A Multi-Level, Multi-Period Nuclear Power Generation Prediction Method Based on Multi-Model Fusion

Abstract

The application relates to a multi-level multi-period nuclear power generation capacity prediction method based on multi-model fusion, which comprises the following steps: inputting historical time series data sets of each nuclear power station into a first time series prediction model to obtain a first power generation capacity prediction sequence; inputting a historical equipment availability rate sequence and a historical power load demand sequence into a second time series prediction model to obtain the equipment availability rate sequence and the power load demand sequence, and obtaining a second power generation capacity prediction sequence according to the equipment availability rate sequence and the power load demand sequence; inputting the historical time series data sets, the equipment availability rate sequence and the power load demand sequence into a third time series prediction model to obtain a third power generation capacity prediction sequence; obtaining a target power generation capacity prediction sequence according to the first power generation capacity prediction sequence, the second power generation capacity prediction sequence and the third power generation capacity prediction sequence; and performing a summary on the target power generation capacity prediction sequence to obtain a multi-level multi-period power generation capacity prediction value, so that the nuclear power generation capacity prediction precision is improved.

Description

Technical Field

[0001] This application relates to the field of power system planning and time-series forecasting technology, and in particular to a multi-level, multi-cycle nuclear power generation forecasting method, apparatus, computer equipment, and computer-readable storage medium based on multi-model fusion. Background Technology

[0002] With the steady increase in installed nuclear power capacity, its role in supporting the base load of the power system is becoming increasingly prominent. Accurate medium- and long-term nuclear power generation forecasts have become a key prerequisite for ensuring the safe and stable operation of the power system, optimizing the allocation of energy resources, and formulating scientific energy policies.

[0003] However, the nuclear power generation prediction methods in related technologies suffer from low accuracy at certain temperatures. Summary of the Invention

[0004] Therefore, it is necessary to provide a multi-level, multi-cycle nuclear power generation prediction method, device, computer equipment, and computer-readable storage medium based on multi-model fusion that can improve the accuracy of nuclear power generation prediction, addressing the aforementioned technical problems.

[0005] Firstly, this application provides a multi-level, multi-cycle nuclear power generation prediction method based on multi-model fusion, the method comprising:

[0006] Obtain historical time-series datasets for each nuclear power plant; these datasets are obtained through preprocessing of historical integrated operational data from each nuclear power plant.

[0007] The historical time series dataset is input into the first time series prediction model to obtain the first power generation prediction sequence for future time periods; the first time series prediction model is a deep learning model that uses an inverted attention mechanism.

[0008] The historical equipment availability sequence and historical power load demand sequence from the historical time series dataset are input into the second time series prediction model to obtain the equipment availability sequence and power load demand sequence for the future time period. Based on the equipment availability sequence and power load demand sequence, the second power generation prediction sequence for the future time period is obtained. The second time series prediction model is a statistical model using an additive decomposition framework.

[0009] Historical time-series datasets, equipment availability sequences, and electricity load demand sequences are input into the third time-series prediction model to obtain the third generation prediction sequence for future time periods; the third time-series prediction model is a machine learning model that adopts a forward distributed learning strategy.

[0010] Based on the first power generation prediction sequence, the second power generation prediction sequence, and the third power generation prediction sequence, the target power generation prediction sequence for each nuclear power plant is obtained.

[0011] Based on the preset regions and preset cycles, the target power generation prediction sequences of each nuclear power plant are summarized to obtain multi-level and multi-cycle power generation prediction values; the preset regions include each nuclear power plant.

[0012] In one embodiment, the first time-series prediction model includes an input layer, a self-attention layer, a feedforward network layer, and an output layer. Historical time-series datasets are input into the first time-series prediction model to obtain a first power generation prediction sequence for future time periods, including:

[0013] The input layer performs dimensionality transpose and feature embedding on the historical time series dataset to obtain the first feature vector sequence.

[0014] By using a self-attention layer, features are extracted from the first feature vector sequence to obtain the second feature vector sequence;

[0015] The second feature vector sequence is extracted through a feedforward network layer to obtain the third feature vector sequence.

[0016] By performing linear mapping and dimension transpose on the third feature vector sequence through the output layer, the first power generation prediction sequence for the future time period is obtained.

[0017] In one embodiment, the historical equipment availability sequence and historical power load demand sequence from the historical time-series dataset are input into a second time-series prediction model to obtain the equipment availability sequence and the power load demand sequence for future time periods, including:

[0018] The second time-series forecasting model decomposes both the historical equipment availability series and the historical power load demand series into trend terms, seasonal terms, holiday terms, and error terms.

[0019] By overlaying the trend term, seasonal term, holiday term, and error term, we can obtain the equipment availability sequence and the power load demand sequence for future time periods.

[0020] In one embodiment, historical time-series datasets, equipment availability sequences, and electricity load demand sequences are input into a third time-series prediction model to obtain a third generation prediction sequence for future time periods, including:

[0021] The historical time series dataset is input into the third time series prediction model to obtain the initial power generation prediction sequence for future time periods;

[0022] Using equipment availability sequence and power load demand sequence as constraints, the initial power generation forecast sequence for future time periods is corrected to obtain the third power generation forecast sequence for future time periods.

[0023] In one embodiment, the target power generation prediction sequence for each nuclear power plant is obtained based on a first power generation prediction sequence, a second power generation prediction sequence, and a third power generation prediction sequence, including:

[0024] Obtain the first historical mean absolute error of the first time series forecasting model, the second historical mean absolute error of the second time series forecasting model, and the third historical mean absolute error of the third time series forecasting model;

[0025] Based on the first historical average absolute error, the second historical average absolute error, and the third historical average absolute error, the first weight of the first time series prediction model, the second weight of the second time series prediction model, and the third weight of the third time series prediction model are obtained.

[0026] Based on the first weight, the second weight, and the third weight, the first power generation prediction sequence, the second power generation prediction sequence, and the third power generation prediction sequence are fused to obtain the fused power generation prediction sequence.

[0027] Obtain the first historical prediction error of the first time series prediction model, the second historical prediction error of the second time series prediction model, and the third historical prediction error of the third time series prediction model;

[0028] The mean error is obtained based on the first historical prediction error, the second historical prediction error, and the third historical prediction error.

[0029] Based on the mean error, the merged power generation prediction sequence is corrected to obtain the target power generation prediction sequence for each nuclear power plant.

[0030] In one embodiment, the multi-level, multi-cycle power generation forecast includes monthly, quarterly, and annual power generation forecasts for various cities and provinces, and after obtaining the multi-level, multi-cycle power generation forecasts, the method further includes:

[0031] Obtain the monthly power generation quotas for each city and each province; if the deviation between the predicted monthly power generation value and the monthly power generation quota for each city exceeds a first deviation threshold, or if the deviation between the predicted monthly power generation value and the monthly power generation quota for each province exceeds a first deviation threshold, perform a retrospective adjustment to the power generation prediction sequence for each nuclear power plant; and / or,

[0032] Obtain the quarterly power generation quotas for each city and each province; if the deviation between the quarterly power generation forecast for each city and its quarterly power generation quota exceeds a second deviation threshold, or if the deviation between the quarterly power generation forecast for each province and its quarterly power generation quota exceeds a second deviation threshold, perform a retrospective adjustment to the power generation forecast sequence for each nuclear power plant; and / or,

[0033] Obtain the annual power generation quotas for each city and each province; if the deviation between the annual power generation forecast for each city and the annual power generation quota for each city exceeds the third deviation threshold, or if the deviation between the annual power generation forecast for each province and the annual power generation quota for each province exceeds the third deviation threshold, backtrack and adjust the power generation forecast sequence for each nuclear power plant.

[0034] In one embodiment, the method further includes:

[0035] If the number of days in the future time period is less than or equal to a preset number, the parameters of the first time series prediction model, the second time series prediction model, and the third time series prediction model are updated at intervals of the first time period.

[0036] If the number of days in the future time period is greater than the preset number, the parameters of the first time series prediction model, the second time series prediction model, and the third time series prediction model are updated at second time intervals.

[0037] In the event of a target event, the parameters of the first time-series forecasting model, the second time-series forecasting model, and the third time-series forecasting model are updated; the target event includes at least one of a major policy adjustment event, an extreme weather event, and a major overhaul event of nuclear power plant equipment.

[0038] Secondly, this application also provides a multi-level, multi-cycle nuclear power generation prediction device based on multi-model fusion, the device comprising:

[0039] The historical time-series dataset acquisition module is used to acquire historical time-series datasets for each nuclear power plant; the historical time-series datasets are obtained by preprocessing the historical comprehensive operation data of each nuclear power plant.

[0040] The first power generation prediction sequence acquisition module is used to input the historical time series dataset into the first time series prediction model to obtain the first power generation prediction sequence for future time periods; the first time series prediction model is a deep learning model using an inverted attention mechanism.

[0041] The second power generation prediction sequence acquisition module is used to input the historical equipment availability sequence and historical power load demand sequence from the historical time series dataset into the second time series prediction model to obtain the equipment availability sequence and power load demand sequence for the future time period. Based on the equipment availability sequence and power load demand sequence, the second power generation prediction sequence for the future time period is obtained. The second time series prediction model is a statistical model using an additive decomposition framework.

[0042] The third power generation prediction sequence acquisition module is used to input historical time series datasets, equipment availability sequences, and power load demand sequences into the third time series prediction model to obtain the third power generation prediction sequence for future time periods; the third time series prediction model is a machine learning model that adopts a forward distributed learning strategy.

[0043] The target power generation prediction sequence acquisition module is used to obtain the target power generation prediction sequence for each nuclear power plant based on the first power generation prediction sequence, the second power generation prediction sequence, and the third power generation prediction sequence.

[0044] The multi-level, multi-cycle power generation prediction module is used to summarize the target power generation prediction sequence of each nuclear power plant according to a preset region and a preset cycle to obtain multi-level, multi-cycle power generation prediction values; the preset region includes each nuclear power plant.

[0045] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method steps of the first aspect.

[0046] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method steps of the first aspect.

[0047] The aforementioned multi-model fusion-based multi-level, multi-cycle nuclear power generation prediction method, device, computer equipment, and computer-readable storage medium acquire historical time-series datasets from each nuclear power plant. These historical time-series datasets are obtained through preprocessing of historical integrated operation data from each nuclear power plant. The historical time-series datasets are then input into a first time-series prediction model to obtain a first power generation prediction sequence for future time periods. The first time-series prediction model is a deep learning model employing an inverted attention mechanism. The historical equipment availability sequence and historical power load demand sequence from the historical time-series datasets are then input into a second time-series prediction model to obtain a future equipment availability sequence and a future power load demand sequence. Based on the equipment availability sequence and the power load demand sequence, a second power generation prediction sequence for future time periods is obtained. The second time-series prediction model is a statistical model using an additive decomposition framework. Historical time-series datasets, equipment availability sequences, and electricity load demand sequences are input into the third time-series prediction model to obtain a third power generation prediction sequence for future time periods. The third time-series prediction model is a machine learning model using a forward distributed learning strategy. Based on the first, second, and third power generation prediction sequences, target power generation prediction sequences for each nuclear power plant are obtained. According to a preset region and preset period, the target power generation prediction sequences for each nuclear power plant are summarized to obtain multi-level, multi-period power generation prediction values. The preset region includes all nuclear power plants. Thus, this application captures the long-term dependencies of historical time-series datasets using the first time-series prediction model, avoiding a significant increase in prediction error as the time span increases. The second time-series prediction model performs independent univariate prediction on regular data with significant periodicity and trends in historical time-series datasets, fully utilizing its adaptive modeling capabilities for seasonality and holiday effects, thereby improving the prediction accuracy of regular data. By using a third-time-series forecasting model, the impact of equipment availability and power load demand on power generation is considered, thus avoiding a disconnect between forecast results and actual operating scenarios. This enables accurate forecasting of nuclear power generation at multiple levels, from nuclear power plants to regions, cities, and provinces, and from monthly to quarterly and annual cycles. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is an application environment diagram of a multi-level, multi-cycle nuclear power generation prediction method based on multi-model fusion in one embodiment;

[0050] Figure 2This is a flowchart illustrating a multi-level, multi-cycle nuclear power generation prediction method based on multi-model fusion in one embodiment.

[0051] Figure 3 This is a flowchart of a multi-level, multi-cycle nuclear power generation prediction method based on multi-model fusion in one embodiment;

[0052] Figure 4 This is a structural block diagram of a multi-level, multi-cycle nuclear power generation prediction device based on multi-model fusion in one embodiment.

[0053] Figure 5 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0055] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0056] The multi-model fusion-based, multi-stage, multi-cycle nuclear power generation prediction method provided in this application can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or placed on a cloud or other network server. Terminal 102 acquires historical time-series datasets from various nuclear power plants; these datasets are obtained through preprocessing of historical integrated operational data from each nuclear power plant; the historical time-series datasets are input into a first time-series prediction model to obtain a first power generation prediction sequence for the future time period; the first time-series prediction model is a deep learning model using an inverted attention mechanism; the historical equipment availability sequence and historical power load demand sequence from the historical time-series dataset are input into a second time-series prediction model to obtain a future equipment availability sequence and a future power load demand sequence; based on the equipment availability sequence and power load demand sequence, a second power generation prediction sequence for the future time period is obtained. The system consists of three time-series power generation prediction sequences. The first time-series prediction sequence uses a statistical model employing an additive decomposition framework. The second time-series prediction model uses a statistical model employing a forward distributed learning strategy. Based on the first, second, and third time-series prediction sequences, power generation prediction sequences for each nuclear power plant are obtained. The power generation prediction sequences for each nuclear power plant are then aggregated according to a preset region and preset period to obtain multi-level, multi-period power generation prediction values. The preset region includes all nuclear power plants. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, drones, low-altitude aircraft, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, projection devices, etc. Portable wearable devices can include smartwatches, smart bracelets, head-mounted devices, etc. Head-mounted devices can be virtual reality (VR) devices, augmented reality (AR) devices, smart glasses, etc. Server 104 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides cloud computing services.

[0057] In one embodiment, such as Figure 2 As shown, a multi-level, multi-cycle nuclear power generation prediction method based on multi-model fusion is provided. This embodiment applies this method to... Figure 1 Taking terminal 102 as an example, the method includes the following steps:

[0058] Step S210: Obtain the historical time-series dataset for each nuclear power plant; the historical time-series dataset is obtained by preprocessing the historical comprehensive operation data of each nuclear power plant.

[0059] The historical comprehensive operation data includes basic data on nuclear power plant operation, power load data, external impact data, and policy and transmission data.

[0060] The basic operational data of nuclear power plants include daily power generation, installed capacity (inherent installed capacity and future planned installed capacity) for the past 5 years or more, equipment availability (utilization hours), and technical route classification (such as "Hualong One" and "Guohe One"), which are derived from the nuclear power plant's SCADA system and production management platform.

[0061] The power load data includes the daily power load demand and load characteristic curves of the region over the past 5 years, and is sourced from the load statistics system of the power grid dispatch center.

[0062] External impact data includes records of extreme weather events (high temperatures, heavy rain, typhoons, etc.), holiday arrangements, and records of major events, sourced from publicly available data from meteorological departments and local government platforms.

[0063] Policy and transmission data include nuclear power policy quotas, inter-regional transmission plans, and provincial and municipal electricity consumption plans, which are derived from policy documents from energy authorities and power grid transmission and dispatch schemes.

[0064] The historical time-series dataset includes historical sequences of dates, daily power generation, installed capacity, equipment availability, power load demand, technology routes, extreme events, and policy quotas.

[0065] In this embodiment of the application, the historical integrated operation data of each nuclear power plant are preprocessed, including data cleaning, data standardization, feature construction, and data partitioning, to obtain the historical time-series dataset of each nuclear power plant.

[0066] Data cleaning includes using the mean replacement method for adjacent time periods to handle missing values, and using cubic spline interpolation to supplement data with missing values ​​for more than 3 consecutive days; identifying outliers using the "3σ criterion," correcting them by combining historical data from the same period, and removing invalid data with extreme anomalies (such as zero power generation records due to equipment failure). Data standardization includes zero-mean standardization of feature data of different dimensions. Feature construction includes constructing time-series features (such as weekly, monthly, and quarterly statistics), trend features (such as the 7-day consecutive power generation growth rate), and correlation features (such as the correlation coefficient between power generation and load demand) based on the original data; and performing one-hot encoding on classification features such as technical routes. Data partitioning includes dividing the preprocessed data into a training set (70%), a validation set (20%), and a test set (10%) in time series order for model training and performance evaluation.

[0067] Step S220: Input the historical time series dataset into the first time series prediction model to obtain the first power generation prediction sequence for future time periods; the first time series prediction model is a deep learning model using an inverted attention mechanism.

[0068] The first time-series prediction model can be the iTransformer model. This model accurately captures the relationships between multiple variables and long-term time-series dependencies by optimizing model dimensionality mapping, layer normalization objects, and attention mechanism adaptability. At the same time, it relies on feedforward networks to deeply mine univariate time-series features to achieve high-quality predictions.

[0069] In this embodiment of the application, a historical time-series dataset is input into a first time-series prediction model, which can output a power generation prediction sequence for the next X days (the first power generation prediction sequence for the future time period). For example, X is the prediction step size, which is 30-90 days or 180-365 days.

[0070] Step S230: Input the historical equipment availability sequence and historical power load demand sequence from the historical time series dataset into the second time series prediction model to obtain the equipment availability sequence and power load demand sequence for the future time period. Based on the equipment availability sequence and power load demand sequence, obtain the second power generation prediction sequence for the future time period. The second time series prediction model is a statistical model using an additive decomposition framework.

[0071] The second time series prediction model can be the Prophet model, which decomposes time series data into four independent components: trend, seasonal, holiday / abnormal event, and error. By overlaying these components, it achieves accurate modeling of complex time series patterns.

[0072] In this embodiment of the application, the historical equipment availability sequence and the historical power load demand sequence are used as inputs to the second time series prediction model, and the second time series prediction model outputs the equipment availability sequence and the power load demand sequence for the next X days.

[0073] Based on the equipment availability sequence and the preset rated power of the nuclear power plant, the available power sequence is obtained; the daily minimum power in the available power sequence and the power load demand sequence is determined; based on the daily minimum power and the future time period, the second power generation forecast sequence for the future time period is obtained.

[0074] For example, if the equipment availability rate on a certain day is 90%, and the preset rated power of the nuclear power plant is 100MW, then the available power is 90MW. If the power load demand during a certain period of the day is 80MW, then the power generation during that period is 80MWh. The power generation during all periods of the day is added together to obtain the predicted power generation value for that day.

[0075] Step S240: Input the historical time series dataset, equipment availability sequence, and power load demand sequence into the third time series prediction model to obtain the third power generation prediction sequence for future time periods; the third time series prediction model is a machine learning model that adopts a forward distributed learning strategy.

[0076] Among them, the third time series prediction model is the XGBoost model, which adopts the Gradient Boosting Decision Tree (GBDT) ensemble framework. It gradually optimizes the objective function through a forward distributed learning strategy to model the complex nonlinear relationship between multiple features and power generation.

[0077] In this embodiment of the application, the historical time series dataset, the equipment availability sequence, and the power load demand sequence are used as inputs to the third time series prediction model. The third time series prediction model outputs the power generation prediction sequence for the next X days (the third power generation prediction sequence for the future time period).

[0078] Step S250: Based on the first power generation prediction sequence, the second power generation prediction sequence, and the third power generation prediction sequence, obtain the target power generation prediction sequence for each nuclear power plant.

[0079] In this embodiment of the application, the first power generation prediction sequence, the second power generation prediction sequence, and the third power generation prediction sequence are fused to obtain the power generation prediction sequence (target power generation prediction sequence) for each nuclear power plant in the next X days.

[0080] Step S260: According to the preset area and preset period, summarize the target power generation prediction sequence of each nuclear power plant to obtain multi-level and multi-period power generation prediction values; the preset area includes each nuclear power plant.

[0081] The pre-defined regions include the geographical regions to which each nuclear power plant belongs, the cities that merge the geographical regions according to their administrative affiliation, and the provinces that merge the cities.

[0082] The preset periods include monthly, quarterly, and annual periods.

[0083] In this embodiment of the application, the target power generation prediction sequences of each basic unit within a geographical area can be merged to obtain the power generation prediction sequence of each geographical area. Here, a basic unit refers to a combination of adjacent nuclear power plants. The power generation prediction sequences of each geographical area are then summarized monthly, quarterly, and annually to obtain the monthly, quarterly, and annual power generation prediction values ​​for each geographical area.

[0084] Based on administrative affiliation, the power generation forecast sequences of various geographical areas within the same prefecture-level city are merged to obtain the power generation forecast sequences of each prefecture-level city. The power generation forecast sequences of each prefecture-level city are then aggregated monthly, quarterly, and annually to obtain the monthly, quarterly, and annual power generation forecast values ​​for each prefecture-level city.

[0085] Based on administrative affiliation, the power generation forecast sequences of all cities within the same province are merged to obtain the power generation forecast sequence for each province. The power generation forecast sequences of each province are then aggregated monthly, quarterly, and annually to obtain the monthly, quarterly, and annual power generation forecast values ​​for each province.

[0086] The aforementioned multi-level, multi-cycle nuclear power generation prediction method based on multi-model fusion obtains historical time-series datasets from each nuclear power plant. These datasets are preprocessed from historical integrated operational data of each plant. The historical time-series datasets are then input into a first time-series prediction model to obtain a first generation prediction sequence for future time periods. This first time-series prediction model is a deep learning model employing an inverted attention mechanism. The historical equipment availability sequence and historical power load demand sequence from the historical time-series datasets are then input into a second time-series prediction model to obtain a future equipment availability sequence and a future power load demand sequence. Based on these sequences, a second generation prediction sequence for future time periods is obtained. This second time-series prediction model is a deep learning model employing an inverted attention mechanism. A statistical model using an additive decomposition framework is employed. Historical time-series datasets, equipment availability sequences, and electricity load demand sequences are input into a third time-series prediction model to obtain a third power generation prediction sequence for future time periods. This third time-series prediction model is a machine learning model employing a forward distributed learning strategy. Based on the first, second, and third power generation prediction sequences, target power generation prediction sequences for each nuclear power plant are obtained. The target power generation prediction sequences for each nuclear power plant are then aggregated according to a preset region and preset period to obtain multi-level, multi-period power generation prediction values. The preset region includes all nuclear power plants. Thus, this application captures the long-term dependencies of historical time-series datasets using the first time-series prediction model, avoiding a significant increase in prediction error with increasing time span. The second time-series prediction model performs independent univariate prediction on regular data with significant periodicity and trends in the historical time-series dataset, fully utilizing its adaptive modeling capabilities for seasonality and holiday effects to improve the prediction accuracy of regular data. By using a third-time-series forecasting model, the impact of equipment availability and power load demand on power generation is considered, thus avoiding a disconnect between forecast results and actual operating scenarios. This enables accurate forecasting of nuclear power generation at multiple levels, from nuclear power plants to regions, cities, and provinces, and from monthly to quarterly and annual cycles.

[0087] In one embodiment, the first time-series prediction model includes an input layer, a self-attention layer, a feedforward network layer, and an output layer. Historical time-series datasets are input into the first time-series prediction model to obtain a first power generation prediction sequence for future time periods, including:

[0088] Step S221: Through the input layer, the historical time series dataset is transposed and embedded with features to obtain the first feature vector sequence.

[0089] Unlike traditional Transformer models that aggregate multiple variables at the same time point into Temporal Tokens, the iTransformer model (the first time-series prediction model) innovatively adopts a "dimensionality inversion" modeling approach. It independently maps the entire time-series sequence of each variable to a Variable Token, and applies layer normalization to the sequence representation of a single variable rather than the traditional representation of multiple variables at the same time point. This design effectively reduces interference from differences in measurement scales of different variables, improving the model's adaptability to non-stationary time-series data. Furthermore, it naturally and deeply uncovers the intrinsic relationships between multiple variables, aligning with the scenario in nuclear power generation prediction where "installed capacity, load demand, and other variables have a long-term synergistic impact on power generation." It fully couples the correlation characteristics between the operating data of each nuclear power plant and external influencing factors, avoiding noise generated by the interaction of time-disaligned variables in traditional modeling.

[0090] In this embodiment, based on preprocessed daily time-series data from the past 5 years or more, an input layer is constructed through a two-step process of "data transposition-feature embedding". First, the original time-series data... ( For the length of the historical time series, (for the number of variables) transpose This approach ensures that each row vector corresponds to a complete time-series sequence of a single variable. The core idea is to treat the historical power generation sequences and regional load demand sequences of each nuclear power plant as independent Variation Tokens, while integrating multi-dimensional key features to form a complete input set. This includes: past daily power generation, inherent and planned installed capacity (segmented and labeled according to commissioning time), equipment availability (utilization hours), and uniquely coded technology routes (such as "Hualong One" and "Guohe One"). Subsequently, an embedding layer (implemented using a 2-layer MLP) is used to embed each Variation Token (dimension...) Mapped to fixed dimensions The vector representation of the final input dimension (the final input dimension is...) This ensures that time-series data from different variables and nuclear power plants have a unified feature dimension, providing a foundation for subsequent attention calculations. The input sequence is ultimately constructed as follows: (First eigenvector sequence), where, The total number of variables (including operating variables of each nuclear power plant and external influence variables), each for 3D feature vectors ( The value can be either 256 or 512, adaptively selected based on the number of variables, and ≥1825 (i.e., more than 5 years of daily data) to ensure long-term dependent capture effect.

[0091] Step S222: Through the self-attention layer, feature extraction is performed on the first feature vector sequence to obtain the second feature vector sequence.

[0092] In this embodiment, the self-attention mechanism of the self-attention layer adopts the scaling dot product attention mechanism to model the correlation between different Variate Tokens. The core is to calculate the attention weight of each variable time series to quantify the influence strength between variables, thereby accurately capturing key correlations such as "installed capacity-power generation", "load demand-power generation", and "coordination of power generation of different nuclear power plants".

[0093] The specific process is as follows: input feature matrix The query matrix is ​​generated through three independent linear projection layers. Key matrix Value matrix ( , (The number of attention heads is 8); calculate the attention score matrix. (dimension) ), elements in the matrix Characterizing the first The variable for the first The dependence strength of each variable. After normalizing the attention score using the Softmax function, it is compared with... Weighted summation yields the attention output, enabling the fusion of multivariate features. The attention score matrix learned by this self-attention mechanism intuitively reflects the correlation patterns among multiple variables and possesses good interpretability.

[0094] Step S223: Extract features from the second feature vector sequence through a feedforward network layer to obtain the third feature vector sequence.

[0095] In this embodiment, a feed-forward network layer is embedded within each Varitae Token to focus on the intrinsic features of a single variable time series through deep learning, compensating for the shortcomings of attention mechanisms in capturing long-term trends of single variables. The feed-forward network (FFN) layer adopts a classic structure of "linear transformation-activation-linear transformation-residual connection," specifically: first, the input feature dimension is mapped from D to 4D (hidden layer dimension) through a linear layer, and the ReLU activation function is used to enhance the model's ability to fit nonlinear features; then, the dimension is mapped back to D through a linear layer, and residual connections are made with the input features to improve the model's training stability. This network can accurately mine three core features of single-variable time series: long-term trends (such as the steady increase in power generation due to installed capacity growth, and the performance degradation trend within the equipment's life cycle), periodic features (such as power generation fluctuations caused by quarterly maintenance, and power generation adjustments caused by annual load changes), and abnormal patterns (such as short-term power output fluctuations caused by extreme weather). By stacking L blocks, FFN works in synergy with the attention mechanism to achieve the dual goals of "deep extraction of univariate temporal features + fusion of multivariate correlation features".

[0096] Step S224: Through the output layer, the third feature vector sequence is linearly mapped and transposed to obtain the first power generation prediction sequence for the future time period.

[0097] In this embodiment, the output layer adopts a "linear mapping-dimensional transpose" structure to convert the high-dimensional features learned by the feedforward network into predicted future power generation values. First, the output of L blocks is processed through a mapping layer (one linear layer). The feature matrix is ​​mapped to ( To predict the step size (i.e., 30-90 days or 180-365 days), the predicted sequences for each variable are obtained; then, the prediction matrix is ​​transposed to obtain... The prediction results are given, with each row corresponding to the predicted value of each variable for a future day. Finally, the power generation variable sequence is extracted as the output of the iTransformer model. .

[0098] In the training phase of the time series model, a sliding window method is used to divide the samples (the window length is set to 90 days to cover quarterly periodic features). The mean absolute error (MAE) is used as the loss function, and the Adam optimizer (with an initial learning rate of 5×10^(-4)∼10^(-3)) is used to iteratively optimize the model parameters. The learnable parameters of the embedding layer, attention mechanism, feedforward network and mapping layer are adjusted through backpropagation. At the same time, an early stopping strategy is adopted (training is stopped if the root mean square error (MAE) of the validation set does not decrease for 10 consecutive rounds) to avoid overfitting and ensure that the model converges to the optimal state.

[0099] This application utilizes the inverted Transformer (iTransformer) model, taking daily time-series data from the past five years or more as input, and integrates all features such as past daily power generation, installed capacity, equipment availability (utilization hours), technology routes (classification), and power load demand. By strengthening the capture of long-term time-series dependencies through the inverted attention mechanism, it can achieve the prediction of power generation on the next X days.

[0100] In one embodiment, the historical equipment availability sequence and historical power load demand sequence from the historical time-series dataset are input into a second time-series prediction model to obtain the equipment availability sequence and the power load demand sequence for future time periods, including:

[0101] Step S231: Using the second time-series prediction model, both the historical equipment availability series and the historical power load demand series are decomposed into trend terms, seasonal terms, holiday terms, and error terms.

[0102] In the embodiments of this application, the trend item A piecewise linear trend model is adopted, which can flexibly capture trend abrupt changes and is well-suited to the sudden changes in availability / load demand caused by equipment maintenance, policy adjustments, and changes in installed capacity in nuclear power scenarios. Its core is to achieve continuous splicing of linear trends from different time periods by pre-setting or automatically identifying trend abrupt change points. The formula for the trend term is:

[0103]

[0104] in, The base growth rate characterizes the core rate of change in the absence of mutations. This is the offset, corresponding to the baseline value at the initial moment; These are trend inflection points (which can be manually specified by analysts as known times such as equipment maintenance or policy releases, or automatically filtered through sparse priors). The growth rate adjustment at the mutation point is obtained through Bayesian inference optimization to ensure that the trend curve before and after the mutation is continuous and smooth.

[0105] Seasonal items Fourier series fitting is employed for multi-period seasonality. Fourier series can flexibly approximate arbitrarily smooth periodic functions, adapting to annual (e.g., peak summer cooling, winter heating load) and quarterly periodic fluctuations in electricity load demand, as well as quarterly maintenance-related periodic changes in equipment availability. The formula for the seasonal term is:

[0106]

[0107] in, For the period (annual cycle) Adaptable to annual load fluctuations; quarterly cycle (Adapting to availability fluctuations caused by quarterly maintenance). The number of Fourier terms (annual cycle) It can accurately fit complex seasonal fluctuations within a year; quarterly cycles (balancing fitting accuracy with the risk of overfitting) The value can be automatically optimized using the AIC criterion; , For seasonal parameters, normal smoothing prior constraints are used to avoid overfitting short-term random fluctuations and improve the generalization ability of seasonal models.

[0108] Holidays / Abnormal Events This refers to modeling sudden, non-periodic events such as extreme weather (high temperatures, heavy rain, typhoons, etc. affecting equipment operation and load demand), planned equipment maintenance, and major holidays (such as load fluctuations during Spring Festival and National Day) using 0-1 variable labeling. It also supports setting event impact windows to capture cascading fluctuations before and after the event. The formula for holiday / abnormal event items is:

[0109]

[0110] in, The event is marked (1 if the event occurs at time t, 0 otherwise; if a window is set, the weight can be gradually set to 0.5-1 for 1-3 days before and after the event to represent the decay process of the influence). The parameter for the intensity of event impact is constrained by normal priors to avoid excessive interference of a single event on the prediction results. Parameters can be set independently for different types of events (such as extreme weather and equipment maintenance) to accurately distinguish the degree of impact of different events.

[0111] Error term This represents unique fluctuations not captured by the model. The default assumption is that the data follows a normal distribution. This assumption can be verified and adjusted through residual analysis to adapt to the fluctuation characteristics of nuclear power-related data.

[0112] Step S232: Overlay the trend term, seasonal term, holiday term, and error term to obtain the equipment availability sequence and the power load demand sequence for the future time period.

[0113] In this embodiment of the application, the calculation formula for the second time-series prediction model is as follows:

[0114]

[0115] Specifically, by overlaying trend terms, seasonal terms, holiday terms, and error terms, a forecast for the next X days is obtained, including the equipment availability sequence. and electricity load demand sequence Meanwhile, the model can output independent prediction curves for each component, allowing analysts to intuitively judge the contribution of each factor to the prediction results, which is convenient for subsequent model optimization.

[0116] In the training phase of the second time-series prediction model, daily data of equipment availability and daily data of electricity load demand from the past five years are used as inputs. The L-BFGS optimization algorithm is employed to solve for the maximum a posteriori estimate, and Bayesian inference is combined to adjust for trend inflection points. Seasonal parameters and Event impact parameters Imposing prior constraints balances model fitting accuracy and generalization ability. To improve training reliability, outlier removal (such as a sudden drop in availability due to equipment failure) of the input sequence is necessary before training to ensure training data quality.

[0117] This application employs the Prophet model to independently predict regular data with significant periodicity and trends, such as equipment availability and power load demand, for the next X days. It fully utilizes Prophet's adaptive modeling capabilities for seasonality and holiday effects to improve the prediction accuracy of regular variables.

[0118] In one embodiment, historical time-series datasets, equipment availability sequences, and electricity load demand sequences are input into a third time-series forecast model to obtain a third generation forecast sequence for future time periods, including:

[0119] Step S231: Input the historical time series dataset into the third time series prediction model to obtain the initial power generation prediction sequence for future time periods;

[0120] Step S232: Using the equipment availability sequence and the power load demand sequence as constraints, the initial power generation prediction sequence for the future time period is corrected to obtain the third power generation prediction sequence for the future time period.

[0121] In this embodiment, a multi-level hard constraint correction mechanism is constructed to ensure that the prediction results conform to actual operating rules and policy requirements: First, the policy quota (annual / quarterly maximum allowable power generation) is decomposed into daily constraint thresholds according to time proportions (annual quotas are allocated according to the proportion of monthly days, and quarterly quotas are refined according to the proportion of daily load), and the inter-regional power transmission plan is transformed into a daily transmission capacity upper limit constraint; the preliminary prediction results output by the XGBoost model are then processed. Daily constraint verification is performed. If the daily forecast value exceeds the policy quota threshold or the transmission capacity limit, it is adjusted according to the constraint boundary value (the boundary value is used when it exceeds the upper limit, and the lower limit is used when it is below the technical lower limit, such as when the equipment is at its minimum output). For sudden changes in forecast values ​​over multiple consecutive days, a smoothing correction strategy is introduced, using a moving average of three adjacent days to correct the adjusted forecast value, so as to avoid excessive fluctuations in power generation that could affect grid stability. Finally, the corrected power generation forecast sequence is output. This ensures that the prediction results are both accurate and feasible for engineering purposes.

[0122] In the training phase of the third time-series prediction model, the objective function consists of a training loss term and a regularization term, balancing fitting accuracy and model simplicity. The specific formula is as follows:

[0123]

[0124] in, To train the loss term, the Mean Absolute Error (MAE) loss function is selected, which is suitable for continuous power generation prediction scenarios and has stronger robustness to extreme outliers. The expression is as follows: ; This is a regularization term used to suppress overfitting, and it includes a penalty for the complexity of the tree structure. ( This represents the number of leaf nodes in the current decision tree. The penalty coefficient for the number of leaf nodes. This is the weight decay coefficient. For the first The output weights of each leaf node are used. During model training, the objective function is approximated by second-order Taylor expansion to quickly solve for the optimal decision tree structure. At the same time, a histogram optimization algorithm is used to accelerate the feature splitting process and improve training efficiency. The key hyperparameters are set as follows: maximum tree depth 6-8 layers (to balance fitting ability and overfitting risk), learning rate 0.05-0.1 (to control the contribution weight of each tree), and subsample ratio 0.8 (to randomly sample training data to enhance generalization).

[0125] This application is based on the XGBoost model, integrates the regularity variable prediction results of Prophet, and combines data such as installed capacity (inherent + planned), technical route, past daily power generation, and extreme events to predict power generation; at the same time, it introduces equipment availability and power load demand as external constraints, and outputs corrected power generation prediction values.

[0126] In one embodiment, obtaining the target power generation prediction sequence for each nuclear power plant based on the first power generation prediction sequence, the second power generation prediction sequence, and the third power generation prediction sequence includes:

[0127] Step S251: Obtain the first historical average absolute error of the first time series prediction model, the second historical average absolute error of the second time series prediction model, and the third historical average absolute error of the third time series prediction model.

[0128] In this embodiment of the application, the validation set data is input into the first time series prediction model, and the first historical average absolute error of the first time series prediction model is calculated based on the output power generation prediction value and the corresponding tag data.

[0129] The validation set data is input into the second time series prediction model. Based on the output equipment availability prediction value and power load demand prediction value, the power generation prediction value is obtained. Based on the power generation prediction value and the corresponding label data, the second historical mean absolute error of the second time series prediction model is calculated.

[0130] The validation set data is input into the third time series prediction model. Based on the output power generation prediction value and the corresponding tag data, the third historical mean absolute error of the third time series prediction model is calculated.

[0131] Step S252: Based on the first historical average absolute error, the second historical average absolute error, and the third historical average absolute error, obtain the first weight of the first time series prediction model, the second weight of the second time series prediction model, and the third weight of the third time series prediction model.

[0132] In this embodiment, the weights are allocated according to the reciprocal of the mean absolute error (MAE), as shown in the formula:

[0133]

[0134] in, These correspond to the iTransformer, Prophet, and XGBoost models, respectively.

[0135] In an optional embodiment, after determining the first, second, and third weights, the weights can be adjusted based on the number of days in the future time period. Specifically, for X = 30-90 days, the second weight corresponding to the Prophet model is increased because short-term regularity characteristics are more significant. For X = 180-365 days, the first weight corresponding to the iTransformer model is increased because long-term, long-period dependencies are more critical.

[0136] Step S253: Based on the first weight, the second weight, and the third weight, the first power generation prediction sequence, the second power generation prediction sequence, and the third power generation prediction sequence are fused to obtain the fused power generation prediction sequence.

[0137] In this embodiment of the application, the expression for the fused power generation prediction sequence is:

[0138]

[0139] in, Indicates the first weight. Indicates the second weight. Indicates the third weight. This represents the first power generation forecast sequence. This represents the second power generation forecast sequence. This represents the third power generation forecast sequence.

[0140] Step S254: Obtain the first historical prediction error of the first time series prediction model, the second historical prediction error of the second time series prediction model, and the third historical prediction error of the third time series prediction model.

[0141] In this embodiment of the application, the validation set data is input into the first time series prediction model, and the first historical prediction error of the first time series prediction model is obtained based on the difference between the output power generation prediction value and the corresponding tag data.

[0142] The validation set data is input into the second time series prediction model. Based on the output equipment availability prediction value and power load demand prediction value, the power generation prediction value is obtained. Based on the difference between the power generation prediction value and the corresponding tag data, the second historical prediction error of the second time series prediction model is obtained.

[0143] The validation set data is input into the third time series prediction model. Based on the output power generation prediction value and the error value of the corresponding tag data, the third historical prediction error of the third time series prediction model is obtained.

[0144] Step S255: Obtain the mean error based on the first historical prediction error, the second historical prediction error, and the third historical prediction error.

[0145] In this embodiment of the application, a normal error distribution model is constructed based on the first historical prediction error, the second historical prediction error, and the third historical prediction error to obtain the mean error of the normal error distribution model.

[0146] Step S256: Based on the mean error, the merged power generation prediction sequence is corrected to obtain the target power generation prediction sequence for each nuclear power plant.

[0147] In this embodiment of the application, the expression for the target power generation prediction sequence of each nuclear power plant is as follows:

[0148]

[0149] in, This represents the mean error.

[0150] The embodiments of this application combine historical mean absolute error (MAE) to adjust the fusion weights of the iTransformer, Prophet and XGBoost models, and correct the fusion results through a normal error distribution model to improve prediction accuracy.

[0151] In one embodiment, the multi-level, multi-cycle power generation forecast includes monthly, quarterly, and annual power generation forecasts for various cities and provinces. After obtaining the multi-level, multi-cycle power generation forecasts, the method further includes:

[0152] Step S310: Obtain the monthly power generation quotas for each city and each province; if the deviation between the predicted monthly power generation value and the monthly power generation quota for each city is greater than a first deviation threshold, or if the deviation between the predicted monthly power generation value and the monthly power generation quota for each province is greater than the first deviation threshold, perform a retrospective adjustment on the power generation prediction sequence for each nuclear power plant; and / or,

[0153] The first deviation threshold can be set according to actual needs.

[0154] The monthly power generation quota refers to the maximum allowed power generation per month.

[0155] In this embodiment of the application, the monthly power generation quota of each city can be obtained through the monthly operation briefing of each city, and the monthly power generation quota of each province can be obtained through the monthly operation briefing of each province.

[0156] If the deviation between the monthly power generation forecast of each city and the monthly power generation quota of each city is greater than the first deviation threshold, or if the deviation between the monthly power generation forecast of each province and the monthly power generation quota of each province is greater than the first deviation threshold, the model parameters of the first time series forecast model, the second time series forecast model and the third time series forecast model are adjusted to retrospectively adjust the power generation forecast sequence of each nuclear power plant.

[0157] Step S320: Obtain the quarterly power generation quotas for each city and each province; if the deviation between the quarterly power generation forecast for each city and its quarterly power generation quota exceeds a second deviation threshold, or if the deviation between the quarterly power generation forecast for each province and its quarterly power generation quota exceeds a second deviation threshold, perform a retrospective adjustment to the power generation forecast sequence for each nuclear power plant; and / or,

[0158] The second deviation threshold can be set according to actual needs.

[0159] The quarterly power generation quota refers to the maximum allowed power generation per quarter.

[0160] In this embodiment of the application, the quarterly power generation quota of each city can be obtained through the quarterly operation briefing of each city, and the quarterly power generation quota of each province can be obtained through the quarterly operation briefing of each province.

[0161] If the deviation between the quarterly power generation forecast of each city and the quarterly power generation quota of each city is greater than the second deviation threshold, or if the deviation between the quarterly power generation forecast of each province and the quarterly power generation quota of each province is greater than the second deviation threshold, the model parameters of the first time series forecast model, the second time series forecast model, and the third time series forecast model will be adjusted to retrospectively adjust the power generation forecast sequence of each nuclear power plant.

[0162] Step S330: Obtain the annual power generation quotas for each city and each province; if the deviation between the annual power generation forecast value of each city and the annual power generation quota of each city is greater than the third deviation threshold, or if the deviation between the annual power generation forecast value of each province and the annual power generation quota of each province is greater than the third deviation threshold, backtrack and adjust the power generation forecast sequence of each nuclear power plant.

[0163] The third deviation threshold can be set according to actual needs.

[0164] The annual power generation quota refers to the maximum permitted power generation per year.

[0165] In this embodiment of the application, the annual power generation quota of each city can be obtained through the annual operation briefing of each city, and the annual power generation quota of each province can be obtained through the annual operation briefing of each province.

[0166] If the deviation between the annual power generation forecast of each city and the annual power generation quota of each city is greater than the third deviation threshold, or if the deviation between the annual power generation forecast of each province and the annual power generation quota of each province is greater than the third deviation threshold, the model parameters of the first time series forecast model, the second time series forecast model, and the third time series forecast model will be adjusted to retrospectively adjust the power generation forecast sequence of each nuclear power plant.

[0167] The embodiments of this application can dynamically optimize the prediction results and improve the accuracy of multi-level and multi-cycle power generation prediction values ​​by performing hierarchical consistency checks on monthly, quarterly and annual data.

[0168] In one embodiment, the method further includes:

[0169] Step S410: If the number of days corresponding to the future time period is less than or equal to a preset number, update the parameters of the first time series prediction model, the second time series prediction model, and the third time series prediction model at intervals of the first time period.

[0170] The preset quantity can be set according to actual needs. Specifically, the preset quantity is 90.

[0171] In this embodiment of the application, when X≤90, the data and model parameters are updated every 7 days. The latest 7 days' actual power generation, load demand and other data are added to the training set, the model is retrained and the prediction results for the remaining time are corrected.

[0172] Step S420: If the number of days corresponding to the future time period is greater than the preset number, update the parameters of the first time series prediction model, the second time series prediction model, and the third time series prediction model at second time intervals.

[0173] In this embodiment of the application, when X > 90, the data and model parameters are updated every 30 days, the latest 30 days of actual running data are added, the model weights and error distribution model parameters are adjusted, and the remaining duration prediction results are dynamically corrected.

[0174] Step S430: In the event of a target event, update the parameters of the first time series prediction model, the second time series prediction model, and the third time series prediction model; the target event includes at least one of a major policy adjustment event, an extreme weather event, and a major overhaul event of nuclear power plant equipment.

[0175] Among them, major policy adjustment events refer to events in which the government or relevant regulatory agencies make significant modifications or adjustments to energy policies, environmental protection policies, tax policies, etc.

[0176] Extreme weather events refer to weather phenomena caused by abnormalities in the natural climate system that have a significant impact on human society, economy, and environment.

[0177] Among them, major maintenance events of nuclear power plant equipment refer to events involving large-scale, long-term maintenance and repair of key equipment in nuclear power plants.

[0178] In this embodiment, when major policy adjustments, extreme weather events, or major overhauls of nuclear power plant equipment occur, the model parameters are updated in real time to ensure that the model can quickly adapt to sudden scenarios.

[0179] This application adopts a differentiated rolling update strategy, updating data and model parameters every 7 days when X≤90 days and every 30 days when X>90 days. This can dynamically correct the remaining duration prediction results and improve the prediction accuracy of multi-level and multi-cycle power generation prediction values.

[0180] To facilitate understanding of the above method embodiments, as follows: Figure 3The diagram illustrates a flowchart of a multi-level, multi-cycle nuclear power generation prediction method based on multi-model fusion. The method includes: Step S1, data acquisition and processing; Step S2, multi-module fusion power generation prediction; Step S3, hierarchical merging; Step S4, consistency verification; and Step S5, rolling update. Through these steps, monthly, quarterly, and annual nuclear power generation prediction results can be obtained for basic regions, cities, and provinces.

[0181] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0182] Based on the same inventive concept, this application also provides a device for predicting multi-level, multi-cycle nuclear power generation based on multi-model fusion, which is used to implement the multi-level, multi-cycle nuclear power generation prediction method based on multi-model fusion described above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the device for predicting multi-level, multi-cycle nuclear power generation based on multi-model fusion provided below can be found in the limitations of the multi-level, multi-cycle nuclear power generation prediction method based on multi-model fusion described above, and will not be repeated here.

[0183] In one exemplary embodiment, please refer to Figure 4 A multi-level, multi-cycle nuclear power generation prediction device based on multi-model fusion is provided. The device includes:

[0184] The historical time-series dataset acquisition module 410 is used to acquire the historical time-series dataset of each nuclear power plant; the historical time-series dataset is obtained by preprocessing the historical comprehensive operation data of each nuclear power plant.

[0185] The first power generation prediction sequence acquisition module 420 is used to input the historical time series dataset into the first time series prediction model to obtain the first power generation prediction sequence for future time periods; the first time series prediction model is a deep learning model using an inverted attention mechanism.

[0186] The second power generation prediction sequence acquisition module 430 is used to input the historical equipment availability sequence and historical power load demand sequence from the historical time series dataset into the second time series prediction model to obtain the equipment availability sequence and power load demand sequence for the future time period. Based on the equipment availability sequence and power load demand sequence, the second power generation prediction sequence for the future time period is obtained. The second time series prediction model is a statistical model using an additive decomposition framework.

[0187] The third power generation prediction sequence acquisition module 440 is used to input historical time series datasets, equipment availability sequences, and power load demand sequences into the third time series prediction model to obtain the third power generation prediction sequence for future time periods; the third time series prediction model is a machine learning model that adopts a forward distributed learning strategy.

[0188] The target power generation prediction sequence acquisition module 450 is used to obtain the target power generation prediction sequence for each nuclear power plant based on the first power generation prediction sequence, the second power generation prediction sequence, and the third power generation prediction sequence.

[0189] The multi-level, multi-cycle power generation prediction module 460 is used to summarize the target power generation prediction sequence of each nuclear power plant according to a preset region and a preset cycle to obtain multi-level, multi-cycle power generation prediction values; the preset region includes each nuclear power plant.

[0190] The modules in the aforementioned multi-model fusion-based multi-level, multi-cycle nuclear power generation prediction device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0191] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 5As shown, the computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When the computer program is executed by the processor, it implements a multi-level, multi-cycle nuclear power generation prediction method based on multi-model fusion. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0192] Those skilled in the art will understand that Figure 5 The structures shown are merely block diagrams of some structures related to the present application and do not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the aforementioned multi-level, multi-cycle nuclear power generation prediction method based on multi-model fusion. The steps of the multi-level, multi-cycle nuclear power generation prediction method based on multi-model fusion described above can be steps from the multi-level, multi-cycle nuclear power generation prediction method based on multi-model fusion described in the various embodiments above.

[0193] In one embodiment, a computer-readable storage medium is provided, storing a computer program that, when executed by a processor, causes the processor to perform the steps of the multi-level, multi-cycle nuclear power generation prediction method based on multi-model fusion described above. The steps of the multi-level, multi-cycle nuclear power generation prediction method based on multi-model fusion described here can be the steps in one of the multi-level, multi-cycle nuclear power generation prediction methods based on multi-model fusion described in the various embodiments above.

[0194] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, causes the processor to perform the steps of the multi-level, multi-cycle nuclear power generation prediction method based on multi-model fusion described above. The steps of the multi-level, multi-cycle nuclear power generation prediction method based on multi-model fusion described here can be the steps in the multi-level, multi-cycle nuclear power generation prediction method based on multi-model fusion described in the various embodiments above.

[0195] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0196] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0197] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0198] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A multi-level, multi-cycle nuclear power generation prediction method based on multi-model fusion, characterized in that, The method includes: Historical time-series datasets for each nuclear power plant are obtained; these historical time-series datasets are obtained by preprocessing historical integrated operational data of each nuclear power plant. The historical time series dataset is input into the first time series prediction model to obtain the first power generation prediction sequence for future time periods; the first time series prediction model is a deep learning model using an inverted attention mechanism. The historical equipment availability sequence and historical power load demand sequence from the historical time series dataset are input into the second time series prediction model to obtain the equipment availability sequence and the power load demand sequence for the future time period. Based on the equipment availability sequence and the power load demand sequence, the second power generation prediction sequence for the future time period is obtained. The second time series prediction model is a statistical model using an additive decomposition framework. The historical time series dataset, the equipment availability sequence, and the power load demand sequence are input into the third time series prediction model to obtain the third power generation prediction sequence for future time periods; the third time series prediction model is a machine learning model that adopts a forward distributed learning strategy. Based on the first power generation prediction sequence, the second power generation prediction sequence, and the third power generation prediction sequence, a target power generation prediction sequence for each of the nuclear power plants is obtained; According to the preset area and preset period, the target power generation prediction sequence of each nuclear power plant is summarized to obtain multi-level and multi-period power generation prediction values; the preset area includes each nuclear power plant.

2. The method according to claim 1, characterized in that, The first time-series prediction model includes an input layer, a self-attention layer, a feedforward network layer, and an output layer. The step of inputting the historical time-series dataset into the first time-series prediction model to obtain a first power generation prediction sequence for a future time period includes: The input layer is used to perform dimensionality transpose and feature embedding on the historical time series dataset to obtain a first feature vector sequence. The self-attention layer is used to extract features from the first feature vector sequence to obtain the second feature vector sequence. The feedforward network layer is used to extract features from the second feature vector sequence to obtain the third feature vector sequence. The output layer performs linear mapping and dimension transpose on the third feature vector sequence to obtain the first power generation prediction sequence for the future time period.

3. The method according to claim 1, characterized in that, The step of inputting the historical equipment availability sequence and historical power load demand sequence from the historical time series dataset into the second time series prediction model to obtain the equipment availability sequence and power load demand sequence for future time periods includes: The second time-series prediction model decomposes both the historical equipment availability sequence and the historical power load demand sequence into trend terms, seasonal terms, holiday terms, and error terms. The trend term, the seasonal term, the holiday term, and the error term are superimposed to obtain the equipment availability sequence and the power load demand sequence for the future time period.

4. The method according to claim 1, characterized in that, The step of inputting the historical time-series dataset, the equipment availability sequence, and the electricity load demand sequence into the third time-series prediction model to obtain the third power generation prediction sequence for future time periods includes: The historical time series dataset is input into the third time series prediction model to obtain the initial power generation prediction sequence for future time periods; Using the equipment availability sequence and the power load demand sequence as constraints, the initial power generation prediction sequence for the future time period is corrected to obtain the third power generation prediction sequence for the future time period.

5. The method according to claim 1, characterized in that, The step of obtaining the target power generation prediction sequence for each of the nuclear power plants based on the first power generation prediction sequence, the second power generation prediction sequence, and the third power generation prediction sequence includes: Obtain the first historical average absolute error of the first time series prediction model, the second historical average absolute error of the second time series prediction model, and the third historical average absolute error of the third time series prediction model; Based on the first historical average absolute error, the second historical average absolute error, and the third historical average absolute error, the first weight of the first time series prediction model, the second weight of the second time series prediction model, and the third weight of the third time series prediction model are obtained. Based on the first weight, the second weight, and the third weight, the first power generation prediction sequence, the second power generation prediction sequence, and the third power generation prediction sequence are fused to obtain a fused power generation prediction sequence. Obtain the first historical prediction error of the first time series prediction model, the second historical prediction error of the second time series prediction model, and the third historical prediction error of the third time series prediction model; The mean error is obtained based on the first historical prediction error, the second historical prediction error, and the third historical prediction error; Based on the mean error, the merged power generation prediction sequence is corrected to obtain the target power generation prediction sequence for each nuclear power plant.

6. The method according to any one of claims 1 to 5, characterized in that, The multi-level, multi-cycle power generation forecast includes monthly, quarterly, and annual power generation forecasts for various cities and provinces. After obtaining the multi-level, multi-cycle power generation forecasts, the method further includes: Obtain the monthly power generation quotas for each city and each province; if the deviation between the predicted monthly power generation value and the monthly power generation quota for each city is greater than a first deviation threshold, or if the deviation between the predicted monthly power generation value and the monthly power generation quota for each province is greater than the first deviation threshold, perform a retrospective adjustment on the power generation prediction sequence for each nuclear power plant; and / or, Obtain the quarterly power generation quotas for each city and each province; if the deviation between the predicted quarterly power generation value and the quarterly power generation quota for each city is greater than a second deviation threshold, or if the deviation between the predicted quarterly power generation value and the quarterly power generation quota for each province is greater than the second deviation threshold, perform a retrospective adjustment on the power generation prediction sequence for each nuclear power plant; and / or, Obtain the annual power generation quotas for each city and each province; if the deviation between the predicted annual power generation value of each city and the annual power generation quota of each city is greater than the third deviation threshold, or if the deviation between the predicted annual power generation value of each province and the annual power generation quota of each province is greater than the third deviation threshold, perform a retrospective adjustment on the power generation prediction sequence of each nuclear power plant.

7. The method according to any one of claims 1 to 5, characterized in that, The method further includes: If the number of days corresponding to the future time period is less than or equal to a preset number, the parameters of the first time series prediction model, the second time series prediction model, and the third time series prediction model are updated at intervals of the first time period. If the number of days corresponding to the future time period is greater than the preset number, the parameters of the first time series prediction model, the second time series prediction model, and the third time series prediction model are updated at second time intervals. In the event of a target event, the parameters of the first time-series prediction model, the second time-series prediction model, and the third time-series prediction model are updated; the target event includes at least one of a major policy adjustment event, an extreme weather event, and a major overhaul event of nuclear power plant equipment.

8. A multi-level, multi-cycle nuclear power generation prediction device based on multi-model fusion, characterized in that, The device includes: The historical time-series dataset acquisition module is used to acquire the historical time-series dataset of each nuclear power plant; the historical time-series dataset is obtained by preprocessing the historical comprehensive operation data of each nuclear power plant. The first power generation prediction sequence acquisition module is used to input the historical time series dataset into the first time series prediction model to obtain the first power generation prediction sequence for future time periods; the first time series prediction model is a deep learning model using an inverted attention mechanism. The second power generation prediction sequence acquisition module is used to input the historical equipment availability sequence and historical power load demand sequence from the historical time series dataset into the second time series prediction model to obtain the equipment availability sequence and power load demand sequence for the future time period, and obtain the second power generation prediction sequence for the future time period based on the equipment availability sequence and the power load demand sequence; the second time series prediction model is a statistical model using an additive decomposition framework; The third power generation prediction sequence acquisition module is used to input the historical time series dataset, the equipment availability sequence, and the power load demand sequence into the third time series prediction model to obtain the third power generation prediction sequence for future time periods; the third time series prediction model is a machine learning model that adopts a forward distributed learning strategy. The target power generation prediction sequence acquisition module is used to obtain the target power generation prediction sequence for each of the nuclear power plants based on the first power generation prediction sequence, the second power generation prediction sequence, and the third power generation prediction sequence. The multi-level, multi-cycle power generation prediction module is used to summarize the target power generation prediction sequence of each nuclear power plant according to a preset region and a preset cycle to obtain multi-level, multi-cycle power generation prediction values; the preset region includes each nuclear power plant.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images