Real-time prediction system for atmospheric dispersion of radioactive substances based on deep neural networks

Abstract

The application discloses a radioactive material atmospheric diffusion real-time prediction system based on a deep neural network, and belongs to the technical field of nuclear safety emergency response. The system comprises a data acquisition module, a multi-scale space-time feature extraction network, a meteorological-terrain coupling analysis module, a physical constraint neural network layer and an adaptive prediction correction module. Fast and accurate prediction is realized through a multi-module deep coupling closed-loop collaborative mechanism. The multi-scale space-time feature extraction network simultaneously captures diffusion features of different spatial and temporal scales. The meteorological-terrain coupling analysis module adaptively fuses meteorological and terrain factors through an attention mechanism. The physical constraint neural network layer ensures that the prediction conforms to the convection-diffusion equation. The adaptive prediction correction module forms a closed-loop feedback through Kalman filtering. The system prediction speed is improved by more than 200 times, and the prediction accuracy is improved by 42% compared with traditional methods. The system provides key technical support for nuclear emergency decision-making.

Description

Technical Field

[0001] This invention belongs to the field of nuclear safety emergency response and atmospheric environment monitoring technology, specifically involving a real-time prediction system for atmospheric diffusion of radioactive materials based on deep neural networks, which is suitable for rapid decision support in nuclear emergency scenarios such as nuclear power plant accidents and radioactive material transportation accidents. Background Technology

[0002] Nuclear energy, as a clean energy source, occupies an important position in the global energy structure. However, once an accident occurs at a nuclear facility leading to the release of radioactive materials, it poses a serious threat to public health and environmental safety. The 2011 Fukushima nuclear accident in Japan demonstrated that rapidly and accurately predicting the atmospheric diffusion path and impact range of radioactive materials in the early stages of an accident is crucial for developing emergency response measures and organizing public evacuations.

[0003] Currently, the prediction of atmospheric diffusion of radioactive materials mainly relies on traditional numerical simulation methods. According to the technology disclosed in Chinese Patent Publication No. CN119692205A, most existing technologies use diffusion models based on numerical methods for solving the problem. This technology discloses a neural network-based method for solving neutron diffusion in hybrid energy spectrum reactors. It utilizes a deep learning model combining convolutional neural networks and long short-term memory networks to predict the coupling response data and neutron diffusion modes between the thermal and fast spectral regions by extracting spatiotemporal feature data. This method overcomes the limitations of traditional numerical simulation methods and possesses strong spatiotemporal feature capture capabilities.

[0004] However, the technology described in CN119692205A mainly addresses the neutron diffusion problem inside reactors, and its technical solution has the following shortcomings: First, this solution only focuses on the extraction of spatiotemporal features and does not consider the coupled influence of meteorological and topographical conditions on diffusion behavior during actual atmospheric diffusion, especially the interaction mechanism between wind field, temperature field, and topographic undulation; Second, the neural network model used in this solution does not incorporate the physical laws constraining atmospheric diffusion, and the purely data-driven approach is prone to producing prediction results that do not conform to physical laws when training samples are insufficient; Third, this solution lacks a real-time feedback correction mechanism and cannot dynamically optimize the prediction model using real-time monitoring data during the accident process, resulting in a decrease in prediction accuracy over time.

[0005] Furthermore, while traditional Lagrange particle diffusion models and Gaussian plume models can accurately describe the diffusion process of radioactive materials, they suffer from high computational complexity, stringent requirements for meteorological data, limited simulation accuracy under complex terrain conditions, and difficulty in achieving sub-second response times. Existing deep learning methods mostly employ single-scale feature extraction, failing to simultaneously capture both local diffusion details and large-scale transport characteristics of radioactive plumes, resulting in accumulated errors when predicting long-term evolution trends.

[0006] Therefore, there is an urgent need to develop a deep neural network prediction system that can integrate meteorological and topographical coupling effects, follow physical constraints, and have real-time feedback correction capabilities to meet the requirements of speed, accuracy, and robustness in nuclear emergency scenarios. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a real-time prediction system for atmospheric diffusion of radioactive materials based on deep neural networks. By constructing a deep coupled closed-loop collaborative system of four core modules—a multi-scale spatiotemporal feature extraction network, a meteorological-topographic coupling analysis module, a physically constrained neural network layer, and an adaptive prediction and correction module—it achieves rapid, accurate, and robust prediction of atmospheric diffusion of radioactive materials, providing key technical support for nuclear emergency decision-making.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is: a real-time prediction system for atmospheric diffusion of radioactive materials based on deep neural networks, including a data acquisition module, a multi-scale spatiotemporal feature extraction network, a meteorological-topographic coupling analysis module, a physical constraint neural network layer, and an adaptive prediction and correction module.

[0009] The data acquisition module acquires source data, meteorological data, and topographic data, providing foundational information for subsequent forecasts. A multi-scale spatiotemporal feature extraction network extracts spatiotemporal features at different scales from the input data through spatial multi-scale convolutional layers and temporal multi-scale recursive layers, capturing both local diffusion details and large-scale transmission patterns. The meteorological-topographic coupling analysis module adaptively fuses meteorological transmission characteristics and topographic influence factors through a coupling analysis network and attention mechanism, generating a feature representation that comprehensively considers the meteorological-topographic coupling effect. A physically constrained neural network layer applies physical constraints based on the convection-diffusion equation during the forecasting process, ensuring that the forecast results conform to the physical laws of atmospheric diffusion. The adaptive forecast correction module receives real-time monitoring data and dynamically adjusts network parameters by calculating forecast bias, forming a closed-loop system of forecasting-monitoring-feedback-correction.

[0010] A deep coupling relationship is formed among the modules: the output of the multi-scale spatiotemporal feature extraction network is directly used as the input of the meteorological-topographic coupling analysis module; the coupled feature tensor generated by the meteorological-topographic coupling analysis module is used by the physically constrained neural network layer for diffusion prediction; and the correction signal of the adaptive prediction correction module adjusts the weight parameters of the multi-scale spatiotemporal feature extraction network and the meteorological-topographic coupling analysis module inversely. This coupled closed-loop collaborative mechanism realizes the mutual promotion and superposition of feature extraction, coupling analysis, physical constraints, and real-time correction, resulting in a non-linear growth characteristic of 1+1>2 in the system's prediction accuracy and robustness.

[0011] Compared with the prior art, the present invention has the following beneficial effects:

[0012] First, multi-scale feature extraction enhances spatial coverage. This invention utilizes parallel processing of spatial multi-scale convolutional layers and temporal multi-scale recurrent layers to simultaneously extract diffusion features at different spatial scales, ranging from hundreds of meters to tens of kilometers, as well as evolution patterns at different temporal scales, from minutes to hours. Experiments show that, compared to single-scale CNN-LSTM networks, this invention reduces the average absolute error by 42% when predicting concentration distribution within a range of 10 to 50 kilometers from the release source, with a particularly significant improvement in prediction accuracy under complex terrain conditions.

[0013] Second, the meteorological-topographic coupling significantly improves prediction accuracy. This invention utilizes a meteorological-topographic coupling analysis module to deeply integrate meteorological elements such as wind field and temperature field with topographic features such as elevation and surface roughness, and adaptively adjusts the contribution weights of both through an attention mechanism. This coupling mechanism can accurately capture complex physical processes such as the guiding effect of topography on the wind field and the lifting effect of mountains on plumes. In test cases in mountainous and hilly terrain, compared to models that do not consider topographic factors, the prediction error of the concentration peak location of this invention is reduced by 65%, effectively solving the problem of inaccurate predictions by traditional models under complex terrain.

[0014] Third, physical constraints ensure the physical plausibility of the prediction results. This invention introduces a physical constraint loss function based on the convection-diffusion equation into the neural network, ensuring that the network adheres to the fundamental physical laws of atmospheric diffusion throughout the training and prediction processes. This constraint mechanism effectively suppresses overfitting and non-physical predictions that occur in purely data-driven methods when training samples are scarce. Even with scarce real nuclear accident data, this invention maintains over 95% prediction reliability, an improvement of 18 percentage points compared to purely data-driven models.

[0015] Fourth, closed-loop feedback enables dynamic adaptive optimization. The adaptive prediction and correction module of this invention uses a Kalman filter algorithm to evaluate prediction bias in real time and feeds the correction signal back to the preceding network to dynamically adjust feature extraction and coupling weights, forming a complete closed-loop system. This mechanism allows the system to continuously optimize the prediction model using monitoring data acquired during the accident. Experiments show that within 3 hours after the accident, through continuous feedback correction, the prediction accuracy is improved by 31% compared to the initial model, demonstrating good adaptability and robustness.

[0016] Fifth, the computational efficiency meets the real-time requirements of emergency decision-making. This invention replaces traditional numerical simulation with deep neural networks. After model training, a single prediction can complete the propagation prediction for the next 72 hours in just 2 to 5 seconds, a speed improvement of more than 200 times compared to the several hours of computation time required by traditional Lagrange particle propagation models. This second-level response capability provides crucial technical support for rapid decision-making in the initial stages of a nuclear emergency, significantly shortening emergency response time and reducing public exposure risks.

[0017] In summary, this invention, by constructing a multi-module deeply coupled closed-loop collaborative system, has achieved significant improvements in multiple dimensions such as prediction accuracy, physical rationality, adaptability, and computational efficiency, providing an innovative solution for predicting the atmospheric diffusion of radioactive materials, and has important application value and social significance. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention;

[0019] Figure 2 This is a schematic diagram of the structure of a multi-scale spatiotemporal feature extraction network;

[0020] Figure 3 This is a schematic diagram of the meteorological-terrain coupling analysis module;

[0021] Figure 4 This is a flowchart of the computation process for physically constrained neural network layers.

[0022] Figure 5 A schematic diagram of the feedback mechanism for the adaptive prediction and correction module;

[0023] Figure 6 Here is a flowchart of the system training process;

[0024] Figure 7 This is a flowchart of the system prediction process. Detailed Implementation

[0025] Please refer to the attached document. Figures 1-7 The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings to make the technical solution of the present invention clearer. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0026] like Figure 1 As shown, the real-time prediction system for atmospheric diffusion of radioactive materials based on deep neural networks provided by this invention includes a data acquisition module 1, a multi-scale spatiotemporal feature extraction network 2, a meteorological-topographic coupling analysis module 3, a physically constrained neural network layer 4, and an adaptive prediction and correction module 5. These modules form a complete closed-loop system through deep coupling at both the parameter and state levels: data input → multi-scale feature extraction → meteorological-topographic coupling → physically constrained prediction → real-time feedback correction.

[0027] The data acquisition module 1 is used to acquire all the input data required for the system to operate, including three main categories: source data, meteorological data, and terrain data.

[0028] Release source data describes the characteristics of radioactive material release, including the type of radionuclide, release rate, and release altitude. The type of radionuclide can be... I, Cs、 Sr and other single or mixed nuclides exhibit different half-lives and deposition characteristics. The release rate represents the activity of radioactive material released into the atmosphere per unit time, measured in Bq / s. Release height refers to the vertical distance from the release point to the ground, measured in meters (m), and this parameter significantly affects the initial dispersion pattern of the plume. In practical applications, release source data can be obtained through the safety monitoring system of the nuclear facility, or estimated using source term inversion algorithms when initial information is incomplete.

[0029] Meteorological data describes the dynamics and thermodynamics of the atmosphere and is a key factor determining the transport and diffusion of radioactive materials. The meteorological data collected in this invention includes wind speed, wind direction, temperature, humidity, and atmospheric stability parameters. Wind speed and direction determine the direction and speed of plume transport; data sources include ground meteorological stations, wind towers, and numerical weather prediction models, with a preferred spatial resolution of 1 km and a preferred temporal resolution of 10 minutes. Temperature and humidity affect atmospheric stratification stability and the wet deposition process of radioactive materials. Atmospheric stability parameters reflect the intensity of atmospheric turbulence and are commonly represented by the Pasquill stability scale, from Class A (extremely unstable) to Class F (extremely stable). This parameter can be calculated using empirical formulas based on solar radiation intensity, cloud cover, and surface wind speed.

[0030] Topographic data describes the surface morphology of the predicted area, including surface elevation and surface roughness. Surface elevation is obtained through a digital elevation model, with a preferred resolution of 30m, and is used to describe the impact of terrain undulations such as mountains and hills on the airflow field. Surface roughness characterizes the obstruction effect of the surface on airflow. Different land use types have different roughness values: approximately 1.0m to 2.0m in urban built-up areas, approximately 0.8m to 1.5m in forest areas, approximately 0.1m to 0.3m in farmland, and approximately 0.001m to 0.01m on water surfaces. Topographic data is typically derived from remote sensing imagery and geographic information system databases.

[0031] Data acquisition module 1 integrates the three types of data mentioned above into a standardized multidimensional tensor format, which serves as the input to the subsequent neural network. Specifically, the spatial grid is defined as... A three-dimensional mesh, in which , , These represent the number of grid cells in the x, y, and z directions, respectively, with the time dimension using... Each time step is represented. The dimension of the input data tensor is... ,in The number of channels includes multiple physical quantities such as wind speed u, wind speed v, wind speed w, temperature T, humidity H, elevation E, and roughness R.

[0032] In a preferred embodiment, the prediction region is set to be centered on the release source. The scope is defined by a spatial grid resolution of 1 km, 20 vertical layers up to 2000 m, a time step of 10 minutes, and a prediction duration of 72 hours. This configuration satisfies the spatial coverage and prediction duration requirements for emergency decision-making while ensuring computational efficiency.

[0033] like Figure 2 As shown, the multi-scale spatiotemporal feature extraction network 2 is one of the core innovative modules of this invention, used to extract multi-scale spatiotemporal features from the input data provided by the data acquisition module 1. This network consists of spatial multi-scale convolutional layers and temporal multi-scale recursive layers, and can simultaneously capture diffusion features at different spatial and temporal scales.

[0034] The spatial multi-scale convolutional layer group adopts a multi-branch parallel network structure, where each branch uses convolutional kernels of different sizes to perform convolution operations on the input data. In a preferred embodiment, three parallel branches are set, with convolutional kernel sizes of [sizes to be filled in]. , and .

[0035] The convolutional branch is used to extract local diffusion features. Due to the small receptive field of the convolutional kernel, this branch can capture local gradients and diffusion details between adjacent grid points, making it suitable for describing the fine structure of plume edges and concentration distribution near the source region. This branch contains four convolutional layers, each followed by batch normalization and ReLU activation functions, with 32, 64, 64, and 128 channels respectively.

[0036] The convolutional branch is used to extract regional diffusion features. A medium-sized receptive field can integrate spatial information within a certain range, making it suitable for describing mesoscale morphological changes and regional transport characteristics of plumes. This branch also contains four convolutional layers, with the number of channels... The branches are the same, namely 32, 64, 64, and 128.

[0037] The convolutional branch is used to extract large-scale transport features. Its large receptive field captures long-distance spatial correlations, making it suitable for describing long-distance transport and regional diffusion patterns of plumes under the influence of large-scale meteorological systems. The network configuration of this branch is consistent with other branches, with 32, 64, 64, and 128 channels.

[0038] The outputs of the three branches are merged through a concatenation operation to obtain a result with dimension . The spatial multi-scale feature map is represented by 384, which is the sum of the number of channels in the three branches. The fused feature map contains complete spatial information from local to global perspectives, providing rich feature representations for subsequent temporal evolution modeling.

[0039] During implementation, in order to maintain the consistency of spatial dimensions of the output feature maps of different branches, the following measures are taken: and Convolutions use padding. The convolution stride is uniformly set to 1, and pooling layers are not used to avoid loss of spatial resolution, which is crucial for accurately predicting the spatial details of concentration distribution.

[0040] The temporal multi-scale recursive layer group is used to extract temporal evolution patterns from spatial multi-scale feature maps. This layer group includes two parallel branches: short-time recursive units and long-time recursive units, which capture dynamic characteristics at different time scales, respectively.

[0041] The short-time recurrent unit employs a long short-term memory network to model short-term diffusion dynamics. The time span of this unit is set to 5 to 30 minutes, enabling it to capture the rapid evolution of the plume in the near term, the trajectory of the concentration peak, and the immediate impact of changes in local meteorological conditions on diffusion. The short-time LSTM consists of two layers, each with 256 hidden units. A dropout mechanism is used to prevent overfitting, with a dropout rate set to 0.2.

[0042] The long-term recurrent unit also employs a long short-term memory network, but with a time span set from 1 to 6 hours to model long-term evolutionary trends. This unit focuses on the large-scale transport patterns of plumes, diurnal variations, and the cumulative effects of persistent meteorological conditions. The long-term LSTM consists of two layers with 256 hidden units and a dropout rate of 0.2.

[0043] The outputs of the two recursive units are weighted and fused to obtain the final temporal multi-scale features. The fusion weights are adaptively determined through a gating mechanism. The gating network is a single-layer fully connected network that takes the concatenation of short-term and long-term features as input and outputs two weight coefficients. and ,satisfy and Features after fusion The calculation is as follows:

[0044] ,

[0045] in, The output characteristics of the short-time recursive unit, This represents the output feature of the long-term recursive unit. This gating fusion mechanism enables the network to automatically adjust the contribution ratio of short-term and long-term features based on the characteristics of the current prediction time.

[0046] The output of the multi-scale spatiotemporal feature extraction network 2 is a feature tensor that integrates spatial and temporal multi-scale information, with a dimension of . This feature tensor comprehensively describes the spatiotemporal evolution of radioactive material diffusion processes at different scales, providing a rich feature foundation for subsequent meteorological and topographic coupling analysis.

[0047] like Figure 3 As shown, the meteorological-topographic coupling analysis module 3 is the second core innovative module of this invention, used to deeply integrate meteorological transmission characteristics and topographic influence factors to generate a feature representation that comprehensively considers the meteorological-topographic coupling effect. This module achieves adaptive fusion of meteorology and topography through a coupling analysis network and an attention mechanism.

[0048] The coupled analysis network receives the feature tensor output by the multi-scale spatiotemporal feature extraction network 2 as input, and decomposes it into two branches: meteorological features and topographic features for processing.

[0049] The meteorological feature branch focuses on extracting features related to atmospheric dynamics and thermodynamic processes, including wind-driven advection transport, turbulent diffusion, and the influence of atmospheric stability on vertical diffusion. This branch employs a fully connected network structure with three layers and 512, 256, and 128 neurons respectively. The LeakyReLU activation function is used to avoid gradient vanishing. The output of the meteorological feature branch is represented as follows: .

[0050] The topographic feature branch focuses on extracting the guiding, blocking, and lifting effects of topography on airflow. In mountainous and hilly terrains, topographic forcing significantly alters the wind field structure, with local circulations such as valley winds and slope winds having a significant impact on plume transport paths. The topographic feature branch also employs a fully connected network with the same configuration as the meteorological branch, containing three layers with 512, 256, and 128 neurons respectively. The output is represented as... .

[0051] The outputs of the two branches are concatenated to obtain the initial coupling characteristics. This stitching operation preserved complete information on both meteorology and topography, providing a foundation for subsequent attention fusion.

[0052] The attention mechanism is used to adaptively adjust the weights of meteorological transmission characteristics and topographic influence factors in the coupled features, enabling the network to dynamically balance the importance of both based on current meteorological and topographic conditions.

[0053] The meteorological attention subnetwork determines the weighting coefficients of meteorological transport characteristics based on current wind speed and atmospheric stability. When wind speed is high, advection transport dominates, and the weight of meteorological transport characteristics should be increased; when the atmosphere is stable, vertical diffusion is suppressed, and the role of meteorological transport characteristics becomes more prominent. The formula for calculating meteorological attention is:

[0054] ,

[0055] in, For meteorological attention weight, It is the sigmoid activation function. This is the weight matrix. For bias vectors, The normalized wind speed scalar (normalized to the 0-1 interval after calculating the Euclidean distance between the horizontal wind speeds u and v). The stability parameter is normalized (by mapping the Pasquill stability level to a value between 0 and 1). In a preferred embodiment, The dimension is Output It is a single scalar.

[0056] The terrain attention sub-network determines the weight coefficients of terrain influence factors based on the degree of terrain undulation. When terrain undulation is severe, the terrain forcing effect is significant, and the weight of the terrain influence factor should be increased; under flat terrain conditions, the terrain influence is smaller, and the weight should be reduced accordingly. The degree of terrain undulation is quantified by calculating the standard deviation of surface elevation within the prediction area. The terrain attention calculation formula is:

[0057] ,

[0058] in, For terrain attention weights, This is the normalized value of the standard deviation of surface elevation. and These are the weight scalar and the bias scalar, respectively.

[0059] To ensure that the sum of the two attention weights is 1, for and Normalization is performed:

[0060] ,

[0061] in, and This represents the normalized attention weights.

[0062] Final Coupled Feature Tensor We obtain the result through weighted summation:

[0063] ,

[0064] This coupling feature tensor The study integrates meteorological transport characteristics and topographic influence factors, and achieves an adaptive balance between the two through an attention mechanism. This tensor, as the output of the meteorological-topographic coupling analysis module 3, is passed to the physically constrained neural network layer 4 for diffusion prediction.

[0065] In practical implementation, the parameters of the attention mechanism , , , The results were obtained through end-to-end training. Experiments show that when the wind speed is greater than 5 m / s, the meteorological attention weights... Typically between 0.6 and 0.8; when the standard deviation of terrain elevation is greater than 200m, the terrain attention weight... Typically between 0.4 and 0.6, this reflects the module's adaptive response capability to different conditions.

[0066] like Figure 4 As shown, the physical constraint neural network layer 4 is the third core innovative module of this invention. It is used to perform diffusion prediction based on coupled feature tensors and to apply physical constraints based on the convection-diffusion equation during the training and prediction process to ensure that the prediction results conform to the physical laws of atmospheric diffusion.

[0067] The physical constraint neural network layer 4 adopts a fully connected neural network structure and receives the coupled feature tensor output by the meteorological-terrain coupling analysis module 3. As input, the output is the predicted three-dimensional concentration field of radioactive materials.

[0068] The network consists of 5 fully connected layers with 256, 512, 1024, and 512 neurons respectively. The first four layers use the LeakyReLU activation function, and the last layer uses the ReLU activation function to ensure that the concentration predictions are non-negative. Batch normalization layers and Dropout layers are added between each layer, with the Dropout rate set to 0.3 to improve the model's generalization ability.

[0069] The network output is a three-dimensional concentration field. , indicating at time Predict the radioactive material concentration at each grid point within the region in real time, in units of Bq / m³. By iteratively predicting at multiple time steps, a complete spatiotemporal evolution sequence can be obtained. ,in This is the predicted time series.

[0070] To ensure that the prediction results conform to the physical laws of atmospheric diffusion, this invention introduces a physical constraint loss function based on the convection-diffusion equation. The diffusion process of radioactive materials in the atmosphere follows the convection-diffusion equation:

[0071] ,

[0072] in, Concentration of radioactive materials For time, , , For spatial coordinates, , , These are the three components of wind speed. , , These are the three components of the diffusion coefficient. It is the radioactive decay constant. For source terms.

[0073] The first term on the left side of the equation is the rate of change of concentration over time, the second to fourth terms are advection terms representing transport caused by the wind field, the first three terms on the right side are turbulent diffusion terms, the second to last term is a radioactive decay term, and the last term is a source term.

[0074] The physical constraint loss function calculates the partial derivatives of the predicted concentration field in space and time, and substitutes them into the aforementioned convection-diffusion equation to calculate the equation residuals. Specifically, for the predicted concentration field... The time and space partial derivatives are calculated using the finite difference method:

[0075] The time partial derivatives are obtained using forward differencing:

[0076] ,

[0077] Spatial partial derivatives are obtained using central difference:

[0078] ,

[0079] in, For time step, This represents the spatial grid spacing. (For...) and The partial derivatives in each direction are calculated in the same way.

[0080] The second-order spatial partial derivative is obtained using the three-point central difference:

[0081] ,

[0082] Substituting the calculated partial derivatives into the convection-diffusion equations, we obtain the physical equation residuals. :

[0083] ,

[0084] in, This represents the concentration field predicted by the neural network. Ideally, if the prediction perfectly conforms to the laws of physics, the residual... It should be zero.

[0085] Physical constraint loss term The mean square error of the residuals is defined as follows:

[0086] ,

[0087] in, Indicates at time Time grid points The physical equation residuals at that location.

[0088] The total loss function is a weighted combination of the data fitting loss term and the physical constraint loss term:

[0089] ,

[0090] in, For data fitting loss, mean squared error is used to measure the difference between predicted and true concentrations:

[0091] ,

[0092] in, To predict concentration, This represents the actual concentration in the training data. The weighting coefficient for the physical constraint loss is used to balance the two objectives of data fitting and physical constraints, and is set to 0.1 in the preferred embodiment.

[0093] By introducing a physically constrained loss function, the neural network not only learns to fit the training data during training but is also constrained to follow the physical laws of the convection-diffusion equation. This design philosophy of physically-informed neural networks enables the model to produce physically plausible predictions even with limited training samples, significantly improving the model's generalization ability and reliability.

[0094] In practical implementation, the diffusion coefficient , , The horizontal diffusion coefficient is calculated using empirical formulas based on atmospheric stability and wind speed.

[0095] ,

[0096] in, Friction speed, in m / s. Vertical diffusion coefficient, measured in meters (m). Based on atmospheric stability classification, the value range under unstable conditions is 10m. / s to 100m / s, with a range of 1m under stable conditions. / s to 10m / s. Radioactive decay constant. Calculated based on the half-life of the specific nuclide, for Cs, s .

[0097] like Figure 5 As shown, the adaptive prediction and correction module 5 is the fourth core innovative module of this invention. It is used to receive real-time monitoring data and dynamically adjust system parameters based on the deviation between the predicted value and the actual monitored value, forming a closed-loop system of prediction-monitoring-feedback-correction.

[0098] The adaptive prediction and correction module 5 continuously receives real-time concentration monitoring data from the monitoring stations. ,in For the first The spatial coordinates of each monitoring station. The monitored values ​​are compared with the predicted values ​​output by the fourth layer of the physical constraint neural network. Compare and calculate the prediction bias:

[0099] ,

[0100] in, For the first Each monitoring station at time The prediction bias. The biases of all monitoring stations are statistically analyzed, and the root mean square error is calculated:

[0101] ,

[0102] in, This represents the number of monitoring sites. The RMSE value reflects the overall error level of the current prediction model.

[0103] The adaptive prediction and correction module 5 uses the extended Kalman filter algorithm to dynamically adjust the system parameters. The feature extraction weights of the multi-scale spatiotemporal feature extraction network 2 and the coupling weights of the meteorological-topographic coupling analysis module 3 are considered as the system state vector. The monitored concentration values ​​are considered as observation vectors. .

[0104] State vector It includes two parts of weight parameters: gated fusion weights in the multi-scale feature extraction network. and And attention weights in the meteorological and topographic coupling module and .therefore, .

[0105] The prediction step of the Kalman filter is as follows:

[0106] ,

[0107] ,

[0108] in, For the first Prior state estimation at time 10:00 For the first Posterior state estimation at time 10:00. Here is the state transition matrix. Let be the prior error covariance matrix. Let be the posterior error covariance matrix. Let be the process noise covariance matrix. In a preferred embodiment, it is assumed that the state parameters change slowly. Set as identity matrix , Set as a diagonal matrix, with diagonal elements as .

[0109] The observation model describes the relationship between the state vector and the observed concentration:

[0110] ,

[0111] in, For the observation vector, The observation matrix describes how the state vector affects the predicted concentration. For observation noise, the covariance matrix is: Observation matrix The Jacobian matrix is ​​obtained by taking the partial derivative of the state vector with respect to the neural network.

[0112] ,

[0113] in, This represents the predicted concentration vector for each monitoring station. The Jacobian matrix can be calculated using automatic differentiation techniques.

[0114] The update step of the Kalman filter is as follows:

[0115] ,

[0116] ,

[0117] ,

[0118] in, Here, represents the Kalman gain, indicating the degree to which the observed data corrects the state estimate. The updated state estimate. This refers to the corrected weight parameters.

[0119] Corrected weight parameters obtained through Kalman filtering The data is fed back to the multi-scale spatiotemporal feature extraction network 2 and the meteorological-topographic coupling analysis module 3, dynamically updating the corresponding parameters in the network. Specifically, the data will be fed back to the multi-scale spatiotemporal feature extraction network 2 and the meteorological-topographic coupling analysis module 3, dynamically updating the corresponding parameters in the network. and Update the gated fusion layer of the multi-scale feature extraction network to and Update the attention mechanism to the meteorological and topographical coupling module.

[0120] Parameter updates employ an exponential moving average strategy to avoid drastic fluctuations.

[0121] ,

[0122] in, For old parameters, New parameters estimated by Kalman filtering, For the final updated parameters, The smoothing coefficient is set to 0.9 in the preferred embodiment.

[0123] Through the aforementioned feedback correction mechanism, the system can continuously optimize model parameters using real-time monitoring data, gradually bringing the prediction results closer to the true values. This closed-loop design significantly improves the system's adaptability and prediction accuracy, especially in situations where meteorological conditions change rapidly or the initial model parameters are inaccurate, enabling it to quickly converge to the optimal state.

[0124] like Figure 6 As shown, the system adopts a two-stage training strategy. The first stage uses simulated data for pre-training, and the second stage uses real data for fine-tuning.

[0125] Phase 1: Simulated Data Pre-training

[0126] Due to the scarcity of real nuclear accident data, this invention first utilizes a traditional atmospheric diffusion model to generate a large amount of simulation training data. The training data generation module simulates the diffusion process of radioactive materials under different release conditions, meteorological conditions, and topographical conditions using a Lagrange particle diffusion model.

[0127] Specifically, the release rate range is set as follows: Bq / s to The concentration field was measured in Bq / s, with release heights ranging from 0m to 200m, wind speeds ranging from 0.5m / s to 15m / s, atmospheric stability categories A to F (all six categories), and terrain types including plains, hills, and mountains. Five thousand different scenarios were generated in the parameter space using the Latin hypercube sampling method. For each scenario, a Lagrange particle diffusion model was run to obtain a 72-hour spatiotemporal evolution sequence of the concentration field.

[0128] The generated training dataset contains 5000 samples, each including an input data tensor and a corresponding concentration field sequence. The input data tensor is in the following format: The concentration field sequence is in the following format: The dataset is divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0129] In the first phase of training, the entire neural network is trained end-to-end using the training set data. The optimization algorithm used is Adam, and the initial learning rate is set to... The batch size is 16. The loss function is a weighted sum of the data fitting loss and the physical constraint loss, with weighting coefficients... Set to 0.1. Training is performed for 200 epochs, during which the loss on the validation set is monitored, and an early stopping strategy is used to prevent overfitting. When the validation set loss does not decrease for 20 consecutive epochs, training is stopped and the model parameters with the minimum validation set loss are saved.

[0130] The first stage of pre-training enables the neural network to learn the basic laws of atmospheric diffusion and diffusion patterns under different conditions, laying the foundation for the second stage of fine-tuning.

[0131] Phase Two: Fine-tuning with Real Data

[0132] The second phase involved fine-tuning the pre-trained model using monitoring data from a small number of real nuclear accidents or radioactive release drills. The real data sources included atmospheric monitoring station data after the Fukushima nuclear accident, historical monitoring data from the Chernobyl accident, and monitoring data from regularly conducted nuclear emergency drills.

[0133] Because the amount of real-world data is limited, typically only tens to hundreds of sets, fine-tuning employs a transfer learning strategy. The parameters of the first few layers of the neural network are frozen, and only the parameters of the later layers and the output layer are fine-tuned to avoid overfitting. The learning rate is reduced to... The training process lasted 50 epochs. The fine-tuned model was able to better adapt to the characteristics of real-world scenarios, improving prediction accuracy.

[0134] By employing a two-stage training strategy, this invention effectively addresses the problem of scarce real training samples. It utilizes simulated data to learn general diffusion patterns while simultaneously fine-tuning with real data to improve the accuracy of practical applications.

[0135] like Figure 7 As shown, the system's real-time prediction process after a nuclear accident is as follows:

[0136] Step S1: Data acquisition module 1 acquires current release source data, meteorological data, and topographic data. Release source data includes known or estimated types of radionuclides, release rates, and release altitudes. Meteorological data is acquired in real-time from numerical weather prediction systems or meteorological observation networks, including wind speed, wind direction, temperature, humidity, and atmospheric stability. Topographic data is retrieved from a geographic information database.

[0137] Step S2: Input the acquired data into the multi-scale spatiotemporal feature extraction network 2, and extract multi-scale spatiotemporal features through a group of spatial multi-scale convolutional layers and a group of temporal multi-scale recursive layers. The computation time for this step is approximately 0.5 to 1 second.

[0138] Step S3: Input the multi-scale spatiotemporal features into the meteorological-topographic coupling analysis module 3, and generate a coupled feature tensor through the coupling analysis network and attention mechanism. The attention mechanism adaptively adjusts the weights of meteorological and topographic features based on the current wind speed, stability, and topographic relief. The computation time for this step is approximately 0.2 to 0.5 seconds.

[0139] Step S4: Input the coupled feature tensor into the physics-constrained neural network layer 4, which outputs the predicted 3D concentration field for the next 72 hours. The physics-constrained loss function ensures that the prediction results conform to the convection-diffusion equation. The computation time for this step is approximately 1 to 2 seconds.

[0140] Step S5: Compare the prediction results with the real-time monitoring station data. The adaptive prediction correction module 5 calculates the prediction deviation and generates the parameter correction amount using the Kalman filter algorithm. The calculation time for this step is approximately 0.3 to 0.5 seconds.

[0141] Step S6: Feed the correction values ​​back to the multi-scale spatiotemporal feature extraction network 2 and the meteorological-terrain coupling analysis module 3 to dynamically update the network parameters. This step is an instantaneous operation and its time consumption is negligible.

[0142] Step S7: Re-predict using the updated network, outputting the corrected concentration distribution, surface deposition distribution, and effective dose distribution. This step repeats steps S2 to S4, with a computation time of approximately 2 to 3 seconds.

[0143] Step S8: Display the prediction results graphically in the emergency decision support system, including concentration contour maps, dose distribution cloud maps, and suggested evacuation areas. Simultaneously output key statistical information, such as maximum concentration value, maximum dose value, and affected area.

[0144] The total computation time for the above prediction process is approximately 2 to 5 seconds, meeting the real-time requirements of nuclear emergency scenarios. The system can be set to automatically update the prediction results every 10 minutes, and the prediction accuracy continues to improve as monitoring data is continuously acquired.

[0145] To verify the performance of the system of the present invention, multiple test cases were used for evaluation, including simulated cases and real cases.

[0146] In simulated case testing, system performance was evaluated using 1000 test samples independent of the training set. The test samples covered different release conditions, weather conditions, and terrain conditions. Evaluation metrics included root mean square error, correlation coefficient, and hit rate.

[0147] Test results show that the root mean square error of the system in this invention is 0.18 Bq / m in plain terrain. The correlation coefficient was 0.96, and the hit rate was 92%. The root mean square error of the prediction in hilly terrain was 0.25 Bq / m. The correlation coefficient was 0.93, and the hit rate was 88%. The root mean square error for prediction in mountainous terrain was 0.32 Bq / m. The correlation coefficient was 0.91, and the hit rate was 85%. The overall prediction accuracy decreased slightly with the increase of terrain complexity, but remained at a high level.

[0148] Comparative experiments show that, compared to a single-scale CNN-LSTM model that does not use multi-scale feature extraction, the root mean square error of the system proposed in this invention is reduced by 42%, the correlation coefficient is increased by 6 percentage points, and the hit rate is increased by 12 percentage points. Compared to models that do not consider terrain factors, the prediction error of the concentration peak location in mountainous terrain is reduced by 65%, significantly improving the prediction ability under complex terrain.

[0149] In a real-world test, monitoring data from the 2011 Fukushima nuclear accident was used for retrospective verification. Monitoring data from the three days prior to the accident was used as input to predict the concentration distribution over the following 72 hours, and then compared with the actual monitoring values.

[0150] Test results show that the root mean square error of the system of this invention in predicting the spread process of the Fukushima nuclear accident is 1.2 Bq / m. The correlation coefficient was 0.89. Although the prediction accuracy of real-world cases was slightly lower than that of simulated cases, considering the complexity of real-world scenarios and the uncertainty of monitoring data, this accuracy is sufficient to meet the needs of emergency decision-making.

[0151] Through continuous optimization of the adaptive prediction and correction module, the prediction error was reduced from the initial 2.5 Bq / m within 6 hours after the accident. Reduced to 1.2 Bq / m The decrease reached 52%, verifying the effectiveness of the closed-loop feedback mechanism.

[0152] In terms of computational efficiency, the system of this invention takes approximately 3 seconds to predict the diffusion process over the next 72 hours on a workstation equipped with an NVIDIA RTX 3090 GPU. In comparison, the traditional Lagrange particle diffusion model requires approximately 10 hours to complete the same prediction under the same hardware conditions, representing a computational speed improvement of more than 200 times.

[0153] In the emergency response drill, the system took no more than 5 seconds from receiving the release source information to outputting the preliminary prediction results, meeting the requirement of the golden 15-minute decision-making window for nuclear emergencies and providing key technical support for the timely activation of emergency response measures.

[0154] In summary, the real-time prediction system for atmospheric diffusion of radioactive materials based on deep neural networks provided by this invention has achieved significant improvements in multiple dimensions such as prediction accuracy, physical rationality, adaptability, and computational efficiency by constructing a deep coupled closed-loop collaborative system of four core modules: a multi-scale spatiotemporal feature extraction network, a meteorological-topographic coupling analysis module, a physical constraint neural network layer, and an adaptive prediction and correction module. This provides strong technical support for nuclear emergency decision-making and has important practical value and social significance.

[0155] Those skilled in the art should understand that the above embodiments are merely illustrative of the technical solutions of the present invention, and not as limitations on the present invention. Any modifications, equivalent substitutions, and improvements made to the technical solutions of the present invention without departing from the spirit and scope of the present invention should be included within the protection scope of the present invention.

Claims

1. A real-time prediction system for atmospheric diffusion of radioactive materials based on deep neural networks, characterized in that, include: The data acquisition module is used to acquire release source data, meteorological data, and topographic data. The release source data includes the type of radionuclide, release rate, and release height. The meteorological data includes wind speed, wind direction, temperature, humidity, and atmospheric stability parameters. The topographic data includes surface elevation and surface roughness. A multi-scale spatiotemporal feature extraction network, connected to the data acquisition module, is used to extract multi-scale spatiotemporal features from the released source data, meteorological data, and topographic data. The multi-scale spatiotemporal feature extraction network includes a spatial multi-scale convolutional layer group and a temporal multi-scale recursive layer group. The spatial multi-scale convolutional layer group uses convolutional kernels of different scales to process in parallel to extract features at different spatial scales. The temporal multi-scale recursive layer group uses recursive units with different time spans to capture the evolutionary patterns at different time scales. The meteorological-topography coupling analysis module is connected to the multi-scale spatiotemporal feature extraction network. It is used to generate a coupling feature tensor by fusing meteorological transmission characteristics and topographic influence factors through the coupling analysis network based on the multi-scale spatiotemporal features. The coupling analysis network adaptively adjusts the weights of meteorological transmission characteristics and topographic influence factors through an attention mechanism. The physical constraint neural network layer, connected to the meteorological-topographic coupling analysis module, is used to perform diffusion prediction based on the coupling feature tensor. During training and prediction, physical constraints based on the convection-diffusion equation are applied, and the prediction results are ensured to conform to the physical laws of atmospheric diffusion through the physical constraint loss function. An adaptive prediction and correction module, connected to the physical constraint neural network layer and the data acquisition module, is used to receive real-time monitoring data and dynamically adjust the feature extraction weights of the multi-scale spatiotemporal feature extraction network and the coupling weights of the meteorological-terrain coupling analysis module based on the deviation between the predicted value and the actual monitoring value.

2. The system according to claim 1, characterized in that, The spatial multi-scale convolutional layer group includes at least three parallel convolutional branches, each using a convolutional kernel of different sizes, namely 3×3, 5×5 and 7×7, which are used to extract local diffusion features, regional diffusion features and large-scale transmission features, respectively.

3. The system according to claim 1, characterized in that, The time-scale recursive layer group includes short-time recursive units and long-time recursive units. The time span of the short-time recursive units is 5 to 30 minutes, which is used to capture short-term diffusion dynamics. The time span of the long-time recursive units is 1 to 6 hours, which is used to capture long-term evolutionary trends.

4. The system according to claim 1, characterized in that, The attention mechanism in the meteorological-terrain coupled analysis module includes a meteorological attention subnetwork and a terrain attention subnetwork. The meteorological attention subnetwork determines the weight coefficients of meteorological transmission characteristics based on the current wind speed and atmospheric stability, while the terrain attention subnetwork determines the weight coefficients of terrain influence factors based on the degree of terrain undulation.

5. The system according to claim 1, characterized in that, The physical constraint loss function includes a data fitting loss term and a physical constraint loss term, which is constructed by calculating the partial derivatives of the predicted concentration field in space and time with the mean square error of the convection-diffusion equation residuals.

6. The system according to claim 1, characterized in that, The adaptive prediction correction module uses the Kalman filter algorithm to calculate the covariance matrix of the prediction bias, and determines the adjustment amount of the feature extraction weights and coupling weights based on the covariance matrix.

7. The system according to claim 1, characterized in that, The system also includes a training data generation module, which generates a training dataset using a Lagrange particle diffusion model. The training dataset includes spatiotemporal sequences of radioactive material concentration distribution under various meteorological, topographical, and release conditions.

8. The system according to claim 7, characterized in that, The system employs a two-stage training strategy. In the first stage, it uses simulated data generated by the training data generation module for pre-training, and in the second stage, it uses real monitoring data from historical nuclear accident events for fine-tuning.

9. The system according to claim 1, characterized in that, The system also includes an uncertainty quantification module, which is used to assess the uncertainty of the prediction results using the Monte Carlo Dropout method and output the confidence interval of the predicted concentration.

10. The system according to claim 1, characterized in that, The output of the physical constraint neural network layer includes the concentration distribution of radioactive material at three-dimensional spatial grid points, the surface deposition distribution, and the effective dose distribution, with a prediction time span of 1 hour to 72 hours after release.

Images