Method and device for constructing health status model of pumped storage unit and predicting performance trend
By establishing a health status model of pumped storage units based on Gaussian process regression and using the CNN-LSTM model for trend prediction, the problem of failure to fully consider the hydraulic-mechanical-electromagnetic coupling factor in the prior art is solved, and the accuracy of health performance trend prediction is improved.
Patent Information
- Application Number
- CN202210156063.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-02-21
- Publication Date
- 2025-04-15
- Estimated Expiration
- 2042-02-21
AI Technical Summary
The prior art fails to fully consider that the operating conditions of pumped storage units are affected by hydraulic-mechanical-electromagnetic coupling factors, resulting in poor health management and fault prediction results.
By obtaining the health status monitoring data of the pumped storage unit under different operating conditions, the key operation monitoring data are determined as the working head, active power and guide vane opening, and combining the unit vibration data, a standard health status model based on Gaussian process regression is established. Then, online monitoring data and health status models are used to extract health performance indicators and trend prediction is performed through the convolution-long short-term memory neural network (CNN-LSTM) model.
The accuracy of predicting the health performance trend of pumped storage units is improved, and the evolutionary law of unit health performance can more accurately reflect the engineering needs of pumped storage units health management.
Smart Images
Figure CN114880917B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of health management of pumped storage units, and in particular to a method and device for constructing a health status model of a pumped storage unit and predicting a performance trend. Background Art
[0002] Pumped storage units have been operating in harsh environments for a long time with complex operating conditions and frequent start-stop changes, which gradually cause fatigue damage and structural degradation of unit components. As the degree of fatigue degradation accumulates, the unit will produce hydraulic imbalance and reduced efficiency, which may cause the unit to fail to operate normally or even cause catastrophic accidents in severe cases. Therefore, in order to ensure the safe and stable operation of the unit under different operating conditions, it is urgent to carry out research on the prediction of unit health performance trends.
[0003] Existing research on health management and fault prediction of pumped storage units focuses on noise reduction filtering and intelligent fault pattern recognition of monitoring signals, but fails to fully consider the impact of hydraulic-mechanical-electromagnetic coupling factors on the unit's operating conditions.
[0004] Therefore, the multi-dimensional condition monitoring data of the unit should be fully considered to effectively characterize the unit health level, and then extract the index factors that can effectively quantify the unit health performance level, build a high-precision prediction model, and meet the engineering needs of pumped storage unit health management. Since the health management and fault prediction of pumped storage units have not received enough attention in past research, the prediction of the health performance trend of pumped storage units is still in the primary research stage. In addition, deep neural network models, especially the combination of CNN and LSTM models, are rarely used in the research of pumped storage unit health performance trend prediction. Summary of the invention
[0005] In view of this, the present invention provides a method and device for constructing a health status model of a pumped storage unit and predicting performance trends, which can solve the technical problem that the prior art fails to fully consider the influence of hydraulic-mechanical-electromagnetic coupling factors on the unit's operating conditions.
[0006] In order to solve the above technical problems, the present invention is implemented as follows.
[0007] A method for constructing a health status model of a pumped storage unit, comprising:
[0008] Step S1: acquiring health status monitoring data of the pumped-storage unit under different working conditions, and determining the key operation monitoring data affecting the equipment status of the pumped-storage unit as working head, active power and guide vane opening based on the status monitoring data; acquiring unit vibration data in the status monitoring data, where unit vibration is a sensitive characteristic quantity that can reflect the operating status of the pumped-storage unit;
[0009] Step S2: Establishing a standard health state model of the pumped-storage unit based on Gaussian process regression based on the key operation monitoring data and unit vibration data of the pumped-storage unit under different working conditions; the standard health state model of the pumped-storage unit is:
[0010] F(t)=f(H(t),P(t),G(t))
[0011] The standard health state model of the pumped storage unit is used to comprehensively characterize the operating characteristics of the unit under different operating conditions.
[0012] Preferably, in the stage of constructing the standard health status model, the correlation coefficient R is selected as an evaluation index of the fitting effect of the constructed standard health status model:
[0013]
[0014] In the formula, c(t) is the model output value, r(t) is the actual measurement value, N is the number of data, t is the time, and are the means of the output value and the measured value respectively. The correlation coefficient R is used to evaluate the fitting effect of the health status model.
[0015] The present invention provides a method for predicting the health performance trend of a pumped storage unit, which is implemented based on the health status model of the pumped storage unit constructed by the method for constructing the health status model of the pumped storage unit as described above, and is characterized in that the health performance trend prediction method comprises the following steps:
[0016] Step S3: Obtain the online monitoring data of the pumped storage unit, set multiple sampling time points, and obtain the online monitoring value V of the pumped storage unit corresponding to each sampling time point. num ;
[0017] Extract the working water head H corresponding to each sampling time point from the online monitoring data num , Active power P num , guide vane opening G num , 1≤num≤tnum, tnum is the number of sampling time points, num is the serial number of the sampling time point; the working water head H corresponding to each sampling time point num , Active power P num , guide vane opening G num Input the health status model of the pumped storage unit to obtain the equipment operation standard value F of the pumped storage unit corresponding to each sampling time point num ;
[0018] Calculate the health performance index value HPI corresponding to each sampling time point num , the health performance index value HPInum The calculation formula is:
[0019]
[0020] Among them, ω1 and ω2 are weights, and their values are 0.8 and 0.2 respectively according to experience. t is the actual normal operation time of the pumped storage unit, T is the maintenance cycle of the pumped storage unit, and k is the fault index of the pumped storage unit, and its value is 2. num is the standard value of the pumped storage unit equipment operation, F 上限 is the upper limit of the pumped storage unit equipment operation value, F 下限 is the lower limit of the pumped storage unit equipment operation value, C is the adjustment coefficient, and its value is:
[0021]
[0022] Each health performance index value HPI num Arrange in chronological order to obtain a health performance index sequence HPI of the pumped storage unit;
[0023] Step S4: normalize the health performance index sequence of the pumped storage unit to obtain a normalized index sequence L num ;
[0024] Step S5: Select the index sequence L num The first m data are used as training data, an input matrix and an output matrix are constructed based on the training data, and a convolutional-long short-term memory neural network prediction (CNN-LSTM) model is trained using the input matrix and the output matrix to obtain a trained CNN-LSTM model;
[0025] Step S6: Based on the indicator sequence L num The input matrix in2 is constructed by using the remaining data after removing the first m data, and the input matrix in2 is input into the trained CNN-LSTM model to obtain the health performance trend prediction result of the pumped-storage unit equipment.
[0026] Preferably, in the convolution-long short-term memory neural network prediction model, CNN is used to extract potential local features in the HPI sequence, and LSTM is used to obtain the mutual dependence between input data; the input data of the CNN-LSTM model is the HPI sequence. The convolution-long short-term memory neural network prediction model extracts local features of the input data through a two-layer one-dimensional convolutional network, then uses LSTM to store key information in the data set, and finally compresses the data dimension through a fully connected network to output the predicted value of the HPI sequence.
[0027] Preferably, the loss function of the CNN-LSTM model is:
[0028]
[0029] In the formula, is the predicted value output by the CNN-LSTM network, y k represents the true value, and K is the number of training data.
[0030] A pumped storage unit health status model building device, comprising:
[0031] The first data acquisition module is configured to acquire health status monitoring data of the pumped-storage unit under different working conditions, and determine the key operation monitoring data affecting the equipment status of the pumped-storage unit as working head, active power and guide vane opening based on the status monitoring data; acquire unit vibration data in the status monitoring data, where unit vibration is a sensitive characteristic quantity that can reflect the operating status of the pumped-storage unit;
[0032] The first model building module is configured to establish a standard health state model of the pumped-storage unit based on Gaussian process regression based on the key operation monitoring data and unit vibration data of the pumped-storage unit under different working conditions; the standard health state model of the pumped-storage unit is:
[0033] F(t)=f(H(t),P(t),G(t))
[0034] The standard health state model of the pumped storage unit is used to comprehensively characterize the operating characteristics of the unit under different operating conditions.
[0035] Preferably, in the stage of constructing the standard health status model, the correlation coefficient R is selected as an evaluation index of the fitting of the constructed standard health status model:
[0036]
[0037] In the formula, c(t) is the model output value, r(t) is the actual measurement value, N is the number of data, t is the time, and are the means of the output value and the measured value respectively. The correlation coefficient R is used to evaluate the fitting effect of the health status model.
[0038] The present invention provides a pumped storage unit health performance trend prediction device, which is implemented based on the pumped storage unit health status model constructed by the pumped storage unit health status model construction device as described above, and the pumped storage unit health performance trend prediction device includes:
[0039] The second data acquisition module is configured to obtain the online monitoring data of the pumped storage unit, set multiple sampling time points, and obtain the online monitoring value V of the pumped storage unit corresponding to each sampling time point. num ;
[0040] Extract the working water head H corresponding to each sampling time point from the online monitoring data num , Active power P num , guide vane opening G num , 1≤num≤tnum, tnum is the number of sampling time points, num is the serial number of the sampling time point; the working water head H corresponding to each sampling time point num , Active power P num , guide vane opening G num Input the health status model of the pumped storage unit to obtain the equipment operation standard value F of the pumped storage unit corresponding to each sampling time point num ;
[0041] Calculate the health performance index value HPI corresponding to each sampling time point num , the health performance index value HPI num The calculation formula is:
[0042]
[0043] Among them, ω1 and ω2 are weights, and their values are 0.8 and 0.2 respectively according to experience. t is the actual normal operation time of the pumped storage unit, T is the maintenance cycle of the pumped storage unit, and k is the fault index of the pumped storage unit, and its value is 2. num is the standard value of the pumped storage unit equipment operation, F 上限 is the upper limit of the pumped storage unit equipment operation value, F 下限 is the lower limit of the pumped storage unit equipment operation value, C is the adjustment coefficient, and its value is
[0044] Each health performance index value HPI num Arrange in chronological order to obtain a health performance index sequence HPI of the pumped storage unit;
[0045] The indicator sequence generation module is configured to normalize the health performance indicator sequence of the pumped storage unit to obtain a normalized indicator sequence L num ;
[0046] Training module: configured to select the indicator sequence L numThe first m data are used as training data, an input matrix and an output matrix are constructed based on the training data, and a convolutional-long short-term memory neural network prediction (CNN-LSTM) model is trained using the input matrix and the output matrix to obtain a trained CNN-LSTM model;
[0047] Prediction module: configured based on the indicator sequence L num The input matrix in2 is constructed by using the remaining data after removing the first m data, and the input matrix in2 is input into the trained CNN-LSTM model to obtain the health performance trend prediction result of the pumped-storage unit equipment.
[0048] Preferably, in the convolution-long short-term memory neural network prediction model, CNN is used to extract potential local features in the HPI sequence, and LSTM is used to obtain the mutual dependence between input data; the input data of the CNN-LSTM model is the HPI sequence. The convolution-long short-term memory neural network prediction model extracts local features of the input data through a two-layer one-dimensional convolutional network, then uses LSTM to store key information in the data set, and finally compresses the data dimension through a fully connected network to output the predicted value of the HPI sequence.
[0049] Preferably, the loss function of the CNN-LSTM model is:
[0050]
[0051] In the formula, is the predicted value output by the CNN-LSTM network, y k represents the true value, and K is the number of training data.
[0052] Beneficial effects:
[0053] (1) The method for constructing a health status model and predicting performance trends of a pumped-storage unit of the present invention obtains health performance indicators of the pumped-storage unit, constructs a trend prediction model based on a deep neural network, and improves the accuracy of health performance trend prediction of the pumped-storage unit under complex hydraulic, mechanical and electromagnetic interference.
[0054] (2) The method for constructing a health status model and predicting performance trends of a pumped storage unit of the present invention considers and analyzes the unit operation monitoring data in the health status model to determine the influence of the working head, active power and guide vane opening on its operating status; constructs an indicator that can accurately quantify the health performance; and establishes a trend prediction model based on CNN-LSTM by integrating the advantages of CNN in local feature extraction and LSTM sequence representation. This can timely predict the future health performance development trend of the unit and effectively reveal the evolution law of the unit's health performance.
[0055] (3) Furthermore, the method for constructing a health status model and predicting performance trends of a pumped storage unit of the present invention takes into account the time-varying and random characteristics of the unit's health performance trends affected by its operating conditions, comprehensively analyzes the mapping relationship between the working head, active power and guide vane opening and the unit's vibration characteristics, and uses historical monitoring data of the unit under different operating conditions to construct a standard health status model based on GPR, and uses online monitoring data to extract the HPI sequence that can quantify the unit's performance level.
[0056] (4) The method for predicting the health performance trend of a pumped storage unit of the present invention combines the advantages of CNN in local feature extraction and the good expressiveness of LSTM in time series, and proposes a unit health performance trend prediction model based on CNN-LSTM, which effectively improves the accuracy of the model's single-step prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0057] Figure 1 A schematic diagram of a process for constructing a health status model of a pumped storage unit provided in an embodiment of the present invention;
[0058] Figure 2 A schematic diagram of a pumped storage unit health performance trend prediction process according to an embodiment of the present invention;
[0059] Figure 3 This is a structural diagram of the CNN-LSTM prediction model in an embodiment of the present invention;
[0060] Figure 4 A specific flow chart of a health performance trend prediction model for a pumped storage unit based on health performance indicators and CNN-LSM provided in an embodiment of the present invention;
[0061] Figure 5 It is a working head diagram of monitoring data of the pumped storage unit in an embodiment of the present invention;
[0062] Figure 6 It is an active power diagram of monitoring data of the pumped storage unit in an embodiment of the present invention;
[0063] Figure 7 The guide vane opening diagram of the monitoring data of the pumped storage unit in the embodiment of the present invention;
[0064] Figure 8 It is a vibration diagram of monitoring data of a pumped storage unit in an embodiment of the present invention;
[0065] Fig. 9 A diagram showing the relationship between the working water head and the vibration of the pumped storage unit in the monitoring data in the embodiment of the present invention;
[0066] Fig.10 A diagram showing the relationship between active power and unit vibration in monitoring data of a pumped storage unit in an embodiment of the present invention;
[0067] Fig.11 A diagram showing the relationship between the guide vane opening and the vibration of the pumped storage unit in the monitoring data in the embodiment of the present invention;
[0068] Fig.12 This is a fitting result diagram of a standard health state model of a pumped storage unit constructed based on Gaussian process regression in an embodiment of the present invention;
[0069] Fig.13 It is a sequence diagram of health performance indicators of pumped storage units extracted in an embodiment of the present invention;
[0070] Fig.14 This is a trend prediction diagram of unit health performance in an embodiment of the present invention. DETAILED DESCRIPTION
[0071] At present, the research on the trend prediction of the health performance of pumped storage units is still in its preliminary stage. This paper summarizes the research experience of related rotating machinery and proposes that the trend prediction of the health performance of pumped storage units can be divided into three steps: 1) constructing a standard health status model; 2) extracting health performance indicators; 3) establishing a trend prediction model. In the first stage, the multidimensional monitoring data associated with the health status of the pumped storage unit should be fully considered and selected to establish a standard health status model. In addition, as the basis for subsequent trend prediction, indicators that can effectively characterize the health performance of the unit should be constructed in a simple way. Trend prediction aims to use historical data to build a prediction model and predict the future health performance evolution trend based on current operating data. The core is to build a trend prediction model. Due to its good expressiveness and nonlinear transformation capabilities, the convolutional neural network (CNN) extracts local features of the sequence through a convolution filter, and the long short-term memory network (LSTM) can learn and store long-term information, which is suitable for processing data with longer intervals in time series.
[0072] The present invention is described in detail below in conjunction with the accompanying drawings and embodiments.
[0073] like Figure 1 As shown, the present invention proposes a method for constructing a health status model of a pumped storage unit, the method comprising the following steps:
[0074] Step S1: acquiring health status monitoring data of the pumped-storage unit under different working conditions, and determining the key operation monitoring data affecting the equipment status of the pumped-storage unit as working head, active power and guide vane opening based on the status monitoring data; acquiring unit vibration data in the status monitoring data, where unit vibration is a sensitive characteristic quantity that can reflect the operating status of the pumped-storage unit;
[0075] Step S2: Establishing a standard health state model of the pumped-storage unit based on Gaussian process regression based on the key operation monitoring data and unit vibration data of the pumped-storage unit under different working conditions; the standard health state model of the pumped-storage unit is:
[0076] F(t)=f(H(t),P(t),G(t))
[0077] Among them, H(t) is the monitoring data of the unit working head, P(t) is the monitoring data of the unit active power, G(t) is the monitoring data of the unit guide vane opening, f is the Gaussian process regression function, F(t) is the unit vibration data; t is the unit status monitoring time; the standard health status model of the pumped storage unit is used to comprehensively characterize the operating characteristics of the unit under different working conditions.
[0078] Furthermore, in this embodiment, the working water head (H), active power (P), and guide vane opening (G) are selected as the input of the GPR model, and the unit vibration (F) is selected as the output of the GPR model, and a standard health status model suitable for different operating conditions of the unit is constructed to obtain the mapping relationship between the unit operating parameter data (H, P, G) and the status monitoring data.
[0079] Furthermore, in the stage of constructing the standard health status model, the correlation coefficient R is selected as an evaluation index of the fitting effect of the constructed standard health status model:
[0080]
[0081] In the formula, c(t) is the model output value, r(t) is the actual measurement value, N is the number of data, t is the time, and are the means of the output value and the measured value respectively. The correlation coefficient R is used to evaluate the fitting effect of the health status model.
[0082] like Figure 2 As shown, the present invention proposes a method for predicting the health performance trend of a pumped storage unit. The method is implemented based on the health status model of the pumped storage unit constructed by the method for constructing the health status model of the pumped storage unit as described above. The health performance trend prediction method comprises the following steps:
[0083] Step S3: Obtain the online monitoring data of the pumped storage unit, set multiple sampling time points, and obtain the online monitoring value V of the pumped storage unit corresponding to each sampling time point. num ;
[0084] Extract the working water head H corresponding to each sampling time point from the online monitoring data num , Active power P num , guide vane opening Gnum , 1≤num≤tnum, tnum is the number of sampling time points, num is the serial number of the sampling time point; the working water head H corresponding to each sampling time point num , Active power P num , guide vane opening G num Input the health status model of the pumped storage unit to obtain the equipment operation standard value F of the pumped storage unit corresponding to each sampling time point num ;
[0085] Calculate the health performance index value HPI corresponding to each sampling time point num , the health performance index value HPI num The calculation formula is:
[0086]
[0087] Among them, ω1 and ω2 are weights, and their values are 0.8 and 0.2 respectively according to experience. t is the actual normal operation time of the pumped storage unit, T is the maintenance cycle of the pumped storage unit, and k is the fault index of the pumped storage unit, and its value is 2. num is the standard value of the pumped storage unit equipment operation, F 上限 is the upper limit of the pumped storage unit equipment operation value, F 下限 is the lower limit of the pumped storage unit equipment operation value, C is the adjustment coefficient, and its value is:
[0088]
[0089] Each health performance index value HPI num Arrange in chronological order to obtain a health performance indicator sequence of the pumped storage unit;
[0090] In this embodiment, after several years of operation, the components of the pumped storage unit will experience fatigue degradation to varying degrees, and the health performance level will also decrease, which will cause the nonlinear mapping relationship between the unit vibration monitoring data V(t) and the working water head, active power and guide vane opening to deviate. Therefore, in order to effectively quantify the health performance level of the unit equipment under different operating conditions, the health performance index is constructed using the health status model and online monitoring data. The unit operating parameter data (H, P, G) is substituted into the health status model, and the theoretical standard value F(t) of the unit equipment status parameter under the current working condition is calculated. The health performance index (Health Performance Index) is defined according to the following formula, where t is the operating time.
[0091] Step S4: normalize the health performance index sequence of the pumped storage unit to obtain a normalized index sequence L num ;
[0092] Step S5: Select the index sequence L num The first m data are used as training data, an input matrix and an output matrix are constructed based on the training data, and a convolutional-long short-term memory neural network prediction (CNN-LSTM) model is trained using the input matrix and the output matrix to obtain a trained CNN-LSTM model;
[0093] Step S6: Based on the indicator sequence L num The input matrix in2 is constructed by using the remaining data after removing the first m data, and the input matrix in2 is input into the trained CNN-LSTM model to obtain the health performance trend prediction result of the pumped-storage unit equipment.
[0094] Furthermore, in step S5, the input matrix is The output matrix is Among them, X i is the index sequence L num The i-th HPI index value, X i ′ is the predicted value of the convolution-long short-term memory neural network prediction model. In the convolution-long short-term memory neural network prediction model, CNN is used to extract potential local features in the HPI sequence, and LSTM is used to obtain the mutual dependence between input data. The input data of the CNN-LSTM model is the HPI sequence. The structural diagram of the convolution-long short-term memory neural network prediction model is as follows Figure 3 As shown in the figure, the local features of the input data are extracted through a two-layer one-dimensional convolutional network, and then the key information in the data set is stored using LSTM. Finally, the data dimension is compressed through a fully connected network to output the predicted value of the HPI sequence.
[0095] Since the extracted health performance index (HPI) sequence has strong nonlinearity and non-stationarity, it is difficult for traditional prediction methods to make accurate predictions on its trend. In order to avoid the loss of key information in the HPI sequence, a prediction model based on convolution-long short-term memory neural network is proposed by utilizing the ability of CNN to extract local features in time series and the good expressiveness of LSTM in time series. This model has good generalization ability when dealing with high-dimensional nonlinear problems and can accurately predict the future health performance trend of unit equipment.
[0096] The loss function of the CNN-LSTM model is:
[0097] The loss function is used to quantify the deviation between the true value and the output value during model training, and to measure the quality of the model's prediction effect. The smaller the loss function value, the better the model performance during training. The mean square error (MSE) is used as the loss function of the CNN-LSTM network:
[0098]
[0099] In the formula, is the predicted value output by the CNN-LSTM network, y k represents the true value, and K is the number of training data.
[0100] During the model training process of the CNN-LSTM network, due to the deviation between the predicted value and the true value, in order to ensure that the proposed CNN-LSTM model has good accuracy and generalization, the Adam optimization algorithm is used to update the network parameters to minimize the model fitting error. At the same time, the gradient of the loss function is used to update the bias and weight of the network.
[0101] In the CNN-LSTM network, the network learning rate was set to 0.01, the maximum number of iterations was 200, and the dropout value in the fully connected layer was set to 0.2 by trial and error.
[0102] In this embodiment, the health performance index construction based on Gaussian process regression and the trend prediction method of the CNN-LSTM model systematically integrates mathematical statistics and deep learning theory to establish a performance trend prediction system for complex mechanical equipment such as pumped storage units with nonlinear dynamic behavior. The specific implementation process is as follows: Figure 4 The proposed method has two main advantages: first, it takes into account the impact of operating monitoring data such as working head, active power and guide vane opening under different operating conditions on the equipment health performance; second, the constructed CNN-LSTM model combines the excellence of CNN in data local feature extraction and the advantages of LSTM in time series prediction, thus improving the nonlinear learning ability of the model.
[0103] The following examples further illustrate the scheme of the present invention.
[0104] Taking a unit of a pumped storage power station as the research object, the effectiveness of the present invention is verified by predicting the health performance trend of the lower frame of the unit. By obtaining the peak-to-peak value of the Y-axis vibration data of the lower frame and the working head, active power, and guide vane opening at the corresponding time, the Figures 5 to 8 As shown in the figure, the correlation between the monitoring data of the unit, such as working water head, active power and guide vane opening, and the Y-direction vibration of the lower frame is comprehensively analyzed, and the following are drawn using data visualization: Figures 9 to 11 The mapping relationship diagram is shown.
[0105] The sample data of the unit in healthy state is used to build a healthy state model based on Gaussian process regression to obtain the multidimensional mapping relationship between the unit operating parameters and the vibration of the lower frame F(t) = f(H(t), P(t), G(t)). The fitting results are shown in Fig.12As shown in Figure 2, the correlation of the fitting results is 0.97, indicating that the GPR-based unit health status model can characterize the operating characteristics of unit components at a higher confidence level.
[0106] The monitoring data of the unit under stable operation under different working conditions are selected to extract the health performance index of the lower frame. Given the actual monitored vibration data V(t), the unit operation monitoring data such as working head, active power and guide vane opening are used as the input of the health state model to calculate the vibration value F(t) of the lower frame under the standard health state. Based on the actual monitored vibration value and the theoretically calculated vibration value, the HPI sequence of the lower frame health performance is obtained:
[0107]
[0108] Among them, ω1 and ω2 are weights, and their values are 0.8 and 0.2 respectively according to experience. t is the actual normal operation time of the pumped storage unit, T is the maintenance cycle of the pumped storage unit, and k is the fault index of the pumped storage unit, and its value is 2. num is the standard value of the pumped storage unit equipment operation, F 上限 is the upper limit of the pumped storage unit equipment operation value, F 下限 is the lower limit of the pumped storage unit equipment operation value, C is the adjustment coefficient, and its value is:
[0109]
[0110] The obtained HPI sequence is displayed in units of days, such as Fig.13 The overall trend of the health performance of the lower rack is increasing, indicating that its health performance is gradually deteriorating over time.
[0111] According to the acquired HPI sequence, the proposed CNN-LSTM model is applied to the performance trend prediction of the next rack. The prediction results of the proposed CNN-LSTM model are as follows: Fig.14 As shown in the figure, except for some mutation points, the predicted values of the model are basically consistent with the actual values, which can reflect the fluctuation of the original HPI sequence to the greatest extent. In addition, the RMSE value and MAE value of the CNN-LSTM model are 2.27 and 1.97, respectively. The results show that the method of the present invention can accurately predict the health performance trend of the pumped storage unit and obtain relatively satisfactory results.
[0112] The present invention also provides a pumped storage unit health status model construction device, comprising:
[0113] The first data acquisition module is configured to acquire health status monitoring data of the pumped-storage unit under different working conditions, and determine the key operation monitoring data affecting the equipment status of the pumped-storage unit as working head, active power and guide vane opening based on the status monitoring data; acquire unit vibration data in the status monitoring data, where unit vibration is a sensitive characteristic quantity that can reflect the operating status of the pumped-storage unit;
[0114] The first model building module is configured to establish a standard health state model of the pumped-storage unit based on Gaussian process regression based on the key operation monitoring data and unit vibration data of the pumped-storage unit under different working conditions; the standard health state model of the pumped-storage unit is:
[0115] F(t)=f(H(t),P(t),G(t))
[0116] The standard health state model of the pumped storage unit is used to comprehensively characterize the operating characteristics of the unit under different operating conditions.
[0117] The present invention provides a pumped storage unit health performance trend prediction device, which is implemented based on the pumped storage unit health status model constructed by the pumped storage unit health status model construction device as described above, and the pumped storage unit health performance trend prediction device includes:
[0118] The second data acquisition module is configured to obtain the online monitoring data of the pumped storage unit, set multiple sampling time points, and obtain the online monitoring value V of the pumped storage unit corresponding to each sampling time point. num ;
[0119] Extract the working water head H corresponding to each sampling time point from the online monitoring data num , Active power P num , guide vane opening G num , 1≤num≤tnum, tnum is the number of sampling time points, num is the serial number of the sampling time point; the working water head H corresponding to each sampling time point num , Active power P num , guide vane opening G num Input the health status model of the pumped storage unit to obtain the equipment operation standard value F of the pumped storage unit corresponding to each sampling time point num ;
[0120] Calculate the health performance index value HPI corresponding to each sampling time point num , the health performance index value HPI num The calculation formula is:
[0121]
[0122] Each health performance index value HPI num Arrange in chronological order to obtain a health performance index sequence HPI of the pumped storage unit;
[0123] The indicator sequence generation module is configured to normalize the health performance indicator sequence of the pumped storage unit to obtain a normalized indicator sequence L num ;
[0124] Training module: configured to select the indicator sequence L num The first m data are used as training data, an input matrix and an output matrix are constructed based on the training data, and a convolutional-long short-term memory neural network prediction (CNN-LSTM) model is trained using the input matrix and the output matrix to obtain a trained CNN-LSTM model;
[0125] Prediction module: configured based on the indicator sequence L num The input matrix in2 is constructed by using the remaining data after removing the first m data, and the input matrix in2 is input into the trained CNN-LSTM model to obtain the health performance trend prediction result of the pumped-storage unit equipment.
[0126] The above specific embodiments only describe the design principle of the present invention. The shapes and names of the components in the description may be different and are not limited. Therefore, those skilled in the art in the field of the present invention may modify or replace the technical solutions recorded in the above embodiments; and these modifications and replacements do not deviate from the creative purpose and technical solutions of the present invention and should all fall within the protection scope of the present invention.
Claims
1. A method for predicting the health performance trend of a pumped storage unit, the method being implemented based on a pumped storage unit health status model constructed by a method for constructing a pumped storage unit health status model, characterized in that: The health performance trend prediction method comprises the following steps: Step S3: Obtain the online monitoring data of the pumped storage unit, set multiple sampling time points, and obtain the online monitoring value V of the pumped storage unit corresponding to each sampling time point. num ; Extract the working water head H corresponding to each sampling time point from the online monitoring data num , Active power P num , guide vane opening G num , 1≤num≤tnum, tnum is the number of sampling time points, num is the serial number of the sampling time point; the working water head H corresponding to each sampling time point num , Active power P num , guide vane opening G num Input the health status model of the pumped storage unit to obtain the equipment operation standard value F of the pumped storage unit corresponding to each sampling time point num ; Calculate the health performance index value HPI corresponding to each sampling time point num , the health performance index value HPI num The calculation formula is: Wherein, ω1 and ω2 are weights respectively, t is the actual normal operation time of the pumped storage unit, T is the maintenance cycle of the pumped storage unit, k is the fault index of the pumped storage unit, and F num is the standard value of the pumped storage unit equipment operation, F 上限 is the upper limit of the pumped storage unit equipment operation value, F 下限 is the lower limit of the pumped storage unit equipment operation value, C is the adjustment coefficient, and its value is: Each health performance index value HPI num Arrange in chronological order to obtain a health performance index sequence HPI of the pumped storage unit; Step S4: normalize the health performance index sequence of the pumped storage unit to obtain a normalized index sequence L num ; Step S5: Select the index sequence L num The first m data are used as training data, an input matrix and an output matrix are constructed based on the training data, and a convolutional-long short-term memory neural network prediction (CNN-LSTM) model is trained using the input matrix and the output matrix to obtain a trained CNN-LSTM model; Step S6: Based on the indicator sequence L num The remaining data after removing the first m data are used to construct an input matrix in2, and the input matrix in2 is input into the trained CNN-LSTM model to obtain a health performance trend prediction result of the pumped storage unit equipment; The method for constructing the health status model of a pumped storage unit includes: Step S1: acquiring health status monitoring data of the pumped-storage unit under different working conditions, and determining the key operation monitoring data affecting the equipment status of the pumped-storage unit as working head, active power and guide vane opening based on the status monitoring data; acquiring unit vibration data in the status monitoring data, where unit vibration is a sensitive characteristic quantity that can reflect the operating status of the pumped-storage unit; Step S2: Establishing a standard health state model of the pumped-storage unit based on Gaussian process regression based on the key operation monitoring data and unit vibration data of the pumped-storage unit under different working conditions; the standard health state model of the pumped-storage unit is: F(t)=f(H(t),P(t),G(t)) Among them, H(t) is the monitoring data of the unit working head, P(t) is the monitoring data of the unit active power, G(t) is the monitoring data of the unit guide vane opening, f is the Gaussian process regression function, F(t) is the unit vibration data; t is the unit status monitoring time; the standard health status model of the pumped storage unit is used to comprehensively characterize the operating characteristics of the unit under different working conditions.
2. The method according to claim 1, characterized in that In the convolution-long short-term memory neural network prediction model, CNN is used to extract potential local features in the HPI sequence, and LSTM is used to obtain the mutual dependence between input data; the input data of the CNN-LSTM model is the HPI sequence; the convolution-long short-term memory neural network prediction model extracts local features of the input data through a two-layer one-dimensional convolutional network, then uses LSTM to store key information in the data set, and finally compresses the data dimension through a fully connected network to output the predicted value of the HPI sequence.
3. The method according to claim 2, characterized in that The loss function of the CNN-LSTM model is: In the formula, is the predicted value output by the CNN-LSTM network, y k represents the true value, and K is the number of training data.
4. The method according to claim 1, characterized in that In the stage of constructing the standard health status model, the correlation coefficient R is selected as the evaluation index of the fitting effect of the constructed standard health status model: In the formula, c(t) is the model output value, r(t) is the actual measurement value, N is the number of data, t is the time, and are the means of the output value and the measured value respectively; the correlation coefficient R is used to evaluate the fitting effect of the health status model.
5. A pumped storage unit health performance trend prediction device, the device is implemented based on the pumped storage unit health status model constructed by the pumped storage unit health status model construction device, characterized in that: The pumped storage unit health performance trend prediction device comprises: The second data acquisition module is configured to obtain the online monitoring data of the pumped storage unit, set multiple sampling time points, and obtain the online monitoring value V of the pumped storage unit corresponding to each sampling time point. num ; Extract the working water head H corresponding to each sampling time point from the online monitoring data num , Active power P num , guide vane opening G num , 1≤num≤tnum, tnum is the number of sampling time points, num is the serial number of the sampling time point; the working water head H corresponding to each sampling time point num , Active power P num , guide vane opening G num Input the health performance model of the pumped storage unit to obtain the equipment operation standard value F of the pumped storage unit corresponding to each sampling time point num ; Calculate the health performance index value HPI corresponding to each sampling time point num , the health performance index value HPI num The calculation formula is: Wherein, ω1 and ω2 are weights respectively, t is the actual normal operation time of the pumped storage unit, T is the maintenance cycle of the pumped storage unit, k is the fault index of the pumped storage unit, and F num is the standard value of the pumped storage unit equipment operation, F 上限 is the upper limit of the pumped storage unit equipment operation value, F 下限 is the lower limit of the pumped storage unit equipment operation value, C is the adjustment coefficient, and its value is: Each health performance index value HPI num Arrange in chronological order to obtain a health performance index sequence HPI of the pumped storage unit; The indicator sequence generation module is configured to normalize the health performance indicator sequence of the pumped storage unit to obtain a normalized indicator sequence L num ; Training module: configured to select the indicator sequence L num The first m data are used as training data, an input matrix and an output matrix are constructed based on the training data, and a convolutional-long short-term memory neural network prediction (CNN-LSTM) model is trained using the input matrix and the output matrix to obtain a trained CNN-LSTM model; Prediction module: configured based on the indicator sequence L num The remaining data after removing the first m data are used to construct an input matrix in2, and the input matrix in2 is input into the trained CNN-LSTM model to obtain a health performance trend prediction result of the pumped storage unit equipment; The device for constructing the health status model of the pumped storage unit includes: The first data acquisition module is configured to acquire health status monitoring data of the pumped-storage unit under different working conditions, and determine the key operation monitoring data affecting the equipment status of the pumped-storage unit as working head, active power and guide vane opening based on the status monitoring data; acquire unit vibration data in the status monitoring data, where unit vibration is a sensitive characteristic quantity that can reflect the operating status of the pumped-storage unit; The first model building module is configured to establish a standard health state model of the pumped-storage unit based on Gaussian process regression based on the key operation monitoring data and unit vibration data of the pumped-storage unit under different working conditions; the standard health state model of the pumped-storage unit is: F(t)=f(H(t),P(t),G(t)) Among them, H(t) is the monitoring data of the unit working head, P(t) is the monitoring data of the unit active power, G(t) is the monitoring data of the unit guide vane opening, f is the Gaussian process regression function, F(t) is the unit vibration data; t is the unit status monitoring time; the standard health status model of the pumped storage unit is used to comprehensively characterize the operating characteristics of the unit under different working conditions.
6. The device according to claim 5, characterized in that In the convolution-long short-term memory neural network prediction model, CNN is used to extract potential local features in the HPI sequence, and LSTM is used to obtain the mutual dependence between input data; the input data of the CNN-LSTM model is the HPI sequence; the convolution-long short-term memory neural network prediction model extracts local features of the input data through a two-layer one-dimensional convolutional network, then uses LSTM to store key information in the data set, and finally compresses the data dimension through a fully connected network to output the predicted value of the HPI sequence.
7. The device according to claim 6, characterized in that The loss function of the CNN-LSTM model is: In the formula, is the predicted value output by the CNN-LSTM network, y k represents the true value, and K is the number of training data.
8. The device according to claim 5, characterized in that In the stage of constructing the standard health status model, the correlation coefficient R is selected as the evaluation index of the fitting effect of the constructed standard health status model: In the formula, c(t) is the model output value, r(t) is the actual measurement value, N is the number of data, t is the time, and are the means of the output value and the measured value respectively; the correlation coefficient R is used to evaluate the fitting effect of the health status model.