Nuclear fuel production demand prediction method and system based on multi-model measurement and calculation

Abstract

The invention discloses a nuclear fuel production demand prediction method and system based on multi-model measurement and calculation. The method comprises the following steps: S1, acquiring basic data required by nuclear fuel production demand prediction; s2, related parameters in nuclear fuel production demand prediction are extracted from the basic data; s3, based on statistical model measurement and calculation and deep learning algorithm measurement and calculation, prediction of nuclear fuel production demands is realized; and S4, comparing and analyzing the measurement and calculation results by using a multi-criterion decision-making method, measuring and calculating the scene according to the measurement and calculation parameters, and recommending an optimal scheme. According to the invention, a brand new thought and method are provided for prediction of nuclear fuel production demands, and the intelligent level of measurement and calculation is greatly improved.

Description

Technical Field

[0001] This invention relates to the field of nuclear fuel production demand forecasting, and in particular to a method and system for forecasting nuclear fuel production demand based on multi-model calculation. Background Technology

[0002] Against the backdrop of accelerated global energy structure adjustment, nuclear energy, with its significant advantages of being clean and efficient, is playing an increasingly important role in the energy system. As a core link in the nuclear energy industry, the level of digitalization in nuclear fuel production directly impacts the safe and efficient development of the nuclear energy industry. Nuclear fuel demand forecasting is a crucial task for nuclear fuel production companies in capacity planning and management. Building a flexible and intelligent forecasting system to achieve intelligent and accurate demand prediction will help optimize nuclear fuel production and distribution, ensuring the sustainable development of the nuclear energy industry.

[0003] However, nuclear fuel production demand originates from nuclear power units that are currently operating or under construction within nuclear power companies. Different nuclear power units employ different reactor technologies, resulting in varying fuel requirements and refueling cycles. Traditional nuclear fuel production demand forecasting is based on standard parameters used in the design of all nuclear power units, including refueling cycles and fuel requirements. Since the design parameters of nuclear power units lag behind market changes during actual operation, continuously using fixed parameters for demand forecasting will lead to significant discrepancies between the calculated results and actual demand, resulting in low accuracy. Methods for forecasting nuclear fuel production demand include statistical methods, system dynamics methods, and machine learning methods. The International Atomic Energy Agency has developed a nuclear fuel cycle simulation system model, and the World Nuclear Association publishes a nuclear fuel supply and demand report every two years, forecasting supply and demand at different stages of the nuclear fuel cycle. However, there are no well-established tools specifically designed for nuclear fuel production demand forecasting in the domestic market. Relying solely on traditional manual forecasting methods is cumbersome and inaccurate. Medium- to long-term forecasting of nuclear fuel production demand requires combining various factors such as policy and market conditions, adjusting different parameter variables, and achieving demand forecasting under different scenarios to find the optimal solution for company capacity allocation. Traditional calculation methods rely heavily on human experience, and the final calculation scheme is determined through manual comparison and decision analysis. This method has limitations in terms of professionalism, accuracy, and sustainability. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for predicting nuclear fuel production demand based on multi-model calculation, which brings a brand-new idea and method to the prediction of nuclear fuel production demand and greatly improves the intelligence level of the calculation.

[0005] The objective of this invention is achieved through the following technical solution: a method for predicting nuclear fuel production demand based on multi-model calculations, comprising the following steps:

[0006] Step S1. Obtain the basic data required for nuclear fuel production demand forecasting;

[0007] Step S2. Extract relevant parameters from the nuclear fuel production demand forecast from the basic data;

[0008] Step S3. Based on statistical model calculations and deep learning algorithm calculations, the nuclear fuel production demand is predicted;

[0009] Step S4. Use the multi-criteria decision-making method to compare and analyze the calculation results, and recommend the optimal solution based on the calculation parameters and the calculation scenario.

[0010] A nuclear fuel production demand forecasting system based on multi-model calculations includes:

[0011] The data acquisition module is used to acquire the basic data required for nuclear fuel production demand forecasting;

[0012] The parameter extraction module is used to extract relevant parameters from the nuclear fuel production demand forecast from the basic data.

[0013] A demand forecasting module is generated, which uses statistical models and deep learning algorithms to predict the demand for nuclear fuel production.

[0014] The comparison and optimization module uses a multi-criteria decision-making method to compare and analyze the calculation results, and recommends the optimal solution based on the calculation parameters and the calculation scenario.

[0015] The beneficial effects of this invention are as follows: Based on a large amount of historical refueling demand data, this invention utilizes statistical regression models and deep learning models, which are different from traditional empirical algorithms, to achieve intelligent generation and comparison of calculation schemes, recommendation and optimization of the best scheme, according to the boundary conditions under different scenarios. This overcomes the lag effect brought about by traditional fixed parameter prediction and breaks the professional limitations of relying on human experience. It brings a brand-new idea and method to the prediction of nuclear fuel production demand and greatly improves the level of intelligence in calculation. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention;

[0017] Figure 2 This is a system functional module diagram in the embodiment. Detailed Implementation

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0019] like Figure 1 As shown, a method for predicting nuclear fuel production demand based on multi-model calculations includes the following steps:

[0020] Step S1. Obtain the basic data required for nuclear fuel production demand forecasting;

[0021] Before forecasting nuclear fuel production demand, the basic data affecting the forecast must be prepared and processed. This basic data includes two categories: standard data for nuclear power units and actual refueling data for nuclear power units.

[0022] Nuclear power unit standard data preparation: Input and prepare the reactor technology, status (in service, under construction, new unit), refueling cycle, and refueling demand (product enrichment, tailings enrichment, demand) of the nuclear power units involved.

[0023] Preparation of actual refueling data for nuclear power units: Input and prepare refueling data for all previous nuclear power units involved, including refueling demand (product enrichment, tailings enrichment, demand), refueling cycle, and refueling batches.

[0024] Step S2. Extract relevant parameters from the nuclear fuel production demand forecast from the basic data;

[0025] For the two types of basic data in data preparation, relevant parameters are extracted to form the nuclear fuel production demand forecast. All parameters are based on nuclear power units, and three types of parameters are extracted for each group of nuclear power units: standard, actual average, and latest actual.

[0026] Standard parameters, namely the theoretical design parameters of nuclear power units, are extracted from the preparation of standard data for nuclear power units;

[0027] Actual average parameters are obtained by using a time series analysis model to calculate the average parameters from historical material change data.

[0028] The latest actual parameters are based on the last refueling data from the actual refueling data of the nuclear power unit.

[0029] Each parameter category includes the material change cycle.

[0030] Step S3. Based on statistical model calculations and deep learning algorithm calculations, the nuclear fuel production demand is predicted;

[0031] When conducting nuclear fuel production demand forecasting, it is necessary to select the time period and unit to be forecasted, and select the refueling cycle parameter values ​​(i.e., standard, actual average, and actual latest) according to the calculation scenario. The sub-steps are as follows:

[0032] Calculation time selection: Select the calculation time range according to the calculation needs, such as 2026 to 2030. The starting point of this calculation time is the time of the last material change in history.

[0033] Unit selection for calculation: Select the nuclear power units involved in the calculation according to the calculation requirements. This data can be filtered from the units in the data preparation. The reactor technology, status and parameters of the unit should be automatically populated.

[0034] Material changeover cycle parameter recommendations: Based on pre-set calculation scenarios, the system will automatically recommend applicable material changeover cycle parameters for calculation. That is, if the scenario is a theoretical calculation, standard parameters will be recommended; if the scenario is a calculation based on the latest conditions, the latest actual parameters will be recommended; if the scenario is a calculation based on a model, the actual average parameters will be recommended.

[0035] Based on statistical model calculations:

[0036] Other parameter adjustments: Based on the selected refueling cycle parameters, this involves using statistical models (multiple linear regression) and uranium enrichment material balance theory, and analyzing the quantitative relationship between product enrichment, tailings enrichment, and fuel demand based on historical nuclear fuel refueling data. Here, the variable relationships are derived using the uranium enrichment material balance formula, and then the impact is quantified through regression analysis. The detailed sub-steps are as follows:

[0037] In nuclear fuel production, natural uranium ( =0.711%, or 0.00711), after concentration, yields the product (nuclear fuel) and tailings, satisfying the material balance: (U-235 mass conservation); (Total uranium mass is conserved). Where F is the amount of natural uranium feed, P is the amount of nuclear fuel products (i.e., nuclear fuel demand), and T is the amount of tailings.

[0038] By eliminating T from both equations simultaneously, the relationship between nuclear fuel demand and product enrichment can be derived. ) and tailings enrichment ( The relationship between )

[0039]

[0040] When the amount of F natural uranium feed is fixed, there is an influence relationship between the demand for P nuclear fuel and the enrichment of the product and the enrichment of the tailings.

[0041] Fifty sets of refueling data were extracted from historical nuclear fuel refueling data to construct a multiple linear regression model. Each data set must include three values: nuclear fuel demand, product enrichment, and tailings enrichment. The model is as follows:

[0042]

[0043] Estimating coefficients using the least squares method , , In the embodiments of this application, the accuracy of the model can be verified by mean squared error. The reliability of the model is ensured by significance testing, goodness-of-fit analysis, and residual analysis.

[0044] In the embodiments of this application, fuel demand can be calculated by adjusting the parameters of product enrichment and tailings enrichment based on the fitted model.

[0045] Scenario 1: Adjusting only the product enrichment ( ), fixed tailings enrichment ( For example, to Increased from 3.0% to 3.5%. The percentage remains unchanged at 0.25%. The product enrichment level is calculated based on the data. The product enrichment (P) is negatively correlated with the nuclear fuel demand P. The higher the value, the less nuclear fuel P is required.

[0046] Scenario 2: Adjust only the enrichment degree of tailings ( ), fixed product enrichment ( For example, to From 0.25% to 0.2%, The percentage remains unchanged at 3.0%. Calculations based on data show that the enrichment degree of the tailings (…) The enrichment of tailings is positively correlated with the nuclear fuel demand P. The higher the value, the greater the nuclear fuel demand (P).

[0047] Scenario 3: Collaboratively adjust the scenario and simultaneously adjust the product enrichment level. ) and tailings enrichment ( Sensitivity analysis was conducted to calculate the impact of product enrichment and tailings enrichment on fuel demand.

[0048] Through the above modeling process, based on historical refueling data, this invention enables flexible adjustment of product enrichment and tailings enrichment, thereby calculating changes in fuel demand.

[0049] Calculation based on deep learning model algorithm:

[0050] By using deep learning models for prediction, a deep neural network model containing at least one long short-term memory (LSTM) layer is constructed, which can capture more complex nonlinear relationships compared to multiple linear regression models.

[0051] C1. Data Acquisition and Preprocessing: Based on the selected nuclear power unit, extract multiple feature data from historical refueling data, including refueling time, product enrichment, tailings enrichment, and nuclear fuel demand.

[0052] Data cleaning and interpolation are performed. The material change time is encoded using either a single thermal encoding or a periodic encoding, and then interpolated to obtain characteristic data under different material change times. Data normalization processing is then carried out.

[0053] The normalization process uses the min-max normalization method to unify data features of different dimensions and orders of magnitude to a specified scale space. The formula is as follows:

[0054]

[0055] in:

[0056] X is the original eigenvalue.

[0057] Xmin is the minimum value of this feature across the entire dataset.

[0058] Xmax is the maximum value of this feature across the entire dataset.

[0059] Xnorm is the normalized value, ranging from [0,1].

[0060] C2. Construct the sequence and convert the time data into a supervised learning format:

[0061] Define a time step N (e.g., N=12 months), and... Normalized feature data up to time t as input sequence And use the nuclear fuel requirement at time t+1 as a tag. ;

[0062] Thus, a sample was obtained. ;

[0063] C3. Repeat C1 multiple times to obtain a sample set, and then use this set to build and train the prediction model.

[0064] C301. A prediction model based on deep learning algorithms is constructed using a stacked long short-term memory network architecture:

[0065] Input layer, with input sequences of time step N. ;

[0066] The first LSTM layer has Each neural unit has an activation function (such as ReLU or Tanh) and its hidden state sequence is output to the next layer with `return_sequences=True` set; (optional) First Dropout layer: randomly deactivates neurons at a certain ratio (e.g., 0.2) to prevent overfitting; Second LSTM layer: has One neural unit, with `return_sequences=False`, outputs only the hidden state of the last time step of the sequence; (optional) Second Dropout layer: further prevents overfitting. Output layer (fully connected layer): with one neural unit and a linear activation function, used to output continuous regression predictions. ;

[0067] C302. Based on regression predictions and tags Choose mean squared error or mean absolute error as the regression loss function; use Adam or RMSprop as the adaptive learning rate optimizer to update the prediction model;

[0068] C303. Model Training: Divide the sample set into a training set and a validation set, and use the samples in the training set to train the prediction model until the loss of the model on the validation set converges.

[0069] C4. Demand Forecasting: Feed the prepared input sequence with time step N into the trained LSTM model, and the model outputs the normalized predicted value. ,right Perform an inverse transformation to obtain a physically meaningful predicted value for nuclear fuel demand.

[0070] For different nuclear power units, the model needs to be constructed according to step S3 and the optimal solution needs to be recommended according to step S4.

[0071] Step S4. Use the multi-criteria decision-making method to compare and analyze the calculation results, and recommend the optimal solution based on the calculation parameters and the calculation scenario.

[0072] Based on the calculation results of different algorithm models, the system records all selection parameters of the scheme. With the parameters fixed, it uses the Multi-Criterion Decision System (TOPSIS) to compare and analyze the results calculated by different model algorithms, and recommends the optimal scheme based on the selection of calculation parameters or the calculation scenario, including:

[0073] S401. Select performance indicators and evaluation indicators for statistical model calculations and deep learning algorithm-based calculations;

[0074] (1) Performance indicators include coefficient of determination, root mean square error, and mean absolute error;

[0075] (1.1) For the prediction model of nuclear fuel production demand based on statistical model: extract the refueling demand from multiple historical refueling data, including product enrichment, tailings enrichment and nuclear fuel demand, and input the product enrichment and tailings enrichment from each historical refueling data into the prediction model of nuclear fuel production demand based on statistical model to obtain multiple nuclear fuel demand prediction results. Combine the nuclear fuel demand from each historical refueling data to calculate the coefficient of determination, root mean square error and mean absolute error.

[0076] (1.2) For the prediction model based on deep learning algorithm, firstly obtain multiple samples according to steps C1~C2; then send the input sequence of each sample into the prediction model based on deep learning algorithm, and calculate the coefficient of determination, root mean square error and mean absolute error of the model output results and the actual labels of each sample.

[0077] Among them, the coefficient of determination The closer the value is to 1, the better the model fit.

[0078] Root mean square error This reflects the average deviation between the predicted value and the actual value.

[0079] Mean Absolute Error , which represents the average absolute deviation between each predicted value and the actual value.

[0080] When using the above formula to calculate the nuclear fuel production demand forecasting model based on statistical models, since multiple historical refueling data points are extracted, each extracted data point will yield a nuclear fuel demand forecast result. The average value of each data point is The actual nuclear fuel requirement for each data point is Where n represents the number of historical material change data entries extracted. These represent the nuclear fuel demand forecast and the actual nuclear fuel demand for the i-th historical refueling data, respectively. (2) Evaluation indicators include computational cost and sample size requirements, which are implemented through customization:

[0081] Calculation cost score: The higher the calculation cost, the lower the score. By customizing the calculation cost score, calculation cost scores can be set for the nuclear fuel production demand prediction model based on the statistical model and the prediction model based on the deep learning algorithm. The calculation cost score of the nuclear fuel production demand prediction model based on the statistical model (e.g., 0.8) is greater than that of the prediction model based on the deep learning algorithm (e.g., 0.2).

[0082] Sample size score: A score needs to be assigned to the samples of the statistical model and the neural network model: the higher the required sample size, the lower the score. By customizing the sample size score, a sample size score can be set for the nuclear fuel production demand prediction model based on the statistical model and the prediction model based on the deep learning algorithm. The sample size score of the nuclear fuel production demand prediction model based on the statistical model (e.g., 0.8) is greater than that of the prediction model based on the deep learning algorithm (e.g., 0.2).

[0083] S402. Based on the scenario of the scheme calculation, select different models to generate demand forecasts. The scenario of the scheme calculation includes conventional production planning and new fuel research and development.

[0084] (1) Let the custom weights of the coefficient of determination, root mean square error, and mean absolute error be 1. The custom weights for calculating the cost score and the sample size score are both... ,and ;

[0085] For routine production plans, it is necessary to set up ;For example, ;

[0086] For the research and development of new fuels, it is necessary to set up ;For example, ;

[0087] (2) In any scenario, the determination coefficient, root mean square error, mean absolute error, cost score, and sample size score of the nuclear fuel production demand forecasting model calculated based on the statistical model are weighted and summed according to the set weights.

[0088] The determination coefficient, root mean square error, mean absolute error, cost score, and sample size score of the prediction model calculated based on deep learning algorithms are weighted and summed according to the set weights.

[0089] The calculation scheme corresponding to the model with the higher weighted summation result is selected as the recommended prediction scheme.

[0090] In the embodiments of this application, result generation and analysis refers to generating a breakdown of the nuclear fuel demand covered by the selected optimal calculation scheme, which can be broken down by nuclear power unit and year. Furthermore, it utilizes visualization charts and other methods to display the time trend analysis and composition analysis of nuclear fuel production demand, assisting nuclear fuel companies in making capacity allocation decisions.

[0091] A nuclear fuel production demand forecasting system based on multi-model calculations includes:

[0092] The data acquisition module is used to acquire the basic data required for nuclear fuel production demand forecasting;

[0093] The parameter extraction module is used to extract relevant parameters from the nuclear fuel production demand forecast from the basic data.

[0094] A demand forecasting module is generated, which uses statistical models and deep learning algorithms to predict the demand for nuclear fuel production.

[0095] The comparison and optimization module compares and analyzes the results calculated using a multi-criteria decision-making method, and recommends the optimal solution based on the selection of calculation parameters or the calculation scenario.

[0096] In the embodiments of this application, based on the system, the architecture in actual application is as follows: Figure 2 As shown, it mainly includes modules for basic information management, material change information management, calculation scheme formulation and management, and visualization analysis. The relationships between these modules are as follows:

[0097] Basic Information Management Module: This module is deployed on the server side and is used to prepare and process basic data for nuclear power units. Its main functions are to provide input, classification, and retrieval of information on nuclear power companies, nuclear power plants, and nuclear power units, as well as to extract and display parameters.

[0098] Refueling Information Management Module: This module is deployed on the server side and is used to prepare and process historical refueling data of nuclear power units. Its main functions are to provide data entry, classification, and retrieval for each refueling operation, as well as to extract and display relevant parameters.

[0099] The calculation scheme formulation and management module receives nuclear power unit, refueling, and parameter data from the basic information management and refueling information management modules, and provides full-process functional support for calculation scheme formulation. Its main functions include calculation preparation, parameter adjustment, and result confirmation. Calculation preparation supports selecting the calculation time and unit range, and recommends parameters for the selected unit; parameter adjustment enables model selection, parameter adjustment, and calculation; and result confirmation generates a complete calculation scheme and allows for scheme comparison and optimization.

[0100] Visualization Analysis Module: This module receives data from the calculation scheme formulation and management module, enabling the visualization and analysis of the calculation results. Its main functions include nuclear fuel production demand trend analysis and composition analysis, assisting nuclear fuel enterprise management decisions.

[0101] Workflow

[0102] The system described in this invention can be deployed on a cloud server and interacts with users through front-end interfaces such as web pages. Its typical workflow is as follows:

[0103] User information entry: Users enter basic information and material change information through the front-end interface. The system extracts relevant parameters and displays and queries them on the front-end interface.

[0104] Calculation Scheme Formulation: Users conduct nuclear fuel production demand forecasting through the front-end interface, selecting the calculation time and scope. The system automatically recommends relevant parameters for the selected unit. Users choose a calculation model based on actual conditions and adjust its parameters. The system automatically generates relevant schemes. Based on factors such as data dependency and applicable scenarios, the system recommends the optimal scheme. Users can compare multiple schemes. The server-side backend system compares the differences between the generated schemes and calculates their rationality, displaying the results intuitively through the front-end interface to assist users in selecting the final scheme.

[0105] Results presentation: After the selected calculation scheme is generated, the server will transmit the data to the corresponding visualization page. The front-end interface will be updated with the latest calculation results data of the calculation scheme, and users can click and view further for deeper interaction.

[0106] In summary, this invention, based on full utilization of historical refueling data, innovatively adopts statistical regression models and deep learning models for nuclear fuel demand forecasting. It comprehensively utilizes the system to achieve model selection, parameter adjustment, and scheme recommendation during the calculation process, improving and enhancing the nuclear fuel production demand forecasting process and greatly increasing forecast accuracy. Furthermore, by using regression models to combine the physical nature of nuclear fuel production with statistical analysis methods, it achieves a balance between interpretability and practicality, overcoming the limitations of traditional empirical forecasting. Finally, by employing neural network models to solve the challenges of nonlinearity, multiple couplings, and extreme scenarios in nuclear fuel demand forecasting, it breaks through the fitting boundaries of traditional linear models, achieving higher-precision forecasting.

Claims

1. A method for predicting nuclear fuel production demand based on multi-model calculations, characterized in that: Includes the following steps: Step S1. Obtain the basic data required for nuclear fuel production demand forecasting; Step S2. Extract relevant parameters from the nuclear fuel production demand forecast from the basic data; Step S3. Based on statistical model calculations and deep learning algorithm calculations, the nuclear fuel production demand is predicted; Step S4. Use the multi-criteria decision-making method to compare and analyze the calculation results, and recommend the optimal solution based on the calculation parameters and the calculation scenario.

2. The method for predicting nuclear fuel production demand based on multi-model calculation according to claim 1, characterized in that: Assuming there are multiple nuclear power units, for each nuclear power unit, it is necessary to obtain the basic data required for production demand forecasting. The basic data includes the standard data of the nuclear power unit and the actual refueling data of the nuclear power unit. The standard data for nuclear power units includes the standard reactor technology, status, refueling cycle, and refueling requirements for nuclear power units. The actual refueling data of the nuclear power unit adopts multiple historical refueling data of the nuclear power unit. Each historical refueling data includes refueling demand, refueling cycle and refueling batch.

3. The method for predicting nuclear fuel production demand based on multi-model calculation according to claim 2, characterized in that: The status includes units in service, under construction, and new units; The refueling demand includes product enrichment, tailings enrichment, and nuclear fuel demand.

4. The method for predicting nuclear fuel production demand based on multi-model calculation according to claim 3, characterized in that: In step S2, for each group of nuclear power units, the extracted relevant parameters include standard parameters, actual average parameters, and actual latest parameters. The standard parameters include the standard refueling cycle for nuclear power units; The actual average parameters are obtained by calculating the average material change cycle through multiple historical material change data. The actual latest parameters are based on the material replacement cycle of the last material replacement in the historical material replacement data.

5. The method for predicting nuclear fuel production demand based on multi-model calculation according to claim 4, characterized in that: Step S3 includes: A1. Select the time and unit to be predicted, and select the value of the refueling cycle parameter according to the calculation scenario. A2. Select the measurement time range, the starting point of which is the time of the last material change in history; A3. Select the nuclear power unit that needs to be measured and screen the basic data of the unit; A4. Material Change Cycle Parameter Determination: Based on the pre-set calculation scenarios, the recommended material change cycle parameters are calculated as follows: That is, if the scenario is based on theoretical calculations, the recommended material change cycle is the standard parameter. If the scenario is based on model calculations, the average value of the material change cycle in the actual average parameters is recommended. If the scenario is calculated based on the latest state, the recommended material change cycle is the last material change in the latest actual parameters; A5. Starting from the beginning of the calculation time range, within the calculation time range, a prediction of nuclear fuel demand is made every other refueling cycle. The prediction methods include calculations based on statistical models and calculations based on deep learning algorithms.

6. The method for predicting nuclear fuel production demand based on multi-model calculation as described in claim 5, characterized in that: The aforementioned calculation based on statistical models involves, after determining the refueling cycle parameters, utilizing statistical models and uranium enrichment material balance theory, and based on historical refueling data of nuclear fuel refueling demand, analyzing the quantitative impact relationship between product enrichment, tailings enrichment, and fuel demand, including: B101. Enrichment of natural uranium products in nuclear fuel production. =0.711%, after enrichment, nuclear fuel and tailings are obtained, satisfying the material balance, including: U-235 mass conservation: ; Total uranium mass conservation ; Where F represents the amount of natural uranium feed, P represents the amount of nuclear fuel products (i.e., nuclear fuel demand), and T represents the amount of tailings. Enrichment level of natural uranium products; By eliminating T using the two equations simultaneously, the nuclear fuel demand and product enrichment can be derived. and the enrichment of tailings The relationship between them: ； When the amount of F natural uranium feed is fixed, there is an influence relationship between the demand for P nuclear fuel and the enrichment of the product and the enrichment of the tailings. B102. Based on the selected nuclear power unit, extract the refueling demand from multiple historical refueling data and construct a multiple linear regression model; Each set of data must include three values: nuclear fuel demand, product enrichment, and tailings enrichment; the model is as follows: ； Estimating coefficients using the least squares method , , That is, to obtain a predictive model of nuclear fuel production demand based on statistical model calculations; The enrichment of products and the enrichment of tailings are input into a predictive model for nuclear fuel production demand based on a statistical model to obtain the calculation results.

7. The method for predicting nuclear fuel production demand based on multi-model calculation according to claim 6, characterized in that: The calculation based on deep learning algorithms includes: C1. Data Acquisition and Preprocessing: Based on the selected nuclear power unit, extract multiple feature data from historical refueling data, including refueling time, product enrichment, tailings enrichment, and nuclear fuel demand. Data cleaning and interpolation are performed. The material change time is encoded using either a single thermal encoding or a periodic encoding, and then interpolated to obtain characteristic data under different material change times. Data normalization processing is then carried out. The normalization process involves using the minimum-maximum normalization method to unify data features of different dimensions and orders of magnitude into a specified scale space. C2. Construct the sequence and convert the time data into a supervised learning format: Define a time step N, and... Normalized feature data up to time t as input sequence And use the nuclear fuel requirement at time t+1 as a tag. ; Thus, a sample was obtained. ; C3. Repeat C1 multiple times to obtain a sample set, and then use this set to build and train the prediction model. C301. A prediction model based on deep learning algorithms is constructed using a stacked long short-term memory network architecture: Input layer, with input sequences of time step N. ; The first LSTM layer has Each neural unit has a hidden state sequence output to the next layer, with ReLU or tanh activation functions. First Dropout layer: Randomly deactivates neurons according to a set ratio to prevent overfitting; Second LSTM layer: has Each neural unit is configured with return_sequences=False to output only the hidden state of the last time step of the sequence. The second Dropout layer randomly deactivates neurons according to a set ratio to further prevent overfitting. The output layer, a fully connected layer, has one neuron and a linear activation function, and is used to output continuous regression prediction values. ; C302. Based on regression predictions and tags Choose mean squared error or mean absolute error as the regression loss function; use Adam or RMSprop as the adaptive learning rate optimizer to update the prediction model; C303. Model Training: Divide the sample set into a training set and a validation set. Use the samples in the training set to train the prediction model until the loss of the prediction model on the validation set converges. C4. Demand Forecasting: Feed the prepared input sequence with time step N into the trained LSTM model, and the model outputs the normalized predicted value. ,right Perform an inverse transformation to obtain a physically meaningful predicted value for nuclear fuel demand.

8. The method for predicting nuclear fuel production demand based on multi-model calculation according to claim 7, characterized in that: Step S4 includes: S401. Select performance indicators and evaluation indicators for statistical model calculations and deep learning algorithm-based calculations; (1) Performance indicators include coefficient of determination, root mean square error, and mean absolute error; (1.1) For the prediction model of nuclear fuel production demand based on statistical model: extract the refueling demand from multiple historical refueling data, including product enrichment, tailings enrichment and nuclear fuel demand, and input the product enrichment and tailings enrichment from each historical refueling data into the prediction model of nuclear fuel production demand based on statistical model to obtain multiple nuclear fuel demand prediction results. Combine the nuclear fuel demand from each historical refueling data to calculate the coefficient of determination, root mean square error and mean absolute error. (1.2) For the prediction model based on deep learning algorithm, firstly obtain multiple samples according to steps C1~C2; then send the input sequence of each sample into the prediction model based on deep learning algorithm, and calculate the coefficient of determination, root mean square error and mean absolute error of the model output results and the actual labels of each sample. (2) Evaluation indicators include computational cost and sample size requirements, which are implemented through customization: Calculation cost score: The higher the calculation cost, the lower the score. By customizing the calculation cost score, calculation cost scores can be set for the nuclear fuel production demand prediction model based on the statistical model and the prediction model based on the deep learning algorithm. The calculation cost score of the nuclear fuel production demand prediction model based on the statistical model is greater than that of the prediction model based on the deep learning algorithm. Sample size score: The statistical model and the neural network model need to be assigned a score: the higher the sample size requirement, the lower the score. By setting the sample size score in a custom way, the sample size score of the nuclear fuel production demand prediction model based on the statistical model and the prediction model based on the deep learning algorithm can be set. The sample size score of the nuclear fuel production demand prediction model based on the statistical model is greater than that of the prediction model based on the deep learning algorithm. S402. Based on the scenario of the scheme calculation, select different models to generate demand forecasts. The scenario of the scheme calculation includes conventional production planning and new fuel research and development. (1) Let the custom weights of the coefficient of determination, root mean square error, and mean absolute error be 1. The custom weights for calculating the cost score and the sample size score are both... ,and ; For routine production plans, it is necessary to set up ; For the research and development of new fuels, it is necessary to set up ; (2) In any scenario, the determination coefficient, root mean square error, mean absolute error, cost score, and sample size score of the nuclear fuel production demand forecasting model calculated based on the statistical model are weighted and summed according to the set weights. The determination coefficient, root mean square error, mean absolute error, cost score, and sample size score of the prediction model calculated based on deep learning algorithms are weighted and summed according to the set weights. The calculation scheme corresponding to the model with the higher weighted summation result is selected as the recommended prediction scheme.

9. A nuclear fuel production demand forecasting system based on multi-model calculations, wherein the method described in any one of claims 1 to 8 is characterized in that: include: The data acquisition module is used to acquire the basic data required for nuclear fuel production demand forecasting; The parameter extraction module is used to extract relevant parameters from the nuclear fuel production demand forecast from the basic data. A demand forecasting module is generated, which uses statistical models and deep learning algorithms to predict the demand for nuclear fuel production. The comparison and optimization module is used to record all selection parameters of different calculation schemes. With the parameters fixed, it compares and analyzes the calculation results using the multi-criteria decision method, and recommends the optimal scheme based on the selection of calculation parameters or the calculation scenario.

Images