Method for constructing robust scheduling model of water-wind-solar complementary system

Abstract

The invention provides a method for constructing a robust scheduling model of a water-wind-light complementary system. Meteorological data, equipment state data, power grid load data and historical operation data are fused to construct a multi-source heterogeneous data set; training a generative artificial intelligence model based on the data set, and generating a climate-energy-load-equipment state composite extreme scene conforming to a physical law; constructing an uncertainty set containing the composite extreme scene, and establishing a two-stage robust optimization framework embedded with physical constraints; and carrying out iterative solution on the robust optimization framework by adopting a column and constraint generation algorithm to generate a scheduling scheme considering both economy and safety. The method can break through the dependence of a traditional method on historical data, generates a multi-dimensional composite extreme scene according with a physical rule, remarkably improves the scheduling robustness and reliability of a water-wind-solar complementary system under an extreme working condition, and achieves the dynamic balance of economy and safety.

Description

Technical Field

[0001] This invention belongs to the technical field of hydro-wind-solar hybrid systems, specifically relating to a method for constructing a robust scheduling model for hydro-wind-solar hybrid systems. Background Technology

[0002] Currently, with the continuous increase in the proportion of renewable energy, the research on the optimal scheduling of integrated control centers for hydropower, wind power, and solar power has shifted from traditional single-energy independent optimization to multi-energy synergy and complementarity across the entire system, resulting in an exponential increase in scheduling complexity. However, existing scheduling technologies face significant theoretical and methodological bottlenecks in dealing with increasingly frequent extreme operating conditions, severely restricting the resilience and reliability of system operation. Existing mainstream scheduling models, whether based on deterministic optimization algorithms or stochastic programming methods that consider uncertainty, still heavily rely on historical observation data and conventional probability distributions (such as normal distribution and Weiber distribution) for predicting renewable energy output and load demand. This data-driven paradigm is still applicable under normal conditions of stable climate and smooth system operation, but its inherent defects are fully exposed under extreme operating conditions.

[0003] First, these models severely underperform in considering low-probability, high-destructive extreme events such as "black swan" and "gray rhino" events. These events include, but are not limited to: extreme weather events (such as a once-in-a-century catastrophic flood impacting a hydropower station, prolonged windless weather causing a complete wind power outage, or extreme cold or heat waves leading to a sharp increase in load), sudden failures triggered by latent defects or aging of equipment (such as a main transformer explosion or turbine blade breakage), and critical control system failures caused by cyberattacks. Existing models either cannot learn effectively due to the scarcity of such samples in historical data, or they ignore the correlations and cascading effects between these events due to overly idealistic assumptions. As a result, optimal scheduling strategies that perform well under normal conditions may quickly fail under these extreme conditions, not only failing to achieve optimization goals but also potentially amplifying accidents and triggering systemic collapse due to erroneous scheduling instructions (such as over-reliance on hydropower during water shortages).

[0004] Secondly, although the academic community has developed methods such as robust optimization to handle uncertainty, existing robust optimization methods still fall short when dealing with the multi-dimensional and strongly coupled uncertainties of hydro-wind-solar hybrid systems. The fundamental limitation lies in the crudeness and rigidity of the uncertainty set construction methods. Most current methods rely on simple interval estimations (such as "wind and solar output is ±30% of the predicted value") or extrapolation based on limited historical extreme samples. This approach has two major drawbacks: First, it is difficult to accurately characterize the dynamic coupling relationship between the uncertainties in the three dimensions of climate, energy, and load. For example, extreme drought can simultaneously lead to increased photovoltaic output (due to increased sunshine), a sharp decrease in hydropower output (due to reduced water inflow), and changes in load (due to increased irrigation demand). This complex correlation cannot be reflected in simple interval models. Second, it lacks the "imagination" for unprecedented but physically possible new extreme operating conditions. For example, a new type of composite extreme event occurring in the context of climate change (such as a strong typhoon accompanied by continuous rain, causing wind power to be simultaneously constrained by extreme cut-off and excessively low output) exceeds the scope of historical data.

[0005] Existing technologies mainly fall into two categories: stochastic programming models based on scenario analysis and robust optimization models using traditional uncertainty sets. Scenario analysis generates a large number of discrete scenarios to approximate uncertainty, but its scenario library is severely limited by historical data or simple Monte Carlo simulations, failing to creatively generate physically realistic extreme scenarios that exceed historical experience, resulting in insufficient protection against "unknown unknowns." Traditional robust optimization models, while providing a feasible solution under the "worst-case" scenario, often have overly broad uncertainty sets or fail to accurately reflect the system's physical constraints under extreme conditions. The resulting scheduling schemes are typically overly conservative, sacrificing significant economic benefits for safety, making them unacceptable in practical applications. Alternatively, their uncertainty sets may fail to cover real extreme risks, resulting in insufficient protection. Therefore, existing technologies exhibit significant deficiencies in robustness, adaptability, and reliability in scheduling decisions when facing the complex, ever-changing, and increasingly extreme operating environments of the real world. An innovative method is urgently needed that can more intelligently and accurately simulate extreme conditions and generate efficient and robust scheduling strategies accordingly. Summary of the Invention

[0006] The present invention aims to at least partially solve one of the technical problems in the related art.

[0007] Therefore, the first objective of this invention is to propose a method for constructing a robust scheduling model for a hydro-wind-solar hybrid system.

[0008] The second objective of this invention is to propose a robust scheduling model construction device for a hydro-wind-solar hybrid system.

[0009] The third objective of this invention is to provide a computer device.

[0010] The fourth objective of this invention is to provide a non-transitory computer-readable storage medium.

[0011] To achieve the above objectives, a first aspect of the present invention proposes a method for constructing a robust scheduling model for a hydro-wind-solar hybrid system, comprising: S1: Acquire meteorological data, energy operation data, load data, and equipment status data, and construct a training dataset containing the coupling relationship between climate, energy, load, and equipment status based on the cleaning, alignment, and feature engineering of multi-source heterogeneous data; S2, a generative AI model is trained based on the training dataset. The generative AI model takes historical weather patterns, equipment operating conditions and load levels as input conditions, and generates multi-dimensional composite extreme operating scenarios with physical rationality through adversarial or denoising training. S3, construct the multi-dimensional composite extreme operating scenario into an extreme scenario uncertainty set, and combine the physical operating constraints of the water, wind and solar system, equipment aging model and power grid safety criteria to establish a two-stage robust optimization framework that considers multi-dimensional uncertainties; S4. The column and constraint generation algorithm is used to iteratively solve the two-stage robust optimization framework. The basic scheduling decision and the worst-case adjustment strategy are generated through the interaction of the main problem and sub-problems. Finally, a pre-decision scheduling scheme that takes into account both economy and robustness is output.

[0012] In one embodiment of the present invention, S1 includes: S11, perform spatiotemporal alignment processing on the meteorological data, and map the discrete meteorological observation point data to a unified spatiotemporal grid through an interpolation algorithm to eliminate the temporal misalignment caused by the difference in data acquisition time interval; S12, quantify the health status of the equipment status data and calculate the equipment health degradation rate using the equipment aging index.

[0013] In one embodiment of the present invention, S2 includes: S21 employs a conditional generative adversarial network architecture, taking historical weather patterns, equipment operating conditions, and load levels as input conditions. Through adversarial training between the discriminator and the generator, it generates a generator that satisfies... Extreme scenarios; S22 generates extreme scenarios through the denoising process of the diffusion model, gradually transforming the initial noise vector into a scenario that conforms to physical constraints.

[0014] In one embodiment of the present invention, S3 further includes: S31, embeds the equipment aging model into the uncertainty set construction, and dynamically adjusts the equipment output boundary constraints through the aging index; S32 introduces power grid safety criteria as scenario selection conditions, performs feasibility verification on the generated extreme scenarios, and ensures that all scenarios meet the power grid safety constraint set.

[0015] In one embodiment of the present invention, S4 includes: S41 employs an improved column and constraint generation algorithm to iteratively generate basic scheduling decisions through the main problem and to find the worst-case scenario based on sub-problems; S42 sets up a convergence acceleration mechanism, which extracts a representative subset of scenarios through scenario clustering in each iteration, reducing the size of the subproblems to reduce computational complexity.

[0016] To achieve the above objectives, a second aspect of the present invention provides a robust scheduling model construction device for a hydro-wind-solar hybrid system, comprising: The data acquisition and processing module acquires meteorological data, energy operation data, load data, and equipment status data, and constructs a training dataset containing the coupling relationship between climate, energy, load, and equipment status based on the cleaning, alignment, and feature engineering of multi-source heterogeneous data. The generative AI training module trains a generative AI model based on the training dataset. The generative AI model takes historical weather patterns, equipment operating conditions and load levels as input conditions, and generates multi-dimensional composite extreme operating scenarios with physical rationality through adversarial or denoising training. The framework module is used to construct and optimize the multi-dimensional composite extreme operating scenarios into a set of extreme scenario uncertainties. Combined with the physical operating constraints of the water, wind and solar systems, equipment aging models and power grid safety criteria, a two-stage robust optimization framework considering multi-dimensional uncertainties is established. The solution and strategy generation module uses a column and constraint generation algorithm to iteratively solve the two-stage robust optimization framework. It generates basic scheduling decisions and worst-case adjustment strategies through the interaction of the main problem and sub-problems, and finally outputs a pre-decision scheduling scheme that balances economy and robustness.

[0017] This invention discloses a method and apparatus for constructing a robust scheduling model for a hydro-wind-solar hybrid system. Addressing the poor adaptability of existing scheduling models to sudden operating conditions such as extreme weather and equipment aging, it constructs a robust optimization framework considering multi-dimensional uncertainties in climate, energy, and load. Furthermore, it innovatively introduces generative AI technology to simulate various extreme operating scenarios as input for model training, thereby significantly improving the robustness and reliability of the scheduling strategy under uncertain environments. This invention effectively enhances the ability of the hydro-wind-solar integrated control system to cope with emergencies.

[0018] To achieve the above objectives, a third aspect of this application provides a computer device, including a processor and a memory; wherein the processor reads executable program code stored in the memory to run a program corresponding to the executable program code, for implementing a robust scheduling model construction method for a hydro-wind-solar hybrid system as described in the first aspect embodiment.

[0019] To achieve the above objectives, the fourth aspect of this application proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a robust scheduling model construction method for a hydro-wind-solar hybrid system as described in the first aspect embodiment.

[0020] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a method for constructing a robust scheduling model for a hydro-wind-solar hybrid system according to an embodiment of the present invention; Figure 2 This is a data infrastructure construction diagram according to an embodiment of the present invention; Figure 3 This is an extreme scenario generation diagram according to an embodiment of the present invention; Figure 4 This is a system integration and application diagram according to an embodiment of the present invention; Figure 5 This is a structural diagram of a robust scheduling model construction device for a hydro-wind-solar hybrid system according to an embodiment of the present invention; Figure 6 It is a computer device according to an embodiment of the present invention. Detailed Implementation

[0022] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0023] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0024] The following describes, with reference to the accompanying drawings, a method and apparatus for constructing a robust scheduling model for a hydro-wind-solar hybrid system according to an embodiment of the present invention.

[0025] Example 1 Figure 1 This is a flowchart of a task-general visual model construction method based on self-supervised representation according to an embodiment of the present invention, such as... Figure 1 As shown, it includes: S1 acquires meteorological data, energy operation data, load data, and equipment status data, and constructs a training dataset containing the coupling relationship between climate, energy, load, and equipment status based on the cleaning, alignment, and feature engineering of multi-source heterogeneous data.

[0026] In some implementations, this step first acquires high-precision gridded climate data from meteorological departments, including key variables such as wind speed, irradiance, precipitation, and temperature. The spatial resolution is typically 1 km × 1 km, and the temporal resolution is 1 hour. The data format conforms to the WGS-84 coordinate standard and the ISO 8601 timestamp specification. Simultaneously, the system integrates historical operational data and real-time status information of the hydro-wind-solar power station cluster, covering unit output, equipment operating parameters, and fault records. Data acquisition frequency ranges from seconds to minutes, and data storage formats include time-series databases (such as InfluxDB) or structured relational databases (such as MySQL).

[0027] During the data cleaning phase, the system employs a combination of rule-based and statistical methods to remove outliers, missing values, and data points with inconsistent timestamps. For example, the 3σ principle is used for anomaly detection in wind speed data, and threshold filtering is used for irradiance data (values ​​exceeding 1200 W / m² are considered outliers). Data alignment is achieved by using timestamp matching and interpolation algorithms (such as linear interpolation and spline interpolation) to fuse multi-source data on a unified time axis. In terms of feature engineering, the system extracts temporal correlations, lag effects, and joint distribution characteristics among multidimensional variables, such as calculating the covariance matrix of wind speed and photovoltaic output, and the lag correlation coefficient of load and temperature, to enhance the model's ability to perceive complex coupling relationships.

[0028] Furthermore, S1 includes: S11, perform spatiotemporal alignment processing on the meteorological data, and map the discrete meteorological observation point data to a unified spatiotemporal grid through an interpolation algorithm to eliminate the temporal misalignment caused by the difference in data acquisition time interval.

[0029] In some implementations, this process typically employs interpolation algorithms, such as Kriging, Inverse Distance Weighted Interpolation (IDW), or Spline Interpolation, to spatially expand meteorological variables such as wind speed, irradiance, and temperature at discrete points, forming two-dimensional or three-dimensional grid data covering the entire area. In the time dimension, time series interpolation methods (such as linear interpolation, spline interpolation, or time window-based moving averages) are used to align data from different timestamps, ensuring that all meteorological variables are updated synchronously at a uniform time step (e.g., every 15 minutes or hour). During the interpolation process, Geographic Information System (GIS) data must also be incorporated to consider the impact of factors such as topography, altitude, and land cover on meteorological variables, thereby improving interpolation accuracy.

[0030] Specifically, spatial interpolation typically sets the grid resolution to 1. km or km, time step is minutes or The time interval is determined by the system's scheduling requirements, specifically within minutes. Interpolation error can be assessed using root mean square error (RMSE) or mean absolute error (MAE), with the goal of keeping the error within RMSE < 2% or MAE < 1.5% to ensure data reliability.

[0031] S12, Quantify the health status of equipment status data using equipment aging index. Calculate the equipment health degradation rate, where For the current time, This is the time since the last maintenance. This refers to the rated lifespan of the equipment.

[0032] In some implementations, this step first obtains the device's current runtime. First, the timestamp is the time from system startup or the start of the last scheduling cycle to the current moment; second, the time of the device's most recent maintenance is read. This point in time is typically recorded by the Equipment Management System (EMS) or Condition Monitoring System (CMMS); finally, it is based on the rated lifespan provided by the equipment manufacturer. Calculate its aging index This index reflects the proportion of equipment used relative to its design life during the current operating cycle. The closer the value is to 1, the closer the equipment is to its life limit and the worse its health condition.

[0033] Specifically, and It is usually quantified in hours or days. This varies depending on the type of equipment. For example, the rated lifespan of a water turbine might be 20,000 hours, while the rated lifespan of a main transformer might be 30 years. Aging Index The value range is [0,1]. In practical applications, a threshold is usually set. When this happens, the system will trigger the equipment health warning mechanism, indicating that maintenance may need to be arranged in advance or the scheduling strategy adjusted to avoid risks.

[0034] S2, a generative AI model is trained based on the training dataset. The generative AI model takes historical weather patterns, equipment operating conditions and load levels as input conditions, and generates multi-dimensional composite extreme operating scenarios with physical rationality through adversarial or denoising training.

[0035] In some implementations, generative AI models can employ advanced architectures such as Conditional Generative Adversarial Networks (cGANs) or Diffusion Models. The model's input includes multidimensional conditional variables, such as historical meteorological data (wind speed, irradiance, precipitation, temperature, etc.), equipment operating status (e.g., main transformer health index, turbine vibration frequency, photovoltaic module aging coefficient, etc.), and load levels (e.g., regional load forecasting, critical load demand change rate, etc.). These conditional variables, after feature engineering, are encoded into a unified vector form and input into the generator network to guide the generator in generating extreme operating scenarios that conform to physical laws.

[0036] In adversarial training, the generator and discriminator are progressively optimized through game theory. The generator aims to produce synthetic data that is statistically consistent with real-world extreme scenarios, while the discriminator is responsible for distinguishing between generated and real data. During training, the loss function typically minimizes the discriminator's ability to differentiate between generated and real data. ,in For discriminator, For generator, The input is noise. Through multiple rounds of iterative training, the model can capture the nonlinear coupling relationships between multidimensional variables, thereby generating complex extreme event scenarios.

[0037] Specifically, the network structures of the generator and discriminator are typically modeled using a multilayer perceptron (MLP) or convolutional neural network (CNN), with the number of hidden layer nodes set between 128 and 512 to accommodate input features of different dimensions. During training, the learning rate is generally set to... to The optimizer can be Adam or RMSProp to improve model convergence efficiency. Furthermore, to ensure the physical plausibility of the generated scenario, the model output needs to be filtered through a pre-defined physical constraint verification module, such as power grid safety operation boundaries, equipment maximum output limits, and load balancing constraints.

[0038] Furthermore, S2 includes: S21 employs a Conditional Generative Adversarial Network (CGAN) architecture to integrate historical weather patterns. Equipment operating conditions and load level As a conditional input, it is passed through the discriminator. With generator Adversarial training to generate satisfactory results In extreme scenarios, among which To eliminate feature scale differences, the spliced ​​features are subjected to layer normalization after the physical rationality constraint set is spliced ​​at the channel level.

[0039] In some implementations, the CGAN model uses multidimensional condition variables. The encoding is incorporated into the input layer of the generator and discriminator to enable conditional control over the generated scene. Generator Receive random noise vector With condition vector Output the extreme operating scenarios of the synthesis Discriminator Then receive the real scene With the generated scene And, combined with the same input conditions, determine its authenticity. This is done by minimizing the generator's loss function. loss function of maximizing the discriminator The model gradually approximates the distribution characteristics of the real scene during adversarial training.

[0040] Specifically, conditional input Typically, it is a standardized multidimensional vector, with its dimensions corresponding to meteorological variables (such as wind speed, irradiance, temperature, etc.), equipment status (such as aging index, failure probability, operating efficiency), and load characteristics (such as peak load, load fluctuation rate, time distribution). The generator's input noise... The standard normal distribution is usually adopted. Its dimensions The value is typically set to 100-200 to ensure the diversity and complexity of the generated scenes. The discriminator's output is a scalar probability value. , representing the probability that the input scenario is real data. During training, the Adam optimizer is typically used, with a learning rate set to . Furthermore, a gradient penalty term is set to enhance model stability.

[0041] S22 generates extreme scenarios through the denoising process of the diffusion model, and the initial noise vector is... Gradually transforming into a scenario that conforms to physical constraints. Each iteration satisfies ,in For time step The cumulative noise figure, This refers to the noise disturbance in the current step.

[0042] In some implementations, the diffusion model uses a series of time steps. The iterative denoising operation gradually approximates the target scene distribution, and each iteration satisfies the following formula:

[0043] in, Indicates time step The cumulative noise figure is used to control the degree of noise attenuation during the iteration process; The noise disturbance in the current step typically follows a standard normal distribution. This process simulates the evolutionary path from a completely noisy state to a physically meaningful scene state, thereby generating statistically reasonable and physically feasible extreme operating scenarios.

[0044] The diffusion model is trained based on the fusion of historical observation data and multi-source heterogeneous data, including meteorological data, equipment status, and load curves. The model progressively adds Gaussian noise to the real-world scene through a forward diffusion process, eventually reaching a state of pure noise. Subsequently, in the reverse process, the model learns how to gradually recover the original scene from the noise. When generating extreme scenarios, the model guides the denoising process with specific input conditions (such as extreme weather patterns and equipment aging status), thereby generating composite extreme event scenarios highly correlated with these conditions.

[0045] Specifically, the key parameters of the diffusion model include the number of time steps. Noise scheduling function Design (e.g., linear, cosine, or exponential decay form), noise perturbation The distribution characteristics of the model, as well as the hidden layer dimension and training loss function. In this invention, It is typically set to 1000 time steps to ensure the smoothness and diversity of the generated scenes; A cosine decay form is adopted to better simulate the gradual changes in physical processes.

[0046] S3. The multidimensional composite extreme operating scenarios are constructed into a set of extreme scenario uncertainties. Combined with the physical operating constraints of the water, wind and solar systems, equipment aging models and power grid safety criteria, a two-stage robust optimization framework considering multidimensional uncertainties is established.

[0047] In some implementations, this step first filters and clusters the multidimensional extreme scenarios output by the generative AI model. Specifically, the system uses the K-means clustering algorithm or the density-based DBSCAN algorithm to reduce the dimensionality of the generated scenarios and extract representative scenarios, ensuring that the uncertainty set can cover various combinations of extreme events while avoiding redundancy and excessive computational complexity. Subsequently, the filtered scenarios are constructed into a compact uncertainty set. Its form can be represented as a joint interval or convex set of multidimensional variables, such as the joint range of change of variables such as wind and solar power output, load demand, and equipment availability.

[0048] Specifically, the construction of this uncertainty set needs to meet the following key indicators: Coverage Index, Scenario Diversity Score, and Physical Feasibility Ratio. Scenario Coverage Index measures the set's ability to cover potential extreme events, typically assessed by its matching rate with historical extreme events. Scenario Diversity Score is quantified by calculating the Euclidean distance or cosine similarity between scenarios. Physical Feasibility Ratio is determined by the satisfaction rate of system operational constraints, such as the minimum / maximum output limits of hydropower stations, the cut-off wind speed threshold of wind farms, and the irradiance-output response curve of photovoltaic power stations.

[0049] Furthermore, S3 includes: S31, Equipment Aging Model Constructed by embedding an uncertainty set, through an aging index Dynamically adjust the equipment output boundary constraints, specifically in the form of ,in This represents the aging degradation coefficient.

[0050] In some implementations, the equipment aging model It is a predictive model built using machine learning or physical modeling methods, based on the equipment's historical operating data, condition monitoring information (such as vibration, temperature, insulation performance, etc.), and known failure modes. This model can output the equipment's performance at different operating times. aging index This reflects the degree of performance degradation. Aging Index Typically, it is a non-negative real number, and a larger value indicates more severe equipment aging. In this invention, the maximum output boundary of the equipment... Dynamically adjusted to ,in Rated power of the equipment This is the aging attenuation coefficient, used to adjust the intensity of the impact of aging on output capacity. This exponential function form can smoothly reflect the nonlinear suppression effect of equipment aging on output capacity, avoiding instability of the scheduling strategy due to sudden changes in aging.

[0051] Specifically, aging degradation coefficient The value needs to be calibrated based on the equipment type, aging mechanism, and operating environment. For example, for wind turbines, Can be set to As for water and electricity equipment, Can be set to This reflects their different aging rates. In addition, the aging index... The update frequency should match the scheduling cycle, typically once per hour or every 15 minutes, to ensure that the scheduling model can respond to changes in device status in real time.

[0052] S32, Introducing Power Grid Safety Standards As a scenario selection criterion, the feasibility of the generated extreme scenarios is verified to ensure that all scenarios meet the requirements. ,in This is the set of power grid security constraints.

[0053] In some implementations, this step involves passing power grid security guidelines. Mapped to a set of constraints and the system physical operation constraint set Intersection operations are performed to filter out extreme scenarios that meet power grid security requirements. Specifically, power grid security criteria... These include, but are not limited to, voltage stability constraints, frequency deviation limits, line power flow safety boundaries, and node power balance conditions. These criteria are typically modeled based on national standards such as the "Guidelines for Power System Security and Stability" (DL / T 755-2001) or the "Design Specification for Power System Dispatch Automation System" (GB / T 26865-2011). In this invention, these criteria are transformed into mathematical constraints and embedded into the feasibility verification process for scenario selection.

[0054] Specifically, the set of power grid security constraints Constraints typically involve multiple dimensions, such as voltage magnitude constraints. Frequency deviation constraint Line power flow constraints and node power balance constraints .in, Indicates the first The network loss of each node can be calculated based on a power flow calculation model. This is achieved by comparing these metrics with generated extreme scenarios. Perform matching verification to ensure all scenarios meet the requirements. This avoids generating extreme scenarios that are unschedulable or violate grid security.

[0055] S4. The column and constraint generation algorithm is used to iteratively solve the two-stage robust optimization framework. The basic scheduling decision and the worst-case adjustment strategy are generated through the interaction of the main problem and sub-problems. Finally, a pre-decision scheduling scheme that takes into account both economy and robustness is output.

[0056] In some implementations, the master problem is responsible for determining basic scheduling decisions in the first phase, such as unit start-up and shutdown status, hydropower station output plans, and energy storage charging and discharging strategies. This problem typically aims to minimize basic operating costs, with variables being adjustable decision variables within the scheduling cycle. Constraints include system power balance, equipment capacity limitations, and scheduling cycle continuity constraints. Subproblems, based on the decisions of the master problem, seek the worst-case scenario and its corresponding adjustment costs; that is, they solve for the scenario that maximizes system adjustment costs within the set of uncertainties. The solutions to the subproblems are added back to the master problem as new constraints, forming an iterative optimization process.

[0057] Specifically, the accuracy of solving the main problem is usually set to a relative error of less than [a certain value]. To ensure the stability of the scheduling scheme. In the sub-problem, the search range for extreme scenarios is defined by the uncertainty set output by the generative AI model, which includes the joint distribution of multi-dimensional variables such as wind and solar power output, load fluctuations, and equipment availability. The calculation of regulation costs needs to consider key parameters such as the start-up cost of standby units, the load reduction penalty coefficient, and the regulation efficiency of the energy storage system. For example, the start-up cost of standby units can be expressed as... The load reduction penalty factor is The adjustment cost function is ,in This refers to the number of times the standby unit has been started. This represents the load reduction amount.

[0058] Furthermore, S4 includes: S41 employs an improved Column and Constraint Generation (C&CG) algorithm to iteratively generate basic scheduling decisions through the master problem. and based on sub-problems Find the worst-case scenario, among which Let be the objective function. This is a collection of extreme scenarios.

[0059] In some implementations, the C&CG algorithm first solves for the current scheduling policy in the main problem. This strategy typically includes key decision variables such as day-ahead unit allocation, hydropower output planning, and energy storage dispatch. The objective function of the main problem is to minimize the basic operating cost, while constraints include system power balance, equipment operating boundaries, and grid security limitations. Subsequently, in the sub-problems, the current dispatch decision is addressed. In extreme scenario sets Searching for the objective function Maximizing the scenario This refers to the worst-case scenario. This subproblem is essentially a problem of maximizing adjustment costs. Its solution is used to generate new constraints, which are fed back into the main problem to correct the scheduling strategy, ensuring its feasibility and optimality in a wider range of extreme scenarios.

[0060] Specifically, the decision variables of the main problem This typically includes unit start-up and shutdown status, hydropower output allocation, and energy storage charging and discharging power, with dimensions depending on the system scale and scheduling time window. The objective function in the sub-problem... To adjust the cost function, its form can be expressed as follows: ,in For time period The amount of power regulation, For load regulation, , For the corresponding adjustment cost coefficient. Extreme scenario set. Generated by generative AI models, typically ranging from hundreds to thousands of scenarios to ensure coverage of typical extreme cases under multidimensional uncertainty.

[0061] S42, set up a convergence acceleration mechanism, which uses scene clustering in each iteration. Extract a representative subset of scenarios to reduce the size of the subproblems. This is to reduce computational complexity.

[0062] In some implementations, scene clustering A distance-based clustering algorithm (such as K-means or hierarchical clustering) is employed, using multi-dimensional feature vectors as input, including but not limited to key variables such as wind speed, irradiance, load demand, and equipment health status. Each scenario is represented as a vector. Its dimension corresponds to the key uncertainty factors in system operation. Clustering algorithms calculate the Euclidean distance or cosine similarity between scenes, grouping scenes with high similarity into one category, and selecting the central scene of each category as a representative scene. The size of the representative scene subset is strictly controlled within a certain range. That is, it does not exceed 30% of the original extreme scenario set, thereby significantly reducing the number of constraints and variable dimensions of subproblems.

[0063] Specifically, the number of clusters in scene clustering Typically, adjustments are made dynamically based on the size of the scenario library and the complexity of the system, for example, in At that time, it can be set This ensures that the size of the subproblems is kept within a reasonable range. Furthermore, the number of iterations and convergence threshold (such as the maximum number of iterations) of the clustering algorithm are also important considerations. Convergence error Parameters such as these also need to be optimized and configured according to actual computing resources and accuracy requirements.

[0064] The present invention discloses a method for constructing a robust scheduling model for a water-wind-solar complementary system, which can construct a unified latent space that simultaneously supports visual generation and perceptual understanding tasks, maintain the semantic structure of the self-supervised encoder and improve reconstruction quality, significantly reduce training and deployment costs, and enhance the transferability and scalability of the model.

[0065] Example 2 The following describes in detail, with reference to the accompanying drawings, a method for constructing a robust scheduling model for a hydro-wind-solar hybrid system according to an embodiment of the present invention.

[0066] This invention aims to provide a robust scheduling model construction method for hydro-wind-solar hybrid systems under extreme conditions, fundamentally improving the system's survivability and stable operation under extreme conditions. The implementation of this method is a systematic, multi-stage process. Its core innovation lies in the deep integration of cutting-edge generative artificial intelligence technology with robust optimization theory. The specific steps are as follows: First, deep fusion and preprocessing of multi-source heterogeneous data are performed. For example... Figure 2 As shown, the system aggregates high-precision gridded climate data (including wind speed, irradiance, precipitation, temperature, etc.) from meteorological departments, historical operating data and real-time status information of hydropower, wind power, and solar power stations, power grid load data, equipment health status monitoring data, and a known fault record database. By cleaning, aligning, and feature-engineering these multi-source heterogeneous data, a high-quality training dataset that comprehensively describes the coupling relationship between "climate-energy-load-equipment status" is constructed.

[0067] Then, a generative AI model for extreme scenarios is built and trained. This is the core breakthrough of this solution. Figure 3As shown, we employ advanced generative model architectures, such as conditional generative adversarial networks or diffusion models, using historical weather patterns, equipment operating conditions, and load levels as inputs for adversarial or denoising training. The core capability of this model is learning the complex nonlinear relationships and joint probability distributions among multidimensional uncertainties. After sufficient training, the engine can overcome the limitations of historical data and generate a large number of physically plausible but historically unprecedented or extremely rare extreme operating scenarios. These scenarios are no longer simple perturbations of single variables, but multidimensional composite extreme events. For example, simulating a prolonged drought causing hydropower reservoir water levels to drop to historical lows, accompanied by regional calm weather causing wind power output to approach zero for an extended period, compounded by a load surge caused by a heat wave; or simulating a critical main transformer suddenly failing and shutting down during peak electricity demand due to long-term aging, while photovoltaic power plants experience a sharp drop in output due to sandstorms, and other complex operating conditions.

[0068] Next, a two-stage robust optimization framework considering multi-dimensional uncertainties is constructed, such as... Figure 4 As shown. In the first stage, the system formulates day-ahead scheduling plans (such as unit combination, hydropower plans, etc.) based on conventional forecast data. In the second stage, a large number of extreme scenarios generated by generative AI are used to construct a compact and representative set of extreme scenario uncertainties. The objective function of this optimization model is usually designed to minimize "basic operating costs + worst-case adjustment costs for dealing with extreme scenarios". Its constraints ensure that, given the fixed decisions in the first stage, there are feasible adjustment strategies (such as activating standby units, adjusting output, shedding load, etc.) for each extreme scenario in the uncertainty set, thereby guaranteeing the safe operating boundary of the system.

[0069] Subsequently, an efficient algorithm is employed to solve the robust optimization model. Given the high complexity of the model and the scale of the uncertainty set, we use a column and constraint generation algorithm. C&CG is an iterative algorithm that solves the basic scheduling decision in the main problem, and in the subproblems, it seeks the worst-case scenario for the current decision and generates corresponding constraints. Through repeated iterations of the main and subproblems, it eventually converges to a pre-decision scheduling scheme that is both economical and robust to all considered extreme scenarios.

[0070] Finally, the complete model system is integrated and applied. The robust scheduling model constructed above is encapsulated as a standardized service and integrated into the digital twin simulation platform or actual energy management system of the hydro-wind-solar integrated control center. This system can run automatically on a regular basis (e.g., daily) or triggered by events (e.g., receiving an extreme weather warning), generating robust scheduling plans for the next 24 hours to several days, and can perform rolling optimization. When real extreme events occur, the system can automatically match the preset robust strategies or provide operators with decision support that has been "stress-tested" under extreme conditions, thereby significantly enhancing the ability of the hydro-wind-solar hybrid system to cope with a highly uncertain future, effectively avoiding system collapse, and maximizing the power supply security of critical loads and the stable operation of the system.

[0071] Example 3 To achieve the above embodiments, such as Figure 5 As shown, this embodiment also provides a robust scheduling model construction device 10 for a hydro-wind-solar hybrid system. The device 10 includes a data acquisition and processing module 100, a generative AI training module 200, an ensemble construction and optimization framework module 300, and a solution and strategy generation module 400.

[0072] The data acquisition and processing module 100 acquires meteorological data, energy operation data, load data, and equipment status data, and constructs a training dataset containing the coupling relationship between climate, energy, load, and equipment status based on the cleaning, alignment, and feature engineering of multi-source heterogeneous data. Generative AI training module 200 trains a generative AI model based on the training dataset. The generative AI model takes historical weather patterns, equipment operating conditions and load levels as input conditions, and generates multi-dimensional composite extreme operating scenarios with physical rationality through adversarial or denoising training. The assembly construction and optimization framework module 300 constructs the multi-dimensional composite extreme operating scenario into an extreme scenario uncertainty set, and combines the physical operating constraints of the water, wind and solar system, equipment aging model and power grid safety criteria to establish a two-stage robust optimization framework that considers multi-dimensional uncertainties. The solution and strategy generation module 400 uses a column and constraint generation algorithm to iteratively solve the two-stage robust optimization framework. It generates basic scheduling decisions and worst-case adjustment strategies through the interaction of the main problem and sub-problems, and finally outputs a pre-decision scheduling scheme that takes into account both economy and robustness.

[0073] Furthermore, the aforementioned data acquisition and processing module 100 is also used for: The meteorological data is subjected to spatiotemporal alignment processing. The discrete meteorological observation point data is mapped to a unified spatiotemporal grid through an interpolation algorithm to eliminate the temporal misalignment caused by the difference in data acquisition time intervals. The health status data of the equipment is quantified, and the health degradation rate of the equipment is calculated using the equipment aging index.

[0074] Furthermore, the aforementioned generative AI training module 200 is also used for: A conditional generative adversarial network (GAN) architecture is adopted, using historical weather patterns, equipment operating conditions, and load levels as input conditions. Through adversarial training between the discriminator and the generator, a generator is generated that satisfies... Extreme scenarios; Extreme scenarios are generated through the denoising process of the diffusion model, gradually transforming the initial noise vector into a scenario that conforms to physical constraints.

[0075] Furthermore, the aforementioned set construction and optimization framework module 300 is also used for: The equipment aging model is constructed by embedding an uncertainty set, and the equipment output boundary constraints are dynamically adjusted through the aging index. Power grid safety criteria are introduced as scenario selection conditions to verify the feasibility of generated extreme scenarios and ensure that all scenarios meet the power grid safety constraint set.

[0076] Furthermore, the aforementioned solution and strategy generation module 400 is also used for: An improved column and constraint generation algorithm is adopted to generate basic scheduling decisions through iterative generation of the main problem, and to find the worst-case scenario based on sub-problems; A convergence acceleration mechanism is set up to extract a representative subset of scenarios through scenario clustering in each iteration, thereby reducing the size of the subproblems and reducing computational complexity.

[0077] The invention provides a robust scheduling model construction device for a hydro-wind-solar hybrid system, which can break through the dependence of traditional methods on historical data, generate multi-dimensional composite extreme scenarios that conform to physical laws, significantly improve the scheduling robustness and reliability of the hydro-wind-solar hybrid system under extreme conditions, and achieve a dynamic balance between economy and safety.

[0078] To implement the methods of the above embodiments, the present invention also provides a computer device, such as... Figure 6 As shown, the computer device 600 includes a memory 601 and a processor 602; wherein, the processor 602 reads executable program code stored in the memory 601 to run a program corresponding to the executable program code, so as to implement the various steps of the method described above.

[0079] To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in the foregoing embodiments.

[0080] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0081] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for constructing a robust scheduling model for a hydro-wind-solar hybrid system, characterized in that, include: S1: Acquire meteorological data, energy operation data, load data, and equipment status data, and construct a training dataset containing the coupling relationship between climate, energy, load, and equipment status based on the cleaning, alignment, and feature engineering of multi-source heterogeneous data; S2, a generative AI model is trained based on the training dataset. The generative AI model takes historical weather patterns, equipment operating conditions and load levels as input conditions, and generates multi-dimensional composite extreme operating scenarios with physical rationality through adversarial or denoising training. S3, construct the multi-dimensional composite extreme operating scenario into an extreme scenario uncertainty set, and combine the physical operating constraints of the water, wind and solar system, equipment aging model and power grid safety criteria to establish a two-stage robust optimization framework that considers multi-dimensional uncertainties; S4. The column and constraint generation algorithm is used to iteratively solve the two-stage robust optimization framework. The basic scheduling decision and the worst-case adjustment strategy are generated through the interaction of the main problem and sub-problems. Finally, a pre-decision scheduling scheme that takes into account both economy and robustness is output.

2. The method as described in claim 1, characterized in that, S1 includes: S11, perform spatiotemporal alignment processing on the meteorological data, and map the discrete meteorological observation point data to a unified spatiotemporal grid through an interpolation algorithm to eliminate the temporal misalignment caused by the difference in data acquisition time interval; S12, quantify the health status of the equipment status data and calculate the equipment health degradation rate using the equipment aging index.

3. The method as described in claim 1, characterized in that, S2 includes: S21 adopts a conditional generative adversarial network architecture, taking historical weather patterns, equipment operating conditions and load levels as conditional inputs, and generating extreme scenarios that meet physical rationality constraints through adversarial training between the discriminator and the generator. S22 generates extreme scenarios through the denoising process of the diffusion model, gradually transforming the initial noise vector into a scenario that conforms to physical constraints.

4. The method as described in claim 1, characterized in that, The S3 further includes: S31, embeds the equipment aging model into the uncertainty set construction, and dynamically adjusts the equipment output boundary constraints through the aging index; S32 introduces power grid safety criteria as scenario selection conditions, performs feasibility verification on the generated extreme scenarios, and ensures that all scenarios meet the power grid safety constraint set.

5. The method as described in claim 1, characterized in that, The S4 includes: S41 employs an improved column and constraint generation algorithm to iteratively generate basic scheduling decisions through the main problem and to find the worst-case scenario based on sub-problems; S42 sets up a convergence acceleration mechanism, which extracts a representative subset of scenarios through scenario clustering in each iteration, reducing the size of the subproblems to reduce computational complexity.

6. A robust scheduling model construction device for a hydro-wind-solar hybrid system, characterized in that, include: The data acquisition and processing module acquires meteorological data, energy operation data, load data, and equipment status data, and constructs a training dataset containing the coupling relationship between climate, energy, load, and equipment status based on the cleaning, alignment, and feature engineering of multi-source heterogeneous data. The generative AI training module trains a generative AI model based on the training dataset. The generative AI model takes historical weather patterns, equipment operating conditions and load levels as input conditions, and generates multi-dimensional composite extreme operating scenarios with physical rationality through adversarial or denoising training. The framework module is used to construct and optimize the multi-dimensional composite extreme operating scenarios into a set of extreme scenario uncertainties. Combined with the physical operating constraints of the water, wind and solar systems, equipment aging models and power grid safety criteria, a two-stage robust optimization framework considering multi-dimensional uncertainties is established. The solution and strategy generation module uses a column and constraint generation algorithm to iteratively solve the two-stage robust optimization framework. It generates basic scheduling decisions and worst-case adjustment strategies through the interaction of the main problem and sub-problems, and finally outputs a pre-decision scheduling scheme that balances economy and robustness.

7. The apparatus as claimed in claim 6, characterized in that, The data acquisition and processing module is also used for: The meteorological data is subjected to spatiotemporal alignment processing. The discrete meteorological observation point data is mapped to a unified spatiotemporal grid through an interpolation algorithm to eliminate the temporal misalignment caused by the difference in data acquisition time intervals. The health status data of the equipment is quantified, and the health degradation rate of the equipment is calculated using the equipment aging index.

8. The apparatus as claimed in claim 6, characterized in that, The generative AI training module is also used for: A conditional generative adversarial network architecture is adopted, which takes historical weather patterns, equipment operating conditions and load levels as conditional inputs, and generates extreme scenarios that meet physical rationality constraints through adversarial training between the discriminator and the generator. Extreme scenarios are generated through the denoising process of the diffusion model, gradually transforming the initial noise vector into a scenario that conforms to physical constraints.

9. A computer device, characterized in that, Including processor and memory; The processor reads the executable program code stored in the memory to run the program corresponding to the executable program code, so as to implement the robust scheduling model construction method of the hydro-wind-solar hybrid system as described in any one of claims 1-5.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements a robust scheduling model construction method for a hydro-wind-solar hybrid system as described in any one of claims 1-5.

Images