Nuclear power plant optimal control method based on adaptive multi-scale and dual-domain view features

Abstract

A kind of nuclear power plant optimization control method based on adaptive multi-scale and dual-domain view feature, after data acquisition and preprocessing in offline stage to construct training set, neural network for training adaptive multi-scale embedding layer, dual-path sequence decomposition layer, dual-domain view feature learning layer and feature fusion layer, realize nuclear power plant operating parameter prediction in online stage through trained neural network.The present application realizes efficient and high-precision prediction of multi-parameter and multi-step by enhancing local information capture through multi-scale embedding, multi-kernel trend-seasonal decomposition modeling, variable and time series feature learning through dual-domain view, and error accumulation avoidance through direct prediction strategy, especially in the scene of nuclear power operation accident.

Description

Technical Field

[0001] This invention relates to a technology in the field of nuclear power plant control, specifically a nuclear power plant optimization control method based on adaptive multi-scale and dual-domain perspective features. Background Technology

[0002] Predicting operating parameters of nuclear power plants is one of the core functions of building a digital twin nuclear power plant system. To ensure the safety and reliability of nuclear power plant operation, it is necessary to establish a data-driven long-term trend prediction model for key operating parameters. This model can simultaneously predict the state of multiple key parameters over several consecutive time steps through a single input and calculation, thereby helping to identify system anomalies early and providing key decision-making basis for fault handling and operation optimization. However, existing prediction methods still lack systematic theoretical and technical support in areas such as historical data modeling of multi-parameter features, interaction and coordination mechanisms between features; especially in long-term prediction tasks, performance bottlenecks such as low computational efficiency and poor prediction accuracy are common. Summary of the Invention

[0003] This invention addresses the non-periodic, nonlinear, and strongly coupled characteristics of key operating parameters in existing nuclear power plants during accident transients, as well as the technical challenge of balancing error accumulation and computational efficiency in long-term prediction. It proposes an optimized control method for nuclear power plants based on adaptive multi-scale and dual-domain perspective features. This method enhances local information capture through multi-scale embedding, employs multi-core trend-seasonal decomposition modeling, learns variables and temporal features from a dual-domain perspective, and uses a direct prediction strategy to avoid error accumulation. This achieves highly efficient and accurate prediction across multiple parameters and steps, particularly in the context of nuclear power plant accident scenarios.

[0004] This invention is achieved through the following technical solution:

[0005] This invention provides a nuclear power plant optimization control method based on adaptive multi-scale and dual-domain perspective features. After collecting and preprocessing data in the offline stage to construct a training set, the training set is used to train a neural network with an adaptive multi-scale embedding layer, a dual-path sequence decomposition layer, a dual-domain perspective feature learning layer, and a feature fusion layer. In the online stage, the trained neural network is used to predict the operating parameters of the nuclear power plant.

[0006] The neural network specifically comprises: an adaptive multi-scale embedding layer, a dual-path sequence decomposition layer, a dual-domain perspective feature learning layer, a feature fusion layer, and an output layer connected in sequence. Specifically: the adaptive multi-scale embedding layer adaptively generates four scale windows based on the input multi-parameter time series data using a wavelet transform binary scale partitioning method, performing multi-scale feature extraction to obtain a multi-scale feature representation; the dual-path sequence decomposition layer performs trend-seasonal decomposition using a configurable kernel moving average filter based on the extracted multi-scale feature information, obtaining trend and seasonal components; the dual-domain perspective feature learning layer models the inter-variable dependencies and performs nonlinear dynamic capture of the time dimension for both the seasonal and trend components; the feature fusion layer performs residual fusion based on the output features of the two paths from the dual-domain perspective feature learning layer; and the output layer uses a generative decoder to perform a one-step direct mapping from the input sequence to the multi-step prediction result, synchronously outputting the predicted values ​​for all future time points through a single forward computation.

[0007] The adaptive multi-scale embedding layer specifically employs a binary scale partitioning method from wavelet transform to adaptively generate four scale windows, with the scale size determined by the formula... Where L is the length of the input sequence, thus enabling four-level feature extraction from macro trends to local fluctuations.

[0008] The dual-path sequence decomposition layer specifically employs a configurable kernel moving average filtering algorithm, extracts the trend component through one-dimensional average pooling, and uses the difference between the original sequence and the trend component as the seasonality component. The specific formula is as follows: , ,in: Refers to the original sequence. Refers to trend components, Refers to seasonal portion This refers to the size of the filter kernel that controls the smoothness of the trend. The trend component and the seasonal component are treated as two independent inputs for further processing, forming a dual-path system.

[0009] The dual-domain feature learning layer includes: two sets of variable self-attention layers and two sets of temporal multilayer perceptron layers. Specifically: the first variable self-attention layer models the inter-variable dependencies based on the seasonality component to obtain a variable domain feature representation; the first temporal multilayer perceptron layer performs non-linear dynamic capture of the time dimension based on the variable domain feature representation to obtain a time domain feature representation; the second variable self-attention layer models the inter-variable dependencies based on the trend component to obtain a variable domain feature representation; the second temporal multilayer perceptron layer performs non-linear dynamic capture of the time dimension based on the variable domain feature representation to obtain a time domain feature representation.

[0010] The feature fusion layer employs a sequential residual design to fuse features learned from both domains. Specifically: , ,in: Using time series as input offers significant advantages in maintaining training stability and model performance.

[0011] Technical effect

[0012] This invention constructs a long-term prediction model for nuclear power plant operating parameters based on adaptive multi-scale modeling and a dual-domain perspective feature fusion neural network. Through an adaptive multi-scale embedding layer, it can simultaneously capture features across multiple time scales, from short-term fluctuations to long-term trends, overcoming the limitations of traditional methods with their single scale. A dual-path sequence decomposition layer decomposes time-series data into trend and seasonal components for separate processing, improving the ability to model complex time-series patterns. The dual-domain perspective feature learning significantly enhances prediction performance when dealing with complex parameter coupling in nuclear energy systems. A direct multi-step prediction strategy avoids the error accumulation problem in traditional recursive prediction. This method is specifically designed for the non-periodic and nonlinear characteristics of transient processes during nuclear power plant accidents, exhibiting superior prediction performance in such scenarios. It achieves high-precision and high-efficiency prediction of long-term multi-parameter forecasts for nuclear power plants, significantly reducing prediction accuracy and computational speed compared to existing prediction models. Attached Figure Description

[0013] Figure 1 This is a flowchart of the present invention;

[0014] Figure 2 Flowchart of the method for constructing prediction samples;

[0015] Figure 3 Flowchart of the adaptive multi-scale embedding method;

[0016] Figure 4 Flowchart of the dual-domain feature extraction mechanism;

[0017] Figure 5 This is a schematic diagram illustrating the effect of an example. Detailed Implementation

[0018] like Figure 1 As shown in this embodiment, a nuclear power plant optimization control method based on adaptive multi-scale and dual-domain perspective features is proposed. After collecting and preprocessing data in the offline stage to construct a training set, a neural network containing an adaptive multi-scale embedding layer, a variable attention layer, and a temporal multilayer perceptron (MLP) layer is trained. In the online stage, the trained neural network is used to predict the operating parameters of the nuclear power plant. Specifically, this includes:

[0019] Step 1: Data Collection. Specifically, this embodiment selects the typical event of a nuclear power plant losing external power (LOOP) as the verification condition. Under this accident condition, the key operating parameters of the nuclear power plant exhibit strong nonlinear and variable coupling characteristics, which can fully verify the effectiveness of the present invention. Data was collected from a real nuclear power plant through the plant's online monitoring system. The sampling interval was set to 10 seconds / time step, and a total of 1600 continuous time step operating parameter records were obtained, fully covering the three stages of accident initiation, development, and stabilization.

[0020] Step 2, Feature Selection: From a nuclear safety design perspective, nuclear power plants typically configure a large number of redundant parameters to ensure the controllability of the system under various accident conditions. In actual monitoring, to optimize resource utilization efficiency, only key parameters are usually selected for focused monitoring. Based on an in-depth analysis of the above-mentioned accident mechanisms, this study selected seven safety-critical parameters with strong coupling relationships from hundreds of monitoring parameters to construct a prediction system, including: hot pipe section temperature (Th), cold pipe section temperature (Tc), primary coolant average temperature (Tav), pressurizer pressure (Ppzr), pressurizer water level (Lpzr), steam generator secondary side feedwater flow rate (WF), and reactor power (PR), which are used to construct multivariate time series data for nuclear power plants.

[0021] Step 3, data preprocessing, specifically: processing the collected multivariate time series data from nuclear power plants. The dataset is divided into training, validation, and test sets at a ratio of 60%, 30%, and 10%, respectively. For batch size, The length of the historical sequence. The number of variables is set; and max-min normalization is performed to eliminate the dimensional differences between different variables. The processed data will be used as the input to the neural network.

[0022] Step 4: Divide the data into prediction samples, such as... Figure 2 As shown, specifically, model input samples are generated using a sliding time window. Each sample consists of a lookback window and a forecasting window. In this embodiment, the lookback window length is set to 30 time steps (corresponding to 5 minutes of historical data), the forecasting window length is set to 20 time steps (corresponding to 200 seconds of future prediction), and the sliding step size is set to 20 time steps to ensure that the predicted samples do not overlap.

[0023] Step 5, Adaptive multi-scale embedding layer, such as Figure 3As shown, specifically, the input multivariate time series data is converted into a multi-scale feature representation, and different features of the sequence at different time scales are extracted. The binary scale division method in wavelet transform is used to adaptively generate four scale windows without requiring additional manual hyperparameter settings. This achieves four-level feature extraction from macro trends to local fluctuations. The specific calculation formula is as follows: ,in: For the time scale, The length of the input sequence. , , , Corresponding to four time scales—long-term, medium-term, short-term, and ultra-short-term—this multi-scale design ensures comprehensive feature capture from macro trends to micro fluctuations.

[0024] Step 6, Dual-Path Sequence Decomposition Mechanism: After extracting multi-scale features and embedding the vectors, a configurable kernel-based moving average filter is used to explicitly decouple the seasonal and trend components of the embedded feature representation, representing high-frequency fluctuations and low-frequency changes respectively, thus achieving explicit separation of different frequency patterns. Specifically, the sequence decomposition uses a moving average filtering algorithm, extracting the trend component through one-dimensional average pooling, and then using the difference between the original sequence and the trend component to extract the seasonal component. , .in: The original sequence, As a trend component, For seasonal components; Padding() is a one-dimensional symmetric padding function that pads both ends of the sequence with zero values ​​to maintain the same length for the input and output sequences; AvgPool() represents a one-dimensional average pooling operation; It serves as a filter kernel, used to control the smoothness of the trend.

[0025] Step 7: Dual-domain perspective feature learning, specifically: learning the seasonal component separately. With trend components Constructing a dual-domain feature learning approach, modeling sequence features from both the time domain and the variable domain, such as... Figure 4 As shown, it contains two core components: a variable-attention layer and a time-MLPLayer layer.

[0026] The variable attention layer is specifically a multi-head attention mechanism. Its core innovation lies in calculating attention weights on the variable dimension to explicitly model the dependency and synergy between different physical parameters. Multi-head self-attention is calculated on the variable dimension. The input of this layer is the seasonal or trend component from the sequence decomposition layer, specifically: (1) First, the input time series X is transformed into a query, key, and value matrix through linear projection. ,in: , , (2) The variable attention is then calculated on the i-th head as follows: ,in: , For model dimensions, The number of heads can be configured for multiple purposes. The core issue here is the attention score matrix. It independently calculates the dependencies between all time points in the sequence on each variable channel. This is fundamentally different from the traditional method of calculating attention in the time dimension. Its function is to explore the deep coupling relationship between different variables at the same time point, rather than the correlation between time points; (3) Finally, the outputs of each attention head are spliced ​​together and fused through linear projection: , To output the learnable weight matrix of the projection layer.

[0027] The aforementioned temporal multilayer perceptron layer functions to capture the nonlinear dynamic evolution of sequences from a temporal dimension, forming a complementary "dual-domain perspective" learning mechanism with the variable attention layer, which models from a variable dimension. The input to this layer is the output Y of the variable attention layer. It consists of a deep feedforward network containing N fully connected layers, and its ability to model complex nonlinear temporal dynamics is enhanced by increasing the network depth. Its first layer is used for input projection and nonlinear transformation. The k-th hidden layer is represented as: The output layer is used to project back to the model dimensions: Where: Y is the input tensor, and N is the total number of network layers. Let H be the Gaussian error linear unit activation function, and H be the output activation value of each layer. Here are the weight matrices for each layer. These are the bias vectors corresponding to each layer.

[0028] Step 8, Feature Fusion Layer: This layer fuses the output features of the two paths using a sequential residual design. The calculation logic is as follows: , ,in: Using time series as input, it offers significant advantages in maintaining training stability and model performance. Through a dual-domain perspective collaborative mechanism, it significantly enhances the ability to model complex interactions in multivariate time series prediction while maintaining computational efficiency, thus solving the representation bottleneck problem caused by spatiotemporal feature coupling in traditional methods.

[0029] Step 9: Project the prediction results. Specifically, a generative decoder is used to directly output the complete 20-step prediction results through a single forward propagation, avoiding the error accumulation problem of traditional recursive prediction.

[0030] The generative decoder, as described, is based on a fully connected network. Its function is to directly convert the integrated feature representation provided by the feature fusion layer into the final multi-step prediction result. Its input is the integrated feature tensor output by the feature fusion layer, and its output is a two-dimensional tensor with dimensions of [prediction step size × number of variables]. This design allows the model to transform a multi-step prediction problem into a single high-dimensional output regression task, thereby achieving explicit modeling of the dependencies between different time steps. It abandons the autoregressive connections required by traditional recursive prediction, and can synchronously generate predictions for all future time moments through a single forward computation, fundamentally eliminating the gradual accumulation of errors.

[0031] Step 10, Model Result Optimization and Evaluation: Specifically, the hyperparameters of the prediction model are optimized and trained. The target dataset is applied to the prediction model for training. The prediction accuracy and generalization ability are evaluated and tested on the test set, specifically by selecting the mean absolute error (MAE), mean square error (MSE), and root mean square error (RMSE).

[0032] The hyperparameter optimization refers to the automated search to determine the key structural parameters of the prediction model, thereby improving its accuracy and generalization ability in long-term multi-parameter prediction tasks for nuclear power plants. This method employs a Bayesian optimization algorithm as the hyperparameter search strategy. This strategy constructs a probabilistic surrogate model of the objective function to adaptively guide the search direction, thus efficiently approximating the global optimal solution of high-dimensional, nonlinear, and nonconvex problems. Specifically, this optimization process uses the mean absolute error (MAE) on the validation set as the core performance indicator and optimization objective, jointly searching for four key hyperparameters of the neural network: model dimension, number of hidden layers, number of attention heads, and batch size. After a preset 100 search iterations, the optimizer finally determined the optimal configuration of the prediction model as: model dimension 64, number of hidden layers 2, number of attention heads 8, and batch size 32. This set of parameters has been verified to enable the model to achieve the best overall prediction performance on the nuclear power plant time-series data.

[0033] The test set results of the evaluation model are used to assess the error between the predicted values ​​of the nuclear power plant prediction model and the actual values ​​of the target parameters, and to measure the prediction accuracy and generalization ability of the digital twin model. Evaluation indicators need to be developed to evaluate the results, specifically selecting Mean Absolute Error (MAE), Mean Square Error (MSE), and Root Mean Square Error (RMSE), where: , , ,in: The actual value of the parameter. This represents the predicted value output by the prediction model. The output is as follows: Figure 5 As shown in the figure, this is a comparison chart of the predicted values ​​and the actual values ​​of the prediction model on the dataset.

[0034] The aforementioned model comparison analysis refers to comparing the accuracy of this invention with several representative model building methods (Transformer, LSTM, GRU). At this point, the construction of the long-term prediction model for nuclear power plant operating parameters based on adaptive multi-scale embedding and multivariate attention is completed.

[0035] Through practical experiments, a neural network-based prediction model was built on the open-source platform PyTorch. Using CUDA acceleration, a model was developed to predict nuclear power plant power loss events. Seven key parameters (a total of 1600 time-point data points, sampled at 10-second intervals) were selected (hot pipe temperature, cold pipe temperature, primary coolant average temperature, pressurizer pressure, pressurizer water level, nuclear power, and secondary feedwater flow rate of the steam generator). A 20-step prediction was performed by inputting 30 time steps. The raw data was preprocessed: the training, validation, and test sets were divided into 60%, 10%, and 30% sets, respectively, and each variable was independently subjected to min-max normalization to eliminate dimensional differences. Subsequently, a prediction model was constructed and trained using the Adam optimizer. The hyperparameters of the neural network were searched using a Bayesian optimization algorithm. After 200 rounds of training, a comparison chart of the predicted and actual values ​​of seven key parameters—heat pipe section temperature, cold pipe section temperature, primary loop coolant average temperature, pressurizer pressure, pressurizer water level, nuclear power, and secondary side feedwater flow rate of the steam generator—is shown below. Figure 5 As shown in the figure, the blue line represents the model's predicted value, the yellow line represents the actual value, the red dashed vertical line is the dividing line between the training set and the validation set, and the green dashed vertical line is the dividing line between the validation set and the test set.

[0036] The results show that the prediction model achieves significant performance on the test set. The error was calculated using evaluation metrics, yielding an average MAE of 0.0667, an MSE of 0.0543, and an RMSE of 0.1949 for the seven parameters. The prediction task took 2.78 ms. These results demonstrate that the proposed method can effectively predict nuclear power plant operating parameters with high accuracy and efficiency.

[0037] This invention compares its accuracy with several representative model building methods (Transformer, LSTM, GRU) on the same dataset. The MAE of the advanced Transformer model (0.7417), the classic GRU model (1.1500), and the LSTM model (1.2837) are all significantly improved. Furthermore, the prediction task takes only 2.78ms, a 26% speedup compared to the Transformer's 3.77ms. Experiments show that this method successfully solves problems such as multivariate coupling and long-term prediction error accumulation. In particular, its prediction accuracy for critical safety parameters in the initial 200 seconds of an accident meets engineering requirements and has significant potential for further development.

[0038] Performance verification results: Under the same hardware configuration and prediction task (30 steps input, 20 steps prediction), this method shows significant advantages: (1) Prediction accuracy: The average prediction error of the seven key parameters of the nuclear power plant (heat pipe section temperature (Th), cold pipe section temperature (Tc), primary loop coolant average temperature (Tav), pressurizer pressure (Ppzr), pressurizer water level (Lpzr), steam generator secondary side feedwater flow rate (WF), and reactor power (PR)) is MAE=0.0667, MSE=0.0543, RMSE=0.1949; significantly improved compared with the comparison models: Transformer (MAE=0.7417), GRU (MAE=1.1500), LSTM (MAE=1.2837); (2) Computational efficiency: The single prediction time is only 2.78ms, which is about 26.0% faster than Transformer (3.77ms). Experiments show that this method achieves both an order-of-magnitude improvement in accuracy and optimization of inference efficiency in long-term multi-parameter prediction tasks.

[0039] Compared with existing technologies, this method fully explores the non-periodicity, high time dependence, and strong variable correlation of nuclear power time series data. It improves the model by enhancing local information capture through multi-scale embedding, modeling time series dynamics through multi-core trend-seasonal decomposition, achieving parameter coordination through inter-variable attention mechanisms, and avoiding error accumulation through direct prediction strategies. This achieves efficient and accurate prediction with multiple parameters and multiple steps, especially in the scenario of nuclear power operation accident. Evaluation indicators show that its results are superior to several representative model construction methods.

[0040] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A nuclear power plant optimization control method based on adaptive multi-scale and dual-domain perspective features, characterized in that, After collecting and preprocessing data in the offline stage to build a training set, a neural network with an adaptive multi-scale embedding layer, a dual-path sequence decomposition layer, a dual-domain viewpoint feature learning layer, and a feature fusion layer is trained. In the online stage, the trained neural network is used to predict the operating parameters of nuclear power plants. The neural network specifically comprises: an adaptive multi-scale embedding layer, a dual-path sequence decomposition layer, a dual-domain perspective feature learning layer, a feature fusion layer, and an output layer connected in sequence. Specifically: the adaptive multi-scale embedding layer adaptively generates four scale windows based on the input multi-parameter time series data using a wavelet transform binary scale partitioning method, performing multi-scale feature extraction to obtain a multi-scale feature representation; the dual-path sequence decomposition layer performs trend-seasonal decomposition using a configurable kernel moving average filter based on the extracted multi-scale feature information, obtaining trend and seasonal components; the dual-domain perspective feature learning layer models the inter-variable dependencies and performs nonlinear dynamic capture of the time dimension for both the seasonal and trend components; the feature fusion layer performs residual fusion based on the output features of the two paths from the dual-domain perspective feature learning layer; and the output layer uses a generative decoder to perform a one-step direct mapping from the input sequence to the multi-step prediction result, synchronously outputting the predicted values ​​for all future time points through a single forward computation. The adaptive multi-scale embedding layer specifically employs a binary scale partitioning method from wavelet transform to adaptively generate four scale windows, with the scale size determined by the formula... Where L is the length of the input sequence, thus achieving four-level feature extraction from macro trends to local fluctuations. These correspond to four time scales: long-term, medium-term, short-term, and ultra-short-term, with i=1,2,3,4.

2. The nuclear power plant optimization control method based on adaptive multi-scale and dual-domain perspective features according to claim 1, characterized in that, The dual-path sequence decomposition layer specifically employs a configurable kernel moving average filtering algorithm, extracts the trend component through one-dimensional average pooling, and uses the difference between the original sequence and the trend component as the seasonality component. The specific formula is as follows: , ,in: Refers to the original sequence. Refers to trend components, Refers to seasonal portion This refers to the size of the filter kernel that controls the smoothness of the trend. The trend component and the seasonal component are treated as two independent inputs for further processing, forming a dual path.

3. The nuclear power plant optimization control method based on adaptive multi-scale and dual-domain perspective features according to claim 1, characterized in that, The dual-domain feature learning layer includes: two sets of variable self-attention layers and two sets of temporal multilayer perceptron layers. Specifically: the first variable self-attention layer models the inter-variable dependencies based on the seasonality component to obtain a variable domain feature representation; the first temporal multilayer perceptron layer performs non-linear dynamic capture of the time dimension based on the variable domain feature representation to obtain a time domain feature representation; the second variable self-attention layer models the inter-variable dependencies based on the trend component to obtain a variable domain feature representation; the second temporal multilayer perceptron layer performs non-linear dynamic capture of the time dimension based on the variable domain feature representation to obtain a time domain feature representation. The dual-domain perspective refers to the variable domain and time domain of a multi-parameter time series.

Images