A Method for Constructing an Intelligent Agent for Small Hydropower Station Operation Based on Existing Automation Systems

Abstract

This invention relates to the field of intelligent operation and maintenance technology for small hydropower stations, and discloses a method for constructing an intelligent agent for the operation of small hydropower stations based on existing automation systems. This method collects data from existing SCADA systems through standard protocols and performs semantic mapping of point tables to construct a three-dimensional semantic model of unit-hydraulic structure-state; it uses a hierarchical state machine to identify operating conditions; it constructs a dynamic safety boundary model based on Mahalanobis distance and introduces a Bayesian posterior update mechanism; it generates operation suggestions through rule-based strategy matching and synchronously outputs structured decision explanation information; and it implements three-level controlled output with safety level as a constraint. This invention achieves a gradual intelligent upgrade with zero modification and low risk.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for small hydropower stations, specifically to a method for constructing an intelligent agent for the operation of small hydropower stations based on existing automation systems. Background Technology

[0002] Small hydropower stations, as an important component of renewable energy, play an irreplaceable role in rural power supply and regional power grid peak shaving. After years of construction, the vast majority of existing small hydropower stations have established automated operation systems centered on programmable logic controllers, speed governors, relay protection devices, and data acquisition and monitoring control systems, basically realizing basic functions such as unit start-up and shutdown control, operating condition regulation, parameter acquisition, and over-limit alarms. However, with the development of the power system, the existing automation systems have exposed several deep-seated problems that urgently need to be addressed in terms of intelligent operation.

[0003] Firstly, regarding the ability to understand operational status, existing automation systems primarily present information through point-based displays and fixed-threshold alarms. Operators are faced with a large number of discrete numerical points and over-limit alarm signals, lacking the ability to automatically identify the evolution of unit operation phases, inflow trends, and comprehensive operating conditions. When multiple parameters change slowly simultaneously but do not reach a single threshold, the system cannot make a comprehensive judgment on this gradual and complex operating condition. Operators still need to rely on personal experience to make inferences, which not only reduces the timeliness of operational condition response but also creates potential risks of human error.

[0004] Secondly, regarding the definition of safe operating boundaries, existing protection systems employ a fixed threshold triggering mechanism, only triggering protection actions when the absolute value of a certain parameter exceeds preset upper or lower limits. This approach is essentially a reactive mechanism. It cannot identify and manage critical operating states in advance, nor does it possess the intelligent judgment capability to determine whether it is appropriate to continue automatic operation under the current conditions. In actual operation, many hazardous conditions are not caused by the exceeding of a single parameter limit, but rather by the coordinated deviation of multiple parameters under specific combinations of conditions. The single-parameter fixed threshold mechanism of existing systems is insufficient to capture such multi-dimensional evolutionary patterns of safety risks.

[0005] Furthermore, regarding operational decision support, the operating procedures and expert experience of small hydropower stations have long existed in the form of paper documents or unstructured electronic texts, unable to directly participate in the decision-making logic of automated systems. When faced with complex operating conditions, operators need to consult procedures, recall experience, and make judgments within a short period of time, resulting in a significant information gap between automated systems and human decision-making. Especially in unattended or minimally staffed scenarios, this gap may lead to decision delays or even the absence of decisions.

[0006] Furthermore, regarding the traceability of operational decisions, existing systems struggle to structurally record the formation process of operational recommendations or decisions. Key information such as why operators chose a particular operational strategy at a specific time, what data and rules were used as the basis for that decision, and the risk assessment at the time are often unverifiable. This hinders operational debriefing, safety oversight reviews, and the transmission and accumulation of experience and knowledge.

[0007] Currently, there are two main technical approaches in the industry to address the aforementioned issues. The first is a comprehensive upgrade solution, which involves a complete replacement and upgrade of the existing automation system. This requires shutting down and dismantling the original programmable logic controller (PLC) and data acquisition and monitoring control system, and installing entirely new equipment. The cost of upgrading a single station typically exceeds 800,000 yuan, with a construction period of at least three months. Furthermore, there is a risk of control logic failure and safety protection interruption during the upgrade process, making it unaffordable for many small hydropower stations with limited economic benefits. The second approach is a monitoring enhancement solution, which adds sensors and alarm platforms to the existing system to expand the monitoring range. However, this solution only provides data display and threshold alarm functions, and cannot generate executable operational decisions, resulting in low human-machine collaboration efficiency. Neither of these solutions effectively solves the core technical problem of how to achieve intelligent decision-making closed-loop while fully preserving the safety of the existing control loop. Therefore, there is an urgent need for a low-cost, low-risk, and highly reliable technical solution that can achieve a gradual upgrade of small hydropower stations from automated to intelligent operation without replacing or modifying the existing automation and protection systems. Summary of the Invention

[0008] The purpose of this invention is to provide a method for constructing an intelligent operating agent for small hydropower stations based on existing automation systems. By building a logically independent intelligent operating layer on top of the existing data acquisition and monitoring control system, the system acquires the ability to semantically understand operating conditions, dynamically determine safe operating boundaries, automatically generate operating decisions, and make the decision-making process explainable and traceable. This improves the safety, controllability, and intelligence level of small hydropower station operation without any modifications. Simultaneously, this invention provides a low-cost, low-risk, and highly reliable intelligent upgrade solution for small hydropower stations, addressing the common industry problem of slow intelligent upgrade progress in existing power stations due to high costs, risks, and uncertain results.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] The method for constructing an intelligent agent for the operation of a small hydropower station based on an existing automation system includes: collecting operational data from the existing automation system of the small hydropower station through a standard communication protocol; performing semantic mapping processing on the collected operational data to construct a three-dimensional semantic model of the unit, hydraulic structure, and state; constructing a hierarchical operational state machine model based on the semantically mapped operational data; extracting temporal features of operational parameters within a sliding time window to identify the current operating condition of the unit and generate corresponding operating condition labels; constructing a dynamic safe operation boundary model based on multi-dimensional operational parameters; dividing the operating state into three safety levels—safe zone, critical zone, and prohibited automatic operation zone—through a weighted distance metric; generating operational suggestions through strategy matching using a rule-based strategy library under the joint constraints of operating conditions and safety levels; generating structured decision interpretation information for the operational suggestions and storing it in a structured manner; and outputting the operational suggestions in a hierarchical and controlled manner according to the safety level.

[0011] The core innovations of this invention are as follows: First, the intelligent agent is deployed independently in a logical bypass manner, accessing existing system data in read-only mode without modifying any control logic, fundamentally eliminating the risk of interference to the existing security protection system caused by intelligent upgrades. Second, discrete industrial data points are transformed into structured knowledge representations with domain semantics through point-table semantic mapping, providing a semantic foundation for subsequent condition identification and decision reasoning. Third, the hierarchical operational state machine model performs state identification at both the macro and micro levels, ensuring both a grasp of the overall operational situation and a fine perception of local parameter changes. Fourth, the dynamic safety boundary model based on Mahalanobis distance introduces a Bayesian posterior update mechanism, enabling the safety boundary to continuously adaptively calibrate with the accumulation of operational data, avoiding the rigidity of fixed threshold mechanisms. Fifth, the decision generation stage synchronously outputs structured explanatory information, making each operational suggestion traceable to specific triggering conditions, matching rules, and data basis, improving the trustworthiness and auditability of human-machine collaboration. Sixth, the hierarchical controlled output mechanism uses security levels as constraints to manage execution permissions in layers. It achieves automatic closed-loop in the safe zone, introduces manual confirmation in the critical zone, and only records and leaves traces in the prohibited zone, thus achieving a balance between security and intelligence.

[0012] Compared with existing technologies, the present invention has at least the following beneficial effects: it achieves intelligent upgrade of operation without modifying existing automation and protection systems; it improves the system's understanding of operating status through operating condition identification; it clarifies the safety stop line of automated operation through dynamic safety boundary modeling; it improves the interpretability and auditability of operating decisions through rule-based decision-making and interpretation mechanisms; and it achieves safe and controllable progressive intelligence through hierarchical controlled output, which is conducive to realizing safe and controllable operation of small hydropower stations with few or no staff. Attached Figure Description

[0013] Figure 1 This is an overall flowchart of the method of the present invention.

[0014] Figure 2 A flowchart for generating and hierarchically controlled outputs for operational decisions. Detailed Implementation

[0015] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings. Those skilled in the art should understand that the following embodiments are only used to illustrate the technical principles and implementation methods of the present invention and are not intended to limit the scope of protection of the present invention. Without departing from the technical concept of the present invention, those skilled in the art can make adaptive adjustments to the specific parameters and implementation details in the following embodiments according to actual application scenarios.

[0016] The overall execution flow of the method of this invention is as follows: Figure 1 As shown, the operating intelligent agent is independently deployed within the production control area of ​​the safety zone of the small hydropower station. It accesses the data bus of the existing automation system in read-only mode via a standard industrial communication protocol, without modifying the control program and protection logic of the existing programmable logic controller. Logically, the operating intelligent agent constitutes an intelligent decision-making layer independent of the existing control loop. Its output transmits operational suggestions to the existing system or operator terminals through a controlled interface. Preferably, the hardware carrier of the operating intelligent agent can be an industrial-grade embedded computing platform or an industrial control server installed within the safety zone, whose computing power should meet the needs of real-time data processing and state reasoning. In one embodiment of the invention, the operating intelligent agent is deployed in a run-of-river small hydropower station with an installed capacity of 2×6.3MW. This station is equipped with two mixed-flow turbine generator units. The existing automation system uses the Modbus TCP protocol for data communication, and the data acquisition and monitoring control system manages approximately 480 data points.

[0017] Step S1: Data Acquisition and Semantic Modeling. The operating agent first acquires operational data from the existing automation system via standard communication protocols. Preferably, the standard communication protocols include the Modbus TCP / RTU protocol and the IEC 60870-5-104 protocol. The operating agent, acting as a communication slave or client, accesses the data bus of the existing system and periodically reads real-time values ​​from the point table register in read-only mode. In one embodiment of the present invention, the operating agent polls the holding register of the existing programmable logic controller at a period of 500ms using the Modbus TCP protocol. The acquired operational data covers parameters such as upstream water level, downstream water level, inflow rate, unit active power, unit reactive power, unit speed, bearing vibration amplitude (including horizontal and vertical components), stator winding temperature, guide vane opening, and gate opening. For rapidly changing dynamic parameters such as bearing vibration amplitude, the sampling period can be set to 0.1s to 1s depending on the monitoring accuracy requirements, with 0.1s to 0.2s being preferred in scenarios where high-frequency vibration characteristics need to be captured. For parameters that change relatively slowly, such as upstream water level and temperature, the sampling period can be set to 1s to 5s to reduce communication load.

[0018] After acquiring the raw operational data, semantic mapping of the data points is performed. In traditional automation systems, each data point exists as a register address and a numerical value, lacking domain semantic information. This step maps each raw data point to a predefined semantic ontology model, forming a semantic entity with domain meaning. Specifically, the semantic mapping process is executed according to the following rules: First, a data point mapping configuration file is established, which defines the correspondence between each register address and a semantic entity; then, for each acquired data point, it is associated with the corresponding attribute in the unit object, hydraulic object, or operational status object according to the mapping configuration file.

[0019] In one embodiment of the present invention, the construction of the three-dimensional semantic model follows a subject-verb-object triple structure. Taking the unit object as an example, if the value read from register address 40001 is 5200.5, and the mapping configuration indicates that this address corresponds to the active power of Unit 1, then the generated semantic triple is <Unit 1, current active power, 5200.5kW>. Similarly, the upstream water level in the hydraulic object can be expressed as <Forebay, current water level, 342.15m>. The operating status object is automatically generated based on the current attribute values ​​of the unit object and the hydraulic object through rule reasoning. For example, when the active power of Unit 1 is greater than 0kW and the circuit breaker is in the closed state, <Unit 1, current operating stage, grid-connected power generation> can be generated through reasoning.

[0020] Preferably, the semantic model also maintains metadata information for each semantic entity, including physical quantity type (e.g., pressure, temperature, displacement), measurement range (e.g., water level range of 330m to 355m), sampling frequency (e.g., water level parameter 1s to 5s or vibration parameter 0.1s to 1s), and data quality identifier (e.g., normal, communication interruption, out of range). The introduction of data quality identifiers enables subsequent operating condition identification and safety boundary determination modules to automatically identify and eliminate the interference of abnormal data on decision-making. In a preferred embodiment of the present invention, when the number of consecutive missing values ​​of a certain parameter exceeds a preset threshold (e.g., no update for 10 consecutive sampling periods), the semantic entity corresponding to the parameter is automatically marked as a data failure state. Subsequent modules will apply a weight reduction or substitution processing to the parameter when performing state reasoning. In the weight reduction processing mode, the weight of the failure parameter in the Mahalanobis distance calculation is set to zero, and the corresponding dimension of the covariance matrix is ​​marginalized, so that the safety boundary determination is performed only based on the remaining valid parameters, thereby ensuring that the system can still maintain basic safety assessment capabilities under conditions of abnormal communication of some sensors. In the alternative processing mode, the system adopts the strategy of keeping the most recent valid value or the strategy of filling with typical values ​​based on similar historical operating conditions to temporarily replace the failed parameter. At the same time, the operation suggestion clearly indicates that the parameter is currently in the alternative mode, prompting the operators to pay attention to the working status of the corresponding sensor.

[0021] Furthermore, to accommodate the differences in point table configurations across different power plants, this invention designs a configurable semantic mapping template mechanism. Standard mapping templates are pre-set for common brands and models of small hydropower station automation systems; for power plants with non-standard configurations, operators can manually establish the mapping relationship between register addresses and semantic entities through the configuration interface. This mechanism ensures the universality of the method of this invention for different models of existing automation systems.

[0022] In the semantic modeling stage, the data preprocessing strategy deserves special attention. Because existing automated systems' communication links may be affected by electromagnetic interference or network congestion, the collected raw data may contain anomalies such as noise spikes, communication delays, or data loss. In one embodiment of this invention, a data preprocessing pipeline is introduced before semantic mapping, performing the following processes sequentially: First, range verification is performed, checking whether the value of each data point falls within the reasonable range of the corresponding physical quantity (e.g., water level values ​​should not exceed the range from the dead water level of the reservoir to the check flood level). Data exceeding the range is marked as outliers and discarded. Second, median filtering is performed; for vibration parameters with high sampling frequencies, the median value within a window of five consecutive sampling points is taken as the output value to eliminate the influence of impulse noise. Finally, timestamp alignment is performed; since different parameters may have different sampling frequencies, linear interpolation is used to uniformly align the data of all parameters to the same time base. Only the preprocessed running data can enter the semantic mapping stage, ensuring the data input quality of subsequent operating condition identification and safety boundary determination modules.

[0023] Step S2: Operating condition identification. Based on the semanticized operating data, a hierarchical operating state machine model is constructed to identify the current operating condition of the unit. The hierarchical operating state machine model consists of two layers: an outer macroscopic state machine and an inner microscopic state machine, with a hierarchical nesting relationship between them.

[0024] The outer macroscopic state machine is used to identify the macroscopic operating stages of the unit. In one embodiment of the present invention, five macroscopic states are defined: shutdown state, start-up transition state, grid-connected generation state, regulating operation state, and shutdown transition state. The transition conditions between macroscopic states are determined by a combination of key signals, specifically including: the condition for transitioning from the shutdown state to the start-up transition state is the detection of a start-up command signal and the start-up of the unit's guide vane opening; the condition for transitioning from the start-up transition state to the grid-connected generation state is that the unit speed reaches more than 95% of the rated speed and the circuit breaker closes successfully; the condition for transitioning from the grid-connected generation state to the regulating operation state is that the absolute value of the active power change rate exceeds 5% / min of the rated power or the guide vane opening adjustment frequency exceeds a preset threshold (e.g., more than 3 adjustments within 5 minutes); the condition for transitioning from the grid-connected generation state or the regulating operation state to the shutdown transition state is the detection of a shutdown command or protection action signal.

[0025] The inner-layer microstate machine further subdivides substates within each macrostate to achieve refined identification of operational details. Taking grid-connected power generation as an example, it can be further divided into steady-state operation substates, load increase substates, load decrease substates, and high vibration substates. Preferably, the transition conditions of the microstates are determined based on the statistical characteristics of the operating parameters within the sliding time window.

[0026] This invention introduces an adaptive sliding time window mechanism for time-series feature extraction. Specifically, within the sliding time window, statistical features such as the mean, standard deviation, and slope of the changing trend of the time series of operating parameters are calculated. Preferably, the window length is adaptively adjusted according to the current macroscopic state: under steady-state operating conditions, the window length is set to a larger value (e.g., 60s to 300s) to filter out short-term disturbances and capture slowly changing trends; under regulating operating conditions, the window length is shortened to 10s to 60s to improve the response sensitivity to regulating actions; under start-stop transition conditions, the window length is further shortened to 5s to 30s to track rapid state changes.

[0027] Based on temporal feature extraction, operating conditions are comprehensively identified by combining state transition rules. The output of operating condition identification is a structured operating condition label, which consists of three parts: a macroscopic state name, a microscopic sub-state name, and additional descriptive information. For example, a typical operating condition label can be described as grid-connected power generation—steady-state operation—stable inflow—normal water level. In one embodiment of the present invention, the rules for generating operating condition labels are as follows: First, the current macroscopic stage is determined based on the macroscopic state machine; then, the current sub-state is determined based on the microscopic state machine; finally, additional descriptions are generated based on the water level change rate (calculated through linear regression slope) and the power output change rate. If the absolute value of the upstream water level change rate is less than 0.01 m / min and the absolute value of the active power change rate is less than 1% / min of the rated power, the additional description is stable inflow—normal water level; if the water level change rate is positive and greater than 0.05 m / min, the additional description is increased inflow—rising water level.

[0028] Preferably, a state transition debouncing mechanism is introduced during the working condition identification process. That is, when a certain state transition condition is met, the state switch is not executed immediately, but the transition condition is required to be met continuously. A state transition can only be triggered after the condition is met continuously within a sampling period. In one embodiment of the present invention, The cycle time is set to 5 to 10 cycles, corresponding to an actual time of approximately 2.5 to 5 seconds. This anti-shake mechanism effectively avoids frequent state switching caused by transient data fluctuations, improving the stability and reliability of operating condition identification.

[0029] Step S3: Dynamic Safe Operation Boundary Modeling and Determination. Based on the identification of operating conditions, a dynamic safe operation boundary model is constructed based on multi-dimensional operating parameters to determine the safety level of the unit's current operating state. Unlike traditional single-parameter fixed threshold protection mechanisms, the safety boundary model of this invention comprehensively considers the joint distribution characteristics of multiple operating parameters and achieves an overall assessment of the operating state through a weighted distance metric in multi-dimensional space.

[0030] First, key parameters closely related to the safe operation of the unit are selected to form a safety boundary parameter vector. In one embodiment of the present invention, the parameter vector... Includes the following The components are: upstream water level deviation (the difference between the current water level and the normal storage water level), unit speed deviation (the relative value of the difference between the current speed and the rated speed), bearing horizontal vibration amplitude, bearing vertical vibration amplitude, stator winding temperature, guide vane adjustment frequency (the number of adjustments per unit time), and active power deviation (the difference between the current power and the target power). Based on this, it is necessary to determine the safety reference center and boundary parameters. The safety reference center... Defined as the statistical mean vector of each parameter under historical normal operating conditions, while the boundary parameters are represented by the covariance matrix. The form represents the joint distribution characteristics among the parameters.

[0031] This invention uses Mahalanobis distance as a weighted distance metric. The formula for calculating Mahalanobis distance is:

[0032] ,

[0033] in: This is the current runtime parameter vector, with dimension . , In this embodiment, the number of parameter components is... ; The safety baseline center vector has dimensions of The values ​​of each component are determined by the sample mean of historical normal operation data, and the units are consistent with the corresponding parameters; Let be the parametric covariance matrix, with dimension . Its diagonal elements are the variances of each parameter, and its off-diagonal elements are the covariances between parameters, reflecting the correlation between the parameters. It is the inverse of the covariance matrix; The Mahalanobis distance between the current operating state and the safety baseline center is a dimensionless scalar value.

[0034] Compared to Euclidean distance, Mahalanobis distance uses the inverse of the covariance matrix to standardize the parameter space, automatically adapting to differences in the dimensions and magnitudes of the parameters while fully considering the correlations between them. For example, when there is a positive correlation between vibration amplitude and rotational speed deviation, Mahalanobis distance can identify the risk inherent in the combined deviation of both being excessively high, while Euclidean distance might miss it because the absolute deviations of both are not large.

[0035] To further illustrate the engineering significance of this distance metric, a typical range of values ​​is given using the parameter vector in this embodiment as an example. Under normal operating conditions, the upstream water level deviation typically fluctuates within ±0.5m, the unit speed deviation is typically within ±0.3%, the normal value of the bearing horizontal vibration amplitude generally does not exceed 0.06mm, the vertical vibration amplitude does not exceed 0.08mm, the stator winding temperature is between 60℃ and 85℃, the guide vane adjustment frequency does not exceed twice every 5 minutes, and the active power deviation is typically within ±3% of the rated power. These parameters have vastly different physical dimensions and numerical magnitudes. If Euclidean distance is used directly, the deviation in vibration amplitude (on the order of 0.01mm to 0.1mm) will be masked by the variation in water level deviation (on the order of 0.1m to 1m). However, the Mahalanobis distance, through the standardization effect of the covariance matrix, gives each parameter a weight that matches its actual importance in the distance calculation, thereby avoiding evaluation bias caused by different dimensions.

[0036] Based on Mahalanobis distance values, the unit's operating status is divided into three safety levels. In one embodiment of the present invention, the three-zone division rules are as follows:

[0037] Safe zone: when At that time, it was determined to be a safe zone, in which The corresponding parameter joint distribution is at twice the standard deviation level, which, under the assumption of a multidimensional normal distribution, covers approximately 95.4% of the normal operating samples. Operating conditions within the safe zone indicate that the joint offset of the parameters is within the normal fluctuation range.

[0038] Critical region: when When this occurs, it is determined to be in the critical region, where This corresponds to a level three standard deviations, covering approximately 99.7% of normally operating samples. Operating conditions in the critical zone indicate that the parameter combination has deviated from the normal range but has not yet reached a dangerous level, requiring attention and enhanced monitoring.

[0039] Automatic operation prohibited zone: When When the system is in a prohibited automatic operation zone, it indicates that the current parameter combination has significantly deviated from the normal operating envelope, posing a high safety risk and making it unsuitable to continue automatic operation.

[0040] Preferably, the threshold and It is not a simple fixed constant, but can be dynamically adjusted according to the current operating condition label. For example, under start-up and shutdown transition conditions, because the unit's operating parameters themselves fluctuate significantly, the threshold can be appropriately relaxed (e.g., by adjusting the threshold). The threshold is adjusted to 2.5 times the standard deviation level to avoid excessive alarms; while under steady-state operating conditions, a relatively strict threshold setting is maintained to ensure safe sensitivity.

[0041] Another important innovation of this invention lies in the introduction of a Bayesian posterior update mechanism for the safety benchmark center. Covariance Matrix Perform online calibration. Specifically, this involves using data calculated based on historical data prior to the system's deployment. and As a prior distribution parameter, with the most recent The data collected within each sampling period is used as new observation samples, and the posterior distribution parameters are calculated using the Bayesian update formula. For the update of the mean vector, a normal-normal conjugate model is employed:

[0042] ,

[0043] in: The safety baseline center vector is updated by Bayesian posterior, with dimension . ; The prior safety baseline center vector is initialized from the sample mean of historical data. The prior equivalent sample size represents the degree of trust in prior information, and its value typically ranges from 100 to 1000. A larger value indicates a higher level of trust in the prior and a slower update speed; This represents the number of new observation samples within the current update window. Let be the mean vector of the new observation sample, with dimension . .

[0044] Similarly, for the covariance matrix It also performs a posteriori update to reflect changes in the parameter correlation structure in the latest runtime data. Preferably, the update window... The value is taken as the sample size of valid operational data from the past 24 to 72 hours, and the update frequency is set to be once every 24 hours. This online calibration mechanism allows the safety boundary to slowly track long-term drift factors such as equipment aging and seasonal hydrological changes, while also ensuring that the prior equivalent sample size is sufficient to maintain the safety margin. The suppression effect is such that it will not cause drastic shifts due to short-term abnormal data, thus ensuring the stability and robustness of the boundary model.

[0045] Step S4: Run the decision generation process, such as... Figure 2 As shown, after obtaining the operating condition label and safety level determination results, the process proceeds to the operating decision generation stage. This step uses a rule-based operating strategy library to perform strategy matching and generate operating suggestions tailored to the current operating condition.

[0046] The rule-based operation strategy library is the decision-making knowledge foundation of the method of the present invention, storing several pre-compiled operation strategy rules. Each strategy rule includes the following fields: strategy number, applicable operating conditions, suggested operation content, expected effect description, risk level rating, and execution authority constraints. Preferably, the strategy library is compiled based on operating procedures stipulated by national and industry standards, historical operating experience of power plants, and knowledge of domain experts, and is added to the library after being reviewed by senior operators. In one embodiment of the present invention, the strategy library contains approximately 120 to 200 strategy rules, covering multiple scenario categories such as normal operation optimization, response to changes in inflow water, handling of abnormal operating conditions, and start-up and shutdown operation assistance.

[0047] For example, a typical normal operation optimization strategy can be described as follows: Strategy No. R-OPT-012, applicable operating conditions: grid-connected power generation—steady-state operation—stable inflow—high water level, recommended operation: moderately increase guide vane opening to improve output, fully utilize surplus inflow to increase power generation, expected effect: output increase of 3% to 5%, water level returns to normal range within 2 hours, risk level is low, execution authority is automatic execution within the safe zone. Another example is an abnormal operating condition handling strategy: Strategy No. R-ABN-035, applicable operating conditions: grid-connected power generation—high vibration—normal water level, recommended operation: reduce active power to below 80% of rated power, observe vibration amplitude trend, expected effect: vibration amplitude decreases to normal range within 10 minutes, risk level is medium, execution authority is manual confirmation required in the critical zone. Each rule in the strategy library has been reviewed by senior operators and technical experts to ensure that the recommended operation is technically reasonable and safety-controllable.

[0048] The execution logic of the strategy matching process is as follows: First, the current operating condition label is compared item by item with the applicable operating condition conditions of each strategy in the strategy library. Preferably, the comparison adopts a multi-attribute matching method, that is, the macro-state, micro-sub-state and additional descriptive information in the operating condition label are matched separately, and the condition satisfaction score of each strategy is calculated. The formula for calculating the condition satisfaction score is as follows:

[0049] ,

[0050] in: The strategy matching score is a dimensionless value ranging from 0 to 1. In this embodiment, to match the total number of attributes, These correspond to four dimensions: macroscopic state, microscopic sub-state, hydrological description, and parameter trend, respectively. For the first The weight coefficients of each matching dimension satisfy... In one embodiment of the present invention (Macroeconomic state weights) (Weights of micro-sub-states) (Hydrological description weight) (Parameter trend weight); For the first Matching results in each dimension, when a complete match is achieved. When partially matched When there is a mismatch .

[0051] After calculating the matching scores of all strategies, the strategy with the highest score and a risk level not exceeding the upper limit allowed by the current safety level is selected as the final matching strategy. Preferably, when the safety level is in the safe zone, strategies with low or medium risk levels are allowed to be matched; when the safety level is in the critical zone, only conservative strategies with low risk levels are allowed to be matched; when the safety level is in the prohibited automatic operation zone, no strategy matching is performed, and only the current operating condition is recorded. If there are multiple candidate strategies with the same matching score, the strategy with the lower risk level is selected first to reflect the principle of safety priority.

[0052] After successful matching, an operational suggestion is generated, which includes the following elements: suggested operation content (operational guidance described in natural language, such as suggesting that the guide vane opening of Unit 1 be increased to 75% to improve output), risk level calibration value (low / medium / high levels), and execution authority identifier (directly mapped from the current safety level: safe zone → automatic execution, critical zone → execution after manual confirmation, prohibited zone → recording only).

[0053] In a preferred embodiment of the present invention, when strategy matching fails to find a satisfaction score exceeding a preset threshold (e.g., ... When a strategy is not found for the current operating condition, the system generates a default conservative suggestion, which is that no clear strategy is matched for the current operating condition. It is recommended to maintain the current operating state and strengthen manual monitoring. At the same time, the unmatched event is recorded in the log for future strategy library optimization reference.

[0054] Step S5, Decision Interpretation, Structured Evidence Recording, and Hierarchical Controlled Output: After the operational decision is generated, structured decision interpretation information is automatically generated for each operational suggestion. The purpose of generating this decision interpretation information is to enable operators to understand the logical basis behind each suggestion, thereby establishing trust in the agent's decisions and providing a traceable chain of evidence for post-operation review and safety supervision.

[0055] The decision interpretation information includes at least the following elements: trigger condition description, i.e., the natural language expansion of the current operating condition label, indicating what operating state the agent has identified; matching rule identifier, i.e., the strategy number and applicable condition summary of the final matching strategy, indicating which rule the agent made the suggestion based on; data basis summary, i.e., the current value, trend of change, and relative position of the key operating parameters that triggered the suggestion with respect to the safety boundary, such as the current horizontal vibration amplitude of the bearing being 0.085mm, which is 18% higher than the average of the past 60 minutes, and still has a margin from the critical zone threshold of 0.12mm; risk assessment description, i.e., the qualitative interpretation of the current safety level determination result and Mahalanobis distance value, such as the current safety distance measurement value being 1.73, which is within the safety zone (threshold 2.0), and the overall operating risk is controllable.

[0056] Preferably, the generation of decision explanation information employs a template-filling mechanism, which predefines several explanation templates and automatically fills in specific values ​​and descriptions based on actual data at runtime. This mechanism ensures consistent format and semantic standardization of the explanation information while avoiding ambiguity issues that may arise from natural language generation.

[0057] The operational condition labels, operational suggestions, decision explanations, and execution results are persistently stored in a timestamp-indexed structured log format. Each record in the structured log includes at least: a timestamp (accurate to milliseconds), an operational condition label, a security level, a Mahalanobis distance value, a matching strategy number, suggested operation content, a risk level, an execution method (automatic / manual confirmation / recording only), and an execution result (executed / confirmed / rejected / not executed). Preferably, the structured log is stored in a relational database or time-series database local to the running agent, supporting multi-dimensional retrieval by time range, operational condition type, and security level, providing data support for operational review analysis and statistical report generation.

[0058] While maintaining structured traces, operational recommendations are output in a hierarchical and controlled manner based on security levels. This hierarchical and controlled output mechanism is a key design feature of this invention, ensuring the security of existing systems. In one embodiment of this invention, the three-level output rules are as follows:

[0059] Level 1 (Automatic Issuance of Safety Zone Command): When the safety level is determined to be within the safety zone, operational recommendations are automatically issued to the existing automation system via a communication interface in the form of controlled execution commands. Preferably, the command content is verified for compliance before issuance, and can only be sent after confirming that it does not exceed the parameter adjustment range acceptable to the existing system. This level is suitable for routine optimization and adjustment operations, such as making small adjustments to the guide vane opening to optimize power distribution.

[0060] Level 2 (Manual Confirmation for Critical Zone): When the safety level is determined to be in the critical zone, operational recommendations are pushed to the operator's terminal with a prominent identification. Execution can only proceed after the operator's explicit confirmation. This level is suitable for scenarios with some risk but where the operation itself is reasonable, such as recommending reduced load operation when vibration is high but has not reached the protection threshold.

[0061] Level 3 (Record Only in Prohibited Zones): When the safety level determines that automatic operation is prohibited, the system does not output any execution instructions to the existing automation system. Instead, it records the current operating status and risk assessment information to a structured log and notifies the operators with a high-priority alarm. This level ensures that the operating agent strictly exits the execution loop under extreme risk conditions, completely returning control to the existing protection system and operators.

[0062] Preferably, the hierarchical controlled output communication interface design follows the principle of least privilege. When the running agent issues execution instructions to the existing automation system, it is only granted write access to specific register addresses, and the range of the written values ​​is protected by hardware-level limiting. For example, the write range of the guide vane opening instruction is limited to ±10% of the current opening value; instructions exceeding this range will be automatically intercepted by the communication interface and recorded in the exception log. Furthermore, the running agent also sets an instruction sending frequency limit; the same type of execution instruction cannot be repeatedly sent within a short period to avoid equipment oscillation caused by the existing system receiving continuous and dense control instructions due to software logic anomalies. In a preferred embodiment of the invention, the minimum sending interval for the same instruction is set to 30 seconds. Through the superposition of the above multiple security constraints, even if the running agent itself experiences software failure or logical error, the possibility of it causing adverse effects on the existing automation system is controlled to an extremely low level.

[0063] This invention also includes a feedback optimization loop to achieve continuous self-improvement of the running agent. The execution logic of the feedback optimization is as follows: periodically (preferably with an analysis cycle of 7 days) statistically analyzes the execution results recorded in the structured log and calculates the execution success rate of each running strategy. The execution success rate is defined as: among all triggered executions (including automatic execution and execution after manual confirmation), the proportion of the running status that improves in the expected direction after execution. In one embodiment of this invention, the improvement judgment criterion is that the target parameter changes in the expected direction within 30 minutes after execution and the change exceeds a preset minimum improvement threshold.

[0064] When the success rate of a certain strategy falls below a preset threshold (e.g., below 0.70), the parameters of that strategy are automatically adjusted: if the success rate is between 0.50 and 0.70, the applicable operating conditions range of the strategy is narrowed to improve matching accuracy; if the success rate is below 0.50, the risk level rating of the strategy is increased by one level to increase execution prudence. Simultaneously, execution deviation data (i.e., the difference between the actual execution effect and the expected effect) is used as additional observation input for the Bayesian posterior update and fed back into the safe operating boundary model. This closed-loop mechanism allows the safety boundary and decision-making strategy to be continuously optimized with the accumulation of operational experience, reflecting the design concept of progressive intelligent upgrade.

[0065] In a preferred embodiment of the present invention, the feedback optimization loop further includes an automatic expansion mechanism for the strategy library. When the operating agent frequently encounters a certain type of operating condition over a long operating period (e.g., 30 consecutive days) but the strategy library lacks a highly matching strategy, the system automatically extracts the operating condition mode and the corresponding actual operation records of the operators as candidate strategy entries. After review and confirmation by the operators, these entries are added to the strategy library. This mechanism allows the strategy library to be gradually enriched and improved as the power plant's operating experience accumulates, thereby continuously improving the coverage and relevance of the operating recommendations.

[0066] Furthermore, the operational data in the feedback optimization loop can also be used for continuous calibration of the operating condition identification model. When the structured log records the operator's correction operation for a certain operating condition label (e.g., the operator believes that the current operating condition should be a regulating operation rather than the steady-state operation identified by the system), this correction information is fed back as a labeled training sample to the transition threshold optimization module of the microstate machine. By statistically analyzing the accumulated correction samples, the parameter threshold values ​​in the state transition conditions are automatically fine-tuned, so that the accuracy of operating condition identification gradually improves during continuous operation. This complete closed loop from identification to decision-making to feedback enables the operating agent to possess the ability to accumulate experience and improve cognition similar to human operators, which is the core embodiment of the progressive intelligence concept of this invention.

[0067] To verify the practical effectiveness of the method of the present invention, a six-month field deployment test was conducted at the aforementioned run-of-river small hydropower station with an installed capacity of 2×6.3MW. During the test, the existing automation system maintained its original operating mode, while the operating intelligent agent operated in parallel in a bypass mode.

[0068] Regarding operational condition identification, approximately 1.56 × 10⁻⁶ valid operational data were collected during the testing period. 8 The data points (calculated based on 480 points, a 500ms sampling period, and 6 months of continuous operation) show that the operational condition identification module has cumulatively output approximately 2.6 × 10⁻⁶ data points. 5The operating condition label was updated. Subsequent verification by operations personnel showed that the accuracy rate of operating condition identification reached 96.3%, with an accuracy rate of 98.7% for grid-connected steady-state operating conditions and 92.1% for start-stop transition operating conditions. Operating condition misjudgments mainly occurred at the boundary states between regulated operation and steady-state operation. After adjusting the transition threshold of the microstate machine, the misjudgment rate of boundary states decreased by approximately 40%.

[0069] Regarding safety boundary determination, the dynamic safety boundary model triggered 47 critical zone alarms and 3 prohibited zone alarms during the testing period. Post-analysis revealed that 42 of the 47 critical zone alarms were reasonable alarms (corresponding to actual abnormal parameter combinations and shifts), while 5 were excessive alarms caused by sudden hydrological changes, resulting in a reasonable alarm rate of 89.4%. The Bayesian posterior update mechanism performed approximately 180 boundary parameter updates during the 6-month operation period. The updated safety benchmark center drifted by approximately 0.3m in the water level deviation dimension compared to the initial value (consistent with the seasonal variation trend of reservoir water level during this period), validating the effectiveness of the online calibration mechanism.

[0070] Regarding operational decision-making, the policy library triggered a total of 1247 operational suggestions during the testing period, including 823 automatically executed suggestions within the safe zone, 382 manually confirmed suggestions within the critical zone, and only 42 suggestions recorded within the prohibited zone. The success rate of automatically executed suggestions was 91.2%, and the adoption rate of manually confirmed suggestions by operators was 76.4%. In terms of feedback optimization, after three optimization cycles, the average condition satisfaction rate of policy matching increased from the initial 0.72 to 0.81, and the execution success rate increased from the initial 88.5% to 91.2%.

[0071] Regarding the interpretability of decisions, 100 operational recommendations were randomly selected for manual review. Among them, 94 recommendations were rated as clear and well-founded by the operators, 6 recommendations were rated as complete but some descriptions could be further simplified, and no errors in the explanations or missing key information were found.

[0072] In summary, the test results show that the method of the present invention can provide small hydropower stations with reliable capabilities for identifying operating conditions, determining safety boundaries, generating operating decisions, and recording decision interpretations without any modification to the existing automation system. All indicators have met the expected design goals, verifying the feasibility and effectiveness of the incremental intelligent upgrade scheme with zero modification, low risk, and high reliability.

[0073] It should be further noted that during the entire test deployment process, the running agent did not cause any interference to the existing automation system. The control logic of the existing programmable logic controller, the operating characteristics of the protection devices, and the human-machine interface functions of the data acquisition and monitoring control system remained completely unchanged. The additional communication load generated by the running agent's read-only access to the communication bus accounted for approximately 3% to 5% of the bus bandwidth, far below the design redundancy capacity of the communication bus, and therefore did not affect the real-time performance of the existing system. Furthermore, when the running agent required maintenance due to its own software failure or hardware malfunction, it could be directly disconnected from the communication network without affecting the operation of the existing automation system, further verifying the security advantages of the logic bypass deployment method.

[0074] In terms of economic benefits, compared with traditional overall renovation schemes, the implementation cost of the method of this invention is approximately 15% to 20% of that of the overall renovation scheme, the deployment cycle is shortened from more than 3 months to approximately 2 weeks, and no downtime is required throughout the process, avoiding the loss of power generation revenue during downtime. Taking the power plant in this embodiment as an example, during the 6-month test operation period, through optimized scheduling based on operational recommendations, the hydropower utilization rate increased by approximately 2.3% compared to before deployment, which translates to an annual power generation gain of approximately 120 MWh. At the same time, due to the introduction of operating condition identification and safety boundary determination functions, the early warning of critical operating conditions reduces the number of unplanned downtimes by approximately 35%, and equipment maintenance costs are correspondingly reduced.

[0075] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A method for constructing an intelligent agent for the operation of a small hydropower station based on an existing automation system, characterized in that: include: Operational data is collected from the existing automation system of small hydropower stations through standard communication protocols. The collected operational data is processed by semantic mapping of points and tables. The original data points are associated with unit objects, hydraulic objects and operation status objects according to preset ontology rules, and a three-dimensional semantic model of unit-hydraulic-state is constructed. Based on the semantic mapping of the operation data according to the three-dimensional semantic model, the semantic operation data output is used to construct a hierarchical operation state machine model composed of an outer macro state machine and an inner micro state machine. The change trend of the operation parameters is extracted within the sliding time window, and the current operation stage of the unit is identified by combining the state transition rules. The corresponding operation condition labels are generated based on the water level change rate, output change rate and regulation behavior. A dynamic safe operation boundary model is constructed based on multidimensional operating parameters. By calculating the weighted distance metric between the current operating parameter vector composed of safety-related parameters selected from the multidimensional operating parameters and the safety benchmark center, the unit's operating status is divided into three safety levels: safe zone, critical zone, and prohibited automatic operation zone. The boundary parameters of the safe operation boundary model are updated using Bayesian posterior time based on the statistical distribution of historical operating data. Under the joint constraints of the operating condition label and the security level determination result, the strategy matching is performed through the rule-based operating strategy library to generate an operating suggestion containing suggested content, risk level and execution permission identifier, wherein the execution permission identifier is directly determined by the security level; The operation suggestions are used to generate structured decision explanation information, which includes at least a description of the triggering condition, a matching rule identifier, a data basis summary, and a risk assessment description. The operation conditions, operation suggestions, execution results, and decision explanation information are stored in a structured log indexed by timestamps, and the operation suggestions are output in a hierarchical and controlled manner according to the security level.

2. The method according to claim 1, characterized in that, Each semantic entity in the three-dimensional semantic model is expressed in the form of a subject-verb-object triple, and the physical quantity type, range, and sampling frequency corresponding to the semantic entity are recorded. The operating data includes upstream water level, downstream water level, inflow rate, unit active power, unit speed, bearing vibration amplitude, stator winding temperature, guide vane opening, and gate opening. The sampling period for the upstream water level is 1s to 5s, and the sampling period for the bearing vibration amplitude is 0.1s to 1s.

3. The method according to claim 1, characterized in that, The weighted distance metric is calculated using Mahalanobis distance. The distance threshold for the safe zone is set to be within 2 times the standard deviation of the historical distribution of the operating parameters. The distance threshold for the critical zone is set to be between 2 and 3 times the standard deviation. Areas exceeding 3 times the standard deviation are identified as prohibited automatic operation zones.

4. The method according to claim 1, characterized in that, The window length of the sliding time window is adaptively adjusted according to the operating condition label. Under steady-state operating conditions, the window length is set to 60s to 300s. Under adjustment operating conditions, the window length is shortened to 10s to 60s. Under unit start-up and shutdown transition conditions, the window length is set to 5s to 30s.

5. The method according to claim 1, characterized in that, The outer macroscopic state machine includes shutdown state, start-up transition state, grid-connected power generation state, regulation operation state, and shutdown transition state. The state transition conditions are jointly determined by the unit circuit breaker signal, speed threshold value, and power threshold value. The inner microscopic state machine further divides each macroscopic state into sub-states. The transition of the sub-states is jointly triggered by the changing trend and duration of the operating parameters.

6. The method according to claim 1, characterized in that, The boundary parameters of the dynamic safe operation boundary model are calibrated online using a Bayesian posterior update method. The statistical distribution of historical operation data is used as the prior distribution, and the operation data within the most recent preset time period is used as the likelihood function input. The mean and covariance matrix of the posterior distribution are calculated as the updated safety benchmark center and boundary parameters.

7. The method according to claim 1, characterized in that, The strategy matching process includes: comparing the current operating condition label with the applicable operating condition conditions of each strategy in the strategy library item by item, calculating the condition satisfaction score, selecting the strategy with the highest condition satisfaction score and the risk level not exceeding the upper limit allowed by the current security level as the final matching strategy, and when there are multiple candidate strategies with the same condition satisfaction score, the strategy with the lower risk level is selected first.

8. The method according to claim 1, characterized in that, The unit objects in the three-dimensional semantic model include unit number, unit type, rated parameters, and current operating parameters. The hydraulic objects include water level and flow parameters of reservoirs, water diversion channels, pressure pipelines, and tailrace channels. The operating status objects include current operating condition labels, safety level, and cumulative running time.

9. The method according to claim 8, characterized in that, Operation suggestions within the safe zone are automatically sent to existing automated systems for execution; operation suggestions within the critical zone can only be executed after manual confirmation; and automatic operation zones are prohibited from merely recording traces without outputting execution instructions. Each record in the structured log includes a timestamp, operating condition tag, safety level, operation suggestion content, decision explanation information, execution method, and execution result. The structured log supports multi-dimensional retrieval and statistical analysis by time range, operating condition type, and safety level.

10. The method according to claim 9, characterized in that, It also includes a feedback optimization step: periodically performing statistical analysis on the execution results recorded in the structured log, calculating the execution success rate of each operating strategy, and automatically adjusting the applicable working condition range or risk level calibration value of the strategy when the execution success rate of a certain strategy is lower than a preset threshold. At the same time, the execution deviation data is used as an additional observation input for the Bayesian posterior update to achieve adaptive calibration of the safe operation boundary model.

Images