CN115659609A - A Noise Prediction Method for Chemical Industry Park Based on DTW-DCRNN - Google Patents

Abstract

The invention discloses a DTW-DCRNN-based chemical industry park noise prediction method, belongs to the field of signal and information processing, and solves the problem of long-term global noise prediction of a chemical industry park. And (3) introducing a time dynamic regularization theory into the diffusion convolution recurrent neural network, and establishing a space-time noise prediction network model consisting of a time dynamic regularization method, the diffusion convolution recurrent neural network and Kalman filtering. Improving a DTW algorithm by using a penalty coefficient, and reconstructing the spatial relationship of the neural network through a DTW model; the output noise prediction of the neural network is dynamically adjusted using a Kalman method in conjunction with the traffic flow characteristics. According to the method, a spatial relationship is reconstructed based on time sequence similarity, and then data of each station are sent into a model for prediction and dynamic correction, so that multi-station noise prediction, long-term noise prediction and global noise level prediction are realized, and risks of noise disturbing residents and health damage can be avoided in advance.

Description

A Noise Prediction Method for Chemical Industry Park Based on DTW-DCRNN

technical field

The invention belongs to the field of signal and information processing, and in particular relates to a DTW-DCRNN-based noise prediction method for chemical industry parks.

Background technique

With the continuous development of industry, the problem of environmental noise pollution has become increasingly severe with the rapid development of urban society and economy, and the main problems in the prevention and control of noise pollution have become more and more obvious. At present, the research on neural network continues to advance, so there is an urgent need for a method based on neural network for long-term prediction of environmental noise.

The rise of the concept of "smart chemical park" has laid a good foundation for the accurate and effective collection of various data. The smart park construction of Dongming Engineering Plastics Industrial Park is relatively complete, and the monitoring data is complete and the data volume is sufficient, which provides a good data basis for the development of noise prediction research in chemical parks. Modeling, analysis and prediction of data based on deep learning methods, and mining of noise data rules can help park managers take preventive measures in advance during periods of high noise decibels, such as: park operators wear noise-canceling headphones during specified time periods, Set up noise baffles in advance or place equipment with excessive production noise away from residential areas. On the one hand, these tasks can provide operators with a relatively safe working environment through predictive reminders and early warning analysis, and on the other hand, they can also reduce the impact of noise on the lives of residents near the park. In addition, different locations in the park have different noise levels, and the global prediction of environmental noise is also of great significance to the assessment of occupational hazards in the park and the prediction of environmental risks.

Contents of the invention

Based on the above problems, the present invention proposes a noise prediction method for chemical industry parks based on DTW-DCRNN. From the perspective of space-time prediction, a space-time noise composed of improved time dynamic warping algorithm, diffusion convolution recurrent neural network and Kalman filter is constructed. A predictive network model is used to make long-term and global predictions of noise in chemical industrial parks.

In order to achieve the above object, the present invention adopts the following technical solutions:

A noise prediction method for chemical industry parks based on DTW-DCRNN, which uses an improved time dynamic warping algorithm and reconstructs the spatial relationship through the graph structure of monitoring stations to realize long-term spatiotemporal prediction of noise in chemical industry parks. It specifically includes the following steps:

Step 1. Collect and process the monitoring site information and vehicle information data in the chemical industry park;

Step 2. Introduce the penalty coefficient to improve the original time dynamic warping algorithm, and introduce it into the diffusion convolution recurrent neural network, and then combine the Kalman filter to construct the noise prediction network model;

Step 3, train the constructed noise prediction network model, and output the trained model;

Step 4. Collect monitoring site information in real time and perform real-time prediction of noise based on the trained model.

Further, in step 1, the road gate and several monitoring stations are set in the chemical industry park for real-time data collection, and the location information data of each monitoring station are recorded synchronously; the road gate and each monitoring station are equipped with sensors; Information data, including license plate number, vehicle entry and exit time, and vehicle flow characteristics of vehicle models; monitoring stations collect real-time noise and natural environment data in the chemical industrial park;

Site location data, noise and natural environment data, and vehicle information data entering and leaving will be transmitted to the gateway device, and then interactively transmitted with the database server. When noise prediction is performed, data is extracted in real time and processed using the 3σ criterion.

Further, the specific process of step 2 is as follows:

Step 2.1, introducing a penalty coefficient to improve the original time dynamic warping algorithm, and calculating the sequence similarity distance;

The principle of the improved time dynamic warping algorithm is: obtain the optimal path and the number of common subsequences to calculate the penalty coefficient, and combine the penalty coefficient to calculate the similarity distance between stations; the specific process is as follows:

Step 2.1.1, calculate the distance matrix;

Suppose there are two noise sequences S and T, such as formula (1):

The length of the sequence S is o, the length of the sequence T is m, o, m∈Z ⁺ , construct a distance matrix d[s][t] based on the noise sequence m×o, where the i-th sequence point s _i (i=1, 2,...,o) and the jth sequence point t _j (j=1,2,...,m) is the value of the element of the matrix (s _i ,t _j ), Euclidean distance: dis(s _i ,t _j )＝(s _i -t _j ) ² , the matrix distance calculation standard is shown in formula (2):

Step 2.1.2, seeking the optimal path of the matrix;

The optimal path is the path with the minimum cumulative distance that satisfies the lower boundary condition. The lower boundary condition is: starting from the point d[o][m] in the upper right corner of the matrix, find the smallest point among the three lower left points as the next node until The d[0][0] matrix point in the lower left corner ends. In addition, continuity and monotonicity must be ensured. A point cannot be crossed or missed for matching. It can only be aligned or matched with adjacent points. The sequence of time series The order cannot be changed;

Step 2.1.3, calculating the number of common subsequences and the length of the optimal path sequence, and then obtaining the weight and penalty coefficient;

Wherein, the optimal path sequence length is the length of the matrix optimal path finally obtained in step 2.1.2;

The calculation process of the number of common subsequences is as follows: firstly construct the record empty list record, loop through the sequences s1 and s2, use the record list to record the length ls1 or ls2 of the common sequence and the position g of the common sequence after the common sequence appears, and transfer the common subsequence from s1 The sequence is extracted and stored in the s_sum array; after the traversal, the length of s_sum is the number of common subsequences; when calculating the number of common subsequences, set the threshold to 1, and count subsequences with a length greater than 1 as common subsequences. Finally, the total number of common subsequences is counted as the required number of common noise subsequences;

The calculation formula of weight w is as formula (3):

Among them, subseq is the length of the common noise subsequence, and seq is the sequence length of the optimal path of the matrix;

The calculation process of setting the penalty coefficient α is shown in formula (4), where x represents the number of common subsequences, w _i represents the weight of the i-th sequence point,

Step 2.1.4. Express the optimal path as (r ₁ ,r ₂ ,…,r _seq ), r _i (i=1,2,…,seq) represents the value of the i-th sequence point in the optimal path sequence , and finally multiply the penalty coefficient with the sum of values in the optimal path calculated by the original DTW to obtain the improved sequence similarity distance dis _DTW ;

Step 2.2, the noise sequence similarity distance between each monitoring site is formed into a matrix and calculated as an adjacency matrix to construct a graph relationship topology; the specific process is:

Step 2.2.1, according to the noise sequence similarity distance information of each monitoring station, calculate the similarity distance matrix M between the stations, the matrix M is as follows:

Among them, a is the number of monitoring stations;

Step 2.2.2, calculate the standard deviation λ of the similarity distance matrix M, M _c is the cth element value in the matrix, N is the number of elements in the M matrix, μ _M is the matrix mean value, and the calculation process is as follows:

的计算过程如下：Step 2.2.3, analyze the similarity between monitoring sites according to the similarity distance matrix M, use the standard deviation λ of all non-infinite numbers in the similarity distance matrix to construct the adjacency matrix M _d , each element in the adjacency matrix M _d

The calculation process is as follows:

越小，设置阈值为0.1，若

＜阈值，则视为两站点间相似度太低，无相互影响关系，视作非邻接站点，在邻接矩阵中权值为零不构成邻接关系；Among them, c represents the serial number of matrix elements, and the larger the similarity distance is, the

The smaller the value, set the threshold to 0.1, if

<Threshold value, it is considered that the similarity between the two sites is too low, there is no mutual influence relationship, it is regarded as a non-adjacent site, and the weight value of zero in the adjacency matrix does not constitute an adjacency relationship;

Step 2.3, introduce the graph convolutional neural network GCN to construct the adjacency matrix graph structure, input the adjacency matrix graph structure into the diffusion convolution recurrent neural network for noise prediction, and obtain preliminary prediction results; the specific process is as follows:

The propagation mode between the network layers of the graph convolutional neural network GCN is formula (12), and L represents the Lth network layer:

为度矩阵，其构成为：

I表示邻接矩阵A第I行，J表示邻接矩阵A第J列，

表示保留站点本身的特征信息，H为当前层提取到的特征，如果是输入层则X＝H；Among them, σ is a nonlinear activation function; W ^L is a trainable weight matrix,

is a degree matrix, and its composition is:

I represents row I of adjacency matrix A, J represents column J of adjacency matrix A,

Indicates to keep the feature information of the site itself, H is the feature extracted by the current layer, if it is the input layer, then X=H;

Based on the GCN graph convolutional neural network, the spatial-temporal relationship of the noise sequence is modeled using a diffusion convolutional recurrent neural network;

The stationary distribution of the diffusion process is expressed as a weighted combination of infinite random walks on the graph, and the diffusion process is expressed as Equation (13) and calculated in closed form:

是出度矩阵的逆矩阵，β∈[0,1]表示重启概率，k是扩散程度，ε表示从节点扩散的可能性，DCRNN模型中扩散过程是双向的；where W is the node similarity matrix,

is the inverse matrix of the out-degree matrix, β∈[0,1] represents the restart probability, k is the diffusion degree, ε represents the possibility of spreading from the node, and the diffusion process in the DCRNN model is bidirectional;

与滤波器f_θ双向扩散卷积运算定义为式(14)：Therefore, in terms of spatial relationship, based on the graph signal feature matrix

The bidirectional diffusion convolution operation with the filter f _θ is defined as formula (14):

与

分别表示扩散与逆扩散的过度矩阵，定义卷积运算后，构建带有映射关系的扩散卷积层如式(15)：Among them, θ is the filter parameter, G represents the graph G, p represents the pth feature dimension, P represents the total number of feature dimensions, K represents the finite K-step truncation of the diffusion process, G represents the diffusion convolution on the graph G, X: , _p represents the convolution operation of all nodes on the p-th feature, θ _k,1 represents the convolution kernel parameter for out-degree calculation, θ _k,2 represents the convolution kernel parameter for in-degree calculation,

and

Represent the transition matrix of diffusion and inverse diffusion respectively. After defining the convolution operation, construct the diffusion convolution layer with mapping relationship as formula (15):

为输入，当前层提取到的特征

为输出，

表示滤波器，Θ为参数张量，σ为激活函数；Among them, q represents the qth feature, Q represents the total number of mapping features, H:, _q represents the diffusion and convolution operation of all nodes on the qth feature, and the feature matrix

As input, the features extracted by the current layer

for the output,

Represents a filter, Θ is a parameter tensor, and σ is an activation function;

In terms of time relationship, the combination of GRU and diffusion convolution is used to construct the diffusion convolution gated recursive unit DCGRU, which is expressed as formula (16):

★G denotes the diffuse convolution on the graph G, r ^(t) denotes the reset gate at time t, Θ _r★G denotes the parameter tensor of the reset gate, X ^(t) denotes the input at time t, H ^{(t- 1)} represents the output at time t-1, b _r represents the bias vector of the reset gate; u ^(t) represents the update gate at time t, Θ _{u G} represents the parameter tensor of the update gate, and b _u represents the update gate C ^(t) represents the hidden state of the next cell at time t, Θ _{C G} represents the parameter tensor of the hidden state, b _C represents the bias vector of the hidden state; H ^(t) represents the cell at time t Output;

Step 2.4, extract the traffic flow characteristics at the gate corresponding to the current time in real time, and use the Kalman filter method to dynamically adjust the DCRNN noise prediction value;

Kalman filtering uses the estimate of the previous state to make a prediction of the current state; finally, the observed value of the current state is used to correct the predicted value obtained in the prediction stage to obtain a new estimated value that is closer to the real value; the specific implementation is The normalized traffic flow characteristics and noise sequence are estimated and predicted by the covariance matrix to calculate the Kalman gain of noise and drift, and the covariance matrix is updated by Kalman gain correction, and then the noise value at the current moment is corrected and predicted .

Further, the specific process of step 3 is as follows:

Extract the historical data of each monitoring station and use the improved DTW algorithm to reconstruct the spatio-temporal relationship of the data, train the noise prediction network model DTW-DCRNN, select 70% of the data set as the training set, 10% as the verification set, and the last 20% as the test set ;In addition, the training results are not recorded in the first 60 Epochs, the training results are recorded every 10 Epochs in the 60-100 Epochs and the model parameters are output, and the training results are recorded and the model is output every 5 Epochs after 100 Epochs parameter; finally stop training early based on the measured validation loss to ensure that the model is caught when it is close to overfitting.

Beneficial technical effects brought by the present invention:

Introducing the penalty coefficient to improve the original time dynamic rule algorithm, constructing the adjacency matrix relationship through the similar distance matrix to represent the spatial distance relationship can better make up for the limitations of the original physical distance method; the prediction result adjusted by the Kalman filter is also closer to the real value , and with compensation properties, the 3σ noise processing method can effectively avoid the large deviation of the model training prediction results. The spatio-temporal noise prediction network model established by the present invention is composed of time dynamic warping method, diffusion convolution recursive neural network, and Kalman filter, which can realize multi-site noise prediction, long-term noise prediction, and global noise level prediction, thereby avoiding noise disturbance in advance, Risk of damage to health.

Description of drawings

Fig. 1 is the flowchart of the chemical industry park noise prediction method based on DTW-DCRNN of the present invention;

Fig. 2 is a schematic diagram of the data collection process of the present invention;

Fig. 3 is a schematic diagram of the noise prediction deep learning network model of the present invention;

Fig. 4 is the schematic diagram of improved DTW algorithm of the present invention;

Fig. 5 is the flowchart of calculating the number of common subsequences of the present invention;

Fig. 6 is the structural representation of the diffusion convolution recursive neural network model of the present invention;

Fig. 7 is a comparison diagram of adjacency matrix changes in Experiment 1 of the present invention; wherein, (a) shows the adjacency matrix built by the original DTW algorithm, and (b) shows the adjacency matrix built by the improved DTW algorithm;

Fig. 8 is the index variation trend figure based on time step in experiment 1 of the present invention;

Fig. 9 is the curve comparison diagram of the performance index MAE and RMSE of the prediction model of the present invention and the DCRNN model in experiment 2 of the present invention;

Fig. 10 is a curve comparison chart of the long-term performance index MAPE between the prediction model of the present invention and the DCRNN model in Experiment 2 of the present invention.

Detailed ways

Below in conjunction with accompanying drawing and specific embodiment the present invention is described in further detail:

The present invention mainly solves the problems existing in the noise prediction of chemical industry parks from the following four aspects: the noise data has zero values, which will greatly interfere with the prediction accuracy. Value filtering can reduce the standard deviation of the sample unbiased ratio, thereby reducing the prediction error. The noise prediction of a single monitoring site is not related to other sites, and the weights cannot be shared. Therefore, in the face of the problem that the data cannot be communicated between the single-site predictions, and the calculation performance of each site is high, the prediction weight of the data information between the sites is realized through the graph neural network. shared. At present, the noise similarity between non-adjacent monitoring stations is not involved in the calculation. Therefore, facing the problem of station similarity calculation between adjacent noise monitoring stations and non-adjacent stations, the temporal dynamic warping DTW algorithm is used to reconstruct from the perspective of time series similarity Spatial Relations. Facing the problem that the noise is additive, and the factors with high correlation to the noise cannot play a role in the spatio-temporal prediction, and the Kalman Kalman filter can adjust and correct the prediction results in combination with other influencing factors, so the present invention adopts the Kalman method Dynamically adjust the prediction results according to the characteristics of the relevant factors.

The specific performance is: the temporal dynamic warping theory is introduced into the diffuse convolution recurrent neural network DCRNN, and a spatio-temporal noise prediction network model composed of the temporal dynamic warping method, the diffuse convolution recurrent neural network, and Kalman filtering is established. The penalty coefficient is used to improve the DTW algorithm, and the spatial relationship of the neural network is reconstructed through the DTW model; the Kalman method is used to dynamically adjust the output noise prediction of the neural network in combination with the traffic flow characteristics. This method first reconstructs the spatial relationship based on time series similarity, and then sends the data of each site into the model for prediction and dynamic correction, so as to realize multi-site noise prediction, long-term noise prediction, and global noise level prediction, and then avoid noise in advance Risk of nuisance and damage to health.

Based on the characteristics of the DTW algorithm and the DCRNN model, the present invention proposes a spatio-temporal noise prediction network model, which makes up for the limitations of constructing spatial relationships based on distance, and can effectively utilize the park location information and the advantages of the two methods to achieve higher-precision noise prediction .

As shown in Figure 1, a DTW-DCRNN-based noise prediction method for chemical industry parks specifically includes the following steps:

Step 1. Collect and process the monitoring site information and vehicle information data in the chemical industry park.

The road gate and several monitoring stations are set up in the chemical industry park for real-time data collection, and the location information data of each monitoring station is recorded synchronously; the road gate and each monitoring station are equipped with sensors; the road gate collects the information data of vehicles entering and leaving, including the license plate number, Traffic flow characteristics such as vehicle entry and exit time, vehicle type, etc.; monitoring stations collect real-time noise and natural environment data in the chemical industrial park;

As shown in Figure 2, the site location data, noise and natural environment data, and vehicle information data entering and leaving will be transmitted to the gateway device, and then interactively transmitted with the database server. When noise prediction is performed, data is extracted in real time, and the 3σ criterion is used for deal with.

The 3σ criterion is to assume that a set of data contains only random errors, calculate the standard deviation, and determine the interval according to the probability. It is believed that any error exceeding this interval is not a random error but a gross error, and the data containing this error should be removed or replaced. , the interval of noise value noise∈(u-3σ _noise , u+3σ _noise ) accounts for about 99.74%, u represents the noise mean, σ _noise is the noise standard deviation, and noise is the noise value. Replace the noise range in 0≤noise<u-3σ _noise (dB) with the mean value.

The original noise sequence collected by the sensor contains sparse mutation points with zero values. After the noise value processed by the above 3σ criterion, the unbiased standard deviation of the sample is reduced, and the mutation value is reduced, which can improve the accuracy of the noise prediction of the neural network.

Step 2. Introduce the penalty coefficient to improve the original time dynamic warping algorithm, and introduce it into the diffusion convolution recurrent neural network, and then combine the Kalman filter to construct the noise prediction network model.

As shown in Figure 3, the principle of the noise prediction network model is as follows: first, based on the extracted data of each monitoring station, the original time dynamic warping algorithm is used to calculate the optimal path, and then the original time dynamic warping algorithm is improved by introducing a penalty coefficient. The temporal dynamic warping algorithm calculates the sequence similarity distance, forms a matrix for each monitoring site, and then constructs an adjacency matrix graph structure; then inputs the adjacency matrix into the diffusion convolution recurrent neural network for preliminary prediction of noise, and finally combines traffic flow characteristics and The Kalman filter dynamically adjusts the noise prediction results to obtain the final noise prediction results; the specific process is as follows:

Step 2.1, introducing a penalty coefficient to improve the original time dynamic warping algorithm, and calculating the sequence similarity distance;

The principle of the improved time dynamic warping algorithm is: obtain the optimal path and the number of common subsequences to calculate the penalty coefficient, and combine the penalty coefficient to calculate the similarity distance between stations.

The main workflow is shown in Figure 4, and the specific process is as follows:

Step 2.1.1, calculate the distance matrix;

Suppose there are two noise sequences S and T, such as formula (1):

The length of the sequence S is o, the length of the sequence T is m, o,m∈Z ⁺ , the traditional Euclidean distance can directly calculate the distance when the peak and trough coincidence degree is high, but it cannot be used for the calculation of the periodic sequence of different phases, so The present invention constructs a distance matrix d[s][t] based on the noise sequence m×o, where the i-th sequence point s _i (i=1,2,...,o) and the j-th sequence point t _j (j= 1,2,...,m) is the value of the matrix (s _i ,t _j ) element, that is, the Euclidean distance: dis(s _i ,t _j )=(s _i -t _j ) ² , the matrix distance The calculation standard is shown in formula (2):

Step 2.1.2, seeking the optimal path of the matrix;

In the distance matrix calculated in 2.1.1 above, the rows of the matrix represent the sequence S, and the columns of the matrix represent the sequence T. Rows and columns correspond to sequence order one by one, but a certain point in S sequence corresponds to multiple points in T sequence, and the optimal path of the matrix is obtained through dynamic regularization. The selection of the path needs to meet the following boundary conditions: starting from the d[o][m] point in the upper right corner of the matrix, find the smallest point among the three lower left points as the next node until the d[0][0] matrix in the lower left corner In addition to ensuring continuity and monotonicity, a point cannot be crossed or missed for matching, only adjacent points can be aligned or matched, and the order of the time series cannot be changed. The optimal path is the path with the minimum cumulative distance that satisfies the above conditions.

The single DTW algorithm to solve the distance is not friendly to the unaligned periodic sequence, and the distance calculation is not ideal. Therefore, it needs to be improved on the basis of DTW algorithm. The improvement idea is to first use the DTW algorithm to generate the distance matrix and find the optimal path: take the point (i, j) as an example to find the minimum value among the three points on the lower left, if the value is the lower left diagonal value, then (i-1, j-1) is the next node; if the value is the left adjacent point, then the next node is (i-1, j); if the value is the lower adjacent point, then (i, j-1) is the next node A node until there are no next nodes. The obtained optimal path provides a verification basis for finding the common subsequence of the noise sequence later.

Step 2.1.3, calculating the number of common subsequences and the length of the optimal path sequence, and then obtaining the weight and penalty coefficient;

The length of the matrix optimal path finally obtained in the calculation step 2.1.2 is the required optimal path sequence length.

The steps for calculating the number of noise common subsequences are shown in Figure 5. First, construct an empty record list (record), loop through the sequences s1 and s2, and use the record list to record the length of the consensus sequence ls1 or ls2 and the sequence position g , extract the common subsequence from the s1 sequence and store it in the s_sum array. After the traversal is completed, the length of s_sum is the number of common subsequences. When calculating the number of common subsequences, set the threshold to 1, that is, subsequences with a length greater than 1 can be counted as common subsequences, and finally count the total number of common subsequences as the required number of common noise subsequences.

See Algorithm 1 for the calculation method:

Use all the common noise subsequence numbers to calculate the penalty coefficient;

First calculate the weight w, the calculation formula is as formula (3):

Among them, subseq is the length of the common noise subsequence, and seq is the sequence length of the optimal path of the matrix;

Then, the calculation process of setting the penalty coefficient α is shown in formula (4), where x represents the number of common subsequences, and w _i represents the weight of the i-th sequence point. The longer the sequence length of the optimal path, the more the number of common noise subsequences , the longer the length, the smaller the penalty coefficient:

Step 2.1.4. Express the optimal path as (r ₁ ,r ₂ ,…,r _seq ), r _i (i=1,2,…,seq) represents the value of the i-th sequence point in the optimal path sequence , and finally the penalty coefficient is multiplied by the sum of values in the optimal path calculated by the original DTW to obtain the improved sequence similarity distance dis _DTW .

In step 2.2, the noise sequence similarity distance between each monitoring site is formed into a matrix and calculated as an adjacency matrix to construct a graph relationship topology. The specific process is:

Step 2.2.1, according to the noise sequence similarity distance information of each monitoring station, calculate the similarity distance matrix M (numerical unit: km) between the stations, the matrix M is as follows:

Among them, a is the number of monitoring stations;

For example, a total of 11 monitoring stations are set up in the chemical industry park in the embodiment of the present invention, and the matrix composed of each station is as follows.

Step 2.2.2, calculate the standard deviation λ of the similarity distance matrix M, M _c is the cth element value in the matrix, N is the number of elements in the M matrix, μ _M is the matrix mean value, and the calculation process is as follows:

的计算过程如下：Step 2.2.3, analyze the similarity between monitoring sites according to the similarity distance matrix M, use the standard deviation λ of all non-infinite numbers in the similarity distance matrix to construct the adjacency matrix M _d , each element in the adjacency matrix M _d

The calculation process is as follows:

越小，设置阈值为0.1，若

＜阈值，则视为两站点间相似度太低，无相互影响关系，视作非邻接站点，在邻接矩阵中权值为零不构成邻接关系。Among them, c represents the serial number of matrix elements, and the larger the similarity distance is, the

The smaller the value, set the threshold to 0.1, if

<Threshold value, it is considered that the similarity between the two sites is too low, there is no mutual influence relationship, and it is regarded as a non-adjacent site, and the weight value of zero in the adjacency matrix does not constitute an adjacency relationship.

The above method steps meet the similarity calculation requirements of periodic noise time series. Using this method to reconstruct the spatial relationship can lay the foundation for providing a more complete graph representation relationship for the subsequent graph convolutional neural network work.

Step 2.3, introduce the graph convolutional neural network GCN to construct the adjacency matrix graph structure, input the adjacency matrix graph structure into the diffusion convolution recurrent neural network for noise prediction, and obtain the preliminary prediction result;

First understand convolution from the perspective of signal processing. The word "Convolve" originally means flip, and is translated as "volume" in convolution. "Product" itself is an operation method, meaning "multiplication". Noise can be classified as discrete noise Or continuous noise. Convolution generally refers to continuous signals. The sound monitoring method is realized through sensors. The noise collected by the sensor is a discrete value at an interval of 30s. Therefore, only the discrete noise convolution sum is discussed here.

n表示位移位数，l表示离散时间，离散噪声x[n]可用式(9)表示：Define the pulse signal as:

n represents the number of bits shifted, l represents the discrete time, and the discrete noise x[n] can be expressed by formula (9):

where x[l] represents the discrete noise at time l;

The discrete noise convolution process can be understood as first flipping the noise sequence, then shifting, and finally multiplying and summing. It means that the response of the system to the input discrete noise sequence x[n] is characterized by the unit impulse response h[n] of the system, and the convolution and y _dis [n] of the discrete noise convolution are formula (10).

Among them, during convolution, first invert the unit impulse response to obtain h[-l], and then shift n to obtain the displaced function h[n-l];

However, the definition of convolution in Convolutional Neural Network (CNN) lacks the process of flipping compared with the above-mentioned convolution principle, and h[n+l] can be obtained by directly shifting n. The convolution and y of the convolutional neural network _cnn [n] is expressed as formula (11),

It is not a complete convolution sum but a cross-correlation in signal processing, so the convolution kernel is also a filter. The fundamental purpose of using convolution in convolutional neural networks is to weight summation and extract features, and flipping is not necessary.

The noise sequences of multiple monitoring sites can form a topological graph, which is irregular and not translation invariant. Therefore, it is necessary to introduce a graph convolutional neural network (GCN) to solve the graph relationship problem. In the embodiment of the present invention, eleven monitoring stations are regarded as eleven nodes, and each station has its own characteristics, and all nodes form a 11×11-dimensional adjacency matrix A, assuming that the length of the noise sequence is Y, then the characteristics of the nodes are The 11*Y-dimensional feature matrix X, X and A matrix are the inputs of the graph convolutional network. The propagation mode between the network layers of the graph convolutional neural network GCN is formula (12), and L represents the Lth network layer:

为度矩阵，其构成为：

I表示邻接矩阵A第I行，J表示邻接矩阵A第J列，

表示保留站点本身的特征信息，I₁₁表示11维的单位矩阵，λ为站点噪声特征权重，λ＝1表示该站点的噪声特征与相邻站点的噪声特征一样重要，H为当前层提取到的特征，如果是输入层则X＝H。Among them, σ is a nonlinear activation function; W ^L is a trainable weight matrix,

is a degree matrix, and its composition is:

I represents row I of adjacency matrix A, J represents column J of adjacency matrix A,

Indicates that the feature information of the site itself is preserved, I ₁₁ indicates the 11-dimensional identity matrix, λ is the weight of the site noise feature, λ=1 indicates that the noise feature of the site is as important as the noise feature of the adjacent site, and H is extracted from the current layer feature, if it is an input layer then X=H.

If the CNN model is not trained, the extracted features are very limited. Compared with the CNN model, the features extracted by GCN using random initialization parameters are extremely good. Using GCN to extract spatial relationships is the most preferred method in current research.

In order to formalize the spatiotemporal noise sequence prediction problem, on the basis of the GCN graph convolutional neural network, the present invention uses a diffusion convolutional recurrent neural network to model the spatiotemporal relationship of the noise sequence. By associating noise monitoring sites with diffusion processes, modeling the spatial relationships that exist between sites can also capture the dynamic, stochastic nature of noise. The DCRNN model structure of the diffusion convolutional recurrent neural network is shown in Figure 6. DCRNN is a recurrent neural network based on diffusion graph convolution. The original input of the model is the spatiotemporal noise based on the combination of the adjacency matrix of the physical distance of the monitoring stations and the time series noise data. Data structure, the invention proposes an improved DTW algorithm for calculating the similarity distance of the sequence, and reconstructs the spatio-temporal relationship between sites based on the similarity distance. The advantage of this model is that the diffusion convolution can take into account the walk relationship between noises, and an encoder is added to the historical noise sequence, and the real value or predicted value of the previous moment is delayed by one time step as the current decoder. The input of the current time noise prediction. Where ReLU is the activation function, if the input is positive, it will output directly, otherwise, it will output zero. The stationary distribution of the diffusion process can be expressed as a weighted combination of infinite random walks on the graph, and the diffusion process can be expressed as Equation (13) and calculated in closed form:

是出度矩阵的逆矩阵，β∈[0,1]表示重启概率，其中k是扩散程度，ε表示从节点扩散的可能性，DCRNN模型中扩散过程是双向的，可以考虑噪声监测站点双边的环境影响。W is the node similarity matrix,

is the inverse matrix of the out-degree matrix, β∈[0,1] represents the restart probability, where k is the degree of diffusion, ε represents the possibility of spreading from the node, the diffusion process in the DCRNN model is bidirectional, and the bilateral noise monitoring site can be considered environmental impact.

与滤波器f_θ(θ为滤波器参数)双向扩散卷积运算可以定义为式(14)：Therefore, in terms of spatial relationship, based on the graph signal feature matrix

The bidirectional diffusion convolution operation with the filter f _θ (θ is the filter parameter) can be defined as formula (14):

与

分别表示扩散与逆扩散的过度矩阵，定义卷积运算后，构建带有映射关系的扩散卷积层(P维到Q维)如式(15)：Among them, G represents the graph G, p represents the pth feature dimension, P represents the total number of feature dimensions, K represents the finite K-step truncation of the diffusion process, G represents the diffusion convolution on the graph G, X:, _p represents all node pairs The convolution operation of the p-th feature, θ _{k, 1} represents the convolution kernel parameter for out-degree calculation, θ _{k, 2} represents the convolution kernel parameter for in-degree calculation,

and

Represent the transition matrix of diffusion and inverse diffusion respectively. After defining the convolution operation, construct the diffusion convolution layer (P dimension to Q dimension) with mapping relationship as formula (15):

为输入，当前层提取到的特征

为输出，

表示滤波器(Θ为参数张量)，σ为激活函数。基于双向扩散卷积的空间关系构建完成。Among them, q represents the qth feature, Q represents the total number of mapping features, H:, _q represents the diffusion and convolution operation of all nodes on the qth feature, and the feature matrix

As input, the features extracted by the current layer

for the output,

Represents a filter (Θ is a parameter tensor), and σ is an activation function. The spatial relationship based on bidirectional diffusion convolution is completed.

In terms of time relationship, the combination of GRU and diffusion convolution is used to construct the diffusion convolution gated recursive unit (DCGRU). The essential principle is to replace the matrix multiplication in GRU with diffusion convolution. Convolution principle, DCGRU can be expressed as formula (16):

G represents the diffuse convolution of graph G, r ^(t) represents the reset gate at time t, Θ _{r G} represents the parameter tensor of the reset gate, X ^(t) represents the input at time t, H ^{(t-1 )} represents the output at time t-1, b _r represents the bias vector of the reset gate; u ^(t) represents the update gate at time t, Θ _{u G} represents the parameter tensor of the update gate, and b _u represents the update gate’s Bias vector; C ^(t) represents the hidden state of the next cell at time t, Θ _{C G} represents the parameter tensor of the hidden state, b _C represents the bias vector of the hidden state; H ^(t) represents the output.

Through spatio-temporal modeling, two-way propagation is used to capture the spatio-temporal correlation of noise sequences to the greatest extent, but the current work is based on the spatial correlation of distance, and the spatial features constructed based on physical distance are difficult to fully represent the correlation of noise sequences between sites, so A better approach to this problem is also needed. To improve the prediction accuracy of spatial relationship, it is necessary to go deep into the spatial relationship to solve the problem.

Step 2.4, extract the traffic flow characteristics at the gate corresponding to the current time in real time, and use the Kalman filtering method to dynamically adjust the DCRNN noise prediction value.

Kalman filtering uses an estimate of the previous state to make a prediction for the current state. Finally, the predicted value obtained in the prediction stage is corrected by using the observed value of the current state to obtain a new estimated value that is closer to the real value. The specific implementation is to calculate the Kalman gain of noise and drift through the prior estimation and prediction of the covariance matrix of the normalized traffic flow characteristics and noise sequence, and use the Kalman gain correction to update the covariance matrix, and then correct and predict the current moment noise value.

Step 3, train the constructed noise prediction network model, and output the trained model. The specific process is:

Extract the historical data of each monitoring station and use the improved DTW algorithm to reconstruct the spatio-temporal relationship of the data, train the noise prediction network model DTW-DCRNN, select 70% of the data set as the training set, 10% as the verification set, and the last 20% as the test set . In addition, the training results are not recorded in the first 60 Epochs, the training results are recorded and model parameters are output every 10 Epochs in the 60-100 Epochs, and the training results are recorded and model parameters are output every 5 Epochs after 100 Epochs . Finally, training is stopped early based on the measured validation loss to ensure that the model is caught when it is close to overfitting.

Step 4. Collect monitoring site information in real time and perform real-time prediction of noise based on the trained model.

The present invention first establishes the similarity distance matrix based on the improved DTW algorithm according to the positions of each site, abandons the traditional method of constructing the matrix based on the physical distance, and constructs the similarity distance matrix as the representation form of the adjacency matrix, and uses the similarity distance matrix based on the similarity distance The spatio-temporal data set is fed into the DCRNN model to obtain the DTW-DCRNN model based on the similarity distance spatial relationship. Finally, combined with the real-time traffic flow data, the Kalman filter is used to dynamically update the prediction output of the DCRNN.

Using the similarity distance matrix to construct the adjacency matrix relationship to represent the spatial distance relationship can better compensate for the limitations of the original physical distance method. The prediction results adjusted by the Kalman filter are also closer to the real value, and have compensation properties. The 3σ noise processing method can effectively avoid the large deviation of the model training prediction results.

In order to prove the feasibility and superiority of the present invention, the following experiments were carried out.

Experiment 1: Comparative experiment before and after DTW algorithm improvement

Figure 7(a) shows an example of the adjacency matrix based on the original DTW algorithm. The improved DTW algorithm as shown in Figure 7(b) increases the correlation between monitoring site 1 and monitoring site 3 compared with the original method.

The experiment of the present invention adopts mean absolute error MAE, mean absolute percentage error MAPE and root mean square error RMSE as evaluation indexes. The DTW algorithm and the improved DTW algorithm are combined with the DCRNN model for pre-training, and the average prediction results of twelve time steps are obtained. The indicators are shown in Table 1. On the results of the pre-training model, the MAE noise is reduced by 0.06 decibels, and the accuracy is increased by 2.8 %, MAPE is reduced by 0.11%, accuracy is increased by 1.9%, RMSE is increased by 0.02dB, and accuracy is decreased by 0.5%. The improved MAE and MAPE are both lower than the original method, and the RMSE index is slightly higher than the original method, that is, the application of the improved method data outliers, the prediction accuracy of outliers is lower than the original method, but the prediction of the overall index new DTW algorithm The accuracy is higher than the original DTW algorithm.

Table 1 Comparison of DTW algorithm before and after improvement

Draw the indicator diagram according to the time step as shown in Figure 8. Further observe the improved DTW algorithm and the original algorithm. The abscissa is the time step, and the ordinate on the right side of the image is the MAPE index. It can be seen from the figure that the improvement starts from the fourth time step The average absolute percentage error of the DWT algorithm is smaller, and the new method performs better on the MAPE index. The ordinate on the left side of the image is the RMSE and MAE indicators, and the unit is decibel (dB). It can be seen from the figure that the root mean square error of the improved DWT algorithm is smaller from the sixth time step, and the new method is better than the RMSE indicator. The performance is better, the average absolute error of the improved DWT algorithm from the fourth time step is smaller, and the new method performs better on the MAE index.

In summary, the improved DTW algorithm not only improves the overall accuracy of prediction, but also does not increase the time cost of algorithm calculation, and can dig deeper sequence correlations, so it can be innovatively applied to graph convolution work to improve prediction accuracy.

Experiment 2: Comparative experiments between the prediction model proposed by the present invention and other models

In order to further verify the prediction performance of the method of the present invention, the three evaluation indicators of RMSE, MAE and MAPE are used to analyze the HA historical average prediction model, VAR vector autoregressive model, STGCN space-time convolutional neural network, DCRNN diffusion convolutional recurrent neural network, DTW-STGCN , DTW-DCRNN(N) and other methods or models for evaluation, where the DTW-DCRNN(N) method is a method that does not use Kalman filtering and dynamic adjustment of traffic flow data. The above experiments are all based on 12 time step predictions, and the average value of the prediction results of 12 time steps is used as an index to measure the effect of the long-term prediction of the model. According to Table 2, the prediction results of the present invention are better than other prediction methods.

HA: Due to the strong periodicity of noise and many mutation points, the prediction accuracy of this model is low and the training speed is fast.

STGCN: Due to the randomness of the noise sequence, the accuracy of the noise prediction is low.

DCRNN: The effect of this model is stronger than that of STGCN, and the time series noise data has a strong dependence on time, which reflects that GRU has greater advantages in time series modeling than CNN.

VAR: Due to the strong periodicity of the noise sequence, the prediction result of the vector autoregressive model is better than that of STGCN, which shows that the neural network does not perform better under any circumstances.

DTW-STGCN: The DTW algorithm is applied to the STGCN model to construct the DTW-STGCN method, and the prediction performance is still stronger than the STGCN model. It is also verified from the side that it is meaningful to consider the spatial relationship from the perspective of time series similarity.

DTW-DCRNN(N): Compared with DTW-DCRNN(N), the method of the present invention adds Kalman filtering, which has the effect of correcting and adjusting the data, and its effect is the best. Compared with the original DCRNN model, the RMSE and MAE accuracy of the method of the present invention are increased by 11.1% and 5%, respectively, and the MAPE is increased by 5.26%.

Table 2 Baseline model prediction results

In addition, in the experiment, it is also found that the method of the present invention performs very well in long-term prediction. The specific data are shown in Table 3. It can be seen that the prediction accuracy of the method of the present invention in the first time step is lower than that of DCRNN, but in the following In 11 time steps, the performance tends to be more stable, and the accuracy is higher.

Table 3 Analysis of long-term forecast results

In order to facilitate the observation of the specific prediction trend, the above table is visualized as an image. Figure 9 is the mean absolute error MAE and root mean square error RMSE curve. The ordinate on the right side of the image is the RMSE index, and the unit is decibel (dB). It can be seen that the prediction of the 2nd, 3rd, and 4th time steps is more accurate. The ordinate on the left side of the image is the MAE index, and the unit is decibel (dB). When the first time is not considered, the errors of the two methods all increase with the increase of the time step, but the error of the DTW-DCRNN method of the present invention Smaller and more precise. Fig. 10 is the MAPE curve graph of the average absolute percentage error, the trend is similar to the MAE index, but the average absolute percentage error of the DTW-DCRNN model of the present invention is smaller. The above three indicators start from the second time step (including the second time step), and the method of the present invention performs better.

In summary, the improved method of the present invention not only improves the overall accuracy of prediction, but also has more advantages in the long-term noise time series prediction work than the original method, and can be used for the spatial-temporal prediction work. And related research provides a new way of thinking and method.

Of course, the above descriptions are not intended to limit the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or replacements made by those skilled in the art within the scope of the present invention shall also belong to the present invention. protection scope of the invention.

Claims (4)

1. A DTW-DCRNN-based chemical industry park noise prediction method is characterized in that an improved time dynamic warping algorithm is adopted, a space relation is reconstructed through a graph structure of a monitoring station, and long-term space-time prediction of the chemical industry park noise is achieved, and the method specifically comprises the following steps:

step 1, collecting and processing monitoring site information and vehicle information data in a chemical industry park;

step 2, introducing a penalty coefficient to improve an original time dynamic warping algorithm, introducing the penalty coefficient into a diffusion convolution recurrent neural network, and constructing a noise prediction network model by combining Kalman filtering;

step 3, training the constructed noise prediction network model and outputting the trained model;

and 4, acquiring information of the monitoring station in real time and predicting noise in real time based on the trained model.

2. The DTW-DCRNN-based chemical industry park noise prediction method according to claim 1, wherein in step 1, a gateway and a plurality of monitoring sites are arranged in the chemical industry park for collecting data in real time and synchronously recording position information data of each monitoring site; sensors are arranged at the gateway and each monitoring station; the gateway collects information data of vehicles entering and leaving, including license plate number, vehicle entering and leaving time and vehicle flow characteristics of vehicle type; the monitoring station collects noise and natural environment data in a chemical industry park in real time;

and the station position data, the noise and natural environment data and the in-out vehicle information data are transmitted to the gateway equipment and then are interactively transmitted with the database server, and when noise prediction is carried out, data extraction is carried out in real time and 3 sigma criterion is adopted for processing.

3. The DTW-DCRNN-based chemical industry park noise prediction method according to claim 1, wherein the specific process of the step 2 is as follows:

step 2.1, introducing a penalty coefficient to improve an original time dynamic warping algorithm, and calculating a sequence similarity distance;

the principle of the improved time dynamic regularization algorithm is as follows: obtaining the number of the optimal paths and the number of the public subsequences to calculate a penalty coefficient, and calculating the similarity distance between stations by combining the penalty coefficient; the specific process is as follows:

step 2.1.1, calculating a distance matrix;

two noise sequences S and T are assumed, as in equation (1):

the length of the sequence S is o, the length of the sequence T is m, o, m belongs to Z ⁺ Constructing a distance matrix d [ s ] based on the noise sequence mxo][t]Wherein the ith sequence point s _i (i =1,2, \8230;, o) and the jth sequence point t _j (j =1,2, \8230;, m) is a matrix(s) _i ,t _j ) The value of the element(s) is,euclidean distance: dis(s) _i ,t _j )＝(s _i -t _j ) ² The matrix distance calculation criterion is shown in equation (2):

step 2.1.2, solving a matrix optimal path;

the optimal path is a path with the minimum accumulation distance meeting the lower boundary condition, and the lower boundary condition is as follows: starting from the d [ o ] [ m ] point at the upper right corner of the matrix, finding the minimum point in the three lower left points as the next node until the d [0] [0] point at the lower left corner is finished, ensuring continuity and monotonicity, not crossing or omitting a certain point for matching, only aligning or matching with adjacent points, and unchangeable time sequence;

step 2.1.3, calculating the number of public subsequences and the optimal path sequence length, and further obtaining weight to obtain a penalty coefficient;

wherein, the optimal path sequence length is the length of the optimal path of the matrix finally obtained in the step 2.1.2;

the calculation process of the number of the common subsequences is as follows: firstly, constructing a recording empty list record, circularly traversing sequences s1 and s2, recording the length ls1 or ls2 of a common sequence and the sequence position g of the common sequence by using the record list after the common sequence appears, and extracting the common subsequence from the s1 sequence and storing the common subsequence in an s _ sum array; the length of s _ sum after traversal is the number of the public subsequences; when calculating the number of the public subsequences, setting a threshold value as 1, calculating the subsequences with the length greater than 1 as the public subsequences, and finally counting the total number of the public subsequences to obtain the required number of the public noise subsequences;

the calculation formula of the weight w is as follows (3):

wherein, the subseq is the length of the public noise subsequence, and the seq is the sequence length of the optimal path of the matrix;

the calculation process for setting the penalty coefficient alpha is shown as formula (4), wherein x represents the number of public subsequences, and w _i Represents the weight of the ith sequence point,

step 2.1.4, express the optimal path as (r) ₁ ,r ₂ ,…,r _seq )，r _i (i =1,2, \8230;, seq) represents the value of the ith sequence point in the optimal path sequence, and finally, the sum of the values in the optimal path calculated by the original DTW is multiplied by a penalty coefficient to obtain the improved sequence similarity distance dis _DTW ；

Step 2.2, forming a matrix by the noise sequence similarity distance among all the monitored sites, calculating the matrix as an adjacent matrix, and constructing a graph relation topological structure; the specific process is as follows:

step 2.2.1, according to the noise sequence similarity distance information of each monitored site, calculating a similarity distance matrix M between the sites, wherein the matrix M is as follows:

wherein a is the number of monitoring stations;

step 2.2.2, calculating the standard deviation lambda, M of the similarity distance matrix M _c Is the c-th element value in the matrix, N is the number of elements in the M matrix, mu _M For matrix mean, the calculation process is as follows:

step 2.2.3, analyzing the similarity between the monitoring stations according to the similar distance matrix M, and constructing an adjacent matrix M by using the standard deviation lambda of all non-infinite numbers in the similar distance matrix _d Of a contiguous matrix M _d Each element in (1)

The calculation process of (c) is as follows:

wherein c represents the sequence number of the matrix element, and the larger the similarity distance is, the larger the similarity distance is

The smaller the threshold is set to 0.1, if

If the similarity between the two sites is too low, no mutual influence relationship exists, the two sites are regarded as non-adjacent sites, and the adjacent relationship is not formed when the weight in the adjacent matrix is zero;

step 2.3, introducing a graph convolution neural network GCN to construct an adjacent matrix graph structure, and inputting the adjacent matrix graph structure into a diffusion convolution recurrent neural network for noise prediction to obtain a preliminary prediction result; the specific process is as follows:

the propagation mode between the GCN network layers of the graph convolution neural network is shown as formula (12), wherein L represents the Lth network layer:

wherein σ is a nonlinear activation function; w ^L Is a matrix of trainable weights to produce a desired weight,

a degree matrix, which is composed of:

i denotes the I th row of the adjacency matrix A, J denotes the J th column of the adjacency matrix A,

the characteristic information of the station is kept, H is the characteristic extracted by the current layer, and if the characteristic is the input layer, X = H;

on the basis of the GCN graph convolution neural network, a time-space relation of a noise sequence is modeled by using a diffusion convolution recursive neural network;

the smooth distribution of the diffusion process is represented as a weighted combination of infinite random walk on the graph, the diffusion process is represented by equation (13), and is calculated in closed form:

wherein W is a node similarity matrix,

is the inverse of the rectangular output matrix, with beta ∈ [0,1 ]]Representing restart probability, k is diffusion degree, epsilon represents the possibility of diffusion from nodes, and the diffusion process in the DCRNN model is bidirectional;

thus, in the spatial relationship, based on the map signal feature matrix

And filter f _θ The bi-diffusion convolution operation is defined as equation (14):

where θ is the filter parameter, G represents graph G, P represents the pth feature dimension, P represents the total number of feature dimensions, K represents the finite K-step truncation of the diffusion process, and G represents the graph GDiffusion convolution, X:, _p representing the convolution of all nodes with the p-th feature, θ _k,1 A convolution kernel parameter, θ, representing a degree calculation _k,2 Represents the parameters of the convolution kernel of the in-degree calculation,

and with

Respectively representing transition matrixes of diffusion and inverse diffusion, and constructing a diffusion convolution layer with a mapping relation after defining convolution operation as shown in a formula (15):

wherein Q represents the qth feature, Q represents the total number of mapped features, H, _q representing the q characteristic diffusion convolution operation of all the node pairs and a characteristic matrix

Features extracted from the current layer as input

In order to be output, the output is,

representing a filter, wherein theta is a parameter tensor, and sigma is an activation function;

in the time relation, a diffusion convolution gating recursion unit DCGRU is constructed by combining GRU and diffusion convolution, and is expressed as an expression (16):

∑ G represents a diffusion convolution of graph G, r is a radical of hydrogen ^(t) Indicating the reset gate at time t, [ theta ] _r★G Door with indication of resetTensor of parameter (c), X ^(t) Indicating input at time t, H ^(t-1) Representing the output at time t-1, b _r A bias vector representing a reset gate; u. of ^(t) Indicating the update Gate at time t, Θ _u★G Tensor of parameters representing the update gate, b _u A bias vector representing an update gate; c ^(t) Indicates the hidden state of the next cell at time t, [ theta ] _C★G Tensor of parameters representing hidden states, b _C A bias vector representing a hidden state; h ^(t) Represents the output at time t;

step 2.4, extracting the traffic flow characteristics at the road gate corresponding to the current time in real time, and dynamically adjusting the DCRNN noise predicted value by using a Kalman filtering method;

kalman filtering uses the estimate of the previous state to make a prediction of the current state; finally, correcting the predicted value obtained in the prediction stage by utilizing the observed value of the current state to obtain a new estimated value which is closer to the true value; the method specifically comprises the steps of calculating Kalman gains of noise and drift by using normalized traffic flow characteristics and noise sequences through priori estimation and covariance matrix prediction, updating a covariance matrix by using Kalman gain correction, and then correcting and predicting a noise value at the current moment.

4. The DTW-DCRNN-based chemical industry park noise prediction method according to claim 1, wherein the specific process of the step 3 is as follows:

extracting historical data of each monitoring station for reconstructing a spatio-temporal relationship of data by an improved DTW algorithm, training a noise prediction network model DTW-DCRNN, selecting 70% of a data set as a training set, 10% as a verification set and finally 20% as a test set; in addition, the training results are not recorded in the first 60 epochs, the training results are recorded once every 10 epochs in the 60-100 epochs and the model parameters are output, and the training results are recorded once every 5 epochs after 100 epochs and the model parameters are output; and finally stopping training in advance according to the measured verification loss to ensure that the model is captured when the model is about to be overfitting.

Images