds, optimization stops; the final robust optimal scheduling strategy matrix is ​​then obtained. The time series is broken down into scheduling plans within the scheduling period.

8. A source-grid-load-storage collaborative scheduling system based on divergence regularized split-bar optimization, used to implement the source-grid-load-storage collaborative scheduling method based on divergence regularized split-bar optimization as described in any one of claims 1-7, characterized in that, include: The modeling module is used to collect power system data, define uncertain random variables, and construct a continuous and differentiable comprehensive loss function and a feasible region with physical hard constraints on decision variables. The robust optimization module is used to construct fuzzy sets based on the generalized Sinkhorn distance, establish the Min-Max scheduling primal problem and transform its duality into a nested bi-level continuous dual problem; The solver module is used to perform inner dual multiplier estimation and outer scheduling policy update using a nested stochastic gradient descent algorithm. The scheduling execution module is used to output and distribute the optimal power grid scheduling strategy when the convergence conditions are met.

9. An electronic device, comprising a memory and a processor, characterized in that, The memory stores a computer program, and when the processor executes the computer program, it implements a source-grid-load-storage collaborative scheduling method based on divergence regularization split-bar optimization as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements a source-grid-load-storage collaborative scheduling method based on divergence regularization split-bar optimization as described in any one of claims 1-7.

Images