Short-time prediction method and device for port tsp concentration and storage medium

Abstract

This invention discloses a method, apparatus, and storage medium for short-term prediction of TSP concentration in ports. The short-term prediction method includes: acquiring historical monitoring data of TSP concentration and cleaning the concentration data; decomposing the cleaned TSP concentration data into N sets of time series data, wherein the first N-1 sets of time series data have periodic variation characteristics, and the Nth set of time series data has random fluctuation characteristics; constructing N-1 periodic time series prediction models and one random time series prediction model for each of the N sets of time series data; and finally, summing the prediction results of the N time series prediction models to obtain the predicted port TSP concentration. This invention can solve the problem of decreased prediction accuracy caused by the non-stationarity and randomness of TSP concentration data over time, improve the reliability of short-term prediction results of port TSP concentration, and is of great significance for the prevention and control of air pollution in ports.

Description

Technical Field

[0001] This invention belongs to the field of air pollution control technology in environmental engineering, and specifically relates to a calculation method for predicting TSP concentration in ports. Background Technology

[0002] Ports are crucial nodes in the global logistics supply chain, supporting the rapid development of regional trade and economy. However, the large amounts of particulate matter generated during the loading, unloading, and storage of dry bulk cargo such as coal and ore have a significant negative impact on the air quality of surrounding areas. Port workers exposed to high concentrations of particulate matter for extended periods may not only have an increased probability of respiratory illnesses but also a higher risk of heart disease. Surveys of various particulate matter concentrations in the ambient air around ports show that total suspended particulate matter (TSP) concentrations generally exceed standards. In particular, some older port areas contribute more than 50% of the total TSP pollution in their respective cities, constituting a serious pollution problem and becoming one of the main challenges restricting the green and sustainable development of ports.

[0003] Therefore, acquiring the time series of TSP concentrations based on the TSP monitoring system deployed at the port, and accurately and efficiently predicting the distribution of air pollutant concentrations in the port to determine targeted TSP pollution control measures, is one of the primary tasks in building a smart and green world-class port. Previous research by the inventors revealed that the time series changes in port TSP concentrations exhibit complex nonlinear chaotic characteristics, making it difficult to accurately predict future short-term concentration data directly using common time series prediction models (such as the ARIMA model). Therefore, a method is developed that can decompose TSP concentrations into periodic stationary data and stochastic nonstationary data based on the non-stationary fluctuations in port TSP concentration data, and construct prediction models for each, significantly improving the accuracy and reliability of short-term predictions of port TSP concentrations. Summary of the Invention

[0004] The technical problem this invention aims to solve is the non-stationary and random characteristics of TSP concentration data over time, which leads to a decrease in prediction accuracy. The goal is to improve the reliability of short-term TSP concentration prediction results in ports, support port air pollution prevention and control, and achieve smart, green, and sustainable development of ports.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0006] This invention first provides a short-term prediction method for TSP concentration in ports, comprising:

[0007] Obtain historical monitoring data on TSP concentrations at the port;

[0008] The acquired historical monitoring data is decomposed into N sets of time series data; among which the first N-1 sets of time series data Fi (t) exhibits periodic variation characteristics, and the Nth set of time series data R N (t) exhibits random fluctuation characteristics;

[0009] For N sets of time series data, N short-term prediction models are constructed respectively; where F is the time series data with periodic variation characteristics of the first N-1 sets. i (t) Construct N-1 periodic time series prediction models P i F (t), using the time series data R of the Nth group of random fluctuation characteristics N (t) Construct a stochastic time series prediction model

[0010] By summing the prediction results of N-1 periodic time series prediction models and 1 stochastic time series prediction model, the predicted concentration of TSP at the port is finally obtained.

[0011] The present invention also provides a short-term prediction device for port TSP concentration, including a processor and a memory; the memory stores a program or instructions, which are loaded and executed by the processor to implement the steps of the short-term prediction method for port TSP concentration.

[0012] The present invention also provides a computer-readable storage medium storing a program or instructions that, when executed by a processor, implement the steps of short-term prediction of the TSP concentration at the port.

[0013] The technical solution adopted in this invention has the following beneficial effects:

[0014] (1) Compared with traditional methods, this invention constructs a decomposition method when analyzing port TSP concentration, which decomposes the time series characteristics of TSP concentration into periodic stationary data series and random non-stationary data series, which can obtain rich TSP concentration fluctuation information and reduce the complexity of predicting non-stationary time series data of port TSP concentration.

[0015] (2) Based on data decomposition, this invention constructs prediction models according to the characteristics of periodic stationary data and random non-stationary data, which can take into account the advantages of both periodic time series and random time series models, and improve the accuracy and reliability of short-term prediction of port TSP concentration. Attached Figure Description

[0016] Figure 1 Basic technical flowchart of the present invention;

[0017] Figure 2 This is historical monitoring data of TSP at a certain port;

[0018] Figure 3 To Figure 2 Cleaned historical TSP monitoring data of hourly TSP concentrations in China;

[0019] Figure 4 The results are the decomposition of TSP hourly concentration data;

[0020] Figure 5 A comparison chart of the predicted values ​​of the prediction model constructed for the time series component F1 and the measured data;

[0021] Figure 6 A comparison chart of the predicted values ​​of the prediction model constructed for the time series component F2 and the measured data;

[0022] Figure 7 A comparison chart of the predicted values ​​of the prediction model constructed for time series component F3 and the measured data;

[0023] Figure 8 A comparison chart of the predicted values ​​of the prediction model constructed for time series component F4 and the measured data;

[0024] Figure 9 A comparison chart of the predicted values ​​of the prediction model constructed for time series component F5 and the measured data;

[0025] Figure 10 A comparison chart of measured data and predicted results. Detailed Implementation

[0026] The following is a further explanation of this plan, in conjunction with the accompanying drawings:

[0027] This embodiment provides a short-term prediction method for TSP concentration in ports, including:

[0028] A. Acquisition of TSP concentration data at ports:

[0029] Port TSP concentration monitoring data can be obtained from the online monitoring system deployed at the port. At least one week's worth of data is required, which, in chronological order of monitoring time, forms M data points X with time-series characteristics. M =(x1,x2,…x M ).

[0030] B. Port TSP concentration data cleaning:

[0031] The method for cleaning port TSP concentration monitoring data is as follows:

[0032] B.1 Abnormal Data Removal. If there are abnormal data in M ​​time series data X... M In the middle, the monitoring data x at the m-th time point m (m∈M) satisfies x m <Amin or x m >A max Then x m As abnormal data to be cleared, A min and A max The calculation method is as follows:

[0033]

[0034] B.2 Data Completion. If monitoring data x at time point m is missing due to reasons such as device disconnection, document transmission loss, or abnormal data deletion, etc. m (m∈M), then the method for completion is:

[0035]

[0036] B.3 Time Interval Adjustment. The port TSP detectors report concentration data at Q intervals per hour (typically every 1 to 5 minutes), while the minimum time interval for TSP concentration data analysis in ambient air quality standards is 1 hour. Therefore, the acquired concentration data needs to be adjusted to 1-hour intervals. The TSP concentration x at time t is obtained by arithmetically averaging the q TSP concentration data points between time t-1 and t. t for:

[0037]

[0038] B.4 Data sequence after cleaning. TSP concentration data from M ports X M =(x1,x2,…x M After the above three cleaning steps, T TSP concentration data sequences X, with time series characteristics and spaced at 1-hour intervals, can be obtained. T =(x1,x2,…x T ) is used for prediction.

[0039] C. Decomposition of TSP concentration data at ports:

[0040] T hours of TSP concentration data sequence X T =(x1,x2,…x T The decomposition resulted in N sets of TSP concentration data sequences. The method is as follows:

[0041] C.1 In the TSP concentration data sequence X T =(x1,x2,…x T Find all local maxima. and local minimum point Thus, A local maxima are obtained. and corresponding time Concentration data sequence and B local minimum points and corresponding time Composition of time series

[0042] C.2 are time series respectively and Construct the upper envelope function f t max (A) and the lower envelope function f t min (B) is calculated as follows:

[0043]

[0044] C.3 Determine the decomposition variables. Calculate the envelope function f. t max (A) and f t min The calculation methods for the average value S1(t) and the decomposition of component H1(t) in (B) are as follows:

[0045]

[0046] H1(t)=X T -S1(t)

[0047] C.4 If the data sequence of TSP concentration data component H1(t) contains a local maximum point The quantity A(H1), local minimum point The quantity B(H1) and x in the data sequence t When the following relationship exists between the number of (H1) = 0, C(H1),

[0048]

[0049] H1(t) is the first component F1(t) of the data sequence, that is,

[0050]

[0051] Otherwise, replace the TSP concentration data sequence X in step C.1 with the time series data component H1(t). T =(x1,x2,…x T Repeat steps C.1 to C.3 until k times. The constraints are satisfied to obtain F1(t) and the remaining concentration data sequence R1(t).

[0052]

[0053] C.5 Repeat steps C.1 to C.4 until N-1 components are obtained. The Nth component satisfies two conditions: Or R N (t) cannot be further decomposed. At this point,

[0054]

[0055] D. Calculation of the parameters of the upper and lower envelope functions:

[0056] D.1 Envelope function f t max (A) The parameter is calculated as follows:

[0057]

[0058] D.2 Envelope function f t min (B) The parameter is calculated as follows:

[0059]

[0060] E. Construction of periodic short-term forecasting models:

[0061] Original TSP concentration time series X T Decomposition of periodically changing data sequence F i The steps for constructing the prediction model for (t) are as follows:

[0062] E.1 Model Input Data Preparation: Using the i-th TSP concentration data sequence F i Taking (t) as an example, Transform it into a two-dimensional matrix with lag time p as the training sample F i train (t) and one-dimensional data Y i O As a tag, the conversion method is as follows:

[0063]

[0064]

[0065] E.2 Data Convolution:

[0066] Let F be the input data for the s-th convolutional layer. i s (t), and when s=1 F i 1 (t)=F i train (t). Let the width of the convolution window in the s-th layer be e. s And the number of neurons is C sThen the weight matrix of the j-th neuron in the s-th layer is for:

[0067]

[0068] Let the bias matrix of the j-th neuron in the s-th layer be... The number of variables is

[0069] Therefore, the input value of the j-th neuron in the s-th layer is:

[0070]

[0071] Then, the output value of the j-th neuron in the s-th layer is:

[0072]

[0073] The output of the convolutional layer is:

[0074]

[0075] E.3 Data Flattening Processing:

[0076] The output of the convolutional layer is used as the input data Y. i s The shape was flattened.

[0077]

[0078] E.4 Data Fully Connected Layer Output:

[0079] Assume the d-th fully connected layer has C... d There are neurons, and the weight matrix is:

[0080]

[0081] The deviation term matrix of the d-th layer is:

[0082]

[0083] Then in the d-th layer, the k-th (k∈[1,C) d The generalized input of ]) neurons is

[0084]

[0085] The output of this neuron is then:

[0086]

[0087] E.5 Prediction Model Parameter Optimization:

[0088] Summarizing the outputs of each neuron yields the prediction result at time t+1 using the initial parameter values.

[0089] The weight matrix and bias term matrix above are all initial random values. To reduce the error between the predicted value and the training value, backpropagation will be used to calculate the error gradient and update the weights and bias terms of the neural network. Through continuous iterative updates, the prediction result Y will be obtained. d With Y i O Satisfy constraints The parameters at that time are the optimal parameter values.

[0090] E.6 Periodic Time Series Model Prediction Results:

[0091] Based on the optimal parameters determined in the above steps, input data F is used. i (t) can be used to predict the TSP concentration data at time t+1 as F. i (t+1).

[0092] F. Construction of stochastic short-term prediction models:

[0093] Original TSP concentration time series X T Random fluctuation data sequence R decomposed in the middle N The steps for constructing the prediction model for (t) are as follows:

[0094] F.1 Historical TSP Concentration Data Acquisition:

[0095] Based on the time series data R of the random fluctuation characteristics decomposed in step four. N (t), where R is the concentration data of the time period t to be predicted in date l. N (l,t). Then, the actual concentration data R for the hour λ preceding the predicted time period t on the current date l. N (l,t-λ) and historical TSP concentration data for the current time period of each day within the previous η days before the current date. N (l-η,t). Generally, it is recommended that λ be 6 hours and η be 7 days.

[0096] F.2 Stochastic Time Series Model Prediction Results:

[0097] For the concentration data R that needs to be predicted at time t+1 N (t+1), the prediction result is:

[0098]

[0099] in, The calculation method is as follows:

[0100]

[0101] G. Short-term prediction results of TSP concentration at the port:

[0102] Based on the obtained M data X with time series characteristics M =(x1,x2,…x M Based on the above prediction process, the prediction results F for N-1 periodic stationary data and 1 random non-stationary data at time t+1 are obtained respectively. i (t+1) and R N (t+1), then the final port TSP concentration C F The (t+1) short-time forecast result is:

[0103] This embodiment also provides a short-term prediction device for port TSP concentration, including a processor and a memory; the memory stores a program or instructions, which are loaded and executed by the processor to implement the steps of the short-term prediction method for port TSP concentration of the embodiment.

[0104] This embodiment provides a computer-readable storage medium storing a program or instructions that, when executed by a processor, implement the steps of short-term prediction of TSP concentration at a port according to the embodiment.

[0105] Example 1

[0106] Based on the invention's content, application examples are used to illustrate its application.

[0107] Step 1: Data Acquisition. Taking TSP concentration data monitored at a certain port as an example, obtain historical TSP monitoring data for the port from March 7th to March 14th, such as... Figure 2 As shown.

[0108] from Figure 2 As can be seen, the TSP concentration data exhibits time-series characteristics, with an average data acquisition interval of 1 minute, and a total of 11,520 concentration data points were acquired. That is, M data sequences X M =(x1,x2,x3,···x M In this case, M is 11520.

[0109] Step Two: Data Cleaning. The acquired concentration data undergoes outlier removal, data completion, and time interval adjustment to generate T T TSP concentration data points with time-series characteristics at 1-hour intervals. T =(x1,x2,x3,···x T The sequence is used for prediction, where T is 192, such as... Figure 3 As shown.

[0110] In the subsequent example analysis, 168 data points from March 3 to March 13 were used for modeling analysis, and 24 data points from March 14 were used to compare and analyze the prediction results of the prediction model of this invention and the prediction results of the classic ARIMA model to verify the effectiveness of the method of this invention.

[0111] Step 3: Data Decomposition. The cleaned TSP concentration data is decomposed to obtain 5 sets of time data series F with periodic variation characteristics. i = (F1, F2, F3, F4, F5), and a time series R1 with a set of random fluctuation characteristics, such as Figure 4 As shown.

[0112] Step 4: Model building.

[0113] Predictive models were constructed for five sets of time series exhibiting periodic variations. During the iterative optimization of the model parameters, the mean absolute deviation (MAD) was used as the error function, and the hyperbolic tangent function was selected as the activation function. The specific parameters and prediction results of the predictive models constructed for the five sets of periodic time series components are as follows:

[0114] The model parameters constructed from the periodic time series component F1 are shown in Table 1:

[0115] Table 1. Parameters of the prediction model constructed from the F1 component of the periodic time series.

[0116]

[0117] The waveform of the prediction model constructed from the F1 component of the periodic time series is compared with the waveform of the measured data. Figure 5 As shown:

[0118] At this point, the prediction results of the periodic time series component F1 for March 14 are shown in Table 2:

[0119] Table 2. Prediction results of the F1 prediction model for time series data.

[0120] time 1 2 3 4 5 6 7 8 concentration 0.16 -3.27 3.33 -0.94 0.16 -2.60 -3.79 -3.60 time 9 10 11 12 13 14 15 16 concentration -0.62 0.52 -3.43 0.91 -0.83 1.53 1.57 2.89 time 17 18 19 20 21 22 23 24 concentration -3.81 0.13 -0.18 -0.75 1.12 -1.15 -5.25 -2.22

[0121] The model parameters constructed from the periodic time series component F2 are shown in Table 3:

[0122] Table 3. Parameters of the prediction model constructed from the periodic time series component F2.

[0123]

[0124] The waveform of the prediction model constructed from the periodic time series component F2 is compared with the waveform of the measured data. Figure 6 As shown:

[0125] At this point, the prediction results of the periodic time series component F2 for March 14 are shown in Table 2:

[0126] Table 4. Prediction results of the time series data prediction model

[0127] time 1 2 3 4 5 6 7 8 concentration -0.96 -1.00 -0.42 0.44 0.80 1.24 1.21 1.42 time 9 10 11 12 13 14 15 16 concentration 2.00 1.03 -0.78 -1.17 -1.65 0.72 2.55 2.30 time 17 18 19 20 21 22 23 24 concentration -0.64 -2.33 -3.55 -2.69 0.44 0.86 1.02 1.28

[0128] The model parameters constructed from the periodic time series component F3 are shown in Table 5:

[0129] Table 5. Parameters of the prediction model constructed from the periodic time series component F3

[0130]

[0131]

[0132] The waveform of the prediction model constructed using the periodic time series component F3 is compared with the waveform of the measured data. Figure 7 As shown:

[0133] At this point, the prediction results of the periodic time series component F3 for March 14 are shown in Table 6:

[0134] Table 6. Prediction results of the F3 prediction model for periodic time series components.

[0135] time 1 2 3 4 5 6 7 8 concentration -10.00 -9.33 -8.61 -7.86 -7.06 -6.18 -5.10 -3.84 time 9 10 11 12 13 14 15 16 concentration -2.44 -1.01 0.17 0.95 1.36 1.49 1.52 1.36 time 17 18 19 20 21 22 23 24 concentration 0.86 0.14 -0.39 -0.57 -0.31 0.07 0.36 0.56

[0136] The model parameters constructed from the periodic time series component F4 are shown in Table 7:

[0137] Table 7. Parameters of the prediction model constructed from the periodic time series component F4

[0138]

[0139] At this point, the prediction results of the periodic time series component F4 for March 14 are shown in Table 8:

[0140] Table 8. Prediction results of the F4 prediction model for time series data.

[0141]

[0142]

[0143] The waveform of the prediction model constructed using the periodic time series component F4 is compared with the waveform of the measured data. Figure 8 As shown:

[0144] The model parameters constructed from the periodic time series component F5 are shown in Table 9:

[0145] Table 9. Parameters of the prediction model constructed from the periodic time series component F5

[0146]

[0147] The waveform of the prediction model constructed using the periodic time series component F5 is compared with the waveform of the measured data. Figure 9 As shown:

[0148] At this point, the prediction results of the periodic time series component F5 for March 14 are shown in Table 10:

[0149] Table 10 Prediction results of the F5 prediction model for periodic time series components

[0150]

[0151]

[0152] The prediction results of the random time series component R1 for March 14 are shown in Table 11:

[0153] Table 11 Prediction Results of the R1 Prediction Model for Time Series Data

[0154] time 1 2 3 4 5 6 7 8 concentration 13.10 13.08 13.06 13.04 13.03 13.01 12.99 12.97 time 9 10 11 12 13 14 15 16 concentration 12.95 12.93 12.91 12.89 12.87 12.85 12.83 12.81 time 17 18 19 20 21 22 23 24 concentration 12.79 12.77 12.75 12.73 12.71 12.69 12.67 12.65

[0155] Step 5: Concentration prediction.

[0156] Summarizing the prediction results of the five periodic time series components and the prediction results of the one random time series component, the prediction results of the prediction model constructed in this invention are shown in Table 12:

[0157] Table 12. TSP Concentration Prediction Results for March 14

[0158] time 1 2 3 4 5 6 7 8 concentration 4.66 1.43 8.91 5.85 7.72 5.91 5.43 6.77 time 9 10 11 12 13 14 15 16 concentration 11.46 12.83 8.06 12.63 10.70 15.47 17.31 18.20 time 17 18 19 20 21 22 23 24 concentration 8.07 9.64 7.65 7.81 13.15 11.74 8.14 11.70

[0159] To demonstrate the accuracy of the prediction model proposed in this invention, the prediction results of the model are compared with the classic time series prediction model, ARIMA. The parameters of the ARIMA model constructed based on TSP concentration data from March 3rd and March 13th are shown in Table 13.

[0160] Table 13 ARIMA Model Parameters

[0161]

[0162] The prediction results of TSP concentration data on March 14 using the ARIMA model are shown in Table 14:

[0163] Table 14 ARIMA Model Prediction Results

[0164] time 1 2 3 4 5 6 7 8 concentration 6.30 5.64 6.90 5.49 5.30 6.55 5.54 6.75 time 9 10 11 12 13 14 15 16 concentration 9.61 12.47 9.47 14.00 10.99 16.28 18.23 20.14 time 17 18 19 20 21 22 23 24 concentration 10.22 10.50 7.86 8.08 14.55 9.63 10.64 13.61

[0165] Comparison of the measured data from March 14th with the prediction results of the two prediction models: Figure 10 As shown:

[0166] The data comparison chart clearly shows that the prediction results of the proposed model are closest to the measured data, and it can predict significant concentration fluctuations. In contrast, the ARIMA model only reflects the general trend of concentration values, failing to effectively predict many fluctuations. Therefore, the ARIMA method is considered more suitable for data with good stationarity, which also demonstrates that the prediction model of this invention can better handle the non-stationarity of concentration time series.

[0167] To quantitatively evaluate the prediction results of the two prediction models, commonly used evaluation metrics, namely Root Mean Square Error (RMSE) and Mean Absolute Error (MAE), are selected, as shown in Table 15. Lower metric values ​​indicate higher prediction accuracy. The formulas for the two evaluation metrics are:

[0168]

[0169]

[0170] In the formula, Y i —True value, Y i '—Predicted value, N—Sample size

[0171] Table 15 Comparison of TSP Concentration Prediction Errors

[0172]

[0173] Comparing the prediction errors of TSP concentrations for the entire day of March 14th, the prediction error of the model proposed in this invention is smaller than that of the ARIMA method. From the evaluation metrics, the MAE of the prediction model proposed in this invention is 14.5% lower than that of ARIMA, and the RMSE is 13.43% lower than that of ARIMA. Therefore, the prediction model proposed in this invention has better prediction performance.

[0174] Due to the non-stationarity and randomness of TSP concentration data in ports, currently used prediction methods are difficult to meet the needs of port dust pollution prevention and control, thus affecting the green and ecological development of ports. This invention mainly designs a short-time prediction method for port TSP concentration data that considers the non-stationarity and randomness characteristics of port TSP concentration. This method can significantly improve the reliability of port TSP concentration prediction and is of great significance to port dust pollution prevention and control.

Claims

1. A method for predicting port TSP concentration considering the impact of dust suppression measures, characterized in that, include: Obtaining port access Data set of TSP concentrations monitored within hours Data set of meteorological factors under the influence of dust suppression measures ; TSP concentration dataset and meteorological factor dataset After cleaning and removing outliers, the data time intervals were adjusted to obtain the TSP concentration dataset for analysis. and meteorological factor dataset ; Construct a feature meteorological factor screening model and analyze the TSP concentration dataset. and meteorological factor dataset The correlation was analyzed to identify a set of characteristic meteorological factors that could influence TSP concentration. ; Construct a multi-input, single-output TSP concentration prediction model based on the TSP concentration dataset. Feature meteorological factor dataset The concentration forecast results of TSP at the port were obtained. ; The method for constructing a multi-input single-output TSP concentration prediction model is as follows: Data transformation: Transforming the cleaned TSP concentration dataset and characteristic meteorological factor dataset This is transformed into a series of lag times. A two-dimensional matrix is ​​used as the training sample and output label: First layer of data convolution: targeting the... training samples Set the input data for the first convolutional layer. for Let the total number of neurons in the first layer be denoted as . Let the first The weight matrix of each neuron is The deviation matrix is Then the first Output value of each neuron for: in, It is the sum of the Hadamard products of the elements of the two matrices; Input data for the first layer The output data after convolution for: Second-level data recursion: For the first... training samples The output data after the first convolution layer Input data for the second layer of data recursion ,Right now Let the number of neurons in the second layer be... Input data After the first After each neuron updates its state, the recursive output data is: The final result of the second layer of data recursion is: In the formula, It is the hyperbolic tangent function; This is the sigmoid growth curve function; , , and These are the set state weight values; , , and These are the set deviation matrix values; For the first training samples The output data obtained after data recursion, when hour It is the identity matrix; Third-layer weight optimization: optimize the output data of the second layer. Input data for third-layer weight optimization ,Right now ; Assume input data With output labels correlation score for: In the formula, This is the weight matrix; The deviation matrix; Relevance score for each input data point By normalizing the exponential function Calculate the input data Weighting indicators for: Based on the input data and weighting indicators The final TSP concentration forecast result is obtained by performing a weighted average summation. : 。 2. The port TSP concentration forecasting method considering the impact of dust suppression measures according to claim 1, characterized in that, Construct a feature screening model for characteristic meteorological factors and analyze the TSP concentration dataset. and meteorological factor dataset The correlation was analyzed to identify characteristic meteorological factors that could influence TSP concentration. In the steps, characteristic meteorological factors are constructed. The method for selecting models is as follows: Linear correlation coefficient calculation: based on the cleaned TSP concentration dataset and A meteorological monitoring dataset Calculate TSP concentration dataset With the Monitoring datasets of meteorological factors linear correlation coefficient between : in, x t For a moment Monitored TSP concentration; f t,k For a moment The monitored Data for individual meteorological factors; Calculate TSP concentration data and all The linear correlation coefficients among the meteorological factors are used to obtain vectors. ; Set a threshold for determining linear correlation ,if Then it is believed that the first Several meteorological factors showed a significant linear correlation with changes in TSP concentration, thus yielding results that met this condition. One characteristic meteorological factor, G≤K ; Nonlinear correlation coefficient calculation: based on the cleaned TSP concentration dataset and meteorological monitoring dataset Calculate TSP concentration data With the Monitoring data of meteorological factors Nonlinear correlation coefficient between : In the formula, For a moment Monitored TSP concentration Meteorological factors Monitoring data The smaller value in; For a certain moment Monitored TSP concentration and meteorological factors Monitoring data The joint probability; TSP concentration data The marginal probability, Meteorological factors Monitoring data The marginal probability; Calculate TSP concentration data and all The nonlinear correlation coefficients among various meteorological factors yield vectors. ; Set a threshold for nonlinear correlation determination ,if Then it is believed that the first Several meteorological factors exhibit a significant nonlinear correlation with changes in TSP concentration, thus yielding results that satisfy this condition. One characteristic meteorological factor, E≤K ; Screening of characteristic meteorological factors: linearly correlated Characteristic meteorological factors and nonlinear correlation The characteristic meteorological factors are combined into One characteristic meteorological factor: Calculation formula based on the screening index of characteristic meteorological factors Determine the screening indicators for meteorological factors ; The calculation formula is: In the formula, For control coefficients; To make the calculation formula The screening criterion with the smallest value; if If the meteorological factor has no significant impact on the change in TSP concentration, it is removed; otherwise, it is considered a characteristic meteorological factor. Determine the factors that ultimately affect TSP concentration The characteristic meteorological factors are ,in .

3. The port TSP concentration forecasting method considering the impact of dust suppression measures according to claim 2, characterized in that, At a certain moment Monitored TSP concentration and meteorological factors Monitoring data joint probability TSP concentration data marginal probability and meteorological factors Monitoring data marginal probability The method for determining this is based on the TSP concentration. and meteorological factors Detection data Within the distribution range of the Cartesian coordinate axes, it is divided into: One grid, for and The value of the number of grid cells occupied divided by the total number of grid cells. for The value of the number of grid cells occupied divided by the total number of grid cells. for The value of the number of grid cells occupied divided by the total number of grid cells.

4. The port TSP concentration forecasting method considering the impact of dust suppression measures according to claim 1, characterized in that, After cleaning TSP concentration data and characteristic meteorological factors This is transformed into a series of lag times. The transformation method for using a two-dimensional matrix as training samples and output labels is as follows: In the formula, For the first One training sample; For the first One output label; =1, 2, ..., Tp .

5. A port TSP concentration forecasting method considering the impact of dust suppression measures according to any one of claims 1-4, characterized in that, The historical monitoring data on port TSP concentrations and meteorological factors under the influence of dust suppression measures include: Port TSP Concentration Data Acquisition: Based on the port's online TSP monitoring system, within a specified time range... Port monitoring obtained internally The TSP concentration data constitute a dataset with time-series characteristics. ; Port meteorological data acquisition: within the time range Internal monitoring is possible A dataset with time-series characteristics composed of meteorological factors. for: in, f m,k For the first The first meteorological factor One monitoring data point, , .

6. A port TSP concentration forecasting method considering the impact of dust suppression measures according to claim 5, characterized in that, The method for cleaning the historical monitoring data of TSP concentration and meteorological factors is as follows: Abnormal data removal: If in Dataset of TSP concentrations In the middle, the first Monitoring data at each time point satisfy or Then As abnormal data, it was cleared. ;right In the meteorological monitoring data, the first various meteorological factors One monitoring data In the middle, the first Monitoring data at each time point satisfy or Then As abnormal data, it was cleared. ;in , , , for: Time interval adjustment: Adjust the acquired TSP concentration dataset and meteorological factor dataset to any time interval. Concentration data; for time and Between The time was obtained by arithmetically averaging the TSP concentration data. TSP concentration ; For time and Between The first The arithmetic mean of the monitoring data of various meteorological factors was obtained. Time of the first Monitoring data of various meteorological factors are ; Data cleaning results: [The data will be used to clean the port's...] TSP concentration data After cleaning, time-series data with 1-hour intervals were obtained. Dataset of TSP concentrations ; Monitoring datasets of meteorological factors After cleaning Monitoring datasets of meteorological factors .

7. The port TSP concentration forecasting method considering the impact of dust suppression measures according to claim 1, characterized in that, The weight matrix and bias matrix of the constructed three-layer multi-input single-output forecast model are both based on measured TSP concentration data. and characteristic meteorological factor data The optimal parameters are determined through iterative updates.

8. A port TSP concentration forecasting device that takes into account the impact of dust suppression measures, characterized in that, It includes a processor and a memory; the memory stores a program or instructions, which are loaded and executed by the processor to implement the steps of the port TSP concentration forecasting method under the influence of any of the dust suppression measures as described in claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the port TSP concentration forecasting method under the influence of any of the dust suppression measures as described in claims 1 to 7.

Images