Evaporative cooling medium gas concentration prediction method based on large hydraulic turbine set

Abstract

According to the method for predicting the gas concentration of the evaporative cooling medium based on the large hydraulic turbine set, besides historical gas concentration data, the influence of other factors such as the active power of the set, the oxygen concentration, the stator winding temperature, the cooling water flow, the liquid level of the evaporative cooling medium and the pressure of an evaporative cooling system on gas concentration prediction is also considered. The method comprises the following steps: firstly, acquiring historical data of gas concentration of an evaporative cooling medium and corresponding unit active power, oxygen concentration, stator winding temperature, cooling water flow, liquid level of the evaporative cooling medium and pressure data of an evaporative cooling system, preprocessing the data, dividing the data into a training set and a test set, and establishing a CNN model; important feature information is extracted through CNN to further construct a time sequence to be input into an ELM model, finally, a prediction result of the gas concentration is output, error analysis is conducted, whether the predicted gas concentration is abnormal or not is judged, and if yes, an alarm is given.

Description

Technical Field

[0001] The invention belongs to the technical field of water turbine units, and in particular relates to a method for predicting the concentration of evaporative cooling medium gas in a large water turbine unit. Background Art

[0002] Large hydro-turbine generator sets are equipped with evaporative cooling systems to cool the stator windings of the generators. If the hydrofluorocarbon Freon refrigerant (HFC4310) medium gas of the evaporative cooling system leaks, the heat generated by the stator windings will increase suddenly, endangering the safety of the hydro-turbine generator set. At the same time, the medium gas will be deposited in the lower part of the generator pit and the waterwheel room, causing the medium content in this area to increase and the oxygen content to decrease. When the unit is running, the temperature inside the generator pit may be higher than the boiling point of the medium, which may cause dizziness, convulsions, suffocation, irregular heartbeat, loss of consciousness and other symptoms. Therefore, high-precision prediction of the medium gas concentration of the evaporative cooling system is the basis for ensuring the safe and reliable operation of large hydro-turbine generators.

[0003] At present, evaporative cooling turbine generators are equipped with medium detection devices at appropriate heights in the generator wind tunnel along the circumferential direction to detect the HFC4310 and oxygen content in the generator wind tunnel. When the medium concentration in the generator wind tunnel reaches 400ppm, the device sends an alarm signal. The gas concentration medium monitoring method for evaporative cooling turbine generators still uses traditional condition monitoring systems. At the same time, some data sources have large changes in data during special time periods. It is difficult to accurately capture these changes using traditional condition monitoring systems, ignoring the impact of unit active power, oxygen concentration, stator winding temperature, cooling water flow, evaporative cooling medium level, and evaporative cooling system pressure on the predicted value. Summary of the invention

[0004] In order to solve the above problems, the present invention aims to provide a method for predicting the gas concentration of evaporative cooling medium based on large-scale turbine units, which can provide early warning before the abnormal concentration of medium gas in the evaporative cooling system occurs, so that gas leakage can be discovered in the embryonic stage.

[0005] The technical solution of the present invention is as follows: A method for predicting the concentration of evaporative cooling medium gas in a large-scale hydro-turbine unit comprises the following steps: Step 1: Obtain historical data on evaporative cooling medium concentration and corresponding unit active power, oxygen concentration, stator winding temperature, cooling water flow, evaporative cooling medium level, and evaporative cooling system pressure data, and select data during stable operation of the unit to avoid the impact of excessive data fluctuations on the predicted value under shutdown and transition conditions.

[0006] Step 2: Data preprocessing. In order to reduce the impact of outliers, obtain stable prediction results, and minimize the impact of nonlinear transformation, the data is normalized and the sample amplitude is normalized to [0, 1], as shown below: ; in is the normalized value, is the characteristic value, is the minimum value, is the maximum value.

[0007] Step 3: Divide the data into training set and test set with a ratio of 9:1, divide the feature value and target value to obtain the training set; input the processed data into the CNN model.

[0008] Step 4: Construct a CNN model. The CNN model is a feedforward neural network, and its structure consists of a framework consisting of 2 convolutional layers, 2 pooling layers and a fully connected layer. CNN has the characteristics of a small number of free parameters, multi-dimensional network learning data features, and good robustness. By alternating convolutional layers and pooling layers, the purpose of reducing the dimension of feature data and improving the quality of feature data is achieved.

[0009] The output feature vector of the CNN is , which can be expressed as: ; ; ; ; ; In the formula, , They are the outputs of convolutional layer 1 and convolutional layer 2 respectively; , They are the outputs of pooling layer 1 and pooling layer 2 respectively; , are weight matrices respectively; , , , , For deviation; through Operation on the output features from the upper neural network Perform convolution operation; is the maximum function; the output length of the CNN layer is , ; The basic structure of the CNN model is as follows Figure 1As shown, the input includes gas concentration, unit active power, oxygen concentration, stator winding temperature, cooling water flow, evaporative cooling medium level, and evaporative cooling system pressure time series.

[0010] Step 5: Establish an ELM model. Use CNN to extract the multidimensional feature vector that affects the gas concentration and construct a time series as the input of the ELM model. The ELM network is a single hidden layer feedforward neural network, and its network model structure is shown in Figure 2.

[0011] Any given Different samples , For the The input variables of samples are , for Dimensional input data, For the corresponding The expected output is , for Dimensional output data, .make and are the weights of neurons connecting the input layer to the hidden layer and the hidden layer to the output layer, respectively. is the hidden layer threshold, is the activation function, then the output function is: ; remember is the network hidden layer output matrix, is the expected output matrix of a given sample, the above formula is simplified to: ; in, ; remember The Moore-Penrose generalized inverse of ,but The solution is: ; Step 6: Denormalize the prediction results of each subsequence to obtain the prediction results of gas concentration, and perform error analysis on the prediction results of gas concentration of the evaporative cooling system.

[0012] In order to accurately verify the prediction performance of the CNN-ELM model proposed in this paper, the mean absolute percentage error (MAPE) and root mean square error (RMSE) are selected as evaluation indicators. The specific formulas are as follows.

[0013] ; ; In the formula, n is the test set capacity; X act ( i )and X pred ( i ) respectively represent i The actual and predicted values ​​of gas concentration at each moment; MAPE and RMSE The smaller the value, the more accurate the prediction.

[0014] Step 7: Based on the final predicted result of the output gas concentration, determine whether the predicted gas concentration is abnormal. If so, an alarm is issued.

[0015] The beneficial effects of the present invention are: Historical gas concentration data plays an important role in the prediction process. In addition to historical gas concentration data, other factors such as unit active power, oxygen concentration, stator winding temperature, cooling water flow, evaporative cooling medium level, and evaporative cooling system pressure will also affect the accuracy of gas concentration prediction. Due to the large volatility and uncertainty of these factors, they are not suitable to be directly used as ELM input. CNN, with its special structure, can effectively mine the potential relationships among the multidimensional factors that affect the change of gas concentration, extract important feature information, and further construct it into a time series input to the ELM model to achieve accurate prediction of gas concentration. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is the CNN model structure; Figure 2 It is the ELM model structure; Figure 3 It is a flow chart of the method for predicting the concentration of evaporative cooling medium gas in large turbine units. DETAILED DESCRIPTION

[0017] In order to better understand the present invention, the present invention is further described in detail below in conjunction with diagrams and examples, but the scope of protection claimed in the present invention is not limited to the following examples.

[0018] A method for predicting the concentration of evaporative cooling medium gas in a large-scale hydro-turbine unit comprises the following steps: Step 1: Obtain historical data on evaporative cooling medium concentration and corresponding unit active power, oxygen concentration, stator winding temperature, cooling water flow, evaporative cooling medium level, and evaporative cooling system pressure data, and select data during stable operation of the unit to avoid the impact of excessive data fluctuations on the predicted value under shutdown and transition conditions.

[0019] Step 2: Data preprocessing. In order to reduce the impact of outliers, obtain stable prediction results, and minimize the impact of nonlinear transformation, the data is normalized and the sample amplitude is normalized to [0, 1], as shown below: ; in is the normalized value, is the characteristic value, is the minimum value, is the maximum value.

[0020] Step 3: Divide the data into training set and test set with a ratio of 9:1, divide the feature value and target value to obtain the training set; input the processed data into the CNN model.

[0021] Step 4: Construct a CNN model. The CNN model is a feedforward neural network, and its structure consists of a framework consisting of 2 convolutional layers, 2 pooling layers and a fully connected layer. CNN has the characteristics of a small number of free parameters, multi-dimensional network learning data features, and good robustness. By alternating convolutional layers and pooling layers, the purpose of reducing the dimension of feature data and improving the quality of feature data is achieved.

[0022] The output feature vector of the CNN is , which can be expressed as: ; ; ; ; ; In the formula, , They are the outputs of convolutional layer 1 and convolutional layer 2 respectively; , They are the outputs of pooling layer 1 and pooling layer 2 respectively; , are weight matrices respectively; , , , , For deviation; through Operations and output features from the upper neural network Perform convolution operation; is the maximum function; the output length of the CNN layer is , .

[0023] The basic structure of the CNN model is as follows Figure 1As shown, the input includes gas concentration, unit active power, oxygen concentration, stator winding temperature, cooling water flow, evaporative cooling medium level, and evaporative cooling system pressure time series.

[0024] Step 5: Establish an ELM model. Use CNN to extract the multidimensional feature vector that affects the gas concentration and construct a time series as the input of the ELM model. The ELM network is a single hidden layer feedforward neural network, and its network model structure is shown in Figure 2.

[0025] Any given Different samples , For the The input variables of samples are , for Dimensional input data, For the corresponding The expected output is , for Dimensional output data, .make and are the weights of neurons connecting the input layer to the hidden layer and the hidden layer to the output layer, respectively. is the hidden layer threshold, is the activation function, then the output function is: ; remember is the network hidden layer output matrix, is the expected output matrix of a given sample, the above formula is simplified to: ; in, ; remember The Moore-Penrose generalized inverse of ,but The solution is: ; Step 6: Denormalize the prediction results of each subsequence to obtain the prediction results of gas concentration, and perform error analysis on the prediction results of gas concentration of the evaporative cooling system.

[0026] In order to accurately verify the prediction performance of the CNN-ELM model proposed in this paper, the mean absolute percentage error (MAPE) and root mean square error (RMSE) are selected as evaluation indicators. The specific formulas are as follows.

[0027] ; ; In the formula, n is the test set capacity; X act ( i )and X pred ( i ) respectively represent i The actual and predicted values ​​of gas concentration at each moment; MAPE and RMSE The smaller the value, the more accurate the prediction.

[0028] Step 7: Based on the final predicted result of the output gas concentration, determine whether the predicted gas concentration is abnormal. If so, an alarm is issued. The CNN-ELM combined model is selected to predict the gas concentration of the evaporative cooling medium. The evaporative cooling medium is HFC4310 (2,3-dihydrodecafluoropentane, molecular formula C5H2F10), with a boiling point of 55°C, colorless and transparent, slight odor, volatile, gaseous density more than 8 times heavier than air (average molecular weight of air 29), and stable performance below 300°C. The CNN-ELM network, ELM network, LSTM network, and RF network prediction models are constructed using the training set sample data. The CNN-ELM network and ELM network are trained 100 times, the learning rate in the LSTM model is set to 0.001, the number of decision trees in the RF network model is 500, and the number of split features is 4.

[0029] The evaluation indicators of the prediction results of different models are shown in Table 1. Table 1 Comparison of different prediction models

[0030] Further quantitative analysis of the effects of the four prediction models shows that the CNN-ELM prediction model proposed in this paper has the smallest error, that is, the model prediction result is the best, based on the calculation results of the mean absolute percentage error (MAPE) and the root mean square error (RMSE).

[0031] The technical solution of the present invention is explained through the above embodiments, but the present invention is not limited to the above embodiments, that is, it does not mean that the present invention must rely on the above specific embodiments to be implemented. Any improvements made by technicians in the relevant field based on the present invention, or equivalent replacement of the materials selected by the present invention, etc., fall within the scope of protection of the patent.

Claims

1. A method for predicting the concentration of evaporative cooling medium gas in a large-scale hydro turbine unit, characterized in that: The method comprises the following steps: Step 1: Obtain historical data on evaporative cooling medium concentration and corresponding unit active power, oxygen concentration, stator winding temperature, cooling water flow, evaporative cooling medium liquid level, and evaporative cooling system pressure data, and select data during stable operation of the unit; Step 2: Data preprocessing: normalize the data to obtain preprocessed data; Step 3: Divide the data into training set and test set, divide the feature value and target value, and obtain the training set; Step 4: Input the processed data into the CNN model to build the CNN model; Step 5: Establish ELM model: Use CNN model to extract multidimensional feature vectors that affect gas concentration and construct them into time series as input of ELM model; Step 6: Denormalize the prediction results of each subsequence in step 5 to obtain the prediction results of gas concentration, and perform error analysis on the prediction results of gas concentration of the evaporative cooling system; Step 7: Based on the final predicted result of the output gas concentration, determine whether the predicted gas concentration is abnormal. If so, an alarm is issued.

2. The method for predicting the concentration of evaporative cooling medium gas based on large-scale hydro turbine units according to claim 1 is characterized in that: Step 2: Normalize the data to normalize the sample amplitude to [0, 1], as shown in the following formula: ； in is the normalized value, is the characteristic value, is the minimum value, is the maximum value.

3. The method for predicting the concentration of evaporative cooling medium gas based on large-scale hydro turbine units according to claim 1 is characterized in that: In step 4, the output feature vector of CNN , expressed as: ； ； ； ； ； In the formula, , They are the outputs of convolutional layer 1 and convolutional layer 2 respectively; , They are the outputs of pooling layer 1 and pooling layer 2 respectively; , are weight matrices respectively; , , , , For deviation; through Operation on the output features from the upper neural network Perform convolution operation; is the maximum function; the output length of the CNN layer is , .

4. The method for predicting the concentration of evaporative cooling medium gas based on large-scale hydro turbine units according to claim 1 is characterized in that: The CNN input in step 4 includes gas concentration, unit active power, oxygen concentration, stator winding temperature, cooling water flow, evaporative cooling medium level, and evaporative cooling system pressure time series.

5. The method for predicting the concentration of evaporative cooling medium gas based on large-scale hydro turbine units according to claim 1 is characterized in that: In step 5, the specific process of establishing the ELM model is as follows: Given Different samples , For the The input variables of samples are , for Dimensional input data, For the corresponding The expected output is , for Dimensional output data, ;make and are the weights of neurons connecting the input layer to the hidden layer and the hidden layer to the output layer, respectively. is the hidden layer threshold, is the activation function, then the output function is: ； remember is the network hidden layer output matrix, is the expected output matrix of a given sample, the above formula is simplified to: ； in, ； remember The Moore-Penrose generalized inverse of ,but The solution is: 。 6. The method for predicting the concentration of evaporative cooling medium gas based on large-scale hydro turbine units according to claim 1 is characterized in that: In step 6, the mean absolute percentage error (MAPE) and the root mean square error (RMSE) are selected as evaluation indicators. The specific formulas are as follows: ； ； In the formula, n is the test set capacity; X act ( i )and X pred ( i ) respectively represent i The actual and predicted values ​​of gas concentration at each moment; MAPE and RMSE The smaller the value, the more accurate the prediction.

Images