Hydropower plant dispatching management method and system considering network security constraints

Abstract

The present application relates to the technical field of hydropower plant management, and provides a hydropower plant scheduling management method and system considering network security constraints, which first automatically extracts a typical attack consequence scene set by unsupervised clustering of historical attack data, to provide data-driven basis for uncertainty modeling; to cope with the core problem that attack occurrence probability is difficult to accurately predict, the idea of distribution robust optimization is introduced, which does not rely on a single nominal probability distribution, but constructs a fuzzy set containing multiple probability distributions, and solves the unified scheduling plan with optimal expected cost under the worst probability distribution in the fuzzy set; the finally generated robust scheduling scheme can perform well under high probability conditions and ensure that the performance will not collapse under the most unfavorable attack conditions, thereby achieving a robust balance between economy and security under uncertain security threats.

Description

Technical Field

[0001] This invention relates to the field of hydropower plant management technology, and in particular to a hydropower plant dispatching and management method and system that takes into account network security constraints. Background Technology

[0002] In recent years, with the continuous penetration of industrial internet technology, the production operation and dispatch management of hydropower plants have fully shifted to a networked architecture centered on cyber-physical systems. Regional dispatch centers, leveraging SCADA systems and their supporting communication networks, implement remote centralized monitoring and automated dispatch control of hydropower plants within their jurisdiction, thereby achieving efficient allocation of hydropower resources and maximizing comprehensive power generation revenue. However, while this highly interconnected automated system significantly improves dispatch efficiency and management sophistication, it inevitably exposes the core production control systems of hydropower plants to increasingly severe cybersecurity risks. Malicious attackers can utilize various cyberattack techniques, such as fake data injection and denial-of-service attacks, to tamper with critical measurement information (e.g., real-time reservoir water levels, generator active power output), or disrupt the normal issuance of dispatch control commands and the timely transmission of plant operating status information. Such cyber intrusions directly interfere with or even mislead the dispatch optimization decision-making process, potentially leading to significant economic losses such as increased water wastage and reduced power generation, or even jeopardizing the equipment integrity of turbine generator units due to erroneous control commands, and even impacting the safe and stable operation of the regional power grid.

[0003] Currently, most existing research and engineering practices addressing the aforementioned security threats remain at the level of passive protection or deterministic security modeling. One typical approach is to deploy intrusion detection systems on the network side, triggering alarms through real-time monitoring of abnormal traffic or data patterns. However, this mechanism is essentially a reactive measure; by the time an anomaly is identified, scheduling decisions based on erroneous data have often already been generated or are in the process of being executed, making it difficult to provide proactive security assurance. Another approach attempts to draw upon the traditional N-1 security check concept of power systems, simplifying the potential impact of network attacks into fixed deterministic security margins or rigid boundary constraints. While this method is relatively simple to implement in engineering, it struggles to effectively address the highly diverse and dynamic characteristics of network attacks in terms of target selection, attack methods, and the severity of consequences. Because attackers' strategies and targets are highly unpredictable, their actual impact on physical systems is complex and varied; a single deterministic model cannot fully characterize this deep-seated uncertainty. This often leads to two adverse consequences: first, the scheduling scheme becomes overly conservative due to excessive security margins, resulting in unnecessary economic losses; second, when the actual attack patterns exceed the coverage of the preset model, the system still exhibits significant vulnerability. Summary of the Invention

[0004] The present invention aims to solve at least one of the problems existing in the prior art, and to provide a method and system for hydropower plant scheduling and management that takes into account network security constraints.

[0005] One aspect of the present invention provides a hydropower plant dispatching and management method considering network security constraints, comprising: Based on an expert knowledge base and historical attack data, a discrete scenario set is constructed, including: vectorizing the attack impact of historical attack data of hydropower plants based on the expert knowledge base to obtain an attack consequence vector set; and performing unsupervised clustering on the attack consequence vector set to obtain a discrete scenario set. Based on discrete scenario sets and expert knowledge bases, the nominal probability of scenarios is calculated from threat intelligence data to obtain the nominal probability distribution; Based on discrete scenario sets, nominal probability distributions, robust radii, and power system models, a final solvable optimal model is constructed. Robust scheduling plan solution is performed on the final solvable optimization model to obtain robust unit combination and robust basic output; The scheduling plan is issued based on robust unit combination and robust basic output to obtain the issued scheduling instructions.

[0006] Optionally, unsupervised clustering is performed on the attack consequence vector set to obtain a discrete scene set, including: Acquire scene data; Unsupervised clustering of the attack consequence vector set based on scene data is performed to obtain K clusters; Calculate the centroid vector of each of the K clusters as K attack scenarios; Create an M-dimensional zero vector as a baseline scenario without attacks; The K attack scenarios and the baseline scenario without attack are merged to obtain a discrete scenario set.

[0007] Optionally, based on a discrete set of scenarios and an expert knowledge base, the nominal probability of the threat intelligence data is calculated to obtain a nominal probability distribution, including: Based on the indicator dictionary, threat intelligence data is parsed and vectorized to obtain evidence vectors; Retrieve the prior probability vector and likelihood matrix associated with the discrete scene set from the expert knowledge base; Calculate the scenario conditional likelihood from the evidence vector and the likelihood matrix to obtain the scenario conditional likelihood vector; The scene conditional likelihood vector and prior probability vector are updated using Bayesian methods and their probabilities are normalized to obtain the nominal probability distribution.

[0008] Optionally, based on the discrete scenario set, nominal probability distribution, robust radius, and power system model, a final solvable optimal model is constructed, including: Based on a discrete scenario set, a power system model, and first-stage decision variables, a scenario cost function is constructed. The first-stage decision variables are the unit start-up and shutdown plan and the basic output plan for the scheduling cycle. The distance matrix is ​​obtained by calculating the scene spacing matrix of the discrete scene set; The scenario cost function, nominal probability distribution, distance matrix, and robust radius are reconstructed dually based on the worst expected cost to obtain the equivalent linear form of the worst cost; Based on the power system model, the equivalent linear form of the worst cost is integrated into the main problem of sub-Bruker optimization to obtain the final solvable optimal model.

[0009] Optionally, robust scheduling plans are solved on the final solvable optimization model to obtain robust unit configuration and robust basic output, including: The final solvable optimal model is used to instantiate the optimization problem and configure the solver to obtain a solver instance; The solver instance is subjected to mixed-integer programming algorithm execution and solution to obtain the original solution object; The validity of the original solution object is verified and the scheduling plan is extracted to obtain the robust unit combination and robust basic output.

[0010] Optionally, a scheduling plan is issued based on robust unit combination and robust basic output to obtain the issued scheduling instructions, including: The robust unit combination and robust foundation output are converted into specific start-up, shutdown and output commands for each hydropower plant, which are then used as the dispatch commands issued.

[0011] Another aspect of the present invention provides a hydropower plant dispatch and management system that takes into account network security constraints, comprising: The scenario construction module is used to construct a discrete scenario set based on an expert knowledge base and historical attack data. This includes: vectorizing the attack impact of historical attack data of hydropower plants based on the expert knowledge base to obtain an attack consequence vector set; and performing unsupervised clustering on the attack consequence vector set to obtain a discrete scenario set. The probability calculation module is used to calculate the nominal probability distribution of threat intelligence data based on discrete scenario sets and expert knowledge bases. The optimal model construction module is used to construct the final solvable optimal model based on the discrete scenario set, nominal probability distribution, robust radius and power system model; The planning and solving module is used to perform robust scheduling plan solving on the final solvable optimization model to obtain robust unit combination and robust basic output; The instruction issuance module is used to issue scheduling plans based on robust unit combination and robust basic output in order to obtain the issued scheduling instructions.

[0012] Optionally, the scenario construction module is used to perform unsupervised clustering on the attack consequence vector set to obtain a discrete scenario set, including: The scenario construction module is specifically used for: Acquire scene data; Unsupervised clustering of the attack consequence vector set based on scene data is performed to obtain K clusters; Calculate the centroid vector of each of the K clusters as K attack scenarios; Create an M-dimensional zero vector as a baseline scenario without attacks; The K attack scenarios and the baseline scenario without attack are merged to obtain a discrete scenario set.

[0013] Optionally, the probability calculation module is specifically used for: Based on the indicator dictionary, threat intelligence data is parsed and vectorized to obtain evidence vectors; Retrieve the prior probability vector and likelihood matrix associated with the discrete scene set from the expert knowledge base; Calculate the scenario conditional likelihood from the evidence vector and the likelihood matrix to obtain the scenario conditional likelihood vector; The scene conditional likelihood vector and prior probability vector are updated using Bayesian methods and their probabilities are normalized to obtain the nominal probability distribution.

[0014] Optionally, the preferred model building module is specifically used for: Based on a discrete scenario set, a power system model, and first-stage decision variables, a scenario cost function is constructed. The first-stage decision variables are the unit start-up and shutdown plan and the basic output plan for the scheduling cycle. The distance matrix is ​​obtained by calculating the scene spacing matrix of the discrete scene set; The scenario cost function, nominal probability distribution, distance matrix, and robust radius are reconstructed dually based on the worst expected cost to obtain the equivalent linear form of the worst cost; Based on the power system model, the equivalent linear form of the worst cost is integrated into the main problem of sub-Bruker optimization to obtain the final solvable optimal model.

[0015] Compared to existing technologies, this invention first utilizes unsupervised clustering technology to systematically summarize historical network attack records, automatically generating a statistically representative attack consequence scenario library, laying a data-driven foundation for subsequent uncertainty quantification analysis. Addressing the critical challenge of accurately obtaining the probability of attack events in practice, this invention introduces a robust optimization theoretical framework, abandoning reliance on a single fixed probability distribution. Instead, it constructs a fuzzy probability set centered on the nominal distribution and defined by a robust radius. Among all possible probability distributions covered by this set, it solves for a unified scheduling scheme that minimizes the total expected operating cost under the most unfavorable distribution conditions. The resulting robust scheduling plan maintains a good level of economic operation under high-probability scenarios of normal power system operation, while also maintaining the safety baseline of power system operation under the most extreme attack scenarios. Thus, it achieves a robust balance between operational economy and power system security in a dynamically changing network security threat environment. Attached Figure Description

[0016] One or more embodiments are illustrated by way of example with reference numerals in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0017] Figure 1 A flowchart illustrating a hydropower plant scheduling and management method considering network security constraints according to an embodiment of the present invention; Figure 2 This is a data flow diagram of a hydropower plant scheduling and management method considering network security constraints according to an embodiment of the present invention; Figure 3 This is a flowchart illustrating a method for hydropower plant scheduling and management considering network security constraints according to an embodiment of the present invention, which calculates the nominal probability distribution of threat intelligence data based on a discrete scenario set and an expert knowledge base to obtain the nominal probability distribution. Figure 4 A flowchart illustrating the construction of a final solvable optimal model based on a discrete scenario set, nominal probability distribution, robust radius, and power system model, according to an embodiment of the present invention, for a hydropower plant dispatching and management method considering network security constraints. Figure 5 This is a block diagram of a hydropower plant dispatch and management system that takes into account network security constraints according to an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details are presented in the embodiments of the present invention to facilitate a better understanding of the invention. However, the technical solutions claimed in the present invention can be implemented even without these technical details and various variations and modifications based on the following embodiments. The division of the following embodiments is for ease of description and should not constitute any limitation on the specific implementation of the present invention. The various embodiments can be combined with and referenced by each other without contradiction.

[0019] To address the technical deficiencies described in the background section, this invention provides a hydropower plant scheduling and management method and system that considers network security constraints. This aims to overcome the insufficient adaptability of traditional scheduling optimization methods when facing the variability of network attack patterns and the high uncertainty of their occurrence probabilities. Specifically, this invention starts with the data-driven representation of attack impacts. It utilizes an expert knowledge system to standardize and quantify the actual impact of various historical attack events on the physical system, forming an attack consequence vector representation. An unsupervised clustering algorithm is then used to automatically extract several statistically representative typical attack consequence patterns. To overcome the core difficulty of accurately calibrating attack probabilities, this invention does not employ a single, potentially biased probability estimate. Instead, it innovatively constructs a distributed fuzzy set centered on the Bayesian-updated nominal probability distribution. Based on this, by solving a distributed robust optimization model, the optimal scheduling strategy under the most unfavorable conditions is found among all probability distributions covered by the fuzzy set, thereby obtaining a unified, robust scheduling plan with built-in attack tolerance. This scheduling plan does not require real-time online determination of whether an attack is occurring. Its decision-making itself contains a margin of defense against uncertain threats, ensuring excellent economic performance under high-probability normal operation conditions. At the same time, it can still safeguard the basic bottom line of safe operation of the power system under the most unfavorable attack scenarios, achieving a robust coordination between the economic benefits of power generation and the safe operation of the power system.

[0020] One embodiment of the present invention provides a hydropower plant scheduling and management method that takes into account network security constraints. Figure 1 This is a flowchart of a hydropower plant scheduling and management method considering network security constraints according to an embodiment of the present invention. Figure 2 This is a data flow diagram illustrating a hydropower plant scheduling and management method considering network security constraints according to an embodiment of the present invention. (In conjunction with...) Figure 1 and Figure 2According to an embodiment of the present invention, a hydropower plant scheduling and management method considering network security constraints includes the following steps: S100, constructing a discrete scenario set based on an expert knowledge base and historical attack data; S200, calculating the nominal probability of the scenarios based on the discrete scenario set and the expert knowledge base to obtain a nominal probability distribution; S300, constructing a final solvable optimal model based on the discrete scenario set, the nominal probability distribution, the robust radius, and the power system model; S400, solving the final solvable optimal model using a robust scheduling plan to obtain a robust unit combination and a robust basic output; S500, issuing a scheduling plan based on the robust unit combination and the robust basic output to obtain the issued scheduling instructions.

[0021] Specifically, in step S100, a discrete scenario set is constructed based on an expert knowledge base and historical attack data. It is recognized that network attacks exhibit significant diversity and randomness in dimensions such as type selection, target orientation, and severity of consequences. Directly applying scattered and unstructured historical attack records to the construction process of a scheduling optimization model not only faces significant technical difficulties but also struggles to effectively cover potential new attack patterns. Therefore, in the technical solution of this invention, a discrete scenario set is constructed based on an expert knowledge base and combined with historical attack data from hydropower plants. Its core purpose is to standardize and quantify the actual impact of each specific attack event on the physical system, such as the extent of tampering with water level measurements at a specific power station or the degree of interference with communication link delays. Furthermore, it aims to automatically and objectively extract several statistically representative typical attack consequence patterns from a large amount of historical data. Therefore, this provides a set of uncertain scenarios with structured features and controllable scale derived from real data for the subsequent establishment of the split-bar optimization model. This effectively enhances the model's accuracy in characterizing actual network attack threats and its engineering practicality, laying a solid foundation for formulating a scheduling scheme with real attack tolerance capabilities.

[0022] More specifically, in this embodiment of the invention, a discrete scenario set is constructed based on an expert knowledge base and historical attack data, including: vectorizing the attack impact of historical attack data of hydropower plants based on the expert knowledge base to obtain an attack consequence vector set; and performing unsupervised clustering on the attack consequence vector set to obtain a discrete scenario set.

[0023] Specifically, based on an expert knowledge base, historical attack data from hydropower plants is vectorized to obtain an attack consequence vector set. It should be noted that the original historical attack data from hydropower plants often comes from diverse sources and has varying recording formats. Most of it exists in unstructured text describing the attack process, making it impossible to directly parse and utilize by the mathematical framework of the scheduling optimization model. This makes it difficult to effectively quantify and analyze the security risk information contained within. Therefore, in the technical solution of this invention, the attack impact vectorization process is further performed on the historical attack data using an expert knowledge base to obtain an attack consequence vector set. The purpose is to accurately map and convert the manifestation of each specific network attack event from its network behavior level into a quantitative deviation value generated by it on the key state variables of the hydropower scheduling physical model. This transforms heterogeneous and fragmented security event logs into a unified and standardized numerical data set that can directly participate in mathematical calculations, providing a standardized data foundation for subsequently using machine learning methods to objectively extract typical patterns of attack behavior.

[0024] More specifically, in a concrete example of the present invention, a historical attack data record is first received, for example, described as a false data injection attack targeting the upstream water level gauge of Power Station A, and key information such as the attack type, attack target, and impact description are identified from it. Next, a knowledge base mapping step is performed. The identified key information is used to query a preset expert knowledge base, which stores causal mapping rules between network attack behaviors and physical system state variables. The query result will clarify that the attack affects the specific physical quantity of the upstream water level of Power Station A. Then, an impact quantification step is performed. Based on the specific description in the data record or the default impact level defined in the expert knowledge base, the impact of this attack on the upstream water level of Power Station A is quantified into a specific numerical value, such as +1.5 meters. Finally, a vector generation step is performed. Based on a predefined vector dimension template containing all key physical state variables (such as the upstream water level deviation of Power Station A and the communication link delay deviation of Power Station B), an attack consequence vector is created. In this attack consequence vector, the component corresponding to the upstream water level deviation of power station A is assigned a value of +1.5, while all other unaffected components (such as the communication link delay deviation of power station B) are assigned a value of 0. This process is repeated for all historical attack data records, and finally aggregated to form an attack consequence vector set.

[0025] Specifically, unsupervised clustering is performed on the attack consequence vector set to obtain a discrete scenario set. It should be understood that the attack consequence vector set generated in the previous step contains the quantified impact of each historical attack event, resulting in a very large number of similar or redundant vectors. Treating each vector as an independent scenario would lead to the dimensionality curse and computational infeasibility of subsequent optimization models. Therefore, in the technical solution of this invention, unsupervised clustering is further performed on the attack consequence vector set to obtain a discrete scenario set. This uses data-driven methods to objectively and automatically summarize the high-dimensional and complex attack impact data space into a few core, statistically representative attack patterns, and further defines an ideal benchmark in the attack-free state. This generates a discrete scenario set with a controllable number, clear structure, and accurate reflection of the main attack threat types. This not only reduces the computational complexity of subsequent sub-optimal models but also ensures the objectivity and typicality of uncertainty modeling, providing high-quality input for the final generation of a scheduling plan that is both economical and secure.

[0026] More specifically, in this embodiment of the invention, unsupervised clustering of the attack consequence vector set to obtain a discrete scene set includes: acquiring scene data; performing unsupervised clustering of the attack consequence vector set based on the scene data to obtain K clusters; calculating the centroid vector of each of the K clusters as K attack scenes; creating an M-dimensional zero vector as a non-attack baseline scene; and merging the K attack scenes and the non-attack baseline scene to obtain a discrete scene set.

[0027] Specifically, scenario data is acquired, and unsupervised clustering is performed on the attack consequence vector set based on the scenario data to obtain K clusters; the centroid vector of each of the K clusters is calculated as the K attack scenarios. It should be recognized that the attack consequence vector set generated in the aforementioned steps must undergo a standardized data preparation and loading process before entering the clustering analysis stage to ensure that subsequent algorithms can run correctly on a complete and formatted dataset. Furthermore, the attack consequence data after vectorization is still a discrete, high-dimensional, and massive set of points. Direct analysis based on this is insufficient to reveal the intrinsic structure and common characteristics of the attack's impact, nor can it form a concise and representative set of scenarios suitable for subsequent model construction. Therefore, in the technical solution of this invention, the acquisition of scene data is further performed, and unsupervised clustering is performed on the attack consequence vector set based on the scene data to obtain K clusters. The centroid vector of each of the K clusters is calculated as K attack scenarios. In this way, the algorithm automatically performs pattern recognition and summarization of the physical impact of all historical attacks in multi-dimensional space, aggregates attack events with similar impacts together, and summarizes them using a centroid vector that best represents the average impact of this type of attack. In this way, the original and scattered attack consequence data points can be refined into a few core, highly representative attack prototypes. This not only achieves dimensionality reduction and abstraction of the source of uncertainty, but also ensures that the generated attack scenarios are an objective reflection based on the statistical characteristics of the data, rather than subjective conjecture.

[0028] More specifically, in a concrete example of the present invention, the data loading module is first activated, whose task is to access a specified data table or file storing the attack consequence vector set output from the previous step. Next, a data reading and aggregation step is performed. The data loading module traverses the data source, reading all the attack consequence vectors stored therein, such as [1.5,0], [0,480], etc., one by one into memory, forming a temporary vector list. Then, a data verification and formatting step is performed. The vector list in memory is checked to verify that each element is a numerical vector conforming to a predetermined dimension (e.g., M dimensions), and all format errors or missing data items are eliminated. Finally, a dataset construction step is performed. All verified attack consequence vectors are organized into a single, structured data matrix, where each row represents the consequence vector of a historical attack event, and each column corresponds to a specific physical state variable. This data matrix is ​​the final scene data matrix that can be directly input into subsequent clustering algorithms, serving as the acquired scene data.

[0029] Then, based on the preset number of clusters K, K attack consequence vectors are randomly selected from the acquired scene data matrix as initial cluster centroids. Next, an iterative execution phase begins, which cyclically executes two steps: allocation and update. In the allocation step, each attack consequence vector in the scene data matrix is ​​traversed, its Euclidean distance to the current K cluster centroids is calculated, and the vector is assigned to the cluster represented by its nearest centroid. In the update step, after all attack consequence vectors have been assigned, the centroid for each cluster is recalculated; the new centroid is the arithmetic mean of all attack consequence vectors within that cluster. This allocation-update iterative process continues until the cluster assignments no longer change or the centroid movement is less than a preset convergence threshold. Finally, the scene extraction step is executed. When the algorithm converges, it outputs the final stable K centroid vectors, which are then identified as the K representative attack scenarios. For example, a centroid vector of [1.4,0] represents a typical false data injection attack pattern targeting the water level gauge of Power Station A.

[0030] Specifically, an M-dimensional zero vector is created as the attack-free baseline scenario; K attack scenarios and the attack-free baseline scenario are merged to obtain a discrete scenario set. It should be understood that the K centroid vectors obtained through clustering in the previous step only represent different types of attack patterns, while a complete risk decision model must simultaneously include the normal operation of the system, i.e., the state where no attacks occur. Otherwise, the economics of the scheduling scheme under normal conditions cannot be evaluated, nor can a complete probability space be constructed. Therefore, in the technical solution of this invention, an M-dimensional zero vector is further created as the attack-free baseline scenario, and the K attack scenarios and the attack-free baseline scenario are merged to obtain a discrete scenario set. This explicitly and standardizedly defines the ideal operating state of the system within the same mathematical framework as the attack scenarios, thereby constructing a complete set covering all analysis scenarios from normal to various anomalies. This provides a complete and comprehensive sample space for subsequent probability allocation and robust optimization, ensuring that the final scheduling decision is the optimal choice made after fully weighing the economic benefits of normal operation against the security costs of resisting various attacks.

[0031] More specifically, in one embodiment of the present invention, the algorithm is first initialized: based on a pre-set number of clusters K, K attack consequence vectors are randomly selected from the acquired scene data matrix to serve as initial cluster centroids. Then, an iterative computation phase is entered, which cyclically executes two sub-operations: sample allocation and centroid update. In the sample allocation sub-operation, each attack consequence vector in the scene data matrix is ​​traversed, and the Euclidean distance between the attack consequence vector and the current K cluster centroids is calculated. The attack consequence vector is then assigned to the cluster category represented by the centroid with the shortest distance. In the centroid update sub-operation, after the assignment of all attack consequence vectors is completed, the coordinates of the centroid are recalculated for each cluster. The new centroid is the arithmetic mean of all member attack consequence vectors within that cluster. This alternating allocation-update iteration continues until the membership of each cluster no longer changes or the displacement of each centroid drops below a preset convergence threshold. Finally, scene extraction is performed: after the algorithm converges, the final stable K centroid vectors are output. These K centroid vectors are then identified as the K representative attack scenes. For example, a certain center point vector is [1.4,0], which physically represents a typical false data injection attack pattern targeting the water level sensor of Power Station A.

[0032] Specifically, in step S200, nominal probability calculations are performed on the threat intelligence data based on a discrete scenario set and an expert knowledge base to obtain a nominal probability distribution. It is recognized that the prior probability of each attack scenario only reflects the average risk level based on long-term historical statistics and cannot capture the dynamically changing short-term attack risk situation revealed by real-time threat intelligence. This makes it difficult for scheduling decisions to effectively allocate defense resources to address the most pressing security threats. Therefore, in the technical solution of this invention, nominal probability calculations are further performed on the threat intelligence data based on a discrete scenario set and an expert knowledge base to obtain a nominal probability distribution. The core objective is to establish a probabilistic reasoning framework that systematically integrates unstructured real-time threat intelligence with structured prior security knowledge, using Bayes' theorem to transform qualitative intelligence information into a quantitative and dynamic adjustment of the probability of each attack scenario. This allows for the generation of a nominal probability distribution that reflects the latest security situation and is dynamically updated. This distribution serves as the central reference for subsequent sub-bar optimization, enabling the final scheduling plan to more effectively prevent current high-probability attack threats. While maintaining sufficient security margins, it avoids unnecessary economic losses during periods of low threat levels, achieving dynamic coordination between security protection and economic operation.

[0033] Figure 3This is a flowchart illustrating a method for hydropower plant scheduling and management considering network security constraints according to an embodiment of the present invention, which calculates the nominal probability distribution of threat intelligence data based on a discrete scenario set and an expert knowledge base. For example... Figure 3 As shown, step S200 includes: S210, performing threat intelligence parsing and evidence vectorization on the threat intelligence data based on the indicator dictionary to obtain evidence vectors; S220, retrieving prior probability vectors and likelihood matrices associated with the discrete scenario set from the expert knowledge base; S230, calculating scenario conditional likelihood on the evidence vectors and likelihood matrix to obtain scenario conditional likelihood vectors; S240, performing Bayesian update and probability normalization on the scenario conditional likelihood vectors and prior probability vectors to obtain nominal probability distributions.

[0034] Specifically, in steps S210 and S220, threat intelligence data is parsed and vectorized into evidence vectors based on an indicator dictionary to obtain evidence vectors; prior probability vectors and likelihood matrices associated with the discrete scenario set are retrieved from an expert knowledge base. It should be understood that the original threat intelligence data is qualitative, unstructured text and cannot directly participate in subsequent probability calculations. Furthermore, the prior knowledge (i.e., the probability estimated empirically without any new evidence) and likelihood knowledge (i.e., the correlation strength between a specific attack and specific evidence) necessary for Bayesian updates also need to be obtained from a pre-defined knowledge system. Therefore, in the technical solution of this invention, threat intelligence data is further parsed and vectorized into evidence vectors based on an indicator dictionary to obtain evidence vectors; prior probability vectors and likelihood matrices associated with the discrete scenario set are retrieved from an expert knowledge base, thereby transforming dynamic, unstructured external security information into a standardized, computable evidence vector, and simultaneously extracting all the basic parameters necessary for probabilistic inference from a static, structured expert knowledge base. This ensures that all the necessary, formatted input data is prepared for subsequent Bayesian update calculations, guaranteeing that the entire probability calculation process is based on solid objective evidence and systematic expert knowledge.

[0035] More specifically, in a concrete example of the present invention, threat intelligence parsing and vectorization are performed first. The parsing module receives a threat intelligence text, such as information about a vulnerability in Rockwell ControlLogix series PLCs that can be used to carry out fake data injection attacks. The parsing module uses natural language processing technology to identify key entities, such as the Rockwell PLC vulnerability and FDIA attack techniques. Next, indicator matching is performed, comparing the identified entities with a preset indicator dictionary, matching indicator I1 (Rockwell PLC vulnerability) and indicator I2 (FDIA attack technique). Finally, the evidence vector generation step is executed, instantiating a binary vector with the same dimension as the indicator dictionary, setting the components corresponding to indicators I1 and I2 to 1, and setting the remaining components to 0, thereby generating the evidence vector. At the same time, a knowledge base retrieval is performed. First, a priori probability retrieval is performed, which queries the expert knowledge base database, requesting the return of a priori probability vector associated with the current discrete scenario set (e.g., including normal, FDIA, and DoS scenarios). The database returns a vector containing three probability values. Next, a likelihood matrix retrieval is performed, querying the expert knowledge base again to request a complete likelihood matrix. Each row of this matrix corresponds to a scenario, and each column corresponds to a threat indicator. The element value in the matrix represents the probability of observing the indicator represented by that column when the scenario represented by that row occurs. These two data structures retrieved from the knowledge base, along with the previously generated evidence vector, will serve as input for subsequent calculations.

[0036] Specifically, in step S230, scenario conditional likelihood calculation is performed on the evidence vector and likelihood matrix to obtain the scenario conditional likelihood vector. It should be noted that the evidence vector and likelihood matrix obtained in the preceding steps are independent at the information level: the former describes the current state of the observed threat indicator combination, while the latter defines the conditional correlation strength between a single indicator and a single scenario. There is a lack of a comprehensive quantitative indicator to measure the interpretability of each preset scenario for the current overall threat intelligence. Therefore, in the technical solution of this invention, scenario conditional likelihood calculation is further performed on the evidence vector and likelihood matrix to obtain the scenario conditional likelihood vector. Its core function is to systematically and quantitatively calculate the joint conditional probability value of observing a specific evidence combination under the premise that the scenario assumption is true for each discrete scenario. Thus, high-dimensional, dispersed evidence information can be condensed into a one-dimensional, intuitive likelihood vector. Each element in this vector accurately quantifies the degree of agreement between the corresponding scenario and the current threat intelligence, providing the most crucial decisive input for the subsequent Bayesian probability update operation.

[0037] More specifically, in a concrete example of the present invention, firstly, input data is acquired. The calculation module is activated to obtain the evidence vector generated in the previous steps, such as [1,1,0], and the likelihood matrix retrieved from the expert knowledge base. Next, an iterative calculation loop is entered, which traverses each discrete scenario, from the normal scenario ξ0 to the last attack scenario ξ. K In each iteration, for example, for FDIA attack scenario ξ1, a conditional probability multiplication step is performed. Based on the assumption that each threat indicator is conditionally independent in a given scenario, each component of the evidence vector is traversed. For components in the evidence vector with a value of 1 (such as the 1st and 2nd components), the corresponding probability value (e.g., 0.90 and 0.80) is extracted from the corresponding row of the likelihood matrix ξ1; for components with a value of 0 (such as the 3rd component), the corresponding probability value (e.g., 0.05) is extracted and its complement is calculated, i.e., 1 minus this value (1-0.05=0.95). Then, all these selected probability values ​​are multiplied together, such as 0.90 × 0.80 × 0.95, to obtain the final conditional likelihood value of scenario ξ1. This calculation process is repeated for all scenarios. Finally, a scenario conditional likelihood vector is constructed. The conditional likelihood values ​​calculated for each scenario are stored in a new vector according to the order of the scenarios. This new vector is the final output scenario conditional likelihood vector.

[0038] Specifically, in step S240, the scenario conditional likelihood vector and prior probability vector are updated using Bayesian methods and normalized to obtain a nominal probability distribution. It should be understood that the scenario conditional likelihood vector calculated in the previous step only reflects the degree of support of current threat intelligence for each scenario, while the prior probability vector represents long-term, static risk perception. These two have not yet been integrated into a unified final probability judgment that can guide current decision-making. Therefore, in the technical solution of this invention, the scenario conditional likelihood vector and prior probability vector are further updated using Bayesian methods and normalized to obtain a nominal probability distribution. This allows for the combination of prior knowledge representing long-term experience with real-time evidence reflecting short-term situations through a rigorous probabilistic reasoning framework, thereby calculating the posterior probability of each scenario occurring after obtaining new intelligence. This generates a dynamically updated scenario probability distribution that includes historical statistical patterns and fully responds to current specific threat intelligence, providing crucial and timely uncertainty parameter inputs for subsequent robust scheduling models.

[0039] More specifically, in a concrete example of the present invention, firstly, input data is acquired. The calculation module acquires the scenario conditional likelihood vector generated in the previous steps, for example [0.000198, 0.684, 0.0005], and the prior probability vector, for example [0.95, 0.04, 0.01]. Next, the unnormalized posterior probability is calculated. The two acquired vectors are multiplied element-wise, that is, for each scenario, its prior probability is multiplied by the corresponding conditional likelihood value to obtain a temporary, unnormalized posterior probability vector, for example [0.0001881, 0.02736, 0.000005]. Then, the normalization factor is calculated. All elements in this posterior probability vector are summed to obtain a marginal likelihood, for example, 0.0275531, which serves as evidence. Finally, probability normalization is performed. Each element in the posterior probability vector is divided by the calculated normalization factor to obtain the final nominal probability distribution. For example, the final probability of an FDIA attack scenario is 0.02736, which, when divided by 0.0275531, is approximately 0.9930. After this process is completed, a final nominal probability distribution vector is output, in which the sum of all elements is exactly 1.

[0040] Specifically, in step S300, a final solvable optimal model is constructed based on the discrete scenario set, nominal probability distribution, robust radius, and power system model. It should be understood that the nominal probability distribution obtained in the previous step is merely the best estimate based on existing information, and it inherently possesses uncertainties and bias risks. If optimization is performed directly based on this single distribution, the resulting scheduling scheme will be vulnerable when the actual probability distribution deviates from the nominal distribution, failing to guarantee system safety under the worst-case scenario. Therefore, in the technical solution of this invention, a final solvable optimal model is further constructed based on the discrete scenario set, nominal probability distribution, robust radius, and power system model. This transforms an abstract min-max robust optimization problem designed to combat probability distribution uncertainty into a large-scale but structurally deterministic mixed-integer programming master problem that can be directly solved by a standard optimization solver, through a series of precise mathematical steps, including cost quantification, definition of scenario geometric relationships, application of duality theory, and integration of physical constraints. In this way, a final optimization model that mathematically guarantees the robustness of the distribution can be generated. The goal of this model is no longer to seek the optimal under a single, unreliable probability assumption, but to find the best scheduling strategy in the worst case within a fuzzy set that includes all probability distributions that are sufficiently close to the nominal distribution. This provides a computable and theoretically complete decision-making basis for solving a scheduling plan that is truly attack-tolerant.

[0041] Figure 4This document presents a flowchart illustrating the construction of a final solvable optimal model based on a discrete scenario set, nominal probability distribution, robust radius, and power system model, according to an embodiment of the present invention for hydropower plant dispatching and management considering network security constraints. Figure 4 As shown, step S300 includes: S310, constructing a scenario cost function based on a discrete scenario set, a power system model, and first-stage decision variables, where the first-stage decision variables are the unit start-up and shutdown plan and the basic output plan for the scheduling cycle; S320, calculating the scenario spacing matrix for the discrete scenario set to obtain a distance matrix; S330, performing dual reconstruction based on the worst-case expected cost on the scenario cost function, nominal probability distribution, distance matrix, and robust radius to obtain an equivalent linear form of the worst-case cost; S340, performing a multi-barrel optimization master problem integration on the equivalent linear form of the worst-case cost based on the power system model to obtain the final solvable optimal model.

[0042] Specifically, in step S310, a scenario cost function is constructed based on the discrete scenario set, the power system model, and the first-stage decision variables, where the first-stage decision variables are the unit start-up and shutdown plan and the basic output plan for the scheduling cycle. It should be understood that the subsequent robust optimization model requires a clear objective function to evaluate the economic performance of each candidate scheduling scheme under different attack scenarios. This performance includes not only the planned generation cost but also the additional emergency costs incurred in responding to attacks. Therefore, in the technical solution of this invention, a scenario cost function is further constructed based on the discrete scenario set, the power system model, and the first-stage decision variables, where the first-stage decision variables are the unit start-up and shutdown plan and the basic output plan for the scheduling cycle. This establishes an accurate and quantifiable total cost calculation formula for each combination of a preset scheduling scheme and possible attack scenarios. This provides a clear and specific evaluation benchmark for the entire optimization problem, enabling the model to find the optimal robust scheduling strategy by minimizing the expected cost in the worst-case scenario while satisfying all physical constraints.

[0043] More specifically, in a concrete example of the invention, firstly, a first-stage cost is determined. Based on the first-stage decision variables, namely the unit start-up and shutdown plan and the basic output plan, a deterministic cost independent of the specific attack scenario is calculated. This cost includes the start-up and shutdown costs of all units within the scheduling cycle, as well as the generation cost (or water consumption cost) incurred by operating according to the basic output plan. Next, a second-stage cost is determined. For each specific scenario in the discrete scenario set, such as a scenario representing a spoofing attack, the amount of power imbalance caused by the attack on the power system is assessed. To restore balance, additional adjustment costs are required, which is the minimum cost of calling up standby units, performing emergency power support, or, in the worst-case scenario, implementing load shedding, where load shedding has a penalty cost. For the baseline scenario without an attack, its second-stage cost is zero. Finally, a total cost function is synthesized. The first-stage cost is added to the second-stage cost for the specific scenario to obtain the total cost function for that scenario. For example, in a scenario where a severe attack causes partial load shedding, the value of its scenario cost function will be higher than the cost function value in the baseline scenario without an attack.

[0044] Specifically, in step S320, the scene spacing matrix is ​​calculated for the discrete scene set to obtain the distance matrix. It should be noted that a key step in constructing the robust optimization model is defining a probability distribution fuzzy set based on the Wasserstein metric. The calculation of the Wasserstein distance itself requires a pre-determined ground distance metric that quantifies the degree of difference between scenes. Without this fundamental metric, the geometric size and boundary shape of the fuzzy set cannot be defined. Therefore, in the technical solution of this invention, the scene spacing matrix is ​​further calculated for the discrete scene set to obtain the distance matrix. The purpose is to calculate a clear, quantified geometric distance value for each pair of scene vectors in the scene space, thereby accurately describing the uncertain spatial structure of the entire scene set with a complete pairwise numerical relation matrix. This provides a crucial basic input parameter for the subsequent dual reconstruction step. This distance matrix is ​​the mathematical foundation for defining the Wasserstein fuzzy set and deriving its dual equivalent form, enabling the entire robust optimization model to be built on the basis of accurate quantification of the geometric relationships in the scene space.

[0045] More specifically, in a concrete example of the invention, firstly, the configuration uses the 1-norm, i.e., Manhattan distance, as the standard for calculating the distance between two scene vectors. The metric is the sum of the absolute values ​​of the differences between the corresponding components of the two scene vectors. Next, matrix initialization is performed. A square matrix of dimension (K+1)×(K+1) is instantiated, and all its elements are initialized to zero; this matrix is ​​the distance matrix to be filled. Then, an iterative calculation loop is entered, which traverses each unique pair of scene vectors in the discrete scene set. In each iteration, for example, when calculating the distance between the non-attack baseline scene vector [0,0] and the scene vector [1.4,0] representing a fake data injection attack, the 1-norm formula is applied, and the result is the sum of the absolute values ​​of the differences between the corresponding components of the two vectors, i.e., |0-1.4| + |0-0|, yielding a distance value of 1.4. This calculated distance value is filled into the distance matrix at the intersection of the row and column corresponding to the two scenes. Since the distance matrix is ​​symmetric, the symmetric position of this intersection is also filled simultaneously. This loop continues until all off-diagonal elements in the square matrix have been calculated and filled. Finally, the fully filled distance matrix is ​​output.

[0046] Specifically, in step S330, the scene cost function, nominal probability distribution, distance matrix, and robust radius are reconstructed dually based on the worst-case expected cost to obtain an equivalent linear form of the worst-case cost. It should be understood that the original mixed-integer optimization model is a min-max bilevel optimization problem, which internally contains a subproblem of finding the worst case in an infinite-dimensional probability distribution space. This structure cannot be directly handled by standard mixed-integer programming solvers, resulting in a theoretically correct but computationally infeasible model. Therefore, in the technical solution of this invention, the scene cost function, nominal probability distribution, distance matrix, and robust radius are further reconstructed dually based on the worst-case expected cost to obtain an equivalent linear form of the worst-case cost. This applies the strong duality principle in optimization theory, accurately and losslessly transforming the internal maximization problem, which is difficult to handle in the original model, into an equivalent, structurally simple minimization problem, characterized by a set of linear constraints. In this way, the entire complex two-layer min-max problem can be reconstructed into a single-layer, unified, large-scale minimization problem with good mathematical properties (such as convexity), thus transforming a theoretical model that was originally computationally difficult into a practical problem that can be efficiently solved by existing commercial optimization solvers.

[0047] More specifically, in a concrete example of this invention, firstly, a non-negative dual variable λ is introduced into the mathematical model. This dual variable can be interpreted in economics as risk pricing or shadow price for the uncertainty of the probability distribution. Simultaneously, a corresponding auxiliary variable s is introduced for each discrete scenario.k Next, the dual objective function is constructed. The objective function of the internal maximization part in the original model is replaced with a simple linear expression consisting of the newly introduced dual and auxiliary variables, namely the product of λ and the robust radius ε, plus the nominal probability p of each scenario. k Its corresponding auxiliary variable s k The sum of products. Then, the core dual constraints are generated. A set of key linear inequality constraints is generated to relate the newly introduced variables to the original scene costs and inter-scene distances. The core form of this constraint is: for any two scenes j and k, the cost function value of scene j is related to the auxiliary variable s of scene k. k The difference must be less than or equal to the distance D between the dual variable λ and scenes j and k. jk The product of these constraints. This constraint needs to be constructed for all possible scenario pairs. Ultimately, the constructed dual objective function, together with this set of core dual constraints, constitutes an equivalent linear form of the original worst-case expected cost, and serves as the output.

[0048] Specifically, in step S340, the equivalent linear form of the worst-case cost is integrated using a partial Brussels bar optimization master problem based on the power system model to obtain the final solvable optimal model. It should be understood that the dual reconstruction form obtained in the previous step is merely an equivalent substitution of the original problem's objective function; it does not contain any physical laws or constraints regarding power system operation. An effective scheduling scheme must simultaneously satisfy economic optimality and physical feasibility. Therefore, in the technical solution of this invention, the equivalent linear form of the worst-case cost is further integrated using a partial Brussels bar optimization master problem based on the power system model to obtain the final solvable optimal model. This systematically merges the power system constraint equations representing the laws governing the physical world with the dual reconstruction objective function and related constraints representing the pursuit of robustness under uncertainty, thereby constructing a single, complete mathematical programming problem that includes all decision variables, objectives, and constraints. In this way, an optimization model that reflects the real power grid operation constraints and mathematically guarantees the robustness of the distribution can be generated. This model is the final and directly solvable model that serves as a bridge between abstract theory and specific scheduling decisions. Its solution is the attack-tolerant scheduling plan pursued by this invention.

[0049] More specifically, in a concrete example of the present invention, firstly, decision variables are aggregated. All variables that need to be solved are uniformly incorporated into the model, including first-stage variables representing physical decisions (start-stop status and base output of all units at all times), as well as mathematical auxiliary variables introduced in the preceding dual reconstruction (dual variable λ and auxiliary variables s for each scenario). kNext, the final objective function is constructed. The equivalent linear form obtained after dual reconstruction, which is the sum of minimizing the product of λ and the robust radius plus the product of the nominal probabilities of each scenario and their auxiliary variables, is set as the sole optimization objective of the entire ensemble model. Then, a comprehensive constraint integration is performed. Two main sets of constraints are incorporated into the model: the first set is physical constraints from the power system model, including but not limited to system power balance constraints (total power generation equals total load), upper and lower limits of unit output constraints, unit ramp rate constraints, and reservoir balance constraints of hydropower plants; the second set is mathematical constraints from dual reconstruction, namely the core dual constraints connecting scenario costs, auxiliary variables, dual variables, and distance matrices. Finally, model encapsulation and output are performed. All aggregated decision variables, the final objective function, and all integrated constraints are encapsulated into a unified mathematical model conforming to the standard mixed integer programming format. This model is the final solvable optimal model and is ready to be fed into the subsequent optimization solver for computation.

[0050] Specifically, in step S400, a robust scheduling plan is solved on the final solvable optimal model to obtain robust unit combination and robust basic output. It should be understood that the final solvable optimal model constructed in the previous step is a complex large-scale mixed integer programming problem containing a large number of variables and constraints. Its optimal solution cannot be obtained through simple algebraic methods or direct derivation and must rely on a professional computing engine for solution. Therefore, in the technical solution of this invention, a robust scheduling plan is further solved on the final solvable optimal model to obtain robust unit combination and robust basic output. This transforms the abstract, theoretically complete mathematical model into a concrete, numerical optimal solution through a standardized and rigorous calculation process, and extracts scheduling instructions with practical guiding significance for the physical world. In this way, the entire complex technical concept can be ultimately realized as a unique, deterministic, and directly executable robust scheduling plan, achieving the optimal strategy with the lowest expected cost in the worst case, thereby materializing the attack tolerance concept into a concrete operational plan.

[0051] More specifically, in this embodiment of the invention, the robust scheduling plan solution for the final solvable optimal model to obtain the robust unit combination and robust basic output includes: instantiating the optimization problem and configuring the solver for the final solvable optimal model to obtain a solver instance; executing and solving the solver instance using a mixed integer programming algorithm to obtain the original solution object; and verifying the validity of the solution and extracting the scheduling plan for the original solution object to obtain the robust unit combination and robust basic output.

[0052] Specifically, the final solvable optimal model is instantiated as an optimization problem and configured with a solver to obtain a solver instance. It should be understood that the final solvable optimal model constructed in the previous step is an abstract mathematical expression, which cannot be directly recognized and processed by commercial or open-source optimization solver software. It must undergo an intermediate process of being converted into an executable format for a specific computing engine. Therefore, in the technical solution of this invention, the final solvable optimal model is further instantiated as an optimization problem and configured with a solver to obtain a solver instance. This precisely maps each component of the model, including all decision variables, objective functions, and constraints, to the specific data structures and function calls specified by the solver software application programming interface (API), and pre-sets the key performance parameters of the solving process. In this way, a solver instance that is fully initialized in the computing environment, contains all model information, and is configured with a clear solution strategy can be generated, providing a direct and unambiguous computational object for the subsequent efficient and reliable execution of large-scale mixed-integer programming algorithms.

[0053] More specifically, in a concrete example of the invention, firstly, an interface module is activated that translates the abstract mathematical model into a sequence of API calls to a specific solver (e.g., Gurobi). This process includes processing each decision variable in the model, such as the unit start / stop variable u. i,t Basic output variable p i,t Dual variable λ and auxiliary variable s k The process begins by calling the solver's variable creation function, explicitly specifying the variable's type (binary, continuous, or non-negative) and value range. Simultaneously, for each constraint in the model, including physical and dual constraints, the solver's constraint addition function is called, constructing them into the model within the solver's memory. Finally, the objective function setting function is called to set the model's optimization objective to minimization. Next, the solver parameter configuration step is executed. Key parameters are set for the soon-to-be-generated solver instance; for example, the mixed-integer programming solution gap is set to a small value to ensure solution accuracy, while a maximum computation time limit is set to ensure a high-quality feasible solution is obtained within a finite time. Finally, the solver object, containing all model information and with its parameters configured, is finalized as the solver instance and output.

[0054] Specifically, a mixed-integer programming algorithm is executed and solved on the solver instance to obtain the original solution object. It should be noted that the solver instance generated in the preceding steps is essentially only a static, unsolved mathematical description of the optimization problem, containing no solution information. The mixed-integer nature of this problem dictates that its solution process is a computationally intensive task requiring a systematic search across a vast feasible domain using complex iterative algorithms. Therefore, in the technical solution of this invention, a mixed-integer programming algorithm is further executed on the solver instance to obtain the original solution object. Its core function is to activate and run the highly optimized algorithm engine within the solver, commanding it to systematically explore the entire decision space based on preset configuration parameters, while satisfying all physical and mathematical constraints, to find the globally optimal solution that minimizes the robust objective function or an approximately optimal solution that meets the accuracy requirements. Thus, a static problem statement can be transformed into a concrete quantitative result containing the optimal values ​​of all decision variables and solution state information through a rigorous and reproducible computational process, providing a direct data source for the extraction of the final scheduling plan.

[0055] More specifically, in a concrete example of the invention, firstly, the solution is initiated. An instruction to execute the solution is issued to the configured solver instance. Next, the mixed-integer programming algorithm engine within the solver is activated and used to perform the core solution operations. This mixed-integer programming algorithm, such as branch and bound, systematically explores all possible combinations of unit start-up and shutdown by constructing a search tree. At each node of the search tree, the algorithm obtains a bound on the optimal cost achievable under the current branch by solving a relaxed linear programming problem with relaxed integer requirements, and uses this bound to prune search branches that have no hope of producing a better solution, thereby improving solution efficiency. The execution of the algorithm stops when preset termination conditions are met, including that a global optimal solution has been found and proven, or the relative gap between the currently found best integer solution and the bound of the theoretical optimal solution is less than a preset solution gap, or the total computation time has reached a preset time limit. Finally, when the algorithm terminates, the solver encapsulates the final results of the computation process, including the solution status code (e.g., optimal or suboptimal), the value of the optimal objective function, and the optimal values ​​of all decision variables (including unit start-up and shutdown, base output, dual variables, etc.), into a single primitive solution object and returns it.

[0056] Specifically, the validity of the original solution object is verified and a scheduling plan is extracted to obtain the robust unit combination and robust basic output. It should be understood that the original solution object returned by the solver in the previous step is a comprehensive data structure containing the solution status code, objective function value, and optimal values ​​of all first-stage and auxiliary variables. It is not a directly usable scheduling plan, and its success and validity must be confirmed before execution. Therefore, in the technical solution of this invention, the validity of the solution and the scheduling plan are further verified and the robust unit combination and robust basic output are obtained by further verifying the validity of the original solution object. This establishes a standardized parsing and transformation process from the original calculation results to the final executable physical instructions. This process first ensures the validity and reliability of the calculation results, and then accurately separates and extracts the core physical variables representing the final scheduling decision from the complex solution data. In this way, a purely mathematical, high-dimensional solution vector can be safely and accurately transformed into a structured, physically meaningful, and directly applicable robust unit combination and basic output plan for guiding the operation of hydropower plants.

[0057] More specifically, in a specific example of the present invention, firstly, the solution status is verified. The verification module first accesses and reads the solution status code recorded in the original solution object. If the status code shows optimal or suboptimal within a preset tolerance, the solution is determined to be successful, and the process continues; if the status code shows infeasible or unbounded, the model is determined to have a problem, an alarm will be triggered, and the subsequent process will be terminated. After successful verification, the scheduling decision variable extraction is performed. Based on the variable name or index, all variable values ​​related to the first-stage decision are accurately extracted, that is, the start / stop status variable u of each unit in each scheduling period. i,t and basic output variable p i,t In this process, all auxiliary mathematical variables used solely for model building, such as the dual variable λ and the auxiliary variable s, are... k If the values ​​are not specified, they are ignored or filtered. Finally, the scheduling plan is formatted and output. The extracted discrete values ​​are filled into two pre-created matrices according to the dimensions of the generating unit and the time period. The rounded start / stop status values ​​are filled into the robust generating unit combination matrix, while the basic output values ​​are filled into the robust basic output matrix. These two matrices together constitute the final, structured robust scheduling plan and serve as the output.

[0058] Specifically, in step S500, a scheduling plan is issued based on the robust unit combination and robust basic output to obtain the issued scheduling instructions. It is worth noting that the issued scheduling instructions here are specific start-up, shutdown, and output instructions for each hydropower plant, derived from the robust unit combination and robust basic output. That is, step S500 includes: converting the robust unit combination and robust basic output into specific start-up, shutdown, and output instructions for each hydropower plant as the issued scheduling instructions. It should be understood that the robust unit combination and robust basic output calculated in the previous step exist in matrix form, representing a high-level, centralized scheduling plan, and are not themselves low-level operation instructions that can be directly parsed and executed by the hydropower plant's field control unit. Therefore, in the technical solution of this invention, a scheduling plan is further issued based on robust unit combination and robust basic output to obtain the issued scheduling instructions. This allows the abstract, global optimal scheduling scheme to be accurately and unambiguously translated and decomposed into a series of executable control commands, conforming to standard industrial communication protocols, for each specific generator unit at each specific time segment. This ensures that the robust scheduling strategy proven optimal in the mathematical model can be accurately understood and implemented by the physical world's control system, thereby ultimately transforming the theoretical advantages of the attack-tolerant scheduling framework into safety and economic benefits in the actual operation of the hydropower plant.

[0059] More specifically, in a concrete example of the present invention, firstly, the instruction generation module is activated. This module reads and parses the robust generator set combination matrix and robust basic output matrix output from the preceding steps. The instruction generation module processes data segment by segment according to the scheduling time sequence. Next, the specific instruction generation and formatting steps are executed. At each time segment, the instruction generation module generates specific instructions for each generator set. For example, if the robust generator set combination matrix shows that the status of a generator set changes from 0 to 1 during that time period, a start-up instruction is generated; if the status is 1, an active power output setting instruction is generated based on the corresponding value in the robust basic output matrix, for example, setting the output to 75 MW. These instructions are formatted into a standard data structure containing generator set identification, instruction type, and parameter values. Then, the formatted instruction data is encapsulated according to the industrial communication protocol used by the SCADA system (e.g., DNP3 or IEC 61850) to generate a complete data message containing the target device address, function code, data content, and checksum. Finally, the instruction transmission step is executed. The SCADA master station of the dispatch center sends these encapsulated command messages sequentially to the remote terminal units (RTUs) or programmable logic controllers (PLCs) of each subordinate hydropower plant through a dedicated communication network. These field devices then ultimately complete the control operations of the generator sets.

[0060] Compared with existing technologies, the present invention provides a hydropower plant scheduling and management method that considers network security constraints. Firstly, it automatically extracts typical attack consequence scenarios by performing unsupervised clustering on historical attack data, providing a data-driven basis for uncertainty modeling. To address the core challenge of accurately predicting attack probabilities, it introduces the idea of ​​robust optimization, constructing a fuzzy set containing multiple probability distributions instead of relying on a single nominal probability distribution. It then solves for the unified scheduling plan with the optimal expected cost under the worst-case probability distribution within this fuzzy set. The resulting robust scheduling scheme performs well under high-probability conditions and ensures that performance does not collapse under the most unfavorable attack scenarios, thus achieving a robust balance between economy and security under uncertain security threats.

[0061] Another embodiment of the present invention also provides a hydropower plant dispatch management system that takes into account network security constraints.

[0062] Figure 5 This is a block diagram of a hydropower plant dispatch and management system considering network security constraints according to an embodiment of the present invention. Figure 5 As shown, a hydropower plant dispatch management system 500 considering network security constraints according to an embodiment of the present invention includes: a scenario construction module 510, used to construct a discrete scenario set based on an expert knowledge base and historical attack data, including: vectorizing the attack impact of historical attack data of the hydropower plant based on the expert knowledge base to obtain an attack consequence vector set; and performing unsupervised clustering on the attack consequence vector set to obtain a discrete scenario set; a probability calculation module 520, used to calculate the nominal probability of scenarios based on the discrete scenario set and the expert knowledge base to obtain a nominal probability distribution; an optimal model construction module 530, used to construct a final solvable optimal model based on the discrete scenario set, the nominal probability distribution, the robust radius, and the power system model; a plan solving module 540, used to perform robust dispatch plan solving on the final solvable optimal model to obtain a robust unit combination and a robust basic output; and an instruction issuance module 550, used to issue a dispatch plan based on the robust unit combination and the robust basic output to obtain issued dispatch instructions.

[0063] For example, the scenario construction module 510 is used to perform unsupervised clustering on the attack consequence vector set to obtain a discrete scenario set, including: the scenario construction module 510 is specifically used to: acquire scenario data; perform unsupervised clustering on the attack consequence vector set based on the scenario data to obtain K clusters; calculate the centroid vector of each of the K clusters as K attack scenarios; create an M-dimensional zero vector as a non-attack baseline scenario; and merge the K attack scenarios and the non-attack baseline scenario to obtain a discrete scenario set.

[0064] For example, the probability calculation module 520 is specifically used for: performing threat intelligence parsing and evidence vectorization on the threat intelligence data based on the indicator dictionary to obtain evidence vectors; retrieving prior probability vectors and likelihood matrices associated with the discrete scenario set from the expert knowledge base; performing scenario conditional likelihood calculation on the evidence vectors and likelihood matrices to obtain scenario conditional likelihood vectors; and performing Bayesian updates and probability normalization on the scenario conditional likelihood vectors and prior probability vectors to obtain nominal probability distributions.

[0065] For example, the preferred model construction module 530 is specifically used for: constructing a scenario cost function based on a discrete scenario set, a power system model, and first-stage decision variables, where the first-stage decision variables are the unit start-up and shutdown plan and the basic output plan for the scheduling cycle; calculating the scenario spacing matrix for the discrete scenario set to obtain a distance matrix; performing dual reconstruction based on the worst-case expected cost on the scenario cost function, nominal probability distribution, distance matrix, and robust radius to obtain an equivalent linear form of the worst-case cost; and performing a multi-barrel optimization master problem integration on the equivalent linear form of the worst-case cost based on the power system model to obtain the final solvable preferred model.

[0066] The specific implementation method of the hydropower plant dispatch management system considering network security constraints provided in this embodiment of the invention can be found in the hydropower plant dispatch management method considering network security constraints provided in this embodiment of the invention, and will not be repeated here.

[0067] A hydropower plant dispatching and management system 500 considering network security constraints according to an embodiment of the present invention can be implemented in various wireless terminals, such as servers having a hydropower plant dispatching and management algorithm considering network security constraints. In one possible implementation, the hydropower plant dispatching and management system 500 considering network security constraints according to an embodiment of the present invention can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the hydropower plant dispatching and management system 500 considering network security constraints can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the hydropower plant dispatching and management system 500 considering network security constraints can also be one of many hardware modules of the wireless terminal.

[0068] Those skilled in the art will understand that the above embodiments are specific embodiments for implementing the present invention, and in practical applications, various changes in form and detail may be made without departing from the spirit and scope of the present invention.

Claims

1. A hydropower plant dispatching and management method considering network security constraints, characterized in that, include: Based on an expert knowledge base and historical attack data, a discrete scenario set is constructed, including: vectorizing the attack impact of historical attack data of hydropower plants based on the expert knowledge base to obtain an attack consequence vector set; and performing unsupervised clustering on the attack consequence vector set to obtain a discrete scenario set. Based on discrete scenario sets and expert knowledge bases, the nominal probability of scenarios is calculated from threat intelligence data to obtain the nominal probability distribution; Based on discrete scenario sets, nominal probability distributions, robust radii, and power system models, a final solvable optimal model is constructed. Robust scheduling plan solution is performed on the final solvable optimization model to obtain robust unit combination and robust basic output; The scheduling plan is issued based on robust unit combination and robust basic output to obtain the issued scheduling instructions.

2. The hydropower plant dispatching and management method considering network security constraints according to claim 1, characterized in that, Unsupervised clustering of the attack consequence vector set is performed to obtain a discrete scene set, including: Acquire scene data; Unsupervised clustering of the attack consequence vector set based on scene data is performed to obtain K clusters; Calculate the centroid vector of each of the K clusters as K attack scenarios; Create an M-dimensional zero vector as a baseline scenario without attacks; The K attack scenarios and the baseline scenario without attack are merged to obtain a discrete scenario set.

3. The hydropower plant dispatching and management method considering network security constraints according to claim 1, characterized in that, Based on discrete scenario sets and expert knowledge bases, the nominal probability of threat intelligence data is calculated to obtain the nominal probability distribution, including: Based on the indicator dictionary, threat intelligence data is parsed and vectorized to obtain evidence vectors; Retrieve the prior probability vector and likelihood matrix associated with the discrete scene set from the expert knowledge base; Calculate the scenario conditional likelihood from the evidence vector and the likelihood matrix to obtain the scenario conditional likelihood vector; The scene conditional likelihood vector and prior probability vector are updated using Bayesian methods and their probabilities are normalized to obtain the nominal probability distribution.

4. The hydropower plant dispatching and management method considering network security constraints according to claim 1, characterized in that, Based on a discrete scenario set, nominal probability distribution, robust radius, and power system model, a final solvable optimal model is constructed, including: Based on a discrete scenario set, a power system model, and first-stage decision variables, a scenario cost function is constructed. The first-stage decision variables are the unit start-up and shutdown plan and the basic output plan for the scheduling cycle. The scene spacing matrix is ​​calculated for a discrete scene set to obtain the distance matrix; The scenario cost function, nominal probability distribution, distance matrix, and robust radius are reconstructed using duality based on worst expected cost to obtain the equivalent linear form of the worst cost; Based on the power system model, the equivalent linear form of the worst cost is integrated into the main problem of Bruker optimization to obtain the final solvable optimal model.

5. The hydropower plant dispatching and management method considering network security constraints according to claim 1, characterized in that, Robust scheduling plan solutions are performed on the final solvable optimization model to obtain robust unit configuration and robust basic output, including: The final solvable optimal model is used to instantiate the optimization problem and configure the solver to obtain a solver instance; The solver instance is subjected to mixed-integer programming algorithm execution and solution to obtain the original solution object; The validity of the original solution object is verified and the scheduling plan is extracted to obtain the robust unit combination and robust basic output.

6. The hydropower plant dispatching and management method considering network security constraints according to claim 1, characterized in that, The scheduling plan is issued based on robust unit combination and robust basic output to obtain the issued scheduling instructions, including: The robust unit combination and robust foundation output are converted into specific start-up, shutdown and output commands for each hydropower plant, which are then used as the dispatch commands issued.

7. A hydropower plant dispatch and management system considering network security constraints, characterized in that, include: The scenario construction module is used to construct a discrete scenario set based on an expert knowledge base and historical attack data. This includes: vectorizing the attack impact of historical attack data of hydropower plants based on the expert knowledge base to obtain an attack consequence vector set; and performing unsupervised clustering on the attack consequence vector set to obtain a discrete scenario set. The probability calculation module is used to calculate the nominal probability distribution of threat intelligence data based on discrete scenario sets and expert knowledge bases. The optimal model construction module is used to construct the final solvable optimal model based on the discrete scenario set, nominal probability distribution, robust radius and power system model; The planning and solving module is used to perform robust scheduling plan solving on the final solvable optimization model to obtain robust unit combination and robust basic output; The instruction issuance module is used to issue scheduling plans based on robust unit combination and robust basic output in order to obtain the issued scheduling instructions.

8. The hydropower plant dispatch management system considering network security constraints according to claim 7, characterized in that, The scenario construction module is used to perform unsupervised clustering on the attack consequence vector set to obtain a discrete scenario set, including: The scenario construction module is specifically used for: Acquire scene data; Unsupervised clustering of the attack consequence vector set based on scene data is performed to obtain K clusters; Calculate the centroid vector of each of the K clusters as K attack scenarios; Create an M-dimensional zero vector as a baseline scenario without attacks; The K attack scenarios and the baseline scenario without attack are merged to obtain a discrete scenario set.

9. The hydropower plant dispatch management system considering network security constraints according to claim 7, characterized in that, The probability calculation module is specifically used for: Based on the indicator dictionary, threat intelligence data is parsed and vectorized to obtain evidence vectors; Retrieve the prior probability vector and likelihood matrix associated with the discrete scene set from the expert knowledge base; Calculate the scenario conditional likelihood from the evidence vector and the likelihood matrix to obtain the scenario conditional likelihood vector; The scene conditional likelihood vector and prior probability vector are updated using Bayesian methods and their probabilities are normalized to obtain the nominal probability distribution.

10. The hydropower plant dispatch management system considering network security constraints according to claim 9, characterized in that, The preferred model building module is specifically used for: Based on a discrete scenario set, a power system model, and first-stage decision variables, a scenario cost function is constructed. The first-stage decision variables are the unit start-up and shutdown plan and the basic output plan for the scheduling cycle. The scene spacing matrix is ​​calculated for a discrete scene set to obtain the distance matrix; The scenario cost function, nominal probability distribution, distance matrix, and robust radius are reconstructed using duality based on worst expected cost to obtain the equivalent linear form of the worst cost; Based on the power system model, the equivalent linear form of the worst cost is integrated into the main problem of Bruker optimization to obtain the final solvable optimal model.

Images