Pollutant diffusion prediction method

Abstract

The invention provides a pollutant diffusion prediction method, and relates to the field of machine learning. The regional climate change data pollutant diffusion prediction process provided by the invention comprises the steps of constructing a regional climate change data set, performing feature selection and layer normalization operation in combination with a Pearson's correlation coefficient, ensuring the correlation and distribution balance of input data, then constructing a plurality of depth coding modules, and predicting the pollutant diffusion of the regional climate change data. Each module comprises an sLSTM, a full-connection neural network and a ReLU activation function, a residual connection mechanism is introduced to capture a complex mode and a potential relationship in data, then the most significant features in a time sequence are extracted by using a maximum pooling technology to generate a final feature vector, and a final prediction result is obtained through the full-connection neural network. Through combination of feature selection, layer normalization, a depth coding module and a maximum pooling technology, the anti-noise capability and prediction precision of the model are effectively enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine learning, and in particular relates to a pollutant diffusion prediction method. Background Art

[0002] Regional climate change has an important impact on the diffusion of pollutants. Meteorological conditions such as temperature, humidity, wind speed, wind direction, and atmospheric pressure lead to diverse changes in the diffusion and deposition patterns of pollutants in the atmosphere. Accurately predicting the diffusion of pollutants under regional climate change conditions is crucial for environmental protection, air quality management, and public health risk assessment. It can help formulate response measures in advance and prevent the negative impact of pollutant diffusion on the environment and people.

[0003] Traditional pollutant diffusion prediction methods usually rely on meteorological data and pollutant emission data for modeling. These methods can provide useful prediction results to a certain extent, but they also have limitations. First, this method has extremely high requirements on the quality and integrity of the data. Any inaccuracy in the data may lead to deviations in the prediction results. Second, in the context of climate change, traditional models often find it difficult to accurately capture the complex impact of climate change on pollution diffusion, thereby reducing the accuracy and reliability of the prediction.

[0004] Modern forecasting technologies, such as machine learning and big data analysis, provide new tools and methods to improve the limitations of traditional models. By utilizing a large amount of regional climate change data and using advanced algorithms to conduct comprehensive analysis of these data, we can more comprehensively and accurately predict the diffusion of pollutants under regional climate change, which can not only improve the accuracy of the prediction, but also enhance the adaptability and flexibility of the model to future changes. Summary of the invention

[0005] The present invention provides a pollutant diffusion prediction method, which aims to introduce the Pearson correlation coefficient for feature selection and effectively eliminate redundant data, thereby reducing data dimensions and improving computing efficiency. Secondly, a deep coding module is used to mine deep features in time series data to capture potential patterns and complex relationships of the data. Finally, feature fusion technology is used to organically integrate feature vectors at different time points to generate the final feature representation for pollutant diffusion prediction, thereby achieving accurate prediction of pollutant diffusion in regional climate change data.

[0006] The technical method adopted by the present invention to achieve the above-mentioned purpose specifically comprises the following steps: S1. Collect regional climate change data, including temperature, humidity, wind speed, wind direction, atmospheric pressure, precipitation, pollutant concentration monitoring data, and emission source intensity, and process the collected data to provide a solid data foundation for subsequent pollutant diffusion prediction; S2. Use the Pearson correlation coefficient to select data and filter out feature data with a high linear correlation with the target prediction value. By removing redundant and irrelevant data, the complexity of the model is simplified and the performance of the model is improved. S3. Apply layer normalization technology to the filtered data to ensure balanced data distribution for subsequent model processing and ensure the stability of network training; S4. Build a deep encoding module to mine the deep features in regional climate change data and capture the complex patterns and long-term dependencies in the data. The specific steps include: S41. Construct a sequence memory module and use an enhanced long short-term memory network sLSTM to capture the complex temporal dependencies in the data, so as to better describe the spatiotemporal dynamics of pollutant diffusion; S42, construct a perception activation module, use a fully connected neural network and ReLU activation function to further mine deep information in data features, and add the feature vector generated by this module to the input data in step S41 element by element to reduce the gradient disappearance and model degradation problems that may occur in the deep network, thereby improving the prediction effect and training efficiency of the model; S5, using the idea of ​​maximum pooling, fuses all feature vectors processed by the deep coding module to extract key feature information; S6. Construct a fully connected neural network, perform further calculations on the feature representation, and output the pollutant diffusion prediction results of regional climate change data.

[0007] Preferably, in step S1, regional climate change data are collected, including 8 types of data, including temperature, humidity, wind speed, wind direction, atmospheric pressure, precipitation, pollutant concentration monitoring data, and emission source intensity. The collected data are cleaned to remove noise and outliers, improve the accuracy and reliability of the data, and finally construct a high-quality regional climate change dataset.

[0008] Preferably, in step S2, the monitoring value sequence of the i-th type of data is input Where T is the sequence length of the monitoring data, is the monitoring value of the i-th type of data at the Tth moment, and the actual pollutant diffusion risk value of the corresponding sequence length is input at the same time in is the pollutant diffusion risk value at time T. The Pearson correlation coefficient is used to calculate the linear correlation between the collected data and the actual pollutant diffusion risk value. The specific calculation process is as follows: In the formula is the Pearson correlation coefficient, ranging from [-1, 1], where indicates a completely positive correlation. indicates a completely negative correlation. represents no linear correlation, is the mean of the i-th category data, is the mean of the real pollutant diffusion risk values.

[0009] Preferably, by using the Pearson correlation coefficient to calculate the linear correlation between the collected regional climate change data and the pollutant diffusion risk value, it is possible to effectively screen the characteristic data that contributes most to the pollutant diffusion prediction, remove redundant or irrelevant characteristic data, and determine the data features with strong correlation at an early stage. This not only simplifies the complexity of the model, but also improves the calculation efficiency and accuracy of the model.

[0010] Preferably, in step S3, data types whose absolute value of the Pearson correlation coefficient is greater than or equal to a certain threshold are selected, and a total of n data types are selected, and time series data is constructed based on the selected data. Where T is the sequence length, n is the number of selected data types, is the data sequence at time t∈[1,T], is the detection value of the i-th category data at the t-th time. The time series data V is layer normalized. The specific calculation process is as follows: Where μ is v t The mean of all values ​​in , σ is v t is the standard deviation of , ε is a small constant used to avoid the denominator being 0, v t The normalized output value, γ and β are learnable scaling and offset parameters respectively.

[0011] Preferably, the filtered data is constructed into time series data V, and then layer normalization is performed to ensure balanced data distribution at each moment and prevent gradient vanishing or gradient exploding problems. At the same time, layer normalization is suitable for small batch training, which can enhance the training stability of the model, increase the convergence speed, and improve the performance of the model in long time series prediction.

[0012] Preferably, in the steps S4 and S41, a sequence memory module is constructed, and the time series data after layer normalization is input into the enhanced long short-term memory network sLSTM. The specific calculation process is as follows: n t =f t n t-1 +i t ; c t =f t c t-1 +i t z t ; In the formula is the learnable weight parameter, is the bias parameter, σ is the Sigmoid function, They are the input gate, forget gate and output gate of LSTM at timestamp t, is the normalized state at timestamp t, is the candidate memory unit at timestamp t, is the memory unit at timestamp t, is the hidden state at timestamp t.

[0013] Preferably, sLSTM enhances the memory capacity of traditional LSTM by introducing exponential activation function and standardized state, especially when dealing with long time series dependencies, it can dynamically adjust the previous storage information, and can achieve better results in tasks that require accurate tracking of state changes.

[0014] Preferably, in steps S4 and S42, a perception activation module is constructed, and a fully connected neural network and a ReLU activation function are used to further t The specific calculation process is as follows: In the formula h t The feature vector after the fully connected neural network and ReLU activation function, is the learnable weight parameter, is the bias parameter, which will be Add the input data in step S41 element by element, use the result of the addition as the input feature of the next step, and repeat step S4 N times, where N is the number of deep encoding modules, to ensure that higher-level features are gradually extracted during the deep learning process.

[0015] Preferably, a fully connected neural network and a ReLU activation function are used to transform the hidden state ht It performs nonlinear transformation and strengthens the nonlinear representation of data. At the same time, it introduces a residual connection mechanism to add the newly generated features to the original input element by element, ensuring that information is not lost in the deep network and enhancing the feature transmission capability, so that the network can better capture complex time series relationships and provide richer and more accurate feature information for subsequent prediction tasks.

[0016] Preferably, in step S5, for the new time series data Use the maximum pooling idea to select the maximum value of each feature dimension element by element along the time axis to form the final feature vector. The specific calculation process is as follows: In the formula A vector consisting of the maximum values ​​of each feature dimension in H is selected element by element. MaxPooling(·) is an operation that selects the maximum value along the time axis for each feature dimension.

[0017] Preferably, this step extracts the most significant features in the time series data through the maximum pooling operation. The operation of selecting the maximum value element by element along the time axis can retain the most important information of each feature dimension at different time steps, reduce the interference of redundant data, and ensure that even short peaks or important changes in long time series can be effectively captured, thereby improving the model's robustness to noise and the accuracy of prediction.

[0018] Preferably, in step S6, f is passed through a fully connected neural network to calculate the final regional climate change data pollutant diffusion prediction value: P = Sigmoid(Linear(f)); Where Linear(·) is a fully connected neural network, Sigmoid(·) is the Sigmoid function, It is the final prediction value of pollutant diffusion based on regional climate change data.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention combines the Pearson correlation coefficient to perform feature selection and layer normalization operations to ensure the correlation and distribution balance of the input data. Secondly, the sLSTM and fully connected neural network are introduced, and the residual connection mechanism is combined to enhance the integrity of feature transfer and the nonlinear expression ability of the model. The maximum pooling is used to extract the most significant features in the time series, enhance the model's anti-noise ability, and ultimately improve the model's prediction accuracy and robustness in complex time series tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A step-by-step diagram of a pollutant diffusion prediction method.

[0021] Figure 2 Feature selection plot for Pearson correlation coefficient.

[0022] Figure 3 This is the deep coding module diagram.

[0023] Figure 4 Merge graph for feature vectors.

[0024] Figure 5 This is a diagram showing the effect of predicting the possibility of pollutant diffusion. DETAILED DESCRIPTION

[0025] The present invention proposes a pollutant diffusion prediction method. The method combines the Pearson correlation coefficient to perform feature selection and layer normalization operations to ensure the correlation and distribution balance of input data. Secondly, multiple deep encoding modules are constructed, each module introduces sLSTM, fully connected neural network and ReLU activation function, and combines the residual connection mechanism to enhance the model's processing ability for long-term series dependence and effectively capture deep-level feature information. Finally, the maximum pooling is used to extract the most significant features of the time series, and the fully connected network is used to achieve accurate prediction of pollutant diffusion in regional climate change data. The technical solution in the embodiment of the present invention is described in detail and completely below, specifically including the following steps: Figure 1 shown.

[0026] S1. Collect 730 days of regional climate change data, including the following 8 types of data: temperature, humidity, wind speed, wind direction, atmospheric pressure, precipitation, pollutant concentration monitoring data, and emission source intensity. Clean the collected data, including removing noise, filling missing values, and removing outliers, to build a high-quality regional climate change dataset. Then divide the dataset into training set and validation set in a ratio of 7:3. The data of each day uses the climate change data of the previous 14 days as input, which serves as the basis for model learning and prediction.

[0027] S2. First, calculate the Pearson correlation coefficient between each feature data and the target prediction value to quantify the linear correlation between each feature and the target prediction value, screen out feature data with a high linear correlation with the target prediction value, remove redundant and irrelevant data, simplify the complexity of the model, and improve the performance of the model.

[0028] Furthermore, the step S2 is as follows: Figure 2 As shown, the monitoring value sequence of the input i-th type of data Among them, 14 is the sequence length of the monitoring data, is the monitoring value of the i-th type of data at the Tth moment, corresponding to the real pollutant diffusion risk value with a sequence length of 14 in is the pollutant diffusion risk value at time T. The Pearson correlation coefficient is used to calculate the linear correlation between the collected data and the actual pollutant diffusion risk value. The specific calculation process is as follows: In the formula is the Pearson correlation coefficient, ranging from [-1, 1], where indicates a completely positive correlation. indicates a completely negative correlation. represents no linear correlation, is the mean of the i-th category data, is the mean of the real pollutant diffusion risk values.

[0029] S3. Apply layer normalization technology to the filtered data, normalize the mean and variance of each input feature, and ensure that the mean of the data in each layer is 0 and the variance is 1.

[0030] Furthermore, in step S3, based on the calculation of the Pearson correlation coefficient, data types with an absolute value of the Pearson correlation coefficient greater than or equal to 0.3 are selected, and a total of n=7 data types with high correlation are selected, and the selected data are constructed as time series data. Among them, 14 is the sequence length, 7 is the number of selected data types, and the data sequence at time t∈[1, T] is For example, is the detection value of the i-th category data at the t-th time. The time series data V is layer normalized. The specific calculation process is as follows: Where μ is v t The mean of all values ​​in , σ is v t is the standard deviation of , ε is a small constant used to avoid the denominator being 0, v t The normalized output value, γ and β are learnable scaling and offset parameters respectively.

[0031] S4. Construct a deep encoding module to mine the deep features in the regional climate change data and capture the complex patterns and long-term dependencies in the data, which specifically includes the following steps: S41. Construct a sequence memory module and use an enhanced long short-term memory network sLSTM to capture the complex temporal dependencies in the data; S42. Build a perception activation module, use a fully connected neural network and ReLU activation function to enhance the nonlinear expression ability of the model, so that the model can better capture complex data features. Finally, add the feature vector generated by this module to the input data in step S41 element by element to form the input feature of the next step.

[0032] Further, the step S4, as Figure 3 As shown, in S41, a sequence memory module is constructed to convert the time series data after layer normalization into Input into the enhanced long short-term memory network sLSTM, the specific calculation process is as follows: n t =f t n t-1 +i t ; c t =f t c t-1 +i t z t ; In the formula is the learnable weight parameter, is the bias parameter, They are the input gate, forget gate and output gate of LSTM at timestamp t, is the normalized state at timestamp t, is the candidate memory unit at timestamp t, is the memory unit at timestamp t, is the hidden state at timestamp t.

[0033] Further, the step S4, as Figure 3 As shown, the perception activation module is constructed in S42, and the fully connected neural network and ReLU activation function are used to further t For further processing, the specific calculation process is as follows: In the formula h t The feature vector after the fully connected neural network and ReLU activation function, is the learnable weight parameter, As the bias parameter, the idea of ​​residual connection is then adopted, which will be Add the input data in step S41 element by element, that is, combine the original input with the newly generated features, and use the result as the input feature of the next step. Repeat step S4 6 times, where 6 is the number of deep encoding modules, to ensure that higher-level features are gradually extracted during the deep learning process.

[0034] S5. Adopt the idea of ​​maximum pooling to fuse all feature vectors processed by the deep coding module, that is, select the values ​​of all moments of each feature dimension element by element along the time axis, select the maximum value for fusion to obtain a new feature vector.

[0035] Furthermore, the step S5 is as follows: Figure 4 As shown, for the new time series data Use the maximum pooling idea to select the maximum value of each feature dimension element by element along the time axis to form the final feature vector. The specific calculation process is as follows: In the formula is a vector consisting of the maximum values ​​of each feature dimension in H selected element by element, and MaxPooling(·) is an operation of selecting the maximum value along the time axis for each feature dimension.

[0036] S6. Construct a fully connected neural network, perform further calculations on the feature representation, and output the pollutant diffusion prediction results of regional climate change data.

[0037] Furthermore, in step S6, f is calculated through a fully connected neural network to obtain the final regional climate change data pollutant diffusion prediction value: P = Sigmoid(Linear(f)); Where Linear(·) is a fully connected neural network, Sigmoid(·) is the Sigmoid function, It is the final prediction value of pollutant diffusion based on regional climate change data.

[0038] Furthermore, the method is implemented in Python 3.9 and the PyTorch framework is used for model building and training in the CUDA 12 environment. The entire training process is run on the NVIDIA RTX 3090 GPU, which is equipped with 24GB of video memory and can handle large-scale deep learning models and accelerate calculations. The batch size set during training is 32 and the learning rate is 1×10 -4 , using the AdamW optimizer.

[0039] Furthermore, the method predicts the effect as Figure 5As shown in the figure, the ordinate is the possibility of pollutant diffusion (%), the abscissa is time (days), the gray dotted line represents the actual pollutant diffusion risk value, and the black straight line represents the pollutant diffusion prediction value. It can be seen from the figure that the pollutant diffusion prediction value has a high degree of fit with the actual pollutant diffusion risk value, which proves the effectiveness of the method.

[0040] The above are only preferred embodiments of the present invention. It should be pointed out that a person skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A pollutant diffusion prediction method, characterized in that: The following steps are involved: S1. Collect regional climate change data, including temperature, humidity, wind speed, wind direction, atmospheric pressure, precipitation, pollutant concentration monitoring data, and emission source intensity, a total of 8 types of data, process the collected data, remove noise, fill in missing values, and eliminate outliers, to provide a solid data foundation for subsequent pollutant diffusion prediction; S2. Use the Pearson correlation coefficient to evaluate the linear correlation between each feature data and the target prediction value, so as to screen out the feature data with a high linear correlation with the target prediction value, remove redundant and irrelevant data, simplify the complexity of the model, and improve the performance of the model; S3. Apply layer normalization technology to the filtered data to ensure balanced data distribution for subsequent model processing and ensure the stability of network training; S4. Build a deep encoding module to mine the deep features in regional climate change data and capture the complex patterns and long-term dependencies in the data. The specific steps include: S41, build a sequence memory module, using enhanced long short-term memory network sLSTM, which captures the complex temporal dependencies in the data through its special gating mechanism, thus better describing the spatiotemporal dynamics of pollutant diffusion; S42, construct a perception activation module, use a fully connected neural network and ReLU activation function to further mine deep information in the data features, and implement residual connection by adding the feature vector generated by this module to the input data in step S41 element by element, thereby alleviating the gradient vanishing and model degradation problems that may occur in the deep network and improving the prediction effect and training efficiency of the model; S5. Adopt the idea of ​​maximum pooling to fuse all feature vectors processed by the deep coding module, process the features at each moment along the time axis, and retain the most significant part of each feature dimension by selecting the maximum value in the feature element by element; S6. Construct a fully connected neural network, further calculate the final feature representation, and output the pollutant diffusion prediction results of regional climate change data.

2. A pollutant diffusion prediction method according to claim 1, characterized in that: In step S2, the monitoring value sequence of the i-th type of data is input. The actual pollutant diffusion risk value corresponding to the sequence length Among them, 14 is the sequence length of the monitoring data, is the monitoring value of the i-th type of data at the T-th moment, is the pollutant diffusion risk value at time T. The Pearson correlation coefficient is used to calculate the linear correlation between the collected data and the actual pollutant diffusion risk value. The specific calculation process is as follows: In the formula is the Pearson correlation coefficient, ranging from [-1, 1], where indicates a completely positive correlation. represents no linear correlation, indicates a completely negative correlation. is the mean of the i-th category data, is the mean of the real pollutant diffusion risk values.

3. A pollutant diffusion prediction method according to claim 2, characterized in that: In step S3, the data types whose absolute value of the Pearson correlation coefficient is greater than or equal to a certain threshold are selected, and a total of n data types are selected. Based on the selected data, a time series data is constructed. Where T is the sequence length, n is the number of selected data types, is the data sequence at time t∈[1,T], where is the detection value of the i-th category data at the t-th moment, and the time series data V is layer normalized. The specific calculation process is as follows: In the formula v t The normalized output value, γ and β are learnable scaling and offset parameters, respectively. v t The mean of all values ​​in , v t is the standard deviation of , and ε is a small constant used to avoid the denominator being zero.

4. A pollutant diffusion prediction method according to claim 3, characterized in that: In the step S4, a deep coding module is constructed. In the specific step S41, a sequence memory module is constructed, and the time series data after layer normalization is processed using an enhanced long short-term memory network sLSTM. The specific calculation process is as follows: n t =f t n t-1 +i t ; c t =f t c t-1 +i t z t ; In the formula is the learnable weight parameter, is the bias parameter, σ is the Sigmoid function, They are the input gate, forget gate and output gate of LSTM at timestamp t, is the normalized state at timestamp t, is the candidate memory unit at timestamp t, is the memory unit at timestamp t, is the hidden state at timestamp t.

5. A pollutant diffusion prediction method according to claim 4, characterized in that: In the step S4, a deep coding module is constructed. In the specific step S42, a perceptual activation module is constructed to further process the hidden state h using a fully connected neural network and a ReLU activation function. t , the specific calculation process is as follows: In the formula h t The feature vector after the fully connected neural network and ReLU activation function, is the learnable weight parameter, is the bias parameter, which will be The input data in step S41 is added element by element to form a new feature representation as the input feature of the next step. Step S4 is repeated N times, that is, N deep encoding modules are constructed to ensure that higher-level features are gradually extracted during the deep learning process.

6. A pollutant diffusion prediction method according to claim 5, characterized in that: In step S5, the maximum pooling idea is used to perform the new time series data. The maximum value is selected element by element for each feature dimension along the time axis to form the final feature vector. The specific calculation process is as follows: In the formula is a vector consisting of the maximum values ​​of each feature dimension in H selected element by element, and MaxPooling(·) is an operation of selecting the maximum value along the time axis for each feature dimension.

7. A pollutant diffusion prediction method according to claim 1, characterized in that: In order to predict the diffusion of pollutants based on regional climate change data, we collected climate change data, including temperature data, humidity data, wind speed data, wind direction data, atmospheric pressure data, precipitation data, pollutant concentration monitoring data, and emission source intensity data. We preprocessed all the data, including removing noise, filling missing values, and removing outliers, to build a high-quality regional climate change dataset. The dataset was then divided into a training set and a validation set as the basis for model learning and pollutant diffusion prediction.

Images