Short-term prediction method for columnar CO concentration
By combining the advantages of the ConvLSTM2D model, LightGBM model and Bayes_CL model, the problem of insufficient accuracy and stability in predicting columnar CO concentrations is solved, and a more efficient and reliable short-term prediction of columnar CO concentrations is achieved.
Patent Information
- Application Number
- CN202411427709.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-14
- Publication Date
- 2025-02-11
AI Technical Summary
The existing ARIMA models have problems such as insufficient prediction accuracy, insufficient processing of nonlinear features and poor short-term prediction stability when predicting columnar CO concentrations.
The combination method of the ConvLSTM2D model, LightGBM model and Bayes_CL model was used to predict the short-term of columnar CO concentration. Specific steps include data segmentation, model training and prediction, residual optimization, and weighted average to generate the final prediction result.
It significantly improves the accuracy and stability of columnar CO concentration prediction, enhances adaptability to complex nonlinear data, and provides more reliable results in short-term predictions.
Smart Images
Figure CN119416084A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of columnar CO concentration prediction, and in particular relates to a columnar CO concentration short-term prediction method. Background Art
[0002] Carbon monoxide (CO) is one of the key indicators of greenhouse gases, and its short-term prediction is of great significance. Currently, columnar CO concentration data is mainly obtained from the Sentinel 5P satellite of the European Space Agency. In the field of atmospheric pollution prediction, traditional time series models such as the ARIMA model are widely used to predict columnar CO concentration.
[0003] However, the ARIMA model has the following disadvantages when predicting columnar CO concentration:
[0004] a. Insufficient prediction accuracy: ARIMA models usually generate R 2 The value is between 0.6 and 0.8, which may not be sufficient in terms of prediction accuracy and stability when dealing with complex environmental data. For example, in the validation of this study, the overall R 2 The value is 0.61. Although it can provide meaningful results in the field of ecological environment, it has limited performance in short-term prediction and spatiotemporal stability.
[0005] b. Insufficient processing of nonlinear characteristics: The changes in columnar CO concentration are affected by the complex interactions between human activities and natural processes, which requires the prediction model to process nonlinear characteristics. The ARIMA model mainly processes linear relationships and has poor adaptability to complex nonlinear data.
[0006] Poor short-term forecast stability: The ARIMA model has poor stability in short-term forecasts (such as a few months) and cannot capture the dynamic changes of environmental data well. Summary of the invention
[0007] In order to solve the problem that the existing ARIMA model for predicting columnar CO concentration has poor short-term prediction stability when predicting columnar CO concentration, the present invention provides a columnar CO concentration short-term prediction method to overcome the limitations of traditional technology and achieve accurate prediction of columnar CO concentration.
[0008] The technical solution of the present invention is as follows:
[0009] A short-term prediction method for columnar CO concentration includes a ConvLSTM2D model, a LightGBM model and a Bayes_CL model. The Bayes_CL model is used to perform short-term prediction of columnar CO concentration. The method is as follows:
[0010] Step 1: Collect columnar CO concentration data, which is divided into training set and test set;
[0011] Step 2: Input the training set into the ConvLSTM2D model and LightGBM model for training. Use the trained ConvLSTM2D model and LightGBM model to predict the columnar CO concentration data and compare them with the test set.
[0012] Step 3: According to the comparison results, the residual data between the model prediction value and the true value is obtained;
[0013] Step 4: The residual data is input into the Bayesian model for optimization to obtain the collaborative residual Bayesian;
[0014] Step 5: Take the weighted average of the output results of the ConvLSTM2D model and the LightGBM model to get the prediction result, and then add the collaborative residual Bayes to the prediction result to get the Bayes_CL model;
[0015] Step 6: Predict the columnar CO concentration based on the Bayes_CL model.
[0016] Preferably, the ConvLSTM2D model and the LightGBM model predict the columnar CO concentration data, and the Bayes_CL model predicts the columnar CO concentration by calculating the coefficient of variation CV, including calculating the coefficient of determination R 2 and root mean square error RMSE;
[0017] Coefficient of determination R 2 The calculation formula is:
[0018]
[0019] Among them, y i is the true value, is the predicted value, is the average of the observed values; when R2 = 1, it means that the model fits the data perfectly; when R 2 =0 means that the model has no explanatory power for the data;
[0020] The formula for calculating the root mean square error is:
[0021]
[0022] Where n is the number of samples, y i is the true value, is the predicted value;
[0023] The smaller the RMSE, the smaller the gap between the model's predicted value and the true value, and the better the model performance.
[0024] Preferably, the generalization ability of the model is also evaluated by performing SCV stratified cross validation; SCV stratified cross validation includes:
[0025] Data stratification: When performing SCV stratified cross-validation, the data set is first stratified according to the category label, that is, the data of the same category is allocated to different training sets and validation sets according to a certain proportion;
[0026] Data division: The stratified data set will be divided into several subsets according to the fold number set by the user. In each iteration, one of the subsets will be selected as the validation set, and the remaining subsets will be used as the training set.
[0027] Cyclic validation: Each iteration trains the model and tests it using the validation set, recording the model’s performance. This layered operation keeps the proportion of each category in the dataset the same during each training and testing.
[0028] Calculate the average performance: After all training and validation, combine the evaluation results of all validation sets and calculate the average performance index to evaluate the generalization ability of the model.
[0029] Compared with the prior art, the present invention has the following beneficial effects:
[0030] (1) Integrate the advantages of different machine learning models:
[0031] LightGBM model: As a machine learning algorithm based on gradient boosting, the LightGBM model performs well in processing irregular functions and large-scale data, and has high prediction accuracy and computational efficiency.
[0032] ConvLSTM2D model: The ConvLSTM2D model combines the advantages of convolutional neural networks (CNN) and long short-term memory networks (LSTM). It can simultaneously capture the spatial and temporal information of CO concentration, has good fitting capabilities for low-frequency signals and long-term dependent data, and is particularly suitable for processing data with significant spatial and temporal changes.
[0033] Bayes_CL model: Through the method of collaborative residual and Bayesian optimization, the Bayes_CL model can combine the advantages of the ConvLSTM2D model and the LightGBM model, integrate the advantages of deep learning and machine learning, and improve prediction stability and accuracy.
[0034] By using these three machine learning models in combination, the present invention can effectively improve the accuracy and stability of columnar CO concentration prediction. Compared with the traditional ARIMA model, the machine learning model can not only handle complex nonlinear relationships, but also provide more reliable results in short-term predictions. It is worth mentioning that the Bayes_C model of collaborative residual and Bayesian optimization combines the model characteristics of the LightGBM model and the ConvLSTM2D model, which can make up for the shortcomings of a single model, provide more comprehensive prediction capabilities, and further optimize the short-term prediction of columnar CO concentration.
[0035] (2) Improved prediction accuracy: Compared with the traditional ARIMA model, the proposed method significantly improves the accuracy of columnar CO concentration prediction, and the determination coefficient R 2 The value increased significantly.
[0036] (3) Enhanced prediction stability: Machine learning models show higher stability when processing environmental data, especially in short-term predictions.
[0037] (4) Adapting to complex data characteristics: The present invention can effectively process nonlinear relationships and spatiotemporal characteristics in the data, making the prediction results more consistent with the actual situation.
[0038] (5) Optimize the prediction process: Integrating the advantages of multiple machine learning models can give full play to the advantages of each model and improve the overall prediction effect. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Figure 1 The technical roadmap for the Bayes_CL model;
[0040] Figure 2 CV graphs of the predicted CO concentration values of the four models and the true columnar CO concentration values;
[0041] Figure 3 This is a comparison chart of the R2 of the Bayes_CL model, LightGBM model, ConvLSTM2D model and ARIMA model in predicting columnar CO from January to June 2023;
[0042] Figure 4 A visual comparison of the original average value of columnar CO concentration and the average value predicted by the four models;
[0043] Figure 5 The columnar CO concentration forecasts of the Bayes_CL model, LightGBM model, ConvLSTM2D model and ARIMA model and the average visualization of the true values from January to June 2023. DETAILED DESCRIPTION
[0044] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
[0045] A short-term prediction method for columnar CO concentration includes a ConvLSTM2D model, a LightGBM model and a Bayes_CL model. The Bayes_CL model is used to perform short-term prediction of columnar CO concentration. The method is as follows:
[0046] Step 1: Collect columnar CO concentration data, which is divided into training set and test set;
[0047] Step 2: Input the training set into the ConvLSTM2D model and LightGBM model for training. Use the trained ConvLSTM2D model and LightGBM model to predict the columnar CO concentration data and compare them with the test set.
[0048] Step 3: According to the comparison results, the residual data between the model prediction value and the true value is obtained;
[0049] Step 4: The residual data is input into the Bayesian model for optimization to obtain the collaborative residual Bayesian;
[0050] Step 5: Take the weighted average of the output results of the ConvLSTM2D model and the LightGBM model to get the prediction result, and then add the collaborative residual Bayes to the prediction result to get the Bayes_CL model;
[0051] Step 6: Predict the columnar CO concentration based on the Bayes_CL model.
[0052] The present invention will be further described below in conjunction with a specific embodiment:
[0053] Step 1. Data acquisition and preprocessing: Columnar CO concentration data were obtained from the European Space Agency’s Sentinel 5P satellite from January 2019 to June 2023. The data were downloaded and obtained using Python on the Google Earth Engine (GEE) platform.
[0054] Step 2. Comparison of prediction methods: Through comparison, it is found that the traditional ARIMA model method has many defects. The LightGBM model and ConvLSTM2D model based on machine learning are better than the ARIMA model method in predicting short-term CO, but each has its own advantages and disadvantages. Therefore, the residuals between the columnar CO concentration values predicted by the LightGBM model and the ConvLSTM2D model and the true value are calculated, and these residuals are combined with the Bayesian model to predict the short-term residuals of the columnar CO concentration. The optimized residuals are integrated into the prediction results of the LightGBM model and the ConvLSTM2D model to generate the Bayes_CL model. Its technical route is as follows Figure 1 shown.
[0055] The Bayes_CL model combines the advantages of the ConvLSTM2D model and the LightGBM model. The detailed Bayes_CL model structure can correspond to the following steps:
[0056] 1. Data introduction: Obtain raster data of columnar CO concentration from the Sentinel 5P satellite. For example, collect the columnar CO concentration data (X) from January to June in a certain area, and the columnar CO concentration data (Y) in July in the same area.
[0057] 2. Model training: After the data introduction part X is injected into the ConvLSTM2D model and the LightGBM model for training, the trained models are used to predict the columnar CO concentration data of the region in July. The results show that the two models have predicted the columnar CO concentration data of the region in July.
[0058] 3. Residual calculation: Since the actual value of the columnar CO concentration in the region in July is already available in the data introduction stage, the prediction results obtained after the above model training are compared with the actual concentration data of the columnar CO in the region in July to obtain the residual between the model prediction value and the actual value.
[0059] 4. Residual optimization: In order to reduce errors, the residual data obtained from the disability calculation step are input into the Bayesian model to obtain the collaborative residual Bayesian.
[0060] 5. Model fusion and training: Finally, the outputs of the ConvLSTM2D model and the LightGBM model are weighted averaged to obtain a prediction result that combines the prediction advantages of both models. Then the collaborative residual Bayes obtained in the residual optimization step is added to the prediction result (that is, the Bayes_CL model is obtained). Finally, the final columnar CO prediction result is obtained.
[0061] In actual training, a stacked training method was used, that is, the CO concentration data from January to June was used for training to predict the CO concentration data in July; the CO concentration data from February to July was used for training to predict the CO concentration data in August. This step-by-step iterative training method improves the prediction accuracy of the model.
[0062] Step 3: Model training and evaluation: The traditional ARIMA model method, the LightGBM model and ConvLSTM2D model based on machine learning, and the proposed Bayes_CL model method are used to train the columnar CO concentration of Sentinel 5P from January 2019 to December 2022. The columnar CO concentration of Sentinel 5P from January 2023 to June 2023 is used as verification, and the ARIMA model, LightGBM model and ConvLSTM2D model are used to predict the columnar CO concentration from January 2023 to July 2023. Finally, the coefficient of variation CV (R 2 and RMSE), such as Figure 2 shown.
[0063] Figure 2 (a) is the coefficient of variation CV between the predicted value of the ARIMA model and the true columnar CO concentration value; (b) is the coefficient of variation CV between the predicted value of the LightGBM model and the true columnar CO concentration value; (c) is the coefficient of variation CV between the predicted value of the ConvLSTM2D model and the true columnar CO concentration value; (d) is the coefficient of variation CV between the predicted value of the Bayes_CL model and the true columnar CO concentration value. R was used 2 , RMSE, n and SCV are used to measure the overall effect of the coefficient of variation CV. The color bar represents the scatter point density.
[0064] As can be seen from the figure, Sentinel 5P provides rich columnar CO concentration data, which can provide enough data for CV even during the validation period. The coefficient of variation CV of all data obtained by the ARIMA model is poor (R 2 =0.61, RMSE = 0.08). The coefficient of variation CV of all data obtained by the LightGBM model is good (R 2 =0.79, RMSE = 0.05). The performance of the ConvLSTM2D model is between the two (R 2=0.75, RMSE=0.06). In addition, in the comparison of SCV, the columnar CO concentration predicted by the ARIMA model has the largest discrete difference from the true value (SCV=34.281), followed by the LightGBM model prediction value (SCV=33.187), and the ConvLSTM2D model has the lowest (SCV=31.861). This means that the columnar CO concentrations simulated by the LightGBM model and the ConvLSTM2D model are closer to the true value. In particular, for the columnar CO concentration obtained by the ConvLSTM2D model, although the R 2 The RMSE and the RMSE are not as good as those of the LightGBM model, but its discreteness is closer to the real columnar CO concentration. This shows that the LightGBM model and the ConvLSTM2D model have their own advantages and disadvantages in simulating columnar CO concentration. In addition, the R 2 =0.8 and RMSE=0.05 are better than the LightGBM model and the ConvLSTM2D model, and SCV=32.095 is between the LightGBM model and the ConvLSTM2D model. This means that the performance of the Bayes_CL model in predicting columnar CO concentration is improved and more stable compared with the LightGBM model and the ConvLSTM2D model.
[0065] In addition, the R of the Bayes_CL model, LightGBM model, ConvLSTM2D model, and ARIMA model in predicting columnar CO concentration from January to June 2023 were compared. 2 The detailed results are as follows Figure 3 As shown:
[0066] Results and Figure 2 The results in the table are similar. The accuracy of the columnar CO concentration predicted by the Bayes_CL model, LightGBM model, and ConvLSTM2D model is better than that of the ARIMA model. Secondly, the LightGBM model is better than the ConvLSTM2D model in most cases, but in the comparison in May 2023, the R 2 The value is higher than that of the LightGBM model. In addition, the R 2 The values are higher than those of the LightGBM model and the ConvLSTM2D model. At the same time, it can be found that as time goes by, the R 2 The value gradually stabilized at around 0.8. This further proves that the Bayes_CL model combines the advantages of the LightGBM model and the ConvLSTM2D model in machine learning and has stronger predictive ability.
[0067] During the test, the average columnar CO concentration forecasts from January to June 2023 by the Bayes_CL model, LightGBM model, ConvLSTM2D model, and ARIMA model were superimposed and compared with the true value of the columnar CO concentration. Figure 4 Results shown:
[0068] Figure 4 (a) represents the original average value of the columnar CO concentration from January to June 2023; (b) represents the average value predicted by the ARIMA model from January to June 2023; (c) represents the average value predicted by the ConvLSTM2D model from January to June 2023; (d) represents the average value predicted by the LightGBM model from January to June 2023. (e) represents the average value predicted by the Bayes_CL model from January to June 2023; the color column represents the concentration value of the columnar CO concentration; the black arrows represent the locations of the outliers predicted by different models; the black boxes represent the sample areas where the ARIMA model predictions are significantly different from the original predictions.
[0069] Compared with the original results, the spatial distribution of columnar CO concentration predicted by the Bayes_CL model, LightGBM model, and ConvLSTM2D model is relatively reasonable. However, in the ARIMA model prediction ( Figure 4 b), there is an obvious anomaly in the box. This does not appear in the predictions of the Bayes_CL model, LightGBM model, and ConvLSTM2D model. The possible reason is that there are certain differences in the inversion values of the ocean and land during the inversion of columnar CO concentration. These differences are amplified in the subsequent ARIMA model training and prediction, and lead to Figure 4 The anomaly in the box in (b) affects the prediction results of the ARIMA model. In addition, Figure 5 The arrows in the figure indicate the hot spots of the true columnar CO concentration. It can be found that the hot spots at arrows 1 and 2 of the true value of the columnar CO concentration are simulated by the Bayes_CL model, the LightGBM model, the ConvLSTM2D model, and the ARIMA model. The ARIMA model and the Bayes_CL model clearly simulate the high CO concentration area in arrow 3. At arrow 4, only the Bayes_CL model shows a trend of simulating high CO concentrations.
[0070] During the test, Figure 4It is difficult to directly distinguish the predictions of the Bayes_CL model, LightGBM model, ConvLSTM2D model, and ARIMA model from the spatial distribution of the true columnar CO concentration values. Therefore, during the test, the predicted values of the Bayes_CL model, LightGBM model, ConvLSTM2D model, and ARIMA model were subtracted from the true values of the columnar CO concentration to obtain Figure 5 .
[0071] Figure 5 In the figure, (a) represents the difference between the predicted value of the LightGBM model and the original columnar CO concentration; (b) represents the difference between the predicted value of the ConvLSTM2D model and the original columnar CO concentration; (c) represents the difference between the predicted value of the Bayes_CL model and the original columnar CO concentration; (d) represents the difference between the predicted value of the ARIMA model and the original columnar CO concentration, a positive value indicates that the predicted value is greater than the original columnar CO concentration, and a negative value indicates that the predicted value is less than the original columnar CO concentration; the prediction period is from January 2023 to June 2023; blue arrows and blue ellipses are hot spots.
[0072] exist Figure 5 The ARIMA model predictions are different from the original bar graph CO concentration at the arrow position ( Figure 5 d) and the circles show significant differences. It is not difficult to see that the difference in the prediction of the ARIMA model is greater than that of the Bayes_CL model, LightGBM model, and ConvLSTM2D model. The Bayes_CL model ( Figure 5 c’s arrow 1”) and the ConvLSTM2D model ( Figure 5 Arrow 1' in b) outperforms the LightGBM model ( Figure 5 In predicting the low values of the columnar CO concentration, the Bayes_CL model ( Figure 5 c’s arrow 2”) and the LightGBM model ( Figure 5 a’s arrow 2), is better than the ConvLSTM2D model ( Figure 5 b) with arrow 2').
[0073] In summary, the Bayes_CL model performs a combination of the ConvLSTM2D model and the LightGBM model in short-term forecasting of Sentinel-5PCO, and is superior to the traditional ARIMA model method.
[0074] The present invention uses the following parameters to illustrate the relationship between the predicted value and the true value of columnar CO concentration:
[0075] 1. Coefficient of determination R 2 , and its calculation formula is:
[0076]
[0077] Among them, y i is the true value, is the predicted value, is the average of the observed values; when R2 = 1, it means that the model fits the data perfectly; when R 2 =0 means that the model has no explanatory power for the data;
[0078] 2. Root mean square error RMSE, the calculation formula is:
[0079]
[0080] Where n is the number of samples, y i is the true value, is the predicted value;
[0081] The smaller the RMSE, the smaller the gap between the model's predicted value and the true value, and the better the model performance.
[0082] 3. The number of samples N: In model training and testing, the number of samples has a great impact on the performance of the model; the larger the sample size, the more likely the model will learn the underlying patterns of the data and have stronger generalization capabilities.
[0083] 4. SCV (Stratified Cross-Validation): The main function of SCV is to help evaluate the generalization ability of the model. Its working principle includes:
[0084] Data stratification: When performing cross-validation, the data set is first stratified according to the category label, that is, data of the same category is allocated to different training sets and validation sets in a certain proportion.
[0085] Data division: The stratified data set will be divided into several subsets according to the number of folds set by the user (such as 5 folds or 10 folds). In each iteration, one of the subsets will be selected as the validation set, and the remaining subsets will be used as the training set.
[0086] Cyclic validation: Each iteration trains the model and tests it using the validation set, recording the model's performance (such as accuracy, F1 value, etc.). This stratified operation keeps the proportion of each category in the data set the same during each training and testing.
[0087] Calculate average performance: After all training and validation, combine the evaluation results of all validation sets to calculate average performance indicators, such as accuracy, precision, recall, F1 value, etc., to evaluate the generalization ability of the model.
[0088] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.
Claims
1. A method for short-term prediction of columnar CO concentration, characterized in that: Including ConvLSTM2D model, LightGBM model and Bayes_CL model, the Bayes_CL model is used to perform short-term prediction of columnar CO concentration, and the method is as follows: Step 1: Collect columnar CO concentration data, which is divided into training set and test set; Step 2: Input the training set into the ConvLSTM2D model and the LightGBM model for training, predict the columnar CO concentration data through the trained ConvLSTM2D model and LightGBM model, and compare them with the test set; Step 3: According to the comparison results, the residual data between the model prediction value and the true value is obtained; Step 4: The residual data is input into the Bayesian model for optimization to obtain the collaborative residual Bayesian; Step 5: Take the weighted average of the output results of the ConvLSTM2D model and the LightGBM model to get the prediction result, and then add the collaborative residual Bayes to the prediction result to get the Bayes_CL model; Step 6: Predict the columnar CO concentration based on the Bayes_CL model.
2. A method for short-term prediction of columnar CO concentration according to claim 1, characterized in that: The ConvLSTM2D model and LightGBM model predict the columnar CO concentration data, and the Bayes_CL model predicts the columnar CO concentration by calculating the coefficient of variation CV, including calculating the determination coefficient R 2 and root mean square error RMSE; Coefficient of determination R 2 The calculation formula is: Among them, y i is the true value, is the predicted value, is the mean of the observed values; when R 2 = 1, indicating that the model fits the data perfectly; when R 2 =0 means that the model has no explanatory power for the data; The formula for calculating the root mean square error is: Where n is the number of samples, y i is the true value, is the predicted value; The smaller the RMSE, the smaller the gap between the model's predicted value and the true value, and the better the model performance.
3. A method for short-term prediction of columnar CO concentration according to claim 1, characterized in that: It also includes the evaluation of model generalization ability by performing SCV stratified cross-validation; SCV stratified cross-validation includes: Data stratification: When performing SCV stratified cross-validation, the data set is first stratified according to the category label, that is, the data of the same category is allocated to different training sets and validation sets according to a certain proportion; Data division: The stratified data set will be divided into several subsets according to the fold number set by the user. In each iteration, one of the subsets will be selected as the validation set, and the remaining subsets will be used as the training set. Cyclic validation: Each iteration trains the model and tests it using the validation set, recording the model’s performance. This layered operation keeps the proportion of each category in the dataset the same during each training and testing. Calculate the average performance: After all the training and validation, the evaluation results of all validation sets are combined to calculate the average performance index to evaluate the generalization ability of the model.