Cascade fluid state prediction method based on LSTM neural network
By adopting a cascaded fluid state prediction method based on LSTM neural network, the technical gap in fluid state prediction in uranium enrichment plants has been filled, enabling early warning of system disturbance trends and ensuring the safe and stable operation of uranium enrichment plants.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-30
- Publication Date
- 2026-04-03
AI Technical Summary
The lack of effective methods for predicting the cascade fluid states in uranium enrichment plants leads to operational errors and affects system stability. Existing technologies rely on human experience and are difficult to accurately judge fluid state trends.
A cascaded fluid state prediction method based on LSTM neural network is adopted. By collecting, parsing and preprocessing pressure measurement data of the cascaded system of uranium enrichment plant, an LSTM neural network model is constructed. The gating mechanism of input gate, forget gate and output gate is used to process long sequence information. The model is trained by combining MAE loss function and Adam optimizer to achieve rapid prediction and early warning of fluid state.
It enables efficient and accurate prediction of the fluid state of the uranium enrichment plant cascade system, reduces the possibility of human error, ensures the safe and stable operation of the system, and lowers the probability of system interlock protection.
Smart Images
Figure CN121787028A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of uranium enrichment control engineering application technology, and in particular to a cascaded fluid state prediction method based on LSTM neural network. Background Technology
[0002] The stability and safety of a uranium enrichment plant are related to many factors, including design, construction, operation, and maintenance. After commissioning, its operational status largely depends on the skill level of the operators. Actual engineering practice shows that human error is the primary cause of anomalies and accidents in uranium enrichment projects. Misjudgments and operational errors by operators can lead to system malfunctions. To reduce the possibility of misoperation, it is necessary to predict system status, assist operators in providing early warnings of potential anomalies, reduce the probability of system cascading protection triggered by "limits," and ensure the long-term safe and stable operation of the system.
[0003] Fluid state is a crucial factor affecting cascade separation efficiency and operational safety. Changes in external temperature, electrical frequency fluctuations, valve status, and even machine malfunctions during production can all alter the fluid state, propagate within the system, and impact system stability. However, currently, no research institutions have conducted predictive studies on the fluid state in centrifugal cascades, and assessments of state trends primarily rely on the experience of operators. Due to the unique characteristics of uranium enrichment plants, existing technologies from other industries cannot be directly applied; targeted research is essential to address key technological challenges.
[0004] Long Short-Term Memory (LSTM) networks, a variant of Recurrent Neural Networks (RNNs), are commonly used techniques for processing time-series data and predicting input signals. Unlike traditional feedforward neural networks, LSTMs possess memory units and gating mechanisms, which help capture and process long-term dependencies in time series data. By processing time-series data as input in sequence according to time steps, LSTMs can learn patterns and trends between data points, using past time-step data to predict future values or states. Based on this, this invention proposes a cascaded fluid state prediction method based on LSTM neural networks to address the technological gap in cascaded fluid state prediction for uranium enrichment plants. Summary of the Invention
[0005] In view of the fact that there is currently no method for predicting the state of fluids in centrifugal cascade systems in the uranium enrichment industry, and the judgment of fluid state trends mainly relies on the experience of operators, which can easily lead to operational errors, system instability, and substandard product enrichment, this invention provides a cascade fluid state prediction method based on LSTM neural network. This method can quickly predict the fluid state in the early stages of system state changes, realize early warning of system disturbance trends, assist operators in judging possible abnormal states in advance, and ensure the long-term safe and stable operation of the system.
[0006] To achieve the above objectives, the technical solution of the present invention is implemented as follows: A cascaded fluid state prediction method based on LSTM neural network is used for predicting the fluid pressure state of a cascaded system in a uranium enrichment plant, enabling early warning of system disturbance trends. The method includes the following steps: S1. Collect real-time data of pressure measuring points in the uranium enrichment plant cascade system. The uranium enrichment plant cascade system consists of N separation stages. Each stage is equipped with a feed main pipe, a concentrate main pipe, and a lean feed main pipe. The pressure measuring points correspond to the pressure monitoring of the feed / concentrate / lean feed main pipes of each separation stage. S2: Analyze and store the collected pressure measurement data; S3: Preprocess the raw pressure data. The preprocessing includes data format conversion, time dimension standardization, array type conversion and normalization, mapping the data to a preset numerical range. S4: Encode the preprocessed time series data, convert it into supervised learning samples, and generate a dataset and a label set. The encoding includes determining the duration of historical data, the duration of predicted data, and segmenting and adjusting the dimensions of the data. S5: Divide the dataset into training and testing sets; S6: Build a state prediction model based on LSTM neural network. The model includes an LSTM computation layer and a fully connected output layer. The LSTM computation layer processes long sequence information through gating mechanisms of input gate, forget gate and output gate, and adopts a regression loss function and an adaptive learning rate optimizer. S7: Input the training set into the LSTM neural network model for training, and minimize the loss function by iteratively updating the model weights; S8: Test on the test set to verify the model's prediction accuracy and show the comparison between the prediction results and the true values; S9: Input the historical pressure data of the measurement point to be predicted into the trained model, predict the fluid pressure state within a preset time period in the future, and store the prediction results.
[0007] Furthermore, in step S1, the number of pressure measuring points is 3N. The data from the 3N pressure measuring points are coupled together and real-time data is obtained through the distributed control system interface. Each measuring point records the pressure value of the feed / refined / lean material main pipe of the corresponding separation stage.
[0008] Furthermore, in step S2, the pressure measurement data is transmitted in a specific format or protocol, and after parsing, it is stored in a format that can support subsequent data processing.
[0009] Furthermore, step S3 specifically includes the following steps: S31: Organize the stored data into an Excel-readable XLS file. The XLS file contains multiple worksheets, and each worksheet records the pressure value data of the measuring points when pressure disturbances occur in different units. S32: Use Python to delete the first column (serial number) of each sub-table in the XLS file, concatenate the columns to form a master table, and each row in the master table represents the pressure value of 3N measuring points at the current time, expressed as follows: , The file represents the lean / feed measurement points of the corresponding unit, and is parsed into CSV format text data and saved locally; S33: Read the CSV file saved in step S32, convert the "Time" column to Pandas date and time format, and set the "Time" column as the row index of the dataset; S34: Convert the processed dataset into a NumPy array, remove the numerical data from the time column, and make all data of type float32; S35: Use the MinMaxScaler function from the sklearn library to perform max-min normalization on the NumPy array, limiting the range of values in each column of the array to between 0 and 1. The formula is as follows: ,in, This represents the pressure value at the nth measuring point at time m. This represents the minimum value at all times for the current measuring point. This represents the maximum value at all times for the current measuring point. This represents the result of the nth measurement point after normalization at time m.
[0010] Furthermore, step S4 specifically includes the following sub-steps: S41: Determine whether the data input to the model is single-point data or multi-point data; S42: Determine the duration of historical data input to the model; S43: Determine the duration of the predicted data output by the model; S44: Transform the time series prediction problem into a supervised learning problem, flatten it into a 1×3840 row vector in row order, and perform the above encoding on the normalized NumPy array at any time to generate a new NumPy array containing multiple 1×3840 row vectors. S45: Determine the types of predicted measurement points output by the model; S46: Perform vector segmentation and reshape operations on the row vectors. Segment the pressure value data of 3N measuring points in the first 60 seconds of the selected time and reshape it into a two-dimensional array with a dimension of 60×3N, generating a dataset with dimensions of [sample size, historical data length, number of measuring points]. Extract the 60-second data of the measuring points to be predicted from the pressure value data of 3N measuring points in the last 60 seconds of the selected time, and reshape it into a one-dimensional array of 1×60, generating a label set with dimensions of [sample size, data length of the predicted measuring point].
[0011] Furthermore, in step S5, the dataset and label set obtained in step S4 are divided into training set and test set in a ratio of 8:2. The dataset is divided into training dataset train_X and test dataset test_X in an 8:2 ratio, and the label set is divided into training label set train_y and test label set test_y in an 8:2 ratio.
[0012] Furthermore, in step S6, the state prediction model of the LSTM neural network includes one data input layer, two LSTM computation layers, and one fully connected Dense output layer; the regression loss function is MAE, and the adaptive learning rate optimizer is Adam. The mathematical relationships between the parameters of the LSTM computation layer are as follows:
[0013]
[0014]
[0015]
[0016]
[0017]
[0018] Finally, a Dense fully connected layer is used to hide the output state. A linear transformation is performed to calculate the output at time step t, as shown in the following formula:
[0019] in, Input the time series data into the LSTM network. For input time series The feature data at a single time step t, denoted as the output time series of the LSTM network. , The output hidden state for time step t. This refers to the cell state at the previous time step. This represents the current state of the cell. For the Gate of Oblivion For input gate, For output gate, Forget gate matrix, The input gate matrix, For the output gate matrix, For the new unit state matrix, , , , For bias, It is the sigmoid activation function. For activation function, This is the weight matrix of the fully connected layer. This is the bias vector in the fully connected layer.
[0020] Furthermore, step S7 specifically includes the following sub-steps: S71: Input the training dataset train_X and the training label set train_y into the constructed LSTM cascaded fluid state prediction network model for training; S72: Define the number of neurons n_neuron in the computational layer of the LSTM network model; S73: Define the number of samples n_batch processed during the iteration of training the LSTM network model; S74: Define the number of epochs n_epochs for training the LSTM network model; S75: The model weights are updated using the Adam optimizer. The principle behind this weight update is as follows:
[0021]
[0022]
[0023]
[0024] Where m is the first-order moment estimator and v is the second-order moment estimator. It is the decay rate estimated by the first moment. It is the attenuation rate estimated by the second moment. It is the gradient of the loss function with respect to the model parameters, and t is the current iteration number. It's the learning rate. It is a small constant to prevent the denominator from being zero; S76: The formula for calculating the MAE loss function is:
[0025] Where n is the number of samples in the training set. The actual label for the corresponding sample. These are the predicted values from the model.
[0026] Furthermore, in step S8, the predicted results and the true values are subjected to an inverse transformation using the MinMaxScaler function to restore the original data dimensions; the RMSE is used to calculate the model prediction error, and the formula for calculating RMSE is:
[0027] Where n is the number of samples in the test set. The actual label for the corresponding sample. The model's predicted values are used; the pyplot library in Python is used to display the predicted results and the true labels on the same graph, and OpenCV is used to convert the predicted / true value comparison graphs of all samples into video for storage.
[0028] Furthermore, in step S9, the preset duration is 60 seconds, the dimension of the historical data input to the model is consistent with that of the training model, and the length of the output predicted data is consistent with the duration determined in step S43; the predicted data is processed using pandas and NumPy libraries, the "time" column is formatted to display only hours, minutes and seconds, and the ExcelWriter library is used to store multiple prediction results in an XLS file in order of indexing the time column.
[0029] Beneficial Effects: This invention provides an efficient and accurate prediction method for uranium enrichment production processes. By combining the powerful sequence modeling capabilities of LSTM with the advantages of deep learning, this invention can predict key parameters of cascaded systems under different operating conditions, ensuring system stability and safety. This invention models complex cascaded systems and, by learning the dynamic changes in the fluid state within the system, achieves prediction of the state in the next 60 seconds. Compared with traditional prediction methods, this invention does not rely on complex physical models, reduces dependence on engineers' professional knowledge, and improves the model's generalization ability. Attached Figure Description
[0030] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart of the main process of the cascaded fluid state prediction method based on LSTM neural network described in this embodiment of the invention. Figure 2 This is a diagram of the LSTM neural network structure involved in the cascaded fluid state prediction method based on LSTM neural network described in the embodiments of the present invention. Figure 3 This is a schematic diagram of the LSTM neural network involved in the cascaded fluid state prediction method based on LSTM neural network described in the embodiments of the present invention. Figure 4 This is a graph showing the loss curves of the training and test sets during the training process of the LSTM neural network involved in the cascaded fluid state prediction method based on LSTM neural network described in this embodiment of the invention. Figure 5 This is a comparison of the LSTM neural network prediction signal data and the real data under steady state in the cascaded fluid state prediction method based on LSTM neural network described in this embodiment of the invention. Figure 6 This is a comparison of the LSTM neural network prediction signal data and the real data under transient conditions in the cascaded fluid state prediction method based on LSTM neural network described in this embodiment of the invention. Figure 7 This is a graph showing the change of MAPE over time in the cascaded fluid state prediction method based on LSTM neural network described in this embodiment of the invention. Detailed Implementation
[0031] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.
[0032] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0033] Example 1 See Figure 1-7 A cascaded fluid state prediction method based on LSTM neural network is proposed for predicting the fluid pressure state of a cascaded system in a uranium enrichment plant, enabling early warning of system disturbance trends. The method includes the following steps: S1. Collect real-time data of pressure measuring points in the uranium enrichment plant cascade system. The uranium enrichment plant cascade system consists of N separation stages. Each stage is equipped with a feed main pipe, a concentrate main pipe, and a lean feed main pipe. The pressure measuring points correspond to the pressure monitoring of the feed / concentrate / lean feed main pipes of each separation stage. S2: Analyze and store the collected pressure measurement data; S3: Preprocess the raw pressure data. The preprocessing includes data format conversion, time dimension standardization, array type conversion and normalization, mapping the data to a preset numerical range. S4: Encode the preprocessed time series data, convert it into supervised learning samples, and generate a dataset and a label set. The encoding includes determining the duration of historical data, the duration of predicted data, and segmenting and adjusting the dimensions of the data. S5: Divide the dataset into training and testing sets; S6: Build a state prediction model based on LSTM neural network. The model includes an LSTM computation layer and a fully connected output layer. The LSTM computation layer processes long sequence information through gating mechanisms of input gate, forget gate and output gate, and adopts a regression loss function and an adaptive learning rate optimizer. S7: Input the training set into the LSTM neural network model for training, and minimize the loss function by iteratively updating the model weights; S8: Test on the test set to verify the model's prediction accuracy and show the comparison between the prediction results and the true values; S9: Input the historical pressure data of the measurement point to be predicted into the trained model, predict the fluid pressure state within a preset time period in the future, and store the prediction results.
[0034] This embodiment collects real-time data from 3N pressure measurement points in a uranium enrichment plant cascade system using a distributed control system. After preprocessing such as parsing, storage, format conversion, and normalization, supervised learning samples are encoded and divided into training and testing sets. A prediction model containing two LSTM computation layers and one fully connected output layer is built and trained using the MAE loss function and Adam optimizer. After testing and verifying the accuracy, the model predicts and stores the fluid pressure state within a preset time period. This embodiment does not rely on complex physical models or human experience, fully considers the coupling relationship between measurement points, can provide early warning of system disturbance trends, has high prediction accuracy, and fast response. It can reduce the workload of operators, reduce human error, and provide technical support for the safe and stable operation of the uranium enrichment plant cascade system.
[0035] In a specific example, in step S1, there are 3N pressure measuring points. The data from the 3N pressure measuring points are coupled together and real-time data is obtained through the distributed control system (DCS) interface. Each measuring point records the pressure value of the feed / refined / lean material main pipe of the corresponding separation stage.
[0036] In practice, real-time data from the 32 most correlated monitoring points in the uranium enrichment plant cascade system are acquired through the Distributed Control System (DCS) interface. The uranium enrichment plant cascade system data includes data from 32 monitoring points covering feed and lean feed pressure at stages 1-16. Each monitoring point records the pressure value of the feed / lean feed main pipe for the corresponding unit. During implementation, staff can select the content and quantity of variable data that meet actual needs based on subjective experience or work requirements.
[0037] In a specific example, in step S2, the pressure measurement data is transmitted in a specific format or protocol, and after parsing, it is stored in a format that can support subsequent data processing.
[0038] In a specific instance, step S3 specifically includes the following steps: S31: Organize the stored data into an Excel-readable XLS file. The XLS file contains multiple worksheets, and each worksheet records the pressure value data of the measuring points when pressure disturbances occur in different units. In practice, staff can collect data on the impact of changes in each variable on the other variables to increase the network's prediction accuracy under different conditions. In this embodiment, data from all measurement points when pressure disturbances occur in four different electric regulating valves are selected to construct a dataset. S32: Use Python to delete the first column (serial number) of each sub-table in the XLS file, concatenate the columns to form a master table, and each row in the master table represents the pressure value of 3N measuring points at the current time, expressed as follows: , The file represents the lean / feed measurement points of the corresponding unit, and is parsed into CSV format text data and saved locally; In practice, Python is used to delete the first column (serial number) of each sub-table within the XLS file, and then concatenate them column-by-column into a master table. Each row of data in the table represents the pressure value of the electric regulating valves at 32 measuring points at the current time, which can be represented as follows: .
[0039] S33: Read the CSV file saved in step S32, convert the "Time" column to Pandas date and time format, and set the "Time" column as the row index of the dataset for processing and analysis of time series data; S34: Convert the processed dataset into a NumPy array, remove the numerical data from the time column, and make all data of type float32; S35: Use the MinMaxScaler function from the sklearn library to perform max-min normalization on the NumPy array, limiting the range of values in each column of the array to between 0 and 1. The formula is as follows: ,in, This represents the pressure value at the nth measuring point at time m. This represents the minimum value at all times for the current measuring point. This represents the maximum value at all times for the current measuring point. This represents the result of the nth measurement point after normalization at time m.
[0040] In a specific instance, step S4 specifically includes the following sub-steps: S41: Determine whether the data input to the model is single-point data or multi-point data, i.e., the number of data columns at each time point; In practice, staff can select variables based on whether they are correlated. If there is no obvious correlation between different variables, univariate data can be selected for input to enable the model to predict the future data of that measurement point. In this embodiment, since there is a clear mutual influence between different measurement points, 32 measurement point data are selected for input into the model simultaneously to predict the future data of a specified single measurement point. S42: Determine the duration of historical data for the input model, i.e., the number of rows in each set of historical data; In practice, staff can customize the historical data duration of a single sample input model based on the available time lengths of various variables and hardware storage conditions in the actual situation; in this embodiment, 60 seconds of pressure data from 32 measuring points is selected as the feature matrix input model for a single sample. S43: Determine the duration of the predicted data output by the model, i.e., the number of rows in each set of predicted data; In practice, staff can modify the duration of the data to be predicted according to project requirements; in this embodiment, the data of the measurement point in the next 60 seconds is selected to meet the actual needs. S44: Transform the time series prediction problem into a supervised learning problem, flatten it into a 1×3840 row vector in row order, and perform the above encoding on the normalized NumPy array at any time to generate a new NumPy array containing multiple 1×3840 row vectors. It should be noted that, under the conditions determined in steps S41, S42, and S43, the data input requirement for the model in this embodiment is that all 32 measuring points' data be input, the historical data duration is 60 seconds, and the prediction data duration is 60 seconds. Therefore, the pressure value data from the 32 measuring points for the first 60 seconds of the selected time and the pressure value data from the 32 measuring points for the last 60 seconds of the selected time are input into a custom `series_to_supervised` function for encoding, transforming the time series prediction problem into a supervised learning problem. The data is then flattened into a 1*3840 row vector in row order. The NumPy array obtained after normalization in step S35 can be encoded in the above manner at any time, ultimately generating a new NumPy array containing multiple 1*3840 row vectors.
[0041] S45: Determine the types of predicted measurement points output by the model; In practice, staff can decide the variables to be predicted; in this embodiment, the pressure data of the electric regulating valve at the feed measurement point of the first unit is selected for prediction.
[0042] S46: Perform vector segmentation and reshape operations on the row vectors. Segment the pressure value data of 3N measuring points in the first 60 seconds of the selected time and reshape it into a two-dimensional array with a dimension of 60×3N, generating a dataset with dimensions of [sample size, historical data length, number of measuring points]. Extract the 60-second data of the measuring points to be predicted from the pressure value data of 3N measuring points in the last 60 seconds of the selected time, and reshape it into a one-dimensional array of 1×60, generating a label set with dimensions of [sample size, data length of the predicted measuring point].
[0043] It should be noted that the row vectors in the NumPy array obtained in step S44 are split and reshaped. In this embodiment, the pressure value data of the 32 measuring points in the first 60 seconds of the selected time are split and reshaped into a two-dimensional array with a dimension of 60*32 to generate a new three-dimensional array to form a dataset. The dimensions of this dataset are [sample size, historical data length, number of measuring points]. In the pressure value data of the 32 measuring points in the last 60 seconds of the selected time, only the 60-second data of the measuring points that need to be predicted in step S45 are extracted and reshaped into a one-dimensional array of 1*60 to generate a new two-dimensional array to form a label set. The dimensions of this label set are [sample size, data length of the predicted measuring points].
[0044] In a specific example, in step S5, the dataset and label set obtained in step S4 are divided into training set and test set in a ratio of 8:2. The dataset is divided into training dataset train_X and test dataset test_X in an 8:2 ratio, and the label set is divided into training label set train_y and test label set test_y in an 8:2 ratio.
[0045] In a specific instance, in step S6, see... Figure 2-3 The LSTM neural network state prediction model includes one data input layer, two LSTM computation layers, and one fully connected Dense output layer; the regression loss function is MAE (mean absolute error), and the adaptive learning rate optimizer is Adam. The mathematical relationships between the parameters of the LSTM computation layer are as follows:
[0046]
[0047]
[0048]
[0049]
[0050]
[0051] Finally, a Dense fully connected layer is used to hide the output state. A linear transformation is performed to calculate the output at time step t, as shown in the following formula:
[0052] in, Input the time series data into the LSTM network. For input time series The feature data at a single time step t, denoted as the output time series of the LSTM network. , The output hidden state for time step t. This refers to the cell state at the previous time step. This represents the current state of the cell. For the Gate of Oblivion For input gate, For output gate, Forget gate matrix, The input gate matrix, For the output gate matrix, For the new unit state matrix, , , , For bias, It is the sigmoid activation function. For activation function, This is the weight matrix of the fully connected layer. This is the bias vector in the fully connected layer.
[0053] It should be noted that LSTM is a special type of recurrent neural network designed to address the long-term dependency problem faced by traditional RNNs. LSTM avoids the vanishing and exploding gradient problems by introducing gating mechanisms, including input gates, forget gates, and output gates, and can effectively process and remember long sequences of information.
[0054] In a specific example, step S7 specifically includes the following sub-steps: S71: Input the training dataset train_X and the training label set train_y into the constructed LSTM cascaded fluid state prediction network model for training; S72: Define the number of neurons n_neuron in the computational layer of the LSTM network model. This parameter determines the complexity of the LSTM model. Too few neurons may cause the model to fail to capture the complex relationships in the data, while too many neurons may cause the model to overfit. In specific implementation, the staff can define the number of neurons themselves. In this embodiment, n_neuron is defined as 52. S73: Define the number of samples n_batch during the iteration of LSTM network model training. Passing the entire training dataset to the neural network at once may result in a large amount of computation and memory requirements. Batch processing divides the dataset into small batches, and each batch is used for forward and backward propagation, thereby reducing the computational burden and providing more stable weight updates during training, which helps the model converge. In specific implementation, the staff can define it themselves; in this embodiment, n_batch is defined as 32. S74: Define the number of training epochs n_epoch for the LSTM network model. One epoch represents one training run of the model on the entire training dataset. Through training over multiple epochs, the model can continuously improve its ability to fit the training data, thereby improving performance. In specific implementations, the operator can define the epoch themselves; in this embodiment, n_epoch is defined as 150. S75: The model weights are updated using the Adam optimizer. The principle behind this weight update is as follows:
[0055] Where m is the first-order moment estimator and v is the second-order moment estimator. It is the decay rate estimated by the first moment. It is the attenuation rate estimated by the second moment. It is the gradient of the loss function with respect to the model parameters, and t is the current iteration number. It's the learning rate. It is a small constant to prevent the denominator from being zero; S76: The formula for calculating the MAE loss function is:
[0056] Where n is the number of samples in the training set. The actual label for the corresponding sample. This represents the model's predicted value. During model training, the loss function between the true value and the model's predicted value for each sample is calculated, and the Adam optimizer in step S75 is used to continuously update the model weights during backpropagation to minimize the MAE loss. This process improves the model's ability to learn temporal features from different measurement points. For example... Figure 4 The diagram shows the loss curves of the training and test sets during the training process of the LSTM neural network in this embodiment.
[0057] In a specific example, in step S8, the predicted results and the true values are inversely transformed using the MinMaxScaler function to restore the original data dimensions; the RMSE (Root Mean Square Error) is used to calculate the model prediction error. The formula for calculating RMSE is:
[0058] Where n is the number of samples in the test set. The actual label for the corresponding sample. The model's predicted values are used; the pyplot library in Python is used to display the predicted results and the true labels on the same graph, and OpenCV is used to convert the predicted / true value comparison graphs of all samples into video for storage.
[0059] It should be noted that after predicting each sample in the test set, the pyplot library in Python is used to display the prediction results and the true labels on the same graph. OpenCV is then used to convert the predicted / true value comparison graphs of all samples into a video for storage. The LSTM neural network involved in this embodiment of the invention can predict data for the next 60 seconds, such as... Figure 5 , Figure 6 As shown, the predicted signal data demonstrates better prediction performance compared to the actual data in both steady-state and transient scenarios. Figure 7 As shown, the mean absolute error percentage (MAPE) curve changes with time after the disturbance occurs. After 20 seconds of the disturbance, the MAPE decreases to about 2%, at which point the predicted value meets the pressure requirements of the centrifugal cascade system.
[0060] In a specific example, in step S9, the preset duration is 60 seconds, the dimension of the historical data input to the model is consistent with that of the training model, and the length of the output predicted data is consistent with the duration determined in step S43; the predicted data is processed using pandas and NumPy libraries, the "time" column is formatted to display only hours, minutes and seconds, and the ExcelWriter library is used to store multiple prediction results in an XLS file in order of indexing the time column.
[0061] The core working principle of this invention is based on the long-sequence processing capability of LSTM neural networks, combined with the coupling characteristics and temporal correlation of 3N pressure measuring points in a uranium enrichment plant cascade system. First, real-time data from the measuring points is collected, preprocessed, and encoded into supervised learning samples. Then, the pressure change patterns in historical data are mined using an LSTM gating mechanism, and the model is optimized using the MAE loss function and the Adam optimizer. Finally, the trained model is used to infer the fluid pressure state within a preset time frame based on historical data, capturing disturbance propagation trends and achieving early warning.
[0062] In summary, this embodiment addresses the complex temporal dynamic relationship between operating data and status in a uranium enrichment plant cascade system during pressure disturbances. It accurately predicts key system parameters for the next 60 seconds, which is crucial for identifying potential problems, preventing abnormal operations, and ensuring safe system operation. The LSTM model proposed in this embodiment uses historical data for prediction, reducing the influence of subjective factors on the prediction results and thus mitigating the potential risk of human error.
[0063] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A cascaded fluid state prediction method based on LSTM neural network, characterized in that, Fluid pressure state prediction for uranium enrichment plant cascade systems, enabling early warning of system disturbance trends, includes the following steps: S1. Collect real-time data of pressure measuring points in the uranium enrichment plant cascade system. The uranium enrichment plant cascade system consists of N separation stages. Each stage is equipped with a feed main pipe, a concentrate main pipe, and a lean feed main pipe. The pressure measuring points correspond to the pressure monitoring of the feed / concentrate / lean feed main pipes of each separation stage. S2: Analyze and store the collected pressure measurement data; S3: Preprocess the raw pressure data. The preprocessing includes data format conversion, time dimension standardization, array type conversion and normalization, mapping the data to a preset numerical range. S4: Encode the preprocessed time series data, convert it into supervised learning samples, and generate a dataset and a label set. The encoding includes determining the duration of historical data, the duration of predicted data, and segmenting and adjusting the dimensions of the data. S5: Divide the dataset into training and testing sets; S6: Build a state prediction model based on LSTM neural network. The model includes an LSTM computation layer and a fully connected output layer. The LSTM computation layer processes long sequence information through gating mechanisms of input gate, forget gate and output gate, and adopts a regression loss function and an adaptive learning rate optimizer. S7: Input the training set into the LSTM neural network model for training, and minimize the loss function by iteratively updating the model weights; S8: Test on the test set to verify the model's prediction accuracy and show the comparison between the prediction results and the true values; S9: Input the historical pressure data of the measurement point to be predicted into the trained model, predict the fluid pressure state within a preset time period in the future, and store the prediction results.
2. The cascaded fluid state prediction method based on LSTM neural network according to claim 1, characterized in that, In step S1, there are 3N pressure measuring points. The data from these 3N pressure measuring points are coupled together and real-time data is obtained through the distributed control system interface. Each measuring point records the pressure value of the feed / refined / lean material main pipe of the corresponding separation stage.
3. The cascaded fluid state prediction method based on LSTM neural network according to claim 1, characterized in that, In step S2, the pressure measurement data is transmitted in a specific format or protocol, and after parsing, it is stored in a format that can support subsequent data processing.
4. The cascaded fluid state prediction method based on LSTM neural network according to claim 1, characterized in that, Step S3 specifically includes the following steps: S31: Organize the stored data into an Excel-readable XLS file. The XLS file contains multiple worksheets, and each worksheet records the pressure value data of the measuring points when pressure disturbances occur in different units. S32: Use Python to delete the first column (serial number) of each sub-table in the XLS file, concatenate the columns to form a master table, and each row in the master table represents the pressure value of 3N measuring points at the current time, expressed as follows: , The file represents the lean / feed measurement points of the corresponding unit, and is parsed into CSV format text data and saved locally; S33: Read the CSV file saved in step S32, convert the "Time" column to Pandas date and time format, and set the "Time" column as the row index of the dataset; S34: Convert the processed dataset into a NumPy array, remove the numerical data from the time column, and make all data of type float32; S35: Use the MinMaxScaler function from the sklearn library to perform max-min normalization on the NumPy array, limiting the range of values in each column of the array to between 0 and 1. The formula is as follows: ,in, This represents the pressure value at the nth measuring point at time m. This represents the minimum value at all times for the current measuring point. This represents the maximum value at all times for the current measuring point. This represents the result of the nth measurement point after normalization at time m.
5. The cascaded fluid state prediction method based on LSTM neural network according to claim 1, characterized in that, Step S4 specifically includes the following sub-steps: S41: Determine whether the data input to the model is single-point data or multi-point data; S42: Determine the duration of historical data input to the model; S43: Determine the duration of the predicted data output by the model; S44: Transform the time series prediction problem into a supervised learning problem, flatten it into a 1×3840 row vector in row order, and perform the above encoding on the normalized NumPy array at any time to generate a new NumPy array containing multiple 1×3840 row vectors. S45: Determine the types of predicted measurement points output by the model; S46: Perform vector segmentation and reshape operations on the row vectors. Segment the pressure value data of 3N measuring points in the first 60 seconds of the selected time and reshape it into a two-dimensional array with a dimension of 60×3N, generating a dataset with dimensions of [sample size, historical data length, number of measuring points]. Extract the 60-second data of the measuring points to be predicted from the pressure value data of 3N measuring points in the last 60 seconds of the selected time, and reshape it into a one-dimensional array of 1×60, generating a label set with dimensions of [sample size, data length of the predicted measuring point].
6. The cascaded fluid state prediction method based on LSTM neural network according to claim 1, characterized in that, In step S5, the dataset and label set obtained in step S4 are divided into training set and test set in a ratio of 8:
2. The dataset is divided into training dataset train_X and test dataset test_X in an 8:2 ratio, and the label set is divided into training label set train_y and test label set test_y in an 8:2 ratio.
7. The cascaded fluid state prediction method based on LSTM neural network according to claim 1, characterized in that, In step S6, the state prediction model of the LSTM neural network includes one data input layer, two LSTM computation layers, and one fully connected Dense output layer; the regression loss function is MAE, and the adaptive learning rate optimizer is Adam. The mathematical relationships between the parameters of the LSTM computation layer are as follows: Finally, a Dense fully connected layer is used to hide the output state. A linear transformation is performed to calculate the output at time step t, as shown in the following formula: in, Input the time series data into the LSTM network. For input time series The feature data at a single time step t, denoted as the output time series of the LSTM network. , The output hidden state for time step t. This refers to the cell state at the previous time step. This represents the current state of the cell. For the Gate of Oblivion For input gate, For output gate, Forget gate matrix, The input gate matrix, For the output gate matrix, For the new unit state matrix, , , , For bias, It is the sigmoid activation function. For activation function, This is the weight matrix of the fully connected layer. This is the bias vector in the fully connected layer.
8. The cascaded fluid state prediction method based on LSTM neural network according to claim 1, characterized in that, Step S7 specifically includes the following sub-steps: S71: Input the training dataset train_X and the training label set train_y into the constructed LSTM cascaded fluid state prediction network model for training; S72: Define the number of neurons n_neuron in the computational layer of the LSTM network model; S73: Define the number of samples n_batch processed during the iteration of training the LSTM network model; S74: Define the number of epochs n_epochs for training the LSTM network model; S75: The model weights are updated using the Adam optimizer. The principle behind this weight update is as follows: Where m is the first-order moment estimator and v is the second-order moment estimator. It is the decay rate estimated by the first moment. It is the attenuation rate estimated by the second moment. It is the gradient of the loss function with respect to the model parameters, and t is the current iteration number. It's the learning rate. It is a small constant to prevent the denominator from being zero; S76: The formula for calculating the MAE loss function is: Where n is the number of samples in the training set. The actual label for the corresponding sample. These are the predicted values from the model.
9. The cascaded fluid state prediction method based on LSTM neural network according to claim 1, characterized in that, In step S8, the predicted results and the true values are subjected to an inverse MinMaxScaler function transformation to restore the original data dimensions; the RMSE is used to calculate the model prediction error, and the formula for calculating RMSE is: Where n is the number of samples in the test set. The actual label for the corresponding sample. The model's predicted values are used; the pyplot library in Python is used to display the predicted results and the true labels on the same graph, and OpenCV is used to convert the predicted / true value comparison graphs of all samples into video for storage.
10. The cascaded fluid state prediction method based on LSTM neural network according to claim 1, characterized in that, In step S9, the preset duration is 60 seconds, the dimension of the historical data input to the model is consistent with that of the training model, and the length of the output predicted data is consistent with the duration determined in step S43; the predicted data is processed using pandas and NumPy libraries, the "time" column is formatted to display only hours, minutes and seconds, and the ExcelWriter library is used to store multiple prediction results in an XLS file in order of indexing the time column.