Atmospheric pollution prediction method based on multi-model fusion
Through the multi-model fusion method, the concentration of atmospheric pollutants is predicted using LSTM, ConvLSTM, Transformer and GNN-Transformer models, which solves the problem of low prediction accuracy of pollution diffusion in the prior art, and achieves high-precision prediction of multi-time scale and multi-space dimensions.
Patent Information
- Application Number
- CN202510555285.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-29
- Publication Date
- 2025-08-15
AI Technical Summary
The existing air pollution prediction methods are difficult to effectively capture the nonlinear timing characteristics of pollutant concentrations and cannot fully integrate multi-site spatiotemporal distribution data, resulting in low prediction accuracy of multi-site coordinated pollution diffusion, and lack the ability to model pollutant transmission laws, making it difficult to meet the needs of refined environmental management.
The LSTM model, ConvLSTM model, Transformer model and GNN-Transformer model are used to predict pollutant concentration data. Combined with site location and meteorological data, and through weighted fusion and Kalman filtering, high-precision prediction of multi-time scale and multi-space dimensions is achieved.
Effectively capture the nonlinear timing characteristics of pollutant concentration and multi-site spatiotemporal distribution data, improve the accuracy of pollution diffusion prediction, support dynamic collaborative prediction, and meet the needs of refined environmental management.
Smart Images
Figure CN120496681A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of atmospheric environment monitoring and artificial intelligence technology, and specifically relates to an atmospheric pollution prediction method based on multi-model fusion. Background Art
[0002] Traditional air pollution prediction methods primarily rely on numerical and statistical models, such as diffusion equations or regression analysis driven by meteorological fields. While these methods can complement physical mechanisms with data-driven approaches and balance computational efficiency and accuracy, they suffer from the following shortcomings:
[0003] (1) Numerical models and statistical models are difficult to effectively capture the nonlinear time series characteristics of pollutant concentrations.
[0004] (2) Numerical models and statistical models cannot fully integrate the spatiotemporal distribution data of multiple sites, resulting in insufficient accuracy in regional pollution diffusion prediction.
[0005] (3) When making long-term trend forecasts, numerical models and statistical models are limited by the complexity of the models and it is difficult to take into account both global dynamics and local details.
[0006] (4) Numerical models and statistical models lack the ability to model the laws of pollutant transmission, and the effect of multi-site collaborative prediction is limited.
[0007] In summary, existing methods have limited accuracy in multi-site collaborative pollution dispersion forecasting due to their inability to effectively capture the nonlinear temporal characteristics of pollutant concentrations, their inability to fully integrate multi-site spatiotemporal distribution data, and their lack of ability to model pollutant transport patterns. Furthermore, existing methods are limited in their ability to post-process and visualize prediction results, making them inadequate for refined environmental management. Therefore, an intelligent approach is urgently needed that integrates the strengths of multiple models, accounts for spatiotemporal characteristics, and supports dynamic collaborative forecasting. Summary of the Invention
[0008] The purpose of the present invention is to solve the problem of low accuracy of pollution diffusion prediction in multi-site collaboration in existing methods, and to propose an atmospheric pollution prediction method based on multi-model fusion.
[0009] The technical solution adopted by the present invention to solve the above technical problems is: an air pollution prediction method based on multi-model fusion, the method specifically comprising the following steps:
[0010] Step 1: Collect pollutant concentration and meteorological data at each station within the target area. The pollutant concentration data collected at each station include CO concentration, SO2 concentration, NO2 concentration, CO concentration, O3 concentration, PM2.5 concentration, and PM10 concentration. The meteorological data collected at each station include wind speed data and humidity data.
[0011] Then the pollutant concentration data collected at each site is processed to obtain processed pollutant concentration data;
[0012] Step 2: Use the processed pollutant concentration data as the input of the LSTM model, and output the pollutant concentration prediction results for each station in the next T days through the LSTM model;
[0013] Step 3: The location data of each station and the processed pollutant concentration data are used as input to the ConvLSTM model, and the ConvLSTM model outputs the pollutant concentration prediction results for each station in the next T days;
[0014] Step 4: The location data of each station and the processed pollutant concentration data are used as input to the Transformer model, and the Transformer model outputs the predicted pollutant concentration results for each station in the next T days;
[0015] Step 5: The location data of each station, the meteorological data collected at each station, and the processed pollutant concentration data are used as inputs to the GNN-Transformer model, and the GNN-Transformer model outputs the pollutant concentration prediction results for each station in the next T days;
[0016] Step 6: Perform weighted fusion on the prediction results of steps 2 to 5, then perform denormalization and Kalman filtering on the weighted fusion results, and use the processed results as the pollutant concentration prediction results for each station in the next T days.
[0017] The beneficial effects of the present invention are:
[0018] The present invention adopts LSTM model, ConvLSTM model, Transformer model and GNN-Transformer model respectively to predict the pollutant concentration results of each station in the future time. The four models comprehensively consider the pollutant concentration data, location data and meteorological data of each station when making predictions. They can effectively capture the nonlinear time series characteristics of pollutant concentration and fully integrate the spatiotemporal distribution data of multiple stations. Finally, the prediction results of the four models are weighted and fused to achieve high-precision prediction of pollutant concentrations in multiple time scales and multiple spatial dimensions. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] Figure 1 This is a flow chart of an air pollution prediction method based on multi-model fusion according to the present invention;
[0020] Figure 2The LSTM model-based PM2.5 single-site 24-hour forecast map provided by the present invention;
[0021] Figure 3 The PM2.5 multi-site forecast map for the next 48 hours based on the ConvLSTM model provided by the present invention;
[0022] Figure 4 This is a graph showing the pollution value prediction results for the next seven days based on the Transformer model provided by the present invention;
[0023] Figure 5 This is a graph of multi-site PM2.5 prediction results based on the GNN-Transformer model provided by the present invention;
[0024] Figure 6 A comparison chart of predicted data and real data based on the GNN-Transformer model provided by the present invention;
[0025] Figure 7 This is a multi-site pollution heat map based on multi-model fusion provided by the present invention. DETAILED DESCRIPTION
[0026] Specific implementation method 1: Combination Figure 1 This embodiment describes an air pollution prediction method based on multi-model fusion, which specifically includes the following steps:
[0027] Step 1: Collect pollutant concentration and meteorological data at each station within the target area. The pollutant concentration data collected at each station include CO concentration, SO2 concentration, NO2 concentration, CO concentration, O3 concentration, PM2.5 concentration, and PM10 concentration. The meteorological data collected at each station include wind speed data and humidity data.
[0028] Then the pollutant concentration data collected at each site is processed to obtain processed pollutant concentration data;
[0029] Step 2: Use the processed pollutant concentration data as the input of the LSTM model, and output the pollutant concentration prediction results for each station in the next T days through the LSTM model;
[0030] Step 3: The location data of each station and the processed pollutant concentration data are used as input to the ConvLSTM model, and the ConvLSTM model outputs the predicted pollutant concentration results for each station in the next T days;
[0031] Step 4: The location data of each station and the processed pollutant concentration data are used as input to the Transformer model, and the Transformer model outputs the predicted pollutant concentration results for each station in the next T days;
[0032] The prediction results of the Transformer model are as follows Figure 4 As shown;
[0033] Step 5: The location data of each station, the meteorological data collected at each station, and the processed pollutant concentration data are used as inputs to the GNN-Transformer model, and the GNN-Transformer model outputs the pollutant concentration prediction results for each station in the next T days;
[0034] Step 6: Perform weighted fusion on the prediction results of steps 2 to 5, then perform denormalization and Kalman filtering on the weighted fusion results, and use the processed results as the pollutant concentration prediction results for each station in the next T days.
[0035] Specific embodiment 2: This embodiment differs from specific embodiment 1 in that the pollutant concentration data collected at each site is processed as follows:
[0036] Interpolation is performed to fill missing values in the pollutant concentration data collected at each station to obtain complete pollutant concentration data after filling;
[0037] The Z-score standardization method is then used to normalize the filled pollutant concentration data to obtain the processed data.
[0038] Other steps and parameters are the same as those in the first embodiment.
[0039] Specific implementation method three: This implementation method is different from specific implementation methods one or two in that the value of T is 7.
[0040] Other steps and parameters are the same as those in the first or second embodiment.
[0041] Specific embodiment 4: This embodiment differs from any one of specific embodiments 1 to 3 in that the LSTM model used in step 2 includes 2 layers of LSTM units.
[0042] The other steps and parameters are the same as those in the first to third embodiments.
[0043] Each LSTM unit layer consists of 50 neurons. The LSTM model inputs 60 hours of historical pollutant concentration data, uses a dropout layer (ratio 0.2) to prevent overfitting, and outputs pollutant prediction results for the next 24 hours. The gating mechanism and state update process of the LSTM unit are as follows:
[0044] The forget gate determines the amount of memory information retained at the previous moment:
[0045]
[0046] The input gate controls the degree to which new information is incorporated:
[0047] i t =σ(W i ·[h t-1 ,x t ]+b i )
[0048] Candidate memories generate new information to be stored:
[0049]
[0050] Combine the forget gate and input gate to update the memory state:
[0051]
[0052] The output gate controls the amount of information that is ultimately output:
[0053] h t =o t ⊙tanh(C t )
[0054] Specific embodiment 5: This embodiment differs from specific embodiments 1 to 4 in that the specific process of step 2 is as follows:
[0055] Step 21: Each station collects historical pollutant concentration data for T0 hours (T0 is 60 in the present invention), and initializes the station count i=1;
[0056] Step 2: All processed pollutant concentration data collected at the i-th station is used as the input of the first-layer LSTM unit in the LSTM model. The output of the first-layer LSTM unit is used as the input of the second-layer LSTM unit. The second-layer LSTM unit outputs the pollutant concentration prediction result of the i-th station in the next 24 hours.
[0057] Step 23, initialize the number of iterations t = 1;
[0058] Step 24: Use the historical pollutant concentration data collected at the i-th station for hours T0-24t (it should be noted that when the value of T0-24t is less than or equal to 0, it means that the predicted time has exceeded T0, and subsequent model inputs will all be predicted data) and the predicted pollutant concentration data for 24t hours as the input of the first-layer LSTM unit in the LSTM model. Use the output of the first-layer LSTM unit as the input of the second-layer LSTM unit, and output the pollutant concentration prediction result for the i-th station from the 24t hour to the 24(t+1) hour in the future through the second-layer LSTM unit;
[0059] Step 25: Determine whether the pollutant concentration at the i-th station within the next T days has been predicted;
[0060] If the pollutant concentration at the i-th station in the next T days has been predicted, then execute step 26;
[0061] If the pollutant concentration at the i-th station in the next T days is not predicted, set t = t + 1 and return to step 24;
[0062] Step 26: Determine whether the pollutant concentrations of all stations in the next T days have been predicted;
[0063] If the pollutant concentrations of all stations in the next T days have been predicted, the process ends;
[0064] If the pollutant concentrations of all stations in the next T days have not been predicted, set i=i+1 and return to step 22.
[0065] The other steps and parameters are the same as those in the first to fourth embodiments.
[0066] Taking any station as an example, the present invention obtains the pollutant concentration data of this station in the past 60 hours. First, the pollutant concentration data of the past 60 hours is used as the input of the LSTM model, and the LSTM model is used to predict the pollutant concentration data for the next 24 hours. Taking PM2.5 as an example, the concentration prediction results of PM2.5 at a single station in the next 24 hours are as follows: Figure 2As shown. The LSTM model then uses the latest predicted 24-hour pollutant concentration data and the historical 36-hour pollutant concentration data as input, and the model predicts another 24 hours of data (based on the current time, these 24 hours of data represent the data from the 24th to 48th hours in the future). The LSTM model then uses the predicted 48-hour pollutant concentration data and the historically collected 12-hour pollutant concentration data as input, predicting the data from the 48th to 72nd hours in the future. This process continues, until the pollutant concentration is predicted for the next T days. It should be noted that regardless of whether the LSTM model inputs historically collected data or predicted data, each input should contain data from a consecutive 60-hour period.
[0067] Specific embodiment 6: This embodiment differs from any one of specific embodiments 1 to 5 in that the ConvLSTM model includes a first convolutional layer, a second convolutional layer, and an LSTM unit in sequence;
[0068] The location data of each site is used as the input of the first convolutional layer, and the output of the first convolutional layer is used as the input of the second convolutional layer, and the spatial features of each site are output through the second convolutional layer;
[0069] The processed pollutant concentration data and the spatial characteristics of each site are then used as the input of the LSTM unit, and the LSTM unit outputs the pollutant concentration prediction results of each site in the next T days.
[0070] The other steps and parameters are the same as those in the first to fifth embodiments.
[0071] Under the ConvLSTM model, the PM2.5 prediction results for multiple sites in the next 48 hours are as follows: Figure 3 As shown in the figure, similar to the LSTM model, subsequent predictions can be made based on the previous prediction results of the ConvLSTM model. In subsequent predictions, the previous prediction results are used as the input of the LSTM unit until the prediction results within the target duration are obtained.
[0072] Specific embodiment 7: This embodiment differs from any one of specific embodiments 1 to 6 in that the specific process of step 5 is as follows:
[0073] Step 5. Calculate the correlation between the i-th station and the j-th station at the t-th hour in history based on the pollutant concentration gradient and meteorological data differences.
[0074] Step 52: Calculate the edge weight between the i-th station and the j-th station at the t-th hour in history based on the wind speed data and humidity data of the i-th station at the t-th hour in history, the wind speed data and humidity data of the j-th station at the t-th hour in history, and the distance between the i-th station and the j-th station. (generated based on the meteorological field advection-diffusion equation);
[0075] Step 53: According to and Calculate the initial weight between the i-th site and the j-th site at the t-th hour in history Then update the weight between the i-th site and the j-th site
[0076]
[0077] in, represents the initial weight between the i-th station and the p-th station at the t-th hour in history, α is the temperature coefficient (set according to actual conditions), and P represents the set of all stations;
[0078] Then calculate the final weight w between the i-th site and the j-th site ij :
[0079]
[0080] Step 54: Treat each site as a node. For the i-th site, calculate the weight between the i-th site and each other site, select the K sites with the largest weights with the i-th site. The selected K sites are the neighboring sites of the i-th site, and construct the adjacency matrix between the sites.
[0081] The site location is used as the initial feature of the site, the initial feature of the site and the constructed adjacency matrix are used as the input of GNN, and the site spatiotemporal encoding result is output through GNN;
[0082] Step 5: Use the spatiotemporal coding results of the stations output by GNN and the processed pollutant concentration data of each station as the input of the Transformer model, and output the pollutant concentration prediction results of each station in the next T days through the Transformer model.
[0083] The other steps and parameters are the same as those in the first to sixth embodiments.
[0084] The GNN-Transformer model of this invention refers to a model that integrates Transformer and graph neural network GNN. The Transformer's multi-head attention splits Q, K, and V into h heads (h is 8), and its formula is as follows:
[0085]
[0086] MultiHead(Q,K,V)=Concat(head1,…,head h )W O
[0087] Among them, MultiHead(Q,K,V) is the result of splicing the output of all heads and performing linear transformation on the splicing result.
[0088] The feedforward neural network (FNN) included in the Transformer applies two layers of fully connected networks independently at each position.
[0089] The multi-site PM2.5 prediction results based on the GNN-Transformer model are as follows Figure 5 As shown in the figure, the comparison between the predicted data and the real data based on the GNN-Transformer model is as follows Figure 6 shown.
[0090] Specific embodiment eight: This embodiment differs from specific embodiments one to seven in that the specific process of step five is as follows:
[0091]
[0092] Among them, TCN(·) represents the temporal convolutional network, Sigmoid(·) represents the nonlinear activation function, || represents the splicing operation, represents the mean value of the pollutant concentration gradient between the i-th station and the j-th station from time tk to time t, represents the mean value of the difference in meteorological data between the i-th station and the j-th station from time tk to time t (for any moment between time tk and time t, calculate the wind speed difference between the i-th station and the j-th station at that moment, calculate the humidity difference between the i-th station and the j-th station at that moment, then take the average of the wind speed difference from time tk to time t, and take the average of the humidity difference from time tk to time t, is a vector consisting of the mean of wind speed difference and the mean of humidity difference).
[0093] The other steps and parameters are the same as those in the first to seventh embodiments.
[0094] Specific embodiment 9: This embodiment differs from any one of specific embodiments 1 to 8 in that the adjacency matrix between sites is constructed as follows:
[0095] For the i-th site, if the j-th site is a neighbor node of the i-th site, the element in the i-th row and j-th column of the adjacency matrix is 1; otherwise, the element in the i-th row and j-th column of the adjacency matrix is 0.
[0096] The other steps and parameters are the same as those in Specific Embodiments 1 to 8.
[0097] By traversing each site, the adjacency matrix can be obtained.
[0098] Specific embodiment 10: This embodiment differs from specific embodiments 1 to 9 in that the specific process of step 6 is as follows:
[0099]
[0100] Among them, f l (X) represents the prediction result of the lth model, w l represents the weight corresponding to the lth model, Indicates the final prediction result.
[0101] The other steps and parameters are the same as those in Specific Embodiments 1 to 9.
[0102] In the present invention, l=1, 2, 3, 4 correspond to LSTM model, ConvLSTM model, Transformer model, GNN-Transformer model respectively, and the weights satisfy And w l ≥0, the specific weight distribution is obtained according to the performance of each model on the validation set. After the present invention obtains the prediction results, it can also display the prediction results and realize the dynamic simulation of pollutant transmission. The multi-site pollution heat map of the fused prediction results is as follows Figure 7 shown.
[0103] The above examples are merely illustrative of the calculation model and process of the present invention and are not intended to limit the embodiments of the present invention. Persons skilled in the art will readily appreciate that other variations or modifications based on the above description are possible. This list of embodiments is not exhaustive; however, any obvious variations or modifications derived from the technical solution of the present invention remain within the scope of protection of the present invention.
Claims
1. An atmospheric pollution prediction method based on multi-model fusion, characterized in that: The method specifically comprises the following steps: Step 1: Collect pollutant concentration and meteorological data at each station within the target area. The pollutant concentration data collected at each station include CO concentration, SO2 concentration, NO2 concentration, CO concentration, O3 concentration, PM2.5 concentration, and PM10 concentration. The meteorological data collected at each station include wind speed data and humidity data. Then the pollutant concentration data collected at each site is processed to obtain processed pollutant concentration data; Step 2: Use the processed pollutant concentration data as the input of the LSTM model, and output the pollutant concentration prediction results for each station in the next T days through the LSTM model; Step 3: The location data of each station and the processed pollutant concentration data are used as input to the ConvLSTM model, and the ConvLSTM model outputs the predicted pollutant concentration results for each station in the next T days; Step 4: The location data of each station and the processed pollutant concentration data are used as input to the Transformer model, and the Transformer model outputs the predicted pollutant concentration results for each station in the next T days; Step 5: The location data of each station, the meteorological data collected at each station, and the processed pollutant concentration data are used as inputs to the GNN-Transformer model, and the GNN-Transformer model outputs the pollutant concentration prediction results for each station in the next T days; Step 6: Perform weighted fusion on the prediction results of steps 2 to 5, then perform denormalization and Kalman filtering on the weighted fusion results, and use the processed results as the pollutant concentration prediction results for each station in the next T days.
2. The method for air pollution prediction based on multi-model fusion according to claim 1, characterized in that: The pollutant concentration data collected at each site is processed as follows: Interpolation is performed to fill missing values in the pollutant concentration data collected at each station to obtain complete pollutant concentration data after filling; The Z-score standardization method is then used to normalize the filled pollutant concentration data to obtain the processed data.
3. The method for air pollution prediction based on multi-model fusion according to claim 1, characterized in that: The value of T is 7.
4. The method for air pollution prediction based on multi-model fusion according to claim 1, characterized in that: The LSTM model used in step 2 includes two layers of LSTM units.
5. The method for air pollution prediction based on multi-model fusion according to claim 4, characterized in that: The specific process of step 2 is as follows: Step 2:
1. Each station collects historical pollutant concentration data for T0 hours and initializes the station count i=1; Step 2: All processed pollutant concentration data collected at the i-th station is used as the input of the first-layer LSTM unit in the LSTM model. The output of the first-layer LSTM unit is used as the input of the second-layer LSTM unit. The second-layer LSTM unit outputs the pollutant concentration prediction result of the i-th station in the next 24 hours. Step 23, initialize the number of iterations t = 1; Step 24: Use the historical pollutant concentration data collected at the i-th station from hours T0 to 24t and the predicted pollutant concentration data for hours 24t as the input to the first-layer LSTM unit in the LSTM model. Use the output of the first-layer LSTM unit as the input to the second-layer LSTM unit. The second-layer LSTM unit outputs the predicted pollutant concentration result for the i-th station from hours 24t to hours 24(t+1) in the future. Step 25: Determine whether the pollutant concentration at the i-th station within the next T days has been predicted; If the pollutant concentration of the i-th station in the next T days has been predicted, then execute step 26; If the pollutant concentration at the i-th station in the next T days is not predicted, set t = t + 1 and return to step 24; Step 26: Determine whether the pollutant concentrations of all stations in the next T days have been predicted; If the pollutant concentrations of all stations in the next T days have been predicted, the process ends; If the pollutant concentrations of all stations in the next T days have not been predicted, set i=i+1 and return to step 22.
6. The method for air pollution prediction based on multi-model fusion according to claim 1, characterized in that: The ConvLSTM model includes a first convolutional layer, a second convolutional layer and an LSTM unit in sequence; The location data of each site is used as the input of the first convolutional layer, and the output of the first convolutional layer is used as the input of the second convolutional layer, and the spatial features of each site are output through the second convolutional layer; The processed pollutant concentration data and the spatial characteristics of each site are then used as the input of the LSTM unit, and the LSTM unit outputs the pollutant concentration prediction results of each site in the next T days.
7. The method for air pollution prediction based on multi-model fusion according to claim 1, characterized in that: The specific process of step five is: Step 5. Calculate the correlation between the i-th station and the j-th station at the t-th hour in history based on the pollutant concentration gradient and meteorological data differences. Step 52: Calculate the edge weight between the i-th station and the j-th station at the t-th hour in history based on the wind speed data and humidity data of the i-th station at the t-th hour in history, the wind speed data and humidity data of the j-th station at the t-th hour in history, and the distance between the i-th station and the j-th station. Step 53: According to and Calculate the initial weight between the i-th site and the j-th site at the t-th hour in history Then update the weight between the i-th site and the j-th site in, represents the initial weight between the i-th station and the p-th station at the t-th hour in history, α is the temperature coefficient, and P represents the set of all stations; Then calculate the final weight w between the i-th site and the j-th site ij : Step 54: Treat each site as a node. For the i-th site, calculate the weight between the i-th site and each other site, select the K sites with the largest weights with the i-th site. The selected K sites are the neighboring sites of the i-th site, and construct the adjacency matrix between the sites. The site location is used as the initial feature of the site, the initial feature of the site and the constructed adjacency matrix are used as the input of GNN, and the site spatiotemporal encoding result is output through GNN; Step 5: Use the spatiotemporal coding results of the stations output by GNN and the processed pollutant concentration data of each station as the input of the Transformer model, and output the pollutant concentration prediction results of each station in the next T days through the Transformer model.
8. The method for air pollution prediction based on multi-model fusion according to claim 7, characterized in that: The specific process of step 5 is as follows: Among them, TCN(·) represents the temporal convolutional network, Sigmoid(·) represents the nonlinear activation function, || represents the splicing operation, represents the mean value of the pollutant concentration gradient between the i-th station and the j-th station from time tk to time t, It represents the mean value of the difference in meteorological data between the i-th station and the j-th station from time tk to time t.
9. The method for air pollution prediction based on multi-model fusion according to claim 7, characterized in that: The adjacency matrix between the constructed sites is specifically: For the i-th site, if the j-th site is a neighbor node of the i-th site, the element in the i-th row and j-th column of the adjacency matrix is 1; otherwise, the element in the i-th row and j-th column of the adjacency matrix is 0.
10. The method for air pollution prediction based on multi-model fusion according to claim 1, characterized in that: The specific process of step six is as follows: Among them, f l (X) represents the prediction result of the lth model, w l represents the weight corresponding to the lth model, Represents the final prediction result.