A PM2.5 short-impending AI prediction method

Abstract

The application provides a PM2.5 short-term AI prediction method, and relates to the technical field of environmental air quality prediction. The method comprises the following steps: obtaining a multi-source data vector of each region at each time according to historical monitoring data of monitoring stations in multiple regions at multiple times; processing the multi-source data vector through a plurality of versions of a double-branch space-time prediction model to obtain a prediction data vector of each region at a plurality of future times; correcting the prediction data vector through an external tree integrated hour rolling correction model to obtain a corrected prediction data vector; determining the weight of each version of the double-branch space-time prediction model according to multi-dimensional meteorological data; and then combining the corrected prediction data vector to obtain a prediction data output vector of each partition. The method can realize dynamic prediction and rolling correction of PM2.5 cross-regional transmission and concentration change, and improve the accuracy, stability and scene adaptability of PM2.5 short-term prediction.

Description

Technical Field

[0001] This invention relates to the field of ambient air quality forecasting technology, and in particular to a short-term AI forecasting method for PM2.5. Background Technology

[0002] Existing technologies typically employ neural network models such as LSTM and static GNNs to predict PM2.5 concentrations in cities at multiple future times. Because these models use a fixed spatiotemporal modeling approach, they struggle to adjust the correlation between regions based on real-time wind field changes. They also struggle to simultaneously capture short-term pollution spikes and maintain stability over longer prediction periods. Furthermore, conventional loss functions tend to bias the model towards a large number of daily low-concentration samples, further weakening its ability to identify heavy pollution peaks and leading to missed or false alarms. Moreover, existing bias correction processes are usually coupled with the backbone prediction model, making it difficult to independently correct prediction biases using the latest real-time data and hindering the analysis and tracing of error sources. Therefore, existing technologies are prone to increased prediction errors, delayed responses to heavy pollution, and insufficient stability of prediction results under complex wind fields and pollution processes.

[0003] The information disclosed in the background section of this application is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0004] This invention provides a short-term AI forecasting method for PM2.5, which can solve the technical problems of existing technologies that are difficult to balance different forecast lead times, prone to missed or false forecasts, and difficult to correct and trace in a timely manner.

[0005] According to a first aspect of the present invention, a short-term AI forecasting method for PM2.5 is provided, comprising: Based on historical monitoring data from monitoring stations in multiple regions at multiple times, obtain multi-source data vectors for each region at each time. By using multiple versions of the bi-branch spatiotemporal prediction model, the multi-source data vectors of each region at each time are processed to obtain the predicted data vectors of each region at multiple future times. By integrating an hourly rolling correction model with an external tree, the predicted data vector is corrected to obtain the corrected predicted data vector. Acquire multidimensional meteorological data for each region, and determine the weights of each version of the bi-branch spatiotemporal prediction model based on the multidimensional meteorological data; Based on the weights of each version of the bi-branch spatiotemporal prediction model and the corrected prediction data vector, the prediction data output vector for each partition is obtained.

[0006] According to the present invention, by processing the multi-source data vector of each region at each time step using multiple versions of a bi-branch spatiotemporal prediction model, prediction data vectors for multiple future time steps of each region are obtained, including: The multi-source data vectors from multiple times in each region are concatenated to obtain the fused data vector for each region; After extracting features from the fused data vector of each region, the data is input into the spatial branch of the dual-branch spatiotemporal prediction model to obtain the first fused feature vector of each region. The first fused feature vector is decomposed to obtain the second fused feature vector at multiple time points for each region; The second fused feature vector of a single region at multiple times is input into the time branch of the dual-branch spatiotemporal prediction model to obtain the prediction feature vector of the single region. The predicted feature vector of a single region is mapped to obtain the predicted data vector of the single region at multiple future times. Obtain the predicted data vectors for multiple future times for each region from the output of each version of the bi-branch spatiotemporal prediction model.

[0007] According to the present invention, after feature extraction of the fused data vector of each region, it is input into the spatial branch of the dual-branch spatiotemporal prediction model to obtain the first fused feature vector of each region, including: Obtain wind speed data, wind direction data, and current PM2.5 concentration data for each region; Based on the geographical directions of the i-th and j-th regions, and the wind direction data of the i-th region, determine the angle between the geographical direction and the wind direction of the i-th region; Based on the included angle, the wind speed data, and the PM2.5 concentration data, determine the degree of influence of the i-th region on the j-th region; Based on the influence level, determine the neighboring regions of the j-th region; By using the spatial branch of the dual-branch spatiotemporal prediction model, attention operations are performed on the feature vector extracted from the fused data vector of each region and the feature vectors of its neighboring regions to obtain the first fused feature vector of each region.

[0008] According to the present invention, determining the influence degree of the i-th region on the j-th region based on the included angle, the wind speed data, and the PM2.5 concentration data includes: Based on the included angle and the wind speed data, determine the wind field influence information of the i-th region on the j-th region; The influence of region i on region j is determined based on the distance between region i and region j, the wind field influence information, and the PM2.5 concentration data.

[0009] According to the present invention, determining the wind field influence information of the i-th region on the j-th region based on the included angle and the wind speed data includes: According to the formula Determine the wind field impact information of region i on region j at the current time. ,in, For preset coefficients, Let i be the wind speed data for the i-th region at the current moment. Let be the angle between the geographical directions of the i-th and j-th regions and the wind direction of the i-th region, max be the function to maximize the value, and t be the current time.

[0010] According to the present invention, determining the degree of influence of the i-th region on the j-th region based on the distance between the i-th region and the j-th region, the wind field influence information, and the PM2.5 concentration data includes: According to the formula Determine the influence of region i on region j at the current time. ,in, Let be the distance between the i-th region and the j-th region. This provides information on the wind field impact of region i on region j at the current moment. Let represent the PM2.5 concentration data for the i-th region at the current time, and max is the function to take the maximum value.

[0011] According to the present invention, acquiring multidimensional meteorological data for various regions and determining the weights of various versions of the bi-branch spatiotemporal prediction model based on the multidimensional meteorological data includes: Set a contribution threshold; Multidimensional meteorological data and corrected forecast data vectors are input into the decision tree model to obtain the contribution scores of each version of the bi-branch spatiotemporal forecast model. A version of the bi-branch spatiotemporal prediction model that masks contribution scores below the contribution threshold; Output the weights of each version of the bi-branch spatiotemporal prediction model.

[0012] According to the present invention, the method further includes: The confidence interval of each predicted data in the predicted data output vector at each time step is determined by at least one of the error statistics method, the integrated extreme value method, and the quantile prediction method.

[0013] According to the present invention, the method further includes: Obtain the error of each predicted data point in the predicted data output vector for each partition; The bi-branch spatiotemporal prediction model for each version is fine-tuned and iterated based on the error.

[0014] According to a second aspect of the present invention, a short-term AI forecasting system for PM2.5 is provided, comprising: The first vector acquisition module obtains the multi-source data vector for each region at each time based on the historical monitoring data of monitoring stations in multiple regions at multiple times. The second vector acquisition module processes the multi-source data vectors of each region at each time step using multiple versions of the bi-branch spatiotemporal prediction model to obtain the predicted data vectors of each region at multiple future times. The third vector acquisition module uses an external tree to integrate an hourly rolling correction model to correct the predicted data vector and obtain the corrected predicted data vector. The weight calculation module acquires multidimensional meteorological data for each region and determines the weights of each version of the bi-branch spatiotemporal prediction model based on the multidimensional meteorological data. The data output module obtains the prediction data output vector for each partition based on the weights of the bi-branch spatiotemporal prediction model for each version and the corrected prediction data vector.

[0015] By adopting the above technical solution, the present invention can achieve the following technical effects: According to this invention, the pollution transport influence between different regions and neighboring regions with a strong possibility of pollution transport can be dynamically determined based on wind direction, wind speed, inter-regional distance, and PM2.5 concentration. This allows the spatial branch to selectively correlate and fuse the feature vectors of the target region itself and the feature vectors of its neighboring regions, resulting in a first fused feature vector that simultaneously includes both the pollution and meteorological characteristics of the target region. Simultaneously, the temporal branch extracts the concentration change characteristics of the same region over consecutive historical time periods, thereby enhancing the model's ability to represent the cross-regional transport and temporal evolution patterns of PM2.5 and improving the accuracy of short-term forecasts. By independently setting the deviation correction process outside the dual-branch spatiotemporal prediction model, hourly residuals are learned based on the obtained real-time monitoring data, meteorological factors, and original prediction results. As new real-time data is generated, the prediction results for subsequent time periods are rolled over for correction. This achieves independent correction of systematic prediction deviations and short-term residuals at stations without changing the overall prediction trend of the backbone model. This approach reduces systematic biases and short-term residuals at stations in the original prediction results without altering the core structure of the bi-branch spatiotemporal prediction model, thereby improving the accuracy, timeliness, and stability of short-term PM2.5 predictions. By inputting multidimensional meteorological data from various regions and the corrected prediction data vectors from each model version into the decision tree model, the predictive contribution of each model version under current meteorological conditions and pollution scenarios can be dynamically evaluated. Models with lower applicability are filtered out using contribution thresholds, and weights are determined based on the contributions of the remaining models. The corrected prediction results from each model version are then weighted and fused, allowing the model to flexibly select the predictive advantages of different versions based on the current prediction scenario. This avoids the problem of single models or fixed fusion rules being unable to adapt to complex pollution processes, improving the scenario adaptability, stability, and reliability of the final PM2.5 prediction results. Based on the forecast timeliness, pollution level, and risk quantification requirements, appropriate methods can be selected from error statistics, integrated extreme value methods, and quantile forecasting methods to determine the lower and upper limits of PM2.5 forecasts for each region at each forecast time. This allows the model to not only output PM2.5 point forecast values ​​but also further characterize the possible fluctuation range and uncertainty of the forecast results. This adapts to different application scenarios such as daily low-concentration forecasts, heavy pollution extreme processes, and short-term near-term forecasts, improving the confidence interval's coverage of the actual concentration change range and providing a more complete and reliable forecast basis for heavy pollution early warning, risk assessment, and dynamic air quality consultation.

[0016] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Other features and aspects of the invention will become clearer from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained based on these drawings without creative effort. Figure 1 A schematic flowchart of a short-term AI forecasting method for PM2.5 according to an embodiment of the present invention is shown as an example. Figure 2 An exemplary schematic diagram of the structure of a dual-branch spatiotemporal prediction model according to an embodiment of the present invention is shown; Figure 3A , Figure 3B and Figure 3C An exemplary diagram is shown comparing the mean absolute deviation of the upper and lower limits and the range coverage of three confidence interval determination methods under different PM2.5 real-time concentration ranges according to embodiments of the present invention. Figure 4 An exemplary diagram of a short-term AI forecasting system for PM2.5 according to an embodiment of the present invention is shown. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0020] Figure 1 An exemplary flowchart of a short-term AI forecasting method for PM2.5 according to an embodiment of the present invention is shown, the method comprising: Step S1: Based on the historical monitoring data of monitoring stations in multiple regions at multiple times, obtain the multi-source data vector of each region at each time. Step S2: Using multiple versions of the bi-branch spatiotemporal prediction model, the multi-source data vectors of each region at each time step are processed to obtain the predicted data vectors of each region at multiple future times. Step S3: The predicted data vector is corrected by integrating an external tree into an hourly rolling correction model to obtain the corrected predicted data vector. Step S4: Obtain multidimensional meteorological data for each region, and determine the weights of each version of the bi-branch spatiotemporal prediction model based on the multidimensional meteorological data. Step S5: Based on the weights of each version of the bi-branch spatiotemporal prediction model and the corrected prediction data vector, obtain the prediction data output vector for each partition.

[0021] According to the PM2.5 short-term AI forecasting method of the present invention, multi-source data vectors for each region at each time moment are obtained based on historical monitoring data from monitoring stations in multiple regions at multiple times. These multi-source data vectors are then processed using multiple versions of a bi-branch spatiotemporal prediction model to obtain predicted data vectors for each region at multiple future times. An external tree-integrated hourly rolling correction model is then used to correct the predicted data vectors, resulting in corrected predicted data vectors. The weights of each version of the bi-branch spatiotemporal prediction model are determined based on multi-dimensional meteorological data. Finally, the corrected predicted data vectors are combined to obtain the predicted data output vectors for each region. This method enables dynamic prediction and rolling correction of PM2.5 cross-regional transport and concentration changes, improves the compatibility of predictions with different timeframes, and enhances the accuracy, stability, and scenario adaptability of short-term PM2.5 forecasts.

[0022] According to an embodiment of the present invention, in step S1, a multi-source data vector for each region at each time moment is obtained based on historical monitoring data from monitoring stations in multiple regions at multiple times. The multiple regions can be different cities or areas where different monitoring stations are located, and the time interval between the multiple times can be one hour. For example, hourly pollutant monitoring data from 334 national-level monitoring stations in 2018 to March 2026 can be obtained, including six types of pollutant parameters: PM2.5, PM10, SO2, NO2, CO, and O3. Thus, each city can obtain pollutant concentration information corresponding to each historical time moment. Correspondingly, meteorological data corresponding to each city and each time moment can also be obtained. For example, meteorological data can be derived from the GFS 0.25° global forecast field, which is reported daily from 00 UTC, and meteorological parameters such as temperature, humidity, wind speed, wind direction, air pressure, and precipitation can be extracted from it. The wind speed and wind direction data can be used to subsequently determine the pollutant transport relationship between different regions. The multi-source data may also include underlying surface data corresponding to each region. For example, underlying surface data includes city coordinates, elevation, land use percentage, and population density. City coordinates can be used to characterize the spatial location of different cities and provide basic data for subsequently determining the geographical distance and direction between different regions. Furthermore, a pollution process case library can be constructed based on historical PM2.5 concentration data. For instance, a pollution process can be defined as a period when PM2.5 concentration reaches a light pollution level for two consecutive hours, and this pollution process can be divided into an increasing phase, a peak phase, and a decreasing phase. Further indicators such as the start time, end time, peak concentration, and cumulative concentration of the pollution process can be extracted.

[0023] After obtaining the aforementioned multi-source data, standardized preprocessing can be performed. Specifically, firstly, detrending processing is performed, for example, using the 2025-2026 average as a benchmark to eliminate concentration shifts caused by annual emission reductions. Then, anomaly truncation processing is applied to pollutant concentrations. Specifically, data with concentrations below 1 are set to 1, and data with concentrations above 1000 are set to 1000. Further, logarithmic transformation is performed on the truncated concentration data to improve the right-skewed long-tail distribution of the concentration data and balance the loss weights corresponding to high and low concentration data. Subsequently, Z-score standardization is performed on the data to make the processed data suitable for subsequent deep learning models. Further, spatiotemporal alignment is performed on different data sources. For GFS meteorological grid data, nearest neighbor interpolation is used to map the corresponding grid data to each monitoring station. At the same time, the UTC time corresponding to the GFS data is converted to the time of the measured area to ensure that the meteorological data and the hourly data of the monitoring stations are consistent in the time dimension.

[0024] According to an embodiment of the present invention, in step S2, the multi-source data vectors of each region at each time step are processed by multiple versions of the bi-branch spatiotemporal prediction model to obtain the predicted data vectors of each region at multiple future time steps. This includes: concatenating the multi-source data vectors of each region at multiple time steps to obtain a fused data vector for each region; extracting features from the fused data vectors of each region and inputting them into the spatial branch of the bi-branch spatiotemporal prediction model to obtain a first fused feature vector for each region; decomposing the first fused feature vectors to obtain a second fused feature vector for each region at multiple time steps; inputting the second fused feature vectors of a single region at multiple time steps into the temporal branch of the bi-branch spatiotemporal prediction model to obtain a predicted feature vector for a single region; mapping the predicted feature vectors of a single region to obtain the predicted data vectors of a single region at multiple future time steps; and obtaining the predicted data vectors of each region at multiple future time steps output by each version of the bi-branch spatiotemporal prediction model.

[0025] According to an embodiment of the present invention, each region corresponds to a multi-source data vector at each historical moment. Following chronological order, the multi-source data vectors of the same region at multiple past moments are concatenated, transforming the data from those moments into a long-dimensional fused data vector. Then, through a feedforward network or other feature mapping network, features are extracted and dimensionality is unified for the fused data vectors of each region. The feature vectors corresponding to all regions are then arranged according to region dimension and input into the spatial branch of the two-branch spatiotemporal prediction model. An attention mechanism is then used to correlate the feature vectors of its neighboring regions, thereby extracting the spatial correlation and transmission features of pollutants between different regions, resulting in a first fused feature vector corresponding to each region. Neighboring regions can be dynamically selected based on their spatial location and real-time meteorological conditions; the specific method for determining neighboring regions will be further explained later. The feature mapping network can be a multilayer perceptron, and the two-branch spatiotemporal prediction model can be a Transformer-based deep learning neural network model. The spatial and temporal branches of the two-branch spatiotemporal prediction model can each employ a Transformer-based encoding network. They are used to independently model the spatial correlation characteristics between regions and the temporal correlation characteristics between different historical moments in the same region, thereby extracting the cross-regional transport characteristics of pollutants and the evolution characteristics of PM2.5 over time.

[0026] The first fusion feature vector is a region-level spatial fusion feature obtained through spatial attention operations. Before inputting the time branch, it can be mapped through a feedforward network to make its dimension correspond to the number of historical moments, and then split according to different historical moments, and reconstituted into multiple second fusion feature vectors arranged in time. Each second fusion feature vector corresponds to a historical moment of the region and contains the regional association information extracted from the spatial branch.

[0027] Furthermore, the second fused feature vectors from multiple moments in a single region are input into the time branch of the dual-branch spatiotemporal prediction model in chronological order. A time attention mechanism is used to analyze the correlation between the same region at different historical moments, and parallel temporal coding is employed to extract the temporal variation characteristics of PM2.5 in continuous historical moments, such as intraday variation patterns and weekly evolution patterns, thereby obtaining the corresponding prediction feature vector for that region. The prediction feature vector of a single region is then mapped to obtain the prediction data vector for multiple future moments in that region. This prediction data vector can include the predicted PM2.5 concentrations for multiple consecutive future moments in that region. Finally, each version of the dual-branch spatiotemporal prediction model predicts data for all regions, outputting its corresponding prediction data vector for multiple future moments in each region.

[0028] The various versions of the bi-branch spatiotemporal prediction model can be trained using different multi-step prediction methods or loss functions. Multi-step prediction methods can include autoregressive (AR) prediction and parallel non-autoregressive (NAR) prediction. When using AR prediction, PM2.5 predictions for subsequent times are obtained by recursively extrapolating hourly based on the prediction time. It has a strong ability to capture concentration abrupt changes during the short-term prediction of 0–8 hours, but it is prone to long-term error accumulation as the extrapolation time increases. In contrast, when using NAR prediction, predictions for all prediction times can be output in parallel at once without hourly extrapolation. Therefore, there is no error accumulation caused by extrapolation during the 9–23 hour prediction period, and the long-term prediction is relatively stable. Furthermore, a differentiated loss function can be used during model training. Specifically, the base loss function can be the mean squared error (MSE) loss. Based on this, a weight greater than 1 is set for high-pollution samples to increase the weight of heavily polluted samples in the loss function, amplifying the gradient corresponding to heavily polluted samples and thus reducing the underreporting of heavy pollution processes. For some model versions, an additional false alarm penalty term is added to the high-pollution weighted loss (e.g., increasing the weight of the loss term corresponding to a false alarm) to suppress false alarms or alarms under clean weather conditions. Furthermore, based on different combinations of the above prediction methods and loss functions, six versions of bi-branch spatiotemporal prediction models can be constructed. The prediction methods and training loss functions for each version of the bi-branch spatiotemporal prediction model include: AR and base loss (Model A); NAR and base loss (Model B); NAR and pollution weighting (Model C); AR and pollution weighting (Model D); AR and pollution weighting plus false alarm penalty (Model E); and NAR and pollution weighting plus false alarm penalty (Model F). Thus, based on different prediction methods and loss functions, six versions of bi-branch spatiotemporal prediction models can be obtained. Each version of the model obtains prediction data vectors for multiple future times in the corresponding region, which serve as the basis for subsequent external tree integration hourly rolling correction and multi-model dynamic fusion.

[0029] According to an embodiment of the present invention, the following describes a method for filtering neighboring regions. After feature extraction of the fused data vector of each region, the data is input into the spatial branch of a bi-branch spatiotemporal prediction model to obtain the first fused feature vector of each region. This includes: acquiring wind speed data, wind direction data, and PM2.5 concentration data at the current time for each region; determining the angle between the geographical direction and the wind direction of the i-th region based on the geographical directions of the i-th and j-th regions and the wind direction data of the i-th region; determining the influence degree of the i-th region on the j-th region based on the angle, the wind speed data, and the PM2.5 concentration data; determining the neighboring regions of the j-th region based on the influence degree; and performing attention operations on the feature vector of the fused data vector of each region and the feature vectors of its neighboring regions through the spatial branch of the bi-branch spatiotemporal prediction model to obtain the first fused feature vector of each region.

[0030] According to an embodiment of the present invention, wind speed data, wind direction data, and current PM2.5 concentration data for each region are acquired. Wind speed and wind direction are used to characterize the possible transport direction and capacity of pollutants, while the current PM2.5 concentration is used to characterize the intensity of pollutants that may be transported outward from the corresponding region. Further, based on the i-th region and its geographical location, a geographical direction from the i-th region to the j-th region is determined, and the angle between this geographical direction and the current wind direction of the i-th region is calculated. The included angle This is used to characterize the degree of consistency between the current airflow direction of region i and the direction of region j, thereby determining whether the airflow in region i is conducive to transporting pollutants to region j.

[0031] Further, based on the included angle, wind speed data, and PM2.5 concentration data, the influence degree of the i-th region on the j-th region is determined. The influence degree comprehensively represents the probability and intensity of pollutant transport from the i-th region to the j-th region under the current wind field conditions. Then, for the j-th region, the influence degrees of other candidate regions on the j-th region are calculated to determine the neighboring regions of the j-th region. For example, in a prediction scenario including 334 regions, the M regions with the highest influence degree can be selected as neighboring regions for each target region. Here, M represents the number of neighboring regions participating in subsequent spatial attention calculations. Neighboring regions are not necessarily the geographically closest regions, but rather regions that, under the current wind direction, wind speed, and pollution concentration conditions, have a strong pollution transport influence on the target region. Since wind direction, wind speed, and PM2.5 concentration change over time, the influence of each candidate region on the target region can be recalculated at hourly intervals. Based on this, the neighboring regions corresponding to each target region are dynamically updated, i.e., the regions with the greatest influence on the target region. This avoids the problem that regional relationships cannot change with real-time pollution transport conditions when using static spatial partitioning or fixed graph topology, achieving wind-driven dynamic spatial neighborhood selection. After determining the neighboring regions of the j-th region, the feature vector obtained by feature extraction from the fused data vector of the j-th region, and the feature vectors obtained by feature extraction from the fused data vectors of each neighboring region, are input together into the spatial branch of the dual-branch spatiotemporal prediction model. The spatial branch analyzes the degree of feature correlation between the target region and each neighboring region through an attention mechanism, and selectively fuses the feature information of the neighboring regions to obtain the first fused feature vector of the j-th region. The first fused feature vector includes both the PM2.5 and meteorological features of the j-th region itself, as well as the spatial correlation features of neighboring regions that may transport pollutants to the j-th region at the current moment. After performing the above attention operation on all regions, the first fused feature vector of each region is obtained. By using the dynamic neighbor region selection method described above, in a prediction scenario containing N regions, it is unnecessary to perform spatial attention calculations between each target region and all other regions. Instead, only the M (M < N) neighbor regions with the greatest influence are retained for subsequent calculations. This reduces the scale of spatial association calculations compared to pairwise associations of all regions. Reduce to While reducing the computational cost of spatial attention, the spatial branch can focus on extracting regional information that has a strong relationship with pollution transport to the target region, thereby enhancing the model's ability to represent cross-regional pollution transport characteristics.

[0032] According to an embodiment of the present invention, the above-mentioned influence degree is calculated as follows: the influence degree of the i-th region on the j-th region is determined based on the included angle, the wind speed data, and the PM2.5 concentration data, including: determining the wind field influence information of the i-th region on the j-th region based on the included angle and the wind speed data; and determining the influence degree of the i-th region on the j-th region based on the distance between the i-th region and the j-th region, the wind field influence information, and the PM2.5 concentration data.

[0033] According to an embodiment of the present invention, after obtaining the angle between the wind direction of the i-th region and the geographical direction from the i-th region to the j-th region, the process of determining the influence of the i-th region on the j-th region can be divided into two steps.

[0034] First, based on the included angle and the current wind speed data of the i-th region, the wind field influence information of the i-th region on the j-th region is determined. This wind field influence information characterizes the ability of airflow to transport pollutants from the i-th region to the j-th region under the current wind direction and speed conditions. Specifically, the relationship between the wind direction and the geographical direction from the i-th region to the j-th region reflects whether the airflow direction is conducive to the transport of pollutants to the j-th region, while the wind speed reflects the current transport capacity of the airflow. Further, based on the wind field influence information, and combining the distance between the i-th and j-th regions and the current PM2.5 concentration data of the i-th region, the degree of influence of the i-th region on the j-th region is determined. Specifically, the inter-regional distance characterizes the distance attenuation effect experienced by pollutants during transport from the i-th region to the j-th region, and the current PM2.5 concentration of the i-th region characterizes the intensity of pollutants that can currently be transported outward from that region. Therefore, the impact degree comprehensively considers the current wind field conditions, the spatial distance between regions, and the pollutant concentration, and is used to characterize the possibility and intensity of the impact of pollutants from region i being transported to region j under the current conditions.

[0035] According to an embodiment of the present invention, determining the wind field influence information of the i-th region on the j-th region based on the included angle and the wind speed data includes: determining the wind field influence information of the i-th region on the j-th region at the current moment according to formula (1). , (1) in, For preset coefficients, Let i be the wind speed data for the i-th region at the current moment. Let be the angle between the geographical directions of the i-th and j-th regions and the wind direction of the i-th region, max be the function to maximize the value, and t be the current time.

[0036] According to an embodiment of the present invention, This indicates the degree to which the current wind speed in the i-th region enhances the pollutant transport capacity. 1 is the base influence value, indicating that when the wind speed is extremely low, the influence of the base wind field is retained. It is the angle between the regional geographical direction and the airflow transport direction. When the angle is close to At that time, the direction of airflow transmission is basically consistent with the direction of the target area. When the angle is close to 1, the transmission effect is strongest, and the included angle is close to 1. The airflow direction is approximately orthogonal to the target area direction, making it difficult for pollutants to travel to the first... The area has an included angle close to The airflow direction is opposite to the target area direction, which is not conducive to the transfer of pollutants to the j-th area. The minimum value used to set the wind direction influence can be expressed as a minimum directional influence that needs to be maintained even if the airflow direction is inconsistent with the target area direction. This indicates that factors such as wind direction fluctuations or pollutant diffusion may still have some impact. Multiplying the above two values ​​together can comprehensively reflect the degree to which the current wind field is beneficial in transporting pollutants from region i to region j; that is, the wind field influence radius. A larger wind field influence information indicates that the current wind field has a strong pollutant transport capacity. When the wind speed is low or the airflow direction deviates from region j, a smaller wind field influence information is obtained, indicating that the current wind field has a weaker promoting effect on the transport of pollutants from region i to region j.

[0037] According to an embodiment of the present invention, determining the influence degree of the i-th region on the j-th region based on the distance between the i-th region and the j-th region, the wind field influence information, and the PM2.5 concentration data includes: determining the influence degree of the i-th region on the j-th region at the current time according to formula (2). , (2) in, Let be the distance between the i-th region and the j-th region. This provides information on the wind field impact of region i on region j at the current moment. Let represent the PM2.5 concentration data for the i-th region at the current time, and max is the function to take the maximum value.

[0038] According to an embodiment of the present invention, in formula (2), the geographical distance between regions and the effective transport scale of the current wind field are first calculated, and the process is as follows: As shown, when When this occurs, it indicates that the j-th region is within the effective transport scale of the current wind field in the i-th region. ,at this time Therefore, the influence of the i-th region on the j-th region is determined to be the maximum value. When the distance exceeds the effective transport scale of the current wind field, it indicates that the j-th region exceeds the effective transport scale of the current wind field. The greater the distance, the greater the degree of diffusion, dilution and sedimentation of pollutants during transport, and therefore the smaller the impact. This can represent the degree of distance attenuation relative to the current wind farm's transport capacity. Furthermore, it utilizes the PM2.5 concentration of the i-th region at the current moment. Adjust the degree of distance attenuation, i.e. This value can represent the relative distance attenuation of pollutant transport. When the PM2.5 concentration in the i-th region is high, it indicates that the region can currently transport more pollutants outward, and the relative distance attenuation is relatively smaller. Conversely, when the PM2.5 concentration in the i-th region is low, it indicates that the region can transport fewer pollutants outward, and the relative distance attenuation is relatively larger. Finally, the adjusted distance attenuation is negatively calculated and exponentially applied to obtain the influence of the i-th region on the j-th region at the current time. The closer the influence degree is to 0, the greater the attenuation effect of pollutants during long-distance transmission; the closer the influence degree is to 1, the smaller the attenuation effect of pollutants during long-distance transmission.

[0039] In this way, based on wind direction, wind speed, inter-regional distance, and PM2.5 concentration, the influence of pollution transport between different regions and neighboring regions with a strong possibility of pollution transport are dynamically determined. This allows the spatial branch to selectively correlate and fuse the feature vectors of the target region itself and the feature vectors of its neighboring regions, so that the resulting first fused feature vector simultaneously includes the pollution and meteorological characteristics of the target region itself. At the same time, the temporal branch is used to extract the concentration change characteristics of the same region over consecutive historical moments, thereby enhancing the model's ability to represent the cross-regional transport and temporal evolution of PM2.5 and improving the accuracy of short-term forecasts.

[0040] According to an embodiment of the present invention, in step S3, the predicted data vector is corrected using an external tree-integrated hourly rolling correction model to obtain a corrected predicted data vector. This external tree-integrated hourly rolling correction model is independently set outside the dual-branch spatiotemporal prediction model and is used to correct the deviation of the original predicted values ​​for each future time period output by the dual-branch spatiotemporal prediction model. Specifically, the original predicted data vector for the next 0-23 hours output by the dual-branch spatiotemporal prediction model, the actual monitoring data for multiple times, and the meteorological factors corresponding to each predicted time period can be input into the external tree-integrated hourly rolling correction model to learn the deviation pattern between the original predicted values ​​and the actual monitoring values, thereby predicting the residuals for each future time period. The predicted residuals are then added to the corresponding original predicted values ​​to obtain the corrected predicted values. For example, the external tree-integrated hourly rolling correction model can employ lightweight tree ensemble models such as Random Forest, XGBoost, or CatBoost to fit the nonlinear relationship between the original predicted results, actual monitoring information, synchronous meteorological factors, and the predicted residuals. As new real-time monitoring data is generated, the updated real-time monitoring information can be used to re-predict the residuals for subsequent prediction times, and the corresponding corrected prediction results can be updated hourly, thus forming an hourly rolling correction process.

[0041] In this approach, by independently setting the bias correction process outside the bi-branch spatiotemporal prediction model, hourly residuals are learned based on the obtained real-time monitoring data, meteorological factors, and original prediction results. As new real-time data is generated, the prediction results for subsequent time periods are continuously corrected. This allows for independent correction of systematic prediction biases and short-term station residuals without altering the overall prediction trend of the main model. Thus, without changing the main structure of the bi-branch spatiotemporal prediction model, systematic biases and short-term station residuals in the original prediction results are reduced, improving the accuracy, timeliness, and stability of short-term PM2.5 predictions.

[0042] According to an embodiment of the present invention, in step S4, multidimensional meteorological data of each region is acquired, and the weights of each version of the bi-branch spatiotemporal prediction model are determined based on the multidimensional meteorological data, including: setting a contribution threshold; inputting the multidimensional meteorological data and the corrected prediction data vector into the decision tree model to obtain the contribution scores of each version of the bi-branch spatiotemporal prediction model; filtering out bi-branch spatiotemporal prediction models with contribution scores lower than the contribution threshold; and outputting the weights of each version of the bi-branch spatiotemporal prediction model.

[0043] According to embodiments of the present invention, two schemes are provided to achieve the fusion of the final prediction results: fixed-rule splicing and machine learning dynamic fusion. In the fixed-rule splicing scheme, a fixed rule based on typical wind direction scenarios can be used for model selection. Specifically, in the southerly wind scenario, corresponding to the northward transport of pollution from the south accompanied by concentration fluctuations, Model A can be used for all prediction periods (i.e., the PM2.5 prediction data output by Model A is used as the final prediction data); in the westerly wind scenario, corresponding to the steady decrease in concentration under the influence of clean airflow from the northwest, Model B can be used for all prediction periods; in the northerly wind scenario, corresponding to the phased changes of convergence in front of mountains, Model A can be used from 0 to 8 hours, Model B from 9 to 13 hours and 21 to 23 hours, and Model D from 14 to 20 hours; in the easterly wind scenario, corresponding to the alternating disturbances of sea and land winds, Model B can be used from 0 to 6 hours and 14 to 20 hours, and Model A from 7 to 13 hours and 21 to 23 hours. The above model selections are merely examples, and the present invention does not limit them.

[0044] Furthermore, in the machine learning dynamic fusion scheme, the weights of each version of the bi-branch spatiotemporal prediction model can be determined using a machine learning dynamic fusion approach. First, a contribution threshold is set. Then, multi-dimensional meteorological data such as real-time wind direction, wind speed, boundary layer height, and precipitation for each region during the current prediction period are acquired. This multi-dimensional meteorological data, along with the prediction data vectors of each version of the bi-branch spatiotemporal prediction model after correction by an external tree ensemble hourly rolling correction model, are input into a pre-trained decision tree model. The decision tree model can be, for example, a LightGBM or CatBoost ensemble model. Based on the current meteorological scenario and the corrected prediction results of each version of the model, a contribution score is obtained for each version. This contribution score characterizes the degree to which the corresponding version of the model contributes to the fusion of the final prediction results under the current meteorological scenario. A higher contribution score indicates that the corresponding version of the model is more suitable for participating in the fusion of prediction results under the current scenario. Furthermore, the contribution scores of each model version are compared with a pre-set contribution threshold. Model versions with contribution scores below the threshold are masked during the current prediction result fusion process, preventing them from participating in the final prediction result fusion for the current scenario. For model versions with contribution scores not lower than the threshold, corresponding fusion weights are determined based on their contribution scores (e.g., normalizing the contribution scores of model versions with scores not lower than the threshold), thus outputting the weights of each version of the bi-branch spatiotemporal prediction model in the current prediction scenario. Subsequently, the corrected prediction data vectors corresponding to each unmasked model version are fused (weighted summation) according to the weights to obtain the final prediction data output vector. In addition, when using a machine learning dynamic fusion method, a permutation importance analysis method can be introduced to analyze the contribution of meteorological features such as real-time wind direction, wind speed, boundary layer height, and precipitation, as well as the prediction features of each model version, to determine the degree of influence of different meteorological features and model features on the dynamic fusion results.

[0045] According to an embodiment of the present invention, in step S5, the prediction data output vector for each partition is obtained based on the weights of each version of the bi-branch spatiotemporal prediction model and the corrected prediction data vector. Specifically, after obtaining the weights of each version of the bi-branch spatiotemporal prediction model and the corrected prediction data vector, the corrected prediction data vectors of each version model are weighted and summed using the weights, and the prediction results of multiple version models are dynamically fused to obtain the prediction data output vector for each partition.

[0046] In this way, by inputting multidimensional meteorological data from various regions and the corrected prediction data vectors of various model versions into the decision tree model, the predictive contribution of each model version under the current meteorological conditions and pollution scenarios can be dynamically evaluated. Models with low applicability are filtered out using contribution thresholds, and the corresponding weights are determined based on the contribution of the remaining models. The corrected prediction results of each model version are then weighted and fused, enabling the model to flexibly select the predictive advantages of different versions according to the current prediction scenario. This avoids the problem that a single model or fixed fusion rules are difficult to adapt to complex pollution processes, and improves the scenario adaptability, stability, and reliability of the final PM2.5 prediction results.

[0047] According to an embodiment of the present invention, the method further includes: determining the confidence interval of each predicted data in the predicted data output vector at each time step by using at least one of the error statistics method, the integrated extreme value method, and the quantile prediction method.

[0048] According to an embodiment of the present invention, for each prediction time in the prediction data output vector of each region, at least one method can be selected from error statistics method, integrated extreme value method and quantile prediction method to determine the confidence interval of each prediction data according to different business scenarios such as prediction timeliness, pollution level and risk quantification requirements. Among them, the error statistics method (Method 1) can statistically analyze the distribution characteristics of errors under different prediction timeframes based on the errors between historical predicted values ​​and corresponding actual monitoring values, determine the error range according to a preset confidence level, and then combine the error range with the current predicted value to obtain a confidence interval, which is suitable for daily low-concentration prediction scenarios; the ensemble extreme value method (Method 2) can obtain multiple corrected predicted values ​​output by multiple tree ensemble correction models for the same prediction time, and use the minimum and maximum values ​​as the lower and upper limits of the confidence interval, respectively, thereby improving the coverage of the true concentration under extreme scenarios such as heavy pollution; the quantile prediction method (Method 3) can use the pinball loss function to train prediction models corresponding to the lower and upper quantiles, respectively, so that the model directly outputs the lower and upper quantiles of the concentration at each prediction time, and uses them as the lower and upper limits of the confidence interval, which is suitable for short-term near-term prediction of 1-6 hours. Figure 3A , Figure 3B and Figure 3C As shown, Method 1, Method 2, and Method 3 represent the error statistics method, the integrated extreme value method, and the quantile prediction method, respectively. By comparing the mean absolute deviation and range coverage of the upper and lower limits of the three methods under different PM2.5 real concentration (y_real) intervals and prediction lead times, the appropriate confidence interval determination method can be selected based on the current pollution level and prediction lead time. The confidence interval refers to the lower and upper prediction limits determined for the predicted PM2.5 value at each prediction time, used to characterize the possible fluctuation range and uncertainty of the corresponding prediction result.

[0049] In this way, appropriate methods can be selected from error statistics, integrated extreme value method, and quantile prediction method according to the forecast timeliness, pollution level, and risk quantification requirements. The lower and upper limits of PM2.5 prediction for each region at each prediction time can be determined. This allows the model to not only output the predicted PM2.5 values, but also further characterize the possible fluctuation range and uncertainty of the prediction results. This makes it suitable for different application scenarios such as daily low concentration prediction, heavy pollution extreme processes, and short-term near-term prediction. It also improves the ability of the confidence interval to cover the actual concentration change range, and provides a more complete and reliable prediction basis for heavy pollution early warning, risk assessment, and dynamic air quality consultation.

[0050] According to an embodiment of the present invention, the method further includes: obtaining the error of each prediction data in the prediction data output vector of each partition; and fine-tuning and iterating the bi-branch spatiotemporal prediction model of each version based on the error.

[0051] According to an embodiment of the present invention, after obtaining the prediction data for each prediction time in each region, the hourly prediction results and pollution process prediction results can be evaluated by combining the real-time monitoring results at the corresponding time and the forecaster's manual evaluation. Forecasters manually score the prediction results and mark omissions, false alarms, and peak deviations, etc., and store the corresponding cases in a database to form manually labeled samples used for subsequent model iterations. Furthermore, a reward model is constructed using the manually labeled samples, and the manually labeled samples and the reward model are used in reverse for the next round of model fine-tuning iterations to continuously update the dual-branch spatiotemporal prediction models of each version. In addition, during the above process, multi-model benchmarking can be performed, abnormal deviations can be automatically alerted, and pollution process reviews can be conducted monthly. The corresponding evaluation and labeling results can continue to be used for subsequent model iterations, thereby forming a continuously iterative and optimized human-machine collaborative business closed loop.

[0052] According to an embodiment of the present invention, a short-term AI forecasting method for PM2.5 dynamically determines the pollution transport influence between different regions and neighboring regions with a strong possibility of pollution transport based on wind direction, wind speed, inter-regional distance, and PM2.5 concentration. This allows the spatial branch to selectively correlate and fuse the feature vectors of the target region itself and its neighboring regions, resulting in a first fused feature vector that simultaneously includes both the pollution and meteorological characteristics of the target region. Simultaneously, the temporal branch extracts the concentration change characteristics of the same region over consecutive historical time periods, thereby enhancing the model's ability to represent the cross-regional transport and temporal evolution patterns of PM2.5 and improving the accuracy of short-term forecasts. By independently setting the deviation correction process outside the dual-branch spatiotemporal prediction model, hourly residuals are learned based on the obtained real-time monitoring data, meteorological factors, and original prediction results. As new real-time data is generated, the prediction results for subsequent time periods are rolled over for correction. This achieves independent correction of systematic prediction deviations and short-term residuals at stations without changing the overall prediction trend of the backbone model. This approach reduces systematic biases and short-term residuals at stations in the original prediction results without altering the core structure of the bi-branch spatiotemporal prediction model, thereby improving the accuracy, timeliness, and stability of short-term PM2.5 predictions. By inputting multidimensional meteorological data from various regions and the corrected prediction data vectors from each model version into the decision tree model, the predictive contribution of each model version under current meteorological conditions and pollution scenarios can be dynamically evaluated. Models with lower applicability are filtered out using contribution thresholds, and weights are determined based on the contributions of the remaining models. The corrected prediction results from each model version are then weighted and fused, allowing the model to flexibly select the predictive advantages of different versions based on the current prediction scenario. This avoids the problem of single models or fixed fusion rules being unable to adapt to complex pollution processes, improving the scenario adaptability, stability, and reliability of the final PM2.5 prediction results. Based on the forecast timeliness, pollution level, and risk quantification requirements, appropriate methods can be selected from error statistics, integrated extreme value methods, and quantile forecasting methods to determine the lower and upper limits of PM2.5 forecasts for each region at each forecast time. This allows the model to not only output PM2.5 point forecast values ​​but also further characterize the possible fluctuation range and uncertainty of the forecast results. This adapts to different application scenarios such as daily low-concentration forecasts, heavy pollution extreme processes, and short-term near-term forecasts, improving the confidence interval's coverage of the actual concentration change range and providing a more complete and reliable forecast basis for heavy pollution early warning, risk assessment, and dynamic air quality consultation.

[0053] Figure 2An exemplary schematic diagram of the structure of the dual-branch spatiotemporal prediction model according to an embodiment of the present invention is shown. Meteorological data, pollutant data, and static data can be preprocessed and then input into the dual-branch spatiotemporal prediction module. First, spatial correlation features between different regions are extracted through the spatial attention branch, and then temporal correlation features between the same region at different historical moments are extracted through the temporal attention branch. Subsequently, the obtained spatiotemporal features are sequentially subjected to layer normalization and mapping processing to obtain prediction results for multiple future moments of the corresponding region.

[0054] Figure 3A , Figure 3B and Figure 3C An exemplary diagram illustrates a comparison of the mean absolute deviations of the upper and lower limits of three confidence interval determination methods under different PM2.5 real-time concentration ranges according to embodiments of the present invention, along with the range coverage. Specifically, it compares the mean absolute deviations of the upper and lower limits of the confidence intervals for prediction lead times of 3h, 6h, 12h, and 23h, as well as the range coverage for each prediction time from 0 to 23h. This is particularly relevant in low-concentration scenarios. Under these conditions, Method 1 has a higher and more stable coverage rate; as PM2.5 concentration increases, and In the pollution scenario, Method 2 showed higher range coverage, but the deviation range of the upper and lower limits of the corresponding confidence interval increased. The three methods showed different performances under different concentration ranges and prediction timelines. Therefore, the appropriate confidence interval calculation method can be selected according to different pollution scenarios to obtain the corresponding hourly predicted concentration and upper and lower confidence boundaries.

[0055] Figure 4 An exemplary schematic diagram of a short-term AI forecasting system for PM2.5 according to an embodiment of the present invention is shown, the system comprising: The first vector acquisition module obtains the multi-source data vector for each region at each time based on the historical monitoring data of monitoring stations in multiple regions at multiple times. The second vector acquisition module processes the multi-source data vectors of each region at each time step using multiple versions of the bi-branch spatiotemporal prediction model to obtain the predicted data vectors of each region at multiple future times. The third vector acquisition module uses an external tree to integrate an hourly rolling correction model to correct the predicted data vector and obtain the corrected predicted data vector. The weight calculation module acquires multidimensional meteorological data for each region and determines the weights of each version of the bi-branch spatiotemporal prediction model based on the multidimensional meteorological data. The data output module obtains the prediction data output vector for each partition based on the weights of the bi-branch spatiotemporal prediction model for each version and the corrected prediction data vector.

[0056] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0057] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are merely examples and do not limit the present invention. The objectives of the present invention have been fully and effectively achieved. The functions and structural principles of the present invention have been demonstrated and explained in the embodiments, and any variations or modifications may be made to the implementation of the present invention without departing from the stated principles.

[0058] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A short-term AI forecasting method for PM2.5, characterized in that, include: Based on historical monitoring data from monitoring stations in multiple regions at multiple times, obtain multi-source data vectors for each region at each time. By using multiple versions of the bi-branch spatiotemporal prediction model, the multi-source data vectors of each region at each time are processed to obtain the predicted data vectors of each region at multiple future times. By integrating an hourly rolling correction model with an external tree, the predicted data vector is corrected to obtain the corrected predicted data vector. Acquire multidimensional meteorological data for each region, and determine the weights of each version of the bi-branch spatiotemporal prediction model based on the multidimensional meteorological data; Based on the weights of each version of the bi-branch spatiotemporal prediction model and the corrected prediction data vector, the prediction data output vector for each partition is obtained.

2. The PM2.5 short-term AI forecasting method according to claim 1, characterized in that, By using multiple versions of the bi-branch spatiotemporal prediction model, the multi-source data vectors for each region at each time step are processed to obtain the predicted data vectors for each region at multiple future time steps, including: The multi-source data vectors from multiple times in each region are concatenated to obtain the fused data vector for each region; After extracting features from the fused data vector of each region, the data is input into the spatial branch of the dual-branch spatiotemporal prediction model to obtain the first fused feature vector of each region. The first fused feature vector is decomposed to obtain the second fused feature vector at multiple time points for each region; The second fused feature vector of a single region at multiple times is input into the time branch of the dual-branch spatiotemporal prediction model to obtain the prediction feature vector of the single region. The predicted feature vector of a single region is mapped to obtain the predicted data vector of the single region at multiple future times. Obtain the predicted data vectors for multiple future times for each region from the output of each version of the bi-branch spatiotemporal prediction model.

3. The PM2.5 short-term AI forecasting method according to claim 2, characterized in that, After extracting features from the fused data vector of each region, the vector is input into the spatial branch of the dual-branch spatiotemporal prediction model to obtain the first fused feature vector of each region, including: Obtain wind speed data, wind direction data, and current PM2.5 concentration data for each region; Based on the geographical directions of the i-th and j-th regions, and the wind direction data of the i-th region, determine the angle between the geographical direction and the wind direction of the i-th region; Based on the included angle, the wind speed data, and the PM2.5 concentration data, determine the degree of influence of the i-th region on the j-th region; Based on the influence level, determine the neighboring regions of the j-th region; By using the spatial branch of the dual-branch spatiotemporal prediction model, attention operations are performed on the feature vector extracted from the fused data vector of each region and the feature vectors of its neighboring regions to obtain the first fused feature vector of each region.

4. The PM2.5 short-term AI forecasting method according to claim 3, characterized in that, Based on the included angle, the wind speed data, and the PM2.5 concentration data, the influence of the i-th region on the j-th region is determined, including: Based on the included angle and the wind speed data, determine the wind field influence information of the i-th region on the j-th region; The influence of region i on region j is determined based on the distance between region i and region j, the wind field influence information, and the PM2.5 concentration data.

5. The PM2.5 short-term AI forecasting method according to claim 4, characterized in that, Based on the included angle and the wind speed data, the wind field influence information of the i-th region on the j-th region is determined, including: According to the formula Determine the wind field impact information of region i on region j at the current time. ,in, For preset coefficients, Let i be the wind speed data for the i-th region at the current moment. Let be the angle between the geographical directions of the i-th and j-th regions and the wind direction of the i-th region, max be the function to maximize the value, and t be the current time.

6. The PM2.5 short-term AI forecasting method according to claim 4, characterized in that, Based on the distance between region i and region j, the wind field influence information, and the PM2.5 concentration data, the degree of influence of region i on region j is determined, including: According to the formula Determine the influence of region i on region j at the current time. ,in, Let be the distance between the i-th region and the j-th region. This provides information on the wind field impact of region i on region j at the current moment. Let represent the PM2.5 concentration data for the i-th region at the current time, and max is the function to take the maximum value.

7. The PM2.5 short-term AI forecasting method according to claim 1, characterized in that, Acquire multidimensional meteorological data for each region, and determine the weights of each version of the bi-branch spatiotemporal prediction model based on the multidimensional meteorological data, including: Set a contribution threshold; Multidimensional meteorological data and corrected forecast data vectors are input into the decision tree model to obtain the contribution scores of each version of the bi-branch spatiotemporal forecast model. A version of the bi-branch spatiotemporal prediction model that masks contribution scores below the contribution threshold; Output the weights of each version of the bi-branch spatiotemporal prediction model.

8. The PM2.5 short-term AI forecasting method according to claim 1, characterized in that, The method further includes: The confidence interval of each predicted data in the predicted data output vector at each time step is determined by at least one of the error statistics method, the integrated extreme value method, and the quantile prediction method.

9. The PM2.5 short-term AI forecasting method according to claim 1, characterized in that, The method further includes: Obtain the error of each predicted data point in the predicted data output vector for each partition; The bi-branch spatiotemporal prediction model for each version is fine-tuned and iterated based on the error.

10. A short-term AI forecasting system for PM2.5, characterized in that, include: The first vector acquisition module obtains the multi-source data vector for each region at each time based on the historical monitoring data of monitoring stations in multiple regions at multiple times. The second vector acquisition module processes the multi-source data vectors of each region at each time step using multiple versions of the bi-branch spatiotemporal prediction model to obtain the predicted data vectors of each region at multiple future times. The third vector acquisition module uses an external tree to integrate an hourly rolling correction model to correct the predicted data vector and obtain the corrected predicted data vector. The weight calculation module acquires multidimensional meteorological data for each region and determines the weights of each version of the bi-branch spatiotemporal prediction model based on the multidimensional meteorological data. The data output module obtains the prediction data output vector for each partition based on the weights of the bi-branch spatiotemporal prediction model for each version and the corrected prediction data vector.

Images