Intelligent peak shaving method for nuclear power plant based on model and data dual drive

Abstract

The invention discloses a nuclear power plant intelligent peak shaving method based on model and data dual drive, which comprises the following steps: 1, carrying out reactor core modeling, and carrying out calculation by using reactor core physical software to obtain reactor core parameters; 2, combining the operation data of the nuclear power plant with the reactor core parameters to construct a data set, and preprocessing the data set; thirdly, training the neural network model; 4, setting a rod position at the next moment, and inputting the to-be-predicted input data into the trained neural network model to predict the axial power offset (AO) of the reactor core at the next moment after calculation and construction of the to-be-predicted input data by using physical software of the reactor core; and 5, judging whether the predicted AO meets requirements or not, if the predicted AO meets the requirements, taking the corresponding rod position as the target rod position of the next moment, otherwise, re-adjusting the rod position and returning to the step 4 to re-predict the AO. According to the invention, the AO can be accurately predicted in real time, the artificial decision risk is reduced, and intelligent peak regulation of the nuclear power plant can be realized on the premise of ensuring the safety of the reactor core.

Description

Technical Field

[0001] The present invention relates to the technical field of nuclear reactor physics and nuclear power plant operation control, and in particular to an intelligent peak-shaving method for a nuclear power plant based on dual drive of a model and data. Background Art

[0002] The rapid development of my country's nuclear power industry, while providing a greater source of clean energy, also places immense pressure on power grids in regions with relatively high nuclear power production and rapidly growing installed capacity. To ensure stable grid operation, nuclear power faces an urgent need to participate in peak load regulation. Furthermore, nuclear power units in some regions of China are already frequently participating in peak load regulation based on grid demand, necessitating urgent research into peak load regulation at nuclear power plants.

[0003] During peaking operations in nuclear power plants, frequent changes in control rod position and power can cause poison oscillations, leading to localized power distortion and uneven core temperature distribution, resulting in reactivity control issues and associated safety risks. Therefore, controlling the core axial power offset (AO) and maintaining power fluctuations within a certain range are key concerns during peaking operations. Currently, nuclear power plants typically employ constant AO control, drawing on the experience of nuclear power plants in France and other countries, to minimize AO fluctuations. Therefore, calculating or predicting AO is crucial. However, due to rapid power fluctuations and frequent rod movements, core operating conditions fluctuate dramatically. Traditional peaking schemes calculated preemptively through core modeling alone are unable to track core conditions in real time and accurately predict AO. Consequently, the actual peaking process relies heavily on individual experience. This not only fails to meet the high-frequency and deep peaking requirements of pressurized water reactors, but also increases decision-making and human risk, hindering core safety. In traditional neural network models, they are mainly driven by individual data, with directly measured macro data as input data. However, these macro data are difficult to accurately reflect the core state and it is difficult to accurately predict AO.

[0004] If we can use both model-driven methods, such as using core physics software to perform core modeling to obtain relevant core parameters that can reflect the core state, and data-driven methods, and then use neural network models for data-driven methods based on core parameters and nuclear power plant operating parameters to efficiently and accurately predict AO, and accurately know the impact of current rod position adjustments on future AO, we can intelligently assist operators in AO control to reduce the risks of human decision-making and intervention, thereby achieving intelligent peak regulation and ensuring core safety. Summary of the Invention

[0005] In order to solve the problems existing in the above-mentioned prior art, the purpose of the present invention is to provide an intelligent peak-shaving method for nuclear power plants based on dual-drive of models and data. The method of the present invention meets the peak-shaving needs of my country's nuclear power plants, ensures the safety of the core, and assists operators in real-time, accurately and efficiently predicting the core axial power offset AO in complex working conditions such as the peak-shaving process, thereby reducing the risks in the peak-shaving process and realizing intelligent peak-shaving of nuclear power plants under the premise of ensuring the safety of the core.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions: A method for intelligent peak-shaving of nuclear power plants based on dual drive of models and data includes the following steps: Step 1: Select the geometric and material parameters of the nuclear power plant reactor core and fuel assemblies for core modeling, and use core physics software to perform simulation calculations to obtain the required but not directly measurable core parameters; Step 2: Select the nuclear power plant operating data from historical peak-shaving and the core parameters calculated by core physics software simulation. The two together constitute the input data set required for neural network model training. The core axial power offset (AO) value at the next moment is used as the label data set. The input data set and the label data set are preprocessed. The present invention combines the core data calculated using professional core physics software with the nuclear power plant operation data, which can largely solve the data source problem and improve the accuracy and robustness of data-driven methods such as neural network models.

[0007] Step 3: Divide the input dataset and label dataset preprocessed in step 2 into a training set, a validation set, and a test set, and use the training set, validation set, and test set to train the neural network model to obtain a trained neural network model with the required accuracy. By properly dividing the training set, validation set, and test set, we can not only improve the expressiveness of the neural network model, but also prevent overfitting, thereby maximizing the accuracy and robustness of the neural network model.

[0008] Step 4: In actual peak shaving, the rod position at the next moment is set. After tracking and simulating the core using core physics software to calculate the core parameters, the obtained core parameters and the actual nuclear power plant operating data during peak shaving are used as the input data to be predicted. The predicted input data is preprocessed and then input into the trained neural network model to predict the core axial power offset AO at the next moment in real time. Step 5: Determine whether the predicted core axial power offset AO meets the requirements. If not, readjust the rod position and return to step 4 to re-predict the core axial power offset AO. If it meets the requirements, use the corresponding rod position as the target rod position for the next moment to assist the operator in decision-making and peak-shaving control, thereby realizing intelligent peak-shaving.

[0009] By presetting and calibrating the control rod position, the operator's expert experience can be incorporated into the rod position selection, and the trained neural network model can be used to determine whether the predicted AO meets the requirements, providing the operator with an intelligent tool to participate in peak-shaving decision-making, assisting him in decision-making and peak-shaving control.

[0010] The core parameters calculated by the core physics software in step 1 include fuel burnup and neutron poison distribution. The advantage of this is that the fuel burnup and neutron poison distribution can be used to characterize the core state during peak regulation, especially the xenon oscillation.

[0011] The nuclear power plant operating data in step 2 includes reactor power, control rod assembly positions, core axial power offset (AO) measurements from external detectors, fuel assembly enrichment, and the amount of burnable poisons. This data is advantageous because nuclear power plants have a large amount of operating data during peak load periods, particularly AO measurements. This provides a rich data source for neural network model training and reduces the error between predicted and actual AO.

[0012] The pretreatment in step 2 and step 4 includes: Dataset reconstruction: All input datasets and label datasets are reconstructed into the data input format required by the neural network model according to the preset order and structure; Dataset normalization: Normalize all reconstructed input datasets and label datasets to the range of 0 to 1 to reduce the impact of dimension and magnitude.

[0013] The neural network model described in step 3 uses the long short-term memory neural network model LSTM. The advantage of the LSTM model is that it can effectively transmit and express information in long time series without causing useful information from a long time ago to be ignored. It can also solve the gradient vanishing problem that exists in conventional recurrent neural networks.

[0014] The neural network model is trained as described in step 3, including forward propagation, calculating the loss function, solving the gradient in reverse to reduce the loss function, and using the early stopping mechanism to interrupt the training process in advance when the conditions are met; The early stopping mechanism interrupts neural network training when the validation set loss function stops decreasing after a certain number of rounds, or when the loss decreases by less than a preset value. The advantage of using the early stopping mechanism is that it maintains the neural network model's feature extraction capabilities while preventing overfitting.

[0015] The loss function uses the mean square error loss function.

[0016] Meeting the requirements in step 5 means that the error between the predicted core axial power offset (AO) and the target core axial power offset (AO) does not exceed 5%. This not only meets the nuclear power plant operating specifications, but also reduces the time required to search for the rod position, improving the operator's decision-making efficiency.

[0017] The present invention first uses a model-driven approach, using core physics software to model the core and obtain relevant core parameters that can reflect the core state. The neural network model is then trained based on the core parameters and nuclear power plant operating parameters. Finally, the trained neural network model is used for data-driven, efficient and accurate prediction of the core axial power offset (AO), accurately determining the impact of current rod position adjustments on future AO. This model- and data-driven hybrid approach to intelligent peak-shaving in nuclear power plants can intelligently assist operators in AO control, reducing the risks of human decision-making and intervention, thereby achieving intelligent peak-shaving and ensuring core safety. Compared with existing technologies, the present invention has the following advantages: 1. The present invention adopts a dual-model and data-driven approach, innovatively adding the core parameters calculated by core physics software simulation to the data source driven by the neural network model. This solves the difficulties of conventional data-driven methods, such as a single data source, few types, and difficulty in accurately characterizing the core state.

[0018] 2. The present invention uses a dual-driven model and data approach to train a neural network model that can predict AO with high precision, and can predict AO efficiently and accurately.

[0019] 3. The intelligent peak-shaving method for nuclear power plants proposed in the present invention can intelligently assist operators in rod position adjustment and AO control, realize intelligent peak-shaving, reduce the risks of human decision-making and intervention during the peak-shaving process, and effectively ensure the safety of the core. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 Flow chart of the method of the present invention.

[0021] Figure 2 This is a schematic diagram of the changes in relative power and R rod position over time during a peak regulation period of a certain M310 unit in February 2023.

[0022] Figure 3 This is a schematic diagram comparing the predicted AO and actual AO during a peak regulation period of an M310 unit in February 2023.

[0023] Figure 4 This is a schematic diagram of the actual R rod position changing with time during a peak regulation period of a certain M310 unit in February 2023.

[0024] Figure 5 This is a schematic diagram of the change of target R rod position over time during a peak regulation period of a certain M310 unit in February 2023. DETAILED DESCRIPTION

[0025] In order to make the principles, methods and other features of the present invention easier to understand, the following specific implementation cases of the present invention are further described in detail with reference to the accompanying drawings. It must be emphasized that the specific implementation methods described here are only used to illustrate the present invention with specific cases, and do not limit the present invention.

[0026] like Figure 1 As shown, the present invention provides a nuclear power plant intelligent peak regulation method based on dual drive of model and data, the method comprising the following steps: Step 1: Select the geometric and material parameters of the nuclear power plant reactor core and fuel assemblies for core modeling. Use the core physics software Bamboo-C to perform simulation calculations to obtain the required but not directly measurable core parameters, including fuel burnup and neutron poison distribution. Step 2: Select the nuclear power plant operating data during historical peak regulation, such as the relative power of the reactor, the position of each control rod group, the core axial power offset AO measurement value of the detector outside the reactor, the enrichment of the fuel assembly and the amount of burnable poison, and the core parameters obtained by simulation calculation of the core physics software Bamboo-C to form the input data set required for neural network model training. The continuous data is sampled at a time interval of 4 minutes, and each peak regulation period is divided into a different number of samples; the core axial power offset AO value at the next moment is used as the label data set; the input data set and the label data set are preprocessed: all input data sets and label data sets are arranged in order and subjected to sliding window operation to process them into the corresponding input and output samples of the neural network model. The reconstructed data are subjected to "Min-Max normalization" and uniformly converted to the range between 0 and 1; Conventional data-driven approaches rely heavily on the data itself. While they can achieve certain results, particularly in the field of nuclear energy applications, they struggle to overcome drawbacks such as a single data source, insufficient data types, and low data quality. In the field of nuclear energy applications, the cost of acquiring core experimental data is very high. Due to limitations in detection methods and measurement technology, the types of measurement data collected during the operation of a single nuclear power plant are limited, making it difficult to accurately characterize the core state during rapid core changes such as peak-shaving operation. The present invention combines core data calculated using specialized core physics software with nuclear power plant operating data, which can largely resolve the data source issue and improve the accuracy and robustness of data-driven approaches such as neural network models.

[0027] Step 3: Divide the input dataset and labeled dataset preprocessed in Step 2 into a training set, a validation set, and a test set. The neural network model is trained using the training set, validation set, and test set. In this example, the neural network model uses the long short-term memory (LSTM) neural network model. Model training includes forward propagation, calculating the loss function, and back-calculating the gradient to reduce the loss function. The loss function uses the mean square error (MSE) loss function. An early stopping mechanism is used, terminating the training process if the validation set loss function stops decreasing within 50 epochs. Step 4: In actual peak regulation, the rod position at the next moment is set, and the core physics software is used to track and simulate the core to obtain the core parameters. The core parameters and the nuclear power plant operating data in actual peak regulation are then used as the input data to be predicted. The input data to be predicted is preprocessed using the same preprocessing method as in Step 2, and then input into the trained neural network model to predict the core axial power offset AO at the next moment in real time. The single prediction time does not exceed 1 second. Step 5: Determine whether the core axial power offset AO predicted in step 4 meets the requirement that the error with the target core axial power offset AO does not exceed 5%. If it does not meet the requirement, readjust the rod position and return to step 4 to re-predict the core axial power offset AO. If it meets the requirement, use the corresponding rod position as the target rod position at the next moment to assist the operator in decision-making and peak-shaving control, thereby realizing intelligent peak-shaving.

[0028] In this example, a peak-shaving period in February 2023 is taken as an example. The actual AO is used as the target AO. The rod position is adjusted under the premise that the error between the predicted AO and the target AO does not exceed 1% to obtain the required target rod position. The adjusted target R rod position is then compared with the actual R rod position to verify the feasibility of the intelligent peak-shaving method described in the present invention.

[0029] in Figure 2 This is a schematic diagram of the change of relative power and R rod position over time during a peak regulation period of a certain M310 unit in February 2023. Figure 3 This is a comparison diagram of the predicted AO and actual AO during the peak regulation process. Figure 2 and Figure 3 There are 4 data points between each two data points that are not shown, that is, the interval between two adjacent data points in the figure is 20 minutes. Figure 2 and Figure 3 It can be seen that during the peak regulation process, the relative power and the rod position are constantly changing, but the AO predicted value of the present invention is in good agreement with the AO measured value, with a small error. Figure 4 and Figure 5 These are respectively the schematic diagrams of actual R stick position variation over time and the schematic diagrams of target R stick position variation over time during a peak regulation period of a certain M310 unit in February 2023. Figure 4 and Figure 5 There are 4 data points between each two data points that are not shown, that is, the interval between two adjacent data points in the figure is 20 minutes. Figure 4 and Figure 5 It can be seen from the figure that the target R-rod position is in good agreement with the actual R-rod position, which proves that the intelligent peak-shaving method for nuclear power plants based on dual-drive of model and data in the present invention can efficiently and accurately predict the core axial power offset AO, and can intelligently assist operators in adjusting the rod position and controlling the core axial power offset AO, thereby reducing the risk of human decision-making and intervention, thereby realizing intelligent peak-shaving and ensuring core safety.

Claims

1. A model- and data-driven intelligent peak-shaving method for nuclear power plants, characterized by: The steps include: Step 1: Select the geometric and material parameters of the nuclear power plant reactor core and fuel assemblies for core modeling, and use core physics software to perform simulation calculations to obtain the required but not directly measurable core parameters; Step 2: Select the nuclear power plant operating data from historical peak-shaving and the core parameters calculated by core physics software simulation. The two together constitute the input data set required for neural network model training. The core axial power offset (AO) value at the next moment is used as the label data set. The input data set and the label data set are preprocessed. Step 3: Divide the input dataset and label dataset preprocessed in step 2 into a training set, a validation set, and a test set, and use the training set, validation set, and test set to train the neural network model to obtain a trained neural network model with the required accuracy. Step 4: In actual peak shaving, the rod position at the next moment is set. After tracking and simulating the core using core physics software to calculate the core parameters, the obtained core parameters and the actual nuclear power plant operating data during peak shaving are used as the input data to be predicted. The predicted input data is preprocessed and then input into the trained neural network model to predict the core axial power offset AO at the next moment in real time. Step 5: Determine whether the predicted core axial power offset AO meets the requirements. If not, readjust the rod position and return to step 4 to re-predict the core axial power offset AO. If it meets the requirements, use the corresponding rod position as the target rod position for the next moment to assist the operator in decision-making and peak-shaving control, thereby realizing intelligent peak-shaving.

2. The intelligent peak-shaving method for nuclear power plants based on dual model and data drive according to claim 1 is characterized by: The core parameters obtained by simulation calculation by the core physics software in step 1 include fuel burnup and neutron poison distribution.

3. The intelligent peak-shaving method for nuclear power plants based on dual model and data drive according to claim 1 is characterized in that: The nuclear power plant operating data in step 2 includes reactor power, the position of each control rod group, the core axial power offset (AO) measurement value of the external detector, the fuel assembly enrichment, and the amount of burnable poison.

4. The intelligent peak-shaving method for nuclear power plants based on dual model and data drive according to claim 1 is characterized in that: The time interval between adjacent times shall not exceed 60 minutes.

5. The intelligent peak-shaving method for nuclear power plants based on dual model and data drive according to claim 1 is characterized in that: The pretreatment in step 2 and step 4 includes: Dataset reconstruction: All input datasets and label datasets are reconstructed into the data input format required by the neural network model according to the preset order and structure; Dataset normalization: Normalize all reconstructed input datasets and label datasets to the range of 0 to 1 to reduce the impact of dimension and magnitude.

6. The intelligent peak-shaving method for nuclear power plants based on dual model and data drive according to claim 1 is characterized by: The neural network model described in step 3 adopts the long short-term memory neural network model LSTM.

7. The intelligent peak-shaving method for nuclear power plants based on dual model and data drive according to claim 1 is characterized in that: The neural network model is trained as described in step 3, including forward propagation, calculating the loss function, solving the gradient in reverse to reduce the loss function, and using the early stopping mechanism to interrupt the training process in advance when the conditions are met; The early stopping mechanism is: when the loss function of the validation set no longer decreases in a certain round, or when the decrease in a certain round is less than the preset value, the training process of the neural network model is interrupted.

8. The intelligent peak-shaving method for nuclear power plants based on dual model and data drive according to claim 7 is characterized in that: The loss function uses the mean square error loss function.

9. The intelligent peak-shaving method for nuclear power plants based on dual model and data drive according to claim 1, characterized in that: The requirement met in step 5 means that the error between the predicted core axial power offset AO and the target core axial power offset AO does not exceed 5%.

Images