Water turbine efficiency online monitoring and operation optimization method and system

Abstract

The invention belongs to the technical field of hydroelectric generating set monitoring, and relates to a water turbine efficiency online monitoring and operation optimization method and system. The operation data of the water turbine are obtained; acquiring an actual efficiency value by using the mechanism data hybrid model according to the water turbine operation data; acquiring equipment state parameters, and generating an operation parameter optimization strategy by using a deep reinforcement learning algorithm according to the actual efficiency value and the equipment state parameters; and performing security verification on the operation parameter optimization strategy to ensure that the optimization strategy conforms to the equipment security boundary. According to the method, multi-source sensor data are fused, a mechanism data hybrid model is constructed to calculate an actual efficiency value, an operation parameter optimization strategy is generated by using a deep reinforcement learning algorithm according to the actual efficiency value and equipment state parameters, and the efficiency of the water turbine is detected by combining collaborative optimization of operation parameters of the water turbine and the equipment state. Adaptation of dynamic working conditions is facilitated, and advanced optimization of the water turbine is facilitated.

Description

Technical Field

[0001] The present invention belongs to the technical field of hydro-generator unit monitoring, and relates to an online monitoring and operation optimization method and system for water turbine efficiency. Background Art

[0002] As the core power equipment in a hydropower station, the core function of a water turbine is to efficiently convert the energy of water flow in nature into rotational mechanical energy to drive the generator to operate, and then into clean electric energy to deliver light and power to thousands of households. Most modern water turbines are installed in hydropower stations to drive generators to generate electricity.

[0003] Water turbine efficiency monitoring refers to measuring and analyzing the efficiency of energy conversion during the operation of a water turbine to evaluate the effectiveness of converting the kinetic energy (or potential energy) of water flow into mechanical energy. Efficiency monitoring can reveal possible minor defects or performance degradation in the design, manufacture, installation, and long-term operation of a water turbine, providing a scientific basis for the operation and maintenance team of the hydropower station to guide them to take targeted maintenance measures, such as optimizing the water flow regulation strategy, implementing a preventive maintenance plan, or carrying out necessary equipment upgrades. This process is a key link in the operation and maintenance of a hydropower station, directly affecting power generation efficiency, equipment life, and economic benefits.

[0004] In the prior art, the monitoring of water turbine efficiency mainly relies on fixed threshold alarms and empirical models, which mainly have problems such as poor adaptability to dynamic operating conditions and lag in water turbine operation optimization. Summary of the Invention

[0005] The purpose of the present invention is to provide an online monitoring and operation optimization method and system for water turbine efficiency to solve the technical problems of poor adaptability of water turbine efficiency monitoring to dynamic operating conditions and lag in water turbine operation optimization.

[0006] To achieve the above purpose, the present invention adopts the following technical solutions: In the first aspect, the present invention provides an online monitoring and operation optimization method for water turbine efficiency, including the following steps: Obtain the operation data of the water turbine; Obtain the actual efficiency value using a mechanism data hybrid model based on the operation data of the water turbine; Obtain the equipment status parameters, and generate an operation parameter optimization strategy using a deep reinforcement learning algorithm based on the actual efficiency value and the equipment status parameters; Verify the safety of the operation parameter optimization strategy.

[0007] Furthermore, the operation data of the water turbine includes: water turbine vibration signals, pressure distribution in the volute, water turbine flow parameters, water turbine speed, and temperature inside the bearing housing.

[0008] Further, the following steps are also included: Perform progressive wavelet denoising, outlier rejection, and time series feature extraction on the operating data of the water turbine.

[0009] Further, the actual efficiency value is obtained by using a mechanism data hybrid model based on the operating data of the water turbine, specifically as follows: Simulate and generate multi-condition data of the water turbine; Construct a physical model based on the multi-condition data, specifically as follows: Use proper orthogonal decomposition to reduce the dimension of the CFD full-order model, establish a theoretical efficiency calculation module, and output the theoretical efficiency; Construct a data-driven model, specifically as follows: Adopt a spatio-temporal attention network STA-LSTM. The input layer fuses the operating data of the water turbine and the theoretical efficiency value, and outputs the efficiency deviation; Couple the STA-LSTM encoder with the physical model through transfer learning to form a hybrid inference module, and output the actual efficiency according to the hybrid inference module in combination with the theoretical efficiency, efficiency deviation, and operating data of the water turbine.

[0010] Further, the calculation formula for the theoretical efficiency is as follows:

[0011]

[0012]

[0013] In the formula, represents the theoretical efficiency, represents the measured temperature of the bearing, represents the calibrated temperature, represents the maximum allowable temperature, γ is an empirical coefficient, represents the output torque, represents the angular velocity, represents the flow rate of the water turbine, represents the temperature compensation coefficient, represents the inlet head, represents the outlet head.

[0014] Further, the spatio-temporal attention network includes: a time attention module and a spatial graph convolution module; The time attention module is used to calculate the weights of each time step:

[0015] The spatial graph convolution module aggregates node features with the sensor topology as the adjacency matrix:

[0016] In the formula, represents the hidden state of the LSTM at time step t, represents the trainable weight matrix, represents the trainable query vector, represents the node feature matrix of the layer, represents the trainable weight matrix, represents the adjacency matrix, represents the degree matrix, represents the weights of each time step, and softmax() represents the normalization of the scores for all time steps, represents the corresponding time step, represents enhancing the non - linear expression ability,

[0017] Furthermore, the STA - LSTM encoder is coupled with the physical model through transfer learning to form a hybrid inference module, specifically as follows: Using multi - condition data as the pre - training data set; Pre - training the STA - LSTM encoder using the pre - training data set, and the loss function is:

[0018] In the formula, represents the efficiency deviation predicted by the STA - LSTM network, represents the efficiency deviation calculated by CFD simulation, The divergence term is approximately calculated by the Jensen - Shannon distance, represents the feature distribution of the simulation data, represents the feature distribution of the real sensor data, the trade - off coefficient; According to the pre - trained STA - LSTM encoder, it is coupled with the physical model to form a hybrid inference module.

[0019] Furthermore, the device state parameters are obtained, and according to the actual efficiency value and the device state parameters, an operation parameter optimization strategy is generated using the deep reinforcement learning algorithm, specifically as follows: The device state parameters include state space parameters, and the state space S is defined as , , , , W], where is the actual efficiency, is the measured temperature of the bearing, is the flow fluctuation coefficient, is the inlet head change rate, W is the wear index; Define action space B = [ ], is the guide vane opening adjustment amount, is the excitation current correction value, is the cooling water flow adjustment amount; Building a reward function ), where Calculated based on the equipment safety status, is the effective value of vibration, Pt ∣ is the L2 norm of the guide vane opening and the excitation current adjustment, is the exponential form of the vibration term, For actual efficiency, is the reward weight, To punish the weight; The policy network is trained using the PPO algorithm based on the state space, action space, and reward function.

[0020] Furthermore, the The calculation method is as follows:

[0021] in, For the Item weight, For the Safety factors include vibration safety factor, temperature safety factor, wear safety factor and pressure pulsation safety factor. Safety rewards include vibration safety rewards, temperature safety rewards, wear safety rewards, and pressure pulsation safety rewards, as follows:

[0022] in, represents the vibration safety factor, Indicates vibration safety reward;

[0023] in, represents the temperature safety factor, For temperature safety rewards, represents the hyperbolic tangent function;

[0024] in, represents the wear safety factor, For wear safety rewards, is the weight coefficient;

[0025] Among them, represents the pressure pulsation safety factor, is the pressure pulsation safety reward.

[0026] In a second aspect, the present invention also discloses an on-line monitoring and operation optimization system for water turbine efficiency, including: A data acquisition module for acquiring the operation data of the water turbine; An actual efficiency module for obtaining the actual efficiency value by using a mechanism data hybrid model according to the operation data of the water turbine; An operation parameter optimization strategy generation module for obtaining the equipment state parameters, and generating an operation parameter optimization strategy by using a deep reinforcement learning algorithm according to the actual efficiency value and the equipment state parameters; A safety verification module for performing safety verification on the operation parameter optimization strategy.

[0027] Compared with the prior art, the present invention has the following beneficial effects: By acquiring the operation data of the water turbine, the present invention provides multi-dimensional data support for subsequent models; by using a mechanism data hybrid model according to the operation data of the water turbine to obtain the actual efficiency value, it breaks through the limitation of a single model, ensures the basic accuracy, and is beneficial to ensuring the accuracy of the actual efficiency value; by obtaining the equipment state parameters, and generating an operation parameter optimization strategy by using a deep reinforcement learning algorithm according to the actual efficiency value and the equipment state parameters, through dynamic learning of the environment feedback, it balances the short-term performance and long-term equipment loss, and quickly adapts to the requirements of complex scenarios such as load fluctuations and power grid dispatching; by performing safety verification on the operation parameter optimization strategy, it ensures that the optimization strategy conforms to the equipment safety boundary. The present invention fuses multi-source sensor data, constructs a mechanism data hybrid model to calculate the actual efficiency value, then generates an operation parameter optimization strategy by using a deep reinforcement learning algorithm according to the actual efficiency value and the equipment state parameters, and at the same time combines the collaborative optimization of the operation parameters and the equipment state of the water turbine to realize the detection of the water turbine efficiency, which is beneficial to adapting to dynamic working conditions and is beneficial to the advanced optimization of the water turbine.

[0028] The present invention constructs a mechanism data hybrid model, the physical model provides a theoretical benchmark, the data-driven model compensates for the unmodeled dynamics, and the comprehensive error of the output actual efficiency is lower than that of a single model.

[0029] The theoretical efficiency calculation module of the present invention is integrated with a temperature compensation term for correcting the influence of friction loss caused by temperature changes of bearings or mechanical components in the calculation of water turbine efficiency, which can realize safety warning, and in the control strategy, the load can be dynamically adjusted to maintain it within a safe range.

[0030] The present invention uses a deep reinforcement learning algorithm to generate an operation parameter optimization strategy, incorporating multi-dimensional indicators such as efficiency, vibration, and temperature, comprehensively considering multiple factors, and the generated optimization strategy is more scientific and reasonable.

[0031] The system of the present invention includes: a data acquisition module, an actual efficiency module, an operation parameter optimization strategy generation module, and a security verification module; the data acquisition module is used to acquire the operation data of the water turbine; the actual efficiency module is used to obtain the actual efficiency value according to the operation data of the water turbine by using a mechanism data hybrid model; the operation parameter optimization strategy generation module is used to obtain the equipment status parameters, and according to the actual efficiency value and the equipment status parameters, use a deep reinforcement learning algorithm to generate an operation parameter optimization strategy; the security verification module is used to perform security verification on the operation parameter optimization strategy. Each module cooperates with each other, can realize the detection of the water turbine efficiency, is beneficial to adapting to dynamic working conditions, and is beneficial to the advanced optimization of the water turbine. Brief Description of the Drawings

[0032] Figure 1 is the method flow chart of an embodiment of the present invention; Figure 2 is the method flow chart of another embodiment of the present invention; Figure 3 is the system module connection diagram of an embodiment of the present invention. Detailed Description of the Embodiment

[0033] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts shall fall within the protection scope of the present invention.

[0034] It should be noted that the terms "first", "second", etc. in the description and drawings of the present invention are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments of the present invention described here can be implemented in an order other than those illustrated or described here. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units does not necessarily have to be limited to those clearly listed steps or units, but may include other steps or units that are not clearly listed or are inherent to these processes, methods, products, or devices.

[0035] The present invention will be further described in detail below in conjunction with the accompanying drawings: Example 1: Refer to Figure 2 , this example discloses an online monitoring and operation optimization method for water turbine efficiency, including the following steps: S1. Obtain the operation data of the water turbine to provide multi-dimensional data support for the subsequent model. The operation data of the water turbine includes: the vibration signal of the water turbine, the pressure distribution of the volute, the flow parameters of the water turbine, the rotation speed of the water turbine, and the temperature inside the bearing housing. The data covers the full operating condition range of the water turbine, avoiding misjudgment of efficiency caused by monitoring blind spots, and at the same time realizing a millisecond-level data refresh rate to ensure the timeliness of monitoring.

[0036] In the embodiment of the present invention, the following steps are further included: Perform progressive wavelet denoising, outlier removal, and time series feature extraction on the operation data of the water turbine.

[0037] S2. Use the mechanism data hybrid model to obtain the actual efficiency value according to the operation data of the water turbine, break through the limitations of a single model, ensure basic accuracy, and be conducive to ensuring the accuracy of the actual efficiency value. Specifically as follows: Simulate and generate multi-condition data of the water turbine; Construct a physical model according to the multi-condition data. Specifically as follows: Use proper orthogonal decomposition to reduce the dimension of the CFD full-order model, establish a theoretical efficiency calculation module, and output the theoretical efficiency. The calculation formula of the theoretical efficiency is as follows:

[0038]

[0039]

[0040] In the formula, represents the theoretical efficiency, represents the measured temperature of the bearing, represents the calibrated temperature, represents the maximum allowable temperature, γ is an empirical coefficient, represents the output torque, represents the angular velocity, represents the flow rate of the water turbine, represents the temperature compensation coefficient, represents the inlet head, represents the outlet head.

[0041] Construct a data-driven model. Specifically as follows: Use the spatio-temporal attention network STA-LSTM. The input layer fuses the operation data of the water turbine and the theoretical efficiency value, and outputs the efficiency deviation. The spatio-temporal attention network includes: a time attention module and a spatial graph convolution module; The time attention module is used to calculate the weights of each time step:

[0042] The spatial graph convolution module aggregates node features with the sensor topology as the adjacency matrix:

[0043] In the formula, represents the hidden state of the LSTM at time step t, represents the trainable weight matrix, represents the trainable query vector, represents the node feature matrix of the layer, represents the adjacency matrix, represents the degree matrix, represents the weights of each time step, and softmax() represents the normalization of the scores for all time steps, represents the corresponding time step, represents enhancing the non-linear expression ability, represents the ReLU activation function.

[0044] The time STA-LSTM encoder is coupled with the physical model through transfer learning to form a hybrid inference module. The actual efficiency is output according to the hybrid inference module in combination with the theoretical efficiency, efficiency deviation, and turbine operation data, as follows: Using multi-condition data as the pre-training dataset; Pre-training the STA-LSTM encoder with the pre-training dataset, and the loss function is:

[0045] In the formula, represents the efficiency deviation predicted by the STA-LSTM network, represents the efficiency deviation calculated by CFD simulation, The divergence term is approximately calculated by the Jensen-Shannon distance, represents the feature distribution of the simulation data, represents the feature distribution of the real sensor data, Trade-off coefficient; Coupling the pre-trained STA-LSTM encoder with the physical model to form a hybrid inference module.

[0046] S3. Obtain the device status parameters. According to the actual efficiency value and the device status parameters, use the deep reinforcement learning algorithm to generate an operating parameter optimization strategy, and through dynamic learning environment feedback, balance short-term performance and long-term device wear, and quickly adapt to the requirements of complex scenarios such as load fluctuations and grid dispatching, specifically as follows: The device status parameters include state space parameters, and define the state space S = , , , , W], where is the actual efficiency, is the measured temperature of the bearing, is the flow fluctuation coefficient, is the change rate of the inlet head, and W is the wear index; Define the action space B = , is the guide vane opening adjustment amount, is the excitation current correction value, is the cooling water flow adjustment amount; Construct the reward function ), where, is calculated based on the device safety status, is the effective value of vibration, Pt ∣ is the L2 norm of the guide vane opening and the excitation current adjustment amount, is the exponential form of the vibration term, is the actual efficiency, is the reward weight, is the punishment weight; Train the policy network according to the state space, action space and reward function and adopt the PPO algorithm.

[0047] In the embodiment of the present invention, the calculation method of the is as follows:

[0048] Where, is the weight of the th item, is the safety factor of the th item. The safety factor includes vibration safety factor, temperature safety factor, wear safety factor and pressure pulsation safety factor, is the safety reward, including vibration safety reward, temperature safety reward, wear safety reward and pressure pulsation safety reward, specifically as follows:

[0049] Where, represents the vibration safety factor, Indicates the vibration safety reward;

[0050] Among them, Indicates the temperature safety factor, is the temperature safety reward, Indicates the hyperbolic tangent function;

[0051] Among them, Indicates the wear safety factor, is the wear safety reward, is the weight coefficient;

[0052] Among them, Indicates the pressure pulsation safety factor, is the pressure pulsation safety reward.

[0053] S4. Perform safety verification on the operation parameter optimization strategy to ensure that the optimization strategy complies with the equipment safety boundary.

[0054] Based on the above method, the present invention also discloses an online monitoring and operation optimization system for water turbine efficiency. Refer to Figure 3 , which includes: a data acquisition module, an actual efficiency module, an operation parameter optimization strategy generation module, and a safety verification module; The data acquisition module is used to acquire the operation data of the water turbine; The actual efficiency module is used to obtain the actual efficiency value according to the operation data of the water turbine by using the mechanism data hybrid model; The operation parameter optimization strategy generation module is used to obtain the equipment state parameters, and generate an operation parameter optimization strategy according to the actual efficiency value and the equipment state parameters by using the deep reinforcement learning algorithm; The safety verification module is used to perform safety verification on the operation parameter optimization strategy.

[0055] Each module of the system of the present invention cooperates with each other, can realize the detection of the water turbine efficiency, is beneficial to adapting to dynamic working conditions, and is beneficial to the advanced optimization of the water turbine.

[0056] Embodiment 2: Refer to Figure 1 , this embodiment discloses an online monitoring and operation optimization method for water turbine efficiency, including the following steps: Step S1: Obtain the original data, and collect the operation data of the water turbine in real time through a multi-source sensor array. The operation data of the water turbine includes vibration signals, pressure distribution, flow parameters, rotational speed, and temperature; Step S2: Preprocess the original data; Step S3: Use the preprocessed data as input and calculate the actual efficiency value through a mechanism data hybrid model, where the mechanism data hybrid model includes: A physical model, a hydrodynamic simulation based on the Navier-Stokes equation; A data-driven model that uses a spatio-temporal attention network to fuse multi-modal time series data; Step S4: Generate an operation parameter optimization strategy using a deep reinforcement learning algorithm based on the actual efficiency value and equipment status parameters; Step S5: Verify the security of the operation parameter optimization strategy.

[0057] In one implementation, the multi-source sensor array includes: a triaxial acceleration sensor arranged in the runner chamber, a piezoelectric pressure sensor arranged circumferentially along the volute, an ultrasonic flowmeter arranged at the draft tube, an optoelectronic encoder for collecting the water turbine speed, and a temperature sensor arranged in the bearing housing.

[0058] In one implementation, the preprocessing in step S2 includes wavelet denoising, outlier removal, and time series feature extraction.

[0059] In one implementation, the construction method of the mechanism data hybrid model in step S3 is as follows: Step S3.1: Generate multi-condition data of the water turbine through CFD software simulation as a pre-training data set; Step S3.2: Construct a physical model, use proper orthogonal decomposition to reduce the dimension of the CFD full-order model, establish a theoretical efficiency calculation module, and output the theoretical efficiency ; Step S3.3: Construct a data-driven model, use a spatio-temporal attention network STA-LSTM, fuse sensor time series data and theoretical efficiency value in the input layer, and output efficiency deviation prediction ; Step S3.4: Couple the STA-LSTM encoder with the physical model through transfer learning to form a hybrid inference module and output the actual efficiency 。

[0060] In one implementation, in step S3.2, the theoretical efficiency calculation module integrates a temperature compensation term, and the formula for theoretical efficiency calculation is expressed as follows;

[0061] In the formula, represents the calibration temperature, represents the maximum allowable temperature, γ Empirical coefficient Denotes the output torque Denotes the angular velocity , Denotes the inlet head Denotes the outlet head Denotes the temperature compensation coefficient

[0062] In one embodiment, the spatio-temporal attention network in step S3.3 includes: A time attention module that calculates the weights for each time step: ; A spatial graph convolutional module that uses the sensor topology as the adjacency matrix to aggregate node features:

[0063] In the formula, Denotes the hidden state of the LSTM at time step t Denotes the trainable weight matrix Denotes the trainable query vector Denotes the Node feature matrix of the layer Denotes the trainable weight matrix Denotes the adjacency matrix Denotes the degree matrix Denotes the weights for each time step, and softmax() denotes the normalization of the scores for all time steps Denotes the measured temperature of the bearing Denotes enhancing the non-linear expression ability Denotes the ReLU activation function

[0064] In one embodiment, the transfer learning in step S3.4 includes: Pre-train the STA-LSTM encoder using the pre-training dataset, and the loss function is: ; In the formula, Denotes the efficiency deviation predicted by the STA-LSTM network Denotes the efficiency deviation calculated by CFD simulation The divergence term is approximately calculated by the Jensen-Shannon distance Denotes the feature distribution of the simulation data Denotes the feature distribution of the real sensor data Trade-off coefficient

[0065] In one embodiment, the specific method for generating an operating parameter optimization strategy using the deep reinforcement learning algorithm in step S4 is as follows: Step S4.1: Define the state space S = , , , ΔH in , W], where is the actual efficiency, is the measured temperature of the bearing, is the flow fluctuation coefficient, ΔH in is the inlet head change rate, and W is the wear index; Step S4.2: Define the action space B = , is the guide vane opening adjustment amount, is the excitation current correction value, is the cooling water flow adjustment amount; Step S4.3: Construct the reward function ), where, St ) is calculated based on the equipment safety state, is the effective vibration value, Pt ∣ is the L2 norm of the guide vane opening and the excitation current adjustment amount, is the exponential form of the vibration term, is the actual efficiency, is the reward weight, is the punishment weight; Step S4.4: Train the policy network using the PPO algorithm.

[0066] In one implementation, the calculation method of the St ) is as follows:

[0067] Where, is the weight of the th term, is the safety factor of the th term, including: Vibration safety reward =

[0068] Where, represents the vibration safety factor; Temperature safety reward = −0.8)), where, represents the temperature safety factor; Wear safety reward = β ⋅ln( +0.1), where, Indicates the wear safety factor; Pressure pulsation safety reward = , where Indicates the pressure pulsation safety factor.

[0069] To achieve the above object, the present invention also provides an on-line monitoring and operation optimization system for water turbine efficiency, including: A data acquisition module, which acquires raw data and real-time collects the operation data of the water turbine through a multi-source sensor array. The operation data includes vibration signals, pressure distribution, flow parameters, rotational speed and temperature; A data preprocessing module: preprocesses the raw data; A hybrid model construction module: takes the preprocessed data as input and calculates the actual efficiency value through a mechanism-data hybrid model. The mechanism-data hybrid model includes: a physical model, a hydrodynamic simulation based on the Navier-Stokes equation; a data-driven model, which uses a spatio-temporal attention network to fuse multi-modal time series data; An optimization strategy module: generates an operation parameter optimization strategy by using a deep reinforcement learning algorithm according to the actual efficiency value and the equipment state parameters; A safety verification module: verifies the safety of the operation parameter optimization strategy.

[0070] Compared with the prior art, the present invention has the following beneficial effects: The present invention fuses multi-source sensor data, constructs a mechanism-data hybrid model to calculate the actual efficiency value, and then generates an operation parameter optimization strategy by using a deep reinforcement learning algorithm according to the actual efficiency value and the equipment state parameters. At the same time, it combines the collaborative optimization of the operation parameters (such as guide vane opening, rotational speed, etc.) and the equipment state (such as vibration, wear) of the water turbine, thus effectively solving the problems of poor adaptability to dynamic working conditions and lag in water turbine operation optimization existing in the prior art.

[0071] The present invention constructs a mechanism-data hybrid model. The physical model provides a theoretical benchmark, and the data-driven model compensates for the unmodeled dynamics. The comprehensive error of the output actual efficiency is lower than that of a single model.

[0072] The theoretical efficiency calculation module of the present invention integrates a temperature compensation term, which is used to correct the influence of friction loss caused by temperature changes of bearings or mechanical components in the calculation of water turbine efficiency, can realize safety warning, and in the control strategy, the load can be dynamically adjusted to maintain it within a safe range.

[0073] The present invention uses a deep reinforcement learning algorithm to generate an operation parameter optimization strategy, incorporates multi-dimensional indexes such as efficiency, vibration, temperature, etc., comprehensively considers multiple factors, and the generated optimization strategy is more scientific and reasonable.

[0074] Example 3: As Figure 1 shown, this embodiment provides an online monitoring and operation optimization method for the efficiency of a hydraulic turbine. The design principle of this method is as follows: Multisource sensor data is fused to construct a mechanism data hybrid model to calculate the actual efficiency value. Then, based on the actual efficiency value and equipment status parameters, a depth reinforcement learning algorithm is used to generate an operation parameter optimization strategy, solving the problems of poor adaptability to dynamic working conditions and lag in the operation optimization of hydraulic turbines existing in the prior art.

[0075] The specific implementation steps of the online monitoring and operation optimization method for the efficiency of the hydraulic turbine provided in this embodiment are as follows: Step S1, obtaining the original data; The operation data of the hydraulic turbine is collected in real time through a multisource sensor array. The operation data includes vibration signals, pressure distribution, flow parameters, rotational speed, and temperature. The multisource sensor array includes: a triaxial acceleration sensor arranged in the runner chamber with a sampling frequency ≥ 10 kHz; piezoelectric pressure sensors arranged circumferentially along the spiral case; an ultrasonic flowmeter arranged at the draft tube with an accuracy class of 0.5; an optoelectronic encoder for collecting the rotational speed of the hydraulic turbine, and the rotational speed signal measured by the optoelectronic encoder with a resolution ≤ 0.1°, and a temperature sensor arranged in the bearing housing.

[0076] Step S2, data preprocessing; The obtained original data is preprocessed. The preprocessing methods include but are not limited to: wavelet denoising, outlier rejection, and time series feature extraction. The above preprocessing methods are all existing mature data preprocessing methods, so they will not be elaborated here.

[0077] Step S3, constructing a mechanism data hybrid model and calculating the actual efficiency value; The preprocessed data is used as the input, and the actual efficiency value is calculated through the mechanism data hybrid model. In this embodiment, the mechanism data hybrid model includes: A physical model, a hydrodynamic simulation based on the Navier - Stokes equation. The physical model provides a theoretical benchmark and outputs a theoretical efficiency value; A data - driven model, which uses a spatio - temporal attention network to fuse multimodal time - series data and outputs a predicted efficiency deviation. The data - driven model compensates for unmodeled dynamics.

[0078] The specific construction method of the mechanism data hybrid model is as follows: Step S3.1: Generate multi-condition data of the water turbine through CFD software simulation as the pre-training data set; generate a high-precision simulation data set based on CFD (Computational Fluid Dynamics) software (such as ANSYS Fluent), simulate different conditions (combinations of flow rate, head, and guide vane opening), and output the flow field pressure distribution, eddy current characteristics, and theoretical efficiency values. The preprocessing method for the multi-condition data of the water turbine is: simulation data enhancement. For example, perform spatial interpolation (cubic spline interpolation) and time series expansion (GAN generative adversarial network) on the CFD results to cover extreme conditions (such as cavitation and transient load impact).

[0079] Step S3.2: Build a physical model, reduce the dimension of the CFD full-order model using proper orthogonal decomposition, establish a theoretical efficiency calculation module, and output the theoretical efficiency ; Build the internal flow field control equation of the water turbine based on the Navier-Stokes equation:

[0080] where u is the flow velocity field, p is the pressure field, μ is the dynamic viscosity, and f 边界 is the boundary condition of the guide vane and runner.

[0081] Extract the main modes of the flow field using proper orthogonal decomposition (POD), compress the CFD full-order model into a low-dimensional state space model (the dimension is reduced from 10 6 to 10 2 ), and achieve real-time simulation.

[0082] In this embodiment, the theoretical efficiency calculation module integrates a temperature compensation term, and the formula for calculating the theoretical efficiency is as follows; The formula for calculating the theoretical efficiency is as follows;

[0083]

[0084]

[0085] In the formula, represents the measured temperature of the bearing, represents the calibrated temperature (such as 25 °C), represents the maximum allowable temperature, γ is the empirical coefficient (determined by fitting through experiments or historical data, for example: 0.06), represents the output torque, represents the angular velocity, represents the inlet head, represents the outlet head, represents the temperature compensation coefficient, To quantify the negative impacts of temperature-sensitive factors such as bearing friction and winding resistance on the actual efficiency of a water turbine. When the temperature rises, a decrease in lubricating viscosity or metal expansion will lead to an increase in mechanical friction. The value decreases, thereby reducing the theoretical efficiency value. When the temperature is at the calibration value (such as normal temperature 25°C), ≈ 1, and no correction is required).

[0086] Step S3.3: Build a data-driven model. Adopt the spatio-temporal attention network STA-LSTM. The input layer fuses the sensor time-series data and the theoretical efficiency value, and outputs the prediction of the efficiency deviation. ; The spatio-temporal attention network STA-LSTM includes: A time attention module that calculates the weights of each time step: softmax ; A spatial graph convolution module that uses the sensor topology as the adjacency matrix to aggregate node features:

[0087] In the formula, represents the hidden state of LSTM at time step t (dimension d×1), represents the trainable weight matrix (dimension d′×d) used for linear transformation of the hidden state, represents the trainable query vector; (dimension d′×1), which determines which feature patterns to focus on, Maps the hidden state to the attention space. tanh() enhances the non-linear expression ability and limits the output range to [-1,1][−1,1], Calculates the dot product (scalar) of the query vector and the transformed hidden state, which characterizes the importance of time step t. softmax() normalizes the scores of all time steps to obtain the weight ∈(0,1); represents the node feature matrix of the layer (dimension N×d), represents the trainable weight matrix (dimension d×d′) used for feature transformation. N represents the number of sensor nodes, and d represents the feature dimension. represents the adjacency matrix (dimension N×N, using the Euclidean distance threshold method: if the sensor spacing < 1m, then ij = 1), which defines the physical connection or signal correlation between sensors. For example: if the vibration sensor and the pressure sensor are installed adjacent to each other, then the corresponding A represents the degree matrix, , Symmetrically normalize the adjacency matrix to address the problem of uneven node degree distribution. Aggregate the features of neighboring nodes. Perform a linear transformation. Let

[0088] Step S3.4: Couple the time STA-LSTM encoder with the physical model through transfer learning to form a hybrid inference module and output the actual efficiency. The actual efficiency is calculated as follows: where is the dynamic confidence weight of temperature T and vibration V (regulated by the Sigmoid function); in this step, the transfer learning strategy includes: Model pre-training stage: Pre-train the STA-LSTM encoder using the pre-training dataset, and the loss function is: ; The design purpose of the loss function is to improve the generalization ability of the model in the real scenario by jointly optimizing the prediction accuracy and distribution alignment.

[0089] where represents the efficiency deviation predicted by the STA-LSTM network, represents the efficiency deviation calculated by CFD simulation, forces the data-driven model to approximate the physical simulation results to ensure the prediction accuracy, and the KL divergence term is approximately calculated by the Jensen-Shannon distance. sim represents the feature distribution of the simulation data (generated by CFD), real represents the feature distribution of the real sensor data, is the trade-off coefficient (usually taken as 0.1 - 0.5, determined by cross-validation) p sim‖ p real) functions to reduce the domain difference (Domain Gap) between the simulation and real data and avoid model overfitting to the simulation data; Model fine-tuning stage: Freeze the parameters of the first three layers of the encoder and only fine-tune the fully connected layer and the attention weight matrix.

[0090] Step S4: Generate an optimization strategy for operating parameters using the deep reinforcement learning algorithm. According to the actual efficiency value and the device status parameters, generate an optimization strategy for operating parameters using the deep reinforcement learning algorithm. In this application, the deep reinforcement learning algorithm is based on the policy network; the specific method is as follows: Step S4.1: Define the state space S = , , , , W], reflecting the real-time operating efficiency of the water turbine and the health status of the equipment, where is the actual efficiency, is the measured temperature of the bearing, is the change rate of the inlet head (change amount per minute, unit: m / s), is the flow fluctuation coefficient (standard deviation of the sliding window, window length 60 s), is the change rate of the inlet head (change amount per minute, unit: m / s), W is the wear index (calculated based on the wavelet packet energy entropy of the acoustic emission signal using an acoustic emission sensor, value range: 0 - 1); the above data are collected and calculated through sensors; Step S4.2: Define the action space B = , is the guide vane opening adjustment amount (-2° to +2°, resolution 0.1°), is the excitation current correction value (-5% to +5% of the rated value, step size 0.5%), is the cooling water flow adjustment amount (-10% to +10%, for temperature control); Step S4.3: Construct the reward function ), where, St ) is calculated based on the equipment safety status, is the effective value of vibration, Pt ∣ is the L2 norm of the guide vane opening and the excitation current adjustment amount, is the exponential form of the vibration term, is the actual efficiency, is the reward weight (for example, the value is 0.5) (for example, the value is 0.1, to suppress mechanical vibration and avoid equipment damage). Pt ∣ is the action smoothing term, which reduces the frequent large adjustments of the control parameters, prolongs the equipment life, and the weight 0.1 balances the efficiency and operation stability, avoiding high-frequency oscillation of the strategy; In this step, a multi-objective trade-off reward function is designed to balance efficiency improvement and equipment safety. To reward safe operation and prevent faults such as overheating and cavitation, gradient-sensitive safety guidance is achieved through piecewise functions and non-linear transformations (such as tanh). Specifically, St ) is calculated as follows: ) =

[0091] is the Item safety factor, including: vibration safety reward = , when the vibration approaches the threshold, apply an exponential penalty to force load reduction, where represents the vibration safety factor, = , takes a value of 5 mm / s, the safety threshold 0; temperature safety reward = St −0.8)), (output range [-1, +1]), feature: a steep gradient area is formed near = 0.8 to prompt the strategy to respond quickly to temperature rise, where t represents the temperature safety factor, t = , T represents the measured temperature of the bearing, represents the calibrated temperature, takes a value of 85 °C, the safety threshold t 0; wear safety reward = β ⋅ln( w + 0.1), function: cumulative penalty for long-term wear to avoid accelerating equipment aging due to short-term efficient operation, where represents the wear safety factor, = W, the safety threshold ≥ 0.7; pressure pulsation safety reward = , function: quadratic function penalty for sudden pressure difference to prevent cavitation damage, where represents the pressure pulsation safety factor, = , represents the pressure pulsation amplitude, represents the rated pressure of the water turbine, the safety threshold 0.

[0092] Step S4.4: Train the policy network using the PPO algorithm.

[0093] The actual efficiency value and equipment status parameters are used as the input of the deep reinforcement learning algorithm, and the optimized strategy of the operating parameters, that is, the optimized action, is output.

[0094] Step S5: Verify the safety of the optimized strategy of the operating parameters Purpose of safety verification: Ensure that the operating parameter adjustment scheme generated by deep reinforcement learning (DRL) can improve efficiency while absolutely guaranteeing equipment safety and operating stability. For example: Prevent equipment damage and failures, including mechanical safety, to avoid fatigue fractures of key components such as bearings and runner wheels due to excessive vibration (e.g., exceeding the ISO 7919 standard), thermal safety: prevent material degradation or insulation failure caused by overheating of bearings or windings, and hydraulic safety: suppress cavitation (e.g., by verifying pressure pulsations ), to avoid cavitation damage on the runner surface; Avoid operating risks, including: Transient process stability: Ensure that optimization actions (such as rapid guide vane adjustment) do not trigger water hammer effects or sudden flow changes, resulting in pipeline rupture, and Extreme operating condition adaptability: Verify the robustness of the strategy under abnormal conditions such as flood periods (high flow rates) and winter (low water temperatures).

[0095] Safety verification can adopt: Digital twin simulation verification, Technical principle: By constructing a high-fidelity digital twin model, simulate the transient response of the water turbine under optimization actions. For example, combine CFD (Computational Fluid Dynamics) to simulate the flow field pressure distribution, and a multi-body dynamics model to predict vibration characteristics; Thermodynamic efficiency test, Parameter measurement: Install high-precision temperature and pressure sensors at the high-pressure side and low-pressure side sections to measure water flow temperature, pressure, and flow rate. Iterative optimization algorithm: Calculate the absolute efficiency of the water turbine through the first law of thermodynamics, and use an iterative algorithm to reduce the number of sensors and improve accuracy; Design of Experiments (DOE) and comparative verification: Orthogonal experiment: Design multi-factor (such as water head, flow rate, guide vane angle) and multi-level experiments to analyze the primary and secondary effects of optimization actions; Comparative verification, Horizontal comparison: Compare with traditional PID control strategies to verify the advantages of deep reinforcement learning (DRL) in terms of annual average efficiency improvement (2.7%) and failure rate reduction (70%), and Vertical comparison: Through historical data backtracking, evaluate the adaptability of optimization actions under different seasons and hydrological conditions; Reverse verification based on historical data, Principle: Use historical operating data (such as fault records, operating condition parameters) to verify the rationality of the optimization strategy, including: Data playback: Input historical operating condition data into the DRL policy network to generate "virtual actions", and Comparative analysis: Compare virtual actions with actual historical operations to evaluate whether the strategy has avoided known faults; It should be noted that the above are examples of existing methods for safety verification, and those skilled in the art can also adopt other existing safety verification methods.

[0096] After the optimization strategy is verified, convert the optimization strategy into control instructions to adjust the guide vane opening and excitation current, and at the same time output visual monitoring results and maintenance warnings.

[0097] Based on the above method, this embodiment provides an online monitoring and operation optimization system for water turbine efficiency, including: A data acquisition module that acquires raw data and real-time collects the operation data of the water turbine through a multi-source sensor array. The operation data includes vibration signals, pressure distribution, flow parameters, rotational speed, and temperature; A data preprocessing module: preprocesses the raw data; A hybrid model construction module: takes the preprocessed data as input and calculates the actual efficiency value through a mechanism-data hybrid model. The mechanism-data hybrid model includes: A physical model, a hydrodynamic simulation based on the Navier-Stokes equation; A data-driven model that uses a spatio-temporal attention network to fuse multi-modal time-series data; An optimization strategy module: generates an operation parameter optimization strategy using a deep reinforcement learning algorithm according to the actual efficiency value and equipment status parameters; A safety verification module: verifies the safety of the operation parameter optimization strategy.

[0098] It should be noted that the above functional modules correspond one by one to the steps of the online monitoring and operation method for water turbine efficiency based on artificial intelligence provided in Embodiment 3. The specific functions implemented are the same as those of the online monitoring and operation optimization of water turbine efficiency provided in Embodiment 3, and the beneficial effects achieved are also the same as those of the online monitoring and operation method for water turbine efficiency based on artificial intelligence provided in Embodiment 3.

[0099] The above content is only to illustrate the technical idea of the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution according to the technical idea proposed by the present invention fall within the protection scope of the present invention.

Claims

1. An on-line monitoring and operation optimization method for the efficiency of a hydraulic turbine, characterized in that, It includes the following steps: Obtain the operation data of the water turbine; Obtain the actual efficiency value by using the mechanism data hybrid model according to the operation data of the water turbine; Obtain the equipment status parameters, and generate an operation parameter optimization strategy by using the deep reinforcement learning algorithm according to the actual efficiency value and the equipment status parameters; Conduct safety verification on the operation parameter optimization strategy.

2. The online monitoring and operation optimization method for the efficiency of a water turbine according to claim 1, wherein The operation data of the water turbine includes: the vibration signal of the water turbine, the pressure distribution of the spiral case, the water turbine flow parameters, the water turbine speed, and the temperature inside the bearing housing.

3. The on-line monitoring and operation optimization method of the water turbine efficiency according to claim 1, characterized in that It also includes the following steps: Perform progressive wavelet denoising, outlier removal, and time series feature extraction on the operation data of the water turbine.

4. The online monitoring and operation optimization method for the efficiency of a hydraulic turbine according to claim 1, wherein The specific method for obtaining the actual efficiency value by using the mechanism data hybrid model according to the operation data of the water turbine is as follows: Simulate and generate the multi-condition data of the water turbine; Construct a physical model according to the multi-condition data, specifically as follows: Use proper orthogonal decomposition to reduce the dimension of the CFD full-order model, establish a theoretical efficiency calculation module, and output the theoretical efficiency; Construct a data-driven model, specifically as follows: Adopt the spatio-temporal attention network STA-LSTM, and the input layer fuses the operation data of the water turbine and the theoretical efficiency value to output the efficiency deviation; Couple the STA-LSTM encoder with the physical model through transfer learning to form a hybrid inference module, and output the actual efficiency according to the hybrid inference module in combination with the theoretical efficiency, the efficiency deviation, and the operation data of the water turbine.

5. The online monitoring and operation optimization method for the water turbine efficiency according to claim 4, characterized in that, The calculation formula of the theoretical efficiency is as follows: Wherein, represents the theoretical efficiency, represents the measured temperature of the bearing, represents the calibrated temperature, represents the maximum allowable temperature, γ empirical coefficient, represents the output torque, represents the angular velocity, represents the flow rate of the water turbine, represents the temperature compensation coefficient, represents the inlet head, represents the outlet head.

6. The on-line monitoring and operation optimization method for the water turbine efficiency according to claim 4, characterized in that, The spatio-temporal attention network includes: a time attention module and a spatial graph convolution module; The time attention module is used to calculate the weights of each time step: The spatial graph convolution module aggregates node features with the sensor topology as the adjacency matrix: In the formula, represents the hidden state of the LSTM at time step t, represents the trainable weight matrix, represents the trainable query vector, represents the node feature matrix of the represents the trainable weight matrix, represents the adjacency matrix, represents the degree matrix, represents the weights for each time step, and softmax() represents the normalization of the scores for all time steps, represents the corresponding time step, represents enhancing the non - linear expression ability, is represented as the ReLU activation function.

7. The on-line monitoring and operation optimization method for the water turbine efficiency according to claim 4, characterized in that The specific method for coupling the STA-LSTM encoder with the physical model through transfer learning to form a hybrid inference module is as follows: Use the multi-condition data as the pre-training data set; Pre-train the STA-LSTM encoder by using the pre-training data set, and the loss function is: In the formula, represents the efficiency deviation predicted by the STA-LSTM network, represents the efficiency deviation calculated by CFD simulation, The divergence term is approximately calculated by the Jensen-Shannon distance, represents the characteristic distribution of simulation data, represents the characteristic distribution of real sensor data, Trade-off coefficient; Couple the pre-trained STA-LSTM encoder with the physical model to form a hybrid inference module.

8. The online monitoring and operation optimization method for the water turbine efficiency according to claim 1, characterized in that, The specific method for obtaining the equipment status parameters, and generating an operation parameter optimization strategy by using the deep reinforcement learning algorithm according to the actual efficiency value and the equipment status parameters is as follows: The device state parameters include state space parameters, and the state space S is defined as , , , , W], where is the actual efficiency, is the measured temperature of the bearing, is the flow fluctuation coefficient, is the change rate of the inlet head, and W is the wear index; Define the action space B = , is the guide vane opening adjustment amount, is the excitation current correction value, is the cooling water flow adjustment amount; Construct the reward function ), where Calculated based on the device safety status is the effective vibration value Pt ∣ is the L2 norm of the guide vane opening and the excitation current adjustment amount is the exponential form of the vibration term is the actual efficiency is the reward weight is the punishment weight; Train the policy network according to the state space, action space, and reward function and by using the PPO algorithm.

9. The online monitoring and operation optimization method for the efficiency of a hydraulic turbine according to claim 8, characterized in that, The calculation method is as follows: Among them, is the th item weight, is the th safety factor. The safety factor includes vibration safety factor, temperature safety factor, wear safety factor and pressure pulsation safety factor. is the safety reward, including vibration safety reward, temperature safety reward, wear safety reward and pressure pulsation safety reward, specifically as follows: Among them, represents the vibration safety factor, represents the vibration safety reward; Among them, represents the temperature safety factor, is the temperature safety reward, represents the hyperbolic tangent function; Among them, represents the wear safety factor, is the wear safety reward, is the weight coefficient; Among them, represents the pressure pulsation safety factor, is the pressure pulsation safety reward.

10. Online monitoring and operation optimization system for hydraulic turbine efficiency, characterized in that, It includes: A data acquisition module, which is used to obtain the operation data of the water turbine; An actual efficiency module, which is used to obtain the actual efficiency value by using the mechanism data hybrid model according to the operation data of the water turbine; An operation parameter optimization strategy generation module, which is used to obtain the equipment status parameters, and generate an operation parameter optimization strategy by using the deep reinforcement learning algorithm according to the actual efficiency value and the equipment status parameters; A safety verification module, which is used to conduct safety verification on the operation parameter optimization strategy.

Images