Hydropower station unit state on-line monitoring system

Abstract

The invention discloses a hydropower station unit state online monitoring system, relates to the technical field of hydropower station unit monitoring control, and adopts a hybrid digital twin modeling technology combining a physical mechanism main model and a liquid neural network residual compensation model to construct a high-fidelity unit operation state model. The system comprises a data acquisition module, a digital twin modeling module, a fault evolution prediction module, a multi-target optimization module and an adaptive control generation module. Multi-source heterogeneous data fusion is realized through a space-time adaptive weight distribution algorithm, and residual compensation modeling is performed by using dynamic time constant characteristics of a liquid neural network. Virtual fault injection and fault evolution trajectory prediction are realized, and passive fault response is converted into active fault prediction. A two-stage optimization strategy is adopted to realize'safety-efficiency-life 'three-dimensional target collaborative optimization, and a continuous and smooth adaptive control parameter trajectory is generated by controlling a liquid neural network. The modeling precision is improved, and the fault early warning time is advanced.

Description

Technical Field

[0001] The invention relates to the technical field of monitoring and control of hydropower station units, in particular to an online monitoring system for the status of hydropower station units. Background Art

[0002] As an important clean energy source, hydropower occupies a crucial position in the nation's energy mix. As the core equipment of hydropower systems, the operating status of hydropower station units directly impacts power generation efficiency, system safety, and equipment lifespan. With the continued growth of hydropower installed capacity and the increasing demand for power supply reliability from the power grid, higher requirements are being placed on hydropower station unit condition monitoring technology.

[0003] Traditional hydropower station unit monitoring systems are primarily based on single physical models or data-driven approaches. Physical modeling approaches describe the unit's operating characteristics by building a mathematical model. However, due to the complex nature of hydropower units, the coupling of multiple physical fields, and the strong nonlinearity, traditional physical models often require numerous simplifying assumptions, resulting in low modeling accuracy, often with a 10-20% modeling error, making it difficult to accurately reflect the unit's actual operating status. While data-driven approaches can effectively fit historical data, they lack the support of physical mechanisms, suffer from "black box" issues, have poor interpretability, and lack adaptability to changing operating conditions.

[0004] Existing technologies manage unit status through traditional threshold alarms and regular maintenance strategies, but there are still certain limitations. For example, fault detection is a passive response mode and can only issue an alarm after a fault occurs, making it impossible to predict and prevent faults in advance; the monitoring system lacks in-depth analysis of the fault evolution process and cannot predict the development trend and cascading impact of faults; the optimization objectives are single, usually only focusing on safety indicators, and failing to achieve coordinated optimization of multi-dimensional objectives such as safety, efficiency, and lifespan; the control strategy is relatively fixed and lacks the ability to adaptively adjust to changes in operating conditions; the human-computer interaction interface is simple and the decision support function is limited, which cannot fully leverage the synergy between human experience and intelligent analysis.

[0005] Therefore, there is an urgent need for an intelligent online monitoring system for hydropower station unit status that can achieve high-precision modeling, fault evolution prediction, multi-objective collaborative optimization and human-machine collaborative decision-making to meet the technical requirements for safe and efficient operation of modern hydropower stations. Summary of the Invention

[0006] The purpose of the present invention is to overcome the deficiencies of the prior art and to provide an online monitoring system for the status of hydropower station units to solve the above-mentioned problems.

[0007] The object of the present invention is achieved through the following technical solutions: an online monitoring system for the status of a hydropower station unit, comprising a data acquisition module, a digital twin modeling module, a fault evolution prediction module, a multi-objective optimization module and an adaptive control generation module; The data acquisition module is used to collect real-time monitoring data of the hydropower units; The digital twin modeling module is configured to build and run a hybrid digital twin model. The hybrid digital twin model consists of a physical mechanism main model and a liquid neural network residual compensation model. The physical mechanism main model outputs a first predicted state based on the unit's rotor dynamics, hydraulics, and electromagnetic field coupling mechanisms. The liquid neural network residual compensation model dynamically models the residual sequence between the first predicted state and real-time monitoring data and outputs a residual compensation value. By fusing the first predicted state and the residual compensation value, a high-fidelity final state of the unit is generated. The fault evolution prediction module is configured to perform online virtual fault injection based on the hybrid digital twin model and use the liquid neural network residual compensation model to deduce and predict the future evolution path, development speed and cross-system cascading impact of the injected fault; The multi-objective optimization module is configured to perform collaborative optimization calculations in a multi-dimensional decision space based on the fault evolution prediction results, with the optimization objectives being unit operational safety, power generation efficiency, and service life of key components, to obtain the optimal operation strategy with the best overall unit benefits. The adaptive control generation module is configured to generate and output an adaptive control parameter trajectory that continuously changes over time for the unit speed governor system or the excitation system based on the optimal operation strategy.

[0008] The liquid neural network residual compensation model is a continuous-time recurrent neural network defined by ordinary differential equations. Its neurons have dynamically changing nonlinear time constants determined by the current input and hidden state, enabling it to accurately capture complex time-varying dynamics that are not fully explained by the main model of the physical mechanism.

[0009] The main physical mechanism model is a multi-physics field coupling model, which includes a rotor dynamics sub-model for describing the vibration and swing of the main shaft, a fluid dynamics sub-model for describing the pressure pulsation of the volute, a fluid dynamics sub-model for describing the pressure pulsation of the top cover, a fluid dynamics sub-model for describing the pressure pulsation of the draft tube, and an electromagnetic field sub-model for describing the electromagnetic force caused by the uneven air gap between the stator and rotor.

[0010] The fault evolution prediction module specifically performs the following steps: injecting a virtual fault initial disturbance into the hybrid digital twin model calibrated with real-time data, using the nonlinear dynamic transfer characteristics of the liquid neural network residual compensation model to deduce the evolution trajectory of the disturbance, and generating a predictive alarm. The predictive alarm is generated in advance when it is deduced that the future value of a certain monitoring parameter will reach its preset alarm limit.

[0011] The multi-objective optimization module adopts a two-stage optimization strategy: in the first stage, a lightweight liquid neural network trained as a hybrid digital twin model proxy model is used to perform a rough search of the large-scale global parameter space; in the second stage, the complete hybrid digital twin model is called within the candidate solution interval determined by the rough search to perform precise evaluation and determine the final optimal operation strategy.

[0012] The operational safety target in the multi-objective optimization module is quantified as the safety margin between the key monitoring parameters and their alarm thresholds. The power generation efficiency target is quantified as the product of the unit's power generation and water energy utilization efficiency. The service life target of key components is quantified as the inverse of the expected fatigue damage accumulation rate of key components calculated based on the fault evolution prediction results.

[0013] The adaptive control generation module includes a dedicated control liquid neural network, which takes the real-time status of the unit and the optimization target as input and outputs a continuous and smooth control parameter adjustment trajectory. The control parameter adjustment trajectory is specifically manifested as a recommended adjustment curve for the proportional-integral-derivative parameters of the speed regulator system, or a recommended adjustment curve for the voltage set value and current set value of the excitation system.

[0014] The real-time monitoring data collected by the data acquisition module include: vibration data of the upper frame of the unit, vibration data of the lower frame of the unit, swing data of the upper guide bearing, swing data of the lower guide bearing, swing data of the water guide bearing, pressure pulsation data of the volute, pressure pulsation data of the top cover, pressure pulsation data of the tailwater pipe, and air gap data of the stator and rotor of the generator.

[0015] The liquid neural network residual compensation model is trained through learning so that the residual compensation value it outputs can bridge the dynamic error between the prediction of the physical mechanism main model and the real-time monitoring data to the greatest extent, thereby achieving the optimal fit between the final state of the unit and the real-time monitoring data.

[0016] It also includes a human-computer interaction module, which presents the control parameter trajectory generated by the adaptive control generation module on the human-computer interaction interface for the operating personnel to conduct final review and confirmation. After obtaining the confirmation instruction, the corrected parameters will be sent to the actual control system of the unit, thus forming a closed-loop intelligent decision-making process in which people are in the loop.

[0017] The beneficial effects of the present invention are: The present invention adopts hybrid digital twin modeling technology, and solves the technical bottlenecks of traditional single modeling methods in terms of accuracy, robustness and interpretability by organically combining the main model of the physical mechanism with the liquid neural network residual compensation model. Traditional single physical models usually have a modeling error of 10-20% due to theoretical simplification and parameter uncertainty, while the present invention improves the modeling accuracy to more than 95%, which is 30-50% higher than the traditional method. The dynamic time constant characteristics of the liquid neural network enable the system to automatically adapt to changes in operating conditions, which is more than 60% more adaptable than the fixed parameter model. It can still maintain stable high-precision prediction capabilities under complex conditions such as operating parameter fluctuations and changes in environmental conditions. At the same time, the present invention maintains the interpretability of the physical model, avoids the "black box" problem of pure data-driven methods, and provides convenience for engineers to understand system behavior and failure mechanisms.

[0018] The present invention has achieved a technological breakthrough in fault evolution prediction, fundamentally changing the passive response mode of traditional monitoring systems. Traditional monitoring systems can only provide alarms after a fault occurs, but the present invention, through virtual fault injection technology and fault evolution trajectory deduction, can predict the occurrence of a fault 6-24 hours in advance, with a prediction accuracy of over 85%. Virtual fault injection technology can simulate various fault scenarios without affecting actual operation, and deduce the disturbance evolution trajectory through the nonlinear dynamic transfer characteristics of the liquid neural network residual compensation model. The system can not only predict the development of a single fault, but also analyze the cascading propagation and cross-system impact of the fault through the Lyapunov index algorithm and graph theory propagation model, realize the quantitative evaluation of the fault development speed and impact range, and provide a scientific basis for the comprehensive prevention and control of complex faults.

[0019] The present invention realizes the coordinated optimization of the three-dimensional goals of "safety-efficiency-lifespan", breaking through the limitations of the single safety goal of traditional monitoring systems. Through a two-stage optimization strategy, the first stage uses a lightweight liquid neural network proxy model to conduct a large-scale global parameter space search, and the second stage calls a complete hybrid digital twin model for precise evaluation. While ensuring computational efficiency, it realizes the search for the global optimal solution, and the optimization efficiency is improved by more than 40%. The time-varying weight adjustment algorithm can dynamically adjust the importance of the optimization target according to the actual operating status, and the comprehensive benefit is improved by more than 25% compared with the fixed weight method. The Pareto dynamic maintenance algorithm ensures the quality and diversity of multi-objective optimization solutions.

[0020] This invention achieves human-machine collaborative decision-making, leveraging the advantages of artificial intelligence analysis and human experience, while avoiding the risks associated with full automation. The continuous, smooth control trajectory generated by the liquid neural network avoids the impact of sudden parameter changes on the equipment, extending its service life by 10-15%. Continuity regularization constraints ensure smooth control and system stability, improving system stability by over 30% compared to traditional control methods.

[0021] The fault evolution prediction function can provide early warning 6-24 hours before an accident occurs, avoiding major safety incidents. Multi-parameter collaborative safety margin assessment improves safety by over 50% compared to traditional single-parameter monitoring. Cross-system cascading impact analysis can effectively prevent small faults from evolving into systemic faults, avoiding serious consequences such as large-scale power outages. The continuous smoothing control strategy avoids sudden changes in control parameters, and multi-objective collaborative optimization ensures stable operation of the system within the safe zone. The adaptive characteristics of the liquid neural network make the system more resistant to external interference, and the hybrid modeling technology improves the system's robustness to parameter uncertainty and measurement noise.

[0022] Precise monitoring and optimized control of the operating status of hydropower units improves grid stability and reliability, reduces the impact of unplanned outages on the grid, and ensures secure power supply. Intelligent control strategies improve power quality, reduce voltage fluctuations and frequency deviations, and provide users with a more stable, high-quality power supply. This system provides a model and a guide for the intelligent transformation of the hydropower industry, promotes the intelligent and high-end development of the hydropower equipment manufacturing industry, and creates new market opportunities for related software development, system integration, and other industries. The application of the system promotes the cultivation of professionals in hydropower operations and maintenance, intelligent control, and other fields, providing a practical platform for teaching and research in related fields at universities. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 The system architecture of the present invention Figure 1 ; Figure 2 The system architecture of the present invention Figure 2 ; Figure 3 This is the system architecture diagram of the present invention. DETAILED DESCRIPTION

[0024] The following will clearly and completely describe the technical solutions of the present invention in conjunction with the embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work shall fall within the scope of protection of the present invention.

[0025] It is to be noted that the directions of "left", "right", "up", "down", "front", "back", "inside" and "outside" in the following schemes are all relative directions and are not listed here one by one.

[0026] Example 1: like Figures 1 to 3 As shown, this embodiment provides an online monitoring system for the status of hydropower station units. The system constructs a high-fidelity hybrid digital twin model by integrating physical mechanism modeling and liquid neural network technology, thereby realizing accurate modeling and real-time monitoring of the operating status of hydropower units.

[0027] The monitoring system of this embodiment includes a data acquisition module, a digital twin modeling module, a fault evolution prediction module, a multi-objective optimization module, and an adaptive control generation module. These modules work together to form a complete unit intelligent monitoring system.

[0028] The data acquisition module is responsible for collecting real-time operational monitoring data from the hydropower unit. Equipped with a variety of sensors and data acquisition equipment, it simultaneously collects vibration data from the unit's upper and lower frames, swing data from the upper and lower guide bearings, swing data from the water guide bearings, pressure pulsation data from the volute, top cover, and draft tube, as well as data on the stator and rotor air gaps between the generators. This monitoring data covers the unit's key operating parameters and provides a comprehensive data foundation for subsequent digital twin modeling. The data acquisition module utilizes high-precision acquisition equipment to ensure that the collected data has sufficient accuracy and time resolution to reflect subtle changes in the unit's operating status.

[0029] To ensure the effective integration of multi-source heterogeneous monitoring data, the data acquisition module uses a spatiotemporal adaptive weight allocation algorithm for data preprocessing. The algorithm calculates the dynamic weight of each monitoring point using the following formula: ; : The weight of the i-th monitoring point at time t, : The real-time noise standard deviation of the i-th monitoring point at time t, : The reliability factor of the i-th monitoring point at time t calculated based on historical data, : The spatial position vector of the i-th monitoring point, : The spatial position vector of the center of the key monitoring area, : spatial correlation attenuation parameter, N: total number of monitoring points, j: summation index variable; The integrated monitoring data after fusion is expressed as: ; : The integrated monitoring data vector after fusion at time t, : The raw data vector of the i-th sensor at time t, : The i-th sensor is based on the sampling frequency f s The time synchronization correction factor, : sampling frequency, i: sensor index; To ensure the effective integration of multi-source heterogeneous monitoring data, the data acquisition module uses a spatiotemporal adaptive weight allocation algorithm for data preprocessing. The algorithm calculates the dynamic weight of each monitoring point using the following formula: The digital twin modeling module builds and runs a hybrid digital twin model. This hybrid digital twin model utilizes a dual modeling architecture, consisting of a primary physical mechanism model and a liquid neural network residual compensation model. This architecture leverages the interpretability of physical modeling and the data fitting capabilities of neural networks, addressing the limitations of traditional single modeling approaches in terms of accuracy, robustness, and interpretability.

[0030] The main physical mechanism model is a multi-physics coupling model that describes the unit's basic operating principles based on the unit's rotordynamics, hydraulics, and electromagnetic field coupling mechanisms. Specifically, this multi-physics coupling model includes a rotordynamics sub-model for describing main shaft vibration and runout, a fluid dynamics sub-model for describing volute pressure pulsation, a fluid dynamics sub-model for describing top cover pressure pulsation, a fluid dynamics sub-model for describing draft tube pressure pulsation, and an electromagnetic field sub-model for describing the electromagnetic force caused by the uneven stator-rotor air gap.

[0031] The rotor dynamics submodel uses the distributed mass-spring-damper coupling equation to describe the dynamic characteristics of the spindle system: : speed-dependent mass matrix, : Damping matrix considering hydraulic damping effect, : displacement-dependent nonlinear stiffness matrix, : rotor displacement vector, : rotor speed vector (first derivative of the displacement vector), : rotor acceleration vector (second derivative of displacement vector), : hydraulic excitation force vector, : electromagnetic excitation force vector, : unbalanced force vector, n: rotor speed; The fluid dynamics submodel is based on a modified three-dimensional incompressible Navier-Stokes equations that takes into account turbulence effects and interactions with solid boundaries: ; : fluid velocity field vector, t: time, p: pressure field, : fluid density, : kinematic viscosity, : gradient operator, : Laplace operator, :based on The turbulent force term of the model, k: turbulent kinetic energy, : turbulent dissipation rate, : fluid-solid coupling force term, : Rotor displacement vector; The prediction of pressure pulsation is achieved through the following equation:

[0032] : pressure pulsation value at time t, K: total number of harmonics, k: harmonic order index, : the amplitude of the kth harmonic, : the phase of the kth harmonic, :Frequency conversion, : spatial attenuation coefficient of the kth harmonic, : spatial position vector, : leaf position vector, : runner radius, : vector norm; The electromagnetic field sub-model establishes an air gap magnetic field distribution model based on Maxwell's equations, taking into account the influence of the relative position change of the stator and rotor: , : magnetic field strength vector, : electric field intensity vector, : magnetic induction intensity vector, : electric displacement vector, : current density vector, : curl operator; The distribution of air gap magnetic induction intensity adopts the improved Fourier series expansion: ; :angle and the air gap magnetic induction intensity at time t, : spatial angle coordinate, : average magnetic induction intensity, N: number of Fourier series expansion terms, n: harmonic order index, : the amplitude of the nth harmonic at time t, : the phase angle of the nth harmonic, : air gap unevenness correction factor, :angle and the real-time air gap length at time t; where, is the average magnetic induction intensity, For the Subharmonic amplitude, is the phase angle, and the air gap non-uniformity correction factor is defined as: , : Design air gap length, : Air gap correction parameter, : air gap correction index parameter; These sub-models are coupled to each other, and a coupled iterative algorithm is used to achieve a collaborative solution of multiple physical fields. Based on these coupled mechanisms, the main physical mechanism model outputs a first predicted state, which includes various unit operating parameters predicted based on physical principles.

[0033] However, due to the complexity of actual engineering systems, it is difficult for the main physical mechanism model to fully and accurately describe all dynamic characteristics, especially some nonlinear and time-varying complex phenomena. To this end, this embodiment introduces a liquid neural network residual compensation model to make up for the shortcomings of the physical model. The liquid neural network residual compensation model is a continuous-time recurrent neural network defined by ordinary differential equations, which has unique dynamic characteristics. Its neurons have a nonlinear time constant that can be dynamically changed and is determined by the current input and hidden state. This design enables the network to accurately capture the complex time-varying dynamics that are not fully explained by the main physical mechanism model.

[0034] The dynamic behavior of the liquid neural network is described by the following set of differential equations: ; : The state of the i-th neuron at time t, : the dynamic time constant of the ith neuron at time t, : the time derivative of the i-th neuron state, N: the total number of neurons in the neural network, M: the total number of external inputs, : The connection weight from the jth neuron to the i-th neuron, : The input weight of the kth external input to the i-th neuron, : neuron activation function, : Input processing function, : the value of the kth external input at time t, : random noise term of the i-th neuron at time t, i: neuron index, j: neuron index in the summation, k: input index; The dynamic time constant is updated via the following adaptive mechanism: ; : basic time constant, : Adaptive adjustment parameters, : weight sensitivity parameter, : Input sensitivity parameter, : hyperbolic tangent function, : the absolute value of the connection weight, : the absolute value of the j-th neuron state, : total input intensity at time t, : The absolute value of the total input intensity; the activation function adopts a piecewise continuous adaptive form: ; : piecewise continuous activation function, x: input variable of the activation function, : Gain parameter in the linear region, : Gain parameter in the saturation region, : Gain parameter in the nonlinear region, : The first threshold parameter (upper limit of the linear region), : The second threshold parameter (upper limit of saturation region), : saturation parameter of the hyperbolic tangent function, : Sign function, when x>0 it is 1, when x<0 it is -1, : Square root of the absolute value of the input, : absolute value of input; The time constant of the liquid neural network is not fixed, but is dynamically adjusted according to the input signal and the internal state of the network. This adaptability enables it to better adapt to changes in the system dynamics.

[0035] The liquid neural network residual compensation model is specifically designed to dynamically model the residual sequence between the first predicted state and the real-time monitoring data. The model analyzes the difference between the output of the physical mechanism master model and the actual monitoring data, learns the dynamic laws of this difference, and outputs the corresponding residual compensation value. The dynamic modeling of the residual sequence uses the following algorithm: , : real-time residual vector at time t, : The actual monitoring data vector at time t, : The physical model prediction value vector at time t; ; : The residual compensation value vector output by the liquid neural network at time t, : the number of neurons in the liquid neural network, : The output weight of the i-th neuron at time t, : the state vector of the i-th neuron at time t, i: neuron index; This learning process is continuous, allowing the model to continuously adapt to changes in system characteristics. Weight updates use an improved real-time recursive least squares algorithm: ;

[0036] : covariance matrix at time t, : The updated covariance matrix at time t+1, : forgetting factor (0<λ≤1), : input vector at time t, : transpose of the input vector, : scalar value; : weight vector at time t, : updated weight vector at time t+1, : The prediction error at time t, defined as

[0037] Through learning and training, the liquid neural network residual compensation model can output residual compensation values ​​that can bridge the dynamic errors between the predictions of the physical mechanism main model and the real-time monitoring data to the greatest extent.

[0038] The hybrid digital twin model generates a high-fidelity final state of the unit by fusing the first predicted state with the residual compensation value. Specifically, the final state of the unit is calculated using the following adaptive weighted fusion algorithm: : The final state vector of the unit at time t, : dynamic weight vector at time t, : all-1 vector, : element-wise multiplication (Hadamard product), : the complement of the weight vector; Dynamic weights are calculated using a confidence evaluation algorithm:

[0039] : The dynamic weight of the i-th parameter at time t, : Adjustment parameters, : exponential function, : The prediction uncertainty (standard deviation) of the physical model for the i-th parameter at time t, : The uncertainty (standard deviation) in the prediction of the residual model for the i-th parameter at time t, : the inverse square of the uncertainty of the physical model prediction (precision), : The inverse square of the uncertainty of the residual model prediction (precision); This fusion approach preserves the interpretability of the physical model while improving prediction accuracy through neural network compensation. The fit between the final state of the unit and the real-time monitoring data is quantified by the following indicators:

[0040] is the mean of historical monitoring data. When this indicator reaches the optimal value, the hybrid model can more accurately reflect the actual operating status of the unit. : the fitness score at time t, : L2 norm (Euclidean norm), : The square of the L2 norm.

[0041] The Fault Evolution Prediction module, based on a hybrid digital twin model, leverages its high-fidelity modeling capabilities to enable predictive analysis of unit faults. This module performs online virtual fault injection based on the hybrid digital twin model and uses a liquid neural network residual compensation model to deduce and predict the future evolution path, development speed, and cross-system cascading impact of the injected fault.

[0042] Based on the fault evolution prediction results, the multi-objective optimization module performs collaborative optimization calculations within a multidimensional decision space, optimizing unit safety, power generation efficiency, and the service life of key components. This module determines the optimal operating strategy for the best overall unit benefits. This module balances multiple competing objectives to find the most optimal operating parameters for the current operating conditions.

[0043] Based on the optimal operating strategy, the adaptive control generation module generates and outputs adaptive control parameter trajectories for the unit's speed governor or excitation system, which continuously change over time. These control parameter trajectories guide the unit's actual control and achieve intelligent operational optimization.

[0044] Through the collaborative operation of the aforementioned modules, this embodiment implements a high-precision unit status monitoring system based on a hybrid digital twin model. This system not only accurately reflects the unit's current operating status but also provides a reliable technical foundation for fault prediction, operational optimization, and intelligent control. The adoption of a hybrid modeling approach significantly improves modeling accuracy and adaptability while maintaining physical interpretability, providing an effective technical solution for the intelligent operation and maintenance of hydropower units.

[0045] Example 2: like Figures 1 to 3 As shown, based on the first embodiment, this embodiment particularly strengthens the function of the fault evolution prediction module to realize the evolution path prediction function of the unit fault, fundamentally changes the traditional passive fault response mode, and realizes the technological leap from "post-diagnosis" to "pre-prediction", providing advanced technical means for predictive maintenance of hydropower units.

[0046] Building on the data acquisition module and digital twin modeling module described in Example 1, the monitoring system of this embodiment specifically enhances the fault evolution prediction module, making it the core highlight of the entire system. The structure and functions of the data acquisition module and digital twin modeling module are the same as those in Example 1, providing comprehensive data support and a high-fidelity modeling foundation for fault evolution prediction.

[0047] The fault evolution prediction module leverages the high-fidelity modeling capabilities of the hybrid digital twin model described in Example 1 to accurately predict the fault evolution path of the unit. The module performs a series of advanced predictive analysis steps that form the complete fault evolution prediction process.

[0048] First, the fault evolution prediction module injects virtual fault initial disturbances into the hybrid digital twin model, which has been calibrated with real-time data. This virtual fault injection allows the system to simulate various possible fault scenarios without affecting actual unit operation. These virtual fault initial disturbances can include a small increase in bearing clearance, an increase in rotor imbalance, a change in the surface roughness of a hydraulic component, or a slight deviation in the electromagnetic field distribution.

[0049] Virtual fault injection uses a multi-level disturbance superposition algorithm, which generates a composite fault disturbance scenario using the following formula:

[0050] in, : the comprehensive virtual fault vector at time t, : number of fault types, i: fault type index, : The time-varying activation weight vector of the i-th fault at time t, : intensity distribution function of type i fault, : location parameter vector of type i fault, : intensity parameter of the i-th type fault, : The fault mapping matrix of the i-th type fault based on the current operating state, : Current benchmark running state vector; The fault intensity distribution function adopts an adaptive Gaussian mixture model: ; : the number of Gaussian components of the i-th type fault, k: Gaussian component index, : The weight of the kth Gaussian component in the i-th type of fault, :The mean is , the covariance is The time-varying Gaussian distribution of : The covariance matrix of the kth Gaussian component in the i-th type fault at time t, : The decay function of the distance from the key position, : The distance between the i-th fault location and the critical location, : key position vector; The time-varying activation weights are dynamically adjusted through the failure probability assessment algorithm: ; : hyperbolic tangent function, : sensitivity coefficient of type i fault, : recent residual vector at time t, : historical residual basis vector, : L2 norm of the recent residual vector, : L2 norm of the historical residual basis vector, : exponential function, : the time of last fault injection, : cooling time constant, t: current moment; Although these disturbances are small, they can gradually amplify and cause serious failures in complex unit systems. By injecting these virtual disturbances into the digital twin model, the system can preemptively assess the development trends of various potential failures.

[0051] Next, the fault evolution prediction module uses the nonlinear dynamic transfer characteristics of the liquid neural network residual compensation model to deduce the evolution trajectory of the disturbance. Due to its unique dynamic time constant characteristics, the liquid neural network residual compensation model can keenly capture and amplify the propagation of small disturbances in the system.

[0052] The evolution trajectory deduction adopts an improved dynamic system analysis method, whose core equation is: in, : The time derivative of the fault state vector (evolution rate), : fault state vector at time t, : fault evolution function, : baseline running state vector, t: time variable, : coupling strength matrix depending on the fault state, : Nonlinear transfer function vector of liquid neural network; The fault evolution function is approximated by piecewise linearization: ; Where, J: total number of segmented regions, j: segmented region index, : the time-varying weight coefficient of the jth region at time t, : linearization matrix of the jth region, : the bias vector of the jth region, : Heaviside step function (1 when the input is greater than 0, otherwise 0), : The threshold vector of the jth region, : The difference vector between the fault state and the threshold; The nonlinear transfer function of the liquid neural network is expressed as: ; in, : total number of neurons in the liquid neural network, i: neuron index, : The output weight of the i-th neuron at time t, : activation function of the i-th neuron, K: input dimension (the dimension of the fault state vector), k: input dimension index, : the connection weight from the k-th input to the i-th neuron, : The value of the kth component of the fault state vector at time t, : The threshold (bias) of the i-th neuron; Through its complex nonlinear dynamic characteristics, the model simulates how disturbances propagate across the unit's subsystems, interact with other system parameters, and evolve over time. This evolutionary trajectory takes into account the system's nonlinear characteristics, coupling effects, and time-varying properties, enabling accurate prediction of fault development paths.

[0053] The fault evolution prediction module can not only predict the evolution trajectory of a single fault, but also analyze the fault's development speed and cross-system cascading impact. The development speed prediction uses an improved algorithm based on the Lyapunov exponent:

[0054] in, : Fault evolution speed at time t, : The limit when the time increment approaches 0, : time increment, : natural logarithm function, : L2 norm of the fault state change, :time The fault state vector, : fault state vector at time t, : The vector of small disturbances at time t, : L2 norm of small perturbation; In order to improve computational efficiency, the finite difference approximation is used: ; :time The fault state vector, : L2 norm of the fault state change in the previous time step; The prediction of the development speed enables the system to determine how long it will take for the fault to develop to a dangerous level. The time arrival prediction formula is: , : predicted time to reach critical state, : The integral from the current moment to the moment of failure, : current moment, : predicted time of failure occurrence, : differential element of the fault state, : integration variable (time), :time The fault evolution speed, : critical fault state vector, :time The fault state vector, : L2 norm of the distance between the current fault state and the critical state; The analysis of cross-system cascading effects uses an improved graph theory propagation model. The coupling relationship between subsystems is represented by a directed graph, and the cascading effects are calculated using the following recursive equation: in, : No. The cascade influence vector of the iteration at time t, : The cascade influence vector of the kth iteration at time t, k: iteration number index, : inter-system coupling matrix, : direct influence vector at time t, : The feedback influence vector of the kth iteration at time t, : cascade propagation weight parameter, : directly affects the weight parameters, : Feedback influence weight parameter; The elements of the coupling matrix are calculated using the following formula: ; in, : the coupling strength from subsystem i to subsystem j in the coupling matrix, : Take the maximum value function (make sure it is non-negative), : correlation coefficient between subsystems i and j, : coupling threshold, : total number of subsystems, k: sum index, : correlation coefficient between subsystems i and k, : exponential function, : the distance between subsystems i and j, : characteristic distance parameter, : subsystem index; The fault evolution prediction module can predict the propagation path and impact of this cascading effect.

[0055] Based on the deduction results of the fault evolution trajectory, the fault evolution prediction module can generate a predictive alarm. This predictive alarm is an important feature of this embodiment. It is generated in advance when it is deduced that the future value of a monitoring parameter will reach its preset alarm limit.

[0056] The predictive alarm generation algorithm uses a multi-level risk assessment mechanism: ; in, : At the moment The probability of an alarm occurring, : Total number of monitoring parameters, i: parameter index, : The weight of the i-th parameter, : probability function, : The i-th parameter at time The fault status value, : The alarm threshold of the i-th parameter, : The current safety margin of the i-th parameter at time t; The safety margin is calculated using the following formula:

[0057] in, : the current value of the i-th parameter at time t, : The absolute difference between the current value and the threshold, : The standard deviation of the safety buffer of the i-th parameter, : safety buffer variance; When the predicted probability exceeds the set threshold, the system generates a predictive alarm: ; : predictive alarm status at time t (1 for alarm, 0 for normal), : Alarm probability threshold; The alarm time advance is determined by the following optimization algorithm: ; in, : Optimal alarm time advance, :To maximize the objective function value, :The time advance is The maintenance utility function when :The time advance is The false alarm cost function is .

[0058] Unlike traditional alarm systems based on current parameter values, predictive alarms issue warnings before a fault actually occurs, providing operators with ample time to respond. Predictive alarms are generated well before the parameter's real-time value reaches the alarm limit. This lead time allows maintenance personnel to take preventative measures to prevent further development of the fault.

[0059] The predictive alarm system can not only predict the anomaly of a single parameter, but also comprehensively consider the correlation between multiple parameters and identify complex patterns that may lead to systemic failures. Multi-parameter correlation analysis uses a feature selection algorithm based on mutual information: ; in, : Mutual information between parameters i and j at time t, i, j: parameter index, : The value of parameter i, : The value of parameter j, : The sum of all possible values ​​of parameter i, : The sum of all possible values ​​of parameter j, : At time t, parameter i takes the value x i and parameter j takes value x j The joint probability density of : At time t, parameter i takes the value x i The marginal probability density of : At time t, parameter j takes the value x j The marginal probability density of : logarithmic function, : The ratio of the joint probability to the product of the marginal probability; By analyzing the evolution trajectory in the multidimensional parameter space, the system can identify potential risks that are difficult to detect with traditional single-parameter monitoring methods. Multidimensional risk assessment uses a combination of principal component analysis and support vector machine methods:

[0060] in, : Multidimensional risk assessment value at time t, : the number of principal components retained in principal component analysis, k: principal component index, : The eigenvalue of the kth principal component (indicating the importance of the principal component), : The score vector of the kth principal component at time t, : the absolute value (or norm) of the kth principal component score vector, : The decision function value of the support vector machine for the kth principal component score, : Support vector machine decision function; The composition and functions of the multi-objective optimization module and the adaptive control generation module are basically the same as those in Example 1. However, in this embodiment, these two modules specifically use the results of fault evolution prediction as important constraints and input information to ensure that the optimization strategy and control parameters can effectively avoid the predicted fault risks, thereby realizing the organic combination of predictive maintenance and intelligent control.

[0061] Through the core function of the fault evolution prediction module, this embodiment has achieved a fundamental change in the maintenance mode of hydropower units. The traditional passive maintenance mode can only respond after the fault occurs, but this embodiment uses advanced fault evolution prediction technology to provide early warning and intervention in the incipient stage of the fault, greatly improving the safety and reliability of the unit operation. This predictive maintenance method can not only reduce unplanned downtime and reduce maintenance costs, but also extend the service life of the equipment and improve the overall economic benefits. The combination of hybrid digital twin models and liquid neural network technology provides strong technical support for fault evolution prediction, enabling complex fault evolution processes to be accurately modeled and predicted, opening up a new technical path for the intelligent operation and maintenance of hydropower units.

[0062] Example 3: like Figures 1 to 3 As shown, based on the first and second embodiments, this embodiment particularly strengthens the functions of the multi-objective optimization module and the adaptive control generation module, and adds a human-computer interaction module, realizing the multi-objective collaborative optimization control and human-computer integrated decision-making functions for "safety-efficiency-lifespan", fundamentally changing the passive risk avoidance operation mode of the traditional monitoring system, and realizing the technical sublimation from "passive monitoring" to "active optimization" and then to "intelligent decision-making".

[0063] Building on the data acquisition module and digital twin modeling module described in Example 1, and the fault evolution prediction module described in Example 2, the monitoring system of this embodiment specifically enhances the functionality of the multi-objective optimization module and the adaptive control generation module, and adds a human-computer interaction module, making it the core link of the entire system, from analysis and decision-making to execution and control. The composition and functions of these modules, as described in Examples 1 and 2, provide comprehensive data support, a high-fidelity modeling foundation, and accurate fault prediction information for multi-objective optimization and intelligent decision-making.

[0064] The multi-objective optimization module is the core component of this implementation. It goes beyond the single risk avoidance function of traditional monitoring systems and achieves global collaborative optimization of unit operations. Based on fault evolution prediction results, the multi-objective optimization module performs collaborative optimization calculations within a multidimensional decision space, with unit operational safety, power generation efficiency, and key component service life as optimization objectives, to determine the optimal operating strategy for the unit's overall benefits.

[0065] The operational safety objective in the multi-objective optimization module is precisely quantified as the safety margin between key monitoring parameters and their alarm thresholds. This quantification approach not only considers traditional single-parameter safety limits but also establishes a systematic safety margin assessment system by comprehensively analyzing the synergistic relationships between multiple key parameters. The safety margin calculation comprehensively considers the current values, changing trends, and historical statistical characteristics of multiple monitoring parameters such as vibration, swing, pressure pulsation, and air gap, forming a comprehensive indicator that reflects the overall safety status of the unit.

[0066] The safety margin assessment adopts a multi-parameter collaborative quantification algorithm. The comprehensive safety margin is obtained by multiplying the safety margins of each monitoring parameter. The safety margin of each parameter is based on the difference between its current monitoring value and the alarm threshold. It is calculated using a Gaussian distribution function and adjusted by a time-varying weight index. Finally, it is multiplied by a parameter coupling correction factor to consider the mutual influence between parameters.

[0067] The parameter coupling correction factor is calculated using a multivariate regression spline function. This factor is based on a base value of 1 plus the coupling contribution between each parameter pair. Each coupling contribution is composed of the product of the coupling strength coefficient, a bivariate B-spline basis function, and a time-varying coupling activation function. The input of the B-spline basis function is the standardized parameter value.

[0068] The time-varying weight index is dynamically adjusted through an adaptive importance evaluation algorithm. The index is obtained by multiplying the basic weight by the change rate sensitivity adjustment term and the memory decay term. The change rate sensitivity adjustment term is based on the ratio of the parameter change rate to the critical change rate, and the memory decay term takes into account the influence of the time since the last alarm.

[0069] The power generation efficiency target is scientifically quantified as the product of the unit's generated power and its water energy utilization efficiency. This quantification method fully reflects the economic nature of hydropower unit operation. The calculation of generated power takes into account the unit's power output characteristics under different loads, while the water energy utilization efficiency comprehensively considers the matching degree of hydraulic parameters such as head, flow rate, and rotational speed.

[0070] The power generation efficiency objective function adopts a multi-dimensional efficiency coupling model, and the comprehensive power generation efficiency is composed of the product of power generation power, water energy utilization efficiency, parameter matching function and dynamic process efficiency correction factor.

[0071] The water energy efficiency is calculated using an improved hydraulic coupling model. This efficiency is equal to the ratio of the theoretical hydraulic power to the water flow input power, multiplied by a loss correction term, which is 1 minus the weighted sum of various loss functions, each of which depends on the head, flow rate, and speed parameters.

[0072] The parameter matching function adopts a multi-objective harmony evaluation algorithm. The function is calculated by an exponential function. The negative value of the index is the weighted sum of the degree of deviation from the optimal value between each control parameter pair, and the weight is determined by the harmony weight matrix between parameters.

[0073] The service life target for key components is quantified as the inverse of the expected fatigue damage accumulation rate for these components, calculated based on the fault evolution prediction results. This quantification method is a key technological breakthrough in this embodiment, as it directly transforms the results of the fault evolution prediction into part of the optimization objective function. The calculation of the fatigue damage accumulation rate is based on material mechanics theory and the actual operating conditions of the unit. By analyzing the stress cycle characteristics experienced by key components under different operating parameters, the accumulation rate of fatigue damage is predicted.

[0074] The service life objective function of key components adopts a multi-scale fatigue damage prediction model. The comprehensive life target value is equal to the inverse of the weighted sum of the fatigue damage accumulation rates of each key component, multiplied by the collaborative damage correction factor between components and the environmental impact factor.

[0075] The fatigue damage accumulation rate is calculated using an improved multiaxial stress rainflow counting algorithm. The rate is the accumulation of contributions from each stress cycle. The contribution of each cycle is composed of the ratio of the cycle count to the fatigue life, the power of the ratio of the equivalent stress to the reference stress, the load history correction function, and the temperature correction function.

[0076] The inter-component collaborative damage correction factor is calculated using a graph theory cascade damage model. This factor is based on a base value of 1, plus the collaborative damage contribution between each component pair. Each contribution consists of the product of the collaborative damage coefficient, the normalized damage product, and the distance attenuation factor.

[0077] The multi-objective optimization module employs a two-stage optimization strategy, effectively addressing the trade-off between computational efficiency and accuracy in the multi-objective optimization of complex systems. In the first stage, a lightweight liquid neural network trained as a proxy model for the hybrid digital twin model is used to perform a large-scale, rough search of the global parameter space. This proxy model approach leverages the compact structure and expressive power of liquid neural networks to compress the complex hybrid digital twin model into an efficient proxy model. In the second stage, the complete hybrid digital twin model is used within the candidate solution interval identified by the rough search to perform a precise evaluation and determine the final optimal operation strategy.

[0078] The first stage rough search adopts an improved multi-objective genetic agent algorithm, and the next generation rough search solution is equal to the current generation solution plus the weighted sum of the gradients of each agent model plus the adaptive mutation vector.

[0079] The agent model is trained using a multi-fidelity reinforcement learning algorithm. The agent model output is the weighted sum of the outputs of each neuron. The output of each neuron is composed of the product of the time-varying output weight, the activation function value, and the dynamic sparsification factor. The input of the activation function is the weighted sum of the connection weight and the input plus the bias term.

[0080] The second stage of precise evaluation uses a multi-objective constrained optimization algorithm to minimize the precise objective function vector based on the complete hybrid digital twin model within the candidate solution interval. This vector contains the negative values ​​of the three objectives of safety, efficiency and lifespan.

[0081] The optimization process is subject to inequality constraints, equality constraints, and variable bound constraints.

[0082] Multi-objective collaborative optimization adopts an improved Pareto frontier dynamic maintenance algorithm, and the new Pareto frontier set is the union of the current set and the new non-dominated solution.

[0083] The optimal operation strategy is determined by a multi-criteria decision fusion algorithm, and the optimal solution is the solution that maximizes the weighted sum of each objective in the Pareto front.

[0084] The time-varying weights are updated through the operating state adaptive adjustment algorithm. The weight of each target is obtained by multiplying the basic weight by the exponential adjustment term based on the deviation between the current target value and the reference value, and then normalizing it.

[0085] The adaptive control generation module is the key link in this embodiment's transition from analysis to closed-loop control. Based on the optimal operating strategy determined by the multi-objective optimization module, this module generates and outputs a continuously changing adaptive control parameter trajectory for the unit's speed governor or excitation system. The core of this module is the introduction of a dedicated control liquid neural network, whose structure is optimized specifically for the control system's characteristics.

[0086] The control liquid neural network uses the unit's real-time status and optimization objectives as input and can generate a continuous and smooth control parameter adjustment trajectory. The control parameter adjustment trajectory is specifically manifested as a recommended adjustment curve for the proportional-integral-differential parameters of the speed regulator system, or a recommended adjustment curve for the voltage setpoint and current setpoint of the excitation system. For the speed regulator system, the adjustment curve generated by the control liquid neural network specifies in detail the dynamic change trajectory of the proportional gain, integral time constant, and differential time constant over a period of time. For the excitation system, the generated adjustment curve covers the time-varying trajectory of the excitation voltage setpoint and excitation current setpoint.

[0087] The mathematical model of controlling the liquid neural network is expressed as the time derivative of the hidden state vector of the control network is equal to the element-by-element product of the negative inverse of the time constant and the hidden state, plus the product of the weight matrix and the state activation function, plus the product of the weight matrix and the optimization target activation function.

[0088] The control time constant is adjusted through a multi-timescale adaptive mechanism. The time constant of each neuron is obtained by multiplying the basic time constant by an adaptive adjustment term based on the hidden state norm and a decay term based on the rate of change of the optimal control parameter.

[0089] The control parameter trajectory is generated using a continuous-time recursive output algorithm. The control parameter trajectory vector is composed of the product of the output weight matrix and the hidden state, the product of the feedforward weight matrix and the optimal control parameter, and the product of the memory weight matrix and the historical hidden state integral.

[0090] For the speed regulator system, the control parameter trajectory is specifically expressed as the proportional gain, integral time constant and differential time constant, which are obtained by multiplying the corresponding trajectory components by the nominal value and the deviation-based adjustment term, respectively.

[0091] For the excitation system, the control parameter trajectory is expressed as the excitation voltage and current set values ​​obtained by multiplying the corresponding trajectory components by the nominal value and the adjustment term based on the deviation, respectively.

[0092] The output of the liquid neural network is characterized by continuity and smoothness, which is determined by the continuous-time characteristics of the liquid neural network. Unlike traditional discrete control strategies, the continuous and smooth control parameter adjustment trajectory can avoid the impact of sudden changes in control parameters on unit operation, ensuring the smoothness of the control process and the stability of the system.

[0093] The smoothness constraint is ensured by the continuity regularization term, and the smoothness loss function is composed of the sum of the square integral of the second-order derivative and the weighted integral of the absolute value of the first-order derivative of each control parameter.

[0094] The human-computer interaction module is a crucial component of this embodiment's collaborative decision-making process. This module displays the control parameter trajectories generated by the adaptive control generation module on a human-computer interaction interface for operator review and confirmation. The design of the human-computer interaction module embodies the advanced "human in the loop" control concept, fully leveraging the intelligent system's analytical and optimization capabilities while retaining the crucial role of human experience and judgment in critical decision-making.

[0095] The human-machine interface utilizes an intuitive graphical display, presenting complex control parameter trajectories to operators in the form of easy-to-understand graphs, trend charts, and data tables. The interface design fully considers the operator's operating habits and cognitive characteristics. Through multi-level information display and interactive operation, operators can quickly understand the system's analysis results and control recommendations.

[0096] Operators can use the human-computer interface to conduct a detailed review of the control parameter trajectory generated by the system, including the rationality of parameter changes, the suitability of time points, the safety of adjustment ranges, and other aspects. The interface provides flexible parameter correction capabilities, allowing operators to adjust the system's recommended control parameters based on actual site conditions and operational experience.

[0097] After receiving confirmation from the operator, the human-machine interaction module transmits the reviewed and possibly revised control parameters to the unit's actual control system. This process is implemented using standard industrial communication protocols and secure data transmission mechanisms, ensuring that control instructions are accurately and promptly transmitted to the appropriate control devices. The control parameter distribution process utilizes a step-by-step execution and real-time monitoring safety mechanism. The system monitors control execution in real time and immediately initiates safety protection measures if any anomalies are detected.

[0098] Through the coordinated action of the human-computer interaction modules, the entire system forms a complete closed-loop intelligent decision-making process with humans in the loop. This process begins with data collection, progresses through digital twin modeling, fault evolution prediction, multi-objective optimization, and adaptive control generation, and ultimately implements decisions through human-computer interaction. This forms a complete intelligent operations and maintenance system from perception to cognition, from analysis to decision-making, and from recommendation to execution.

[0099] This embodiment fundamentally changes the functional positioning and technical architecture of hydropower unit monitoring systems by implementing multi-objective optimization and human-machine collaborative decision-making. Traditional monitoring systems can only provide status information and passive alarms. However, this embodiment, through advanced multi-objective optimization technology and human-machine interaction mechanisms, achieves a comprehensive upgrade from passive monitoring to active optimization, from single risk avoidance to comprehensive optimization, and from manual decision-making to human-machine collaborative decision-making.

[0100] The foregoing description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein and should not be construed as excluding other embodiments. Rather, the present invention can be used in various other combinations, modifications, and environments and can be modified within the scope of the concept described herein through the above teachings or techniques or knowledge in the relevant field. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention are intended to be protected by the appended claims.

Claims

1. An online monitoring system for the status of a hydropower station unit, characterized in that: It includes data acquisition module, digital twin modeling module, fault evolution prediction module, multi-objective optimization module and adaptive control generation module; The data acquisition module is used to collect real-time monitoring data of the hydropower unit; The digital twin modeling module is configured to build and run a hybrid digital twin model, which is composed of a physical mechanism main model and a liquid neural network residual compensation model. The physical mechanism main model outputs a first predicted state based on the rotor dynamics, hydraulics and electromagnetic field coupling mechanism of the unit. The liquid neural network residual compensation model dynamically models the residual sequence between the first predicted state and real-time monitoring data and outputs a residual compensation value. By fusing the first predicted state and the residual compensation value, a high-fidelity final state of the unit is generated; The fault evolution prediction module is configured to perform online virtual fault injection based on the hybrid digital twin model, and use the liquid neural network residual compensation model to deduce and predict the future evolution path, development speed and cross-system cascading impact of the injected fault; The multi-objective optimization module is configured to perform collaborative optimization calculations in a multi-dimensional decision space with the optimization objectives of unit operation safety, power generation efficiency and service life of key components based on the fault evolution prediction results, so as to obtain the optimal operation strategy with the best overall benefits of the unit; The adaptive control generation module is configured to generate and output an adaptive control parameter trajectory that continuously changes over time and is oriented to the speed governor system or the excitation system of the unit according to the optimal operation strategy.

2. The online monitoring system for the status of a hydropower station unit according to claim 1, characterized in that: The liquid neural network residual compensation model is a continuous-time recurrent neural network defined by ordinary differential equations. Its neurons have dynamically changing nonlinear time constants determined by the current input and hidden state, enabling it to accurately capture complex time-varying dynamics that are not fully explained by the main model of the physical mechanism.

3. The online monitoring system for the status of a hydropower station unit according to claim 2, characterized in that: The main model of the physical mechanism is a multi-physics field coupling model, which includes a rotor dynamics sub-model for describing the vibration and swing of the main shaft, a fluid dynamics sub-model for describing the pressure pulsation of the volute, a fluid dynamics sub-model for describing the pressure pulsation of the top cover, a fluid dynamics sub-model for describing the pressure pulsation of the draft tube, and an electromagnetic field sub-model for describing the electromagnetic force caused by the uneven air gap between the stator and rotor.

4. The online monitoring system for the status of a hydropower station unit according to claim 3, characterized in that: The fault evolution prediction module specifically performs the following steps: injecting a virtual fault initial disturbance into a hybrid digital twin model calibrated with real-time data, using the nonlinear dynamic transfer characteristics of the liquid neural network residual compensation model to deduce the evolution trajectory of the disturbance, and generating a predictive alarm. The predictive alarm is generated in advance when it is deduced that the future value of a certain monitoring parameter will reach its preset alarm limit.

5. The online monitoring system for the status of a hydropower station unit according to claim 4, characterized in that: The multi-objective optimization module adopts a two-stage optimization strategy: in the first stage, a lightweight liquid neural network trained as a hybrid digital twin model proxy model is used to perform a rough search in the large-scale global parameter space; In the second stage, the complete hybrid digital twin model is called to perform precise evaluation within the candidate solution range determined by the rough search to determine the final optimal operation strategy.

6. The online monitoring system for the status of a hydropower station unit according to claim 5, characterized in that: The operational safety target in the multi-objective optimization module is quantified as the safety margin between the key monitoring parameters and their alarm thresholds; the power generation efficiency target is quantified as the product of the unit's power generation power and the water energy utilization efficiency; and the key component service life target is quantified as the inverse of the expected fatigue damage accumulation rate of the key components calculated based on the fault evolution prediction results.

7. The online monitoring system for the status of a hydropower station unit according to claim 1, characterized in that: The adaptive control generation module includes a dedicated control liquid neural network, which takes the real-time status of the unit and the optimization target as input and outputs a continuous and smooth control parameter adjustment trajectory. The control parameter adjustment trajectory is specifically manifested as a recommended adjustment curve for the proportional-integral-differential parameters of the speed regulator system, or a recommended adjustment curve for the voltage setting value and current setting value of the excitation system.

8. The online monitoring system for the status of a hydropower station unit according to claim 7, characterized in that: The real-time monitoring data collected by the data acquisition module include: vibration data of the upper frame of the unit, vibration data of the lower frame of the unit, swing data of the upper guide bearing, swing data of the lower guide bearing, swing data of the water guide bearing, pressure pulsation data of the volute, pressure pulsation data of the top cover, pressure pulsation data of the draft pipe, and air gap data between the stator and rotor of the generator.

9. The online monitoring system for the status of a hydropower station unit according to claim 8, characterized in that: The liquid neural network residual compensation model is trained through learning so that the residual compensation value it outputs can bridge the dynamic error between the physical mechanism main model prediction and the real-time monitoring data to the greatest extent, thereby achieving the optimal fit between the final state of the unit and the real-time monitoring data.

10. The online monitoring system for the status of a hydropower station unit according to claim 9, characterized in that: It also includes a human-computer interaction module, which presents the control parameter trajectory generated by the adaptive control generation module on the human-computer interaction interface for the operating personnel to conduct final review and confirmation. After obtaining the confirmation instruction, the corrected parameters are sent to the actual control system of the unit, thus forming a closed-loop intelligent decision-making process in which people are in the loop.

Images