A new energy power generation base water-light storage day-ahead stochastic scheduling method for extreme scenarios

Abstract

The application discloses a new energy power generation base water-light-storage day-ahead random dispatching method for coping with extreme scenes, and relates to the technical field of new energy and power system optimization dispatching. The method generates a large number of extreme scenes through a Copula-GAN model, and utilizes MCMC technology to screen out a simplified scene set with statistical representation from the extreme scenes; further, the SAO optimization algorithm is adopted to realize the day-ahead random dispatching of water-light-storage collaborative complementation, the search range is widened to avoid local optimum; at the same time, based on the sensitivity analysis engine and the Bayesian deep learning, the robustness is evaluated, and a hierarchical risk early warning mechanism is established; finally, the SGA optimization algorithm is utilized to quickly correct the external sending dispatching plan by using the extreme scene risk information, and the safety and stability of long-term operation of the power system are ensured. The application effectively solves the problems of extreme scene modeling loss in traditional dispatching, low multi-energy collaborative efficiency, insufficient dynamic early warning mechanism and dispatching correction lag and the like.

Description

Technical Field

[0001] This invention relates to the field of new energy and power system optimization scheduling technology, specifically to a random scheduling method for hydropower, solar power, and energy storage bases in response to extreme scenarios. Background Technology

[0002] With the large-scale integration of new energy sources such as hydropower and photovoltaics into power generation bases, their output is significantly affected by extreme weather (such as extreme cold without sunlight, extreme heat without wind, etc.), leading to huge challenges for traditional dispatching methods, mainly manifested in the following ways:

[0003] First, there is a lack of extreme scenario modeling. Existing scenario generation methods are unable to accurately capture the spatiotemporal dependence and extreme fluctuation characteristics of wind and solar resources, resulting in insufficient adaptability of scheduling plans to rare high-risk events.

[0004] Secondly, the multi-energy coordinated scheduling is inefficient. The water-solar-storage complementary model gets stuck in local optima due to high-dimensional nonlinear constraints (water balance, power limitation, etc.). Traditional optimization algorithms have slow convergence speed and unstable solution quality.

[0005] Furthermore, the dynamic risk warning mechanism is lacking, the fixed threshold warning model cannot quantify the uncertainty of scheduling parameters, has a high false alarm rate, and lacks the ability to sensitively locate vulnerable features.

[0006] Finally, the delay in external dispatching correction, and the reliance on manual experience for adjusting external power in extreme scenarios, can easily lead to a chain of failures due to response delays.

[0007] To address the aforementioned issues, a new scheduling method is urgently needed to resolve the problems inherent in traditional methods and achieve safe and stable operation of the power system under extreme scenarios. Summary of the Invention

[0008] The purpose of this invention is to provide a random scheduling method for hydropower, solar power, and energy storage bases in response to extreme scenarios, so as to overcome the problems of lack of extreme scenario modeling, low efficiency of multi-energy collaborative scheduling, lack of dynamic risk early warning mechanism, and lag in external transmission scheduling correction in traditional scheduling methods.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] A stochastic scheduling method for hydro-solar-storage days in new energy power generation bases to cope with extreme scenarios includes:

[0011] S1: Multi-scene generation is performed based on the Copula-GAN model, and redundant scenes are reduced using the MCMC method to obtain typical extreme scene samples;

[0012] S2: Establish a day-ahead optimization scheduling model for the synergistic complementarity of hydro-solar-storage with the goal of minimizing total operating costs, and solve it using the snow ablation optimization algorithm (SAO).

[0013] S3: Based on the Bayesian deep learning module, the robustness of the scheduling plan is evaluated, and a dynamic hierarchical risk warning mechanism is established in combination with the sensitivity analysis engine;

[0014] S4: Based on the day-ahead scheduling plan and risk warning results, a fast correction model for new energy transmission scheduling considering extreme scenario risks is established, and the Snow Goose Optimization Algorithm (SGA) is used to solve it, finally outputting a robust scheduling plan.

[0015] Furthermore, the multi-scene generation based on the Copula-GAN model includes:

[0016] Collect historical time-series data and actual data on hydropower output, photovoltaic power, and energy storage equipment;

[0017] Marginal distribution fitting is performed for each feature dimension; kernel density estimation is used for continuous variables, and empirical cumulative distribution function is used for discrete variables.

[0018] Perform Copula dependency structure modeling, construct a loss function that includes adversarial loss and regularization term, and train Copula-GAN;

[0019] A uniform vector is synthesized using a trained generator, and then the final scene sample is generated through inverse transformation.

[0020] Furthermore, the scenario for reducing redundancy using the MCMC method includes:

[0021] Assign uniform probability weights to the initial large-scale scene set;

[0022] Construct a Markov chain, define the state space and transition probabilities, and generate a scenario through Gaussian perturbation;

[0023] Define a scene distance function to measure similarity and an acceptance rate function to control the probability of acceptance of similar scenes;

[0024] The scene weights are updated by iterative sampling, and finally the top N representative scenes are selected and their weights are renormalized.

[0025] Furthermore, the objective function of the day-ahead optimization scheduling model for the synergistic complementarity of hydropower, solar power and energy storage is to minimize the total cost, including hydropower operation cost, reserve cost and curtailment cost;

[0026] The constraints that the model needs to satisfy include: system power balance constraints, water balance constraints, reservoir water level constraints, hydropower station discharge capacity constraints, hydropower station output limits, unit head constraints, photovoltaic safe operation range constraints, curtailment equation constraints, and energy storage system operation constraints.

[0027] Furthermore, the method of using the Snow Ablation Optimization Algorithm (SAO) for solving the problem specifically includes:

[0028] Initialize the random group and calculate the initial particle positions;

[0029] Gaussian Brownian motion is used to model the diffusion characteristics of snow or meltwater vaporization and update particle positions, while the global search range is expanded by combining the centroid of the elite set.

[0030] The snow melting model is used for refined search and constraint processing. Elite particles are retained and the global optimal solution is updated until the maximum number of iterations or the cost convergence threshold is reached.

[0031] Furthermore, the Bayesian deep learning robustness evaluation module employs a Bayesian neural network trained with Monte Carlo Dropout, whose input is scheduling plan parameters and output is a system risk score and uncertainty measure.

[0032] The sensitivity analysis engine consists of three parts: feature perturbation analysis, global Sobol exponent calculation, and adversarial robustness testing.

[0033] Furthermore, the dynamic hierarchical risk early warning mechanism sets three levels of early warning thresholds based on the expected risk value, prediction variance, and sensitivity index: Level 1 low-risk early warning prompts monitoring, Level 2 medium-risk early warning prompts adjustment of scheduling plans, and Level 3 high-risk early warning triggers the activation of emergency response plans.

[0034] The warning threshold is calibrated based on historical accident data and dynamically adjusted through a sliding window mechanism.

[0035] Furthermore, the new energy transmission dispatching fast correction model aims to minimize the total cost after correction, and adds transmission channel risk constraints, transmission power allocation constraints, and corrected power balance constraints.

[0036] The total cost includes an outbound risk penalty based on a risk correction factor, which is determined by the risk warning level.

[0037] Furthermore, the method of using the Snow Goose Optimization Algorithm (SGA) to solve the corrected model specifically includes:

[0038] The initialization phase defines the population position and velocity matrices;

[0039] The introduction of a heading angle function control algorithm marks the transition from the exploration phase to the development phase.

[0040] During the exploration phase, a V-shaped flight strategy was adopted, and the population was divided into elite, weak, and intermediate groups according to fitness and updated accordingly.

[0041] During the development phase, a linear flight strategy is adopted, combined with Brownian random walks and collective guidance mechanisms to escape local optima.

[0042] In summary, the present invention has at least one of the following beneficial technical effects:

[0043] This invention proposes a day-ahead stochastic scheduling method for new energy power generation bases that addresses extreme scenarios, overcoming problems such as lack of extreme scenario modeling, low efficiency of multi-energy collaborative scheduling, lack of dynamic risk early warning mechanism, and lag in external transmission scheduling correction in traditional scheduling methods. Specifically:

[0044] (1) A method for generating and reducing multiple scenarios based on Copula-GAN model and MCMC is proposed. The method combines prediction data of water, light, and storage, uses Copula-GAN method to generate multiple scenarios, and uses MCMC to reduce scenarios, generating a statistically representative set of extreme scenarios.

[0045] (2) A random scheduling method for hydro-solar-storage synergy based on SAO optimization algorithm is proposed. The initial day-ahead prediction data and scenario set are obtained, hydro-solar-storage synergy modeling is performed, and the SAO algorithm is used to solve the problem, which expands the search range and avoids getting trapped in local optima.

[0046] (3) A risk warning method based on sensitivity analysis engine and Bayesian deep learning robustness assessment is proposed. By outputting risk expectation and variance through Monte Carlo Dropout, the prediction blind zone is identified, uncertainty is quantified, and a hierarchical warning mechanism is established by combining feature perturbation analysis and global Sobol index.

[0047] (4) A fast correction method for new energy transmission scheduling is adopted using extreme scenario risk and SGA optimization algorithm. SGA realizes population grouping and parallel optimization, improves computational efficiency, and accelerates model solving. Adding correction factor coupled risk warning realizes automatic reduction of transmission power in high-risk situations, ensuring the safety and stability of the power system in long-term operation. Attached Figure Description

[0048] Figure 1 This is a flowchart of the present invention;

[0049] Figure 2 This is a schematic diagram of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0051] like Figure 1 and Figure 2 As shown, a stochastic scheduling method for hydro-solar-storage days in new energy power generation bases to cope with extreme scenarios includes:

[0052] S1: Multi-scene generation is performed based on the Copula-GAN model, and redundant scenes are reduced using the MCMC method to obtain typical extreme scene samples.

[0053] Multi-scene generation based on Copula-GAN model:

[0054] First, historical time-series data and actual data of hydropower output, photovoltaic power output, and energy storage equipment were collected. Historical time-series data of each node in the base power grid were collected, including historical data of hydropower output, photovoltaic power output, and energy storage equipment output, as well as corresponding actual data of hydropower output, photovoltaic power output, and energy storage equipment output.

[0055] Secondly, marginal distribution fitting is performed for each feature dimension. Kernel density estimation is used for continuous variables, and empirical cumulative distribution function is used for discrete variables.

[0056] Marginal distribution fitting is performed, fitting a cumulative distribution function (CDF) separately for each feature dimension d (e.g., hydropower forecasting, photovoltaic forecasting, etc.). For continuous variables, kernel density estimation (KDE) can be used instead of empirical CDF to obtain a smooth distribution. For discrete variables, empirical CDF is used directly. Empirical distribution function (EDF) estimation is then performed. CDF:

[0057] (1)

[0058] In the formula, N is the number of samples; It is the marginal distribution function; This is an indicator function (1 when the condition is met, 0 otherwise); Let d represent the d-th feature value of the i-th sample.

[0059] Mapping the original data to a uniform distribution space separates marginal distributions from dependency structures. For each sample... calculate:

[0060] (2)

[0061] at this time It follows a D-dimensional standard uniform distribution [0,1]D.

[0062] Next, Copula dependency structure modeling is performed, a loss function containing adversarial loss and regularization term is constructed, and Copula-GAN is trained.

[0063] Perform Copula dependency structure modeling, and complete Copula selection and parameter estimation:

[0064] (3)

[0065] (4)

[0066] (5)

[0067] in, is the parameter vector of the Copula function; u is the cumulative probability vector of the marginal distribution; lnL represents the log-likelihood value; k is the number of parameters.

[0068] When training Copula-GAN, it is necessary to construct a loss function.

[0069] (6)

[0070] (7)

[0071] In Equation 6-7, The total loss function is composed of the sum of the first two adversarial loss terms and the third regularization term. For the discriminator to analyze real data samples The output probability; E[.] represents the synthesized data generated by the generator based on random noise z and uniformly distributed variable u; E[.] represents the mathematical expectation. The Copula regularization term forces the load of the distributed structure of the generated data to be pre-defined. This represents the regularization coefficient, which controls the strength of regularization.

[0072] Finally, uniform vectors are synthesized using the trained generator, and then inversely transformed to generate the final scene samples.

[0073] Generate new samples that conform to the statistical characteristics of the original data, use the trained generator G to generate a synthetic uniform vector, and perform an inverse transformation on each dimension d to synthesize the final sample, thus completing scene generation.

[0074] (8)

[0075] The scenario for reducing redundancy using the MCMC method is as follows:

[0076] First, uniform probability weights are assigned to the initial large-scale scene set.

[0077] Assume the initial scene set is (M is extremely large), each scenario This is a multi-dimensional time series (hydropower, photovoltaic, energy storage power). Initial uniform probability weights are assigned:

[0078] (9)

[0079] Secondly, a Markov chain is constructed, the state space and transition probabilities are defined, and the scenario is generated through Gaussian perturbation.

[0080] Define the state space and design the transition probabilities Each state corresponds to a scenario. .

[0081] (10)

[0082] in, Indicates from arrive The proposed distribution: , The covariance matrix for each dimension of the scene is used to control the perturbation amplitude; For Gaussian perturbation terms; This represents the acceptance rate adjustment factor.

[0083] Next, we define a scene distance function to measure similarity and an acceptance rate function to control the acceptance probability of similar scenes.

[0084] Define the distance between scenes and the acceptance rate. The scene distance function measures similarity, and the acceptance rate adjustment factor controls the probability of acceptance of similar scenes.

[0085] (11)

[0086] (12)

[0087] In equations (11) and (12), T is the time step; D is the feature dimension. For feature weights; For temperature parameters; This is the scene distance function.

[0088] Finally, the scene weights are updated through iterative sampling, and the top N representative scenes are selected and their weights are renormalized. Specifically:

[0089] (1) Iterate from k=1 to K (preset number of iterations):

[0090] a. From the current scenario Generate candidate scenes (through proposal distribution) ).

[0091] b. Calculate the probability of acceptance .

[0092] c. Based on probability accept Set to the new state; otherwise, retain the previous state. .

[0093] d. Update scene weights:

[0094] (13)

[0095] in, This represents the scene weight after the k-th iteration.

[0096] (2) Based on the final weight Sort in descending order and select the first N scenes (N M).

[0097] (3) Renormalize the weights:

[0098] (14)

[0099] Obtain typical extreme scenario samples. After iteratively calculating all weight values, obtain the actual data portion and obtain extreme scenario and extreme scenario subclass data with probability weights.

[0100] S2: Establish a day-ahead optimization scheduling model for the synergistic complementarity of hydro-solar-storage with the goal of minimizing total operating costs, and solve it using the snow ablation optimization algorithm (SAO).

[0101] The objective function of the day-ahead optimization scheduling model for hydropower-solar-storage synergy is to minimize the total cost, including hydropower operating cost, reserve cost, and curtailment cost. The constraints that the model must satisfy include: system power balance constraints, water balance constraints, reservoir water level constraints, hydropower station discharge capacity constraints, hydropower station output limits, unit head constraints, photovoltaic safe operation range constraints, curtailment equation constraints, and energy storage system operation constraints.

[0102] First, data such as initial storage capacity, inflow rate during each period of the scheduling period, daily load demand curve, and photovoltaic power forecast curve are obtained.

[0103] Secondly, establish mathematical and physical models related to hydropower, photovoltaics, and energy storage to formulate day-ahead scheduling plans that integrate hydropower, photovoltaics, and energy storage.

[0104] The objective function of the day-ahead optimization scheduling model for synergistic and complementary hydro-solar-storage systems is as follows:

[0105] (15)

[0106] (16)

[0107] In equation (15-16), The total cost of optimizing the scheduling in the current day; , and These are the operating costs of hydropower, standby costs, and costs of curtailed solar power. and The water value coefficient for the operation of hydropower units; , and These are the capacity cost coefficients for upper reserve, lower reserve, and emergency reserve, respectively. The average power of hydropower unit i during time period t; , and These are the upper reserve capacity, lower reserve capacity, and emergency reserve capacity of unit i, respectively. The penalty factor for discarding light is yuan / MW; This represents the amount of solar power curtailment from unit j; For time period sets, A collection of hydroelectric power units; A collection of photovoltaic units.

[0108] The day-ahead optimization scheduling model for hydro-solar-storage complementary systems should satisfy constraints such as water balance, reservoir water level, and power balance.

[0109] Power balance constraints:

[0110]

[0111] in, This represents the local load demand during time period t.

[0112] Water balance constraints:

[0113] (17)

[0114] in, The storage capacity at the beginning of time period t is expressed in m³. The average inflow rate of unit i during time period t is expressed in m³ / s. The average power generation flow of unit i during time period t is expressed in m³ / s. The time step for decision-making is represented by h.

[0115] Hydropower conversion relationship:

[0116] (18)

[0117] (19)

[0118] in, The overall unit output coefficient representing time period t is related to the power generation flow rate and the power generation head. The average head of the hydropower station during time period t, in meters; Represents the upstream reservoir water level at the end of time period t; This represents the downstream tailwater level at the end of time period t.

[0119] Reservoir water level constraints:

[0120] (20)

[0121] (twenty one)

[0122] in, and Represents the lower and upper limits of the reservoir's water level; This represents the relationship between the reservoir capacity and the reservoir water level of a hydropower station.

[0123] Permissible discharge constraints for hydropower stations:

[0124] (twenty two)

[0125] (twenty three)

[0126] in, and These represent the minimum and maximum allowable discharge flows of the hydropower station, respectively. The following functions represent the relationship between the flow rate and the tailrace level of the power plant.

[0127] Hydropower station output limitations:

[0128] (twenty four)

[0129] in, and These represent the minimum and maximum allowable output of the hydropower station, respectively.

[0130] Unit head constraints:

[0131] (25)

[0132] in, and These represent the minimum and maximum generating heads of the hydropower station, respectively.

[0133] Constraints on the safe operating range of photovoltaic systems:

[0134] (26)

[0135] in, Let be the minimum output of the i-th photovoltaic unit at time t; Let be the maximum output of the i-th photovoltaic unit at time t; This represents the actual output of the i-th photovoltaic unit during time period t; The predicted output of the i-th photovoltaic unit during time period t.

[0136] And it is subject to the light loss equation constraint:

[0137] (27)

[0138] Energy storage system operating constraints:

[0139] (28)

[0140] in, It is a binary number representing the start-up and shutdown status of the generator set.

[0141] (29)

[0142] (30)

[0143] In equation (29-30), and Let represent the discharge power and charging power of unit i at time t, respectively; and The [0,1] variables represent the charging and discharging states, respectively. A value of 1 indicates discharge. A value of 1 indicates charging; This represents the charging and discharging power of unit i.

[0144] (31)

[0145] In the formula, Let be the state of charge of the i-th energy storage unit at time t+1; Let be the state of charge of the i-th energy storage unit at time t; Let be the charging efficiency of the i-th energy storage unit at time t; Let be the discharge efficiency of the i-th energy storage unit at time t; and These represent the low state of charge (SOC) value of the i-th energy storage unit at time t and the high state of charge (SOC) value of the i-th energy storage unit at time t, respectively.

[0146] A Snow Ablation Optimization Algorithm (SAO) is employed to solve the aforementioned day-ahead optimal scheduling model for hydro-solar-storage complementary systems. First, a random swarm is initialized, and initial particle positions are calculated. Gaussian Brownian motion is used to model the diffusion characteristics of snow or meltwater vaporization and update particle positions, while the centroid of the elite set is combined to expand the global search range. A refined search is performed using the snow ablation model, and constraints are applied, retaining elite particles and updating the global optimal solution until the maximum number of iterations or the cost convergence threshold is reached.

[0147] The Snow Demolition Optimization (SAO) algorithm is inspired by the three state changes that snow undergoes within an ecosystem. Snow can transform into two states: liquid water (through melting) and gaseous vapor (through sublimation). The SAO algorithm comprises three main stages:

[0148] 1. Initialization stage: Set the boundary of the optimization parameters and randomly generate the initial particle swarm according to formula (32).

[0149] 2. Exploration Phase: This phase simulates the high diffusion characteristics exhibited when snow or meltwater transforms into steam. Gaussian Brownian motion is used to update the search position, thereby expanding the exploration range of the algorithm.

[0150] 3. Development Phase: This phase uses the snow melting model (day-to-day method) to focus on the current optimal solution for a refined search, aiming to obtain a better solution rather than continuing the highly diffuse behavior of the previous phase.

[0151] In the first phase, initialize the random group:

[0152] (32)

[0153] Where Z is the initial particle position; L is the lower limit of each parameter to be optimized; U is the upper limit of each parameter to be optimized; and θ is a random number in [0,1].

[0154] In the second stage, Gaussian Brownian motion is used to model the high diffusion characteristics exhibited during snow or meltwater vaporization. This mechanism updates the location of the search agent, effectively broadening the algorithm's global exploration capability. The location update formula is as follows:

[0155] (33)

[0156] The formula defines the following core variables: This represents the position vector of the m-th particle during the k-th iteration. It is a random vector generated by a Gaussian distribution, used to simulate Brownian motion. (Symbol) This represents a row-wise multiplication operation. θ(i) is a random scalar uniformly distributed in the interval [0,1]. G(k) refers to the current best solution; Elite(k) is a random individual in an elite set. Let be the centroid of the entire group. The corresponding mathematical formula is as follows:

[0157] (34)

[0158] (35)

[0159] In the formula, and The positions of the second and third best particles; The centroid position of the particles whose fitness values ​​rank in the top 50%; The total number of individuals in the group.

[0160] In the third stage, development is based on the current optimal solution, and the specific formula is as follows:

[0161] (36)

[0162] In the formula, This represents a random number selected from [-1, 1]; M is the snowmelt rate, which is a day-by-day snowmelt model, and its expression is as follows:

[0163] (37)

[0164] Where k is the current iteration number; This represents the maximum number of iterations.

[0165] The specific implementation process of the SAO optimization algorithm is as follows: First, initialize the particle swarm and randomly generate initial positions within the parameter boundary according to formula (32). Each particle represents a water-light-storage scheduling scheme. Second, conduct fitness evaluation and use the total cost function (15) of the day-ahead optimization scheduling model as the evaluation index. Third, enter the exploration stage, simulate the diffusion characteristics of snow vaporization, use formula (33) to update the particle positions through Gaussian Brownian motion, and combine the centroid formula (34)(35) of the elite set to expand the global search range. Fourth, enter the development stage, focus on the neighborhood of the current optimal solution, use the snow melting model formula (36)(37) to conduct a refined search, and control the local optimization intensity through the day-ahead method. Fifth, implement constraint processing and repair solutions that violate the conditions of water balance (17), water level constraint (20)(21), power limit (24)(26). Finally, retain the elite particles of each generation and update the global optimal solution until the maximum number of iterations or the cost convergence threshold is reached.

[0166] S3: Based on the Bayesian deep learning module, the robustness of the scheduling plan is evaluated, and a dynamic hierarchical risk warning mechanism is established in combination with the sensitivity analysis engine.

[0167] The Bayesian deep learning robustness evaluation module uses a Bayesian neural network trained with Monte Carlo Dropout. Its input is the scheduling plan parameters, and its output is the system risk score and uncertainty measure.

[0168] First, it receives the extreme scenario set generated by S1 and the day-ahead scheduling plan output by S2, and obtains real-time monitoring data: the actual operating power of hydropower / photovoltaic / energy storage, load demand, meteorological data, etc. in the current period.

[0169] Secondly, a robustness evaluation module for Bayesian deep learning is constructed, using a Bayesian neural network (BNN) as the prediction core, with the input layer consisting of scheduling plan parameters (hydropower output). Photovoltaic power output Energy storage charging and discharging The output layer is a system risk score. .

[0170] Uncertainty quantization is achieved through Monte Carlo Dropout training, with the objective function being:

[0171] min (38)

[0172] in, Let be the likelihood function. ; For KL regularization terms; This represents the set of weight parameters.

[0173] Forward propagation execution Then calculate the uncertainty measure:

[0174] Expected risk: (39)

[0175] Prediction variance: (40)

[0176] in, For the first The output risk score of the next forward propagation.

[0177] The sensitivity analysis engine comprises three parts: feature perturbation analysis, global Sobol exponent calculation, and adversarial robustness testing. The sensitivity analysis engine performs multi-dimensional sensitivity detection through these three components:

[0178] The first part is the characteristic perturbation analysis, which consists of two parts:

[0179] 1. Gaussian noise injection: Construct perturbed samples by adding noise only to the d-th dimension feature. Other dimensions remain unchanged.

[0180] 2. Sensitivity index calculation: Use the difference approximation partial derivative to calculate the model prediction. The rate of change along the d-th dimension feature. The specific formula is as follows:

[0181] Gaussian noise injection: (41)

[0182] Sensitivity index calculation: (42)

[0183] in, Represents the original input vector; This represents the perturbation input after adding noise in the d-th dimension; This indicates that the noise follows a pattern with a mean of 0 and a variance of . Gaussian distribution; Represents a noise scalar; Indicates the noise variance; Represents the unit direction vector; is the sensitivity index for the d-th dimension feature.

[0184] The second part involves performing global Sobol analysis to calculate the Sobol exponent of feature d:

[0185] First-order Sobol index: (43)

[0186] Total effect index: (44)

[0187] in, This represents the total variance of the expected value of risk; This represents the expectation of all features except d; Indicates to The variance.

[0188] The third part is the adversarial robustness test, which generates adversarial examples based on FGSM:

[0189] FGSM attack generation: (45)

[0190] Robustness rating: (46)

[0191] in, This indicates the manually set perturbation step size; MM represents the number of test samples. The robustness score represents the model's performance. This represents an adversarial example of the original test sample; Represents the loss function The gradient with respect to the input x; It is a symbolic function.

[0192] After obtaining the index data from the sensitivity analysis, a tiered early warning mechanism is established, based on a threshold system. Calibrate using historical accident data as a benchmark:

[0193] (47)

[0194] In the formula, This is the cumulative distribution function of the corresponding indicator; This represents the permissible threshold for extreme risk probability. This is the variance threshold; Sensitivity threshold; This is the risk threshold.

[0195] The dynamic risk tiered early warning mechanism sets three levels of warning thresholds based on expected risk, prediction variance, and sensitivity indicators: Level 1 low-risk warning prompts monitoring; Level 2 medium-risk warning prompts adjustments to scheduling plans; and Level 3 high-risk warning triggers the activation of emergency response plans. The warning thresholds are calibrated based on historical accident data and dynamically adjusted using a sliding window mechanism.

[0196] Logic for generating risk warnings:

[0197] Level 1 Warning (Low Risk): When At that time, a message will appear stating "Uncertainty has increased and monitoring is required."

[0198] Level 2 Warning (Medium Risk): When At that time, the message "Vulnerable features exposed 0, adjust scheduling" will be displayed.

[0199] Level 3 Warning (High Risk): When When the "extreme risk is triggered, activate the contingency plan" is activated.

[0200] Finally, dynamic threshold adjustment is performed, updating the threshold based on a sliding window, using the following formula:

[0201] (48)

[0202] Event severity function: (49)

[0203] In equations (48-49), This represents the current threshold of the i-th feature at time t; The smoothing factor is α (0 ≤ α ≤ 1). The value is the sliding window size; j is the event index. This represents the j-th event within the sliding window; This function represents the severity of an event and outputs a numerical value based on the event type.

[0204] S4: Based on the day-ahead scheduling plan and risk warning results, a fast correction model for new energy transmission scheduling considering extreme scenario risks is established, and the Snow Goose Optimization Algorithm (SGA) is used to solve it, finally outputting a robust scheduling plan.

[0205] First, the risk warning method in S3 is used to identify potential risks in the system, and the risk warning signal generated by it serves as the input to S4. Based on the day-ahead water-solar-storage coordinated scheduling plan formulated in S2, S4 proposes a rapid correction scheme for renewable energy scheduling and transmission under the extreme scenario constraints introduced in S1, in order to meet the requirements for safe transmission.

[0206] Secondly, the rapid correction model for new energy transmission scheduling aims to minimize the corrected total cost, and adds risk constraints on transmission channels, power allocation constraints, and corrected power balance constraints. The total cost includes a transmission risk penalty term based on a risk correction factor, which is determined by the risk warning level.

[0207] The new energy transmission dispatch correction method is based on the overall day-ahead dispatch model. To meet transmission demand, internal generator output plans need to be readjusted under extreme conditions. A tiered early warning strategy is used based on risk scores to set corresponding correction factors for each risk level and time period. By pre-setting correction factors for generator units, rapid correction of the day-ahead dispatch plan by region can be achieved.

[0208] Consider providing adjustable hydropower units, photovoltaic units, and energy storage units to minimize the cost of the revised day-ahead dispatch plan.

[0209] (50)

[0210] in, Indicates the risk cost coefficient; This represents the expected value of the system risk output by S3; This represents the complement of the risk correction factor; when the risk is high, Smaller, then The larger the penalty, the more likely people are to reduce delivery orders. This represents the power transmitted during time period t.

[0211] The revised model should satisfy the newly added outbound transmission channel risk constraints, outbound power allocation constraints, and the revised power balance constraints:

[0212] (51)

[0213] (52)

[0214] (53)

[0215] in, Indicates the maximum transmission capacity of the external transmission channel; , , Represents the water / light / storage external transmission allocation coefficient, and satisfies ; The risk correction factor for time period t is determined by the risk warning level S3, and its mathematical expression is:

[0216] (54)

[0217] Finally, a Snow Goose Optimization Algorithm (SGA) is used to accelerate the solution of the modified hydro-solar-storage day-ahead scheduling model, obtaining the hydro-solar-storage day-ahead scheduling plan. The initialization phase defines the population position and velocity matrices;

[0218] A heading angle function control algorithm is introduced to transition from the exploration phase to the development phase. In the exploration phase, a zigzag flight strategy is adopted, and the population is divided into elite, weak and intermediate groups according to fitness and updated respectively. In the development phase, a straight flight strategy is adopted, which combines Brownian random walk and collective guidance mechanism to escape local optima.

[0219] The snow goose optimization algorithm simulates the characteristic V-formation and straight-line formation processes during snow goose migration. In these behavioral patterns, the lead goose plays a crucial role in the flock, its position creating favorable aerodynamic conditions for the entire group, thus significantly reducing the drag coefficient of subsequent individuals and improving the flock's endurance efficiency. During the algorithm initialization phase, each individual (snow goose) in the population is considered a candidate solution. The position and motion state of each solution in the d-dimensional continuous solution space are mathematically represented by the position matrix and velocity matrix, respectively.

[0220] , (55)

[0221] Here, variable n represents the total number of snow geese in the population, while d corresponds to the number of variables associated with the optimization problem.

[0222] The algorithm uses the heading angle of the snow goose migration process to represent the transition of the snow goose population from the exploration phase to the development phase:

[0223] (56)

[0224] in, This represents the maximum number of iterations. As long as the iterations performed by function U satisfy ( If the iteration count exceeds a certain value, the SGA operation will continue. Once the iteration count is exceeded, the SGA algorithm will stop executing. This represents the transition parameter from exploration (herringbone) to development (linear).

[0225] When designing the function, the challenges associated with local optima must be considered. In this case, the snow goose adopts a two-stage flight strategy, starting with exploration in a V-formation and transitioning to straight-line formation development. The velocity update formula during the exploration phase is:

[0226] (57)

[0227] (58)

[0228] (59)

[0229] in, denoted as the next generation velocity; 'a' represents acceleration; 'c' represents the velocity decay weight, which first increases and then decreases with each iteration, simulating the process of population energy accumulating and then decaying. Indicates the current speed; Let represent the acceleration of the i-th snow goose; 1.29 kg / m³ is the air density constant.

[0230] V-shaped flight: The population was divided into three groups according to fitness, from best to worst: the top 20%, the bottom 20%, and the middle 60%. The following methods were used for each group:

[0231] Elite Group: (60)

[0232] Weak group: (61)

[0233] the remaining: (62)

[0234] (63)

[0235] (64)

[0236] (65)

[0237] (66)

[0238] Where b, d, and e are random weights; Indicates the fitness-weighted population center location; This represents the fitness value of the i-th individual.

[0239] Different types of snow geese are updated using different formulas to modify their positions. Each component requires a weighting coefficient to represent its contribution to the overall formula; these weighting coefficients are established using b, d, and e. Here, variables b, d, and e correspond to empirical values ​​obtained from experiments. This indicates the position corresponding to the candidate solution, which represents the position of the lowest-ranking snow goose after group classification.

[0240] Straight-line flight: When the angle between snow geese exceeds 1 / 2, they enter the second stage of flight. In this stage, the algorithm prioritizes escaping local optima rather than precise navigation. Snow geese adopt a straight-line flight pattern:

[0241] (67)

[0242] In the formula, 1. Indicates element-wise multiplication; Brownian(d) represents the d-dimensional Brownian random walk directional quantity; r is a random number uniformly distributed in [0,1]. ① Collective guidance: If the random number r>0.5, the snow goose follows experienced and physically strong companions to collectively find the best destination. ② Random behavior: When r≤0.5, if trapped in a local solution, the snow goose exhibits random behavior similar to Brownian motion.

[0243] The specific solution process is as follows: First, randomly generate the position matrix (candidate solution) and velocity matrix of the snow goose population, and calculate the fitness value; Second, dynamically adjust the strategy according to the angle parameter, divide the population into three groups, and update the position according to the differentiated formula: the elite group is searched through global attraction and velocity guidance, the weakest group introduces the negative perturbation of the center position, and the middle group superimposes the escape term of the worst position, while updating the velocity; Third, update the position using a random guidance or Brownian motion strategy. If the random number r>0.5, follow the elite individual; otherwise, introduce Brownian perturbation; After each iteration, recalculate the fitness and update the global best and worst positions. The termination condition is reaching the maximum iteration or convergence threshold.

[0244] In summary, this invention proposes a stochastic scheduling method for hydro-solar-storage-energy storage bases to cope with extreme scenarios. By using Copula-GAN and MCMC for scenario generation and reduction, SAO algorithm for efficient scheduling solution, Bayesian deep learning-driven dynamic risk warning, and SGA-optimized fast transmission correction, the safe and stable operation of the power system under extreme scenarios can be achieved.

[0245] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0246] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0247] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0248] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0249] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. A new energy power generation base water-light storage day-ahead stochastic scheduling method for extreme scenarios, characterized in that, Comprise: S1: Based on Copula-GAN model for multi-scenario generation, and using MCMC method to reduce redundant scene, obtain typical extreme scene sample; S2: Establish a water-light-storage collaborative complementary day-ahead optimization scheduling model with the objective of minimizing the total operation cost, and solve it by using the snow ablation optimization algorithm SAO; S3: Based on Bayesian deep learning module for scheduling plan robustness evaluation, and combined with sensitivity analysis engine to establish dynamic hierarchical risk early warning mechanism; S4: Based on day-ahead scheduling plan and risk warning results, a new energy export scheduling quick correction model considering extreme scenario risk is established, and the snow goose optimization algorithm SGA is used to solve it, and finally the robust scheduling plan is output. 2.The new energy power generation base water-light storage day-ahead stochastic scheduling method for coping with extreme scenarios according to claim 1, characterized in that, The Copula-GAN model is used for multi-scenario generation, which comprises: Collecting historical time series data and actual data of hydropower, photovoltaic power and energy storage equipment; Fit the marginal distribution of each feature dimension, use kernel density estimation for continuous variables, and use empirical cumulative distribution function for discrete variables; Carry out Copula dependent structure modeling, construct a loss function containing adversarial loss and regularization term and carry out Copula-GAN training; Synthesize uniform vectors through the trained generator, and generate final scene samples through inverse transformation. 3.The new energy power generation base water-light storage day-ahead stochastic scheduling method for coping with extreme scenarios according to claim 1, characterized in that, The MCMC method is used to reduce redundant scenes, which comprises: Assign uniform probability weights to the initial large-scale scene set; Construct Markov chain, define state space and transition probability, and generate scenes by Gaussian disturbance; Define scene distance function to measure similarity and acceptance rate function to control acceptance probability of similar scenes; Update scene weights by iterative sampling, and finally select the top N representative scenes and normalize the weights.

4. The new energy power generation base water-light-storage day-ahead stochastic scheduling method for extreme scenarios according to claim 1, characterized in that: The objective function of the water-light-storage collaborative complementary day-ahead optimization scheduling model is to minimize the total cost, including hydropower operation cost, standby cost and light rejection cost; The constraints that the model needs to meet include: system power balance constraint, water balance constraint, reservoir water level constraint, reservoir discharge capacity constraint, hydropower station output limit, unit head constraint, photovoltaic safe operation interval constraint, light rejection equality constraint and energy storage system operation constraint.

5. The new energy power generation base water-light storage day-ahead stochastic scheduling method for extreme scenarios according to claim 1, characterized in that, The snow ablation optimization algorithm SAO is used for solving, which specifically comprises: Initialize random group and calculate initial particle position; Model the diffusion characteristics of snow or water vaporization by Gaussian Brown motion and update particle position, and expand the global search range by combining elite set centroid; Use the snow melting model for fine search and implement constraint processing, retain elite particles and update global optimal solution until the maximum iteration number or cost convergence threshold is reached.

6. The new energy power generation base water-light-storage day-ahead stochastic scheduling method for extreme scenarios according to claim 1, characterized in that: The Bayesian deep learning robustness evaluation module uses a Bayesian neural network trained with Monte Carlo Dropout, which takes scheduling plan parameters as input and outputs system risk score and uncertainty measure; The sensitivity analysis engine includes three parts: feature perturbation analysis, global Sobol index calculation, and adversarial robustness testing.

7. The new energy power generation base water-light storage day-ahead stochastic scheduling method for coping with extreme scenarios according to claim 1, characterized in that: The dynamic grading risk early warning mechanism sets three early warning thresholds according to risk expectation value, prediction variance and sensitivity index: first-level low-risk early warning prompts monitoring, second-level medium-risk early warning prompts adjustment of scheduling plan, and third-level high-risk early warning triggers emergency plan to start; The early warning threshold is calibrated based on historical accident data and dynamically adjusted through a sliding window mechanism.

8. The new energy power generation base water-light storage day-ahead stochastic scheduling method for coping with extreme scenarios according to claim 1, characterized in that: The new energy sending scheduling quick correction model takes the minimization of the corrected total cost as the goal, adds risk constraints of sending channels, sending power allocation constraints and corrected power balance constraints; The total cost includes a sending risk penalty term based on a risk correction factor, which is determined by the risk warning level. 9.The new energy power generation base water-light storage day-ahead stochastic scheduling method for coping with extreme scenarios of claim 1, wherein, The snow goose optimization algorithm SGA is used to solve the correction model, specifically including: The initialization stage defines the population position and speed matrix; The heading angle function control algorithm is introduced to control the transition from the exploration stage to the development stage; The exploration stage adopts a herringbone flight strategy, divides the population into elite group, weak group and intermediate group according to fitness, and updates them respectively; The development stage adopts a straight-line flight strategy, combined with Brownian random walk and collective guidance mechanism to escape from local optimum.

Images