A wind-solar energy storage device capacity management method and system using a particle swarm algorithm

Abstract

The application discloses a kind of wind and light energy storage equipment capacity management method and system using particle swarm algorithm, it is related to new energy power system planning and operating technical field;Its technical points are: the method includes multi-source heterogeneous data fusion and double population initialization, joint fitness evaluation and subordinate dispatch optimization driven by business resilience, target-oriented dynamic scenario set collaborative evolution, risk perception adaptive particle swarm optimization and mode switching and decision output and rolling check;Through the overall cooperation of double population collaborative evolution, business resilience quantification, risk perception adaptive optimization and dynamic scenario stress test, the technical problems of income risk assessment misalignment under extreme coupling conditions and particle swarm optimization unable to adaptively adjust are solved, the collaborative optimization of capacity configuration and dispatching strategy is realized, and the risk resistance ability of system full life cycle is improved.

Description

Technical Field

[0001] This invention relates to the field of new energy power system planning and operation technology, specifically to a method and system for capacity management of wind and solar energy storage equipment using particle swarm optimization algorithm. Background Technology

[0002] Capacity management of wind, solar and energy storage equipment is a complex decision-making process that scientifically decides on installed capacity, rated power and energy configuration in the planning and operation phases of a hybrid energy system composed of wind power, photovoltaic and energy storage. By optimizing capacity ratio and operation strategy, it achieves a balance between economy, reliability and risk controllability throughout the entire life cycle.

[0003] Existing capacity management technologies for wind, solar and energy storage equipment are mostly based on generating static scene sets using typical daily curves or fixed probability distributions. They combine traditional particle swarm optimization algorithms to perform one-time offline optimization with the goal of maximizing net revenue over the entire life cycle. Such methods typically treat wind and solar power output, electricity price fluctuations and battery degradation as independent variables, and use fixed algorithm inertia weights and constant economic penalty parameters to complete capacity configuration.

[0004] In the complex commercial operation scenario of high-frequency fluctuations in the electricity spot market and accelerated degradation of energy storage batteries, the aforementioned existing technologies still have some problems: First, the existing scenario set cannot capture the low-probability high-loss events of low wind and solar output, deep electricity price troughs, and nonlinear battery degradation coexisting. It lacks a convolutional quantization mechanism for the joint probability distribution, which makes the optimized capacity scheme face a commercial revenue drawdown far exceeding expectations under real extreme operating conditions. Second, the exploration convergence strategy and objective function weights of the traditional particle swarm optimization algorithm are preset and fixed. It cannot dynamically switch between economic priority and risk defense modes according to the market volatility index and the equipment degradation slope. As a result, the algorithm is prone to premature convergence of local fragile solutions in high-risk coupled scenarios, and the output capacity strategy cannot adapt to the time-varying characteristics of long-term market and equipment status. Summary of the Invention

[0005] To achieve the above objectives, the present invention provides the following technical solution: A capacity management method for wind and solar energy storage devices using particle swarm optimization includes the following steps: Multi-source heterogeneous data from wind, solar and energy storage systems are collected and standardized to obtain a standardized database. Based on the standardized database, capacity particle swarm optimization and scenario particle swarm optimization are initialized respectively. Capacity particle swarm optimization represents candidate capacity configuration schemes, and scenario particle swarm optimization represents joint uncertain time-series scenarios. Based on the scenario particle swarm optimization, the capacity particle swarm is subordinately scheduled and optimized. The downside risk of revenue is quantified by a multi-dimensional joint probability calculation method to obtain the business resilience index. The fitness of each capacity particle is calculated by integrating the business resilience index. Based on the fitness distribution on the scene particle swarm, the scene particle swarm is dynamically evolved. Combined with the real-time health status of the energy storage device, vulnerable scene fragments are identified and adversarial augmentation scenes are generated. The scene particle swarm is then filtered and updated. The fitness is recalculated based on the updated scenario particle swarm. The capacity particle swarm is iteratively optimized using a risk-aware particle swarm algorithm. The risk status is determined based on the power market volatility index and the energy storage device attenuation slope. The risk mode is switched according to the risk status to adjust the penalty weight and correct the fitness. The inertia weight is dynamically adjusted according to the risk dispersion of the scenario particle swarm and the power market volatility index to update the particle swarm algorithm and obtain the converged optimal capacity particle swarm. Cluster the Pareto non-dominated solution set in the optimal capacity particle swarm, output the capacity management strategy package, and periodically collect new data in actual operation to perform rolling verification and full life cycle adaptive update.

[0006] Furthermore, the multi-source heterogeneous data includes meteorological data, wind and solar power output data, load data, electricity market data, energy storage battery degradation parameter data, and real-time health status data of energy storage devices; The capacity particle adopts a two-layer encoding structure, with the outer layer encoding representing the capacity configuration decision variables and the inner layer encoding representing the scheduling strategy parameter set. The initialization of the scene particle swarm includes: fitting the marginal probability distribution of each variable in the standardized database using a kernel density estimation model, and using a conditional generative adversarial network to learn the joint distribution relationship between the variables, thereby generating a baseline scene sequence covering different seasonal characteristics and typical meteorological patterns.

[0007] Furthermore, the subordinate scheduling optimization process includes: with the goal of minimizing the total operating cost within the scheduling cycle, solving for the near-optimal scheduling strategy under a fixed capacity configuration, and obtaining technical and economic performance data including wind curtailment rate, solar curtailment rate, load shedding probability, energy storage cycle loss conversion cost, electricity market revenue, and ancillary service market revenue.

[0008] Furthermore, the process of calculating the fitness of each capacity particle by integrating the business resilience index includes: performing multi-dimensional joint probability calculation based on the energy storage battery health index decay distribution, electricity price volatility distribution, and ancillary service call probability distribution to obtain a joint probability distribution; calculating the expected downside risk value based on the joint probability distribution as the business resilience index; calculating the net present value of each operating mode; correcting the net present value by the business resilience index and the penalty weight coefficient; and finally weighting and summing according to the proportion of each mode's occurrence time to obtain the fitness.

[0009] Furthermore, the process of dynamically evolving the scene particle swarm includes: calculating the mean and variance of the fitness of all capacity particles under each scene particle, comparing it with global statistics to obtain the degree of scene deviation, marking time segments with deviation exceeding a preset deviation threshold as vulnerable scene segments; calculating the effective available capacity decay coefficient under the vulnerable scene segment based on the battery health index decay model, and adversarially amplifying the vulnerable scene segment according to the effective available capacity decay coefficient to generate amplified scene particles.

[0010] Furthermore, the process of screening and updating the scene particle swarm includes: calculating the distinguishing contribution score of each scene particle, which is determined based on the change in the ranking of the Pareto non-dominated solution set and the change in fitness before and after removing the scene particle; retaining and replicating scene particles with scores higher than the high screening threshold, eliminating scene particles with scores lower than the low screening threshold for multiple generations, and replacing them with the highest-scoring particle among the expanded scene particles; and setting a timeliness score and a lifespan counter for each scene particle, eliminating scene particles with timeliness scores that are consistently lower than the timeliness threshold or whose lifespan exceeds the effective period.

[0011] Furthermore, the process of updating the particle swarm optimization algorithm includes: calculating a comprehensive risk adjustment factor based on the risk dispersion of the scenario particle swarm and the power market volatility index, and dynamically adjusting the inertia weight according to a linear mapping; when the power market volatility index and the energy storage device attenuation slope both exceed the warning threshold, increasing the penalty weight to increase the penalty intensity of the business resilience index in the fitness calculation, thereby switching the optimization objective from the conventional economic trade-off mode to the extreme resilience defense mode; and transforming the degree of violation of the business model's entry condition boundary into a dynamic penalty gradient, embedding it into the speed update rule of the particle swarm optimization algorithm.

[0012] Furthermore, the process of outputting the capacity management strategy package includes: performing K-means clustering based on the expected net present value, commercial resilience index, and wind-solar-storage ratio characteristics, mapping them to high-risk high-return preferences, robust equilibrium preferences, and policy arbitrage preferences, and outputting a strategy package for each preference. The strategy package includes a wind-solar-storage ratio recommendation list, a multi-mode scheduling strategy table, a life cycle cash flow forecast curve, and capacity adjustment trigger threshold parameters. The rolling verification and full lifecycle adaptive update process includes: periodically collecting new data to incrementally update the standardized database, re-executing all steps, applying perturbation to initialize the capacity particle swarm based on the historical best solution, and outputting configuration adjustment instructions according to the fitness deviation.

[0013] A capacity management system for wind and solar energy storage devices using particle swarm optimization algorithm, comprising: The data preprocessing and dual-population initialization module collects multi-source heterogeneous data from the wind-solar-storage system and performs standardization processing to obtain a standardized database. Based on the standardized database, the capacity particle swarm and the scene particle swarm are initialized respectively. The fitness assessment and business resilience calculation module performs subordinate scheduling optimization of capacity particle swarm based on scene particle swarm, quantifies the downside risk of revenue through multi-dimensional joint probability calculation method to obtain business resilience index, and integrates business resilience index to calculate the fitness of each capacity particle. The dynamic scene evolution module identifies vulnerable scene fragments based on the fitness distribution on the scene particle swarm and the real-time health status of the energy storage device, generates adversarial augmented scenes, and filters and updates the scene particle swarm to obtain the updated scene particle swarm. The risk-aware particle swarm optimization module recalculates the fitness of each capacity particle based on the updated scenario particle swarm. It uses the risk-aware particle swarm algorithm to iteratively optimize the capacity particle swarm. It determines the risk status based on the power market volatility index and the energy storage device attenuation slope to switch the risk mode and adjust the penalty weight. At the same time, it dynamically adjusts the inertia weight based on the risk dispersion of the scenario particle swarm and the power market volatility index to obtain the converged optimal capacity particle swarm. The strategy output and rolling verification module clusters the Pareto non-dominated solution set in the optimal capacity particle swarm, outputs the capacity management strategy package, and periodically collects new data to perform rolling verification and full lifecycle adaptive updates during actual operation.

[0014] Furthermore, in the data preprocessing and dual-population initialization module, the capacity particles adopt a dual-layer encoding structure. The outer layer encoding includes wind power installed capacity, photovoltaic installed capacity, energy storage rated power, and energy storage rated energy. The inner layer encoding includes charging and discharging power and state-of-charge operating range parameters under different seasonal phase modes and business models. The initialization of the scene particle swarm uses a kernel density estimation model to fit the marginal probability distribution of each variable and uses a conditional generative adversarial network to learn the joint distribution relationship between variables, generating a baseline scene sequence covering different seasonal characteristics and typical meteorological patterns. In the fitness evaluation and business resilience calculation module, a quadratic programming solver is used for subordinate scheduling optimization to obtain wind curtailment rate, solar curtailment rate, load shedding probability, energy storage cycle loss conversion cost, electricity market revenue, and ancillary service market revenue. Based on the battery health index decay distribution, electricity price volatility distribution, and ancillary service call probability distribution, a multi-dimensional joint probability calculation is performed to obtain the business resilience index, and then the fitness is obtained by weighted summation according to the mode weight. In the dynamic scene evolution module, the calculation of each The fitness mean and variance of all capacity particles under each scenario particle are used to obtain the degree of scenario deviation. Vulnerable scenario segments are marked. The effective usable capacity decay coefficient is calculated based on the battery electrochemical decay model, and low decay coefficient segments are adversarially amplified. Then, high-value scenario particles are selected based on the differentiation contribution score, and low-contribution or outdated scenario particles are eliminated. In the risk perception particle swarm optimization module, the comprehensive risk adjustment factor is calculated by weighting the risk dispersion of the scenario particle swarm and the power market volatility index, and the inertia weight is obtained by linear mapping. When the power market volatility index and the energy storage device decay slope exceed the warning threshold at the same time, the penalty weight is increased to switch to the extreme resilience defense mode. The degree of violation of the business model access condition boundary is transformed into a dynamic penalty gradient embedding speed update rule. In the strategy output and rolling verification module, K-means clustering is performed according to the net present value expectation, business resilience index and wind-solar-storage ratio characteristics to map into multiple risk preference strategy packages. New data is collected periodically to re-trigger the execution of all modules, and the capacity particle swarm is perturbed and initialized with the historical optimal solution as the center.

[0015] This invention provides a method and system for capacity management of wind and solar energy storage devices using particle swarm optimization, which has the following advantages: 1. This invention uses a three-dimensional convolution combining battery degradation distribution, electricity price volatility, and ancillary service call probability to accurately extract the low quantile of returns and generate a business resilience index, shifting capacity decision-making from static economic assessment to dynamic risk defense. Simultaneously, it maps inertial weights in real time based on scenario risk dispersion and market volatility characteristics, and automatically switches optimization modes and dynamically adjusts penalty weights based on dual-indicator thresholds. This allows the algorithm to strengthen global exploration and resilience priority during high-risk periods and accelerate local convergence during stable periods. It establishes a closed loop of uncertainty quantification, risk state perception, and algorithm parameter adaptation, ensuring that the output capacity strategy possesses both economic optimality and resilience against extreme operating conditions.

[0016] 2. This invention accurately captures vulnerable time segments that lead to performance degradation of capacity solutions by using normalized Euclidean distance, and uses the effective usable capacity decay coefficient calculated by the electrochemical decay model as a physical weight to drive the generative adversarial network to perform nonlinear weighted amplification, breaking through the statistical boundaries of historical data to generate high-value extreme stress test scenarios; further, it introduces a distinguishing contribution score to quantify the marginal value of each scenario in ranking non-dominated solutions, realizing the elite replication of high-value scenarios and the timely elimination of ineffective scenarios; it significantly improves the ability of the scenario set to target the weaknesses of capacity solutions, provides a challenging environment for continuous evolution in the optimization process, and effectively overcomes the overfitting defects caused by traditional fixed scenario sets.

[0017] 3. This invention constructs a rolling verification and business strategy clustering output system for closed-loop management throughout the entire life cycle. After the algorithm converges, it performs multi-dimensional clustering of the Pareto front based on the expected net present value, business resilience index, and wind-solar-storage ratio characteristics, directly mapping it to a capacity management strategy package with differentiated risk preferences. In the actual commissioning stage, it triggers dual-population reinitialization and incremental model training through periodic incremental measured data, applies perturbations around the historical optimal solution to quickly restart optimization, and automatically outputs configuration correction instructions based on the deviation of the resilience index. This realizes the leap from one-time static planning to data-driven dynamic evolution in capacity decision-making, significantly reducing the full life cycle operation and maintenance costs and the risk of commercial asset stranding, and meeting the long-term refined operation and management needs of wind, solar, and storage assets in a complex power market environment. Attached Figure Description

[0018] Figure 1 This is a flowchart of a method according to an embodiment of the present invention.

[0019] Figure 2 This is a flowchart illustrating the dynamic evolution of the particle swarm in the scenario described in this invention.

[0020] Figure 3 This is a flowchart of the risk-aware adaptive particle swarm optimization process of the present invention.

[0021] Figure 4 This is a system framework diagram of Embodiment 2 of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1: Please see Figures 1 to 3 This embodiment provides a capacity management method for wind and solar energy storage equipment using particle swarm optimization algorithm, including steps 1 to 5.

[0024] Step 1: Multi-source heterogeneous data fusion and dual-population initialization: Collect, fuse and standardize the basic data related to the wind and solar energy storage system in the benchmark area, and initialize the capacity particle swarm and scene particle swarm respectively to provide the input data and initial solution set required for subsequent optimization steps.

[0025] Step 1.1: Multi-source heterogeneous data acquisition and preprocessing.

[0026] For specific geographical areas where wind, solar, and energy storage systems will be deployed or are already in operation, i.e., target areas, comprehensive historical data and real-time monitoring data will be collected, including: Obtain multi-year historical meteorological data of the target area through meteorological monitoring equipment, including wind speed time series data, light intensity time series data and ambient temperature time series data; The wind power output time series data and photovoltaic power output time series data for the corresponding time period are obtained through power generation monitoring equipment; Load demand curve data is obtained through the power grid system.

[0027] Obtain energy storage battery degradation parameter data through the battery management system or manufacturer's specifications, including charge and discharge efficiency curve data of the battery under different health conditions, data on the relationship between battery polarization internal resistance and cycle number, data on the relationship between battery polarization internal resistance and ambient temperature, and battery calendar degradation coefficient data. Electricity market data is obtained through the electricity trading platform, including time-series data of electricity spot market prices, historical data of ancillary service market clearing, and carbon trading policy data. Among them, the time-series data of electricity spot market prices is the historical electricity price time-series data, and the historical data of ancillary service market clearing records the winning bid prices and call volumes of ancillary services in different periods in the past.

[0028] If the wind-solar-storage energy system is already in operation, real-time health status monitoring data of the energy storage equipment will be collected simultaneously, including battery health index, current ambient temperature, and recent charge-discharge cycle depth statistics.

[0029] Outlier removal and missing value imputation were performed on all the collected data. Outlier removal used the Isolation Forest algorithm, which constructs multiple binary trees by randomly selecting features and random split values. The path length of each data point in the tree is calculated. If the path length deviates significantly from the mean, such as more than 3 times the standard deviation, the sample is identified as an outlier and removed. Missing value imputation used the linear interpolation method.

[0030] Then, a standardization transformation is performed, using the minimum-maximum normalization method to uniformly transform all data into the range of 0 to 1, so as to eliminate the impact of dimensional differences on subsequent calculations; after processing, a standardized database is obtained.

[0031] Step 1.2: Capacity Particle Swarm Initialization: Construct the initial capacity particle swarm.

[0032] The capacity particle swarm represents a set of candidate capacity configuration schemes. Each capacity particle adopts a two-layer coding structure, as follows: The outer layer coding represents the decision variables for wind, solar and energy storage capacity configuration, which includes four variables: wind power installed capacity variable, in MW; solar power installed capacity variable, in MW; energy storage rated power variable, in MW; and energy storage rated energy variable, in MWh. The inner layer encoding represents the parameter set of scheduling strategies for different operating modes under a given outer layer capacity constraint. The operating modes include seasonal phase modes and business models. Seasonal phase modes include: spring mode, summer mode, autumn mode, winter mode, and extreme weather resilience phase mode. Business models include: independent energy storage peak-valley arbitrage mode, shared energy storage leasing mode, microgrid self-balancing mode, and frequency regulation ancillary service mode. The two sets of modes are combined, such as spring peak-valley arbitrage mode and summer frequency regulation ancillary service mode. The inner layer encoding contains the following parameters: energy storage charging and discharging power variables for each time period, in MW; state of charge operating range parameters, i.e., the lower limit and upper limit of the state of charge, dimensionless.

[0033] During initialization, the outer and inner codes of all particles are randomly generated within the preset constraint boundaries. The constraint boundaries of the outer codes are set according to the site resource conditions and equipment technical specifications of the target area, while the constraint boundaries of the inner codes are set according to the physical characteristics of the energy storage equipment and the power grid dispatching procedures. After processing, an initial capacity particle swarm is obtained, with each particle carrying complete decision information of the outer capacity code and the inner dispatching code.

[0034] Step 1.3: Scene Particle Swarm Initialization: Construct an initial scene particle swarm representing the uncertainty of the future operating environment.

[0035] A scene particle swarm represents a collection of jointly uncertain temporal scenes. Each scene particle contains a continuous temporal curve, including the following variables: wind speed time series data, in m / s; and light intensity time series data, in W / m². 2 Ambient temperature time series data, in °C; load time series data, in MW; electricity market price time series data, in yuan / MWh.

[0036] The initial scene particle swarm is generated based on a standardized database, as follows: First, the marginal probability distribution of each variable in the standardized database is fitted using the kernel density estimation model. The kernel density estimation model obtains a continuous probability density curve by placing a Gaussian kernel function on each data point and summing the results. Specifically, for a set of sample points of a certain variable, the kernel density estimate is equal to the sum of the kernel function values ​​corresponding to each sample point divided by the number of samples and then divided by the bandwidth parameter. Then, a conditional generative adversarial network (GAN) model is used to learn the complex nonlinear joint distribution relationship between variables such as wind and solar power output, load demand, ambient temperature, and electricity market price. The GAN consists of a generator and a discriminator. The generator receives a random noise vector and conditional variables and outputs the generated joint time series curve. The discriminator receives the generated curve and the real historical curve and outputs the probability that the curve is real data. The two are trained alternately. The generator tries to make the discriminator unable to distinguish between the generated data and the real data, and the discriminator tries to correctly distinguish between the two, until the generator can generate a joint time series curve that is indistinguishable from the statistical characteristics of the real historical data.

[0037] Using a trained conditional generative adversarial network model, a baseline scene sequence covering different seasonal characteristics and typical weather patterns is generated. Each generated scene is called a scene particle, and the collection of all scene particles constitutes the initial scene particle swarm.

[0038] Step 2: Business Resilience-Driven Joint Fitness Evaluation and Subordinate Scheduling Optimization: Calculate the overall fitness of each particle in the capacity swarm. The fitness value represents the overall merit of the capacity scheme and provides a basis for particle swarm updates.

[0039] Step 2.1: Subordinate scheduling strategy optimization: Find the near-optimal short-term scheduling strategy for each capacity particle under different operating modes.

[0040] For each capacity particle in the current generation of capacity particle swarm, first fix the capacity configuration scheme determined by its outer layer encoding, including wind power installed capacity, photovoltaic installed capacity, energy storage rated power and energy storage rated energy; for each operating mode in the inner layer encoding of the particle, such as the spring peak-valley arbitrage mode, use the embedded lightweight fast optimizer, such as a quadratic programming solver, to solve the near-optimal scheduling strategy under the capacity configuration and the operating mode.

[0041] The quadratic programming solver aims to minimize the total operating cost. The objective function includes the cost of wind and solar curtailment penalties, the cost of energy storage cycle losses, the negative value of electricity market revenue, and the negative value of ancillary service market revenue. The constraints include power balance constraints, dynamic constraints of energy storage state of charge, upper and lower limits of charging and discharging power constraints, and capacity constraints determined by the outer layer encoding.

[0042] After the optimization is completed, the optimal scheduling strategy and techno-economic performance data for the capacity particle under each operating mode are obtained. The techno-economic performance data includes wind curtailment rate, solar curtailment rate, load shedding probability, energy storage cycle loss conversion cost, electricity market revenue, and ancillary service market revenue: wind curtailment rate is the ratio of curtailed wind power to total generateable power, dimensionless; solar curtailment rate is the ratio of curtailed solar power to total generateable power, dimensionless; load shedding probability is the proportion of periods with insufficient power supply to the total number of periods, dimensionless; energy storage cycle loss conversion cost is in yuan; electricity market revenue is in yuan; and ancillary service market revenue is in yuan.

[0043] Step 2.2: Calculation of Business Resilience Index: The business resilience index is obtained by quantifying the downside risk of returns under extreme conditions through a multidimensional joint probability calculation method.

[0044] The Business Resilience Index is calculated based on a multidimensional joint probability calculation method, and the specific calculation process is as follows.

[0045] First, determine the probability distributions of the three inputs: The decay distribution of the effective usable capacity of the energy storage battery: Based on the energy storage battery decay parameter data collected in step 1.1, a battery health index decay model is established; the battery health index is the ratio of the current usable capacity of the battery to the nominal capacity, with a value range of 0 to 1; the battery health index decay model takes the number of cycles, ambient temperature and charge / discharge rate as input variables, and outputs the rate of decrease of the battery health index under this operating condition; based on the battery health index decay model, a large number of battery aging trajectory samples are generated by Monte Carlo simulation, and the kernel density estimation is fitted to the battery health index samples of each year to obtain the probability density function of the corresponding year; Distribution of electricity market price volatility: Based on the time series data of electricity spot market prices collected in step 1.1, the logarithmic return series of electricity prices is calculated, and then the standard deviation of the series within the sliding time window is calculated to obtain the electricity price volatility series; kernel density estimation is performed on the volatility series to obtain the probability density function of electricity price volatility; Distribution of ancillary service call probability: Based on the historical data of ancillary service market clearing collected in step 1.1, the frequency of ancillary services being called under various weather conditions is statistically analyzed to obtain the distribution of ancillary service call probability.

[0046] Assuming the three variables are independent, their joint effect can be characterized by multiplying their probability density functions. Using a multidimensional joint probability calculation method, a joint probability distribution of the system's net income over its entire lifecycle is constructed. Specifically, the possible ranges of the battery health index, electricity price volatility, and ancillary service call probability are each discretized into a finite number of continuous numerical intervals. All possible combinations of these three numerical intervals are iterated. For each combination, the available energy storage capacity is first adjusted based on the battery health index, and the electricity market revenue and ancillary service market revenue in step 2.1 are adjusted based on the electricity price volatility and ancillary service call probability to calculate the system's net income over its entire lifecycle. Then, the probability density values ​​of the battery health index, electricity price volatility, and ancillary service call probability in the corresponding intervals are multiplied, and then multiplied by the width of each interval on the original scale to obtain the probability mass of the system's net income under this combination. The probability masses of all combinations are summed to obtain the joint probability distribution of the system's net income over the entire value space.

[0047] In the joint probability distribution, the probability mass is accumulated in ascending order of returns. The return value corresponding to the first time the accumulated probability reaches 10% is found. If the accumulated probability just exceeds 10%, the corresponding return value is determined by linear interpolation between two adjacent return values. This value is the lower 10th percentile of the return distribution. The selection of this quantile is based on the Value at Risk (VaR) method commonly used in financial risk management. The return value represents that, at a 90% confidence level, the actual return of the system will not be lower than this value. Then, the expected return value under all possible scenarios is calculated, i.e., the weighted average return. The difference between the weighted average return value and the return value corresponding to the lower 10th percentile is the expected downside risk value. The expected downside risk value reflects the expected decline in return relative to the average level when the system faces the most unfavorable 10% extreme adverse scenario.

[0048] The Business Resilience Index is equal to the expected downside risk, expressed in yuan. The higher the index value, the greater the expected economic loss that the capacity plan may suffer under extremely adverse conditions throughout its life cycle, i.e., the weaker its ability to resist business risks.

[0049] Step 2.3: Calculate the joint fitness function: Combine operational performance and business resilience index to obtain the comprehensive fitness value.

[0050] First, the comprehensive technical and economic performance data of each operating mode output in step 2.1 are summarized. For each operating mode, the net present value is calculated. The net present value is equal to the sum of the electricity market revenue and the ancillary service market revenue over the entire life cycle of the mode, minus the investment cost, operation and maintenance cost and the discounted cost of energy storage cycle loss.

[0051] Secondly, a risk adjustment is made to the net present value (NPV). The adjusted NPV is equal to the NPV of the model minus the product of the business resilience index and the penalty weight coefficient. The penalty weight coefficient is initially set by the investor based on their risk appetite. Risk appetite reflects the investor's aversion to downside risk and ranges from 0 to 1. For example, when the investor is aggressive and seeks high returns, the risk appetite value is a lower value of 0.1; when the investor is conservative, the risk appetite value is a middle value of 0.5; and when the investor is extremely risk-averse, the risk appetite value is a higher value of 0.9. The penalty weight coefficient is equal to the risk appetite value.

[0052] Then, the weight coefficient of each operating mode is calculated. The weight coefficient of each mode is the proportion of the actual time of each operating mode in the whole year to the total time. By statistically analyzing the occurrence time of each mode in the meteorological data, load data and electricity price data of the target area over many years, the occurrence time of each mode is divided by the total time to obtain the weight coefficient of each mode. The weight coefficients of all modes are added together to equal 1.

[0053] Finally, the overall fitness value of each capacity particle is calculated. The overall fitness value is equal to the sum of all operating modes. The summation term is the corrected net present value multiplied by the corresponding mode weight coefficient. For simplicity, the overall fitness value will be referred to as fitness from now on.

[0054] Step 3: Goal-oriented dynamic scene set co-evolution: The scene particle swarm output in Step 1 is dynamically updated and evolved, enabling the scene set to effectively stress test the weaknesses of the current capacity scheme, and providing more challenging scene inputs for fitness evaluation in Step 4.

[0055] Step 3 and Step 2 are executed alternately and iteratively to drive the dynamic evolution of the scene particle swarm, enabling it to more effectively stress test the capacity scheme and eliminate invalid scenes.

[0056] Step 3.1: Vulnerable Scene Fragment Identification: Analyze the fitness distribution of the current capacity particle swarm on the scene set and identify the temporal fragments that cause significant degradation in the fitness of the capacity scheme.

[0057] Iterate through each scene particle in the current scene particle swarm output in step 1; for each scene particle, summarize the comprehensive fitness value of all capacity particles in that scene; for each scene particle, calculate two statistics: the mean and variance of the fitness of all capacity particles in that scene; then, compare these two statistics with the global mean and global variance of fitness in all scenes to calculate the degree of scene deviation; the degree of deviation is measured using normalized Euclidean distance, specifically by dividing the difference between the scene mean and the global mean, and the difference between the scene variance and the global variance by their respective global ranges for normalization, and then calculating the square root of the sum of squares.

[0058] The time segment corresponding to the scene particle whose deviation exceeds the preset deviation threshold is marked as a vulnerable scene segment. The deviation threshold is set according to the significant difference judgment standard, with a value range of 0.5 to 0.8. This scheme takes 0.7. When the deviation exceeds 0.7 times the standard deviation of the overall mean, the scene is considered to have significant vulnerability. The characteristics of vulnerable scene segments include: wind and solar power output being at a low level for several consecutive days and load demand being at a high level, or the electricity market price being at a low level plus the above conditions.

[0059] The processed data results in a collection of fragile scene fragments, each fragment containing a starting time index, length, and corresponding scene particle identifier.

[0060] Step 3.2: Adversarial amplification of physical information correction: Combine the battery health status to calculate the attenuation coefficient, and generate composite risk extreme scenarios by weighted amplification of vulnerable fragments.

[0061] First, for each vulnerable scenario segment identified in step 3.1, real-time health status monitoring data of the energy storage device is read simultaneously, including battery health index data, current ambient temperature data, and recent cycle depth statistics data.

[0062] Secondly, the effective available capacity decay coefficient is calculated using a pre-established battery health index decay model. The battery health index decay model is a physics-based mathematical model. Its inputs are ambient temperature time series data, estimated charge / discharge rate sequence data, estimated state of charge operating interval sequence data, cumulative cycle depth equivalent value, and current battery health index data in the vulnerable scenario segment. The output is the effective available capacity decay coefficient under the vulnerable scenario. The effective available capacity decay coefficient represents the future rate of change of the battery's available capacity under a given operating condition.

[0063] The estimated charge / discharge rate sequence is obtained by dividing the charge / discharge power in the optimal scheduling strategy for this scenario in step 2.1 by the rated energy storage power; the effective available capacity decay coefficient is equal to the ratio of the remaining schedulable energy to the nominal energy after the battery has been subjected to a given temperature sequence and charge / discharge rate sequence under the current battery health index. The effective available capacity decay coefficient ranges from 0 to 1, and the lower the value, the more severe the performance degradation of the battery under the corresponding operating conditions.

[0064] Then, the effective usable capacity decay coefficient is used as a physical weighting factor. For vulnerable scenario segments where the effective usable capacity decay coefficient is lower than the preset decay threshold, the probability of its adversarial amplification is set to a high probability value, such as 0.9. The decay threshold is based on the lithium battery health status grading standard and ranges from 0.3 to 0.5. This scheme uses 0.4. When the health index decays to below 0.4, the battery enters the rapid aging zone. The amplification amplitude is nonlinearly amplified according to the reciprocal of the effective usable capacity decay coefficient. The amplification amplitude includes two dimensions: the increase in fluctuation duration, in hours; and the increase in fluctuation amplitude, which is a dimensionless multiple. The larger the reciprocal of the decay coefficient, the larger the amplification amplitude.

[0065] The augmentation operation is implemented using a conditional generative adversarial network (GAN). The original vulnerable scene fragment is taken as input, and the generator outputs augmented new scene particles. During augmentation, the scene is constrained by statistical consistency; that is, the generated fluctuation patterns still conform to the marginal distribution characteristics described by kernel density estimation, but the fluctuation intensity and duration exceed the historical maximum values. The augmentation process iterates multiple times, generating a certain number of new scene particles.

[0066] After processing, a set of extreme augmented scene particles with physical weight factor labels are obtained.

[0067] Step 3.3: Pareto Front Differentiation Contribution Evaluation and Screening: Evaluate the value of each scenario particle in differentiating excellent candidate solutions, and screen scenarios accordingly.

[0068] First, obtain the Pareto non-dominated solution set of the particle swarm with the current capacity. A Pareto non-dominated solution is a solution in multi-objective optimization that is not inferior to it in all objectives and is superior to it in at least one objective. The Pareto non-dominated solution set is the set of all non-dominated solutions.

[0069] The process of obtaining Pareto non-dominated solutions is as follows: traverse all capacity particles in the current generation, and compare the superiority relationship between any two capacity particles one by one based on the comprehensive fitness value; if one capacity particle is not inferior to the other capacity particle in all optimization objectives, and is superior in at least one optimization objective, then the other capacity particle is determined to be a dominated particle and is removed; after the traversal is completed, all the capacity particles that have not been removed together form the Pareto non-dominated solution set of the current generation.

[0070] Secondly, the contribution of each scene particle in the scene particle swarm is calculated sequentially. The calculation steps are as follows: Given the presence of particles in the current scene, calculate the fitness value of each non-dominated solution and sort them from high to low fitness value to obtain sorting sequence A; The scene particle is temporarily removed from the scene set evaluation system. That is, the fitness value of each non-dominated solution is recalculated using only the remaining scene particles to obtain new fitness values ​​and reorder them to obtain sorted sequence B. The degree of difference between sorted sequences A and B is obtained by weighted summation of two parts: the first part is the proportion of non-dominated solutions whose sorting positions have changed to the total number of non-dominated solutions; the second part is the mean of the absolute changes in the fitness values ​​of all non-dominated solutions, normalized to a minimum and maximum value, where the minimum and maximum values ​​are the global maximum and minimum values ​​of the fitness values ​​of all non-dominated solutions in the current generation. The weighting coefficients of the two parts sum to 1, with the first part having a weighting coefficient of 0.6 and the second part having a weighting coefficient of 0.4, indicating that changes in sorting are more important than changes in numerical values. The degree of difference is the distinguishing contribution score of the particle in this scenario; the higher the distinguishing contribution score, the higher the value of the particle in distinguishing the relative merits of different non-dominated solutions.

[0071] Then, the scene particles are filtered based on their contribution scores: Based on the Pareto front sensitivity analysis of the scenarios, the screening thresholds were selected. The low screening threshold was 0.2, and the high screening threshold was 0.6. The high threshold of 0.6 can allow about 30% of high-value scenarios to enter the next generation, while the low threshold of 0.2 can eliminate long-term ineffective scenarios. For scene particles with a contribution score > 0.6, retain them and make a copy in the next generation of scene particle swarm; For scene particles whose contribution score is consistently <0.2 for multiple generations, they are not immediately eliminated. The system records the number of generations in which the scene particle is continuously below the elimination threshold. If the score of the scene particle is below the elimination threshold for three consecutive iterations, it is marked as a redundant scene particle and eliminated at the end of the third iteration. For scene particles with a differentiation contribution score of 0.2 or less and a difference contribution score of 0.6 or less, they are retained but not copied; that is, only the original number of scene particles are retained.

[0072] The eliminated scene particles are replaced by the particles with the highest contribution score predictions among the new scene particles generated in step 3.2. The new scene particles inherit the scene particle identifier of the eliminated particles but reset their lifecycle counters.

[0073] After processing, a set of filtered scene particles is obtained, with each particle accompanied by a distinguishing contribution score label.

[0074] Step 3.4: Scene Effectiveness Lifecycle Management: Set an effectiveness score and lifecycle counter for each scene particle, and eliminate scene particles that have not contributed for a long time or are outdated.

[0075] Assign two management parameters to each scene particle in the scene particle swarm: a timeliness rating parameter and a lifecycle counter parameter.

[0076] The timeliness score parameter is calculated as follows: The trend of the impact of particles on the performance of capacity schemes in the most recent three consecutive iterations is statistically analyzed; the performance impact is the absolute value of the difference between the fitness value of the best capacity particle and the average fitness value of all capacity particles in the given scenario; the change in this impact over the three consecutive iterations is calculated; the change is normalized to obtain the timeliness score parameter. The timeliness score ranges from 0 to 1, with a higher value indicating that the particles in the scenario can still effectively challenge and distinguish capacity schemes at the current stage.

[0077] The lifecycle counter parameter records the number of generations a particle in this scene has existed from its generation to the current generation. The lifecycle counter increments by one after each iteration.

[0078] The preset time threshold is set to 0.1, and the effective period is set to 50 generations. This setting is based on the following: In pre-conducted sensitivity simulation tests, effective periods of 20, 50, and 100 generations were set respectively, and the fitness convergence curves of the final capacity scheme were compared. Simulation results show that: when the effective period is set to 20 generations, scene particles are updated too quickly, and the capacity scheme is eliminated before it can adapt to new scenes, leading to increased fitness oscillations; when set to 100 generations, old scene particles remain for a long time, and their average distinguishing contribution score drops to below 18% of the initial value, slowing down the evolution speed by about 40%; when set to 50 generations, the fitness convergence curve is smooth and the computation time is optimal. After 50 generations, the average distinguishing contribution score of scene particles drops to below 20% of the initial value, which conforms to the empirical rule of the optimal number of iterations in evolutionary algorithms; therefore, an effective period of 50 generations is chosen.

[0079] Elimination criteria: If the timeliness score parameter of a scene particle is consistently lower than the timeliness threshold and remains below this value for three consecutive generations; or if the lifecycle counter parameter exceeds the effective cycle generation value; if either condition is met, the scene particle is marked as an outdated particle and eliminated.

[0080] Eliminated scene particles are replaced by the particle with the highest contribution score among the newly generated scene particles from step 3.2. The newly generated scene particles inherit the scene particle identifier of the eliminated particles but reset their lifecycle counters.

[0081] The resulting updated scene particle swarm maintains a constant size but its composition continues to evolve.

[0082] Step 4: Risk-aware adaptive particle swarm optimization and mode switching: Combine the fitness value calculated in Step 2 with the scene particle swarm updated in Step 3 to drive the capacity particle swarm to iteratively update its position and velocity according to the improved particle swarm algorithm, and dynamically adjust the search parameters and optimize the target weights according to the external risk status.

[0083] Step 4.1: Dynamic adjustment of inertia weight: Dynamically adjust the particle swarm inertia weight based on the risk dispersion of the particle swarm in the current scenario and the power market volatility index.

[0084] During the update of the position and velocity of the capacity particle swarm, the inertial weight is no longer set to a fixed constant.

[0085] First, calculate the risk dispersion of the scene particle swarm; the risk dispersion is the variance of the discrimination contribution score of each scene particle in step 3.3; the larger the variance, the greater the difference in the ability of each scene particle to distinguish the merits and demerits of the capacity scheme, which means that there is a large uncertainty in the current test environment.

[0086] Secondly, calculate the electricity market volatility index; based on the time series data of electricity spot market prices collected in step 1.1, select the electricity price data of the most recent 30 trading days, calculate the rolling standard deviation of its daily return, and then take the average of the rolling standard deviation as the electricity market volatility index; the higher the volatility index, the more volatile the market price.

[0087] Then, the risk dispersion value and the power market volatility index value are respectively subjected to minimum-maximum normalization so that both fall within the range of 0 to 1; the comprehensive risk adjustment factor is calculated by weighted summation; the weighting coefficient can be preset according to the system characteristics, for example, each is taken as 0.5.

[0088] The inertia weights are updated as follows: Suppose that the inertial weight has two boundary parameters: an initial value and a final value. The initial value is w. initial =0.9, final value w final =0.4. The selection of these two values ​​is based on the classic parameters of the particle swarm optimization algorithm. The initial value of 0.9 ensures the early global search capability, and the final value of 0.4 ensures the local convergence accuracy in the later stage. The current value of the inertia weight is updated at the beginning of each iteration based on the inertia weight value of the previous generation. The update rule is: the current inertia weight w of the kth generation. k =w final +(w initial -w final)×min(1, comprehensive risk adjustment factor×adjustment coefficient); where the adjustment coefficient is 0.5, which ensures that when the comprehensive risk adjustment factor is 1, the inertial weight w k It is still greater than the final value of 0.4, which is within a reasonable range.

[0089] This update method ensures that: when the comprehensive risk adjustment factor is large, i.e. in a high-risk environment, the inertial weight approaches the initial value of 0.9, enhancing the global exploration capability and expanding the capacity configuration search range; when the comprehensive risk adjustment factor is small, i.e. in a low-risk environment, the inertial weight approaches the final value of 0.4, accelerating local convergence and improving search accuracy; the weight is always constrained within the interval [0.4, 0.9] to ensure computational stability.

[0090] After processing, the adaptive inertia weight coefficients for each iteration are obtained.

[0091] Step 4.2: Business Model Soft Constraint Mapping: Transform the admission conditions of different business models into dynamic penalty gradients and embed the velocity update rules of particle swarm optimization.

[0092] For different business models, corresponding admission criteria boundary values ​​are set as follows.

[0093] Based on the power market grid connection and ancillary service assessment standards, the frequency regulation response time threshold is set at 2 seconds, the minimum lease utilization rate is set at 80%, and the tracking error threshold is set at 2%. For the independent energy storage peak-valley arbitrage model, it is required that the response time of the energy storage system's charging and discharging commands in the frequency regulation response of the electricity market shall not exceed the frequency regulation response threshold. Particle schemes whose response time exceeds the threshold will have their fitness deducted as a penalty value. For the shared energy storage leasing model, the annual utilization rate of the energy storage system must not be lower than the minimum leasing utilization rate. For particle schemes with a utilization rate lower than this minimum, the fitness will be penalized. For frequency regulation ancillary service mode, the power tracking error of the energy storage system is required to be no higher than the tracking error threshold. For particle schemes whose error exceeds the threshold, the fitness will be penalized.

[0094] The degree of violation of the aforementioned boundary is transformed into a dynamic penalty gradient. Specifically, when the particle's position vector (i.e., the capacity configuration scheme) deviates from a certain business model boundary after a velocity update, the ratio of the deviation to the boundary range is calculated. The deviation ratio is then squared and multiplied by a preset penalty coefficient to obtain a penalty value. This penalty value is subtracted from the overall fitness value calculated in step 2.3 to obtain the fitness after soft constraint correction, which is used for particle swarm iteration optimization. The penalty coefficient ranges from 0 to 10; this scheme uses a value of 5. This value is based on optimization sensitivity analysis, ensuring that a 10% deviation results in a fitness reduction of approximately 0.5, matching the magnitude of the business resilience index.

[0095] After processing, we obtain the particle fitness value and velocity update direction with constraint correction.

[0096] Step 4.3: Adaptive switching of risk mode: Real-time monitoring of the electricity market volatility index and the deterioration acceleration slope of energy storage equipment. When both exceed the warning threshold at the same time, the optimization objective will be switched from the conventional economic trade-off mode to the extreme resilience defense mode.

[0097] The system continuously monitors two key risk signals: the first is the electricity market volatility index, which is calculated in the same way as in step 4.1; the second is the rate of acceleration of energy storage device degradation, which is calculated as follows: based on the historical data of battery health index collected in step 1.1, the rate of decline of health index within the most recent quarterly time window is calculated, that is, the change in health index per unit time. If the change is negative, the larger the absolute value, the more obvious the device degradation acceleration; the change in health index is the amount of reduction in battery health index per unit time.

[0098] The logic for judging based on two indicators is as follows: When the electricity market volatility index exceeds a preset high volatility threshold, such as 0.7, and the absolute value of the accelerated degradation slope of energy storage devices also exceeds a preset high slope threshold, such as a 2% monthly decrease, the system is determined to have entered a high-risk period, automatically triggering a risk mode switch. At this time, the fitness penalty weight is switched from the conventional economic trade-off mode to the extreme resilience defense mode, significantly increasing the commercial resilience penalty weight. The high volatility threshold of 0.7 corresponds to a medium-to-high level of electricity price volatility, based on statistics of extreme market events. The 2% monthly slope threshold is based on experimental inflection point data of batteries entering the accelerated aging zone. This combined threshold setting is based on the review of historical extreme coupled events. When both indicators exceed the threshold simultaneously, the probability that the system is in a high-risk zone exceeds 80%.

[0099] Under the conventional economic trade-off model, the penalty weight coefficient of the business resilience index in step 2.3 is set to the default value, such as 0.2. Under the extreme resilience defense model, this penalty weight coefficient is significantly increased to between 3 and 5 times the default value. This solution uses 4 times, i.e., 0.8. The selection of this multiple is based on the following sensitivity analysis: backtesting was performed on historical extreme coupling events, testing the penalty weight multiples of 2, 3, 4, and 5 times respectively; at 2 times, the optimized solution failed to effectively avoid revenue drawdown in extreme scenarios, and the net present value decreased by 28% under extreme conditions. At a ratio of 4, the net present value decreases to 12%, while the economic loss under normal operating conditions is controlled within 5%. At a ratio of 5, the economic loss under normal operating conditions increases to 9%, and entering the resilience defense mode too early causes the convergence speed to decrease by about 30%. Therefore, a ratio of 4 is a better value that achieves a balance between defense effect and economy. After increasing the penalty weight, the algorithm will deduct the commercial resilience index value more significantly when calculating the overall fitness, thus giving priority to capacity schemes that can maintain basic power supply reliability and limited economy even under extreme combined operating conditions.

[0100] When the electricity market volatility index falls below the low volatility threshold, such as 0.4, or when the absolute value of the acceleration slope of energy storage device degradation falls below the low slope threshold, such as a 1% decrease per month, meaning that the two indicators are no longer in a high-risk state at the same time, the system will gradually revert the penalty weight coefficient back to the default value and restore the normal economic trade-off optimization mode; the reversion adopts a linear decreasing method, reducing the increment by 10% in each generation.

[0101] After processing, the optimization mode identifier and penalty weight coefficient corresponding to the current iteration are obtained.

[0102] Step 4.4: Particle Swarm Iteration and Convergence Judgment: Based on the updated parameters from Steps 4.1 to 4.3, perform velocity and position iteration calculations for the capacity particle swarm and determine the global convergence condition.

[0103] The standard velocity and position update formula of the particle swarm algorithm is used to iteratively update the capacity particle swarm. The velocity update of each particle is as follows: New generation velocity = inertia weight × current generation velocity + individual learning factor × first random factor × (the particle's historical best position - current position) + social learning factor × second random factor × (population global best position - current position).

[0104] The individual learning factor and the social learning factor are set by the user according to the classic design of the particle swarm optimization algorithm before optimization begins. They are usually each set to 2.0. The individual learning factor controls the tendency of particles to move towards their own historical best position, while the social learning factor controls the tendency of particles to move towards the global best position of the population. The first and second random factors are set according to the randomness of the algorithm and are randomly generated each time they are updated. They follow a uniform distribution in the interval [0, 1]. The use of a uniform distribution on [0, 1] indicates that the contribution weights of the individual's historical best position and the global best position to the particle's new velocity are random and unbiased each time they are updated.

[0105] The position update of each particle is obtained by accumulating the velocity update, that is: the position of the new generation is equal to the position of the current generation plus the velocity of the new generation.

[0106] During the iteration process, the soft constraints of the business model in step 4.2 indirectly affect the direction of velocity update by influencing the individual's historical best position and the global best position; the dynamic inertia weight in step 4.1 directly adjusts the velocity update; and the risk mode in step 4.3 changes the selection of the individual's historical best position and the global best position by influencing the fitness calculation.

[0107] The global convergence condition is set to satisfy both of the following conditions simultaneously: First, iterate continuously for multiple generations, with a preset generation value of 10 generations, and the absolute value of the fitness change of the optimal capacity particle is lower than the preset fitness change threshold of 0.001. Second, the composition of the scene particle swarm has not changed significantly in multiple consecutive iterations, that is, the net change rate of newly injected scene particles and eliminated scene particles in step 3.4 is less than the preset net change rate threshold of 5%. The net change rate is calculated as follows: the sum of the differences between the number of newly injected scene particles and the number of eliminated scene particles in each of the 10 consecutive generations is divided by the total size of the scene particle swarm, and then divided by 10. When this ratio is less than 5%, the scene set is considered stable.

[0108] If the convergence condition is not met, return to step 2 for the next iteration. If the convergence condition is met, terminate the iteration and proceed to step 5.

[0109] After processing, we obtain the converged optimal capacity particle swarm and scene particle swarm.

[0110] Step 5: Decision Output and Rolling Verification: Output the converged optimization results as an executable capacity management scheme, and establish a rolling update mechanism during the actual operation of the system to achieve closed-loop adaptive management of capacity decision-making throughout its entire lifecycle.

[0111] Step 5.1: Pareto optimal solution set clustering and strategy output: Cluster the converged non-dominated solution set according to business risk preferences and output a differentiated capacity management strategy package.

[0112] After iterative convergence, extract all capacity configuration schemes in the current Pareto nondominated solution set and their corresponding comprehensive fitness value, expected net present value, and business resilience index.

[0113] Multidimensional label clustering was performed based on the business risk appetite dimension, which includes three types: For those with a high-risk, high-return preference, prioritize options with high expected net present value but also high business resilience index. A robust and balanced approach is preferred, selecting options that strike a good balance between expected net present value and business resilience index; Policy arbitrage preferences prioritize strategies that leverage ancillary services markets and carbon trading policies to enhance returns.

[0114] The clustering uses the K-means clustering algorithm. First, a feature vector is extracted for each capacity scheme. The feature vector includes dimensions such as expected net present value, business resilience index, rated energy of energy storage, and the ratio of wind power installed capacity to photovoltaic installed capacity. Then, the feature vector is input into the K-means clustering algorithm, and the number of clusters is set to 3. The algorithm iteratively calculates the cluster centers and clusters the solution set into 3 categories. After the clustering is completed, the categories are mapped to the above 3 risk preference types according to the mean of the feature vector of each category.

[0115] For each risk appetite, a corresponding capacity management strategy package is output, including: A recommended list of wind, solar and energy storage ratios, which includes recommended ranges for wind power capacity, solar power capacity, rated energy storage power and rated energy storage capacity. A multi-mode scheduling strategy table, which provides recommended values ​​for charging and discharging power and state of charge operating range under different seasonal phases and business models; Lifecycle cash flow forecast curve, i.e., the expected value of net income each year and the lower bound of 5% confidence level and the upper bound of 95% confidence level; The capacity adjustment trigger threshold parameter is the specific value that suggests re-optimization when the business resilience index or equipment degradation slope rises to a certain critical value.

[0116] The processed data yields three strategy packages categorized by risk appetite, each of which can be deployed independently.

[0117] Step 5.2: Rolling verification and full lifecycle adaptive update: In actual operation, new data is collected periodically and optimization is re-executed to realize dynamic updates of capacity management strategies throughout their entire lifecycle.

[0118] The strategy package is deployed into the actual wind, solar and energy storage capacity management system, which then generates an equipment procurement list, grid connection application parameter files and commercial operation baseline plan files.

[0119] After the project is put into actual operation, the latest measured data on wind and solar power output, grid operation data, electricity market price data, and health status monitoring data of energy storage equipment, including battery health index and temperature data, are collected according to a preset collection cycle, such as quarterly. These new data are fed back to the standardized database in step 1.1 to incrementally update the original database. At the same time, the new data are used to incrementally train the conditional generative adversarial network model in step 1.3 to update the probability distribution parameters of scenario generation and to calibrate the parameters of the battery health index decay model in step 2.2.

[0120] After each data update, the entire process of dynamic scene evolution and capacity optimization from steps 1 to 4 is automatically triggered. When re-executing the capacity particle swarm initialization in step 1.2, random initialization is used for the first run. In subsequent rolling verifications, the optimal solution output in the previous round is used as the center of the initial seed for perturbation initialization, rather than being generated completely randomly, in order to speed up the convergence speed. The updated comprehensive fitness value is then recalculated.

[0121] The updated fitness value is compared with the baseline value output in step 5.1 to calculate the deviation. The deviation is calculated as follows: the difference between the updated business resilience index and the baseline business resilience index, divided by the baseline business resilience index. If the deviation exceeds the preset tolerance range, such as ±10%, or the energy storage device degradation slope exceeds the expected range, such as a monthly decrease of more than 3%, then an automatic capacity configuration adjustment suggestion or scheduling strategy correction instruction is output, and the operating parameters of the capacity management system are updated; if the deviation is within the tolerance range, the original scheme is maintained, and only the prediction model parameters are updated.

[0122] Rolling verification is carried out periodically to achieve closed-loop capacity management throughout the entire lifecycle; after the above processing, continuously updated capacity management strategies and operating parameters are obtained.

[0123] The working process of this method is as follows: First, multi-source heterogeneous data of wind, solar and energy storage systems are collected and standardized. Capacity particle swarm representing capacity configuration schemes and scenario particle swarm representing uncertain time series scenarios are initialized respectively. For each capacity particle, subordinate scheduling optimization is performed under fixed capacity. The business resilience index is obtained by fusing battery degradation, electricity price fluctuation and ancillary service call probability in three-dimensional convolution. It is then incorporated into the fitness calculation to accurately quantify the downside risk of revenue under extreme conditions. Subsequently, based on the fitness distribution of the current capacity particle swarm on the scene particle swarm, vulnerable scene segments that lead to performance degradation are identified, and adversarial amplification scenes are generated in combination with the battery electrochemical decay model. This drives the scene set to stress test the weaknesses of the capacity scheme. At the same time, high-value scene particles are selected based on the differentiation contribution, and redundant scenes are eliminated, solving the problem that the scene set cannot capture low-probability high-loss events. Based on this, a risk-aware particle swarm optimization algorithm is adopted to dynamically determine the risk status according to the power market volatility index and the attenuation slope of energy storage equipment. It automatically switches between conventional economic optimization and extreme resilience defense modes, and adaptively adjusts the inertia weight and penalty weight to balance global exploration and local convergence, avoiding premature convergence. Finally, the clustering of the converged Pareto non-dominated solution set outputs a capacity management strategy package with multiple risk preferences, and the rolling verification is triggered by periodically collecting new data to achieve adaptive updates throughout the entire lifecycle.

[0124] This method enables the optimized capacity scheme to maintain acceptable commercial returns under the combined extreme conditions of wind and solar power fluctuations, electricity price volatility, and battery aging, significantly improving the system's long-term resilience.

[0125] Example 2: Please see Figure 4 Based on Embodiment 1, this embodiment also provides a wind and solar energy storage capacity management system using particle swarm optimization algorithm, including: The data preprocessing and dual-population initialization module collects multi-source heterogeneous data from wind, solar and energy storage systems and performs standardization processing to obtain a standardized database. Based on the standardized database, the module initializes the capacity particle swarm and the scenario particle swarm respectively. The capacity particle swarm represents the candidate capacity configuration scheme, and the scenario particle swarm represents the joint uncertain time series scenario. The fitness assessment and business resilience calculation module receives the capacity particle swarm and scene particle swarm output by the data preprocessing and dual population initialization module. Based on the scene particle swarm, it performs subordinate scheduling optimization on the capacity particle swarm. It quantifies the downside risk of revenue through a multi-dimensional joint probability calculation method to obtain the business resilience index, and integrates the business resilience index to calculate the fitness of each capacity particle. The dynamic scenario evolution module receives the fitness and scenario particle swarm output by the fitness assessment and business resilience calculation module. Based on the fitness distribution on the scenario particle swarm, it identifies vulnerable scenario fragments in combination with the real-time health status of the energy storage device, generates adversarial augmentation scenarios, filters and updates the scenario particle swarm, obtains the updated scenario particle swarm, and sends it back to the fitness assessment and business resilience calculation module. The risk-aware particle swarm optimization module receives the fitness and updated scenario particle swarm output from the fitness assessment and business resilience calculation module. Based on the updated scenario particle swarm, it recalculates the fitness of each capacity particle and uses the risk-aware particle swarm algorithm to iteratively optimize the capacity particle swarm. It determines the risk status based on the power market volatility index and the energy storage device attenuation slope to switch the risk mode and adjust the penalty weight. At the same time, it dynamically adjusts the inertia weight based on the risk dispersion of the scenario particle swarm and the power market volatility index to obtain the converged optimal capacity particle swarm. The strategy output and rolling verification module receives the optimal capacity particle swarm output by the risk-aware particle swarm optimization module, clusters the Pareto non-dominated solution set in the optimal capacity particle swarm, outputs the capacity management strategy package, and periodically collects new data during actual operation and sends it back to the data preprocessing and dual-population initialization module to perform rolling verification and full life cycle adaptive update.

[0126] Furthermore, in the data preprocessing and dual-population initialization module, multi-source heterogeneous raw data is received from external data sources, such as meteorological monitoring equipment, power generation monitoring equipment, power grid systems, battery management systems, power trading platforms, and real-time health status monitoring equipment for energy storage devices. The capacity particles adopt a dual-layer coding structure. The outer layer coding includes wind power installed capacity, photovoltaic installed capacity, rated power of energy storage, and rated energy of energy storage. The inner layer coding includes charging and discharging power and state-of-charge operating range parameters under different seasonal phase modes and business models. The initialization of the scene particle swarm uses a kernel density estimation model to fit the marginal probability distribution of each variable and uses a conditional generative adversarial network to learn the joint distribution relationship between variables, generating a benchmark scene sequence covering different seasonal characteristics and typical meteorological patterns.

[0127] In the fitness assessment and business resilience calculation module, a quadratic programming solver is used to optimize subordinate scheduling to obtain the wind curtailment rate, solar curtailment rate, load shedding probability, energy storage cycle loss conversion cost, electricity market revenue, and ancillary service market revenue. Based on the battery health index decay distribution, electricity price volatility distribution, and ancillary service call probability distribution, a multi-dimensional joint probability calculation is performed to obtain the business resilience index. Finally, the fitness is obtained by weighting and summing according to the mode weights.

[0128] In the dynamic scenario evolution module, the current capacity particle swarm and its fitness value are received from the fitness assessment and commercial resilience calculation module, a standardized database for battery degradation model is received from the data preprocessing and dual-population initialization module, and real-time health status monitoring data of energy storage devices are received from the outside. The mean and variance of the fitness of all capacity particles under each scenario particle are calculated to obtain the degree of scenario deviation, vulnerable scenario fragments are marked, the effective usable capacity degradation coefficient is calculated based on the battery electrochemical degradation model, and low degradation coefficient fragments are adversarially amplified. Then, high-value scenario particles are selected based on the differentiation contribution score, and low-contribution or outdated scenario particles are eliminated.

[0129] In the risk-aware particle swarm optimization module, real-time risk indicators such as the power market volatility index and the energy storage device attenuation slope are received from the outside. The comprehensive risk adjustment factor is calculated based on the risk dispersion of the scenario particle swarm and the power market volatility index, and the inertia weight is dynamically adjusted. When the power market volatility index and the energy storage device attenuation slope both exceed the warning threshold, the penalty weight is increased to switch to the extreme resilience defense mode. The degree of violation of the business model's access condition boundary is transformed into a dynamic penalty gradient embedding speed update rule. The current generation capacity particle swarm is output to the fitness assessment and business resilience calculation module for fitness recalculation in the next iteration.

[0130] In the strategy output and rolling verification module, a standardized database is received from the data preprocessing and dual-population initialization module for incremental updates, and new data collected periodically during actual operation is received from the outside. K-means clustering is performed according to the expected net present value, commercial resilience index and wind-solar-storage ratio characteristics to map into multiple risk preference strategy packages, and new data is collected periodically to re-trigger the execution of all modules, and the capacity particle swarm is perturbed and initialized with the historical optimal solution as the center.

[0131] In the complex commercial operation scenario of high-frequency fluctuations in the electricity spot market and accelerated degradation of energy storage batteries, this system first collects and standardizes multi-source heterogeneous data through a data preprocessing and dual-population initialization module, generating capacity particle swarms and scenario particle swarms. Subsequently, the fitness assessment and business resilience calculation module performs subordinate scheduling optimization of capacity particles based on the scenario particle swarms, obtains the business resilience index and calculates fitness through multi-dimensional joint probability calculation, and solves the problem of missing quantification of low-probability high-loss events. The dynamic scenario evolution module identifies vulnerable scenario fragments based on fitness distribution, generates adversarial amplification scenarios in conjunction with the battery degradation model, and selects high-value scenario particles, enabling the scenario set to actively attack the weaknesses of the capacity scheme. The risk-aware particle swarm optimization module dynamically determines the risk status based on the electricity market volatility index and equipment degradation slope, automatically switches between economic priority and risk defense modes, and adjusts inertia weight and penalty weight to avoid premature convergence. Finally, the strategy output and rolling verification module clusters and outputs capacity management strategy packages with multiple risk preferences, and triggers rolling verification by periodically collecting new data to achieve adaptive updates throughout the entire life cycle. The capacity strategy ultimately output by the system can maintain commercial resilience under extreme conditions such as fluctuating electricity prices and aging batteries, and significantly reduce the risk of revenue drawdown.

[0132] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0133] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0134] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A capacity management method for wind and solar energy storage equipment using particle swarm optimization algorithm, characterized in that: Includes the following steps: Multi-source heterogeneous data from wind, solar and energy storage systems are collected and standardized to obtain a standardized database. Based on the standardized database, capacity particle swarm optimization and scenario particle swarm optimization are initialized respectively. Capacity particle swarm optimization represents candidate capacity configuration schemes, and scenario particle swarm optimization represents joint uncertain time-series scenarios. Based on the scenario particle swarm optimization, the capacity particle swarm is subordinately scheduled and optimized. The downside risk of revenue is quantified by a multi-dimensional joint probability calculation method to obtain the business resilience index. The fitness of each capacity particle is calculated by integrating the business resilience index. Based on the fitness distribution on the scene particle swarm, the scene particle swarm is dynamically evolved. Combined with the real-time health status of the energy storage device, vulnerable scene fragments are identified and adversarial augmentation scenes are generated. The scene particle swarm is then filtered and updated. The fitness is recalculated based on the updated scenario particle swarm. The capacity particle swarm is iteratively optimized using a risk-aware particle swarm algorithm. The risk status is determined based on the power market volatility index and the energy storage device attenuation slope. The risk mode is switched according to the risk status to adjust the penalty weight and correct the fitness. The inertia weight is dynamically adjusted according to the risk dispersion of the scenario particle swarm and the power market volatility index to update the particle swarm algorithm and obtain the converged optimal capacity particle swarm. Cluster the Pareto non-dominated solution set in the optimal capacity particle swarm, output the capacity management strategy package, and periodically collect new data in actual operation to perform rolling verification and full life cycle adaptive update.

2. The method for capacity management of wind and solar energy storage equipment using particle swarm optimization algorithm according to claim 1, characterized in that: The multi-source heterogeneous data includes meteorological data, wind and solar power output data, load data, electricity market data, energy storage battery degradation parameter data, and real-time health status data of energy storage devices. The capacity particle adopts a two-layer encoding structure, with the outer layer encoding representing the capacity configuration decision variables and the inner layer encoding representing the scheduling strategy parameter set. The initialization of the scene particle swarm includes: fitting the marginal probability distribution of each variable in the standardized database using a kernel density estimation model, and using a conditional generative adversarial network to learn the joint distribution relationship between the variables, thereby generating a baseline scene sequence covering different seasonal characteristics and typical meteorological patterns.

3. The method for capacity management of wind and solar energy storage equipment using particle swarm optimization algorithm according to claim 1, characterized in that: The subordinate scheduling optimization process includes: under a fixed capacity configuration, with the goal of minimizing total operating costs, solving for near-optimal scheduling strategies, and obtaining technical and economic performance data including wind curtailment rate, solar curtailment rate, load shedding probability, energy storage cycle loss conversion cost, electricity market revenue, and ancillary service market revenue.

4. The method for capacity management of wind and solar energy storage equipment using particle swarm optimization algorithm according to claim 1, characterized in that: The process of obtaining the fitness of each capacity particle includes: performing multi-dimensional joint probability calculation based on the energy storage battery health index decay distribution, electricity price volatility distribution and ancillary service call probability distribution to obtain a joint probability distribution, calculating the expected downside risk value based on the joint probability distribution as a business resilience index, calculating the net present value of each operating mode, correcting the net present value through the business resilience index and penalty weight coefficient, and then weighting and summing according to the proportion of each mode's occurrence time to obtain the fitness.

5. The method for capacity management of wind and solar energy storage equipment using particle swarm optimization algorithm according to claim 1, characterized in that: The process of dynamically evolving the scene particle swarm includes: calculating the mean and variance of the fitness of all capacity particles under each scene particle, comparing it with global statistics to obtain the degree of scene deviation, marking time segments with deviation exceeding a preset deviation threshold as vulnerable scene segments; calculating the effective available capacity decay coefficient under the vulnerable scene segment based on the battery health index decay model, and adversarially amplifying the vulnerable scene segment according to the effective available capacity decay coefficient to generate amplified scene particles.

6. The method for capacity management of wind and solar energy storage equipment using particle swarm optimization algorithm according to claim 1, characterized in that: The process of filtering and updating the scene particle swarm includes: calculating the distinguishing contribution score for each scene particle, which is determined based on the changes in the ranking of the Pareto non-dominated solution set and the change in fitness before and after removing the scene particle; retaining and replicating scene particles with scores higher than the high screening threshold, eliminating scene particles with scores lower than the low screening threshold for multiple generations, and replacing them with the highest-scoring particle from the expanded scene particles; and setting a timeliness score and a lifespan counter for each scene particle, eliminating scene particles with timeliness scores that are consistently lower than the timeliness threshold or whose lifespan exceeds the effective period.

7. The method for capacity management of wind and solar energy storage equipment using particle swarm optimization algorithm according to claim 1, characterized in that: The process of updating the particle swarm optimization algorithm includes: calculating a comprehensive risk adjustment factor based on the risk dispersion of the scenario particle swarm and the power market volatility index, and dynamically adjusting the inertia weight according to a linear mapping; when the power market volatility index and the energy storage device attenuation slope both exceed the warning threshold, adjusting the penalty weight to increase the penalty intensity of the business resilience index in the fitness calculation, switching the optimization objective from the conventional economic trade-off mode to the extreme resilience defense mode; and converting the degree of violation of the business model's entry condition boundary into a dynamic penalty gradient, embedding it into the speed update rule of the particle swarm optimization algorithm.

8. The method for capacity management of wind and solar energy storage equipment using particle swarm optimization algorithm according to claim 1, characterized in that: The process of outputting the capacity management strategy package includes: performing K-means clustering based on the expected net present value, commercial resilience index, and wind-solar-storage ratio characteristics, mapping them to high-risk-high-return preferences, robust equilibrium preferences, and policy arbitrage preferences, and outputting a strategy package for each preference. The strategy package includes a wind-solar-storage ratio recommendation list, a multi-mode scheduling strategy table, a life cycle cash flow forecast curve, and capacity adjustment trigger threshold parameters. The rolling verification and full lifecycle adaptive update process includes: periodically collecting new data to incrementally update the standardized database, re-executing all steps, applying perturbation to initialize the capacity particle swarm based on the historical best solution, and outputting configuration adjustment instructions according to the fitness deviation.

9. A capacity management system for wind and solar energy storage devices using particle swarm optimization algorithm, characterized in that: include: The data preprocessing and dual-population initialization module collects multi-source heterogeneous data from wind, solar and energy storage systems and performs standardization processing to obtain a standardized database, outputting initial capacity particle swarm and scene particle swarm. The fitness assessment and business resilience calculation module receives capacity particle swarm and scene particle swarm, performs subordinate scheduling optimization on the capacity particle swarm, quantifies the downside risk of revenue through a multi-dimensional joint probability calculation method to obtain the business resilience index, and calculates the fitness of each capacity particle by integrating the business resilience index. The dynamic scenario evolution module receives fitness and scenario particle swarm. Based on the fitness distribution on the scenario particle swarm, it identifies vulnerable scenario fragments in conjunction with the real-time health status of the energy storage device, generates adversarial augmentation scenarios, filters and updates the scenario particle swarm, obtains the updated scenario particle swarm, and sends it back to the fitness assessment and business resilience calculation module. The risk-aware particle swarm optimization module receives the updated scene particle swarm, recalculates the fitness of each capacity particle, and uses the risk-aware particle swarm algorithm to iteratively optimize the capacity particle swarm. It determines the risk status based on the power market volatility index and the energy storage device attenuation slope to switch the risk mode and adjust the penalty weight. At the same time, it dynamically adjusts the inertia weight based on the risk dispersion of the scene particle swarm and the power market volatility index to obtain the converged optimal capacity particle swarm. The strategy output and rolling verification module receives the optimal capacity particle swarm, performs Pareto non-dominated solution set clustering, outputs the capacity management strategy package, and periodically collects new data during actual operation and sends it back to the data preprocessing and dual-population initialization module to perform rolling verification and full lifecycle adaptive updates.

10. A capacity management system for wind and solar energy storage equipment using particle swarm optimization algorithm according to claim 9, characterized in that: In the data preprocessing and dual-population initialization module, the capacity particles adopt a dual-layer coding structure with outer-layer coding capacity configuration and inner-layer coding scheduling strategy. The scene particle swarm uses kernel density estimation to fit the edge distribution and combines it with a conditional generative adversarial network to generate a baseline scene sequence. In the fitness assessment and business resilience calculation module, a quadratic programming solver is used to optimize subordinate scheduling, and a multi-dimensional joint probability calculation is performed based on the battery health index decay distribution, electricity price volatility distribution and auxiliary service call probability distribution to obtain the business resilience index. The fitness is obtained by weighted summation according to the mode weight. In the dynamic scene evolution module, the degree of scene deviation is calculated by fitness mean and variance to mark vulnerable scene segments. The low decay coefficient segments are adversarially amplified using the battery electrochemical decay model. High-value scene particles are selected and low-contribution or outdated scene particles are eliminated based on the differentiation contribution score. In the risk-aware particle swarm optimization module, a comprehensive risk adjustment factor is calculated by weighting the risk dispersion and the power market volatility index to obtain the inertial weight through linear mapping. When the market volatility index and the decay slope exceed the threshold at the same time, the penalty weight is increased to switch to the extreme resilience defense mode. At the same time, the degree of violation of the business model boundary is transformed into a dynamic penalty gradient embedded in the speed update rule. In the strategy output and rolling verification module, K-means clustering is performed based on net present value expectation, commercial resilience index and wind-solar-storage ratio characteristics to map into multiple risk preference strategy packages. New data is collected periodically to re-trigger the execution of all modules, and the capacity particle swarm is perturbed and initialized with the historical optimal solution as the center.

Images