Ozone prediction method, electronic device, and storage medium

Abstract

This application provides an ozone forecasting method, electronic device, and storage medium, belonging to the field of artificial intelligence technology. The method includes: determining an interpolated smoothed concentration matrix based on observed ozone concentrations at ground observation points; constructing a grid masking matrix based on observation points and air pressure data in a grid; generating a grid ozone concentration prediction set using an ozone concentration generation model on the smoothed concentration matrix, grid masking matrix, first time-coded information, and first meteorological ensemble forecast data; and performing an ozone ensemble forecast using an ozone forecasting model on the grid ozone concentration prediction set, grid masking matrix, second time-coded information, and second meteorological ensemble forecast data. This application combines real-time ground observation data and meteorological data through an ozone concentration generation model to predict ozone concentration without data assimilation. The ozone forecasting model uses predicted ozone concentrations and meteorological ensemble forecasts to achieve ozone ensemble forecasting, improving the efficiency and accuracy of ozone forecasting.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an ozone forecasting method, electronic device, and storage medium. Background Technology

[0002] Currently, existing ozone forecasting methods employ a Weather Research and Forecasting coupled with GEOS-Chem (WRF-GC) model, a numerical chemical transport model. Specifically, WRF-GC uses the Weather Research and Forecasting (WRF) model to simulate the meteorological field for future periods based on initial atmospheric conditions. It uses previous simulation results as the initial ozone field, or generates an initial ozone field through data assimilation (i.e., iteratively solving a cost function minimization problem (e.g., 4D-Var). This involves combining multiple data sources (e.g., weather forecast data, ground observation station data) with the background field of the WRF-GC model (e.g., the currently estimated ozone field) and iteratively correcting the shortcomings of the current estimated ozone field to generate an initial ozone field closer to reality). Based on this initial ozone field, the meteorological field-driven chemistry module (GEOS-Chem) solves continuous equations encompassing the transport, diffusion, chemical reactions, and deposition processes of pollutants (e.g., ozone) to simulate the evolution of ozone pollution, ultimately obtaining gridded ozone forecast results under different atmospheric conditions. However, in the process of generating the initial ozone field using the numerical chemical transport model employed by existing technologies, it is usually necessary to continuously iterate and optimize the generated initial ozone field through data assimilation. Each iteration requires a complete run of the high-resolution WRF-GC ozone field generation process, which makes the initial ozone field generation time-consuming, resulting in slow ozone forecasting speed and failing to meet the real-time requirements of ozone forecasting. Furthermore, since solving the evolution of ozone pollution uses a large amount of computing resources, this numerical chemical transport model is unable to have sufficient computing power to fully quantify meteorological uncertainties in the ozone forecasting process, further leading to poor accuracy of ozone forecasts. Summary of the Invention

[0003] The main objective of this application is to provide an ozone forecasting method, electronic device, and storage medium, which aim to improve the efficiency and accuracy of ozone forecasting.

[0004] To achieve the above objectives, a first aspect of this application proposes an ozone saturation forecasting method, the method comprising:

[0005] Obtain the ozone concentration at the ground observation point of the target grid, and construct the observation point concentration matrix based on the ozone concentration at the ground observation point. Then, perform interpolation processing on the observation point concentration matrix to obtain a smoothed concentration matrix.

[0006] The observation point concentration matrix is ​​masked to obtain a grid observation point masking matrix;

[0007] The air pressure data in the target grid is obtained from the preset meteorological ensemble forecast, and a grid air pressure masking matrix is ​​constructed based on the air pressure data in the target grid; wherein, each element in the grid air pressure masking matrix is ​​used to characterize the relationship between the air pressure value and the preset air pressure threshold.

[0008] Acquire the first meteorological ensemble forecast data containing multiple ensemble members of the target grid and the first time encoding information of the first meteorological ensemble forecast data; wherein, the first meteorological ensemble forecast data is meteorological field data from historical time points to the current time point;

[0009] The ozone concentration prediction set of the target grid is obtained by using a pre-trained ozone concentration generation model to generate ozone concentrations from the smoothed concentration matrix, the grid observation point masking matrix, the grid pressure masking matrix, the first time-coded information, and the first meteorological set forecast data.

[0010] Acquire the second meteorological ensemble forecast data containing multiple ensemble members of the target grid and the second time encoding information of the second meteorological ensemble forecast data; wherein, the second meteorological ensemble forecast data is meteorological field data at future time points;

[0011] The ozone ensemble forecast data for the target grid is obtained by using the fine-tuned ozone forecast model to perform ozone ensemble forecasts on the ozone concentration prediction set of the target grid, the grid observation point masking matrix, the grid pressure masking matrix, the second time-coded information, and the second meteorological ensemble forecast data.

[0012] In some embodiments, before generating ozone concentration prediction sets for the target grid by using a pre-trained ozone concentration generation model on the smoothed concentration matrix, the grid observation point masking matrix, the grid pressure masking matrix, the first time-coded information, and the first meteorological ensemble forecast data, the method further includes:

[0013] Obtain the first training set of the target grid and the original ozone concentration generation model; wherein, the first training set includes a training observation point concentration matrix, a training smoothed concentration matrix after interpolation of the training observation point concentration matrix, the current noisy ozone concentration, the target true ozone concentration, a training grid observation point masking matrix, a training grid barometric pressure masking matrix, training first meteorological forecast data, and training first time encoding information, and the training observation point concentration matrix, the current noisy ozone concentration, and the target true ozone concentration are obtained from preset simulation data;

[0014] The ozone concentration is generated by using the original ozone concentration generation model to generate ozone concentration from the training smooth concentration matrix, the training grid observation point masking matrix, the training grid pressure masking matrix, the training first meteorological forecast data, the training first time encoding information, and the current noisy ozone concentration, thus obtaining the training predicted ozone concentration of the target grid.

[0015] The concentration loss is calculated based on the predicted ozone concentration from the training and the actual target ozone concentration to obtain the concentration loss value.

[0016] The original ozone concentration generation model is trained based on the concentration loss value to obtain a pre-trained ozone concentration generation model.

[0017] Obtain the second training set of the target grid and the original ozone forecast model; wherein, the second training set includes the training grid observation point masking matrix, the training grid pressure masking matrix, training second time coding information, training second meteorological forecast data, current true ozone concentration, future noisy ozone concentration and true ozone forecast data, and the training second meteorological forecast data, the current true ozone concentration, the future noisy ozone concentration and the true ozone forecast data are obtained from the simulation data;

[0018] The original ozone forecast model is used to forecast ozone based on the training grid observation point masking matrix, the training grid pressure masking matrix, the training second meteorological forecast data, the training second time coding information, the future noisy ozone concentration, and the current real ozone concentration, to obtain the training ozone forecast data for the target grid.

[0019] The forecast loss is calculated based on the training ozone forecast data and the actual ozone forecast data to obtain the training forecast loss value.

[0020] The original ozone forecasting model is trained based on the training forecasting loss value to obtain the pre-trained ozone forecasting model.

[0021] In some embodiments, after training the original ozone forecasting model based on the training forecasting loss value to obtain a pre-trained ozone forecasting model, the method further includes:

[0022] Obtain a fine-tuning training set; wherein, the fine-tuning training set includes the training grid observation point masking matrix, the training grid barometric pressure masking matrix, fine-tuned first meteorological forecast data, fine-tuned first time encoding information, fine-tuned second meteorological forecast data, fine-tuned second time encoding information, a first-time smoothed concentration matrix and a second-time smoothed concentration matrix obtained and interpolated from the ground observation stations, and the fine-tuned first meteorological forecast data and the fine-tuned second meteorological forecast data are obtained from preset real meteorological forecast data;

[0023] The ozone concentration is generated by using the pre-trained ozone concentration generation model to generate ozone concentration from the training grid observation point masking matrix, the training grid pressure masking matrix, the fine-tuned first meteorological forecast data, the fine-tuned first time encoding information, and the first time smoothed concentration matrix, thus obtaining the initial fine-tuned ozone concentration of the target grid.

[0024] The ozone concentration is generated by using the pre-trained ozone concentration generation model to generate ozone concentration from the training grid observation point masking matrix, the training grid pressure masking matrix, the fine-tuned second meteorological forecast data, the fine-tuned second time encoding information, and the second time-smoothed concentration matrix, thus obtaining the target fine-tuned ozone concentration of the target grid.

[0025] The ozone forecast model is used to perform ozone forecasting on the training grid observation point masking matrix, the training grid pressure masking matrix, the fine-tuned second meteorological forecast data, the fine-tuned second time coding information, and the initial fine-tuned ozone concentration to obtain the fine-tuned ozone forecast data of the target grid.

[0026] The forecast loss is calculated based on the fine-tuned ozone forecast data and the target fine-tuned ozone concentration to obtain the fine-tuned forecast loss value.

[0027] The pre-trained ozone forecasting model is fine-tuned based on the fine-tuned forecasting loss value to obtain the fine-tuned ozone forecasting model.

[0028] In some embodiments, the ozone concentration prediction set for the target grid is obtained by using a pre-trained ozone concentration generation model to generate ozone concentrations from the smoothed concentration matrix, the grid observation point masking matrix, the grid pressure masking matrix, the first time-coded information, and the first meteorological ensemble forecast data, including:

[0029] The smoothed concentration matrix, the grid observation point masking matrix, the grid pressure masking matrix, and the first meteorological ensemble forecast data are fused to obtain the target fused features;

[0030] The number of diffusion steps is determined based on the ozone concentration generation model and the target fusion features.

[0031] Noise data is acquired, and the noise data is gradually diffused into the ozone concentration generation model based on the number of diffusion steps to obtain a target noisy feature map;

[0032] The target noisy feature map and the first temporal encoding information are convolved to obtain a convolutional noisy feature map;

[0033] Based on the number of diffusion steps, the target noisy feature map and the convolutional noisy feature map are subjected to stepwise denoising processing to obtain the ozone concentration prediction set of the target grid.

[0034] In some embodiments, the step of progressively diffusing the noise data into the ozone concentration generation model based on the number of diffusion steps to obtain a target noisy feature map includes:

[0035] Based on the number of diffusion steps, the noise data and the target fusion feature are fused to obtain the fused noise feature;

[0036] The fused noise features are first encoded using the ozone concentration generation model to obtain a first noisy feature map.

[0037] The first noisy feature map is downsampled to obtain a first noisy dimensionality-reduced noisy feature map.

[0038] The noise intensity of the noise data is updated based on the number of diffusion steps to obtain updated noise data. The updated noise data is then fused with the first noisy dimensionality-reduced noisy feature map based on the number of diffusion steps to obtain updated fused noise features.

[0039] The updated fused noise features and the target fused features are encoded a second time to obtain a second noisy feature map;

[0040] The second noisy feature map is downsampled to obtain the target noisy feature map.

[0041] In some embodiments, the step of first encoding the fused noise features using the ozone concentration generation model to obtain a first noisy feature map includes:

[0042] The fused noise features are convolved to obtain a fused noise feature map;

[0043] The fused noise feature map is randomly discarded to obtain a discarded feature map;

[0044] The discarded feature map is convolved to obtain the enhanced feature map;

[0045] The enhanced feature map is then normalized to obtain a normalized feature map;

[0046] The standardized feature map is activated to obtain the first noisy feature map.

[0047] In some embodiments, the stepwise denoising process performed on the target noisy feature map and the convolutional noisy feature map based on the number of diffusion steps to obtain the ozone concentration prediction set of the target grid includes:

[0048] The target noisy feature map and the convolutional noisy feature map are transposed and convolved to obtain the first upsampled feature map;

[0049] Based on the number of diffusion steps, the first upsampled feature map, the second noisy feature map, and the updated noise data are subjected to a first denoising process to obtain a first denoised feature map;

[0050] The first denoised feature map is subjected to a second transposed convolution to obtain a second upsampled feature map;

[0051] Based on the diffusion steps, the second upsampled feature map, the first noisy feature map, and the noise data are subjected to a second denoising process to obtain the ozone concentration prediction set of the target grid.

[0052] In some embodiments, the interpolation of the concentration matrix at the observation points to obtain a smoothed concentration matrix includes:

[0053] Obtain each target interpolation point of the target grid, and perform Euclidean distance calculation on each target interpolation point and the ground observation point to obtain the target distance of the interpolation observation point;

[0054] Based on the preset power parameter, the ozone concentration at the observation point, and the target distance of the interpolation observation point, ozone interpolation processing is performed on each of the target interpolation points to obtain the smoothed concentration matrix.

[0055] To achieve the above objectives, a second aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect.

[0056] To achieve the above objectives, a third aspect of the present application provides a storage medium storing a computer program that, when executed by a processor, implements the method of the first aspect described above.

[0057] The ozone forecasting method, electronic device, and storage medium proposed in this application first construct an observation point concentration matrix of ozone concentration at ground observation points. This matrix is ​​then interpolated to obtain a smoothed concentration matrix. By combining this matrix with observation point masking and constructing a grid pressure masking matrix based on the pressure data in the target grid, spatial constraints based on observation station sparsity and topographic pressure characteristics are provided for the subsequent ozone concentration generation model, thus improving the accuracy of subsequent ozone forecasts. Secondly, by utilizing the ozone concentration generation model to generate an ozone concentration set containing multiple set members from the smoothed concentration matrix, the grid observation point masking matrix, the grid pressure masking matrix, the first time-coded information, and the first meteorological ensemble forecast data, a predicted ozone concentration set containing multiple set members for the target grid is obtained. This method can directly predict ozone concentration through a generative model, eliminating the need for the time-consuming data assimilation and iteration process of traditional numerical models. This effectively saves model computational resources and improves the timeliness of ozone concentration generation, thus enhancing the timeliness of subsequent ozone forecasts. It can also promptly introduce new concentration information and respond quickly when there are significant discrepancies between ozone forecasts and actual observations. Finally, by fine-tuning the ozone forecast model, ozone ensemble forecasts are performed on the target grid's ozone concentration prediction set, grid observation point masking matrix, grid pressure masking matrix, second time-coded information, and second meteorological ensemble forecast data. This allows for direct integration of the ozone concentration prediction set from the ozone concentration generation model to perform ozone ensemble forecasts containing multiple ensemble members. This leverages the saved model computational resources to fully quantify meteorological uncertainties, thereby improving the efficiency and accuracy of ozone forecasts. Attached Figure Description

[0058] Figure 1 This is a flowchart of the ozone forecasting method provided in the embodiments of this application;

[0059] Figure 2 yes Figure 1 The flowchart of step S101 in the text;

[0060] Figure 3 This is another flowchart of the ozone forecasting method provided in the embodiments of this application;

[0061] Figure 4 This is another flowchart of the ozone forecasting method provided in the embodiments of this application;

[0062] Figure 5 yes Figure 1 The flowchart of step S105 in the process;

[0063] Figure 6 yes Figure 5 The flowchart of step S503 in the process;

[0064] Figure 7 yes Figure 6 The flowchart of step S602 in the document;

[0065] Figure 8 yes Figure 5 The flowchart of step S505 in the document;

[0066] Figure 9 This is an example diagram of the execution flow of the ozone concentration generation model provided in the embodiments of this application;

[0067] Figure 10 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0068] 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.

[0069] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0070] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0071] First, let's analyze some of the terms used in this application:

[0072] Artificial intelligence (AI) is a new branch of computer science that studies, develops, and applies theories, methods, technologies, and systems to simulate, extend, and expand human intelligence. It aims to understand the essence of intelligence and produce intelligent machines that can react in a way similar to human intelligence. Research in this field includes robotics, speech recognition, image recognition, natural language processing, and expert systems. AI can simulate the information processes of human consciousness and thought. Furthermore, AI utilizes digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceiving the environment, acquiring knowledge, and using that knowledge to achieve optimal results.

[0073] This application provides an ozone forecasting method, electronic device, and storage medium, aiming to improve the efficiency and accuracy of ozone forecasting.

[0074] The ozone forecasting method, electronic device, and storage medium provided in this application are specifically described through the following embodiments. First, the ozone forecasting method in this application embodiment is described.

[0075] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0076] Foundational technologies for artificial intelligence generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0077] The ozone forecasting method provided in this application relates to the field of artificial intelligence technology. The ozone forecasting method provided in this application can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc.; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application implementing the ozone forecasting method, but is not limited to the above forms.

[0078] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0079] Figure 1 This is an optional flowchart of the ozone forecasting method provided in the embodiments of this application. Figure 1 The method may include, but is not limited to, steps S101 to S107.

[0080] Step S101: Obtain the ozone concentration at the ground observation point of the target grid, construct the observation point concentration matrix based on the ozone concentration at the ground observation point, and perform interpolation processing on the observation point concentration matrix to obtain a smooth concentration matrix.

[0081] Step S102: Perform observation point masking on the observation point concentration matrix to obtain the grid observation point masking matrix.

[0082] Step S103: Obtain air pressure data in the target grid from the preset meteorological ensemble forecast, and construct a grid air pressure masking matrix based on the air pressure data in the target grid; wherein, each element in the grid air pressure masking matrix is ​​used to characterize the relationship between the air pressure value and the preset air pressure threshold.

[0083] Step S104: Obtain the first meteorological ensemble forecast data containing multiple ensemble members of the target grid and the first time encoding information of the first meteorological ensemble forecast data; wherein, the first meteorological ensemble forecast data is meteorological field data from historical time points to the current time point.

[0084] Step S105: Using the pre-trained ozone concentration generation model, ozone concentration is generated from the smoothed concentration matrix, the grid observation point masking matrix, the grid pressure masking matrix, the first time encoding information, and the first meteorological set forecast data, to obtain the ozone concentration prediction set of the target grid.

[0085] Step S106: Obtain the second meteorological ensemble forecast data containing multiple ensemble members of the target grid and the second time encoding information of the second meteorological ensemble forecast data; wherein, the second meteorological ensemble forecast data is meteorological field data at future time points.

[0086] Step S107: Using the fine-tuned ozone forecast model, ozone ensemble forecasts are performed on the ozone concentration prediction set of the target grid, the grid observation point masking matrix, the grid pressure masking matrix, the second time coding information, and the second meteorological ensemble forecast data to obtain the ensemble ozone forecast data of the target grid.

[0087] Steps S101 to S107 of this embodiment first construct an ozone concentration matrix for ground observation points. This matrix is ​​then interpolated to obtain a smoothed concentration matrix. By combining this matrix with observation point masking, and constructing a grid pressure masking matrix based on the pressure data in the target grid, spatial constraints based on observation station sparsity and topographic pressure characteristics are provided for the subsequent ozone concentration generation model, thus improving the accuracy of subsequent ozone forecasts. Secondly, by utilizing the ozone concentration generation model to generate an ozone concentration set containing multiple set members from the smoothed concentration matrix, the grid observation point masking matrix, the grid pressure masking matrix, the first time-coded information, and the first meteorological ensemble forecast data, a predicted ozone concentration set containing multiple set members for the target grid is obtained, which can directly... Ozone concentration prediction is achieved through a generative model, eliminating the time-consuming data assimilation and iteration process of traditional numerical models. This effectively saves model computational resources and improves the timeliness of ozone concentration generation, thus enhancing the timeliness of subsequent ozone forecasts. When there is a significant deviation between ozone forecasts and actual observations, new concentration information can be promptly introduced for a rapid response. Finally, the fine-tuned ozone forecast model performs ensemble ozone forecasting on the target grid's ozone concentration prediction set, grid observation point masking matrix, grid pressure masking matrix, second time-coded information, and second meteorological ensemble forecast data. This allows for direct integration with the ozone concentration prediction set from the ozone concentration generation model to perform ensemble ozone forecasts containing multiple ensemble members. This leverages the saved model computational resources to fully quantify meteorological uncertainties, thereby improving the efficiency and accuracy of ozone forecasting.

[0088] In step S101 of some embodiments, specifically, the target grid refers to a two-dimensional grid consisting of multiple location points that covers the geographical area of ​​ozone forecasting.

[0089] For example, a 100x100 target grid covering the target city area, where each grid point represents a location point.

[0090] Specifically, the ozone concentration at the observation point refers to the ozone concentration value actually measured by the ground observation stations actually set up in the target grid.

[0091] For example, ground observation station A observed an ozone concentration of 120 μg / m³ at 14:00 in the afternoon.

[0092] Specifically, the observation point concentration matrix is ​​a matrix with the same size as the target grid, and the ozone value corresponding to the ground observation point in the grid is the measured ozone concentration value, while the values ​​of the other grid points are empty or zero.

[0093] In this embodiment, an observation point concentration matrix is ​​constructed based on the ozone concentration at the ground observation points. This matrix maps discrete ground observation point data onto a regular grid, forming a spatially sparse but highly accurate matrix, providing structured input for subsequent interpolation processing.

[0094] Please see Figure 2 In some embodiments, step S101 may include, but is not limited to, steps S201 to S202:

[0095] Step S201: Obtain each target interpolation point of the target grid, and calculate the Euclidean distance between each target interpolation point and the ground observation point to obtain the target distance of the interpolation observation point.

[0096] Step S202: Based on the preset power parameter, the ozone concentration at the observation point, and the target distance of the interpolation observation point, ozone interpolation processing is performed on each target interpolation point to obtain a smooth concentration matrix.

[0097] In step S201 of some embodiments, specifically, the target interpolation point refers to the location point in the target grid where no ground observation station is set up and the ozone concentration needs to be supplemented by interpolation calculation.

[0098] Specifically, the target distance of the interpolation observation point refers to the Euclidean distance between each target interpolation point and the ground observation point.

[0099] Specifically, by traversing all locations in the target grid, locations that are not marked as ground observation points can be selected as target interpolation points. For example, in a 100x100 grid, after excluding 10 ground observation points, the remaining 9990 points can be identified as target interpolation points.

[0100] Specifically, for each target interpolation point, the Euclidean distance between the target interpolation point and all ground observation points can be calculated and determined as the target distance of the interpolation observation point.

[0101] For example, the Euclidean distance between the target interpolation point (e.g., m1=20, n1=30) and the ground observation point (m2=15, n2=25) can be calculated by taking the square root of [(20-15)²+(30-25)²] to obtain the target distance of the interpolation observation point as 7.07 km.

[0102] In step S202 of some embodiments, specifically, the preset power parameter is a hyperparameter used in the interpolation algorithm to adjust the degree of influence of distance weight, usually taking a value of 2 (i.e., inverse square weight), which is used to indicate that the influence of distance on weight is a quadratic decay.

[0103] Specifically, the smoothed concentration matrix refers to the complete ozone concentration matrix with the same size as the target grid, formed by filling the missing values ​​in the concentration matrix of the observation points through an interpolation algorithm.

[0104] Specifically, the ozone interpolation process can be represented by the following formula:

[0105]

[0106] in, This represents the smoothed concentration matrix of the target interpolation point V, where V represents the target interpolation point. This represents the g-th ground observation point. Indicates the target interpolation point V and the ground observation point The interpolated observation point target distance, where n represents the number of ground observation points. Let represent the ozone concentration at the g-th ground observation point, and l represent the power parameter.

[0107] Steps S201 to S202 fill the gaps in areas without observation points, transforming sparse and discrete ground station observation data into a spatially continuous ozone concentration field based on a grid scale that can cover the entire forecast area, thus providing data support for subsequent ozone concentration generation.

[0108] In step S102 of some embodiments, specifically, the grid observation point masking matrix is ​​a binary matrix with the same size as the target grid (i.e., each element in the matrix is ​​0 or 1).

[0109] Specifically, if an element in the grid observation point masking matrix represents a ground observation point, then the element is 1; otherwise, the element is 0.

[0110] In this embodiment, by masking the observation point concentration matrix, a grid observation point masking matrix is ​​obtained. This helps the subsequent ozone concentration generation model to identify reliable observation point ozone concentrations, thereby assigning higher weights to the ozone concentration generation process and improving the accuracy of ozone concentration generation.

[0111] In step S103 of some embodiments, specifically, the meteorological ensemble forecast data refers to the result set used to characterize the possible range and probability of future meteorological distribution and weather conditions of the target grid, and the meteorological ensemble forecast data includes multiple meteorological values ​​such as temperature, air pressure, humidity, wind speed and wind direction, and the meteorological forecast corresponding to each meteorological value is called a forecast member.

[0112] For example, if we predict the weather changes over the next 24 hours based on an initial meteorological value (such as 20.1℃, 1012 hPa, 63% humidity, 5.5 m / s wind speed, or 355 degrees wind direction), then the meteorological forecast data such as temperature, air pressure, humidity, wind speed, or wind direction for the next 24 hours are the forecast members.

[0113] Specifically, the grid pressure masking matrix refers to another binary matrix with the same size as the target grid (such as WxH). Each element in the grid pressure masking matrix is ​​used to characterize the relationship between the pressure value and the preset pressure threshold.

[0114] Specifically, the preset air pressure threshold is usually determined based on the standard isobaric surface corresponding to the terrain height. That is, the preset air pressure threshold usually includes a first air pressure threshold of 925 hPa and a second air pressure threshold of 850 hPa. Therefore, the grid air pressure masking matrix includes a first air pressure masking matrix corresponding to the first air pressure threshold and a second air pressure masking matrix corresponding to the second air pressure threshold.

[0115] Specifically, in plains areas, the surface atmospheric pressure is usually higher than 925 hPa. If the pressure value represented by an element in the first pressure shielding matrix is ​​greater than or equal to 925 hPa, then the element is 1; if the pressure value represented by an element in the first pressure shielding matrix is ​​less than 925 hPa, then the element is 0. In mountainous areas, the surface atmospheric pressure is usually lower than 850 hPa. If the pressure value represented by an element in the second pressure shielding matrix is ​​greater than or equal to 850 hPa, then the element is 1; if the pressure value represented by an element in the second pressure shielding matrix is ​​less than 850 hPa, then the element is 0.

[0116] In this embodiment, a grid pressure masking matrix is ​​constructed based on the air pressure data in the target grid, which can help the ozone concentration generation model distinguish the differences in the impact of different altitude terrain regions on ozone generation, thereby further improving the accuracy of ozone forecasting.

[0117] In step S104 of some embodiments, specifically, the first meteorological ensemble forecast data refers to the meteorological field data of the target grid from historical time points to the current time point, and the meteorological field data contains multiple ensemble members, and each ensemble member is composed of the forecast values ​​of multiple forecast members.

[0118] For example, the first meteorological ensemble forecast data may include 2m temperature, 2m relative humidity, 10m east-west wind speed, 10m north-south wind speed, instantaneous surface air pressure, 925hpa east-west wind speed, 925hpa north-south wind speed, 925hpa geopotential height, 850hpa east-west wind speed, 850hpa north-south wind speed, 850hpa geopotential height, and 6-hour cumulative precipitation, etc.

[0119] Specifically, first-time encoded information refers to converting time features (such as hours and days of the week) into numerical vectors that the model can recognize through sine / cosine encoding.

[0120] For example, 14 o'clock can be converted into a two-dimensional vector (sin(2πx14 / 24), cos(2πx14 / 24)) using hour sine / cosine encoding, and Wednesday can be converted into a two-dimensional vector (sin(2πx3 / 7), cos(2πx3 / 7)) using week number sine / cosine encoding.

[0121] In this embodiment, by acquiring the first meteorological ensemble forecast data containing multiple ensemble members of the target grid and the first time encoding information of the first meteorological ensemble forecast data, the model can learn the influence of meteorological information and time patterns on ozone concentration changes, which helps to further improve the accuracy of subsequent ozone forecasts. Multiple meteorological ensemble members help to assess the impact of meteorological uncertainties on ozone forecasts.

[0122] Please see Figure 3 In some embodiments, prior to step S105, the ozone forecasting method further includes, but is not limited to, steps S301 to S308:

[0123] Step S301: Obtain the first training set of the target grid and the original ozone concentration generation model; wherein, the first training set includes the training observation point concentration matrix, the training smoothed concentration matrix after interpolation of the training observation point concentration matrix, the current noisy ozone concentration, the target true ozone concentration, the training grid observation point masking matrix, the training grid air pressure masking matrix, the training first meteorological forecast data, and the training first time encoding information, and the training observation point concentration matrix, the current noisy ozone concentration, and the target true ozone concentration are obtained from the preset simulation data.

[0124] Step S302: Using the original ozone concentration generation model, ozone concentration is generated from the training smooth concentration matrix, the training grid observation point masking matrix, the training grid pressure masking matrix, the training first weather forecast data, the training first time encoding information, and the current noisy ozone concentration to obtain the training predicted ozone concentration of the target grid.

[0125] Step S303: Calculate the concentration loss based on the predicted ozone concentration from the training and the actual target ozone concentration to obtain the concentration loss value.

[0126] Step S304: Train the original ozone concentration generation model based on the concentration loss value to obtain a pre-trained ozone concentration generation model.

[0127] Step S305: Obtain the second training set of the target grid and the original ozone forecast model; wherein, the second training set includes the training grid observation point masking matrix, the training grid pressure masking matrix, the training second time coding information, the training second meteorological forecast data, the current true ozone concentration, the future noisy ozone concentration and the true ozone forecast data, and the training second meteorological forecast data, the current true ozone concentration, the future noisy ozone concentration and the true ozone forecast data are obtained from the simulation data.

[0128] Step S306: Ozone forecasting is performed on the training grid observation point masking matrix, training grid pressure masking matrix, training second meteorological forecast data, training second time coding information, future noisy ozone concentration and current real ozone concentration using the original ozone forecasting model to obtain the training ozone forecast data for the target grid.

[0129] Step S307: Calculate the forecast loss based on the training ozone forecast data and the actual ozone forecast data to obtain the training forecast loss value.

[0130] Step S308: Train the original ozone forecast model based on the training forecast loss value to obtain a pre-trained ozone forecast model.

[0131] In step S301 of some embodiments, specifically, the first training set refers to the dataset used to train the original ozone concentration generation model. The first training set includes a training observation point concentration matrix (such as the observed ozone concentration on the grid where 10 ground observation points are located at time t-3h and time t), a training smoothed concentration matrix after interpolation of the training observation point concentration matrix, a training grid observation point masking matrix (such as a binary matrix that identifies the location of ground observation stations), a training grid pressure masking matrix (such as a binary matrix that identifies pressures greater than atmospheric pressures of 850hPa and 925hPa), training first meteorological forecast data (such as meteorological forecast data (i.e., meteorological field) at time t-3h and time t), training first time-coded information, current noisy ozone concentration and target true ozone concentration, and the training observation point concentration matrix, current noisy ozone concentration and target true ozone concentration can be obtained from the WRF-GC simulation data for ozone distribution.

[0132] Furthermore, the current noisy ozone concentration refers to the ozone field with random noise added to the ozone concentration of the target grid from historical time points to the current time point. It is used to simulate the uncertain input that the ozone concentration generation model may face in practical applications.

[0133] Specifically, the current concentration of ozone containing noise can be determined using the following formula:

[0134]

[0135] in, , The signal retention coefficient for the target true ozone concentration. The noise accumulation product coefficient is used to control the proportion of the target true ozone concentration signal retained, where s is the time step. The strength of random noise, To randomly add noise, The target is the actual ozone concentration.

[0136] Furthermore, the target true ozone concentration is the true ozone field of the target grid from a historical time point to the current time point.

[0137] Furthermore, the true ozone forecast data is the actual ozone field of the target grid at the future forecast time point.

[0138] Specifically, the ozone concentrations of the target grid at times t-3h and t obtained from WRF-GC can be determined as the target true ozone concentration, and the current noisy ozone concentration can be determined by adding random noise to the target true ozone concentration.

[0139] Specifically, the original ozone concentration generation model is an untrained neural network model with randomly initialized weight parameters (such as a diffusion model). This model includes an encoding layer, a bottleneck layer, and a decoding layer. The encoding layer is used to progressively diffuse noisy data into the preset ozone concentration generation model to extract the training noisy feature map of the first training set and achieve preliminary localization of the noisy region. The bottleneck layer is used to further convolve the training noisy feature map to generate a potential representation of the noise distribution. The decoding layer is used to progressively denoise the convolved training convolutional noisy feature map to generate a denoised ozone concentration, thereby achieving ozone concentration prediction for the target grid at the current time.

[0140] In this embodiment, by acquiring the first training set of the target grid, the ozone concentration generation model can be improved by using a high-resolution grid training set as input, given that ozone has a spatial transport and accumulation effect. The first training set also includes training data covering multiple times, weather conditions, and noise, providing the model with rich samples to learn from sparse observations to the complete field. This ensures that the generation model can learn ozone prediction tasks under various seasons and weather scenarios, which helps to improve the accuracy of ozone forecasts.

[0141] In step S302 of some embodiments, specifically, the first training set can be subjected to progressive noise encoding, convolution, and progressive denoising processing through the original ozone concentration generation model to generate a denoised training predicted ozone concentration.

[0142] In step S303 of some embodiments, specifically, the concentration loss value refers to the difference between the training predicted ozone concentration and the target true ozone concentration.

[0143] Specifically, the concentration loss value can be calculated using the following loss function:

[0144]

[0145] Where L represents the concentration loss value, This represents the feature weight matrix of the first training set. The feature weight matrix only applies to the target feature, namely the target ozone concentration. Y represents the value in the i-th row and j-th column of the weight matrix, and Y represents the target true ozone concentration. This indicates the training predicted ozone concentration. This represents element-wise multiplication. This represents the Frobenius norm.

[0146] In step S304 of some embodiments, specifically, the original ozone forecast model can be updated based on the concentration loss value to obtain an updated ozone concentration generation model. Based on the updated ozone concentration generation model, the process of generating ozone concentration using the original ozone concentration generation model is repeated, using the training smooth concentration matrix, the training grid observation point masking matrix, the training grid pressure masking matrix, the training first meteorological forecast data, the training first time-encoded information, and the current noisy ozone concentration, until the concentration loss value meets the preset concentration loss condition, thus obtaining a pre-trained ozone concentration generation model.

[0147] The concentration loss condition is the convergence criterion for the ozone concentration generation model. If the training rounds reach 10, the ozone concentration generation model will stop training.

[0148] Specifically, if the concentration loss value meets the concentration loss condition, the trained predicted ozone concentration can be determined as the target trained predicted ozone concentration. If the concentration loss value does not meet the concentration loss condition, the parameters of the ozone concentration generation model (such as the model convolution kernel weights) can be adjusted according to the concentration loss value through the backpropagation algorithm until the concentration loss value is less than the concentration loss condition, so as to obtain the pre-trained ozone concentration generation model.

[0149] In step S305 of some embodiments, specifically, the second training set refers to the dataset used to train the original ozone forecast model. The second training set includes a training grid observation point masking matrix, a training grid pressure masking matrix, training second time coding information, training second weather forecast data, current true ozone concentration, future noisy ozone concentration, and true ozone forecast data. The training second weather forecast data, current true ozone concentration, future noisy ozone concentration, and true ozone forecast data can be obtained from the WRF-GC simulation data on ozone distribution.

[0150] Furthermore, the training second weather forecast data refers to the weather forecast data at the target grid at the future forecast time point. It is used to forecast the distribution of future meteorological elements and weather conditions, and includes, but is not limited to, two-dimensional grid data such as temperature, humidity, and wind field.

[0151] Furthermore, the future noisy ozone concentration refers to the ozone field to which random noise interference is added to the ozone concentration of the target grid at future forecast time points.

[0152] Specifically, the ozone concentration of the target grid at time t+3h obtained from WRF-GC can be determined as the true ozone forecast data, and the future noisy ozone concentration can be determined by adding random noise to the true ozone forecast data.

[0153] Specifically, the ozone forecasting model is also a diffusion model, which includes a coding layer, a bottleneck layer, and a decoding layer. Its function is consistent with that of the specific network layers in the ozone concentration generation model, and will not be elaborated here.

[0154] In step S306 of some embodiments, specifically, the second training set can be encoded, convolved, and decoded using the original ozone forecast model to generate training ozone forecast data for future times.

[0155] In step S307 of some embodiments, specifically, the training forecast loss value refers to the difference between the training ozone forecast data and the actual ozone forecast data.

[0156] Specifically, the calculation process for the training forecast loss value is the same as that for the concentration loss value, and will not be repeated here.

[0157] In step S308 of some embodiments, specifically, if the training forecast loss value meets the preset forecast loss condition, the original ozone forecast model is determined as the pre-trained ozone forecast model. If the training forecast loss value does not meet the forecast loss condition, the parameters of the ozone forecast model can be adjusted according to the training forecast loss value through the backpropagation algorithm, and the process returns to execute the step of ozone forecasting using the original ozone forecast model on the training grid observation point masking matrix, the training grid pressure masking matrix, the training second meteorological forecast data, the training second time coding information, the future noisy ozone concentration, and the current real ozone concentration, until the training forecast loss value meets the forecast loss condition, so as to obtain the pre-trained ozone forecast model.

[0158] Specifically, the forecast loss condition is the convergence criterion of the ozone forecast model. If the training rounds reach 10, the ozone forecast model will stop training.

[0159] Through steps S301 to S308, the ozone concentration generation model generates a high-accuracy ozone field using prior knowledge of sparse observation stations and shielding space constraints. It can directly generate ozone concentration through a generative model without going through the time-consuming data assimilation and iteration process of traditional numerical models, thus achieving timely ozone concentration generation. Furthermore, by using an ozone forecasting model based on the ozone field, future meteorological conditions, and shielding space constraints, it further ensures that the prediction results conform to the physical laws of topography and meteorology, thereby improving the timeliness and accuracy of ozone forecasts.

[0160] Please see Figure 4 In some embodiments, after step S308, the ozone forecasting method further includes, but is not limited to, steps S401 to S406:

[0161] Step S401: Obtain the fine-tuning training set; wherein, the fine-tuning training set includes the training grid observation point masking matrix, the training grid pressure masking matrix, fine-tuned first meteorological forecast data, fine-tuned first time coding information, fine-tuned second meteorological forecast data, fine-tuned second time coding information, first-time smoothed concentration matrix and second-time smoothed concentration matrix obtained and interpolated from ground observation stations, and the fine-tuned first meteorological forecast data and the fine-tuned second meteorological forecast data are obtained from preset real meteorological forecast data.

[0162] Step S402: The ozone concentration is generated by using the pre-trained ozone concentration generation model to generate ozone concentration from the training grid observation point masking matrix, the training grid air pressure masking matrix, the fine-tuned first weather forecast data, the fine-tuned first time encoding information, and the first time smoothed concentration matrix, so as to obtain the initial fine-tuned ozone concentration of the target grid.

[0163] Step S403: The ozone concentration is generated by using the pre-trained ozone concentration generation model to generate ozone concentration from the training grid observation point masking matrix, the training grid pressure masking matrix, the fine-tuned second weather forecast data, the fine-tuned second time coding information, and the second time smoothed concentration matrix, so as to obtain the target fine-tuned ozone concentration of the target grid.

[0164] Step S404: Using the pre-trained ozone forecast model, ozone forecasts are performed on the training grid observation point masking matrix, the training grid pressure masking matrix, the fine-tuned second meteorological forecast data, the fine-tuned second time coding information, and the initial fine-tuned ozone concentration to obtain the fine-tuned ozone forecast data for the target grid.

[0165] Step S405: Calculate the forecast loss based on the fine-tuned ozone forecast data and the target fine-tuned ozone concentration to obtain the fine-tuned forecast loss value.

[0166] Step S406: Fine-tune the pre-trained ozone forecast model based on the fine-tuned forecast loss value to obtain the fine-tuned ozone forecast model.

[0167] In step S401 of some embodiments, specifically, the fine-tuning training set is used as the dataset for pre-training the fine-tuning ozone forecast model. The fine-tuning training set includes a training grid observation point masking matrix, a training grid pressure masking matrix, fine-tuning first meteorological forecast data, fine-tuning first time-coded information, fine-tuning second meteorological forecast data, fine-tuning second time-coded information, a first-time smoothed concentration matrix and a second-time smoothed concentration matrix obtained and interpolated from ground observation stations, and the fine-tuning first meteorological forecast data and the fine-tuning second meteorological forecast data are obtained from preset real meteorological forecast data.

[0168] Specifically, the fine-tuned first weather forecast data for the times t-6h and t-3h, as well as the fine-tuned second weather forecast data for the times t and t+3h, can be obtained from real weather forecast data.

[0169] In step S402 of some embodiments, for example, the ozone concentration generated from the fine-tuned first meteorological forecast data for times t-6h and t-3h can be determined as the initial fine-tuned ozone concentration, and this initial fine-tuned ozone concentration is used as the input to the subsequently pre-trained ozone forecast model.

[0170] Specifically, a pre-trained ozone concentration generation model can be used to progressively encode noise, convolve, and progressively denoise the fine-tuning training set to generate a denoised initial fine-tuned ozone concentration.

[0171] In step S403 of some embodiments, for example, the ozone concentration generated from the fine-tuned second meteorological forecast data for times t and t+3h can be determined as the target fine-tuned ozone concentration, and the target fine-tuned ozone concentration can be used as the fine-tuning benchmark for the subsequent pre-trained ozone forecast model.

[0172] In step S404 of some embodiments, the fine-tuning training can be encoded, convolved, and decoded using a pre-trained ozone forecasting model to generate fine-tuned ozone forecast data for future times.

[0173] In step S405 of some embodiments, specifically, the fine-tuning forecast loss value refers to the difference between the fine-tuned ozone forecast data and the target fine-tuned ozone concentration, which is used to characterize the forecast accuracy of the pre-trained ozone forecast model.

[0174] In step S406 of some embodiments, specifically, if the fine-tuned forecast loss value meets the forecast loss condition, the pre-trained ozone forecast model is determined as the fine-tuned ozone forecast model. If the fine-tuned forecast loss value does not meet the forecast loss condition, the parameters of the pre-trained ozone forecast model can be fine-tuned according to the fine-tuned forecast loss value through the backpropagation algorithm, and the process returns to execute the steps of ozone forecasting using the pre-trained ozone forecast model on the training grid observation point masking matrix, the training grid pressure masking matrix, the fine-tuned second meteorological forecast data, the fine-tuned second time coding information, and the initial fine-tuned ozone concentration, until the fine-tuned forecast loss value meets the forecast loss condition, so as to obtain the fine-tuned ozone forecast model.

[0175] Specifically, the forecast loss condition can be the convergence criterion for fine-tuning the ozone forecast model. For example, if the ozone forecast model stops fine-tuning after 10 training rounds, it can be used as a benchmark.

[0176] Through steps S401 to S406, in this embodiment, the fine-tuned ozone forecast model learns information that is difficult to measure or parameterize directly when fine-tuning using nearby historical ground observations, such as the spatiotemporal changes in ozone precursor emissions. That is, by combining ozone forecast data from real meteorological events to fine-tune the ozone forecast model, the accuracy of the ozone forecast model can be further improved.

[0177] Please see Figure 5 In some embodiments, step S105 includes, but is not limited to, steps S501 to S505:

[0178] Step S501: The smoothed concentration matrix, the grid observation point masking matrix, the grid pressure masking matrix, and the first meteorological ensemble forecast data are fused to obtain the target fused features.

[0179] Step S502: Determine the number of diffusion steps based on the ozone concentration generation model and target fusion characteristics.

[0180] Step S503: Obtain noise data and gradually diffuse the noise data into the ozone concentration generation model based on the diffusion steps to obtain the target noisy feature map.

[0181] Step S504: Perform convolution processing on the target noisy feature map and the first time-encoded information to obtain the convolutional noisy feature map.

[0182] Step S505: Based on the number of diffusion steps, perform stepwise denoising on the target noisy feature map and the convolutional noisy feature map to obtain the ozone concentration prediction set of the target grid.

[0183] In step S501 of some embodiments, in one embodiment of this application, before feature fusion, the smoothed concentration matrix features (i.e., two-dimensional vectors) of the smoothed concentration matrix and the first meteorological forecast features (i.e., two-dimensional vectors) of the first meteorological ensemble forecast data can be extracted. The smoothed concentration matrix features, the first meteorological forecast features, the grid observation point masking matrix, and the grid pressure masking matrix are then subjected to zero-mean normalization to obtain the normalized smoothed concentration matrix features, the grid observation point masking matrix, the grid pressure masking matrix, and the first meteorological forecast features.

[0184] Specifically, zero-mean standardization can be represented by the following formula:

[0185]

[0186] in, The sample features m before standardization (including smoothed concentration matrix features, grid observation point masking matrix, grid pressure masking matrix and first weather forecast features) are the k-th sample features. The average value of sample feature m. Let m be the standard deviation of the sample feature m.

[0187] Specifically, fusion refers to the three-dimensional comprehensive feature vector representation obtained by splicing together the smoothed concentration matrix features, the grid observation point masking matrix, the grid pressure masking matrix, and the first meteorological forecast features.

[0188] In step S502 of some embodiments, specifically, the diffusion step number is a key parameter in the ozone concentration generation model, which is used to control the number of iterations of noise injection intensity during the generation of the target noisy feature map.

[0189] Specifically, the number of diffusion steps is determined by the ozone concentration generation model and target fusion features. This number of diffusion steps is used to integrate the accuracy requirements of ozone concentration generation with the real-time status to ensure the efficiency and accuracy of ozone concentration generation.

[0190] For example, the diffusion step count can be 100.

[0191] In step S503 of some embodiments, specifically, the noise data refers to a randomly generated Gaussian noise matrix with the same feature size as the fused noise.

[0192] Specifically, the method for generating noise data is the same as the method for generating the current noisy ozone concentration, and will not be repeated here.

[0193] Please see Figure 6 In some embodiments, step S503 may include, but is not limited to, steps S601 to S606:

[0194] Step S601: Based on the diffusion steps, the noise data and the target fusion features are fused to obtain the fused noise features.

[0195] Step S602: The fused noise features are first encoded using the ozone concentration generation model to obtain the first noisy feature map.

[0196] Step S603: The first noisy feature map is downsampled to obtain the first noisy dimensionality-reduced noisy feature map.

[0197] Step S604: Update the noise intensity of the noise data based on the number of diffusion steps to obtain updated noise data, and fuse the updated noise data with the first noisy dimensionality-reduced noisy feature map based on the number of diffusion steps to obtain updated fused noise features.

[0198] Step S605: Perform a second encoding on the updated fused noise features and the target fused features to obtain a second noisy feature map.

[0199] Step S606: Downsample the second noisy feature map to obtain the target noisy feature map.

[0200] In step S601 of some embodiments, specifically, the fused noise feature is a feature representation after mixing the noise data of the current diffusion step with the target fused feature.

[0201] Specifically, different intensities of noise can be gradually injected into the target fusion feature through diffusion steps, so that the fusion feature is gradually covered by noise during the diffusion process.

[0202] For example, if the current diffusion step number is 10 and the current noise level is 10, then the target fusion feature accounts for 90% of the fused noise features, and the noise proportion of the current diffusion step number can be 10%.

[0203] Please see Figure 7 In some embodiments, step S602 includes, but is not limited to, steps S701 to S705:

[0204] Step S701: Perform convolution processing on the fused noise features and noise data to obtain the fused noise feature map.

[0205] Step S702: Randomly discard the fused noise feature map to obtain the discarded feature map.

[0206] Step S703: Perform convolution processing on the discarded feature map to obtain the enhanced feature map.

[0207] Step S704: Perform feature standardization on the enhanced feature map to obtain a standardized feature map.

[0208] Step S705: Activate the standardized feature map to obtain the first noisy feature map.

[0209] In step S701 of some embodiments, specifically, the fused noise feature map is a primary feature representation for extracting fused noise features.

[0210] Specifically, the fused noise features can be processed by convolution in the first convolutional layer of the first coding block to output a fused noise feature map.

[0211] For example, if the size of the fused noise feature is 160x224x64, then a first convolutional layer with 64 3x3 convolutional kernels, a stride of 1, and padding of 1 can be used to extract features and output a fused noise feature map with a size of 160x224x64, where C=64, HxW=160x224, C is the number of channels, HxW is the grid resolution, H is the grid height, and W is the grid width.

[0212] In this embodiment, by performing convolution processing on the fused noise features, a preliminary feature map can be extracted from the noisy input without losing spatial resolution, providing a noise signal data foundation for subsequent encoding.

[0213] In step S702 of some embodiments, specifically, the discarded feature map is a feature map obtained by setting some channels or spatial positions of the fused noise feature map to zero, and the size of the feature map is consistent with the size of the fused noise feature map.

[0214] Specifically, random dropout (such as Dropout) technology can be used to randomly select 10% of the channel feature values ​​of the fused noise feature map and set them to zero according to a preset dropout rate (such as 0.1).

[0215] In step S703 of some embodiments, specifically, the enhanced feature map is a feature map extracted by performing a second convolution on the discarded feature map, and the size of the feature map is also consistent with the size of the fused noise feature map.

[0216] Specifically, the discarded feature map (e.g., size 160x224x64) can be input into the second convolutional layer (e.g., 64 3x3 convolutional kernels, stride 1, padding 1) in the first coding block to extract the enhanced feature map (size 160x224x64), and the enhanced feature map can be activated by an activation function (e.g., SiLU) to highlight the enhanced features.

[0217] Specifically, although the parameters of the second convolutional layer are the same as those of the first convolutional layer, the second convolutional layer is used to enhance key features in the discard feature map (such as the grid pressure mask matrix feature of "low ozone concentration in low pressure areas").

[0218] In step S704 of some embodiments, specifically, the standardized feature map refers to the feature map after normalizing the channel dimension of the enhanced feature map.

[0219] Specifically, the mean and variance of the enhanced feature map are calculated for each channel, with the mean set to 0 and the variance set to 1. The mean and variance distribution are then adjusted using the model's learnable scaling and translation parameters to maintain the model's expressive power.

[0220] In step S705 of some embodiments, specifically, the first noisy feature map is a feature map after activating the normalized feature map, and the size of this feature map is also the same as that of the fused noise feature map. Figure 1 To (e.g., 160x224x64).

[0221] Specifically, activation can be performed using the following formula:

[0222]

[0223] in, Represents the standardized feature map conduct The first noisy feature map after function activation. Represents a standardized feature map. This represents the activation function. This indicates a target for the standardized feature map. It is an exponentially decaying function.

[0224] Through steps S701 to S705, multi-layer coding enables the model to capture noisy features at multiple scales. By combining multiple downsampling to gradually expand the receptive field, the accuracy of extracting noisy features is ensured while reducing the amount of computation. Furthermore, by combining feature reuse to retain key details, the accuracy of the model in capturing noisy features at different scales is improved.

[0225] In step S603 of some embodiments, specifically, the first noisy dimensionality reduction noisy feature map is a feature representation after reducing the spatial resolution of the first noisy feature map.

[0226] Specifically, downsampling can be represented by the following expression:

[0227]

[0228] in, This indicates the position of the first noisy, dimensionality-reduced, noisy feature map output after max pooling. eigenvalues, This represents the four feature values ​​of the first noisy feature map within a 2x2 window, located at local offsets (p, q) within a sliding window, starting from (2i, 2j). Here, i is the height coordinate of the feature map, j is the width coordinate of the feature map, p is the relative offset in the height direction, and q is the relative offset in the width direction.

[0229] Specifically, the first noisy feature map (160x224x64) can be subjected to max pooling (e.g., a 2x2 max pooling window with a stride of 2) to halve the spatial resolution of the first noisy feature map while keeping the number of channels unchanged, thus generating the first noisy dimensionality-reduced noisy feature map (80x112x64).

[0230] In this embodiment, by downsampling the first noisy feature map, significant features of local regions can be extracted while ignoring detailed noise, effectively reducing the amount of computation.

[0231] In step S604 of some embodiments, specifically, updating the noise data means enhancing the Gaussian noise matrix according to the current diffusion step number, and the updated noise data has the same size as the noise data.

[0232] For example, if the diffusion step count is 50, the strength of the random noise can be adjusted. To enhance the current noise level to 50, and to determine the updated noise data after adjusting the noise intensity.

[0233] Specifically, the updated fused noise features are the feature representation after mixing the updated noise data of the current diffusion step number with the first noisy dimensionality reduction noise feature map.

[0234] For example, if the current diffusion step number is 50 and the updated noise level is 50, then the proportion of the target fusion feature in the updated fusion noise feature can be 50%, and the noise proportion corresponding to diffusion step number 50 can be 50%.

[0235] In step S605 of some embodiments, specifically, the second noisy feature map is a feature representation further extracted from the multi-source input by increasing the number of channels.

[0236] For example, the first noisy, dimensionality-reduced noisy feature map (e.g., 80x112x64) can be upsampled to 160x224x64. The target fusion feature (160x224x64) and the updated noise data (160x224x64) are then adjusted to 128 channels and concatenated according to the channel dimension to form an enhanced input (160x224x128). The second noisy feature map (80x112x128) is then extracted from the enhanced input through the second coding block.

[0237] In this embodiment, by performing a second encoding on the updated fusion noise features and the target fusion features, the target fusion features and the updated noise data can be combined to supplement the details that may be lost during downsampling, thereby improving the completeness of the feature expression and further ensuring the accuracy of ozone concentration generation.

[0238] In step S606 of some embodiments, specifically, the target noisy feature map is a feature representation after reducing the spatial resolution of the second noisy feature map.

[0239] Specifically, since the method for downsampling the second noisy feature map is the same as that for downsampling the first noisy feature map, it will not be described again here.

[0240] For example, the second noisy feature map (80x112x128) can be max-pooled to extract the target noisy feature map (40x56x128).

[0241] Steps S601 to S606 preserve the spatial constraints of the target fusion features and add noise interference of different intensities during ozone generation, thereby improving the prediction accuracy of the model under conditions of few samples and complex weather.

[0242] In an optional embodiment of this application, the number of coding blocks included in the coding layer is not limited here. In order to further improve the accuracy of the ozone concentration generated by the model, the coding layer may also include a third coding block and a fourth coding block. The third coding block can be used to encode the target noisy feature map, the target fusion feature, and the noise data updated again based on the number of diffusion steps (e.g., 75) (e.g., noise level 75) to obtain a third noisy feature map (e.g., 40x56x256). The third noisy feature map is then downsampled to obtain a third dimensionality-reduced noisy feature map. The fourth coding block can be used to encode the third dimensionality-reduced noisy feature map (e.g., 20x28x256), the target fusion feature, and the noise data updated again based on the number of diffusion steps (e.g., 100) (e.g., noise level 100) to obtain a fourth noisy feature map (e.g., 20x28x512). The fourth noisy feature map is then downsampled to obtain a fourth dimensionality-reduced noisy feature map. The fourth dimensionality-reduced noisy feature map is determined as the final target noisy feature map (10x14x512).

[0243] In step S504 of some embodiments, specifically, the convolutional noisy feature map is a feature representation after fusing temporal information with the target noisy feature map.

[0244] Specifically, the bottleneck layer can be used to expand the first-time encoded information into a target dimension vector through a fully connected layer and broadcast it to every spatial location of the target noisy feature map. Then, convolution processing is used to fuse temporal and spatial features to output a preliminary convolutional noisy feature map. Further dilated convolution processing is performed on the preliminary convolutional noisy feature map to obtain the convolutional noisy feature map.

[0245] For example, the first temporal encoding information of a 6-dimensional vector can be expanded into a 512-dimensional vector through a fully connected layer using a bottleneck layer. This vector is then broadcast to every spatial location of the target noisy feature map (e.g., 10x14x512). Temporal and spatial features are then fused using a 3×3 convolution (e.g., 1024 channels, stride 1, padding 1) to output a preliminary convolutional noisy feature map (e.g., 10x14x1024). This preliminary convolutional noisy feature map is further processed with dilated convolution (e.g., dilation rate 2) to expand the receptive field, ultimately resulting in a 10x14x1024 convolutional noisy feature map.

[0246] Specifically, dilated convolution can be represented by the following formula:

[0247]

[0248] in, This represents the feature value of the noisy convolutional feature map output after dilated convolution at position (i,j). This represents the weight of the dilated convolution kernel at the local offset (p, q), where p is the relative offset in the height direction and q is the relative offset in the width direction. This represents the activation feature value of the initial convolutional noisy feature map at position (i+r,ir), where i is the height coordinate of the feature map, j is the width coordinate of the feature map, and r is the dilation rate.

[0249] Please see Figure 8 In some embodiments, step S505 includes, but is not limited to, steps S801 to S804:

[0250] Step S801: Perform transpose convolution on the target noisy feature map and the convolutional noisy feature map to obtain the first upsampled feature map.

[0251] Step S802: Based on the number of diffusion steps, perform a first denoising process on the first upsampled feature map, the second noisy feature map, and the updated noise data to obtain a first denoised feature map.

[0252] Step S803: Perform a second transpose convolution on the first denoised feature map to obtain a second upsampled feature map.

[0253] Step S804: Perform a second denoising process on the second upsampled feature map, the first noisy feature map, and the noise data to obtain the ozone concentration prediction set of the target grid.

[0254] In step S801 of some embodiments, specifically, the first upsampled feature map is a feature map that upsamples low-resolution features through transposed convolution.

[0255] Specifically, the transposed convolution can be represented by the following formula:

[0256]

[0257] in, This indicates that the first upsampled feature map is at position eigenvalues, This represents the weight of the transposed convolution kernel at the local offset (p, q), where p is the relative offset in the height direction and q is the relative offset in the width direction. This represents the activation feature value of the target noisy feature map and the convolutional noisy feature map at position (i / 2, j / 2), where i is the height coordinate of the feature map, j is the width coordinate of the feature map, -p is the reverse offset in the height direction, and -q is the directional offset in the width direction. -p and -q are used to characterize the zero-padding and sliding process of the transposed convolution.

[0258] For example, the target noisy feature map (e.g., 40x56x128) can be concatenated with the convolutional noisy feature map (e.g., 40x56x128), and then input into the decoding layer for transposed convolution with zero padding and convolution operations (e.g., stride 2, padding 1, number of channels 64), outputting the first upsampled feature map (e.g., 80x64x128) to restore the resolution before downsampling of the second coding block.

[0259] In step S802 of some embodiments, specifically, the first noise reduction feature map is the ozone concentration feature after fusing upsampling features and coding layer features through skip connections.

[0260] For example, the first upsampled feature map (80x64x128), the updated noise data (160×224×64) determined based on the number of diffusion steps, and the second noisy feature map of the second coding block (e.g., 80x64x128) can be concatenated by channel (128 channels) through skip connections. The features are then fused using a 3×3 convolutional layer (128 channels) and the SiLU activation function to output the first denoised feature map (e.g., 80x64x128) corresponding to each diffusion step.

[0261] In this embodiment, by performing a first denoising process on the first upsampled feature map, the second noisy feature map, and the updated noise data, the feature map of the skip connection multiplexing coding layer can be combined to make up for the loss of details in the upsampling process. At the same time, noise injection enables the model to further learn the ability to extract real signals from complex interference, thereby improving the spatial accuracy of the decoded features.

[0262] In step S803 of some embodiments, specifically, the second upsampled feature map is a feature map whose resolution is enhanced by transposing the first denoised feature map through convolution.

[0263] Specifically, the method of performing a second transposed convolution on the first denoised feature map is the same as the method of performing a transposed convolution on the target noisy feature map and the convolutional noisy feature map, and will not be repeated here.

[0264] For example, the first denoised feature map (e.g., 80x64x128) can be transposed and convolved to obtain the second upsampled feature map as 160x224x64.

[0265] In step S804 of some embodiments, specifically, the ozone concentration prediction set is the ozone concentration field of all set members of the final target grid output by the decoding layer at the current moment, and there is corresponding ozone concentration prediction data for each set member.

[0266] Specifically, the method for performing the second denoising process on the second upsampled feature map, the first noisy feature map, and the noise data is the same as the method for performing the first denoising process on the first upsampled feature map, the second noisy feature map, and the noise data, and will not be described again here.

[0267] Through steps S801 to S804, upsampling is performed by transposed convolution, which can gradually improve the resolution of the feature map, ensure that the feature map size matches the coding layer, and reuse the features of the coding layer, effectively making up for the loss of details caused by upsampling. Furthermore, based on the number of diffusion steps, multiple rounds of denoising are performed on the noisy features of the coding layer, which effectively enhances the model's denoising ability and ensures the accuracy of ozone concentration generation.

[0268] By combining steps S501 to S505, the spatial constraints of grid location and grid pressure shielding matrix can be combined to improve the model's prediction accuracy for terrain regions. Ozone concentration can be predicted directly through generative models without going through the time-consuming data assimilation and iteration process of traditional numerical models, effectively saving model computing resources and improving the timeliness of ozone concentration generation.

[0269] Please see Figure 9The specific encoding process of the middle encoding layer in the ozone concentration generation model is as follows: Input parameters including meteorological parameters, ozone concentration information of the grid where the ground observation station is located, grid pressure masking matrix, and noisy ozone concentration field can be input into the encoding layer. Based on the current noise level determined by the number of diffusion steps, the noise is gradually diffused into the encoding layer for encoding by encoding block 1. The output channel C is 64, and the first noisy feature map is 160x224 with H (feature map height) x W (feature map width). The first noisy feature map is then subjected to max pooling with a pooling window of 2x2 and a step size of 2, and the output channel is 64, resulting in an 80x112 first noisy dimensionality-reduced noisy feature map. The 80x11 first noisy dimensionality-reduced noisy feature map is then processed by encoding block 2. The second encoding is performed, outputting a second noisy feature map of 80x112 with 128 channels. Max pooling is then performed on the second noisy feature map to output a second dimensionality-reduced noisy feature map of 40x56 with 128 channels. The second dimensionality-reduced noisy feature map is then encoded a third time through encoding block 3, outputting a third noisy feature map of 40x56 with 256 channels. Max pooling is then performed on the third noisy feature map to output a third dimensionality-reduced noisy feature map of 20x28 with 256 channels. The third dimensionality-reduced noisy feature map is then encoded a fourth time through encoding block 4, outputting a fourth noisy feature map of 20x28 with 512 channels. Max pooling is then performed on the fourth noisy feature map to output a fourth dimensionality-reduced noisy feature map of 10x14 with 512 channels.

[0270] Furthermore, the specific execution process of the bottleneck layer is as follows: the fourth dimension-reduced noisy feature map with 512 channels and the time information (hours, days of the week) after sine and cosine encoding are input into the bottleneck layer for convolution, and the output is a preliminary convolutional noisy feature map with 1024 channels of 10x14. Then, by performing dilated convolution on the preliminary convolutional noisy feature map, a 10x14 convolutional noisy feature map with 1024 channels is output.

[0271] Furthermore, the specific execution process of the decoding layer is as follows: A 10x14 convolutional noisy feature map with 1024 channels and the current noise level determined based on the number of diffusion steps are subjected to a first transposed convolution to output a 20x18 first upsampled feature map with 512 channels. The decoding block and the encoding block are aligned via skip connections. Based on the number of diffusion steps, the encoding block 4 performs a first denoising process on the 20x18 first upsampled feature map with 512 channels, the current noise level, and the 20x28 fourth noisy feature map with 512 channels, outputting a 20x28 first denoised feature map with 512 channels. A second transposed convolution is performed on the first denoised feature map to output a 40x56 second upsampled feature map with 256 channels. The decoding block 2 performs a second denoising process on the second upsampled feature map, the current noise level, and the 40x56 third noisy feature map, outputting a 40x56 third upsampled feature map with 256 channels. The second denoised feature map is subjected to a third transposed convolution, outputting an 80x112 third upsampled feature map with 128 channels. Based on the diffusion steps, decoding block 2 performs a third denoising process on the third upsampled feature map, the current noise level, and the 80x112 second noisy feature map, outputting an 80x112 third denoised feature map with 128 channels. The third denoised feature map is subjected to a fourth transposed convolution, outputting a 160x224 fourth upsampled feature map with 64 channels. Based on the diffusion steps, decoding block 4 performs a final denoising process on the fourth upsampled feature map, the current noise level, and the 160x224 first noisy feature map, outputting a 160x224 target denoised feature map with 64 channels. The target decoded feature map is then convolved by a convolutional layer, resulting in a 160x210 feature map. This feature map is then activated to obtain the ozone concentration prediction set x0.

[0272] In step S106 of some embodiments, specifically, the second meteorological ensemble forecast data is meteorological field data at a future time point in the target grid, and the meteorological forecast field data also includes multiple ensemble members.

[0273] Specifically, the second time encoding information encodes the time characteristics of future time points, such as sin / cos encoding "future Wednesday, t+3h=17:00".

[0274] In this embodiment, by acquiring the second meteorological ensemble forecast data of the target grid and the second time encoding information of the second meteorological ensemble forecast data, the ozone forecast model can learn the influence of the evolution of each ensemble member in the meteorological field over time on the ozone concentration, so as to predict the accumulation or diffusion trend of ozone.

[0275] In step S107 of some embodiments, specifically, the ensemble ozone forecast data refers to the complete ozone concentration distribution of the target grid at future time points output by the fine-tuned ozone forecast model, and each ensemble member corresponds to a ensemble ozone forecast data.

[0276] Specifically, if the target grid is a certain urban area (e.g., 100km × 100km), it is divided into 100 × 100 grids (i.e., each grid is 1km × 1km). The fine-tuned ozone forecast model is based on the ozone concentration prediction set at the current time (e.g., t=14:00), the grid observation point masking matrix (a binary matrix identifying the locations of 10 ground observation stations), the grid pressure masking matrix (a binary matrix identifying areas with atmospheric pressure above 925hPa and 850hPa), the second time-coded information (e.g., the time characteristics of t+3h=17:00), and the second meteorological ensemble forecast data (e.g., the temperature field, wind field, etc. at 17:00) to perform ozone ensemble forecast. The output ensemble ozone forecast data is a 100 × 100 numerical matrix, with some grid points... The quantitative results are as follows: The ozone forecast concentration at 17:00 for grid point a (observation station location, i=20, j=30) is 156 μg / m³; the ozone forecast concentration at 17:00 for grid point b (urban plain area, i=40, j=50, grid pressure masking matrix value is 1) is 142 μg / m³ (affected by the "southwest wind 3 m / s at 17:00" in the second meteorological ensemble forecast data, pollutants from the upstream industrial area are transported here, and the concentration is higher than that of the surrounding mountainous areas); the ozone forecast concentration at 17:00 for grid point c (urban mountainous area, i=60, j=70, pressure masking matrix value is 0, altitude is higher than 925 hPa) is 89 μg / m³ (due to the high terrain, the boundary layer rise leads to the diffusion of pollutants, and the concentration is significantly lower than that of the plain area).

[0277] In this embodiment, the ozone forecast model with fine-tuning is used to generate an ensemble of ozone forecast data. By leveraging the rapid reasoning capabilities of the generative model, ozone forecasts can be directly combined with the ozone concentration prediction ensemble, grid locations, and pressure shielding matrix. This ensures that the predicted ozone concentration distribution remains physically plausible even in areas with complex terrain. The saved model computational resources can be used to fully quantify meteorological uncertainties, thereby improving the efficiency and accuracy of ozone forecasts.

[0278] This application first constructs an ozone concentration matrix at ground observation points, then interpolates this matrix to obtain a smoothed concentration matrix. By combining this matrix with observation point masking and constructing a grid pressure masking matrix based on the pressure data in the target grid, spatial constraints based on observation station sparsity and topographic pressure characteristics are provided for the subsequent ozone concentration generation model, helping to improve the accuracy of subsequent ozone forecasts. Secondly, by using the ozone concentration generation model to generate an ozone concentration set containing multiple set members from the smoothed concentration matrix, grid observation point masking matrix, grid pressure masking matrix, first-time encoded information, and first meteorological ensemble forecast data, a predicted ozone concentration set containing multiple set members for the target grid is obtained. This can be directly implemented using a generative model. The current ozone concentration prediction eliminates the need for the time-consuming data assimilation and iteration process of traditional numerical models, effectively saving model computational resources and improving the timeliness of ozone concentration generation. This contributes to the improved timeliness of subsequent ozone forecasts, allowing for the timely introduction of new concentration information and rapid response when there are significant deviations between ozone forecasts and actual observations. Finally, by using the fine-tuned ozone forecast model to perform ozone ensemble forecasts on the target grid's ozone concentration prediction set, grid observation point masking matrix, grid pressure masking matrix, second time-coded information, and second meteorological ensemble forecast data, ozone ensemble forecasts can be directly combined with the ozone concentration prediction set of the ozone concentration generation model to perform ozone ensemble forecasts containing multiple ensemble members. This fully quantifies meteorological uncertainties by utilizing the saved model computational resources, thereby improving the efficiency and accuracy of ozone forecasts.

[0279] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described ozone forecasting method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0280] Please see Figure 10 , Figure 10 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes:

[0281] The processor 1001 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0282] The memory 1002 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 1002 can store the processing system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1002 and called and executed by the processor 1001 using the ozone forecasting method of the embodiments of this application.

[0283] Input / output interface 1003 is used to implement information input and output;

[0284] The communication interface 1004 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0285] Bus 1005 transmits information between various components of the device (e.g., processor 1001, memory 1002, input / output interface 1003, and communication interface 1004);

[0286] The processor 1001, memory 1002, input / output interface 1003 and communication interface 1004 are connected to each other within the device via bus 1005.

[0287] This application embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the above-described ozone forecasting method.

[0288] Memory, as a non-transitory storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0289] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0290] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0291] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0292] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0293] The terms "first," "second," "third," "fourth," etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data used are interchangeable where appropriate so that the embodiments of this application described herein can describe an implementation sequence other than those illustrated. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

Claims

1. A method for ozone forecasting, characterized in that, The method includes: Obtain the ozone concentration at the ground observation point of the target grid, and construct the observation point concentration matrix based on the ozone concentration at the ground observation point. Then, perform interpolation processing on the observation point concentration matrix to obtain a smoothed concentration matrix. The observation point concentration matrix is ​​masked to obtain a grid observation point masking matrix; The air pressure data in the target grid is obtained from the preset meteorological ensemble forecast, and a grid air pressure masking matrix is ​​constructed based on the air pressure data in the target grid; wherein, each element in the grid air pressure masking matrix is ​​used to characterize the relationship between the air pressure value and the preset air pressure threshold. Acquire the first meteorological ensemble forecast data containing multiple ensemble members of the target grid and the first time encoding information of the first meteorological ensemble forecast data; wherein, the first meteorological ensemble forecast data is meteorological field data from historical time points to the current time point; The ozone concentration prediction set of the target grid is obtained by using a pre-trained ozone concentration generation model to generate ozone concentrations from the smoothed concentration matrix, the grid observation point masking matrix, the grid pressure masking matrix, the first time-coded information, and the first meteorological set forecast data. Acquire the second meteorological ensemble forecast data containing multiple ensemble members of the target grid and the second time encoding information of the second meteorological ensemble forecast data; wherein, the second meteorological ensemble forecast data is meteorological field data at future time points; Using the fine-tuned ozone forecast model, ozone ensemble forecasts are performed on the ozone concentration prediction set of the target grid, the grid observation point masking matrix, the grid pressure masking matrix, the second time-coded information, and the second meteorological ensemble forecast data to obtain the ensemble ozone forecast data of the target grid. The ozone concentration prediction set for the target grid is obtained by using a pre-trained ozone concentration generation model to generate ozone concentrations from the smoothed concentration matrix, the grid observation point masking matrix, the grid pressure masking matrix, the first time-coded information, and the first meteorological ensemble forecast data, including: The smoothed concentration matrix, the grid observation point masking matrix, the grid pressure masking matrix, and the first meteorological ensemble forecast data are fused to obtain the target fused features; The number of diffusion steps is determined based on the ozone concentration generation model and the target fusion features. Noise data is acquired, and the noise data is gradually diffused into the ozone concentration generation model based on the number of diffusion steps to obtain a target noisy feature map; The target noisy feature map and the first temporal encoding information are convolved to obtain a convolutional noisy feature map; Based on the number of diffusion steps, the target noisy feature map and the convolutional noisy feature map are subjected to stepwise denoising processing to obtain the ozone concentration prediction set of the target grid.

2. The method according to claim 1, characterized in that, Before generating ozone concentration prediction sets for the target grid by using a pre-trained ozone concentration generation model to generate ozone concentrations from the smoothed concentration matrix, the grid observation point masking matrix, the grid pressure masking matrix, the first time-coded information, and the first meteorological ensemble forecast data, the method further includes: Obtain the first training set of the target grid and the original ozone concentration generation model; wherein, the first training set includes a training observation point concentration matrix, a training smoothed concentration matrix after interpolation of the training observation point concentration matrix, the current noisy ozone concentration, the target true ozone concentration, a training grid observation point masking matrix, a training grid barometric pressure masking matrix, training first meteorological forecast data, and training first time encoding information, and the training observation point concentration matrix, the current noisy ozone concentration, and the target true ozone concentration are obtained from preset simulation data; The ozone concentration is generated by using the original ozone concentration generation model to generate ozone concentration from the training smooth concentration matrix, the training grid observation point masking matrix, the training grid pressure masking matrix, the training first meteorological forecast data, the training first time encoding information, and the current noisy ozone concentration, thus obtaining the training predicted ozone concentration of the target grid. The concentration loss is calculated based on the predicted ozone concentration from the training and the actual target ozone concentration to obtain the concentration loss value. The original ozone concentration generation model is trained based on the concentration loss value to obtain a pre-trained ozone concentration generation model. Obtain the second training set of the target grid and the original ozone forecast model; wherein, the second training set includes the training grid observation point masking matrix, the training grid pressure masking matrix, training second time coding information, training second meteorological forecast data, current true ozone concentration, future noisy ozone concentration and true ozone forecast data, and the training second meteorological forecast data, the current true ozone concentration, the future noisy ozone concentration and the true ozone forecast data are obtained from the simulation data; The original ozone forecast model is used to forecast ozone based on the training grid observation point masking matrix, the training grid pressure masking matrix, the training second meteorological forecast data, the training second time coding information, the future noisy ozone concentration, and the current real ozone concentration, thereby obtaining the training ozone forecast data for the target grid. The forecast loss is calculated based on the training ozone forecast data and the actual ozone forecast data to obtain the training forecast loss value. The original ozone forecasting model is trained based on the training forecasting loss value to obtain the pre-trained ozone forecasting model.

3. The method according to claim 2, characterized in that, After training the original ozone forecast model based on the training forecast loss value to obtain a pre-trained ozone forecast model, the method further includes: Obtain a fine-tuning training set; wherein, the fine-tuning training set includes the training grid observation point masking matrix, the training grid barometric pressure masking matrix, fine-tuned first meteorological forecast data, fine-tuned first time encoding information, fine-tuned second meteorological forecast data, fine-tuned second time encoding information, a first-time smoothed concentration matrix and a second-time smoothed concentration matrix obtained and interpolated from the ground observation points, and the fine-tuned first meteorological forecast data and the fine-tuned second meteorological forecast data are obtained from preset real meteorological forecast data; The ozone concentration is generated by using the pre-trained ozone concentration generation model to generate ozone concentration from the training grid observation point masking matrix, the training grid pressure masking matrix, the fine-tuned first meteorological forecast data, the fine-tuned first time encoding information, and the first time smoothed concentration matrix, thus obtaining the initial fine-tuned ozone concentration of the target grid. The ozone concentration generation model is used to generate ozone concentrations from the training grid observation point masking matrix, the training grid pressure masking matrix, the fine-tuned second meteorological forecast data, the fine-tuned second time-encoded information, and the second time-smoothed concentration matrix, thereby obtaining the target fine-tuned ozone concentration for the target grid. The ozone forecast model is used to perform ozone forecasting on the training grid observation point masking matrix, the training grid pressure masking matrix, the fine-tuned second meteorological forecast data, the fine-tuned second time coding information, and the initial fine-tuned ozone concentration to obtain the fine-tuned ozone forecast data for the target grid. The forecast loss is calculated based on the fine-tuned ozone forecast data and the target fine-tuned ozone concentration to obtain the fine-tuned forecast loss value. The pre-trained ozone forecasting model is fine-tuned based on the fine-tuned forecasting loss value to obtain the fine-tuned ozone forecasting model.

4. The method according to claim 1, characterized in that, The step of progressively diffusing the noise data into the ozone concentration generation model based on the number of diffusion steps to obtain a target noisy feature map includes: Based on the number of diffusion steps, the noise data and the target fusion feature are fused to obtain the fused noise feature; The fused noise features are first encoded using the ozone concentration generation model to obtain a first noisy feature map. The first noisy feature map is downsampled to obtain a first noisy dimensionality-reduced noisy feature map. The noise intensity of the noise data is updated based on the number of diffusion steps to obtain updated noise data. The updated noise data is then fused with the first noisy dimensionality-reduced noisy feature map based on the number of diffusion steps to obtain updated fused noise features. The updated fused noise features and the target fused features are encoded a second time to obtain a second noisy feature map; The second noisy feature map is downsampled to obtain the target noisy feature map.

5. The method according to claim 4, characterized in that, The first encoding of the fused noise features using the ozone concentration generation model to obtain a first noisy feature map includes: The fused noise features are convolved to obtain a fused noise feature map; The fused noise feature map is randomly discarded to obtain a discarded feature map; The discarded feature map is convolved to obtain the enhanced feature map; The enhanced feature map is then normalized to obtain a normalized feature map; The standardized feature map is activated to obtain the first noisy feature map.

6. The method according to claim 4, characterized in that, The stepwise denoising process, based on the diffusion steps, of the target noisy feature map and the convolutional noisy feature map to obtain the ozone concentration prediction set of the target grid includes: The target noisy feature map and the convolutional noisy feature map are transposed and convolved to obtain the first upsampled feature map; Based on the number of diffusion steps, the first upsampled feature map, the second noisy feature map, and the updated noise data are subjected to a first denoising process to obtain a first denoised feature map; The first denoised feature map is subjected to a second transposed convolution to obtain a second upsampled feature map; Based on the diffusion steps, the second upsampled feature map, the first noisy feature map, and the noise data are subjected to a second denoising process to obtain the ozone concentration prediction set of the target grid.

7. The method according to any one of claims 1 to 6, characterized in that, The process of interpolating the concentration matrix at the observation points to obtain a smoothed concentration matrix includes: Obtain each target interpolation point of the target grid, and perform Euclidean distance calculation on each target interpolation point and the ground observation point to obtain the target distance of the interpolation observation point; Based on the preset power parameter, the ozone concentration at the observation point, and the target distance of the interpolation observation point, ozone interpolation processing is performed on each of the target interpolation points to obtain the smoothed concentration matrix.

8. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the ozone forecasting method according to any one of claims 1 to 7.

9. A storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the ozone forecasting method according to any one of claims 1 to 7.

Images