CN-122020024-A - Nuclear accident radiation field feedback correction model based on self-adaptive depth assimilation network

CN122020024ACN 122020024 ACN122020024 ACN 122020024ACN-122020024-A

Abstract

The invention discloses a nuclear accident radiation field feedback correction model based on a self-adaptive depth assimilation network, which comprises a data receiving module, an initial radiation field prediction module, an ADAN module, an intelligent optimization and feedback correction module and a real-time updating and visualization display module, wherein the data receiving module, the initial radiation field prediction module, the intelligent optimization and feedback correction module and the real-time updating and visualization display module are all connected with the ADAN module. The model utilizes a self-adaptive depth assimilation network to fuse and feedback and correct the real-time gamma radiation monitoring data and the data of the nuclear accident radiation field prediction model, thereby realizing high-precision dynamic prediction of radiation field distribution. The method comprises the steps of carrying out a depth assimilation process, wherein an adaptive parameter optimization and prediction result feedback correction mechanism is introduced in the depth assimilation process, so that a model can dynamically adjust assimilation network parameters according to real-time monitoring data, and carrying out iterative correction on the prediction result of an initial physical model chain, thereby forming a closed loop updating mechanism of radiation field prediction, assimilation correction and feedback optimization. The method is suitable for the fields of nuclear emergency management, nuclear accident result evaluation, radiation diffusion simulation and the like, and has the remarkable effect of improving the accuracy and timeliness of emergency response.

Inventors

LIAN BING
TIAN ZHIJIE
HUANG SHA
ZHAO DAN
QIU ZHIXIN
YANG BIAO
ZHAO DUOXIN
LV MINGHUA
LI MINGYE
ZHANG JUNFANG
GUO HUAN
Lv Xindong
PAN YANHUI
NIU YANJING
SONG YU

Assignees

中国辐射防护研究院

Dates

Publication Date: 20260512
Application Date: 20251231

Claims (10)

1. The nuclear accident radiation field feedback correction model based on the self-adaptive depth assimilation network is characterized by comprising a data receiving module, an initial radiation field prediction module, an ADAN module, an intelligent optimization and feedback correction module and a real-time updating and visualization display module, wherein the data receiving module, the initial radiation field prediction module, the intelligent optimization and feedback correction module and the real-time updating and visualization display module are all connected with the ADAN module; The data receiving module acquires real-time monitoring data of the radiation field of each monitoring point through a field monitoring network, and transmits the real-time monitoring data to the ADAN module through high-speed communication after preprocessing; The initial radiation field prediction module uses initial environmental parameters and source item information of the nuclear accident to call a nuclear accident physical model chain to generate initial radiation field distribution, and the initial radiation field distribution is used as a background field to be input into the ADAN module; The ADAN module receives the real-time monitoring data from the data receiving module and the initial radiation field distribution data from the initial radiation field prediction module, fuses the two types of data through deep assimilation to extract spatial characteristics and learn time sequence variation trend of the data, and generates corrected radiation field distribution data; The intelligent optimization and feedback correction module analyzes the evolution characteristics of prediction errors based on assimilation residual error change conditions between radiation field prediction results and real-time monitoring data output by the ADAN module at a plurality of continuous moments, and accordingly performs on-line self-adaptive adjustment on key parameters of the ADAN module, so that model parameter configuration can be dynamically updated along with accident environment changes, and the stability and the accuracy of radiation field prediction are improved; And the real-time updating and visual display module receives the latest radiation field distribution data from the ADAN module and immediately updates the visual content of the emergency command interface.
2. The nuclear accident radiation field feedback correction model based on the adaptive deep assimilation network of claim 1, wherein the assimilation process of the ADAN module is expressed as the following functional form: Wherein, the Indicating the initial radiation field distribution and, Real-time monitoring data distribution representing the radiation field of the monitoring point, A set of model parameters representing an ADAN network, A corrected radiation field distribution map representing an assimilation output, the formula representing a field distribution after assimilation output correction by effectively fusing an initial prediction result with observed data through a network f The function form is used for describing the calculation process of nonlinear fusion of the prediction radiation field and the monitoring radiation data in engineering realization of the ADAN module, and the output result is directly used for updating the radiation field situation display and the subsequent decision module in the nuclear accident emergency system.
3. The nuclear accident radiation field feedback correction model based on the self-adaptive depth assimilation network of claim 2, wherein the ADAN module comprises a CNN layer and an RNN layer, the CNN layer performs spatial feature extraction on input radiation field grid data, identifies correlation among monitoring points, captures a spatial mode of monitoring point distribution and model deviation, the RNN layer learns time sequence variation trend of data, combines current time and past time sequence data, refines information of dynamic evolution of a radiation field, and outputs corrected radiation field distribution which is closer to real conditions.
4. The nuclear accident radiation field feedback correction model based on the adaptive depth assimilation network of claim 3, wherein said CNN layer is composed of a plurality of layers of convolution units, each layer is followed by a ReLU activation function and a pooling layer to compress feature dimensions, said convolution units are capable of generating feature maps comprising spatial relationships of monitoring points.
5. The nuclear accident radiation field feedback correction model based on the adaptive depth assimilation network of claim 4, wherein the multi-layer convolution unit comprises two-dimensional convolution layers, the convolution kernel size of the first two-dimensional convolution layer is 3×3 for extracting local spatial features, and the convolution kernel size of the second two-dimensional convolution layer is 5×5 for extracting wider spatial correlations.
6. The nuclear accident radiation field feedback correction model based on the adaptive depth assimilation network of claim 5, wherein the spatial features are extracted from the input by two-dimensional convolution operation, and the two-dimensional convolution operation formula is: Wherein, the Representing the input feature map at the C-th channel and position The value at which the value is to be calculated, Representing the corresponding weight parameter values of the convolution kernel at the kth output channel, input channel C, relative positions u, v, Representing the offset term corresponding to the convolution kernel for the kth output channel, Representing the values of the output profile obtained by the convolution operation at the kth channel, positions i, j.
7. The nuclear accident radiation field feedback correction model based on the adaptive deep assimilation network of claim 3, wherein the RNN layer captures time series characteristics by using a long-short-term memory network or a gating circulation unit, the RNN layer inputs spatial characteristics extracted by the CNN layer at each time step, namely, a characteristic diagram of the radiation field, and the implicit state size is set to 128 dimensions so as to balance the expression capacity and the calculated amount.
8. The nuclear accident radiation field feedback correction model based on the adaptive deep assimilation network of claim 1, wherein the initial radiation field prediction module comprises a nuclear accident physical model chain including a forecasting model WRF, a diagnosis wind field model CALMET and a Lagrange particle diffusion model.
9. The nuclear accident radiation field feedback correction model based on the adaptive depth assimilation network of claim 1, wherein the intelligent optimization and feedback correction module is characterized in that the intelligent optimization and feedback correction module is used for adaptively optimizing key parameters of the ADAN module, and particularly comprises the steps of extracting variation trend and spatial distribution characteristics of a prediction error based on assimilation residual errors between corrected radiation field distribution output by the ADAN module at a plurality of continuous moments and corresponding real-time monitoring data, and triggering on-line adaptive adjustment of the key parameters of the ADAN module when continuous deviation or systematic variation of the prediction error is detected so as to optimize radiation field assimilation and prediction accuracy at subsequent moments.
10. The nuclear accident radiation field feedback correction model based on the adaptive depth assimilation network of claim 1, wherein the visual contents comprise a two-dimensional or three-dimensional dose rate distribution map of a nuclear accident influence area, an isodose line superimposed map, a key monitoring point reading dynamic display and a radiation cloud diffusion animation with time.

Description

Nuclear accident radiation field feedback correction model based on self-adaptive depth assimilation network Technical Field The invention relates to the technical field of dynamic monitoring and correction of nuclear radiation fields, in particular to a nuclear accident radiation field feedback correction model based on a self-adaptive depth assimilation network. Background The nuclear accident result evaluation system is widely applied to emergency response after nuclear accidents, and model chains such as WRF, CALMET, lagrange particle diffusion modes and the like are generally adopted to predict radiation field distribution and diffusion trend of the accident scene. These models rely on meteorological data and initial parameter settings at the time of the incident to provide, to some extent, predictions of the radiation field distribution at the beginning of the incident. However, the nuclear accident emergency environment is complex and changeable, the radiation field may be affected by many factors to generate significant deviation, and in the emergency environment, the radiation field distribution often cannot be kept stable along with the continuous change of the accident scene. Particularly, under complex meteorological conditions or changeable terrain environments, the prediction result of the model chain has larger deviation from the actual situation, and accurate support for real-time decision making is difficult to provide. The prior art has the following characteristics in terms of real-time correction and dynamic feedback: (1) Static model chain prediction, namely an existing nuclear accident evaluation model usually adopts a static physical model chain to conduct initial radiation field prediction, does not have the capability of real-time data assimilation, fixes model parameters after initial setting of an accident, and cannot conduct self-adaptive adjustment in the accident development process. (2) And (3) transmitting data by the radiation monitoring equipment, namely arranging gamma radiation monitoring equipment at the accident site, and returning real-time radiation data through a data transmission network (such as 5G). While the monitoring data can provide radiation at the scene of the accident, it is difficult to fuse with the model chain in real time. (3) The radiation field distribution map is often obtained by overlapping the monitoring data with the prediction data in the prior art, however, the method lacks the accuracy of data assimilation, and the radiation field meeting the actual situation cannot be generated. In the nuclear accident emergency response, the traditional nuclear accident result evaluation system can generate a preliminary prediction result of an accident radiation field, but is greatly influenced by environmental factors in a dynamic environment. The drawbacks of the prior art are mainly manifested in the following aspects: (1) Failure to achieve real-time dynamic feedback Existing physical model chains typically rely on incident initiation conditions to generate radiation field prediction results, lacking the ability to feed back real-time data and dynamically correct. After a nuclear accident, the accident scene may change continuously with time, and changes in meteorological conditions such as wind direction, air pressure, precipitation, etc. affect the tendency and speed of radiation diffusion. However, the conventional model cannot correct and optimize the prediction result by using on-site real-time monitoring data, so that the accuracy of radiation field distribution prediction in a complex environment is difficult to meet the actual requirements. (2) The fixed parameters lead to poor model adaptability In nuclear accident emergency response, radiation spread at the accident site may be affected by a variety of factors including meteorological conditions, topographical features, accident source intensity, etc. In the conventional evaluation model, parameters of the model chain are usually kept fixed after initial setting, and cannot be adaptively adjusted according to on-site monitoring data. Such fixed parameter settings make it difficult for the model to cope with changes in the accident scene environment, affecting the accuracy and practicality of the predictions. (3) Depth fusion in the absence of multi-source data A plurality of gamma radiation monitoring devices are usually deployed at the nuclear accident site, and radiation dose data in the accident area can be obtained through a real-time monitoring model. However, the fusion of the conventional evaluation model to the multi-source data is limited to simple superposition, and the spatial and temporal characteristics of the monitored data cannot be deeply analyzed and utilized, so that the prediction result is difficult to accurately reflect the actual distribution of the radiation field. (4) The data assimilation algorithm lacks intelligence and adaptability When the traditional model is used