Mid-term pm2.5 prediction method fusing meteorological guidance, transmission prior and reliability-aware posterior evidence

Abstract

This invention relates to the field of meteorological forecasting technology, specifically to a medium-term PM2.5 forecasting method that integrates prior meteorological guidance transmission with post-verification reliability perception. It addresses the problems of existing traditional statistical models failing to effectively capture multivariate coupling and non-stationary time-series characteristics, while physical models suffer from heavy computational burdens and large deviations in characterizing nonlinear processes, and the lack of reliability quantification and post-verification mechanisms in forecast results. The method includes the following steps: S1: Multi-source data acquisition and feature generation. Meteorological information is acquired using a meteorological lidar array, an IoT sensor gateway, and a satellite data receiver to form multi-source data. Then, a dynamic feature vector is generated through a dynamic feature generation unit. This invention eliminates input noise, reduces transmission path errors, and achieves high-precision forecasting. Through a reliability perception and residual feedback closed loop, it enhances reliability in extreme events, providing highly reliable decision support for cross-regional joint prevention and control.

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting technology, specifically to a medium-term PM2.5 forecasting method that integrates prior meteorological guidance transmission data with post-reliability perception data. Background Technology

[0002] With PM2.5 pollution becoming a major factor affecting ambient air quality, its forecasting is crucial for pollution control and public health. Meteorological conditions (such as wind speed and pressure gradient) dominate the transport and diffusion processes of pollutants, and medium-term forecasts of 7-14 days can provide early warnings of regional pollution events. However, traditional methods relying on statistical models or purely physical simulations are difficult to handle highly nonlinear atmospheric systems, resulting in insufficient accuracy under sudden weather events or complex terrain. Therefore, there is an urgent need for new solutions that integrate meteorological prior mechanisms with data-driven technologies.

[0003] Existing traditional statistical models cannot effectively capture the characteristics of multivariate coupling and non-stationary time series. At the same time, physical models have a heavy computational burden and large bias in characterizing nonlinear processes, and the prediction results lack reliability quantification and posterior calibration mechanisms. Therefore, they do not meet the current needs. To address this, we propose a medium-term PM2.5 prediction method that integrates meteorological guided transmission prior and reliability perception posterior verification evidence. Summary of the Invention

[0004] The purpose of this invention is to provide a medium-term PM2.5 prediction method that integrates prior data from meteorological guidance transmission with post-verification data from reliability perception, in order to solve the problems mentioned in the background art, such as the inability of existing traditional statistical models to effectively capture multivariate coupling and non-stationary time series characteristics, the heavy computational burden of physical models and the large deviation in characterizing nonlinear processes, and the lack of reliability quantification and post-verification mechanisms in the prediction results.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a medium-term PM2.5 prediction method integrating prior data from meteorological guidance transmission and post-verification data from reliability perception, comprising the following steps: S1: Multi-source data acquisition and feature generation. Meteorological information is acquired and multi-source data is formed through the multi-source data aggregation unit, and then dynamic feature vectors are generated through the dynamic feature generation unit. S2: Meteorological-guided prior modeling is used to solve the advection-diffusion equation and simulate the PM2.5 transport process using a meteorological-diffusion simulation unit and update the predicted concentration distribution in real time. Then, the key pathways for cross-regional pollution transport are determined by embedding prior knowledge into the unit. S3: Multimodal fusion prediction, which determines the basic prediction sequence through the spatiotemporal prediction computing unit and outputs physically interpretable fusion prediction using the multimodal fusion unit; S4: Reliability-aware calibration, which uses an uncertainty quantization unit to generate a predicted distribution and display the CSI in real time, and uses an adaptive calibration unit to trigger an adaptive correction mechanism to transmit the residual signal; S5: Closed-loop feedback optimization calibrates the residual signal and feeds it back to the source term inversion ASIC, re-injecting prior features into the prediction model to form an autonomous optimization closed loop of acquisition, simulation, prediction, calibration and feedback, completing a full iteration every 6 hours.

[0006] Preferably, the meteorological lidar array, the Internet of Things sensor gateway, and the satellite data receiver constitute a multi-source data aggregation unit, and the dynamic feature generation unit consists of a VMD signal decomposer and a meteorological feature calculation engine. The VMD signal decomposer is used to decompose the PM2.5 sequence, and the meteorological feature calculation engine is used to calculate TPI and generate dynamic feature vectors through a Transformer encoder. The multi-source data aggregation unit and the dynamic feature generation unit constitute a data fusion and feature extraction module. The meteorological lidar array uses a 532nm / 1064nm dual-wavelength laser emitter to monitor the vertical distribution of aerosols in real time. The IoT sensor gateway receives ground PM2.5 sensor data through an edge computing node integrating the LoRaWAN protocol. The satellite data receiver contains an X-band parabolic antenna and an SDR module. The satellite data receiver is used to decode the raw data stream from the Fengyun-4 meteorological satellite.

[0007] Preferably, the TPI is the atmospheric transport potential index, which is used to quantify the intensity of regional pollutant transport. The larger the value of the atmospheric transport potential index, the stronger the cross-regional transport potential. ; Where u is the wind speed vector. For pressure gradient, This refers to air density.

[0008] Preferably, the meteorological-diffusion simulation unit consists of a high-performance computing cluster and a source term inversion acceleration card, and the prior knowledge embedding unit consists of a graph neural network accelerator and a transmission path identification unit. The graph neural network accelerator is used to construct a spatial correlation graph from the simulation results of the meteorological-diffusion simulation unit, and the transmission path identification unit is used to perform multi-head attention calculation and identification. The meteorological-diffusion simulation unit and the prior knowledge embedding unit together form a meteorological prior modeling module.

[0009] Preferably, the advection-diffusion equation is: ; Where C represents the pollutant concentration, in μg / m³. 3t is time, v is the flow velocity vector. Let D be the concentration gradient and D be the diffusion coefficient. Let be the Laplacian operator, and denote the mathematical form of the diffusion effect.

[0010] Preferably, the spatiotemporal prediction computation unit consists of an encoder and a decoder. The encoder is used to drive a bidirectional ConvLSTM and extract local spatiotemporal features. The decoder is used to deploy a Transformer decoding layer and generate a basic prediction sequence for 7-14 days. The multimodal fusion unit consists of a dynamic gating controller and a physical constraint coprocessor. The spatiotemporal prediction computation unit and the multimodal fusion unit together form the core prediction model module.

[0011] Preferably, the dynamic gating controller is used for weight calculation; ; in, To integrate weights, It is the Sigmoid activation function. For learnable weight matrix, For dynamic feature vectors, These are the prior feature vectors.

[0012] Preferably, the physical constraint coprocessor is used to verify the continuity equation logic; ; Where u, v, and w are the components of wind speed in the meridional, zonal, and vertical directions, respectively. Let be the velocity gradient in the x-direction, and represent the spatial rate of change of the meridional wind. Let be the velocity gradient in the y-direction, representing the spatial rate of change of the zonal wind. Let z be the velocity gradient in the z-direction, and let g be the spatial rate of change of the vertical wind.

[0013] Preferably, the uncertainty quantization unit consists of a probability calculation array and a reliability monitoring terminal. The probability calculation array is used for inference and generating a prediction distribution, and the reliability monitoring terminal is used for real-time display of the CSI on an industrial-grade HMI touchscreen. The adaptive calibration unit consists of a reinforcement learning actuator and a feedback control interface. The reinforcement learning actuator is used to run the PPO reinforcement learning algorithm and dynamically calculate the calibration amount ΔY based on the CSI. The trigger condition for the adaptive correction mechanism is CSI < 0.2. The feedback control interface is used to transmit the residual signal. The uncertainty quantization unit and the adaptive calibration unit together form a posterior calibration optimization module.

[0014] Preferably, the CSI is a confidence score, where a CSI value of 1 indicates high confidence and a CSI value less than 0.2 indicates low confidence. ; in, The interquartile range is used to calculate the range of the 25th to 75th quantiles in the predicted distribution and to measure the bandwidth of uncertainty. This is a sequence of predicted PM2.5 concentrations. This is the median of the predicted values.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention uses a dynamic threshold and a spatiotemporal alignment algorithm in the multi-source cleaning unit of the data fusion module to eliminate input noise and ensure feature reliability. The physical simulation unit of the meteorological prior module inverts the emission source intensity in real time, effectively reducing the transmission path identification error. The spatiotemporal sequence unit of the core prediction module can capture medium- and long-term fluctuations. The multimodal fusion unit dynamically balances prior and real-time data through gating weights, so that the average absolute error of the seven-day prediction is stably reduced to meet the timeliness requirements of pollution prevention and control. 2. This invention constructs a closed-loop optimization mechanism through a posterior calibration module, uses the change in confidence score value to trigger the reinforcement learning strategy of the adaptive calibration unit, dynamically corrects deviation outliers, and feeds the residual signal back to the source term inversion unit of the meteorological prior module through the bus to update emission parameters in real time. At the same time, it drives the physical constraint coprocessor of the core prediction module to strengthen the verification of the continuity equation. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall structure of the present invention; Figure 2 This is a flowchart of the entire invention. Detailed Implementation

[0017] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0018] Please see Figure 2 The present invention provides an embodiment of a medium-term PM2.5 prediction method that integrates prior data from meteorological guided transmission and post-verification data from reliability perception, comprising the following steps: S1: Multi-source data acquisition and feature generation. Meteorological information is acquired and multi-source data is formed through the multi-source data aggregation unit, and then dynamic feature vectors are generated through the dynamic feature generation unit. S2: Meteorological-guided prior modeling is used to solve the advection-diffusion equation and simulate the PM2.5 transport process using a meteorological-diffusion simulation unit and update the predicted concentration distribution in real time. Then, the key pathways for cross-regional pollution transport are determined by embedding prior knowledge into the unit. S3: Multimodal fusion prediction, which determines the basic prediction sequence through the spatiotemporal prediction computing unit and outputs physically interpretable fusion prediction using the multimodal fusion unit; S4: Reliability-aware calibration, which uses an uncertainty quantization unit to generate a predicted distribution and display the CSI in real time, and uses an adaptive calibration unit to trigger an adaptive correction mechanism to transmit the residual signal; S5: Closed-loop feedback optimization calibrates the residual signal and feeds it back to the source term inversion ASIC, re-injecting prior features into the prediction model to form an autonomous optimization closed loop of acquisition, simulation, prediction, calibration and feedback, completing a full iteration every 6 hours.

[0019] Please see Figure 1 The system consists of a multi-source data aggregation unit and a dynamic feature generation unit, forming a data fusion and feature extraction module. The multi-source data aggregation unit comprises a meteorological lidar array, an IoT sensor gateway, and a satellite data receiver. The meteorological lidar array uses a 532nm / 1064nm dual-wavelength laser emitter for real-time monitoring of the vertical distribution of aerosols. The IoT sensor gateway uses an edge computing node with integrated LoRaWAN protocol to receive ground PM2.5 sensor data. The satellite data receiver contains an X-band parabolic antenna and an SDR module. The satellite data receiver is used to decode the raw data stream from the Fengyun-4 meteorological satellite. The dynamic feature generation unit consists of a VMD signal decomposer and a meteorological feature calculation engine. The VMD signal decomposer is used to decompose the PM2.5 sequence, and the meteorological feature calculation engine is used to calculate the TPI and generate dynamic feature vectors through the Transformer encoder. The TPI is the atmospheric transport potential index, which is used to quantify the intensity of regional pollutant transport. The larger the value of the atmospheric transport potential index, the stronger the cross-regional transport potential. ; Where u is the wind speed vector. For pressure gradient, This refers to air density.

[0020] Furthermore, the meteorological-diffusion simulation unit consists of a high-performance computing cluster and a source term inversion acceleration card, while the prior knowledge embedding unit consists of a graph neural network accelerator and a transmission path identification unit. The graph neural network accelerator is used to construct a spatial correlation graph from the simulation results of the meteorological-diffusion simulation unit, and the transmission path identification unit is used to perform multi-head attention calculation and identification. The meteorological-diffusion simulation unit and the prior knowledge embedding unit together form a meteorological prior modeling module. The advection-diffusion equation is: ; Where C represents the pollutant concentration, in μg / m³. 3 t is time, v is the flow velocity vector. Let D be the concentration gradient and D be the diffusion coefficient. Let be the Laplacian operator, and denote the mathematical form of the diffusion effect.

[0021] Furthermore, the spatiotemporal prediction computation unit consists of an encoder and a decoder. The encoder is used to drive the bidirectional ConvLSTM and extract local spatiotemporal features, while the decoder is used to deploy the Transformer decoding layer and generate a 7-14 day basic prediction sequence. The multimodal fusion unit consists of a dynamic gating controller and a physical constraint coprocessor. The spatiotemporal prediction computation unit and the multimodal fusion unit together form the core prediction model module. Dynamic gating controllers are used for weight calculation; ; in, To integrate weights, It is the Sigmoid activation function. For learnable weight matrix, For dynamic feature vectors, These are the prior feature vectors.

[0022] The physical constraint coprocessor is used to verify the logic of the continuity equation; ; Where u, v, and w are the components of wind speed in the meridional, zonal, and vertical directions, respectively. Let be the velocity gradient in the x-direction, and represent the spatial rate of change of the meridional wind. Let be the velocity gradient in the y-direction, representing the spatial rate of change of the zonal wind. Let z be the velocity gradient in the z-direction, and let g be the spatial rate of change of the vertical wind.

[0023] Furthermore, the uncertainty quantification unit consists of a probability calculation array and a reliability monitoring terminal. The probability calculation array is used for inference and generating the prediction distribution, and the reliability monitoring terminal is used for real-time display of CSI on an industrial-grade HMI touch screen. The adaptive calibration unit consists of a reinforcement learning actuator and a feedback control interface. The reinforcement learning actuator is used to run the PPO reinforcement learning algorithm and dynamically calculate the calibration amount ΔY according to CSI. The trigger condition for the adaptive correction mechanism is CSI < 0.2. The feedback control interface is used to transmit the residual signal. The uncertainty quantification unit and the adaptive calibration unit together form a posterior calibration optimization module. CSI is the confidence score. A CSI value of 1 indicates high confidence, while a CSI value less than 0.2 indicates low confidence. ; in, The interquartile range is used to calculate the range of the 25th to 75th quantiles in the predicted distribution and to measure the bandwidth of uncertainty. This is a sequence of predicted PM2.5 concentrations. This is the median of the predicted values.

[0024] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A medium-term PM2.5 forecasting method integrating prior data from meteorological guided transmission and post-verification data from reliability perception, characterized in that... Includes the following steps: S1: Multi-source data acquisition and feature generation. Meteorological information is acquired and multi-source data is formed through the multi-source data aggregation unit, and then dynamic feature vectors are generated through the dynamic feature generation unit. S2: Meteorological-guided prior modeling is used to solve the advection-diffusion equation and simulate the PM2.5 transport process using a meteorological-diffusion simulation unit and update the predicted concentration distribution in real time. Then, the key pathways for cross-regional pollution transport are determined by embedding prior knowledge into the unit. S3: Multimodal fusion prediction, which determines the basic prediction sequence through the spatiotemporal prediction computing unit and outputs physically interpretable fusion prediction using the multimodal fusion unit; S4: Reliability-aware calibration, which uses an uncertainty quantization unit to generate a predicted distribution and display the CSI in real time, and uses an adaptive calibration unit to trigger an adaptive correction mechanism to transmit the residual signal; S5: Closed-loop feedback optimization calibrates the residual signal and feeds it back to the source term inversion ASIC, re-injecting prior features into the prediction model to form an autonomous optimization closed loop of acquisition, simulation, prediction, calibration and feedback, completing a full iteration every 6 hours.

2. The medium-term PM2.5 prediction method according to claim 1, which integrates prior data from meteorological guidance transmission and post-verification data from reliability perception, is characterized in that: The multi-source data aggregation unit consists of a meteorological lidar array, an IoT sensor gateway, and a satellite data receiver. The dynamic feature generation unit consists of a VMD signal decomposer and a meteorological feature calculation engine. The VMD signal decomposer is used to decompose the PM2.5 sequence, and the meteorological feature calculation engine is used to calculate TPI and generate dynamic feature vectors through a Transformer encoder. The multi-source data aggregation unit and the dynamic feature generation unit together form a data fusion and feature extraction module. The meteorological lidar array uses a 532nm / 1064nm dual-wavelength laser emitter to monitor the vertical distribution of aerosols in real time. The IoT sensor gateway receives ground PM2.5 sensor data through an edge computing node integrating the LoRaWAN protocol. The satellite data receiver contains an X-band parabolic antenna and an SDR module. The satellite data receiver is used to decode the raw data stream from the Fengyun-4 meteorological satellite.

3. The medium-term PM2.5 prediction method according to claim 2, which integrates prior data from meteorological guidance transmission and post-verification data from reliability perception, is characterized in that: The TPI is the atmospheric transport potential index, which is used to quantify the intensity of regional pollutant transport. The larger the value of the atmospheric transport potential index, the stronger the cross-regional transport potential. ； Where u is the wind speed vector. For pressure gradient, This refers to air density.

4. The medium-term PM2.5 prediction method according to claim 1, which integrates prior data from meteorological guidance transmission and post-verification data from reliability perception, is characterized in that: The meteorological-diffusion simulation unit consists of a high-performance computing cluster and a source term inversion accelerator card. The prior knowledge embedding unit consists of a graph neural network accelerator and a transmission path identification unit. The graph neural network accelerator is used to construct a spatial correlation graph from the simulation results of the meteorological-diffusion simulation unit. The transmission path identification unit is used to perform multi-head attention calculation and identification. The meteorological-diffusion simulation unit and the prior knowledge embedding unit together form a meteorological prior modeling module.

5. The medium-term PM2.5 prediction method according to claim 4, which integrates prior data from meteorological guidance transmission and post-verification data from reliability perception, is characterized in that: The advection-diffusion equation is: ； Where C represents the pollutant concentration, in μg / m³. 3 t is time, v is the flow velocity vector. Let D be the concentration gradient and D be the diffusion coefficient. Let be the Laplacian operator, and denote the mathematical form of the diffusion effect.

6. The medium-term PM2.5 prediction method according to claim 1, which integrates prior data from meteorological guidance transmission and post-verification data from reliability perception, is characterized in that: The spatiotemporal prediction computation unit consists of an encoder and a decoder. The encoder is used to drive a bidirectional ConvLSTM and extract local spatiotemporal features. The decoder is used to deploy a Transformer decoding layer and generate a 7-14 day basic prediction sequence. The multimodal fusion unit consists of a dynamic gating controller and a physical constraint coprocessor. The spatiotemporal prediction computation unit and the multimodal fusion unit together form the core prediction model module.

7. The medium-term PM2.5 prediction method according to claim 6, which integrates prior data from meteorological guidance transmission and post-verification data from reliability perception, is characterized in that: The dynamic gating controller is used for weight calculation; ； in, To integrate weights, It is the Sigmoid activation function. For learnable weight matrix, For dynamic feature vectors, These are the prior feature vectors.

8. The medium-term PM2.5 prediction method according to claim 7, which integrates prior data from meteorological guidance transmission and post-verification data from reliability perception, is characterized in that: The physical constraint coprocessor is used to verify the continuity equation logic; ； Where u, v, and w are the components of wind speed in the meridional, zonal, and vertical directions, respectively. Let be the velocity gradient in the x-direction, and represent the spatial rate of change of the meridional wind. Let be the velocity gradient in the y-direction, representing the spatial rate of change of the zonal wind. Let z be the velocity gradient in the z-direction, and let g be the spatial rate of change of the vertical wind.

9. The medium-term PM2.5 prediction method according to claim 1, which integrates prior data from meteorological guidance transmission and post-verification data from reliability perception, is characterized in that: The uncertainty quantization unit consists of a probability calculation array and a reliability monitoring terminal. The probability calculation array is used for inference and generating a prediction distribution, and the reliability monitoring terminal is used for real-time display of the CSI on an industrial-grade HMI touchscreen. The adaptive calibration unit consists of a reinforcement learning actuator and a feedback control interface. The reinforcement learning actuator is used to run the PPO reinforcement learning algorithm and dynamically calculate the calibration amount ΔY based on the CSI. The trigger condition for the adaptive correction mechanism is CSI < 0.

2. The feedback control interface is used to transmit the residual signal. The uncertainty quantization unit and the adaptive calibration unit together form a posterior calibration optimization module.

10. The medium-term PM2.5 prediction method according to claim 9, which integrates prior data from meteorological guidance transmission and post-verification data from reliability perception, is characterized in that: The CSI is a confidence score. A CSI value of 1 indicates high confidence, and a CSI value less than 0.2 indicates low confidence. ； in, The interquartile range is used to calculate the range of the 25th to 75th quantiles in the predicted distribution and to measure the bandwidth of uncertainty. This is a sequence of predicted PM2.5 concentrations. This is the median of the predicted values.

Images