Method for quickly generating static stress data of runner of hydroelectric generating unit based on neural network
By combining fully connected residual networks with deep learning and CFD methods, the real-time monitoring challenge of static stress analysis of hydropower turbine runner blades was solved, enabling rapid and accurate generation of static stress data, which is suitable for real-time monitoring and condition prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HOHAI UNIV
- Filing Date
- 2024-12-20
- Publication Date
- 2026-06-26
AI Technical Summary
In existing technologies, static stress analysis of hydropower turbine runner blades cannot be monitored in real time or at high frequency. Traditional methods have low computational efficiency and high equipment requirements, which cannot meet the needs of real-time monitoring.
By employing a neural network-based approach, particularly a fully connected residual network, combined with a deep learning model, high-precision data is obtained through finite element analysis and CFD methods. This enables the rapid generation of static stress data under unknown working conditions.
It achieves efficient and accurate static stress data generation, reduces the requirements for hardware equipment, is suitable for real-time monitoring and condition prediction, and improves calculation speed and accuracy.
Smart Images
Figure CN120030875B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for rapidly generating static stress data of hydropower turbine runners based on neural networks, belonging to the technical field of hydropower turbine runner static stress data generation. Background Technology
[0002] As a crucial component of hydropower stations, the static stress analysis of the turbine blades is essential for ensuring the safety and operational efficiency of the equipment. Under the influence of high-intensity water flow, the blades endure complex stresses, directly impacting the turbine's operating status and lifespan. As a renewable energy power generation device, the turbine system generates a large amount of status data during operation. This data covers key indicators such as turbine stress, speed, and water pressure, providing information about the turbine's operational status.
[0003] In current engineering practice, the static stress of hydroelectric turbine blades cannot usually be directly measured by sensors. Traditional static stress calculation methods typically rely on physical models and experimental data. These methods require finite element analysis and complex mesh generation, followed by high-precision auxiliary calculations using CFD. However, these methods are computationally inefficient, time-consuming, and place high demands on computing equipment, making them unsuitable for real-time or high-frequency monitoring needs. Summary of the Invention
[0004] Purpose of the invention: To address the problems and shortcomings of existing technologies, this invention provides a method for rapidly generating static stress data of hydropower turbine runners based on neural networks. Compared with traditional finite element analysis methods, this method utilizes deep learning models to capture complex nonlinear relationships in the data, achieving efficient and high-precision static stress prediction with less computational resources. This improves computational speed and reduces hardware requirements for equipment, providing an innovative solution for real-time monitoring and state prediction of hydropower units.
[0005] Technical solution: A method for rapidly generating static stress data of hydropower unit runners based on neural networks, comprising the following steps:
[0006] Step 1: Input the static stress data of the hydropower unit runner blades;
[0007] Step 2: Process the hydropower unit data input in Step 1;
[0008] Step 3: Construct a static stress generation model based on a fully connected residual network;
[0009] Step 4: Train the static stress generation model;
[0010] Step 5: Calculate the relative error generated by the model. If the relative error exceeds the preset threshold, proceed to step 6; otherwise, return to step 4.
[0011] Step 6: Input the parameters of the unknown arbitrary working condition and the point data on the hydropower unit into the trained model to obtain the static stress distribution of the runner under the unknown working condition.
[0012] In step 1, the static stress data of the hydropower turbine runner blades requires high-precision auxiliary calculation using a mesh-based finite element method and a CFD method. Data was collected from the hydropower turbine after meshing using the finite element method, and then calculated using the CFD method. A total of 40 million data points for 73 operating conditions of the hydropower turbine were collected. Each data point includes three-dimensional data of the turbine blade location, hydropower turbine operating condition data, and hydropower turbine static stress data.
[0013] The three-dimensional data of the turbine blade locations are as follows: Y Z This indicates the position coordinates of a point on the hydroelectric turbine runner; the hydroelectric unit operating data includes the turbine's operating degree. and water head This indicates the operating status of the hydroelectric generator; the static stress data at the hydroelectric generator unit points represents the magnitude of the static stress at the points on the hydroelectric generator runner blades. ).
[0014] Step 2 specifically refers to:
[0015] The hydropower unit data input in step 1 is processed as follows: First, statistical analysis is performed on the hydropower unit data to calculate the range of static stress values; second, the operating condition data of the hydropower unit and the three-dimensional data of the runner position are normalized; finally, based on the statistical data of the static stress values, when the static stress variance is greater than... The transformation is used to map the range of static stress values to a smaller range, while reducing the order of magnitude difference between static stress data, and the relative error after transformation is evaluated; otherwise, normalization is performed.
[0016] The normalization process is as follows:
[0017]
[0018] in, This is data that needs to be normalized. yes The mean, yes variance yes The normalized value.
[0019] Using power transform Mapping static stress to a smaller interval is specifically as follows:
[0020]
[0021] in, It is the original static stress. It is the static stress obtained after power transformation. After the transformation, the difference between static stresses is reduced.
[0022] The relative error is evaluated, and the range of static stress values is taken as... , The interval length is calculated as The expanded interval is divided into several sub-intervals. .
[0023]
[0024] in, .
[0025] If the model-generated values are within the corresponding range, the relative error rate The range of the maximum value is:
[0026]
[0027] when At that time, relative error rate ;
[0028] If the model-generated values are not in the corresponding interval, and differ by m intervals, the relative error rate is... The maximum value is:
[0029]
[0030] or
[0031]
[0032] The fewer the number of difference intervals, the lower the relative error rate, and the better the model's generation effect.
[0033] Step 3, constructing the static stress generation model based on a fully connected residual network, specifically involves:
[0034] The residual module originates from ResNet in convolutional neural networks. Since neural network models are non-convex functions, and deep networks often suffer from the vanishing gradient problem due to excessive layers, making optimization difficult, the residual module alleviates this problem through its internal residual structure and skip connections. The static stress generation model based on a fully connected residual network is a fully connected network built upon the residual module. The residual module is incorporated into a multilayer perceptron model to capture deep connections between data.
[0035] The residual module can alleviate the gradient problem. Specifically, the residual module is:
[0036]
[0037]
[0038] in, These are the input values for the residual module. , It is a weight matrix. , It corresponds to the bias. It is an activation function.
[0039] Step 4 specifically includes:
[0040] First, the overall model is trained. The dataset is divided into training, validation, and test sets based on the location of all work conditions, with each set representing 80%, 10%, and 10% of the total data points. The Model-Specific Error (MSE) loss function is used during training, and the model is fine-tuned based on the results of the loss function to complete the training.
[0041] The loss function MSE is the mean squared error, specifically:
[0042]
[0043] Where Q represents the number of operating conditions for the hydropower unit, and N represents the total number of points for each operating condition. Indicates the first Type of working condition Actual values of static stress at each point; Indicates the first Type of working condition Predicted static stress values at each point.
[0044] Compared with the prior art, the present invention has the following beneficial effects:
[0045] I. Wide Range and Reliable Data Generation: The rapid generation method for static stress data of hydropower turbine runners based on neural networks in this invention is built upon high-quality data obtained from deep learning algorithms and high-precision CFD calculations, ensuring the reliability of the input data. A fully connected network model using residual modules alleviates the gradient problem. Through data transformation, the model deeply learns the inherent mapping relationship between operating condition data and point coordinates to static stress data, ensuring the reliability and accuracy of the generated data. Simultaneously, this invention overcomes the drawbacks of mesh generation in the finite element method, enabling the generation of static stress data at all points on the runner, data under different operating conditions, and static stress data under unknown operating conditions.
[0046] Fast data generation: Compared with traditional calculation methods, the model generates data quickly and has low requirements for hardware, making it highly practical in real-world applications. Attached Figure Description
[0047] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;
[0048] Figure 2 This is a schematic diagram of the residual module in the static stress generation model based on a fully connected residual network in an embodiment of the present invention;
[0049] Figure 3 This is a schematic diagram of the static stress generation model structure based on a fully connected residual network according to an embodiment of the present invention;
[0050] Figure 4 This is a schematic diagram illustrating the evaluation of static stress generated by the static stress generation model based on a fully connected residual network according to an embodiment of the present invention. Figure (a) is a schematic diagram of the actual static stress data on the turbine blade, and Figure (b) is a schematic diagram of the static stress data generated by the model on the turbine blade. The values of the colored bands in the schematic diagrams are... ,in It is the magnitude of the static stress at that point on the blade of the hydroelectric generator unit;
[0051] Figure 5 This is a 3D model diagram of static stress generation for an unknown operating condition of a hydroelectric generator unit in an example of the present invention. Detailed Implementation
[0052] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.
[0053] While CFD-assisted calculations based on mesh generation using the finite element method offer high computational accuracy, data processing is slow. Deep learning, with its powerful feature extraction capabilities, has gained favor among researchers. Combining this with static stress data obtained from high-precision assisted calculations, a rapid generation model for static stress data of hydroelectric turbine runners was established to generate static stress data for unknown operating conditions. A specific implementation method is provided below:
[0054] like Figure 1 As shown, the method for rapidly generating static stress data of hydropower unit runners based on neural networks includes the following steps:
[0055] Step 1: Collection of static stress data for hydropower unit runner blades
[0056] Static stress data of hydropower turbine runners cannot be directly obtained through sensors. To acquire this data, high-precision auxiliary calculations using mesh-based finite element analysis (FEM) and CFD methods are necessary. To meet the requirements of convenient and rapid model generation, the acquired hydropower turbine runner static stress data only includes the turbine's opening degree, head, three-dimensional coordinates of the runner points, and static stress values at those points, totaling 40 million data points for 73 different operating conditions. Notably, the turbine's opening degree, head, and three-dimensional coordinates of the runner points are directly obtainable data, eliminating the need for cumbersome calculations.
[0057] The three-dimensional data of the turbine blade locations are as follows: Y Z This indicates the position coordinates of a point on the hydroelectric turbine runner; the hydroelectric unit operating data includes the turbine's operating degree. and water head This indicates the operating status of the hydroelectric generator; the static stress data at the hydroelectric generator unit points represents the magnitude of the static stress at the points on the hydroelectric generator runner blades. ).
[0058] The operating conditions of hydropower units are far more than the 73 operating conditions we collected. The operating conditions we collected are shown below. The hydropower unit opening includes: 0%, 27.86%, 33.33%, 38.40%, 39.32%, 46.89%, 53.80%, 54.51%, 69.00%, 69.92%, 77.23%, 100%, etc.; the hydropower unit head includes: 164m, 180m, 195m, 205m, 216m, 222m, 230m, 240m, 251m.
[0059] Step 2: Analyze and process the hydropower unit data.
[0060] In this example, firstly, statistical analysis is performed on the hydropower unit data, and the range of static stress values is as follows: The mean static stress is The variance is .
[0061] Secondly, the operating data of the hydropower unit and the three-dimensional data of the runner location are normalized and standardized.
[0062] Finally, based on the variance of the static stress in the impeller, a power transform is used to map the range of static stress values to a smaller interval. This also reduces the order-of-magnitude difference between static stress data and evaluates the relative error after transformation.
[0063] The normalization process is as follows:
[0064]
[0065] in, This is data that needs to be normalized. yes The mean, yes variance yes The normalized value.
[0066] In this example, the calculated variance of the static stress is: The data processing method using power transformation is used to map the original static stress value range to a smaller range, thereby reducing the difference between larger and smaller data and avoiding the loss function's bias towards larger data.
[0067] The relative error is evaluated, and the static stress range is denoted as... , The length of the interval is denoted as After expansion, the interval is divided into several sub-intervals. .
[0068]
[0069] in,
[0070] If the model-generated values are within the corresponding range, the relative error rate The range of the maximum value is:
[0071]
[0072] when At that time, relative error rate ;
[0073] If the model-generated values are not in the corresponding interval, and differ by m intervals, the relative error rate is... The maximum value is:
[0074]
[0075] or
[0076]
[0077] The fewer the number of difference intervals, the lower the relative error rate, and the better the model's generation effect.
[0078] By dividing the generation interval into several sub-intervals, the static stress generated by the model can meet the requirement of a low relative error if it falls within the corresponding interval.
[0079] Using power transform Mapping static stress to a smaller interval is specifically as follows:
[0080]
[0081] in, It is the original static stress. It is the static stress obtained after power transformation. After the transformation, the difference between static stresses is reduced.
[0082] Step 3: Construct a static stress generation model;
[0083] The specific structure of the model is as follows Figure 3 As shown. To meet the requirement of rapid static stress generation, the generation model simplifies the model structure. The model mainly consists of two parts: residual modules and fully connected modules.
[0084] Residual module mechanism such as Figure 2 As shown. First, the input to the residual module is... ,
[0085]
[0086]
[0087] in, , , , , , It is the weight matrix of the two fully connected layers of the residual module. , These are the bias vectors of two fully connected layers. This is the input batch size. It is the length of the input vector, which is present in each fully connected layer. A number of neurons are used to ensure that the shape of the input X remains unchanged after passing through a fully connected layer. The original input is fused through two consecutive weighted layers, and then the fusion weights are obtained through an activation function.
[0088] The fully connected module is constructed based on neurons, and the input to the fully connected layer is... The output is ,
[0089]
[0090] in, , , , It is the weight matrix of the fully connected layer. It corresponds to the bias vector. This is the input batch size. It is the length of the input vector. It represents the number of neurons in the fully connected layer, and is a learnable parameter.
[0091] Static stress generation model structure as follows Figure 3 As shown, the basic architecture is a residual module. The specific network uses two residual blocks repeated 3 times, for a total of 6 residual blocks, and the model structure has a total of 31 layers.
[0092] The specific architecture of the model is as follows: input layer, two fully connected layers, residual blocks repeated twice, a single fully connected layer, a residual block repeated twice, a single fully connected layer, a residual block repeated twice, a fully connected layer, and an output layer. The maximum width of the model is 256, and the activation function used is... .
[0093] Step 4: Train the generative model built in Step 3;
[0094] For the constructed model, a training set, a validation set, and a test set are built. The model is initialized and trained using the training set data, the model parameters are adjusted using the validation set and evaluation metrics, and the model performance is tested using the test set.
[0095] In this embodiment, the proportions of the training dataset, validation dataset, and test dataset are 80%, 10%, and 10%, respectively.
[0096] The model is trained using the loss function MSE and then fine-tuned to complete the training. Finally, the static stress values at the points on the turbine runner of the hydroelectric generator are output.
[0097] The loss function MSE is the mean squared error, specifically:
[0098]
[0099] Where Q represents the number of operating conditions for the hydropower unit, and N represents the total number of points for each operating condition. Indicates the first Type of working condition Actual values of static stress at each point; Indicates the first Type of working condition Predicted static stress values at each point.
[0100] like Figure 4 As shown, after the generative model was trained, the static stress generation results were compared using the model on the blades of a hydroelectric generator. Figure (a) is a schematic diagram of the actual static stress data on the blades of the hydroelectric generator, and Figure (b) is a schematic diagram of the static stress data generated by the model on the blades of the hydroelectric generator. The values of the colored bands in the schematic diagrams are... ,in This represents the magnitude of the static stress at that point on the turbine blade. Comparing Figure (a) and Figure (b), the model generation result in Figure (b) matches the result in Figure (a).
[0101] Step 5: Calculate the relative error generated by the model. If the relative error exceeds the preset threshold, proceed to step 6; otherwise, return to step 4.
[0102] The reliability of the model-generated values is tested using relative error, evaluating both the average relative error on the dataset and the proportion of data with low error rates. After generating output values from the model, the static stress values of the impeller are obtained using the inverse power transform. ,
[0103]
[0104]
[0105] in, It is the model obtained from step 4. These are the input values (rotor 3D coordinates, hydropower unit head, hydropower unit opening degree). It is the inverse power transform.
[0106] The generated relative error is calculated as follows:
[0107]
[0108] It is the true value of static stress. It is the value generated by the model for static stress. It is used to evaluate the relative error of the predicted value.
[0109] In this example, the model trained in step 4 is evaluated on both the training and validation sets, and the results are as follows:
[0110] Table 1. Relative error of static stress generated by the model
[0111]
[0112] The model's average error on the hydropower turbine runner validation set was 2.13%, with 98.54% of the data points having a relative error of less than 10%, and only 0.09% of the data points having a relatively high relative error. The relative error was low, and in the vast majority of the data points, the relative error was less than 10% or even lower.
[0113] Step 6: Input the parameters of the unknown arbitrary working condition and the point data on the hydropower unit into the trained model to obtain the static stress distribution of the runner under the unknown working condition.
[0114] In step 4, the model is trained using known operating conditions. In step 5, the reliability of the model is determined based on the relative error performance in the validation set. The generation speed of the deep learning model is much faster than the high-precision CFD calculation using the finite element method, quickly obtaining the static stress distribution under unknown operating conditions. In this example, a head of 135m and an unknown opening of 56% are selected. The model is used to generate the static stress distribution on the impeller and the blades, and the results are as follows. Figure 5 .
[0115] In terms of generation speed, the model generates static stress data for a single working condition on a personal computer in seconds, which is much faster than current traditional calculation methods.
Claims
1. A method for rapidly generating static stress data of a hydroelectric generator runner based on neural networks, characterized in that, Includes the following steps: Step 1: Input the static stress data of the hydropower unit runner blades; Step 2: Process the hydropower unit data input in Step 1; Step 3: Construct a static stress generation model based on a fully connected residual network; The static stress generation model based on the fully connected residual network is a fully connected network built on residual modules, with the residual modules added to the multilayer perceptron model. Step 4: Train the static stress generation model; Step 5: Calculate the relative error generated by the model. If the relative error exceeds the preset threshold, proceed to step 6; otherwise, return to step 4. Step 6: Input the parameters of the unknown arbitrary working condition and the point data on the hydropower unit into the trained model to obtain the static stress distribution of the runner under the unknown working condition. In step 2, the hydropower unit data from step 1 is processed; first, statistical analysis is performed on the hydropower unit data to calculate the range of static stress values. Secondly, the operating condition data of the hydropower unit and the three-dimensional data of the turbine blade locations were normalized and standardized. Finally, based on the statistics of the static stress data of the hydropower unit locations, when the static stress variance is greater than... Transformation is used to map the range of static stress values to a smaller range, thereby reducing the order-of-magnitude differences between static stress data; conversely, the static stress point data of hydropower units are normalized.
2. The method for rapidly generating static stress data of a hydroelectric generator runner based on neural networks according to claim 1, characterized in that, In step 1, the finite element method and CFD method with mesh generation are used to assist in the calculation of the static stress data of the turbine runner blades of the hydropower unit; the hydropower unit data after auxiliary calculation is obtained, and each data includes three-dimensional data of the turbine blade points, hydropower unit operating condition data and hydropower unit static stress data. The three-dimensional data of the turbine blade locations are as follows: Y Z , represents the position coordinates of a point on the hydroelectric turbine runner; the hydroelectric generator operating condition data includes the opening degree and head of the hydroelectric generator, indicating the operating status of the hydroelectric generator; the hydroelectric generator point static stress data is the magnitude of the static stress at a point on the hydroelectric turbine runner blade.
3. The method for rapidly generating static stress data of a hydroelectric generator runner based on a neural network according to claim 1, characterized in that, The normalization process is as follows: in, This is data that needs to be normalized. yes The mean, yes variance yes The normalized value.
4. The method for rapidly generating static stress data of a hydroelectric generator runner based on a neural network according to claim 1, characterized in that, The power transform is used to map the range of static stress values to a smaller range, specifically: in, It is the original static stress. It is the static stress obtained after power transformation. After the transformation, the difference between static stresses is reduced.
5. The method for rapidly generating static stress data of a hydroelectric generator runner based on a neural network according to claim 1, characterized in that, Step 4 specifically includes: First, the overall model is trained. The data of all working conditions in the dataset are divided into training set, validation set and test set according to the location of the data points, with 80%, 10% and 10% respectively. During the training process, the loss function MSE is used, and the model is fine-tuned based on the result of the loss function to complete the training of the model.