Dynamic compensation method for mechanism model of water turbine governor based on deep residual learning

Abstract

A dynamic compensation method for a mechanism model of a water turbine governor based on deep residual learning comprises the steps that firstly, multi-dimensional real-time data in the operation process of the water turbine governor is collected, and a mechanism model covering the electrical-mechanical-hydraulic multi-physics field coupling effect is constructed; comparing the guide vane opening degree predicted by the mechanism model with the actually measured guide vane opening degree to obtain a residual error as a training sample; a dual-channel deep residual learning model including a TCN module, a 1D-CNN module, a channel attention module and a bidirectional LSTM module is constructed, and a composite loss function fusing data fitting loss and physical constraint loss is adopted for training; and superposing the output of the residual error model to the output of the mechanism model to realize high-precision dynamic compensation. And a performance monitoring mechanism is set, and when the mean absolute error MAE of the opening degree of the guide vane exceeds 0.2% and is continuously overtime, the model is automatically triggered to be retrained. According to the method, the precision, robustness and adaptive capacity of the speed regulator simulation model are effectively improved, and the method is suitable for a hydraulic turbine speed regulator digital twin system.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of water turbine governor, and particularly relates to a water turbine governor mechanism model dynamic compensation method based on deep residual learning. BACKGROUND

[0002] The water turbine governor is an important control system of the hydropower station, and its performance directly affects the stability and economy of power production. The operation and maintenance of the governor involves the interweaving of knowledge of water power, machinery, electricity, and hydraulic pressure, and the system structure is complex, the operation mode is various, and the construction of an accurate governor simulation model to reproduce the dynamic characteristics of the real system can provide guiding suggestions for the operation and maintenance of the speed regulation system and improve the technical level and operation and maintenance reliability of the speed regulation system. The traditional governor simulation model usually has large model simplification error and poor adaptability of time-varying parameters, and is difficult to adapt to the nonlinear, time-varying, and high-frequency coupling scenarios, and cannot realize dynamic high-fidelity mapping between the physical world and the digital space. With the deep penetration of digital twin technology in the field of industrial control, the traditional model has been difficult to meet the high-precision simulation demand. The application proposes a water turbine governor mechanism model dynamic compensation method based on deep residual learning, which combines mechanism modeling and data-driven methods, establishes a deep learning residual model according to the difference between the actual output value of the governor system and the output value of the mechanism model, compensates the precision error caused by unknown dynamics, and continuously monitors the model performance to realize dynamic compensation of the water turbine governor mechanism model.

[0003] The traditional modeling method of the water turbine governor abstracts the complex physical system into a linear time-invariant (LTI) model, converts the differential equation into an algebraic equation through Laplace transform, and thus establishes a simulation model. Since the water turbine governor is a complex system with electrical-mechanical-hydraulic coupling, the number of equipment components is large, the traditional simulation model ignores the complex characteristics and time-varying factors such as nonlinear friction, oil pressure and temperature, and component wear, and it is difficult to accurately reflect the actual dynamic characteristics of the governor under various working conditions. The main problems are large model simplification error, poor adaptability of time-varying parameters, distortion of multi-physical field decoupling, insufficient coverage of special working conditions, and inability to support online optimization.

[0004] The data-driven model constructed by learning a large amount of governor operation data shows great potential in the digital twin system, but its interpretability has always been a problem in engineering practice, and there are structural defects that cannot be ignored, such as complete loss of simulation in working conditions not covered by the training data, and it is easy to fall into the double dilemma of lack of physical priori and insufficient dynamic adaptability. SUMMARY

[0005] The technical problem solved by the present application is to provide a dynamic compensation method for a hydraulic turbine governor mechanism model based on deep residual learning, aiming to solve the problems of high-dimensional difficulty in analyzing traditional simulation models, precision error caused by model simplification, and other problems, overcome the defects of poor interpretability of data-driven models and poor adaptability to unlearned samples, improve the simulation model precision of the hydraulic turbine governor, reproduce the dynamic characteristics of the real system, meet the demand for online adjustment of the model, and continuously improve the accuracy of the model in practical application.

[0006] To solve the above technical problems, the technical solution adopted by the present application is: A dynamic compensation method for a hydraulic turbine governor mechanism model based on deep residual learning, comprising the following steps: (1) Collecting multi-dimensional real-time data in the operation process of the hydraulic turbine governor, wherein the multi-dimensional real-time data at least includes a guide vane opening measured value, a PID control output signal Ypid, an operating water head, an oil tank oil pressure, and environmental parameters; (2) Constructing a hydraulic turbine governor mechanism model covering the electrical-mechanical-hydraulic multi-physical field coupling effect, wherein the hydraulic turbine governor mechanism model is established based on physical laws, contains a flow continuity equation, a momentum equation, and a motion equation, and accurately restores the dynamic characteristics of each key component of the hydraulic turbine governor; (3) Inputting the PID control output signal Ypid into the hydraulic turbine governor mechanism model to obtain the guide vane opening predicted by the hydraulic turbine governor mechanism model, and calculating the residual error between the guide vane opening predicted by the hydraulic turbine governor mechanism model and the guide vane opening measured value as a training sample; (4) Constructing a double-channel deep residual learning model, (5) Training the double-channel deep residual learning model using the training sample, wherein a composite loss function is used in the training process, the composite loss function includes a data fitting loss and a physical constraint loss, and the weights of the data fitting loss and the physical constraint loss are dynamically adjusted during the training process; (6) Superimposing the output of the double-channel deep residual learning model to the output of the hydraulic turbine governor mechanism model to obtain the compensated output of the hydraulic turbine governor model; (7) Continuously monitoring the performance indicators of the compensated output of the hydraulic turbine governor model, and when the guide vane opening average absolute error MAE exceeds the preset threshold 0.2% and continuously exceeds the preset time threshold, triggering the retraining process of the double-channel deep residual learning model to realize dynamic compensation of the hydraulic turbine governor mechanism model.

[0007] Preferably, the steady-state residual is a low-frequency and long-period model deviation of the hydro-turbine governor caused by equipment aging, component wear or environmental slow changes in the balanced working condition, and the TCN module adopts an expanded convolution structure with an expansion factor set as 1, 2, 4 and 8 in an exponential growth manner, and a convolution kernel size of 3.

[0008] Preferably, the transient residual is a high-frequency and short-time prediction deviation of the hydro-turbine governor generated in the process of load mutation, control instruction switching or disturbance response, and the 1D-CNN module adopts a multi-scale convolution kernel with a convolution kernel size set as {3, 5, 7} to capture transient dynamic characteristics of different frequencies.

[0009] Preferably, the channel attention mechanism realizes feature fusion in the following manner: The feature channels output by the TCN module and the feature channels output by the 1D-CNN module are respectively subjected to global average pooling, the pooling results are input into two fully connected layers and subjected to ReLU activation, and then the weights of each channel are generated through a Sigmoid function, the original output features of the TCN module and the 1D-CNN module are weighted by using the weights, and adaptive fusion of steady-state features and transient features is realized.

[0010] Preferably, the physical constraint loss includes an energy conservation constraint violation degree loss and a dimension consistency loss, and is used to constrain the output of the dual-channel deep residual learning model to comply with physical laws.

[0011] Preferably, the retraining process includes: taking the parameters of the current dual-channel deep residual learning model as pre-training weights, fine-tuning training on a dataset composed of newly added hydro-turbine governor operation data, and evaluating the performance of the fine-tuned model on a validation set, and only when the performance of the fine-tuned model is better than the current deployed model, the dual-channel deep residual learning model is updated and deployed.

[0012] Preferably, when the prediction confidence of the dual-channel deep residual learning model is lower than a set threshold or the output of the dual-channel deep residual learning model exceeds a safety boundary, the output of only the hydro-turbine governor mechanism model is automatically switched to be used to ensure the operation safety of the hydro-turbine governor system.

[0013] Preferably, the dual-channel deep residual learning model includes: a time domain convolution network (TCN) module for extracting steady-state residual features, a one-dimensional convolutional neural network (1D-CNN) module for extracting transient residual features, an attention module for adaptively weighting and fusing features output by the TCN module and the 1D-CNN module based on a channel attention mechanism, a bidirectional LSTM module for modeling the time sequence dependency relationship of the fused features by the attention module; A hydro-turbine governor digital twin system, which adopts the deep residual learning-based hydro-turbine governor mechanism model dynamic compensation method to construct a high-fidelity simulation model for reproducing the dynamic characteristics of a real hydro-turbine governor and supporting online operation decision-making and fault early warning.

[0014] The present application can achieve the following beneficial effects: 1. The present application combines mechanism modeling and data-driven, based on real physical laws, and is constructed by flow continuity equation, momentum equation and motion equation, etc. to build a hydro-turbine governor mechanism model that can accurately simulate the electro-hydraulic multi-physical field coupling effect. The dynamic error not covered by the mechanism model is compensated by deep learning, a dynamic compensation mechanism is introduced, the model performance is continuously monitored, parameter changes, equipment degradation, equipment maintenance or other influencing factors are captured and model retraining is triggered, thereby realizing dynamic compensation of the simulation model. Real-time data enables the model to maintain environmental awareness, the mechanism model represents system behavior, the residual compensation model recognizes limitations, and the optimized model gives the system the ability of self-adaptive evolution.

[0015] 2. The governor simulation model constructed by the method can reproduce the dynamic characteristics of the real system through digital twin technology, has high fidelity, good time-varying adaptability and real-time performance, effectively reduces the simulation error, and can describe the characteristics of the simulation model, reduces the "illusion", uncertainty and unexplainability brought by the data-driven model, and lays a foundation for the application of digital twin technology in the hydro-turbine governor. BRIEF DESCRIPTION OF DRAWINGS

[0016] The present application will be further described below in conjunction with the drawings and examples: Figure 1 The overall architecture of the residual model of the present application is shown in Figure 2 The method flowchart of the present application is shown in DETAILED DESCRIPTION

[0017] The preferred scheme is shown in Figures 1-2 A deep residual learning-based hydro-turbine governor mechanism model dynamic compensation method is as follows: 1. Actual data acquisition A multi-dimensional data acquisition system is constructed to realize real-time capture of key parameters through a high-precision sensor network, including but not limited to operating water head, guide vane opening, PID control output (Ypid), oil tank pressure, and other environmental parameters. The Ypid input is used to drive the mechanism model of the speed regulator, which in turn drives the high-precision simulation of the guide vane opening. The residual data between the measured guide vane opening and the guide vane opening output by the mechanism model constitute the "training sample library" for deep learning, and also serve as an important indicator for evaluating the accuracy of the model and iterating the model.

[0018] 2. Mechanism model construction The water turbine governor is the core control unit of hydroelectric power generation, mainly including PID controllers, proportional servo valves, hand-automatic switching valves, servo motor linear displacement converters, guide valves, emergency stop valves, main pressure regulating valves, accident pressure regulating valves, segmented closing valves, servomotors, pressure tanks, oil pump groups, oil collecting tanks, and other physical components. The modeling of the governor needs to consider fluid dynamics, mechanical dynamics, and control system theory in three major physical fields.

[0019] The mechanism model of the water turbine governor established based on physical laws has the advantages of clear model structure, high extrapolation reliability, and the ability to provide theoretical basis for anomaly detection. When constructing the mechanism model, all key characteristics of the necessary physical components need to be fully considered, such as flow characteristics, pressure characteristics, dynamic response characteristics, oil pipe pressure and flow loss characteristics, oil system oil balance characteristics, etc. Based on the key physical parameters, each component in the governor is accurately restored, and the introduction of related equipment such as water turbines, generators, and water conduits should also be considered. A differential equation system with strict physical meaning is constructed to ensure that the simulation model covers the multi-physical field coupling effects between electrical, mechanical, and hydraulic.

[0020] 3. Residual model construction The residual is the difference between the actual value and the predicted value, i.e., the difference between the actual system output and the mechanism model output: (1) By analyzing the residual, the accuracy of the mechanism model can be evaluated, the prediction ability of the model can be improved, and factors not considered by the mechanism model, such as equipment aging and unmodeled dynamics, can be compensated. The architecture of the deep learning model of the residual is shown in Figure 1 The TCN and 1D-CNN are used to extract residual features, adaptive attention fusion, and LSTM time series dynamic modeling, respectively, to achieve accurate simulation of the mechanism model residual.

[0021] 1) Data preparation The input data includes the input parameters of the original mechanism model, steady-state residual sequences, transient residual sequences, and other factors that may affect the residual (such as the operating state of the governor and environmental conditions); the target data is the total residual at each time step.

[0022] 2) Steady-state residuals Steady-state residuals describe the persistent, low-frequency model errors of a system under equilibrium conditions. (2) Steady-state residuals, caused by long-term factors, carry crucial information such as system performance degradation and the emergence of anomalies. These signals typically manifest as low-frequency, long-period, and strongly causal physical processes. Traditional time-series models often perform poorly in extracting these features, while Temporal Convolutional Networks (TCNs), with their unique mathematical structure, become the optimal solution.

[0023] TCN employs convolutional operations, enabling efficient parallel computation and offering advantages in training and inference on large-scale data. TCN can extract features at different scales by stacking multiple convolutional layers, with each layer processing the input through convolutional kernels of different sizes, thereby capturing features across different time ranges in the sequence data. This multi-scale information extraction capability makes it more sensitive to local dependencies in the sequence data. TCN can accept fixed-length or variable-length sequence data as input, making it more convenient to process sequence data of different lengths.

[0024] TCN captures long-term dependencies through dilated convolution: (3) in, Let d be the hidden state of layer l at time t; l σ is the dilation factor of the l-th layer (set to 1, 2, 4, 8); σ is the ReLU activation function; k is the kernel size (set to 3).

[0025] 3) Transient residual Transient residuals describe the short-term, high-frequency prediction bias of a system during dynamic processes. (4) Transient residuals reflect the distortion of a model's response to rapid disturbances and control defects. These signals differ fundamentally from steady-state residuals and require specialized modeling methods. One-dimensional convolutional neural networks (1D-CNNs) are a variant of convolutional neural networks specifically designed for processing sequential data, such as time series and text. They possess three core advantages: local perception mechanism, hierarchical feature decoupling, and edge enhancement characteristics, perfectly matching the physical nature of transient residuals. Therefore, 1D-CNNs can be used to extract high-frequency dynamic features and capture residual changes caused by rapid disturbances (such as load changes and control responses).

[0026] Use multi-scale convolution to capture transient features at different frequencies: (5) Where, k m ∈{3,5,7} respectively capture transient changes of different frequencies.

[0027] 4) Channel attention mechanism Channel attention mechanisms focus on the weights of different channels in a feature map. Each channel can be viewed as a "filter" for a specific feature, capturing a particular pattern or feature in the image. In standard convolutional neural networks, the feature importance of all channels is usually equal. However, there are differences in feature importance between channels, with some channels contributing more to the current task. Channel attention mechanisms automatically learn the importance of each channel and adjust the representation of the input data accordingly, thereby increasing the model's focus on key information, helping the model better capture crucial information, and further improving model performance.

[0028] By introducing a channel attention mechanism, multi-source features are intelligently fused and their importance weights are learned. When steady-state data dominates, the weights of the TCN branch are amplified; when transient changes are significant, the role of the CNN branch is strengthened. (6) (7) (8) Where c is the feature channel index; δ is the ReLU activation function; σ is the Sigmoid activation function; W1 and W2 are weight matrices.

[0029] 5) Time series modeling Residual signals in industrial scenarios contain two key time-dependent patterns: delayed response and periodic fluctuations. LSTM is a special type of RNN that can learn long-term dependencies. It controls information flow through gating mechanisms (forget gate, input gate, and output gate), solving the gradient vanishing problem of traditional RNNs. Its cell state acts as a "memory line," storing key information long-term. The forget gate selectively discards useless information (such as noise), the input gate updates the state to store new features (such as fault impacts), and the output gate controls the intensity of information output (such as the residual decay process). LSTM offers advantages in industrial time-series tasks, including adaptive learning of arbitrary-length delayed dependencies, high robustness with small samples, and millisecond-level real-time inference. In residual prediction, bidirectional LSTM addresses the dynamic asymmetry of residual signals through a dual-path time-series modeling mechanism. The forward propagation path accurately captures the hysteresis response during residual generation, while the backward propagation path learns the decay inertia during the residual dissipation phase, adding time-series dependency modeling capabilities to the fused features.

[0030] Using bidirectional LSTM to learn the time-dependent patterns of residuals: (9) (10) (11) 6) Training and Validation The residual dataset is divided into training and validation sets, and a loss function guides model training. The loss function mainly consists of two parts: data fitting loss, which is the mean squared error (MSE) between the output residual and the actual residual; and physical constraint loss, such as the degree of violation of energy conservation constraints and dimensional consistency loss. The weights of the loss function are dynamically adjusted during the training process: early on, the focus is on data fitting, and later on, physical constraints are strengthened. This is done in three stages, gradually increasing the complexity of the training data (historical data > recent data > special operating condition data), while simultaneously adjusting the weights of the learning rate and physical constraints.

[0031] After training, the model performance is validated on the test set. Performance metrics include mean square error of guide vane opening (MSE) and mean absolute error of guide vane opening (MAE) to ensure that the model has good generalization ability.

[0032] 4. Compensation governor model The main body of the governor model is the governor mechanism model. The residual model is used to compensate for factors not considered or errors not captured by the mechanism model. Finally, the output of the model is: Speed ​​governor model output = Speed ​​governor mechanism model output + Residual model output.

[0033] 5. Dynamic compensation process 1) Continuous monitoring Changes in time-varying parameters, equipment degradation, equipment maintenance, or other influencing factors can lead to a decrease in the accuracy of the governor model. Therefore, a continuous monitoring mechanism is introduced to monitor the performance indicators and data drift of the governor model in the production environment in real time. An automated monitoring and alarm system is established to track the residuals of the hybrid model in real time and set threshold alarms for key indicators, including mean absolute error, maximum frequency deviation, and maximum guide vane opening deviation. The system checks whether the model output exceeds the safety boundary, and automatically triggers a switch to the mechanistic model when an anomaly occurs, switching to the mechanistic model output when the prediction confidence is low.

[0034] 2) Retraining Model retraining / update is a continuous cyclical process that triggers optimization processes based on monitoring feedback and data changes to adapt to changes in data distribution.

[0035] When the monitoring system detects a decline in model performance that persists for more than one day (e.g., MAE (guide vane opening) > 0.2%), it triggers an anomaly warning and model retraining. New data is collected, the existing model is used as pre-training weights, and the residual model is trained again on the new dataset. The performance of the new residual model is then validated. The output of the adjusted governor model is as follows: Speed ​​governor model output = Mechanism model output + New residual model output.

[0036] This invention has the following advantages: 1. Multiphysics-based full-dimensional mechanism modeling: Based on real physical laws and intrinsic working principles, and according to the flow continuity equation, momentum equation, and motion equation, the key characteristics of necessary physical components are fully considered, such as flow characteristics, pressure characteristics, dynamic response characteristics, oil pipeline pressure and flow loss characteristics, and oil system quantity balance characteristics. A differential equation system covering the electrical-mechanical-hydraulic coupling effect is established to accurately reproduce each component in the turbine governor and construct the mechanism model of the governor.

[0037] 2. Dual-channel physical constraint residual deep learning: For the residual between the predicted and actual values ​​of the governor mechanism model, a deep learning model (TCN-CNN-LSTM-Attention) is constructed to capture the system complexity that is difficult to describe precisely with physical laws. Features are extracted for steady-state residuals and transient residuals respectively. Adaptive channel attention is used to intelligently fuse multi-source features. The temporal dependencies between features are learned through LSTM. At the same time, a physical constraint layer is introduced, such as Hamiltonian constraints based on energy conservation and dimensional consistency constraints, to ensure that the model output conforms to physical laws.

[0038] 3. Continuous monitoring and dynamic compensation mechanism for the model: Continuously monitor the performance of the governor model, identify abnormal performance phenomena (MAE (guide vane opening) > 0.2%), trigger model retraining when performance deteriorates, collect new data, retrain the residual model, and verify the performance of the new model to achieve dynamic compensation of the governor model.

[0039] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A dynamic compensation method for a turbine governor mechanism model based on deep residual learning, characterized in that, Includes the following steps: Collect multi-dimensional real-time data during the operation of the turbine governor. The multi-dimensional real-time data includes at least the measured value of guide vane opening, PID control output signal Ypid, operating head, oil pressure in the pressure tank, and environmental parameters. Construct a mechanism model for a hydro turbine governor that encompasses the multi-physics coupling effects of electrical, mechanical, and hydraulic systems; The PID control output signal Ypid is input into the turbine governor mechanism model to obtain the guide vane opening predicted by the turbine governor mechanism model. The residual between the guide vane opening predicted by the turbine governor mechanism model and the measured value of the guide vane opening is calculated and used as a training sample. Construct a dual-channel deep residual learning model. The dual-channel deep residual learning model is trained using training samples. A composite loss function is used during the training process, which includes data fitting loss and physical constraint loss. The weights of the data fitting loss and physical constraint loss are dynamically adjusted during the training process. The output of the dual-channel deep residual learning model is superimposed on the output of the turbine governor mechanism model to obtain the compensated turbine governor model output. The performance indicators output by the compensated turbine governor model are continuously monitored. When the mean absolute error (MAE) of the guide vane opening exceeds the preset threshold and continues to exceed the preset time threshold, the retraining process of the dual-channel deep residual learning model is triggered to achieve dynamic compensation of the turbine governor mechanism model.

2. The dynamic compensation method for the turbine governor mechanism model based on deep residual learning as described in claim 1, characterized in that, The dual-channel deep residual learning model includes: The temporal convolutional network TCN module is used to extract steady-state residual features; A one-dimensional convolutional neural network (1D-CNN) module for extracting transient residual features; An attention module that adaptively weights and fuses the features output by the TCN module and the 1D-CNN module based on a channel attention mechanism; A bidirectional LSTM module is used to model the temporal dependencies of features fused by the attention module.

3. The dynamic compensation method for the turbine governor mechanism model based on deep residual learning as described in claim 2, characterized in that, The steady-state residual is the low-frequency, long-period model deviation of the turbine governor under balanced operating conditions caused by equipment aging, component wear, or slow environmental changes. The TCN module adopts an dilated convolution structure, with the dilation factor set to 1, 2, 4, and 8 in an exponential growth pattern, and the convolution kernel size is 3.

4. The dynamic compensation method for the turbine governor mechanism model based on deep residual learning as described in claim 2, characterized in that, The transient residual is the high-frequency, short-time prediction deviation generated by the turbine governor during load changes, control command switching, or disturbance response. The 1D-CNN module uses multi-scale convolution kernels with a kernel size set of {3, 5, 7} to capture transient dynamic features of different frequencies.

5. The dynamic compensation method for the turbine governor mechanism model based on deep residual learning as described in claim 2, characterized in that, The channel attention mechanism achieves feature fusion in the following ways: Global average pooling is performed on the feature channels output by the TCN module and the 1D-CNN module respectively. The pooling results are input into two fully connected layers and activated by ReLU. The weights of each channel are generated by the Sigmoid function. The original output features of the TCN module and the 1D-CNN module are weighted using the weights to achieve adaptive fusion of steady-state features and transient features.

6. The dynamic compensation method for the turbine governor mechanism model based on deep residual learning as described in claim 1, characterized in that, The physical constraint loss includes the energy conservation constraint violation degree loss and the dimensional consistency loss, which are used to constrain the output of the dual-channel deep residual learning model to conform to physical laws.

7. The dynamic compensation method for the turbine governor mechanism model based on deep residual learning as described in claim 1, characterized in that, The retraining process includes: using the parameters of the current dual-channel deep residual learning model as pre-training weights, fine-tuning the training on a dataset composed of newly added turbine governor operation data, evaluating the performance of the fine-tuned model on a validation set, and updating and deploying the dual-channel deep residual learning model only when the performance of the fine-tuned model is better than the currently deployed model.

8. The dynamic compensation method for the turbine governor mechanism model based on deep residual learning as described in claim 1, characterized in that, When the prediction confidence of the dual-channel deep residual learning model is lower than a set threshold or the output of the dual-channel deep residual learning model exceeds the safety boundary, the system automatically switches to using only the output of the turbine governor mechanism model to ensure the operational safety of the turbine governor system.

9. The dynamic compensation method for the turbine governor mechanism model based on deep residual learning as described in claim 1, characterized in that, The mechanism model of the turbine governor is based on physical laws, including the flow continuity equation, momentum equation and motion equation, and accurately reproduces the dynamic characteristics of each key component of the turbine governor.

10. A digital twin system for a water turbine governor, characterized in that, The digital twin system for the turbine governor uses the dynamic compensation method based on deep residual learning for the turbine governor mechanism model as described in any one of claims 1 to 9 to construct a high-fidelity simulation model, which is used to reproduce the dynamic characteristics of the real turbine governor and support online operation and maintenance decision-making and fault early warning.

Images