Method for predicting time sequence data of nuclear power main transformer of coding and decoding mechanism
By using an improved Seq2Seq framework and attention mechanism, combined with teacher-forced techniques and early shutdown mechanisms, the long-term dependence and multivariate prediction problems of nuclear power main transformer time series data were solved, achieving high-precision nuclear power equipment status monitoring and fault early warning, thus improving the safety and economic benefits of nuclear power plants.
Patent Information
- Application Number
- CN202511540862.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-27
- Publication Date
- 2026-02-13
AI Technical Summary
Existing time series forecasting methods are difficult to effectively capture nonlinear patterns and multivariate correlations in nuclear power plant main transformer data, especially in terms of long-term dependence and multi-step forecasting, and cannot meet the high-precision requirements of nuclear power plant equipment condition monitoring.
An improved Seq2Seq framework is adopted, combining attention mechanism and teacher-forced technique. A codec is constructed through a bidirectional LSTM network to predict the time series data of nuclear power main transformer. An early stopping mechanism is introduced to prevent overfitting. The Adam optimizer and mean squared error loss function are used to optimize the model.
It significantly improves the prediction accuracy of nuclear power main transformer time series data, solves the problem of long-term dependence, supports multi-variable and multi-step prediction, enhances real-time monitoring capabilities, reduces maintenance costs, and improves the safety and economic benefits of nuclear power plants.
Smart Images
Figure CN121524985A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of coding and decoding, and particularly relates to a method for predicting time sequence data of nuclear power main transformers based on an encoding and decoding mechanism. BACKGROUND
[0002] The nuclear power main transformer is a core device of a nuclear power plant, and its operating state directly affects the safety and stability of the nuclear power plant. By predicting its time sequence data, potential abnormalities can be identified in advance, and maintenance strategies can be optimized. However, these data usually have characteristics such as multivariate, high noise, complex nonlinear relationship and long-term dependence, which bring challenges to accurate prediction.
[0003] Traditional time series prediction methods, such as ARIMA and exponential smoothing method, are difficult to effectively capture the nonlinear patterns and multivariate correlations in the data due to their dependence on statistical assumptions. Although deep learning methods such as RNN and LSTM have shown advantages in processing time series data, they still have limitations in solving long-term dependence problems and improving multi-step prediction accuracy.
[0004] The encoding and decoding mechanism (Encoder-Decoder) has been successful in natural language processing, but there is no successful application case in predicting nuclear power main transformer time sequence data. Existing methods lack targeted optimization of nuclear power data characteristics, making it difficult to meet the high accuracy requirements of practical applications. Therefore, an improved encoding and decoding prediction method is needed to improve prediction accuracy and provide support for device state monitoring and intelligent maintenance. SUMMARY
[0005] The purpose of the present application is to provide a method for predicting nuclear power main transformer time sequence data based on an encoding and decoding mechanism, which can improve prediction accuracy and provide support for device state monitoring and intelligent maintenance.
[0006] The technical solution of the present application is as follows: a method for predicting nuclear power main transformer time sequence data based on an encoding and decoding mechanism, including modeling process and prediction process, the modeling process is used to build and train a time series prediction model based on an encoding and decoding mechanism; the prediction process is used to predict future values and performance of nuclear power main transformer time sequence data using the trained model.
[0007] The modeling process includes: Step 1: data input and preprocessing; Step 2: model architecture; Step 3: model training.
[0008] The step 1 includes: Step 11: data collection Collecting nuclear power main transformer measurement point data, the data of each time step can include multiple variable characteristics, forming a multi-dimensional feature vector, and the collected data is used to construct a complete time series data set, denoted as , wherein is the multi-variable feature vector of the tth time step, T is the sequence length, and d is the variable dimension; Step 12: data standardization The standardization process includes calculating the mean and standard deviation σ of each feature, and then converting the original data to standardized data: This conversion makes each feature have a standard normal distribution with a mean of 0 and a standard deviation of 1, which helps to improve the stability and convergence speed of model training; Step 12: data segmentation The standardized data set is divided into training set and validation set according to time window, and sliding window technology is used to construct input-output pairs, input is historical sequence, and output is future sequence, training set is used for model parameter learning, and validation set is used for model performance evaluation and hyperparameter adjustment.
[0009] The step 2 includes using an improved Seq2Seq framework, including an encoder, an attention mechanism and a decoder; The encoder uses a bidirectional LSTM network structure to capture forward and backward dependency features of the time series, and the input standardized time series X is processed through the LSTM layer to output the hidden state The hidden state of the last time step of the encoder is converted into the initial context vector through a fully connected layer, which is used for subsequent attention mechanism: , wherein is an affine transformation function responsible for mapping the hidden state to the context space, is the hidden state of the last time step of the encoder; The attention mechanism is used to dynamically calculate the similarity between the current hidden state of the decoder and the encoder output H, and dynamically generate the attention weight , the calculation formula is: , wherein is an affine transformation function, is an affine transformation parameter, is a weight normalization function; the finally generated context vector is the weighted sum of the encoder outputs: The decoder generates the predicted sequence step by step based on the LSTM network structure, combining the context vector s, the current input, and the previous hidden state. The output of the decoder is calculated as follows: where, is the time series feature extraction function, is the prediction value at step t.
[0010] The step 3 includes: Step 31: Establishing a loss function Set the loss function as the mean square error, which measures the difference between the predicted value and the true value : where N is the number of samples; Step 32: Optimization algorithm selection Use the Adam optimizer to update the model parameters. The Adam optimizer has the ability to dynamically adjust the learning rate, and the initial learning rate is set to 0.001 to ensure the stability and efficiency of the training; Step 33: Training strategy selection Introduce teacher forcing, use the true value as the decoder input in the training process with a certain probability, reduce the prediction error, and accelerate the model convergence; Step 34: Training process In the training loop, iterate through multiple training rounds, perform encoding, attention calculation, decoding, and loss calculation for each batch of data, and update the model parameters through backpropagation; Step 35: Early stopping mechanism Use the early stopping mechanism. If the validation set loss does not decrease for several consecutive rounds, stop training and save the best model weights to prevent overfitting.
[0011] The prediction process includes: Step 1: Data preparation and preprocessing; Step 2: Model loading and prediction.
[0012] The step 1 includes: Step 11: Load test data Load the test set data and use the same standardization processing method as in the training phase: Construct the input sequence according to the time window to ensure consistency with the training format; Step 11: Constructing input sequence The input sequence is constructed according to the time window, and the format is consistent with the training to facilitate the prediction of the model.
[0013] The step 2 comprises: Step 21: Loading model The best model weight saved in the training stage is loaded, and the model is set to evaluation mode to ensure that no parameter update occurs during the prediction process; Step 22: Prediction The test data is input into the encoder to generate the hidden state and context vector; the weight is calculated through the attention mechanism to generate the context vector input into the decoder; the decoder uses the previous prediction value as the next input in a self-recursive manner to generate a future multi-step prediction sequence; Step 23: Result processing The prediction result is processed by inverse standardization to restore the original data range.
[0014] The present application has the beneficial effects that: the present application improves the Seq2Seq framework and the attention mechanism, significantly improves the prediction accuracy, solves the long-time dependence problem, supports multi-variable multi-step prediction, and adapts to complex operating environment. The teacher forcing technology is introduced to optimize the training efficiency, and the early stopping mechanism is combined to effectively avoid overfitting. The self-recursive prediction enhances the real-time monitoring capability, meets the low delay requirement, provides visual results and evaluation indexes, provides decision support for operation and maintenance, discovers fault risks in advance, reduces maintenance cost, improves the safety and economic benefit of nuclear power plants, and provides an innovative solution for intelligent monitoring and predictive maintenance of nuclear power equipment. BRIEF DESCRIPTION OF DRAWINGS
[0015] Figure 1 A flowchart of a method for predicting nuclear power main transformer time series data according to the encoding and decoding mechanism provided by the present application; Figure 2 The real value and predicted value comparison and residual comparison graph of short-term prediction for predicting the first time point (10 minutes) is shown; Figure 3 The real value and predicted value comparison and residual comparison graph of medium-term prediction for predicting the sixth time point (60 minutes) is shown; Figure 4 The real value and predicted value comparison and residual comparison graph of long-term prediction for predicting the twelfth time point (120 minutes) is shown. DETAILED DESCRIPTION
[0016] The present application will be further described in detail below in combination with the drawings and specific embodiments.
[0017] The application provides a method for nuclear power main transformer time series data prediction based on an encoding-decoding mechanism, which is suitable for multi-step prediction of nuclear power main transformer equipment. By modeling and analyzing the time series data of key parameters of nuclear power main transformer, the application aims to improve prediction accuracy and provide technical support for state monitoring and fault warning of nuclear power plant equipment. The method combines the Seq2Seq model and attention mechanism in deep learning, which can effectively solve the long-time dependence problem in complex time series data and enhance the accurate prediction ability of the equipment running state.
[0018] As shown in Figure 1 , a method for nuclear power main transformer time series data prediction based on an encoding-decoding mechanism includes modeling and prediction processes, as follows: The modeling process is to build and train a time series prediction model based on an encoding-decoding mechanism, with the following specific steps: Step 1: Data input and preprocessing Step 11: Data collection Collect nuclear power main transformer measurement point data, such as A-phase oil temperature, B-phase oil temperature, etc., in the form of time series. Each time step data can include multiple variable features, forming a multi-dimensional feature vector. The collected data is used to build a complete time series dataset, denoted as , where is the multi-variable feature vector at the t-th time step, T is the sequence length, and d is the variable dimension.
[0019] Step 12: Data standardization To reduce the impact of noise and outliers on model learning, standardize the data. The standardization process includes calculating the mean and standard deviation σ of each feature, and then converting the original data to standardized data: This conversion makes each feature have a standard normal distribution with a mean of 0 and a standard deviation of 1, which helps improve the stability and convergence speed of model training.
[0020] Step 12: Data segmentation Divide the standardized dataset into training and validation sets according to the time window. Use sliding window technology to build input-output pairs, with the input being the historical sequence and the output being the future sequence. The training set is used for model parameter learning, and the validation set is used to evaluate model performance and adjust hyperparameters.
[0021] Step 2: Model architecture In the embodiment of the application, an improved Seq2Seq framework is used, which includes an encoder (Encoder), an attention mechanism (Attention), and a decoder (Decoder).
[0022] The encoder adopts a bidirectional LSTM network structure, which can capture the forward and backward dependency features of time series. After the input normalized time series X is processed by the LSTM layer, the hidden state is output. The hidden state of the last time step of the encoder is converted into the initial context vector through a fully connected layer, which is used for subsequent attention mechanism: where is an affine transformation function responsible for mapping the hidden state to the context space. is the hidden state of the last time step of the encoder.
[0023] The attention mechanism is used to dynamically calculate the similarity between the current hidden state of the decoder and the encoder output H, and dynamically generate attention weights , the calculation formula is: where, and are affine transformation parameters, is a weight normalization function. The final generated context vector is the weighted sum of the encoder output: The decoder is based on the LSTM network structure, combined with the context vector s, the current input and the last hidden state, to gradually generate the predicted sequence. The output of the decoder is calculated as follows: where, is a time series feature extraction function, is the predicted value at step t.
[0024] Step 3: Model training Step 31: Establish loss function Set the loss function as Mean Squared Error (MSE) to measure the difference between the predicted value and the true value : where N is the number of samples.
[0025] Step 32: Optimization algorithm selection The model parameters are updated using the Adam optimizer, which has the ability to dynamically adjust the learning rate. The initial learning rate is set to 0.001 to ensure the stability and efficiency of the training.
[0026] Step 33: Training strategy selection The Teacher Forcing technique is introduced, which uses real values as decoder input during training with a certain probability, reducing prediction errors and accelerating model convergence.
[0027] Step 34: Training process In the training loop, multiple training epochs are traversed, and encoding, attention calculation, decoding, and loss calculation are performed for each batch of data. Model parameters are updated through backpropagation.
[0028] Step 35: Early stopping mechanism The Early Stopping mechanism is adopted. If the validation set loss does not decrease for a certain number of rounds (e.g., 10 epochs), training is stopped and the best model weights are saved to prevent overfitting.
[0029] The prediction process uses the trained model to predict future values and evaluate performance for nuclear power transformer time series data. The specific steps are as follows: Step 1: Data preparation and preprocessing Step 11: Load test data Load test set data and use the same standardization processing method as in the training phase: Construct input sequences according to time windows to ensure consistency with training formats.
[0030] Step 11: Construct input sequences Construct input sequences according to time windows to ensure consistency with training formats to facilitate model prediction.
[0031] Step 2: Model loading and prediction Step 21: Load model Load the best model weights saved during training and set the model to evaluation mode to ensure no parameter updates during prediction.
[0032] Step 22: Prediction Input test data into the encoder to generate hidden states and context vectors. Calculate weights through the attention mechanism to generate context vector input into the decoder. The decoder generates a future multi-step prediction sequence in a self-recursive manner (using the previous prediction value as the next input).
[0033] Step 23: Result processing The prediction results are de-standardized to recover the original data range. Specific embodiments Implementation background The main transformer of nuclear power is the core equipment of the nuclear power station, and its running state is directly related to the safety and stability of the power station. Oil temperature is a key monitoring indicator, and abnormal changes may indicate equipment failure or potential risks. This implementation case uses a time series data prediction method based on an encoding-decoding mechanism (Seq2Seq model) to predict and detect abnormalities in nuclear power main transformer oil temperature data to verify the application value of the model in actual operation and maintenance scenarios.
[0035] Implementation object and data This case selects the data of three sensor points on the nuclear power main transformer for prediction and anomaly detection, including A-phase oil temperature 1, A-phase oil temperature 2, and B-phase oil temperature 1. The data collection interval is 10 minutes, and the historical data is divided into training set and test set after preprocessing, and the test set contains 1000 samples. The input is the past oil temperature value, and the output is the future oil temperature prediction value.
[0036] Implementation steps Modeling process: In the data input and preprocessing stage, first collect the historical oil temperature data of the nuclear power main transformer, record at 10-minute intervals to ensure the continuity and integrity of the time series; then standardize the data to convert it to a distribution form with a mean of 0 and a standard deviation of 1 to eliminate the dimension effect and improve the stability of model training; finally, the data set is divided into training set (80%) and validation set (20%) in proportion, which are used for model parameter learning and performance evaluation respectively to ensure that the model is not over-fitted and has generalization ability.
[0037] In the model architecture design stage, a Seq2Seq model based on the encoding-decoding mechanism is constructed, the encoder uses a bidirectional LSTM structure to extract deep features of time series data, the decoder uses a unidirectional LSTM structure to generate prediction sequences, and the attention mechanism is introduced to dynamically allocate weights, focus on key information, improve long-term dependency capture ability, and improve long-time step prediction performance.
[0038] In the model training stage, the Adam optimizer is selected, the learning rate is set to 0.001 to ensure fast convergence, and the mean square error (MSE) is selected as the loss function to quantify the prediction error; the batch size is set to 32 during training, the number of training rounds is 50, the teacher forcing strategy is used to assist learning, and the early stopping mechanism is introduced to prevent overfitting; after training, the best model weight is saved according to the validation set performance to ensure that the prediction task uses the optimal parameters.
[0039] Prediction process: In the data preparation and preprocessing stage, a test set of data containing 1000 samples is loaded and standardized to be consistent with the training data; then, an input sequence is constructed, which contains oil temperature values at several past time points, to predict future oil temperature values and provide a data foundation for the prediction task.
[0040] During the model loading and prediction phase, the best model weights saved during the training phase are loaded to ensure the use of optimal parameters. Then, an autoregressive prediction method is used to predict the oil temperature values at 1 time point (10 minutes), 6 time points (60 minutes), and 12 time points (120 minutes), respectively. Multi-step prediction is achieved by gradually updating the input sequence to evaluate the prediction performance of the model in different time ranges.
[0041] Test Results and Analysis Model performance is measured using the following metrics: MSE (mean squared error) is used to evaluate the average of the squared errors between the predicted and true values; MAE (mean absolute error) is used to measure the average of the absolute errors between the predicted and true values; MAPE (mean absolute percentage error) is used to evaluate the percentage error between the predicted and true values; and R² (coefficient of determination) is used to measure how well the model fits the data, with a value closer to 1 indicating a better fit.
[0042] The test results are as follows: Predict the first time point (10 minutes). Table 1. Sensor Prediction Performance Evaluation Results at the First Time Point (10 Minutes) Predict the 6th time point (60 minutes). Table 2. Sensor performance evaluation results at the 6th time point (60 minutes). Predict the 12th time point (120 minutes). Table 3. Sensor performance evaluation results at the 12th time point (120 minutes). Results Analysis like Figures 2-4 As shown, the model exhibits extremely high fitting ability in short-time step (10 minutes) predictions, with R² values for all sensor data approaching 1, indicating that the model can accurately fit the data and has very high prediction accuracy. Furthermore, the MSE and MAE values for short-time step predictions are also extremely low, further demonstrating that the error between the model's predicted values and the actual values is minimal, resulting in outstanding performance.
[0043] As the prediction time step lengthens (e.g., 60 minutes and 120 minutes), although the error increases, especially the R² value decreases at 120 minutes prediction, the model still shows significant advantages in medium and long-term prediction. The model can better capture the main trend of the data and keep consistent with the overall trend of the true value, which indicates that the model has strong trend prediction ability. Even if there is a certain deviation in the prediction value at 120 minutes prediction, the model can still maintain relatively stable prediction performance in the environment with large data fluctuations, showing its robustness and adaptability in medium and long-term prediction. Although the influence of data fluctuations on medium and long-term prediction exists, the model can still provide valuable reference results in medium and long-term prediction by capturing key change trends.
[0044] Overall, the model not only performs outstandingly in short-term prediction, but also exhibits unique advantages and potential in medium and long-term prediction, providing reliable support for complex time series prediction tasks.
Claims
1. A method for predicting time-series data of nuclear power main transformers using an encoding and decoding mechanism, characterized in that: It consists of two parts: a modeling process and a prediction process. The modeling process is used to build and train a time series prediction model based on an encoding and decoding mechanism. The prediction process is used to use the trained model to predict future values and evaluate the performance of nuclear power main transformer time series data.
2. The method for predicting nuclear power main transformer time series data using an encoding and decoding mechanism as described in claim 1, characterized in that, The modeling process includes: Step 1: Data input and preprocessing; Step 2: Model Architecture; Step 3: Model training.
3. The method for predicting nuclear power main transformer time series data using an encoding and decoding mechanism as described in claim 2, characterized in that, Step 1 includes: Step 11: Data Acquisition Data is collected from measuring points of the main transformer in the nuclear power plant. Data at each time step can include multiple variable features, forming a multidimensional feature vector. The collected data is used to construct a complete time-series dataset, denoted as... ,in Let be the multivariate feature vector at the t-th time step, where T is the sequence length and d is the variable dimension; Step 12: Data Standardization The standardization process includes calculating the mean of each feature. And the standard deviation σ, then transform the original data into standardized data: This transformation results in each feature having a standard normal distribution with a mean of 0 and a standard deviation of 1, which helps improve the stability and convergence speed of model training. Step 12: Data Segmentation The standardized dataset is divided into training and validation sets according to time windows. A sliding window technique is used to construct input-output pairs, with the input being historical sequences and the output being future sequences. The training set is used to learn the model parameters, and the validation set is used to evaluate the model performance and adjust the hyperparameters.
4. The method for predicting nuclear power main transformer time series data using an encoding and decoding mechanism as described in claim 2, characterized in that: Step 2 includes adopting an improved Seq2Seq framework, which consists of three parts: an encoder, an attention mechanism, and a decoder. The encoder employs a bidirectional LSTM network structure to capture the forward and backward dependency features of the time series. The input is a standardized time series X, which, after processing by the LSTM layers, outputs the hidden state. The hidden state of the encoder at the last time step is transformed into the initial context vector through a fully connected layer. This is used for subsequent attention mechanisms. in, It is an affine transformation function responsible for mapping the hidden state to the context space. This represents the hidden state of the encoder at the last time step. The attention mechanism is used to dynamically calculate the similarity between the current hidden state of the decoder and the encoder output H, and to dynamically generate attention weights. The calculation formula is: in, Let be the affine transformation function. For affine transformation parameters, This is the weight normalization function; the final generated context vector It is a weighted sum of the encoder outputs: The decoder, based on an LSTM network structure, generates a prediction sequence step by step by combining the context vector s, the current input, and the previous hidden state. The decoder output is calculated as follows: in, For time series feature extraction function, This is the predicted value at step t.
5. The method for predicting nuclear power main transformer time series data using an encoding and decoding mechanism as described in claim 2, characterized in that, Step 3 includes: Step 31: Establish the loss function Set the loss function to mean squared error, and measure the predicted value. Compared with the true value Differences: Where N is the number of samples; Step 32: Optimization Algorithm Selection The Adam optimizer is used to update model parameters. The Adam optimizer has the ability to dynamically adjust the learning rate. The initial learning rate is set to 0.001 to ensure the stability and efficiency of training. Step 33: Training Strategy Selection Introducing teacher coercion, during the training process, the true value is used as the decoder input with a certain probability, which reduces prediction error and accelerates model convergence; Step 34: Training Process In the training loop, multiple training rounds are traversed, and encoding, attention calculation, decoding, and loss calculation are performed on each batch of data. The model parameters are then updated through backpropagation. Step 35: Early Stop Mechanism An early stopping mechanism is adopted: if the validation set loss does not decrease for several consecutive rounds, training is stopped and the optimal model weights are saved to prevent overfitting.
6. The method for predicting nuclear power main transformer time series data using an encoding and decoding mechanism as described in claim 1, characterized in that, The prediction process includes: Step 1: Data preparation and preprocessing; Step 2: Model loading and prediction.
7. The method for predicting nuclear power main transformer time series data using an encoding and decoding mechanism as described in claim 6, characterized in that, Step 1 includes: Step 11: Load test data Load the test set data and apply the same standardization process as during the training phase: Construct the input sequence according to the time window to ensure that it is consistent with the format during training; Step 11: Construct the input sequence The input sequence is constructed according to the time window to ensure that it is consistent with the format during training, so that the model can make predictions.
8. The method for predicting nuclear power main transformer time series data using an encoding and decoding mechanism as described in claim 6, characterized in that, Step 2 includes: Step 21: Load the model Load the best model weights saved during the training phase and set the model to evaluation mode to ensure that no parameter updates occur during the prediction process; Step 22: Prediction The test data is input into the encoder to generate hidden states and context vectors; weights are calculated through an attention mechanism to generate context vectors which are then input into the decoder; the decoder uses the previous step's prediction value as the next step's input in an autoregressive manner to generate a multi-step prediction sequence. Step 23: Result Processing The prediction results are destandardized to restore them to the original data range.