Source load risk management and control method for cooperative operation of multiple types of adjusting power supplies and new energy

Abstract

The invention discloses a source load risk management and control method for cooperative operation of multiple types of adjusting power supplies and new energy, and the method comprises the steps: depicting a prediction error tail dependency relationship through employing an extreme value theory and a Copula function, constructing a distribution robust optimization model in combination with a Wasserstein distance, and obtaining the standby capacity sharing, triggering threshold and risk index of the multiple types of adjusting power supplies; performing conformal prediction based on the prediction residual error to generate a dynamic confidence interval, and mapping the dynamic confidence interval into a pipeline radius of model prediction control; executing rolling MPC containing information bottleneck regularization under the constraint of the reserve capacity and the pipeline, and generating a robust scheduling strategy; and constructing a polyhedral feasible region model of the flexible load group according to the scheduling strategy and the risk index, and performing scaling to obtain a group control action. And the new energy consumption level and the system operation safety are obviously improved.

Description

Technical Field

[0001] This invention relates to the field of risk management technology, specifically a source-load risk management method for the coordinated operation of multiple types of regulating power sources and new energy sources. Background Technology

[0002] With the rapid growth of new energy installed capacity, the proportion of wind power, photovoltaic power and other power sources in the power system is constantly increasing. However, the output of such power sources is greatly affected by weather conditions, and has significant uncertainty and volatility, which puts pressure on system frequency and voltage stability, and rigidly increases the demand for reserve power.

[0003] At the same time, various types of regulating power sources (such as pumped storage, battery storage, gas turbine peak shaving and user-side flexible loads) are gradually being introduced into the system. Their operating characteristics differ significantly, including different response speeds, regulation durations, recovery costs and constraints. Without coordinated and optimized scheduling, resource utilization is often insufficient or excessive.

[0004] Existing methods largely rely on optimization under single forecasts and average scenarios, making it difficult to characterize tail risks under extreme weather and load resonance scenarios. Furthermore, traditional rolling MPC exhibits overfitting when facing forecast error drift, and control strategies struggle to remain robust across multiple scenarios. Additionally, the response of flexible load groups is typically modeled using empirical aggregation, lacking scalable geometric representations and risk-driven scaling mechanisms, resulting in insufficient executability at large scale. Summary of the Invention

[0005] In view of the above-mentioned problems, the present invention is proposed.

[0006] Therefore, the technical problem solved by this invention is: the difficulty in quantifying tail risks, the tendency of control strategies to overfit, and the insufficient executability of flexible loads in the coordinated operation of new energy sources and multiple types of regulating power sources.

[0007] To address the aforementioned technical problems, this invention provides the following technical solution: a source-load risk management method for the coordinated operation of multiple types of regulating power sources and new energy sources, comprising:

[0008] Collect new energy and load forecast data, use the Copula function to characterize the tail dependency relationship of forecast error, and construct a split-bar optimization model by combining Wasserstein distance to determine the reserve capacity allocation and trigger threshold of multiple types of regulating power sources, and output the risk index.

[0009] Based on the prediction residuals, conformal prediction is performed to generate dynamic confidence intervals, and the interval width is adjusted according to the risk index. The interval is mapped to the pipeline radius constraint of rolling MPC, forming a dynamically shrinking feasible region.

[0010] Under the constraints of reserve capacity and pipeline, a rolling MPC with information bottleneck regularization is executed, and a robust scheduling strategy is generated by combining multi-scenario perturbation sampling and majority decision aggregation.

[0011] Based on the scheduling strategy and risk index, a polyhedral feasible region model of the flexible load group is constructed; the polyhedral feasible region model of the flexible load group is scaled, and the group control action is obtained through linear programming.

[0012] By combining the cross-domain invariant risk model with time-series logic-based adversarial stress testing, the tail-dependent model, conformal prediction, and feasible domain structure are modified, thereby forming a risk management chain of extreme modeling, dynamic control, group execution, and closed-loop correction.

[0013] As a preferred embodiment of the source-load risk management method for the coordinated operation of multiple types of regulating power sources and new energy sources described in this invention, the prediction error sample set is constructed by the measured output data of new energy sources and the AMI data of the load side. When the tail data is insufficient, the extreme value theory is used to fit the residual of the threshold and generate supplementary samples to enhance the coverage of extreme events.

[0014] Based on the supplemented samples, the Copula function is used to characterize the tail dependency relationship between new energy sources and load error;

[0015] The Copula distribution is used as the center of the empirical distribution, and the uncertainty set is constructed by the Wasserstein distance, the radius of which is set according to the sample size and confidence level.

[0016] The output of the uncertainty set is the set of possible distributions, which serves as the input to the pluripotent optimization model.

[0017] As a preferred embodiment of the source-load risk management method for the coordinated operation of multiple types of regulating power sources and new energy sources described in this invention, the sub-Bruker optimization model includes the following: taking the uncertainty set as input, the objective function is composed of reserve cost and imbalance condition risk value, and the constraints include regulating power source capacity, ramp rate, energy storage SOC, cycle life, pumped storage water level and grid power flow limit.

[0018] The model is solved using Benders decomposition and scenario reduction methods, and outputs the reserve capacity share, trigger threshold, and risk index of various types of regulating power supplies.

[0019] The reserve share and trigger threshold are input into the capacity constraint of the rolling MPC; the risk index is input into the conformal prediction and flexible load feasible domain scaling process.

[0020] As a preferred embodiment of the source-load risk management method for the coordinated operation of multiple types of regulating power sources and new energy sources described in this invention, the conformal prediction model takes historical prediction data and the risk index as input, clusters historical samples based on similar weather patterns and daily load patterns, and performs online residual calibration within the clusters to obtain the confidence intervals of new energy sources and loads.

[0021] The risk index is used to dynamically adjust the confidence interval width to match the system risk level;

[0022] The generated confidence interval is mapped to the pipeline radius of the rolling MPC and then generalized to the upper bound using PAC-Bayes to ensure that the risk of exceeding the limit remains under control in unseen scenarios.

[0023] As a preferred embodiment of the source-load risk management method for the coordinated operation of multiple types of regulating power sources and new energy sources described in this invention, the robust control strategy includes simultaneously receiving the reserve capacity share, the trigger threshold, and the pipeline radius constraint as inputs to the rolling MPC; and making the output curves of multiple types of regulating power sources output by the MPC.

[0024] To reduce overfitting of the control to the prediction noise, a mutual information regularization approximation term is added to the objective function of minimizing the operating cost. In each rolling cycle, the confidence interval boundary and tail dependency parameters are perturbed and sampled to form multiple scenario subproblems that are solved in parallel. A robust scheduling strategy is obtained by majority aggregation.

[0025] The scheduling strategy includes the target output sequence of various regulating power sources and the power reduction / transfer requirements required by the load side.

[0026] As a preferred embodiment of the source-load risk management method for the coordinated operation of multiple types of regulating power sources and new energy sources described in this invention, the polyhedral feasible domain model of the flexible load group includes: using the load-side power demand in the robust scheduling strategy and the risk index, analyzing the power response speed, sustainability duration and recovery cost of different types of controllable loads, forming piecewise linear response curves and aggregating them into a multidimensional polyhedral model.

[0027] The risk index is used to determine the feasible region scaling ratio, and the power demand is used as the scaling target to jointly drive the linear programming solution to solve the specific action plan of the flexible load group.

[0028] The solution ensures execution stability through low-rank decomposition, monotonic constraints, and hysteresis mechanisms, and generates execution feedback data after actual execution.

[0029] As a preferred embodiment of the source-load risk management method for the coordinated operation of multiple types of regulating power sources and new energy sources described in this invention, the closed-loop correction includes: using the execution feedback data and system operation monitoring data, extracting consistent features across meteorological and load environments using an invariant risk minimization method, and using the consistent features as input to correct the conformal prediction model and the polyhedral feasible domain model of the flexible load group, so as to reduce the impact of environmental differences on the accuracy of prediction and modeling.

[0030] The adversarial stress test based on time-series logic generates the worst-case perturbation sequence that meets preset conditions, and feeds the worst-case perturbation sequence back to the tail dependency relation, the sub-Bruker optimization model, and the polyhedral feasible domain model of the flexible load group for retraining and recalibration.

[0031] A source-load risk management system for the coordinated operation of multiple types of regulating power sources and new energy sources using any of the methods described in this invention, wherein: a data acquisition unit collects new energy and load forecast data, uses the Copula function to characterize the tail dependency relationship of the forecast error, and constructs a split-bar optimization model in combination with Wasserstein distance to determine the reserve capacity allocation and trigger threshold of multiple types of regulating power sources, and outputs a risk index;

[0032] The constraint unit generates a dynamic confidence interval by performing conformal prediction based on the prediction residual, and adjusts the interval width according to the risk index, mapping the interval to the pipeline radius constraint of rolling MPC, forming a dynamically shrinking feasible region.

[0033] The strategy analysis unit performs rolling MPC with information bottleneck regularization under the constraints of reserve capacity and pipeline, and generates robust scheduling strategies by combining multi-scenario disturbance sampling and majority decision aggregation.

[0034] The action generation unit constructs a polyhedral feasible region model of the flexible load group based on the scheduling strategy and risk index; it scales the polyhedral feasible region model of the flexible load group and obtains the group control actions through linear programming.

[0035] The feedback optimization unit executes the feedback input cross-domain invariant risk model and the adversarial stress test based on time-series logic to correct the tail dependency model, conformal prediction and feasible domain structure, thereby forming a risk management chain of extreme modeling - dynamic control - group execution - closed-loop correction.

[0036] A computer device includes: a memory and a processor; the memory stores a computer program, wherein: when the processor executes the computer program, it implements the steps of the method described in any one of the present invention.

[0037] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method described in any one of the present invention.

[0038] The beneficial effects of this invention are as follows: The source-load risk management method for the coordinated operation of multiple types of regulating power sources and new energy sources provided by this invention enhances the tail risk characterization by introducing extreme value theory and Copula function, and constructs a partial Bruker optimization model by combining Wasserstein distance to achieve robust allocation of reserve capacity of multiple types of regulating power sources; it uses conformal prediction and PAC-Bayes calibration to generate dynamic confidence intervals, avoiding the amplification effect of prediction error drift on control; it introduces information bottleneck regularization and multi-scenario disturbance sampling mechanism in rolling MPC, effectively reducing the overfitting dependence of scheduling strategies on a single prediction scenario; by constructing a multidimensional polyhedral model of flexible load groups and scaling it with a risk index, it can transform macro-scheduling objectives into executable group actions, improving load-side response capabilities; finally, it uses cross-domain invariant risk modeling and adversarial stress testing based on time-series logic to achieve closed-loop correction, enabling the system to have self-learning and adaptive capabilities. Compared with existing methods, this invention not only improves the robustness and safety margin of the system under extreme conditions, but also significantly improves the utilization efficiency of regulating power sources and flexible loads. Attached Figure Description

[0039] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 The overall flowchart of the source-load risk management method for the coordinated operation of multiple types of regulating power sources and new energy sources provided by the present invention is shown. Detailed Implementation

[0041] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0042] Reference Figure 1 As one embodiment of the present invention, a source-load risk management method for the coordinated operation of multiple types of regulating power sources and new energy sources is provided, including:

[0043] S1: Collect new energy and load forecast data, use the Copula function to characterize the tail dependency relationship of forecast error, and construct a split-bar optimization model by combining Wasserstein distance to determine the reserve capacity allocation and trigger threshold of multiple types of regulating power sources, and output the risk index.

[0044] In step S1, a prediction error sample set for both new energy sources and load needs to be constructed first. This sample set is formed by combining measured new energy output data and load-side AMI data. A residual sequence is obtained by comparing the predicted and measured values. For the tail data in the residual sequence that exceeds the threshold, when their quantity is insufficient to support risk modeling, extreme value theory is used for fitting. Extreme value theory generates additional extreme residual samples by fitting the residuals exceeding the threshold to a generalized Pareto distribution, thus compensating for the insufficiency of the original tail data. After this step, an expanded sample set containing extreme samples is obtained.

[0045] After obtaining the expanded sample set, the Copula function is used to characterize the tail dependency between the power output error and the load error of new energy sources. By selecting an appropriate family of Copulas (such as t-Copula), the joint dependency of the error variables in the tail region can be accurately expressed while keeping the marginal distribution unchanged. The resulting Copula distribution not only reflects the statistical characteristics of the margins but also captures the linkage risk under extreme conditions; this distribution is used as the center point of the empirical distribution.

[0046] Building upon this, the Wasserstein distance is introduced to construct an uncertainty set, used to characterize the deviation between the empirical distribution and the potential true distribution. Let the empirical distribution be... Candidate distribution is The Wasserstein distance is defined as follows:

[0047]

[0048] in, Describe the candidate distribution. Let denot be the empirical distribution based on Copula, π represent the joint distribution, Ξ represent the support set of the random variable, ξ represent the value of the random variable under the candidate distribution, ξ′ represent the value of the random variable under the empirical distribution, and ||ξ-ξ′|| represent the norm distance under a given metric. The calculation result of the Wasserstein distance characterizes the minimum transport cost of two distributions in the sample space. The uncertainty set constructed based on this is:

[0049]

[0050] in, The sphere represents the Wasserstein sphere centered on the empirical distribution, and ε represents the radius, the value of which is set according to the sample size and confidence level.

[0051] This set is ultimately used as the input to the sub-Bruker optimization model. In the sub-Bruker optimization modeling process, with the uncertainty set as input, the equipment allocation vector as r, and the risk penalty level as t, the objective function is:

[0052]

[0053] Where r represents the reserve capacity vector of various types of regulating power sources, C represents the reserve cost vector of regulating power sources, κ represents the risk penalty coefficient, t represents the auxiliary variable in CVaR, β represents the confidence level, K represents the sample size, and z k Let represent the excess loss variable in the k-th scenario, ε represent the Wasserstein radius, and λ represent the Lipschitz constant related to the dual constraint. The objective function consists of three parts: The term represents the reserve cost, the term in parentheses represents the CVaR risk value, and the final κελ represents the robustness penalty due to the set of uncertainties.

[0054] The constraints include: power supply capacity constraints, i.e., the actual output does not exceed the capacity limit; ramp rate constraints, to ensure that the output changes are within the allowable range; energy storage SOC constraints, to ensure that the energy storage state of charge is reasonable; cycle life constraints, to avoid excessive wear and tear on energy storage equipment; pumped storage water level constraints, to ensure water level safety; and power flow constraints, to ensure the safe operation of the lines.

[0055] The optimization problem is solved using Benders decomposition and scenario reduction methods. The resulting solution consists of three parts: the reserve capacity share of various regulating power sources, the trigger threshold, and the system risk index. The obtained reserve capacity share and trigger threshold are used as capacity constraint inputs for rolling MPC to ensure sufficient safety margins are reserved in real-time scheduling. The risk index is then input into the conformal prediction model and the flexible load feasible region scaling stage, serving as the basis for dynamically adjusting the confidence interval and scaling magnitude.

[0056] S2: Based on the prediction residual, perform conformal prediction to generate a dynamic confidence interval, and adjust the interval width according to the risk index. Map the interval to the pipeline radius constraint of rolling MPC to form a dynamically shrinking feasible region.

[0057] Intra-cluster online residual calibration is performed and basic confidence intervals are generated:

[0058] Step 1: Cluster historical samples based on similar weather patterns and daily load patterns:

[0059] Construct a feature vector for clustering. Given a sample index t, concatenate weather elements with the daily load pattern as follows:

[0060] f t =[w t ,d t ]

[0061] Subsequently, K-means or Gaussian mixture was used for clustering, and the cluster labels were denoted as:

[0062]

[0063] Where f t w represents the clustering feature vector of the t-th sample. t d represents the weather feature vector corresponding to this sample. t This represents the load pattern encoding vector corresponding to the sample, c(t) represents the cluster number to which the t-th sample is assigned, C represents the total number of clusters, and μ j Let ||j|| represent the center vector of the j-th cluster, and ||·||2 represent the L2 norm.

[0064] Step 2: Calculate residuals within the cluster and perform conformal calibration:

[0065] Point forecast of new energy output or load Compared with the measured y t Forming residuals In the calibration window C of cluster c c Internally defined inconsistent scores:

[0066]

[0067] Set the confidence level to 1-α and calculate the quantiles:

[0068] q c =Quantile 1-α ({s τ |τ∈C c})

[0069] The basic confidence interval is obtained:

[0070] When c(t) = c

[0071] in Let y represent the point prediction at time t. t Let r represent the measured value at time t. t Let s represent the prediction residual at time t. τ Let C represent the non-consistency score of the τ-th sample within the cluster. c Let α represent the set of calibration samples for cluster c, α represent the default rate target for the interval, and q represent the set of calibration samples for cluster c. cThis represents the inconsistency threshold obtained by dividing the cluster by 1-α. This represents the fundamental conformal confidence interval at time t in cluster c. A quantile is a value that, when a set of numbers or the values ​​of a random variable are arranged in ascending order, represents the quantile at a specific proportion.

[0072] Adaptively widen the confidence interval using a risk index:

[0073] Let the risk index output by upstream S1 be denoted as λ. r Define a monotonically amplifying function:

[0074] γ(λ r ) = clip(1+aλ r ,γ min ,γ max )

[0075] And widen the quantile threshold and interval:

[0076]

[0077] Where λ r γ(λ) represents the system risk index. r ) represents the interval amplification factor positively correlated with the risk index, a represents the sensitivity coefficient from risk to the amplification factor, and γ represents the risk level. min γ represents the lower limit of the amplification factor. max This indicates the upper limit of the magnification factor. This represents the risk-adjusted quantile threshold at time t. This represents the risk-adjusted confidence interval at time t.

[0078] Map the confidence interval to the pipeline radius of the rolling MPC:

[0079] Assume the net load or net output forecast is And there are intervals Define the radius of symmetry:

[0080]

[0081] And the uncertainty set of the prediction error is denoted as:

[0082]

[0083] For linearized system models:

[0084] x t+1 =Ax t +Bu t +Ew t

[0085] By employing the compaction approach of pipeline MPC, the influence of the uncertainty set on the constraints is transformed into a radius vector:

[0086] ρ t =Sδ t

[0087] Then tighten the control domain to:

[0088]

[0089] Where δ t Let represent the scalar radius induced by the confidence interval at time t. Let w represent the set of disturbance uncertainties at time t, and let x represent the disturbance variable. t Indicates the system state, u t Let A represent the control input, B represent the state transition matrix, E represent the disturbance input matrix, and ρ represent the control input matrix. t Let S represent the shrinkage radius vector over the controllable feasible region, and let S represent the mapping matrix from the scalar derived from sensitivity or robust optimization to the shrinkage radius of each resource. This represents the lower bound vector of the control quantity. represents the upper bound vector of the control quantity, m represents the safety margin coefficient vector on the lower bound side, n represents the safety margin coefficient vector on the upper bound side, and ⊙ represents element-wise multiplication. Simultaneously, upstream reserve capacity allocation and trigger threshold are incorporated into the capacity constraint.

[0090]

[0091] Where r represents the reserve capacity retention vector, and τ represents the reserve release threshold vector. This represents an indicator vector used to determine whether the risk index exceeds a threshold based on the elements.

[0092] Using the PAC-Bayes upper bound to generalize and calibrate the pipe radius, we define empirical risks for out-of-bounds events caused by confidence intervals:

[0093]

[0094] Applying the PAC-Bayes upper bound to the pipeline radius strategy distribution Q and prior P:

[0095]

[0096] And select the radius family or scaling factor with the goal of minimizing the bound:

[0097]

[0098] It's important to understand that for steps like conformal forecasting (or uncertain pipelines), to meet the predetermined default rate (e.g., 99% coverage or a certain confidence level), a fixed interval radius isn't directly defined. Instead, from a set of candidate radii (a family of radii), scaling methods such as generalization upper bounds (e.g., PAC-Bayes bounds) are used to select the smallest radius that satisfies the target generalized default probability, which is then used for final control / constraint. In other words, this "family of radii" is a set of multiple possible "confidence interval scales / pipeline contraction radii," and the goal is to find the optimal radius among these candidate parameters that guarantees future risk remains within the target threshold.

[0099] in Let R represent the empirical violation rate observed under the radius policy R, N represent the number of samples used for evaluation, 1 represent the indicator function, and violate(R,t) indicate whether the radius policy R leads to constraint violation at time t. Let L(R) represent the expectation of the posterior distribution Q, L(R) represent the true risk, KL(Q||P) represent the relative entropy between the posterior and prior distributions, P represent the prior distribution, Q represent the posterior distribution, δ represent the confidence parameter, and R0 represent the confidence parameter. * This represents the radius strategy that minimizes the upper bound.

[0100] In practice, the prior can be taken from the empirical distribution of the historical best radius in adjacent seasons or adjacent regions, and the posterior can be obtained by fitting the current data. If the threshold exceeds the target default rate threshold, then γ(λ) is increased. r Increase m and n to tighten the contraction until the threshold does not exceed the target.

[0101] In practice, the prior can be taken from the empirical distribution of the historical best radius in adjacent seasons or adjacent regions, and the posterior can be obtained by fitting the current data. If the threshold exceeds the target default rate threshold, then γ(λ) is increased. r Alternatively, increase m and n to tighten the contraction until the threshold does not exceed the target.

[0102] Execution in rolling MPC:

[0103] In each rolling cycle, a finite-time domain optimization is constructed, incorporating the aforementioned contraction and capacity reserve constraints. The objective function includes running cost and information bottleneck regularization terms. The control quantity at the current moment is obtained and executed, and then updated in the next cycle. With ρ t Continue rolling. This transforms the conformal region into a verifiable pipeline radius through risk adaptation and PAC-Bayes calibration, and constrains the available capacity in accordance with the upstream reserve allocation logic, ultimately forming a dynamically shrinking feasible region.

[0104] S3: Execute rolling MPC with information bottleneck regularization under the constraints of reserve capacity and pipeline, and combine multi-scenario disturbance sampling and majority decision aggregation to generate a robust scheduling strategy.

[0105] The robust control strategy includes simultaneously receiving the reserve capacity share, the trigger threshold, and the pipeline radius constraint as inputs to the rolling MPC; and adjusting the power output curves of multiple types of MPC outputs. A mutual information regularization approximation term is added to the objective function of minimizing operating costs to reduce overfitting of the control to prediction noise. Furthermore, perturbation sampling is performed on the confidence interval boundaries and tail dependency parameters in each rolling cycle to form multiple scenario subproblems that are solved in parallel. A robust scheduling strategy is obtained through majority aggregation.

[0106] The scheduling strategy includes the target output sequence of various regulating power sources and the power reduction / transfer requirements required by the load side.

[0107] Specifically:

[0108] Optimization goals and information bottleneck regularization for rolling MPC:

[0109] Solve the finite-time optimization problem H at each rolling time t:

[0110]

[0111] Where u k c represents the control vector for the k-th time period. k Let y represent the marginal cost vector for the k-th time period, α represent the weighting coefficient of the tracking deviation penalty, and y represent the marginal cost vector for the k-th time period. k This represents the observable output quantity obtained from the system state mapping. Represents the reference trajectory for net load or safe operation, ||·||1 represents the 1-norm of the vector, η represents the weighting coefficient of the ramp smoothing penalty, and u k-1 λ represents the control vector of the previous time period. IB The weight coefficients representing information bottleneck regularization. Represents the control strategy distribution π k Distribution with baseline strategy The KL divergence.

[0112] The information bottleneck is concretized into the KL closed form of the Gaussian strategy:

[0113]

[0114] in Σ represents the baseline control vector for the k-th time period. u Let represent the covariance matrix of the control noise. Thus, the objective function is equivalent to adding a quadratic regularization to the economic and safety objectives to suppress excessive control response to noise or random features.

[0115] System and disturbance model and pipeline contraction constraint: linearized or piecewise linearized time-domain model is used.

[0116] x k+1 =Ax k +Bu k +Ew k ,y k =Cx k

[0117] Where x k Let A represent the system state vector at the k-th time interval, B represent the control input matrix, E represent the disturbance input matrix, and w represent the system state vector at the k-th time interval. k Let y represent the prediction error perturbation vector for the k-th time period. k Let C represent the observable output vector, and let C represent the mapping matrix from state to output.

[0118] Map the conformal region generated by S2 to the pipe radius and shrink the control domain:

[0119] u min +m⊙ρ k ≤u k ≤u max -n⊙ρ k

[0120] Where u min u represents the lower control bound vector for each resource. max ρ represents the upper control bound vector for each resource. k Let m represent the feasible region shrinkage radius vector obtained by mapping the confidence interval radius of the k-th time period, m represent the shrinkage weight coefficient vector on the lower bound side, n represent the shrinkage weight coefficient vector on the upper bound side, and ⊙ represent element-wise multiplication.

[0121] Embedding of reserve capacity and trigger threshold: Incorporate the reserve allocation and threshold of S1 into the capacity reserve and conditional release.

[0122]

[0123] Where u i,k This represents the control amount of resource i during time period k. r represents the physical upper limit of resource i. i This represents the share of spare capacity for resource i. The indicator function λ represents the risk index exceeding a threshold. r τ represents the systemic risk index output by the upstream DRO. i This represents the reserve release threshold for resource i. This expression ensures that the resource is left unused when not triggered, and the reserved capacity can only be accessed after a trigger.

[0124] Multi-scenario perturbation sampling and parallel subproblems: neighborhood sampling around confidence intervals and tail dependency parameters.

[0125] θ (s) =(γ) (s) ,ψ (s) ),s=1,…,K

[0126] And solve the corresponding sub-MPC for each scenario:

[0127]

[0128] Where θ (s) Let γ represent the parameter set for the s-th perturbation scenario. (s) This represents the perturbation sampling of the interval amplification factor, ψ (s) This indicates perturbation sampling of the Copula tail parameters, where K represents the number of parallel scenarios. Let J(·) represent the optimal control in the k-th time period under the s-th scenario, and let J(·) represent the aforementioned objective function that includes information bottleneck and contraction constraints. Let r represent the shrinkage radius obtained from the confidence interval under scenario s, r represent the reserve share vector, τ represent the trigger threshold vector, and λ represent the truncation radius. r This represents the risk index. A set of candidate control sequences is obtained through parallel solution.

[0129] Perform majority voting or coordinate median aggregation on the current execution step size:

[0130]

[0131] in This represents the aggregation control quantity at time t, and median(·) represents the robust aggregation operator that takes the median of the components. This represents the first-step control solution for scenario s. This aggregation suppresses overfitting to a particular anomalous scenario.

[0132] Additional considerations regarding hill climbing and energy constraints:

[0133]

[0134] Among them ||·|| ∞ Let R represent the infinity norm, and let R represent the maximum ramp vector for each resource. η represents the energy state of the stored energy at time k, Δt represents the time step, and η represents the energy state of the stored energy at time k. ch Indicates charging efficiency. Indicates the charging power control quantity, η dis Indicates discharge efficiency. This represents the discharge power control quantity. A linearized power flow inequality can be incorporated by combining safety power flow and voltage constraints.

[0135] Only the first step of control is issued during the rolling execution. Refresh ρ using t←t+1 and the latest predictions and measurements. k With λ r Repeat the above solution process. The optimized overall strategy provides two types of outputs: one is the target output sequence of various regulating power sources in the prediction time domain, used for upstream coordination and downstream verification; the other is the reduced or transferred power demand on the load side in each time period, used for the flexible load geometric scaling and combination execution in S4. In this way, reserve space and triggering, bottleneck suppression of information from intervals to pipelines to prevent overfitting, and scenario majority decision robustness are consistently implemented within the same rolling framework, forming an executable and disturbance-resistant robust scheduling strategy.

[0136] S4: Based on the scheduling strategy and risk index, construct a polyhedral feasible region model of the flexible load group; scale the polyhedral feasible region model of the flexible load group and obtain the group control action through linear programming.

[0137] Furthermore, the polyhedral feasible region model of the flexible load group includes analyzing the power response speed, sustainability duration, and recovery cost of different types of controllable loads using the load-side power demand in the robust scheduling strategy and the risk index, forming piecewise linear response curves and aggregating them into a multidimensional polyhedral model. The risk index is used to determine the feasible region scaling ratio, with the power demand as the scaling target, jointly driving linear programming to solve for the specific action scheme of the flexible load group. The scheme ensures execution stability through low-rank decomposition, monotonic constraints, and hysteresis mechanisms, and generates execution feedback data after actual execution.

[0138] It's important to know that the specific construction process is as follows:

[0139] First, the polyhedral feasible region is derived from the equipment to the aggregation. Different types of controllable loads are first processed into piecewise linear responses at the equipment level, and then aggregated to obtain a multidimensional polyhedron. The aggregated time-period decision vector is denoted as x. t The feasible region can be written as a standard polyhedron:

[0140]

[0141] Where x t Let A represent the flexible load aggregation action vector at time t, A represent the coefficient matrix formed by stacking inequalities such as power upper limit, duration, energy budget, and recovery upper limit, and b represent the resource upper bound vector and budget vector paired with matrix A. ≤ represents the baseline feasible region set without risk scaling, and ≤ represents the element-wise inequality constraint.

[0142] To ensure that the linear solution size is controllable, a low-rank structure approximation is introduced, and the action vector is expressed using latent variables:

[0143] x t =Vy t Uy t ≤b

[0144] Where y t Let V represent the low-dimensional latent decision vector at time t, V represent the loading matrix mapping the latent space to the actual device aggregation space, U represent the constraint matrix projected onto the latent space after approximate elimination of the original constraints, and ≤ denotes an element-wise inequality. The b on the right-hand side of the equality sign has the same meaning as in the above equation. This transformation corresponds to an approximation of A as follows: The low-rank decomposition significantly reduces the dimensionality of linear programming variables.

[0145] Power combining between device categories is given by a linear mapping:

[0146]

[0147] Where p t denoted by , where represents the total reduction or transfer power that can be provided by the aggregator at time t, and g represents the weight vector that converts various flexible actions into equivalent active power. Indicates the transpose operation, x t This represents the aforementioned aggregated action vector.

[0148] The next step is to geometrically scale the polyhedron using a risk index. The risk index λ obtained upstream is then... r Convert to scaling factor σ t And expand the feasible region:

[0149] σ t =clip(σ0+κλ) r,t ,σ min ,σ max )

[0150] in σ represents the scaled feasible region at time t. t The scaling factor at time t is represented by `clip(·)`, which is a truncation operator that restricts the value within the parentheses to a given upper and lower bound. σ0 represents the baseline scaling level at risk-free conditions, κ represents the risk-to-scaling sensitivity coefficient, and λ... r,t σ represents the risk index at time t. min σ represents the minimum value of the scaling factor. max This represents the maximum value of the scaling factor; other symbols retain their previous meanings. This formula achieves monotonic scaling, where the wider the feasible region, the greater the risk.

[0151] To avoid hysteresis and rate limiting caused by frequent jitter:

[0152]

[0153] Where s t Indicates the actual scaling instruction used for execution, s t-1 This indicates the scaling instruction from the previous moment. This indicates an increase in the component. This indicates the downward adjustment component, ≥0 indicates a non-negativity constraint, ρ0 represents the allowable range of change in the base, ρ1 represents the weighting coefficient for the range of change with risk, and w t Represents the relationship between |λ and |λ| r,t -λ r,t-1 |A linear upper bound variable is used to inject risk changes into the rate allowance. Set the execution scaling to s. t =σ t Soft tracking, and in the target Adding penalties can create a binary-free hysteresis and smoothing mechanism.

[0154] Then, linear programming is used with scheduling requirements as the objective. The load-side requirements output from the previous rolling MPC step are written as follows: Solve linear programming problems within a time-domain window t = t0, ..., t0 + H-1:

[0155]

[0156] Ax t ≤s t b, x t =Vy t Uy t ≤b,

[0157] Where the objective function is φr represents the marginal cost of the action at time t. t This indicates a penalty for the loss of energy or comfort during recovery. This represents a smooth penalty for changes in scaling instructions; `min` indicates minimization; curly braces represent the decision set in the time domain; Ax in the first constraint t ≤s t 'b' indicates that the scaling command is applied to the polyhedron, in the second constraint. This indicates that the polymerization power should meet or exceed the requirements proposed by the MPC, specifically the x in the third and fourth constraints. t =Vy t With Uy t ≤b represents the equivalent variables and constraints after applying the low-rank approximation.

[0158] ∑ in the last constraint t Γx t≤h represents the energy budget or daily cumulative comfort limit across time periods, Γ represents the coefficient matrix projecting actions from each time period onto the cumulative energy or comfort level, and h represents the corresponding upper limit vector. (Symbols) Represents the total reduction or transfer demand at time t, μ represents the smoothing penalty weight, φ represents the recovery or comfort penalty weight, and r t This represents a linear measure of the recovery or comfort loss determined by the action, and its linearity can be obtained by weighting and summing the recovery coefficients of the equipment class. All inequalities are element-wise.

[0159] Device class embedding for piecewise linear response: To avoid the introduction of integer variables, the convex hull method is used to linearize the relationship between the power and duration of the device class.

[0160] Class variables for a certain type of device at time t Represented by segment weights:

[0161]

[0162] in θ represents the equivalent action quantity of category j at time t. jlt L represents the convex combination weight of the l-th segment of category j at time t. j α represents the number of segments in category j. jl This indicates the power or power-duration trade-off point corresponding to that segment. By categorizing various types... Stack into x t By incorporating the aforementioned linear relationship into matrix A and vector b, the piecewise response can be incorporated into the aggregated polyhedron. It's important to note that a "motion quantity" is a scalar, referring to a single adjustment value of an object at a given moment; a "motion vector" is a set of motion quantities stacked along its dimensions. The two have a "component-vector" relationship, and mathematically, linear mapping or projection is often used to restore a vector to the desired motion quantity.

[0163] Furthermore, the monotonic constraint and execution stability mechanism includes using a linear approximation of the monotonicity from risk to scaling:

[0164] s t ≥s min +κ min λ r,t ,s t ≤s max +κ max λ r,t

[0165] Where s t Indicates scaling, λ r,t Indicates the risk index, s min κ represents the minimum feasible scaling at zero risk.min This represents the minimum slope approximation, s max κ represents the maximum feasible scaling at zero risk. max This represents the maximum slope approximation. The dual inequality at λ... r Within the working range, an upper and lower envelope is formed, thereby achieving a solvable monotonicity constraint. (As mentioned above...) The rate limits work together to create hysteresis and smooth execution without integer variables.

[0166] Linear programming provides the solution for x over the entire time domain. t With the corresponding p t And perform scaling s t x at a given moment t Control commands are mapped to specific load groups and issued, while the actual achieved power and comfort or process indicators are recorded, forming execution feedback for subsequent closed-loop processes. Feedback data and x t The corresponding segment weight θ jlt Writeback is available for updating α jl The calibration of matrices A, U, and V maintains the accuracy of the polyhedron under equipment aging and seasonal drift.

[0167] S5: The cross-domain invariant risk model with feedback input and adversarial stress testing based on time-series logic are used to correct the tail-dependent model, conformal prediction and feasible domain structure, thereby forming a risk management chain of extreme modeling - dynamic control - group execution - closed-loop correction.

[0168] The closed-loop correction includes using the execution feedback data and system operation monitoring data, employing an invariant risk minimization method to extract consistent features across meteorological and load environments, and using these consistent features as input to correct the conformal prediction model and the polyhedral feasible region model of the flexible load group, thereby reducing the impact of environmental differences on prediction and modeling accuracy. Based on temporal logic, an adversarial stress test generates a worst-case perturbation sequence that meets preset conditions, and this worst-case perturbation sequence is fed back to the tail dependency relation, the bibliometric optimization model, and the polyhedral feasible region model of the flexible load group for retraining and recalibration.

[0169] It's important to understand that the power tracking error generated by the execution layer, along with equipment execution deviations and field monitoring data, constitutes the feedback data. This feedback data is fed into two parallel calibration branches. The first is an invariant risk minimization branch, aiming to extract causal features that are stable across weather and load conditions, and use these features to recalibrate conformal prediction and the feasible region of flexible loads, thereby improving generalization stability under normal conditions. The second is a time-series logic-based adversarial stress testing branch, aiming to actively synthesize worst-case disturbance sequences, write extreme operating conditions back to the tail-dependent modeling and the radius and feasible region boundary of the Bruker bar, and strengthen extreme safety margins. The new parameters generated by the two branches are updated and consistent, then written back to the upstream modeling, prediction, and execution modules, and are directly used in the next rolling cycle, forming a closed loop of extreme modeling, dynamic control, group execution, and closed-loop correction.

[0170] First, implement the invariant risk minimization branch. Divide the feedback samples according to environment labels and learn invariant representations across multiple environments using the original training samples. The surrogate objective is adopted as follows:

[0171]

[0172] Where Φ represents the mapping from the original features to the invariant features, w represents the prediction head that works on the invariant features, M represents the number of environments, and R i (·) represents the empirical risk in the i-th environment, ° represents the composition of functions, and β represents the weight of the gradient penalty. Let ||·||² represent the gradient of the prediction head after fixing it at a calibration point, and let ||·||² represent the L2 norm. By minimizing this objective, the optimal head gradient can be made consistent across different environments, forcing Φ to learn a cross-environment stable factor.

[0173] Recalibrate and shape-preservingly predict using invariant features. For samples belonging to the same invariant feature cluster, update quantiles using the new residuals added during execution:

[0174]

[0175] in Denotes the new inconsistency threshold for the h-th cluster, Quantile 1-α The coverage rate is represented by the quantile operator minus alpha, where α represents the target default rate and y represents the measured target quantity. This represents point predictions using invariant features as input, where (x, y) represents a feature-label pair. This represents the cluster set augmented with execution feedback and new environment samples. Writing back will give you the updated interval based on consistent features.

[0176] The invariant characteristics allow the polyhedral parameters of the flexible load to adapt to the environment.

[0177] Using a linear homeomorphic form:

[0178]

[0179] Where A(z) represents the constraint coefficient matrix of a given invariant feature z, b(z) represents the constraint upper bound vector of a given invariant feature z, A0 represents the constraint coefficient matrix of the baseline environment, b0 represents the upper bound vector of the baseline environment, L represents the dimension of the invariant feature, and z l Let A represent the ell-th component of the invariant feature. l b represents the matrix correction along the ell-th characteristic direction. l This represents the upper bound correction along the ell-th feature direction. This parameterization allows environmental differences to be mapped to changes in the shape and capacity of the geometrically feasible region.

[0180] Further implementation of the stress testing branch. Transform operational concerns into time-series logic formulas and use robust semantics to generate worst-case perturbations:

[0181]

[0182] Where ω * Let represent the optimal parameters for the worst-case perturbation, ω represent the adjustable parameter vector generated by the perturbation, and Ω represent the feasible set of perturbation parameters. Representing sequential logic formulas Robustness from the initial time step on the system trajectory generated by the parameter omega. This indicates that the robustness is negated to maximize the violation, and s(ω) represents the simulated system trajectory driven by the disturbance parameters. ω is obtained from... * Then, a combined disturbance sequence of weather, load, and grid structure that conforms to the preset extreme mode can be obtained.

[0183] The tail samples generated by the worst perturbation are augmented to tail dependency modeling, and the new Copula parameters are estimated using maximum likelihood estimation.

[0184]

[0185] Where θ * Let θ represent the optimal Copula parameters, and θ represent the parameter vector to be estimated. c represents the augmented dataset consisting of the original pseudo-observations and adversarial examples. θ (u,v) represents the value of the Copula density with parameter theta on the pseudo-observation pair, and log represents the natural logarithm. Let θ... * Writing back allows for a better match between the joint tail dependence and extreme resonance.

[0186] Adjusting the sub-Blule radius based on robustness statistics on adversarial examples to amplify the uncertain set:

[0187]

[0188] Where ε new Denotes the new Wasserstein radius, ε base κ represents the basic radius. ρ The sensitivity coefficient representing the robustness to radius. This represents the average robustness on a set of worst-case perturbation samples, and max(0,·) indicates that only negative violations are amplified. The radius after writing back transforms the risk of distribution mismatch into geometric blank space.

[0189] The conformal threshold is updated using quantiles including adversarial examples, in the same way as the previous equation, with the same notation and meaning. The flexible load polyhedron is conservatively processed under extreme execution conditions:

[0190] b new =b(z)-Δh

[0191] Where b new Let b(z) represent the adjusted upper bound vector, b(z) represent the vector based on the query ChatGPT, and Δh represent the deduction vector calculated from the peak duration caused by the adversarial trajectory and the cumulative energy gap. The write-back boundary provides additional buffering during extreme periods.

[0192] Unify the above multi-source updates to the parameter layer and use exponential moving average for stable implementation:

[0193]

[0194] in This indicates a specific type of model or hyperparameter currently in effect. This represents the corresponding value from the previous period. Let represent the optimal value obtained in this round of training, and η represent the update step size. By uniformly referring to Copula parameters, Wasserstein radius, conformal threshold, and polyhedral system number, a stable and non-overshooting parameter write-back can be obtained.

[0195] Finally, the engineering implementation was defined. A two-stage cycle, fast and slow, was adopted on the timeline. The fast cycle was synchronized with the rolling MPC, using the parameters written back after the previous closed-loop correction to complete real-time scheduling and execution. The slow cycle was triggered every few rolling cycles, aggregating execution feedback and monitoring data to complete invariant feature learning, conformal recalibration, and polyhedral adaptation. Simultaneously, it performed temporal logic adversarial stress testing, Copula re-estimation, and extreme conservatism of the Wasserstein radius and polyhedral boundaries. All new parameters were exponentially smoothed and consistently written back. The next fast cycle directly read the new parameters, and the width of the interval, the radius of the pipeline, capacity margins, and load scaling were all adaptively adjusted. In this way, data was processed into parameter updates along two branches, and the parameters were written back to the upstream model. The model immediately took effect in the next rolling control, resulting in a rigorous closed-loop logic and a clear implementation path.

[0196] On the other hand, this embodiment also provides a source-load risk management system for the coordinated operation of multiple types of regulating power sources and new energy sources, which includes:

[0197] The data acquisition unit collects new energy and load forecast data, uses the Copula function to characterize the tail dependency relationship of the forecast error, and constructs a bibliometric optimization model based on the Wasserstein distance to determine the reserve capacity allocation and trigger threshold of various types of regulating power sources, and outputs the risk index.

[0198] The constraint unit generates a dynamic confidence interval by performing conformal prediction based on the prediction residual, and adjusts the interval width according to the risk index, mapping the interval to the pipeline radius constraint of rolling MPC, forming a dynamically shrinking feasible region.

[0199] The strategy analysis unit performs rolling MPC with information bottleneck regularization under the constraints of reserve capacity and pipeline, and generates robust scheduling strategies by combining multi-scenario disturbance sampling and majority decision aggregation.

[0200] The action generation unit constructs a polyhedral feasible region model of the flexible load group based on the scheduling strategy and risk index; it scales the polyhedral feasible region model of the flexible load group and obtains the group control actions through linear programming.

[0201] The feedback optimization unit executes the feedback input cross-domain invariant risk model and the adversarial stress test based on time-series logic to correct the tail dependency model, conformal prediction and feasible domain structure, thereby forming a risk management chain of extreme modeling - dynamic control - group execution - closed-loop correction.

[0202] If the above functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0203] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0204] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0205] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0206] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A source-load risk management method for the coordinated operation of multiple types of regulating power sources and new energy sources, characterized in that, include: Collect new energy and load forecast data, use the Copula function to characterize the tail dependency relationship of forecast error, and construct a split-bar optimization model by combining Wasserstein distance to determine the reserve capacity allocation and trigger threshold of multiple types of regulating power sources, and output the risk index. Based on the prediction residual, a conformal prediction is performed to generate a dynamic confidence interval. The interval width is adjusted according to the risk index, and the interval is mapped to the pipeline radius constraint of rolling MPC to form a dynamically shrinking feasible region. Under the constraints of reserve capacity and pipeline, a rolling MPC with information bottleneck regularization is executed, and a robust scheduling strategy is generated by combining multi-scenario perturbation sampling and majority decision aggregation. Based on the scheduling strategy and risk index, a polyhedral feasible region model of the flexible load group is constructed; the polyhedral feasible region model of the flexible load group is scaled, and the group control action is obtained through linear programming. By combining the cross-domain invariant risk model with time-series logic-based adversarial stress testing, the tail-dependent model, conformal prediction, and feasible domain structure are modified, thereby forming a risk management chain of extreme modeling, dynamic control, group execution, and closed-loop correction.

2. The source-load risk management method for the coordinated operation of multiple types of regulating power sources and new energy sources as described in claim 1, characterized in that: The prediction error sample set is constructed by combining measured data of new energy output with load-side AMI data. When tail data is insufficient, extreme value theory is used to fit the residuals exceeding the threshold and generate supplementary samples to enhance coverage of extreme events. Based on the supplemented samples, the Copula function is used to characterize the tail dependency relationship between new energy sources and load error; The Copula distribution is used as the center of the empirical distribution, and the uncertainty set is constructed by the Wasserstein distance, the radius of which is set according to the sample size and confidence level. The output of the uncertainty set is the set of possible distributions, which serves as the input to the pluripotent optimization model.

3. The source-load risk management method for the coordinated operation of multiple types of regulating power sources and new energy sources as described in claim 2, characterized in that: The sub-Bruker optimization model includes an objective function consisting of backup cost and imbalance condition risk value, with the uncertainty set as input, and the constraints including regulating power supply capacity, ramp rate, energy storage SOC, cycle life, pumped storage water level and grid power flow limit. The model is solved using Benders decomposition and scenario reduction methods, and outputs the reserve capacity share, trigger threshold, and risk index of various types of regulating power supplies. The reserve share and trigger threshold are input into the capacity constraint of the rolling MPC; the risk index is input into the conformal prediction and flexible load feasible domain scaling process.

4. The source-load risk management method for the coordinated operation of multiple types of regulating power sources and new energy sources as described in claim 3, characterized in that: The conformal prediction model takes historical prediction data and the risk index as input, clusters historical samples based on similar weather patterns and daily load patterns, and performs online residual calibration within the clusters to obtain the confidence intervals of new energy sources and loads. The risk index is used to dynamically adjust the confidence interval width to match the system risk level; The generated confidence interval is mapped to the pipeline radius of the rolling MPC and then generalized to the upper bound using PAC-Bayes to ensure that the risk of exceeding the limit remains under control in unseen scenarios.

5. The source-load risk management method for the coordinated operation of multiple types of regulating power sources and new energy sources as described in claim 4, characterized in that: The robust control strategy includes simultaneously receiving the reserve capacity share, the trigger threshold, and the pipeline radius constraint as inputs to the rolling MPC; and enabling the MPC output to have multiple types of adjustable power output curves. To reduce overfitting of the control to the prediction noise, a mutual information regularization approximation term is added to the objective function of minimizing the operating cost. In each rolling cycle, the confidence interval boundary and tail dependency parameters are perturbed and sampled to form multiple scenario subproblems that are solved in parallel. A robust scheduling strategy is obtained by majority aggregation. The scheduling strategy includes the target output sequence of various regulating power sources and the power reduction / transfer requirements required by the load side.

6. The source-load risk management method for the coordinated operation of multiple types of regulating power sources and new energy sources as described in claim 5, characterized in that: The polyhedral feasible domain model of the flexible load group includes analyzing the power response speed, sustainability duration and recovery cost of different types of controllable loads using the load-side power demand and the risk index in the robust scheduling strategy, forming piecewise linear response curves and aggregating them into a multidimensional polyhedral model. The risk index is used to determine the feasible region scaling ratio, and the power demand is used as the scaling target to jointly drive the linear programming solution to solve the specific action plan of the flexible load group. The solution ensures execution stability through low-rank decomposition, monotonic constraints, and hysteresis mechanisms, and generates execution feedback data after actual execution.

7. The source-load risk management method for the coordinated operation of multiple types of regulating power sources and new energy sources as described in claim 6, characterized in that: The closed-loop correction includes using the execution feedback data and system operation monitoring data to extract consistent features across meteorological and load environments using an invariant risk minimization method, and using the consistent features as input to correct the conformal prediction model and the polyhedral feasible domain model of the flexible load group, so as to reduce the impact of environmental differences on prediction and modeling accuracy. The adversarial stress test based on time-series logic generates the worst-case perturbation sequence that meets preset conditions, and feeds the worst-case perturbation sequence back to the tail dependency relation, the sub-Bruker optimization model, and the polyhedral feasible domain model of the flexible load group for retraining and recalibration.

8. A source-load risk management system for the coordinated operation of multiple types of regulating power sources and new energy sources using the method described in any one of claims 1-7, characterized in that: The data acquisition unit collects new energy and load forecast data, uses the Copula function to characterize the tail dependency relationship of the forecast error, and constructs a split-bar optimization model by combining the Wasserstein distance to determine the reserve capacity allocation and trigger threshold of multiple types of regulating power sources, and outputs the risk index. The constraint unit generates a dynamic confidence interval by performing conformal prediction based on the prediction residual, and adjusts the interval width according to the risk index, mapping the interval to the pipeline radius constraint of rolling MPC, forming a dynamically shrinking feasible region. The strategy analysis unit performs rolling MPC with information bottleneck regularization under the constraints of reserve capacity and pipeline, and generates robust scheduling strategies by combining multi-scenario disturbance sampling and majority decision aggregation. The action generation unit constructs a polyhedral feasible region model of the flexible load group based on the scheduling strategy and risk index; it scales the polyhedral feasible region model of the flexible load group and obtains the group control actions through linear programming. The feedback optimization unit executes the feedback input cross-domain invariant risk model and the adversarial stress test based on time-series logic to correct the tail dependency model, conformal prediction and feasible domain structure, thereby forming a risk management chain of extreme modeling - dynamic control - group execution - closed-loop correction.

9. A computer device, comprising: A memory and a processor; the memory stores a computer program, characterized in that: when the processor executes the computer program, it implements the steps of the method 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 the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-7.

Images