Hydroelectric generating unit runner static stress data rapid generation method based on neural network

Abstract

The invention discloses a hydroelectric generating set runner static stress data rapid generation method based on a neural network. The method comprises the steps of 1, inputting hydroelectric generating set runner blade static stress data; 2, processing the hydroelectric generating set data input in the step 1; 3, constructing a static stress generation model based on the full-connection residual network; 4, training a static stress generation model; 5, calculating a relative error generated by the model, if the relative error exceeds a preset threshold value, executing the step 6, otherwise, returning to the step 4; and 6, inputting parameters of any unknown working condition and point location data on the hydroelectric generating set into the trained model to obtain static stress distribution of the rotating wheel under the unknown working condition. According to the method, the complex nonlinear relation in the data is captured, efficient and high-precision static stress prediction is achieved with few computing resources, the computing speed is increased, the hardware requirement for equipment is reduced, and an innovative solution is provided for real-time monitoring and state prediction of the hydroelectric generating set.

Description

Technical Field

[0001] The invention relates to a method for quickly generating static stress data of a hydroelectric unit runner based on a neural network, and belongs to the technical field of generating static stress data of a hydroelectric unit runner. Background Art

[0002] As an important part of a hydropower station, the static stress analysis of the runner blades of a hydroelectric generator is crucial to ensure the safety and operating efficiency of the equipment. The blades are subjected to complex stresses under the action of high-intensity water flow, and their stress conditions directly affect the operating status and life of the hydroelectric generator. As a renewable energy power generation equipment, the runner system of a hydroelectric generator will generate a large amount of status data during its operation. These data cover key indicators such as runner force, speed, water pressure, etc., and provide us with information on the running status of the runner.

[0003] In existing engineering practices, the static stress of hydropower turbine blades cannot usually be directly measured by sensors. Traditional static stress calculation methods usually rely on physical models and experimental data. Such methods need to be implemented through finite element calculations and complex meshing, and then CFD is used for high-precision auxiliary calculations. However, these methods have low calculation efficiency, long time consumption, and high requirements for computing equipment, and are not suitable for real-time or high-frequency monitoring needs. Summary of the invention

[0004] Purpose of the invention: In view of the problems and shortcomings of the prior art, the present invention provides a method for quickly generating static stress data of a hydropower unit impeller based on a neural network. Compared with the traditional finite element analysis method, this method uses a deep learning model to capture the complex nonlinear relationship in the data, and achieves efficient and high-precision static stress prediction with fewer computing resources, thereby improving the calculation speed and reducing the hardware requirements for the equipment, providing an innovative solution for real-time monitoring and status prediction of hydropower units.

[0005] Technical solution: A method for quickly generating static stress data of a hydropower unit runner based on a neural network, comprising the following steps:

[0006] Step 1: Input the static stress data of the runner blades of the hydropower unit;

[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: training the static stress generation model;

[0010] Step 5: Calculate the relative error generated by the model. If the relative error exceeds the preset threshold, execute step 6; otherwise, return to step 4.

[0011] Step 6: Input the parameters of the unknown arbitrary working conditions 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 conditions.

[0012] In the step 1, the static stress data of the runner blade of the hydropower unit needs to be assisted by the finite element calculation method of mesh division and the CFD method for high-precision auxiliary calculation. The data obtained by the CFD method auxiliary calculation after the hydropower unit divides the grid points based on the finite element calculation method is collected, and a total of 40 million hydropower unit point data of 73 working conditions of the hydropower unit are collected, and each data includes three-dimensional data of the blade point of the hydropower unit, working condition data of the hydropower unit, and static stress data of the hydropower unit point.

[0013] The three-dimensional data of the blade point of the hydropower unit are X Location (m), Y Location (m), and Z Location (m), which represent the position coordinates of the point on the hydropower unit runner; the operating data of the hydropower unit include the opening (Opening) and water head (Height) of the hydropower unit, indicating the operating status of the hydropower unit; the static stress data of the hydropower unit point is the size of the static stress (Equivalent (von-Mises) Stress (Pa)) on the point on the hydropower unit runner blade.

[0014] The step 2 is specifically as follows:

[0015] The hydropower unit data input in step 1 are processed. First, the hydropower unit data are statistically analyzed to calculate the value range of the static stress value. Secondly, the operating data of the hydropower unit and the three-dimensional data of the runner point are normalized. Finally, according to the statistical data of the static stress value, when the static stress variance is greater than 10 10 , use transformation to map the static stress value interval to a smaller interval, while reducing the order of magnitude difference between the static stress data, and evaluate the relative error after transformation; otherwise, perform normalization and standardization.

[0016] Normalization and standardization processing, specifically:

[0017]

[0018] Among them, x is the data to be normalized, x mean is the mean of x, x var is the variance of x, and x′ is the normalized value of x.

[0019] The power transformation F(x) is used to map the static stress to a smaller interval, specifically:

[0020]

[0021] Among them, y is the original static stress, and y′ is the static stress obtained after power transformation. After the transformation, the difference between the static stresses is reduced.

[0022] The relative error is evaluated and the static stress value interval is calculated as The interval length is denoted as s = ba, and the expanded interval is divided into several sub-intervals [a, t 1 ],[t 2 ,t 3 ],……,[t n ,b].

[0023]

[0024] Among them, t 0 =a=1,t n =m n ,t n+1 =b=m n+1 .

[0025] If the model generates values ​​in the corresponding interval, the relative error rate e i The maximum value range is:

[0026]

[0027] When n>100, the relative error rate e i <0.02;

[0028] If the model generates a value that is not in the corresponding interval, the difference is m intervals, and the relative error rate is e i The maximum value of is:

[0029]

[0030] or

[0031]

[0032] The fewer the number of phase difference intervals, the lower the relative error rate and the better the model generation effect.

[0033] The step 3 of constructing a static stress generation model based on a fully connected residual network is specifically as follows:

[0034] The residual module comes from Res Net in the convolutional neural network. Since the neural network model is a non-convex function and the gradient vanishing problem occurs due to too many layers in the deep network, it is difficult to optimize. The residual module alleviates the gradient vanishing problem in the model through the internal residual structure and jump connection design. The static stress generation model based on the fully connected residual network is a fully connected network built based on the residual module. The residual module is added to the multi-layer perceptron model to obtain the deep connection between the data.

[0035] The residual module can alleviate the gradient problem. The residual module is specifically:

[0036] F(x)=ReLU(W 2 ReLU(W 1 x+b 1 )+b 2 +x)

[0037] ReLU(x)=max{0,x}

[0038] Among them, x is the input value of the residual module, W 1 , W 2 is the weight matrix, b 1 、b 2 is the corresponding bias, and ReLu(·) is the activation function.

[0039] The step 4 is specifically as follows:

[0040] First, the overall model is trained, where the data of all working conditions in the data set are divided into training set, validation set, and test set according to the point positions of 80%, 10%, and 10%. The loss function MSE is used in the training process, and the model is fine-tuned according to the result of the loss function to complete the model training.

[0041] The loss function MSE is the mean square error, specifically:

[0042]

[0043] Among them, Q is the number of hydropower unit operating conditions, N is the total number of points for each operating condition, and y i,j It represents the actual value of static stress at the jth point in the i-th working condition; It represents the predicted value of static stress at the jth point under the i-th working condition.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] 1. The generated data has a wide range and is reliable: The method for quickly generating static stress data of the hydropower unit runner based on neural network of the present invention is based on high-quality data obtained by deep learning algorithm and high-precision CFD calculation, which ensures the reliability of input data. The fully connected network model using the residual module alleviates the gradient problem, and utilizes the data transformation method to enable the model to deeply learn the intrinsic mapping relationship between the working condition data and the point coordinates to the static stress data, thus ensuring the reliability and accuracy of the generated data. At the same time, the present invention eliminates the drawbacks of the finite element method for dividing the grid, and can generate static stress data at all points on the runner, data of different working conditions, and static stress data under unknown working conditions.

[0046] Fast data generation speed: Compared with traditional calculation methods, the model generation speed is fast and the requirements for hardware equipment are low, making it more practical in actual applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0048] Figure 2 Schematic diagram of a residual module in a static stress generation model based on a fully connected residual network in an embodiment of the present invention;

[0049] Figure 3 Schematic diagram of the structure of a static stress generation model based on a fully connected residual network according to an embodiment of the present invention;

[0050] Figure 4 Schematic diagram of static stress evaluation generated by a static stress generation model based on a fully connected residual network in an embodiment of the present invention. FIG. (a) is a schematic diagram of true value static stress data on a hydropower unit blade, and FIG. (b) is a schematic diagram of model-generated static stress data on a hydropower unit blade. The values ​​of the color bands in the schematic diagram are log 10 F, where F is the static stress of the blade of the hydropower unit at that point;

[0051] Figure 5 A 3D model diagram is generated for the static stress of the unknown working condition of the hydropower unit in the example of the present invention. DETAILED DESCRIPTION

[0052] The present invention is further explained below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0053] Although the CFD auxiliary calculation based on the finite element method has high calculation accuracy, the data calculation is slow. Deep learning is favored by researchers for its powerful feature extraction ability. Combined with the static stress data obtained by high-precision auxiliary calculation, a rapid generation model of static stress data of the runner of a hydropower unit is established to generate static stress data of unknown working conditions. The following is a specific implementation method:

[0054] like Figure 1 As shown, the method for quickly generating static stress data of a hydropower unit runner based on a neural network includes the following steps:

[0055] Step 1: Collection of static stress data of hydropower unit runner blades

[0056] The static stress data of the runner of a hydropower unit cannot be directly obtained through sensors. In order to obtain the static stress data of the runner blades, it is necessary to use the finite element calculation method of meshing and the CFD method for high-precision auxiliary calculations. In order to meet the needs of convenient and fast model generation, the obtained static stress data of the runner of the hydropower unit only include the opening of the hydropower unit, the head of the hydropower unit, the three-dimensional coordinates of the runner point, and the static stress value of the runner point. A total of 40 million hydropower unit point data for 73 working conditions of the hydropower unit. Among them, the opening of the hydropower unit, the head of the hydropower unit, and the three-dimensional coordinates of the runner point are all data that can be directly obtained, and there is no need to obtain them through tedious data calculations.

[0057] The three-dimensional data of the blade point of the hydropower unit are X Location (m), Y Location (m), and Z Location (m), which represent the position coordinates of the point on the hydropower unit runner; the operating data of the hydropower unit include the opening (Opening) and water head (Height) of the hydropower unit, indicating the operating status of the hydropower unit; the static stress data of the hydropower unit point is the size of the static stress (Equivalent (von-Mises) Stress (Pa)) on the point on the hydropower unit runner blade.

[0058] The operating conditions of hydropower units are far more than the 73 conditions we have collected. Some of the operating conditions we have collected are shown below: the openings of hydropower units include: 0%, 27.86%, 33.33%, 38.40%, 39.32%, 46.89%, 53.80%, 54.51%, 69.00%, 69.92%, 77.23%, 100%, etc.; the heads of hydropower units include: 164m, 180m, 195m, 205m, 216m, 222m, 230m, 240m, 251m.

[0059] Step 2: Analyze and process the hydropower unit data

[0060] In this example, firstly, the data of hydropower units are statistically analyzed, and the static stress value range is [10 4 ,10 8 ], the average static stress is 1.6×10 7 , with a variance of 1.1×10 14 .

[0061] Secondly, the operating data of the hydropower unit and the three-dimensional data of the runner position are normalized;

[0062] Finally, according to the variance of the static stress of the runner, the power transformation is used to map the value interval of the static stress to a smaller interval [1, 100], while reducing the order of magnitude difference between the static stress data, and evaluating the relative error after the transformation.

[0063] Normalization and standardization processing, specifically:

[0064]

[0065] Among them, x is the data to be normalized, x mean is the mean of x, x var is the variance of x, and x′ is the normalized value of x.

[0066] In this example, the variance of the calculated static stress is 10 14 The power transformation method is used to process the data to map the original static stress range to a smaller range, while reducing the difference between large data and small data to avoid the loss function preferring larger data.

[0067] The relative error is evaluated and the static stress interval is recorded as The interval length is recorded as s = ba, and the expanded interval is divided into several sub-intervals [a, t 1 ],[t 2 ,t 3 ],……,[t n ,b].

[0068]

[0069] Among them, t 0 =a=1,t n =n 2 ,t n+1 =b=(n+1) 2

[0070] If the model generates values ​​in the corresponding interval, the relative error rate e i The maximum value range is:

[0071]

[0072] When n>100, the relative error rate e i <0.02;

[0073] If the model generates a value that is not in the corresponding interval, the difference is m intervals, and the relative error rate is e i The maximum value of is:

[0074]

[0075] or

[0076]

[0077] The fewer the number of phase difference intervals, the lower the relative error rate and the better the model generation effect.

[0078] By dividing the generation interval into several subintervals, the static stress generated by the model can meet the lower relative error requirement if it is in the corresponding interval.

[0079] The power transformation F(x) is used to map the static stress to a smaller interval, specifically:

[0080]

[0081] Among them, y is the original static stress, and y′ is the static stress obtained after power transformation. After the transformation, the difference between the static stresses is reduced.

[0082] Step 3: Construct static stress generation model;

[0083] The specific structure of the model is as follows Figure 3 In order to meet the demand for rapid generation of static stress, the generative model simplifies the model structure, which mainly consists of two parts: the residual module and the fully connected module.

[0084] The residual module mechanism is as follows Figure 2 As shown. First, the input of the residual module is

[0085] F(X) = ReLU(ReLU(XW 1 +b 1 )W 2 +b 2 +X)

[0086] ReLU(x)=max{0,x}

[0087] in, W 1 , W 2 is the weight matrix of the two fully connected layers of the residual module, b 1 , d 2is the bias vector of the two fully connected layers, N is the batch size of the input, d is the length of the input vector, and each fully connected layer has d neurons, so that the shape of the input X remains unchanged after passing through the fully connected layer. The weight layers of two consecutive layers are fused with the original input, and the fusion weight is obtained through the activation function.

[0088] The construction of the fully connected module is based on neurons, and the input of the fully connected layer is The output is

[0089] Y = ReLU(XW+b)

[0090] in, W is the weight matrix of the fully connected layer, b is the corresponding bias vector, N is the batch size of the input, d is the length of the input vector, and M is the number of neurons in the fully connected layer, which is a learnable parameter.

[0091] The static stress generation model structure is as follows Figure 3 As shown, the basic architecture is a residual module, and the specific network uses two residual blocks repeated 3 times, a total of 6 residual blocks, and a total of 31 layers in the model structure.

[0092] The specific architecture of the model structure is: input layer, two fully connected layers, residual blocks repeated twice, a single fully connected layer, residual modules repeated twice, a single fully connected layer, residual modules repeated twice, a fully connected layer, and an output layer. The maximum width of the model is 256, and the activation function used is ReLU(·).

[0093] Step 4: Train the generative model built in step 3;

[0094] For the constructed model, construct training set, validation set and test set, use the training set data to initialize and train the model, use the validation set and evaluation indicators to adjust the model parameters, and use the test set to test the model effect;

[0095] In this embodiment, the proportions of the training data set, the validation data set, and the test data set are 80%, 10%, and 10% respectively.

[0096] The loss function MSE is used for training, and the model is fine-tuned to complete the model training, and finally the static stress values ​​of the points on the runner of the hydropower unit are output.

[0097] The loss function MSE is the mean square error, specifically:

[0098]

[0099] Among them, Q is the number of hydropower unit operating conditions, N is the total number of points for each operating condition, and y i,jIt represents the actual value of static stress at the jth point in the i-th working condition; It represents the predicted value of static stress at the jth point under the i-th working condition.

[0100] like Figure 4 As shown in the figure, after the model is trained, the model is used to compare the static stress generation results on the blades of a hydropower unit. Figure (a) is a schematic diagram of the real value static stress data on the blades of a hydropower unit, and Figure (b) is a schematic diagram of the static stress data generated by the model on the blades of a hydropower unit. The values ​​of the color bands in the schematic diagram are log 10 F, where F is the static stress of the blade of the hydropower unit at that point. Comparing Figure (a) and Figure (b), the model generation results in Figure (b) are consistent with the results in Figure (a).

[0101] Step 5: Calculate the relative error generated by the model. If the relative error exceeds the preset threshold, execute step 6; otherwise, return to step 4.

[0102] The reliability of the model-generated values ​​is tested using relative errors, and the average relative error on the data set and the proportion of data with low error rates are evaluated. After the model generates output values, the inverse transformation of the power transformation is used to obtain the rotor static stress generation value Y hat ,

[0103] y hat =G(H(x)) 2

[0104] G(x)=(100x) 2

[0105] Among them, H(x) is the model trained in step 4, x is the input value (three-dimensional coordinates of the rotor, water head of the hydropower unit, and opening of the hydropower unit), and G(x) is the inverse transformation of the power transformation.

[0106] The resulting relative error calculation is as follows:

[0107]

[0108] y is the true value of the static stress, y hat is the value generated by the model for static stress, and δ is the relative error used to evaluate the predicted value.

[0109] In this example, the model trained in step 4 is evaluated on the training set and the validation set, and the results are:

[0110] Table 1 Relative error of static stress generated by the model

[0111]

[0112]

[0113] The average error of the model on the hydropower unit runner validation set is 2.13%. The data points with relative errors below 10% account for 98.54%, and the data points with high relative errors account for only 0.09%. The relative error is low. At the same time, the relative errors of most points are less than 10% or lower.

[0114] Step 6: Input the parameters of the unknown arbitrary working conditions 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 conditions.

[0115] In step 4, the model is trained with known working conditions. In step 5, the reliability of the model is obtained 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, and the static stress distribution performance of the unknown working condition is quickly obtained. In this example, the water head is selected as 135m and the unknown opening is 56%. The model is used to generate the static stress distribution on the runner and the static stress distribution on the blade, and the results are as follows: Figure 5 .

[0116] In terms of generation speed, the model can generate static stress data for a single working condition on a personal computer in seconds, which is much faster than the current traditional calculation methods.

Claims

1. A method for quickly generating static stress data of a hydropower unit runner based on a neural network, characterized in that: The steps include: Step 1: Input the static stress data of the runner blades of the hydropower unit; 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 based on the residual module, and the residual module is added to the multi-layer perceptron model; Step 4: training the static stress generation model; Step 5: Calculate the relative error generated by the model. If the relative error exceeds the preset threshold, execute step 6; otherwise, return to step 4. Step 6: Input the parameters of the unknown arbitrary working conditions 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 conditions.

2. The method for rapidly generating static stress data of a hydropower unit runner based on a neural network according to claim 1 is characterized in that: In the step 1, a finite element calculation method of meshing and a CFD method are used to perform auxiliary calculation on the static stress data of the runner blade of the hydropower unit; the hydropower unit data after auxiliary calculation are obtained, each data including three-dimensional data of the blade point of the hydropower unit, the operating data of the hydropower unit and the static stress data of the point of the hydropower unit; The three-dimensional data of the blade point of the hydropower unit are X Location (m), Y Location (m), and Z Location (m), which represent the position coordinates of the point on the hydropower unit runner; the operating data of the hydropower unit include the opening and water head of the hydropower unit, indicating the operating status of the hydropower unit; the static stress data of the hydropower unit point is the magnitude of the static stress on the point on the hydropower unit runner blade.

3. The method for rapidly generating static stress data of a hydropower unit runner based on a neural network according to claim 1 is characterized in that: In the step 2, the hydropower unit data of step 1 is processed; first, the hydropower unit data is statistically analyzed to calculate the value range of the static stress value; Secondly, the operating data of the hydropower unit and the three-dimensional data of the blade positions of the hydropower unit are normalized. Finally, according to the statistics of the static stress data of the hydropower unit points, when the static stress variance is greater than 10 10 , use transformation to map the value interval of static stress to a smaller interval, and reduce the order of magnitude difference between static stress data; conversely, the static stress point data of hydropower units are normalized.

4. The method for rapidly generating static stress data of a hydropower unit runner based on a neural network according to claim 3 is characterized in that: Normalization and standardization processing, specifically: Among them, x is the data to be normalized, x mean is the mean of x, x var is the variance of x, and x′ is the normalized value of x.

5. The method for rapidly generating static stress data of a hydropower unit runner based on a neural network according to claim 3 is characterized in that: The power transformation F(x) is used to map the static stress value interval to a smaller interval, specifically: Among them, y is the original static stress, and y′ is the static stress obtained after power transformation. After the transformation, the difference between the static stresses is reduced; The relative error is evaluated and the static stress value interval is recorded as The interval length is denoted as s=ba, and the enlarged interval is divided into several sub-intervals [a, t1], [t2, t3], ..., [t n , b]; Where, t0=a=1, t n =n 2 , t n+1 =b=(n+1) 2 If the model generates values ​​in the corresponding interval, the relative error rate e i The maximum value range is: When n>100, the relative error rate e i <0.02; If the model generates a value that is not in the corresponding interval, the difference is m intervals, and the relative error rate is e i The maximum value of is: or 6. The method for rapidly generating static stress data of a hydropower unit runner based on a neural network according to claim 1 is characterized in that: The step 4 is specifically as follows: First, the overall model is trained, where the data of all working conditions in the data set are divided into training set, validation set, and test set according to the point positions at 80%, 10%, and 10%; The loss function MSE is used in the training process. The model is fine-tuned according to the result of the loss function to complete the model training.

Images