Water pump full characteristic curve prediction method based on multi-machine learning model fusion

Abstract

The invention discloses a multi-machine learning model fusion-based water pump full characteristic curve prediction method, which comprises the following steps of: firstly, constructing a regression pool comprising a plurality of machine learning algorithms, and collecting full characteristic curve data of pumps with different specific speeds; then, an improved whale optimization algorithm is adopted to screen out machine learning algorithms from the regression pool to serve as a basic model combination, a K-fold cross validation training strategy is adopted to carry out K rounds of alternate training on each machine learning algorithm in the basic model combination, and then error correlation verification is carried out on the trained basic model combination; performing secondary training on the verified model combination by using a secondary learning device, calculating the Rperformance of the basic model combination, and screening an optimal basic model combination; and training hyper-parameters of the optimal basic model combination by adopting an improved whale optimization algorithm to obtain a multi-machine learning algorithm fusion prediction model which is used for predicting a to-be-tested pump in a full working condition range. The method is high in prediction precision and saves computing resources.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of hydraulic machinery characteristics research and data-driven prediction, and particularly relates to a pump full characteristic curve prediction method based on multi-machine learning model fusion. BACKGROUND

[0002] In the process of global energy transformation to renewable energy, as a mature large-scale energy storage technology, the full characteristic curve of the core equipment of the water power unit is the key basis for the safe and stable operation of the system, the hydraulic transient analysis and the optimal regulation of the power station.

[0003] The current mainstream method of obtaining pump full characteristic curve has significant defects: first, experimental measurement is reliable, but the equipment investment is large, the cycle is long, the cost is high, and there is often a lack of prototype data in the engineering design stage; second, computational fluid dynamics (CFD) simulation requires high-quality grid and rich modeling experience, and the full working condition calculation resource demand is large, and the prediction reliability in extreme working conditions needs to be verified; third, in the theoretical prediction method, the interpolation error based on the similarity principle is large, and the model based on physical equations needs to assume the hydraulic loss coefficient, and the matching degree of the control equation and the actual curve is low.

[0004] In recent years, machine learning has provided a new path for this problem, but the existing technology still has deficiencies: single model has algorithm bias, it is difficult to capture the complex nonlinear and multi-modal characteristics of the curve, it is sensitive to noise and easy to overfit; traditional ensemble methods (such as average method, weighted method) do not fully utilize the complementarity of models, and the prediction accuracy is limited. Therefore, there is an urgent need for a low-cost, high-efficiency, high-precision pump full characteristic curve prediction method to meet the actual engineering needs. SUMMARY

[0005] In view of the problems of high cost and low efficiency of traditional pump full characteristic curve acquisition method, and the problems of insufficient accuracy and poor robustness of existing machine learning methods (single model, traditional integration), the present application provides a pump full characteristic curve prediction method based on multi-machine learning model fusion, which realizes the input of limited working condition data (third quadrant) and the high-precision output of full working condition (0°-360°) dimensionless head and torque.

[0006] The purpose of the present application is achieved by the following technical solutions:

[0007] A pump full characteristic curve prediction method based on multi-machine learning model fusion, comprising the following steps:

[0008] Step one: construct a regression pool of multiple machine learning algorithms including tree-based ensemble methods, linear models, kernel-based methods, neural network methods, probabilistic models, and instance-based learning methods; the input of the machine learning algorithms is the specific speed of the pump and the third quadrant pump operating characteristic curve data, and the output is the dimensionless head and dimensionless torque of the pump in the full operating range;

[0009] Step two: collect full characteristic curve data of pumps with different specific speeds, convert them into dimensionless parameters, and divide them into training set and test set;

[0010] Step three: select no less than two machine learning algorithms from the regression pool as the base model combination using the improved whale optimization algorithm, adopt the K-fold cross-validation training strategy, divide the training set into K subsets, select K-1 subsets for training and 1 subset for validation in each round, perform K rounds of alternating training on each machine learning algorithm in the base model combination, obtain the trained base model combination, save the K prediction results of each trained machine learning algorithm on the validation set and their corresponding true value labels to form the training data set of the secondary learner, and perform error correlation verification on the model; then, use the secondary learner to perform secondary training on the base model combination that passes the error correlation verification, calculate the R² performance of the base model combination, and select the one with the largest R² as the optimal base model combination;

[0011] The improvements of the improved whale optimization algorithm include: using Logistic-Sine-Cosine chaotic mapping function for population initialization, and introducing adaptive weight coefficient;

[0012] Step four: train the hyperparameters of the optimal base model combination using the improved whale optimization algorithm to obtain a multi-machine learning algorithm fusion prediction model;

[0013] Step five: input the specific speed and third quadrant pump operating characteristic curve data of the pump to be predicted into the multi-machine learning algorithm fusion prediction model to obtain the dimensionless head and dimensionless torque of the pump in the full operating range.

[0014] Further, the tree-based ensemble methods include random forest RF, extreme gradient boosting XGBoost, LightGBM, and decision tree DT, the linear models include linear regression LR, ridge regression Ridge, and Bayesian ridge regression, the kernel-based methods include linear kernel SVR and radial basis function kernel SVR-RBF, the neural network methods include multilayer perceptron MLP, the probabilistic models include Gaussian process regression GPR, and the instance-based learning methods include specific speed-based weighted K-nearest neighbor NsKNN.

[0015] Further, in the step three, the error correlation verification is specifically: calculating the error correlation coefficient of any two models in the basic model combination, and only when the error correlation coefficients of all model pairs in the basic model combination are not greater than the error correlation threshold, the basic model combination passes the error correlation verification; wherein the error correlation coefficient of model i and model j is equal to the ratio of the covariance of the error sequence of the two models to the product of the respective standard deviations.

[0016] Further, the ridge regression is selected as the secondary learner.

[0017] Further, the full characteristic curve data of different specific speed pumps collected in the step two cover centrifugal pumps, mixed flow pumps and axial flow pumps.

[0018] Further, the adaptive weight coefficient ranges from 0.2 to 0.9.

[0019] Further, the K-fold cross-validation is specifically 5-fold cross-validation.

[0020] Further, in the step three, the machine learning algorithm selected from the regression pool by the improved whale optimization algorithm is three.

[0021] A water pump full characteristic curve prediction device based on multi-machine learning model fusion comprises one or more processors for realizing a water pump full characteristic curve prediction method based on multi-machine learning model fusion.

[0022] A computer readable storage medium has a program stored thereon, and the program is executed by a processor to realize a water pump full characteristic curve prediction method based on multi-machine learning model fusion.

[0023] Compared with the prior art, the beneficial effects of the present application are as follows:

[0024] (1) The water pump full characteristic curve prediction method based on multi-machine learning model fusion realizes accurate mapping from limited operating condition data to full operating condition characteristics through heterogeneous basic model fusion, intelligent optimization and secondary learner integration, without expensive experimental equipment or a large amount of CFD calculation resources, and is nearly 10 times faster than traditional grid search, and can quickly output full characteristic curves.

[0025] (2) The present application can effectively avoid overfitting through the complementarity of heterogeneous basic models and K-fold cross-validation, and can maintain stable accuracy in the prediction of various types of pumps.

[0026] (3) The method can provide reliable data support for hydraulic machinery design, hydraulic transient analysis and pump station optimization control, and help the safe and efficient operation of pumped storage power stations. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 A schematic diagram of a water pump full characteristic curve prediction method based on multi-machine learning model fusion according to an embodiment of the present application.

[0028] Figure 2 A water pump full characteristic curve and a partition diagram according to an embodiment of the present application, wherein the left diagram is a water pump full characteristic curve and the right diagram is a partition diagram.

[0029] Figure 3 A flowchart of an improved whale optimization algorithm according to an embodiment of the present application.

[0030] Figure 4 A schematic diagram of a secondary learner prediction according to an embodiment of the present application.

[0031] Figure 5 A comparison diagram of actual values and predicted values of a pump with a specific speed of 37.5 according to an embodiment of the present application, wherein the upper diagram is a comparison diagram of actual values and predicted values of WH and the lower diagram is a comparison diagram of actual values and predicted values of WB.

[0032] Figure 6 A comparison diagram of actual values and predicted values of a pump with a specific speed of 134.5 according to an embodiment of the present application, wherein the upper diagram is a comparison diagram of actual values and predicted values of WH and the lower diagram is a comparison diagram of actual values and predicted values of WB.

[0033] Figure 7 A comparison diagram of actual values and predicted values of a pump with a specific speed of 209.3 according to an embodiment of the present application, wherein the upper diagram is a comparison diagram of actual values and predicted values of WH and the lower diagram is a comparison diagram of actual values and predicted values of WB.

[0034] Figure 8 WH predicted values of pumps with specific speeds of 37.5, 134.5 and 209.3 respectively obtained by using single models and optimal combination models.

[0035] Figure 9 WH predicted values of pumps with specific speeds of 37.5, 134.5 and 209.3 respectively obtained by using single models and optimal combination models. DETAILED DESCRIPTION

[0036] The present application will be described in detail below with reference to the accompanying drawings and preferred embodiments, and the purposes and effects of the present application will become more apparent. It should be understood that the specific embodiments described herein are merely intended to explain the present application and are not intended to limit the present application.

[0037] In order to solve the problems of high cost, low efficiency, large error and the like existing in traditional pump full characteristic curve acquisition methods (experimental measurement, CFD simulation and theoretical prediction), the present application provides a water pump full characteristic curve prediction method based on multi-machine learning model fusion, the core of the method is to provide a multi-machine learning algorithm fusion prediction model based on an improved whale optimization algorithm, that is, through heterogeneous basic model fusion, intelligent optimization and secondary learner integration, accurate mapping from limited working condition data to full working condition characteristics is realized.

[0038] As Figure 1 shown, one embodiment of the present application provides a water pump full characteristic curve prediction method based on multi-machine learning model fusion, including the following steps:

[0039] Step one: build a regression pool of multiple machine learning algorithms including tree-based ensemble methods, linear models, kernel-based methods, neural network methods, probability models and instance-based learning methods; the input of the machine learning algorithm is the specific speed of the pump and the third quadrant pump working condition characteristic curve data, and the output is the dimensionless head and dimensionless torque of the pump in the full working condition range.

[0040] Among them, the tree-based ensemble method includes random forest RF, extreme gradient boosting XGBoost, LightGBM and decision tree DT;

[0041] The linear model includes linear regression LR, ridge regression Ridge and Bayesian ridge regression;

[0042] The kernel-based method includes linear kernel SVR and radial basis function kernel SVR-RBF;

[0043] The neural network method includes multilayer perceptron MLP;

[0044] The probability model includes Gaussian process regression GPR;

[0045] The instance-based learning method includes weighted K-nearest neighbor NsKNN based on specific speed, which is based on the engineering knowledge that pumps with similar specific speeds have similar characteristic curves, selects neighbor samples through specific speed similarity measurement, and uses specific speed distance reciprocal as weight for weighted interpolation prediction.

[0046] Step two: collect full characteristic curve data of pumps with different specific speeds, convert them into dimensionless parameters, and divide them into training set and test set.

[0047] In this embodiment, the full characteristic curve data of 23 different specific speed pumps (including centrifugal pumps, mixed flow pumps, and axial flow pumps) are collected, with a data collection interval of 5°, covering the full operating condition range of 0°-360°. Each sample contains the mapping relationship of dimensionless head (WH), dimensionless torque (WB), and corresponding angle parameter θ.

[0048] wherein the specific speed N s is defined as:

[0049]

[0050] wherein N s is the specific speed, H R is the rated head, with a unit of m; N R is the rated rotational speed, with a unit of rpm; Q R is the rated flow rate, with a unit of m 3 / s.

[0051] In addition, the pump full characteristic curve is an important tool for describing the performance change of the pump in the full operating condition range, and its mathematical expression usually uses dimensionless parameters to achieve unified representation under different sizes and operating conditions, such as Figure 2 as shown in the full characteristic curve and zoning diagram of the water pump in this embodiment.

[0052] The traditional pump characteristic curve is characterized by four dimensionless parameters: rotational speed ratio α, flow ratio v, head ratio h, and torque ratio β. These parameters are respectively defined as:

[0053]

[0054] wherein H, T, Q, and N represent the head, torque, flow rate, and rotational speed of the pump, respectively, and H R , T R , Q R , and N R represent the rated head, rated torque, rated flow rate, and rated rotational speed, respectively.

[0055] However, in the four-quadrant analysis of the pump, when the parameters approach zero, the above ratio tends to infinity, leading to numerical calculation difficulties. Secondly, the full characteristic curve shows an "S" type characteristic in the water pump working condition and the reverse working condition area, and the parameters corresponding to the curve exist cross and overlap. The multi-value of flow and torque will bring inconvenience to the transition process analysis. To solve this problem, Suter proposed an improved dimensionless parameter system (Zheng X B, Guo P C, Tong H Z, et al. Improved Suter-transformation for complete characteristic curves of pump-turbine[J]. IOP Conference Series: Earth and Environmental Science, 2012, 15(6): 062015. doi: 10.1088 / 1755-1315 / 15 / 6 / 062015.), namely Suter transformation, which introduces three dimensionless parameters:

[0056]

[0057]

[0058] Where θ is an angle parameter representing the working state of the pump, with a value range of 0° to 360°, covering the entire working area of the pump; WH(θ) and WB(θ) are dimensionless head and torque parameters, respectively. The advantage of Suter transformation is that it avoids the singular point problem in traditional expression, making the pump characteristic curve continuous in the full working condition range.

[0059] Therefore, in this embodiment, after collecting 23 full characteristic curve data of pumps with different specific speeds, the data is first converted into dimensionless parameters by the above formula, and then the data set is divided. According to the theory of hydraulic machinery (centrifugal pump 9≤Ns<81, mixed flow pump 81≤Ns<174, axial flow pump Ns≥174), three representative specific speed samples (centrifugal pump Ns=37.5, mixed flow pump Ns=134.5, axial flow pump Ns=209.3) are selected as independent reserved test sets (not participating in any training and parameter optimization), and the remaining 20 samples are used for model training and K-fold cross-validation. Therefore, the entire data set is divided into training data set and test data set.

[0060] Step three: use the improved whale optimization algorithm to screen not less than two machine learning algorithms from the regression pool as a basic model combination, use K-fold cross-validation training strategy to divide the training set into K subsets, select K-1 subsets for training and 1 subset for validation in each round, perform K rounds of alternating training on each machine learning algorithm in the basic model combination to obtain the trained basic model combination, save the K prediction results of each trained machine learning algorithm on the validation set and the corresponding true value labels to form a training data set of the secondary learner, and perform error correlation verification on the model; then, use the secondary learner to perform secondary training on the basic model combination that passes the error correlation verification, calculate the R² performance of the basic model combination, and select the one with the largest R² as the optimal basic model combination; the improvements of the improved whale optimization algorithm include: using a Logistic-Sine-Cosine chaotic mapping function for population initialization, and introducing an adaptive weight coefficient.

[0061] In the traditional Whale Optimization Algorithm (WOA), the position of each whale represents a potential solution, and the goal is to approach the global optimal solution by simulating the behavior of a humpback whale hunting prey. The hunting behavior of a whale can be divided into three main steps: surrounding prey, bubble net attacking prey, and searching for prey. The traditional Whale Optimization Algorithm performs well in global search and local development, but still has some shortcomings. In this invention, in order to further enhance the global search ability of the algorithm and the diversity of the solution, chaotic mapping and adaptive weight mechanism are introduced into the traditional Whale Optimization Algorithm, and the specific improvements are as follows:

[0062] (1) Chaotic mapping initialization: the initialization position of the whale population is generated by the Logistic-Sine-Cosine chaotic mapping function, and its formula is:

[0063]

[0064] Where X(t) represents the current position of the whale, X(t+1) represents the position of the whale in the next iteration; r is a random number.

[0065] Applying compound chaotic mapping to the initialization of the whale population in the Whale Optimization Algorithm can make the initial population distribution more uniform, thereby improving the algorithm performance.

[0066] (2) Adaptive weight mechanism: by introducing an adaptive weight coefficient, the algorithm balances between global search and local search to improve the diversity of solutions and search efficiency in the optimization process, and the adaptive weight equation is:

[0067]

[0068] wherein, W min is the minimum weight, and in this embodiment, the value is 0.2, W max is the maximum weight, and in this embodiment, the value is 0.9. T is the total number of iterations, t is the current iteration number, λ = 10 is the steepness of the control Sigmoid function, a is the weight factor, and β is the coefficient for controlling the exponential decay rate, a = 0.7, and β = 3.

[0069] Therefore, the improved whale position update formula is as follows:

[0070]

[0071] wherein, represents the position of the optimal whale in the current iteration, and X(t+1) represents the whale position in the next iteration; D p represents the distance between the whale and the prey, D r represents the distance between the current whale individual and the randomly selected whale individual, b is a constant, and l is a random number in (−1, 1).

[0072] The process of iterative optimization of the improved whale optimization algorithm is as shown in Figure 3 .

[0073] In this embodiment, the improved whale optimization algorithm is used to screen three machine learning algorithms trained once from the regression pool as the basic model combination.

[0074] In this embodiment, K = 5 in K-fold cross-validation, that is, the 20 training samples collected in step two are randomly divided into 5 subsets (keeping the proportion of each type of pump balanced), and each time 4 subsets (16 samples) are used as the training set to train the three machine learning algorithms trained once by the improved whale optimization algorithm, and 1 subset is used as the validation set to generate the prediction result (i.e., the “meta-feature”). After 5 iterations, all meta-features are integrated to form the training data set of the secondary learner, effectively avoiding information leakage and overfitting.

[0075] The detailed process of 5-fold cross-validation is as follows:

[0076] Let the original training set be , wherein X i represents the input features of the i-th sample, and y i represents the true value of the i-th sample. Divide D into D1, D2, D3, D4, and D5. In the first round, the training set is , and the validation set is D1. Taking a single model GPR in it as an example, train GPR1, predict the samples in D1, and obtain ; in the second round, the training set is , the validation set is D2, train GPR2, predict the samples in D2, get , the 3rd-5th round in this way, each sample is predicted once, finally get the complete prediction set of GPR , the three basic model training processes in this way, the complete prediction set of RF , the complete prediction set of NsKNN , the training data set [basic model prediction value, real label] of the secondary learner, input feature matrix X meta is:

[0077]

[0078] label vector Y meta is , then the secondary learner learns the mapping .

[0079] Given the training data set , where x i represents the input features (including specific curve data in the third quadrant range and the specific speed), y i represents the output target (complete full-condition characteristic curve), and the simple schematic diagram of the secondary learner prediction is shown in Figure 4 . Assuming that there are M basic models , the prediction process of this step can be represented as:

[0080]

[0081] where g is the secondary learner, is the final prediction result.

[0082] The secondary learner can be selected from linear regression (LR), support vector regression (SVR), and ridge regression (Ridge). Train the three secondary learners with the training data set, and calculate the RMSE and MAE, as shown in Table 1. As can be seen from Table 1, compared with linear regression (LR) and support vector regression (SVR), the prediction error of ridge regression (Ridge) is significantly lower, with an average reduction of 27.6% RMSE and 28.5% MAE compared with LR, and an average reduction of 36.4% RMSE and 37.2% MAE compared with SVR.

[0083] Table 1 Prediction error of different secondary learners

[0084]

[0085] Therefore, the ridge regression is selected as the secondary learner in this embodiment to handle the multicollinearity between the prediction results of the basic models.

[0086] In this step, the error correlation test of the model is performed as follows:

[0087] Calculate any two models M in the combination i and M j The error correlation coefficient :If and only if all model pairs in the basic model combination satisfy Less than or equal to the error correlation coefficient threshold ρ threshold When , the combination is verified by error correlation.

[0088] Error correlation coefficient between model i and model j It is equal to the ratio of the covariance of the error series of the two models to the product of their respective standard deviations, where the covariance is the sum of the products of the error deviations on all n samples, that is, the error of model i on each sample k Subtract the mean error Error with model j e j,k Subtract the mean error The product accumulation of the standard deviation is the square root of the sum of squares of the error deviations of each model; the average error and is the arithmetic mean of the error of the corresponding model on n samples, expressed as follows:

[0089]

[0090]

[0091]

[0092] Where i and j represent different machine learning model numbers, k represents the sample number in the dataset, k ranges from 1 to n, and represents the kth validation sample. For model M i The average error between the predicted value and the true value on the validation set; For model M j The average error between the predicted value and the true value on the validation set. The smaller the error correlation coefficient, the better the complementary effect of subsequent model fusion.

[0093] The secondary learners are trained separately for each candidate combination that passes the error correlation constraint to obtain multiple fusion models. Finally, the determination coefficient R² of each fusion model is calculated on the validation set, and the basic model combination corresponding to the fusion model with the largest R² is selected as the optimal basic model combination.

[0094] In this embodiment, the error correlation threshold ρ is set threshold= 0.7. After the improved whale optimization algorithm search, three candidate combinations of triplets were generated. Among them: combination 1 (RF+GPR+NsKNN) average error correlation coefficient 0.360 < 0.7, passed the screening; combination 2 (MLP+DT+RF) average error correlation coefficient 0.764 > 0.7, was eliminated; combination 3 (LR+XGB+GRR) average error correlation coefficient 0.498 < 0.7, passed the screening. Then the training data set of the selected combination was used to train the secondary learner, and the R² value of the trained basic model combination was calculated, as shown in Table 2.

[0095] Therefore, based on the maximum principle of the coefficient of determination R² of the fusion model, the optimal basic model combination selected in this embodiment is NsKNN+RF+GPR, the average error correlation of this combination is 0.360, and the maximum R² of the fusion model is 0.869, which ensures high complementarity between models.

[0096] Table 2 Performance of different combination strategies

[0097]

[0098] Step four: using the improved whale optimization algorithm to train the hyperparameters of the optimal basic model combination to obtain a multi-machine learning algorithm fusion prediction model.

[0099] In this embodiment, the key hyperparameters of the optimal basic model combination NsKNN+RF+GPR are the K value of NsKNN, the number of decision trees of RF, and the kernel function parameter of GPR. When the basic model combination is trained for the first time, the default hyperparameters or randomly selected hyperparameters are usually used, which may not be optimal; here, after obtaining the optimal basic model combination, the hyperparameters of the optimal basic model combination need to be trained again, so that each machine learning algorithm can play the best performance on the basis of the optimal cooperation effect.

[0100] In this embodiment, the performance of the multi-machine learning model fusion prediction model is also evaluated by using three independent reserved test sets (centrifugal pump N s = 37.5, mixed flow pump N s = 134.5, and axial flow pump N s = 209.3). The root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) are used as evaluation indexes, and the formulas are as follows:

[0101]

[0102]

[0103]

[0104] wherein, , and represent actual value, predicted value and average value, respectively. In addition, n represents the number of samples in the data set.

[0105] Table 3 shows the comparison of prediction errors of three integration methods on three test samples. The results show that the multi-machine learning model fusion prediction model of the embodiment achieves the optimal performance under all test conditions. Compared with the average method (Average), the Stacking model reduces the RMSE by 25.8% and the MAE by 22.8%; compared with the weighted method (Weight), the Stacking model reduces the RMSE by 23.1% and the MAE by 20.3%, showing a significant improvement in performance.

[0106] Table 3 Comparison of prediction errors of three integration methods on three test samples

[0107]

[0108] Figure 5 , Figure 6 , Figure 7 are the actual value and predicted value comparison charts of the pumps with specific speed of 37.5, 134.5 and 209.3 of the embodiment of the application, in each chart, the upper chart is the actual value and predicted value comparison chart of WH, and the lower chart is the actual value and predicted value comparison chart of WB, each chart gives the predicted value, true value and predicted value comparison of the fusion model of three single base models, in general, the fusion model shows the best prediction accuracy. The fusion model realizes the dual improvement of prediction accuracy and model robustness by learning the prediction reliability of each base model under different angle intervals and different pump types. The advantages of integrated learning prediction are as shown in Figure 6 In the 90°-130° area in the WH prediction, the predicted values of GPR and NsKNN are smaller and the predicted value of RF is larger, however, the proposed fusion model can well fuse the results of the base models to maintain the smallest estimation error. In the 60°-100° area of the WB curve, the predictions of the three base models have large errors, the predicted values of GPR and RF are uniformly larger and the predicted value of NsKNN is smaller, the fusion model effectively reduces the prediction error through the fusion strategy, and the fusion model can still accurately capture the overall trend.

[0109] In addition, in order to more clearly study the differences in prediction performance of different base models for each pump type, the bias chart is used to quantitatively compare the prediction performance of the fusion model and the base model, wherein the closeness of the data points to the diagonal line represents the matching degree of the model prediction and the actual result. Figure 8 andFigure 9 The bias distribution characteristics of different models in WH and WB prediction tasks on three representative specific speed pumps are respectively shown. The bias figure provides a more intuitive prediction accuracy evaluation. In different working conditions, the fusion model is better than the optimal single model, and the data points of the fusion model are more closely distributed around the diagonal line, with smaller variance and bias than the single model.

[0110] The verification result shows that the R² of the model on the three pump types is more than 0.949, compared with the optimal single base model, 4.2% R² improvement, 22.5% RMSE reduction and 27.8% MAE improvement are realized.

[0111] Corresponding to the foregoing embodiment of the water pump full characteristic curve prediction method based on multi-machine learning model fusion, the application also provides an embodiment of a water pump full characteristic curve prediction device based on multi-machine learning model fusion.

[0112] The water pump full characteristic curve prediction device based on multi-machine learning model fusion provided in the embodiment of the application comprises one or more processors for realizing the water pump full characteristic curve prediction method based on multi-machine learning model fusion in the foregoing embodiments.

[0113] The embodiment of the water pump full characteristic curve prediction device based on multi-machine learning model fusion can be applied to any device with data processing capability, which can be a device or apparatus such as a computer. The device embodiment can be realized by software, or by hardware or a combination of software and hardware. Taking software realization as an example, as a logical device, it is formed by reading the corresponding computer program instructions in the non-volatile memory into the memory and running by the processor of the device with data processing capability. From the hardware level, in addition to the processor, memory, network interface and non-volatile memory, the device with data processing capability in the embodiment usually includes other hardware according to the actual function of the device with data processing capability, and details are not repeated.

[0114] The implementation process of the functions and roles of each unit in the above device is specifically described in the implementation process of the corresponding steps in the above method, and is not repeated here.

[0115] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts are described in the part of the method embodiments. The device embodiments described above are only illustrative, wherein the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., can be located in one place or distributed on multiple network units. Some or all of the modules can be selected to achieve the purpose of the present application according to actual needs. Those skilled in the art can understand and implement it without creative labor.

[0116] The embodiment of the present application also provides a computer readable storage medium, which stores a program, and the program is executed by a processor to realize the method for predicting a water pump full characteristic curve based on multi-machine learning model fusion in the above embodiment.

[0117] The computer readable storage medium can be an internal storage unit of any device with data processing capability, such as a hard disk or a memory. The computer readable storage medium can also be an external storage device, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. Further, the computer readable storage medium can include both an internal storage unit of any device with data processing capability and an external storage device. The computer readable storage medium is used to store the computer program and other programs and data required by the device with data processing capability, and can also be used to temporarily store data that has been output or will be output.

[0118] Those skilled in the art can understand that the above description is only a preferred example of the application, and is not used to limit the application, although the application is described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for some technical features. Any modification, equivalent replacement, etc. within the spirit and principles of the application shall be included in the protection scope of the application.

Claims

1. A method for predicting the full characteristic curve of a water pump based on the fusion of multiple machine learning models, characterized in that: The steps include: Step 1: Construct a regression pool of multiple machine learning algorithms, including tree-based ensemble methods, linear models, kernel-based methods, neural network methods, probabilistic models, and instance-based learning methods; the machine learning algorithms are fed with the pump's specific speed and third-quadrant pump operating characteristic curve data as input, and output the dimensionless head and dimensionless torque of the pump over the full operating range; Step 2: Collect the full characteristic curve data of pumps with different specific speeds, convert them into dimensionless parameters, and divide them into training sets and test sets; Step 3: Use the improved whale optimization algorithm to screen out no less than two machine learning algorithms from the regression pool as the basic model combination, adopt the K-fold cross-validation training strategy, divide the training set into K subsets on average, select K-1 subsets for training in each round, and 1 subset for verification, perform K rounds of alternating training on each machine learning algorithm in the basic model combination to obtain the trained basic model combination, and save the K prediction results of each trained machine learning algorithm on the verification set and their corresponding true value labels to form a training data set for the secondary learner, and perform error correlation verification on the model; Then, the secondary learner is used to perform secondary training on the basic model combination verified by error correlation, and the R² performance of the basic model combination is calculated. The one with the largest R² is selected as the optimal basic model combination. The improvements of the improved whale optimization algorithm include: using the Logistic-Sine-Cosine chaotic mapping function to initialize the population and introducing an adaptive weight coefficient; Step 4: Use the improved whale optimization algorithm to train the hyperparameters of the optimal basic model combination to obtain a multi-machine learning algorithm fusion prediction model; Step 5: Input the specific speed of the pump to be predicted and the pump operating characteristic curve data in the third quadrant into the multi-machine learning algorithm fusion prediction model to obtain the dimensionless head and dimensionless torque of the pump within the full operating range.

2. The method for predicting the full characteristic curve of a water pump based on the fusion of multiple machine learning models according to claim 1, characterized in that: The tree-based ensemble methods include random forest RF, extreme gradient boosting XGBoost, LightGBM and decision tree DT, the linear models include linear regression LR, ridge regression Ridge and Bayesian ridge regression, the kernel-based methods include linear kernel SVR and radial basis function kernel SVR-RBF, the neural network methods include multi-layer perceptron MLP, the probability model includes Gaussian process regression GPR, and the instance-based learning methods include weighted specific speed-based K Nearest Neighbor NsKNN.

3. The method for predicting the full characteristic curve of a water pump based on the fusion of multiple machine learning models according to claim 1 is characterized in that: In step 3, the error correlation verification is specifically as follows: calculating the error correlation coefficient of any two models in the basic model combination; if and only if the error correlation coefficients of all model pairs in the basic model combination are not greater than the error correlation threshold, then the basic model combination passes the error correlation verification; Among them, the model i With model j The error correlation coefficient ρᵢ,ⱼ It is equal to the ratio of the covariance of the error series of the two models to the product of their respective standard deviations.

4. The method for predicting the full characteristic curve of a water pump based on the fusion of multiple machine learning models according to claim 1 is characterized in that: Ridge regression is chosen as the secondary learner.

5. The method for predicting the full characteristic curve of a water pump based on the fusion of multiple machine learning models according to claim 1 is characterized in that: The full characteristic curve data of pumps with different specific speeds collected in step 2 covers centrifugal pumps, mixed flow pumps and axial flow pumps.

6. The method for predicting the full characteristic curve of a water pump based on the fusion of multiple machine learning models according to claim 1, characterized in that: The adaptive weight coefficient ranges from 0.2 to 0.

9.

7. The method for predicting the full characteristic curve of a water pump based on the fusion of multiple machine learning models according to claim 1 is characterized in that: The K-fold cross validation is specifically a 5-fold cross validation.

8. The method for predicting the full characteristic curve of a water pump based on the fusion of multiple machine learning models according to claim 1 is characterized in that: In the step three, the improved whale optimization algorithm selects three machine learning algorithms from the regression pool.

9. A device for predicting the full characteristic curve of a water pump based on the fusion of multiple machine learning models, characterized in that: It includes one or more processors for implementing the water pump full characteristic curve prediction method based on the fusion of multiple machine learning models as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that A program is stored thereon, and when the program is executed by the processor, the method for predicting the full characteristic curve of a water pump based on the fusion of multiple machine learning models as described in any one of claims 1 to 8 is implemented.

Images