A deep learning-based pollutant sea flux monitoring and prediction method

Abstract

The present application relates to the technical field of sea flux monitoring, and particularly relates to a pollution sea flux monitoring and prediction method based on deep learning. The method comprises the following steps: collecting historical flow and pollutant concentration data of a target area at different time intervals, and locking the area to be analyzed accordingly; then obtaining water samples in the area, extracting nitrite, ammonia nitrogen, nitrate concentration and flow data, and constructing a training set; after standardizing the training set and real-time flow test set, performing deep time sequence learning with a TCN model to predict future flow and pollutant concentration; and finally calculating the sea flux based on the prediction result to realize high-precision sea flux monitoring and prediction of pollutants. The present application improves the prediction accuracy and efficiency of sea flux of pollutants in complex terrain, extreme climate and densely populated areas by using time sequence convolution network to perform deep causal convolution and residual feature extraction on long time sequence data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of sea flux monitoring, and particularly relates to a pollution sea flux monitoring and prediction method based on deep learning. BACKGROUND

[0002] The pollution sea flux refers to the amount of a certain pollutant entering the sea environment through rivers and other input channels in a certain region per unit time. However, the traditional pollution sea flux prediction method will face some problems in practical application, especially when dealing with long time series data, there is certain limitation. The traditional pollution sea flux prediction method mostly depends on physical numerical models or statistical regression models. These models are mainly based on the hydrological process of the basin and the change law of the weather, and deduce the prediction of the sea flux. Common physical numerical models include hydrological models, weather models, etc., which usually need a large number of parameters and accurate boundary conditions and initial conditions for accurate modeling.

[0003] In practical application, it is often challenging to obtain accurate boundary conditions and initial conditions, especially under the influence of complex topography, climate conditions and human activities, which leads to unsatisfactory prediction results of these models in complex topographic environments, extreme climate environments, and densely populated areas. On the other hand, statistical regression models rely on the statistical relationship between data, which usually assumes a certain linear relationship or fixed distribution pattern between data, which deviates greatly from the actual situation.

[0004] In summary, it is necessary to develop a pollution sea flux prediction method based on a TCN model to overcome the limitations of the prior art and improve the prediction accuracy and efficiency. SUMMARY

[0005] Therefore, it is necessary to provide a pollution sea flux monitoring and prediction method based on deep learning to solve at least one of the above technical problems.

[0006] To achieve the above-mentioned purpose, a pollution sea flux monitoring and prediction method based on deep learning, the method comprising the following steps:

[0007] Step S1: collecting historical water flow data of a target region based on a first time interval; collecting historical pollutant concentration data of the target region based on a second time interval; determining a pollution analysis region according to the historical water flow data and the historical pollutant concentration data;

[0008] Step S2: obtaining a regional water sample based on the pollution analysis region; processing the regional water sample to obtain nitrite concentration data, ammonia nitrogen concentration data and nitrate concentration data respectively, and combining the historical water flow data as a model training set;

[0009] Step S3: pre-process the model training set to obtain a standardized model training set, and take the real-time collected water flow data as a model test set; train a TCN model using the standardized model training set and the model test set, and use the trained TCN model to predict the flow and pollutant concentration of the pollutant analysis area to obtain water flow prediction data and pollutant concentration prediction data;

[0010] Step S4: calculate the pollutant sea flux prediction value according to the water flow prediction data and the pollutant concentration prediction data to perform the pollutant sea flux monitoring and prediction operation.

[0011] Preferably, the step S1 of determining the pollutant analysis area according to the historical water flow data and the historical pollutant concentration data comprises:

[0012] Taking the maximum value and the minimum value of the historical pollutant concentration data as the center, the first monitoring station position and the second monitoring station position closest to the maximum value and the minimum value are determined respectively;

[0013] According to the first monitoring station position and the second monitoring station position, the monitoring radius of the monitoring station is analyzed to obtain the first monitoring station monitoring radius and the second monitoring station monitoring radius;

[0014] Based on the first monitoring station monitoring radius and the second monitoring station monitoring radius, the initial pollutant analysis area between the first monitoring station position and the second monitoring station position is determined;

[0015] The Pearson correlation coefficient between the historical water flow data and the historical pollutant concentration data is calculated;

[0016] The Pearson correlation coefficient is compared with a preset correlation threshold value, and when the Pearson correlation coefficient is greater than or equal to the preset correlation threshold value, the third monitoring station position is determined based on the Pearson correlation coefficient, and the initial pollutant analysis area is re-screened according to the third monitoring station position to obtain the pollutant analysis area.

[0017] Preferably, the step of determining the third monitoring station position based on the Pearson correlation coefficient and re-screening the initial pollutant analysis area according to the third monitoring station position comprises:

[0018] The initial pollutant analysis area is gridded, and the Pearson correlation coefficient of each grid cell in the initial pollutant analysis area is determined;

[0019] The grid cells with a Pearson correlation coefficient greater than or equal to the preset correlation threshold value are screened to obtain a pollutant transport sensitive area;

[0020] Based on the edge of the pollutant transport sensitive area, the third monitoring station position closest to the edge is determined;

[0021] According to the third monitoring station position, a section position collected by the third monitoring station is obtained, and a flow velocity gradient of the river section is calculated in combination with a historical water flow to determine a pollutant migration space range;

[0022] A region intersection calculation is performed on the pollutant migration space range and the initial pollutant region to be analyzed to obtain the pollutant region to be analyzed.

[0023] Preferably, when the Pearson correlation coefficient is less than a preset correlation threshold, the method further comprises:

[0024] When the Pearson correlation coefficient is less than a preset correlation threshold, the initial pollutant region to be analyzed determined based on the first monitoring station position and the second monitoring station position is divided into a plurality of grid cells of equal size;

[0025] For each grid cell, a difference between a historical pollutant concentration average value of the corresponding monitoring station and a global historical average concentration of the target region is calculated to obtain a concentration deviation value of the grid cell.

[0026] Grid cells with a concentration deviation value greater than a preset deviation threshold are screened as high-concentration candidate cells, wherein the preset deviation threshold is greater than the global average concentration + 2 times the standard deviation.

[0027] A cluster analysis is performed on the high-concentration candidate cells in space to extract pollutant aggregation sub-regions that are adjacent to each other and have consistent deviations.

[0028] Based on the extracted pollutant aggregation sub-regions, an outer polygon boundary thereof is used as the pollutant region to be analyzed.

[0029] Preferably, the step S2 of obtaining a regional water sample based on the pollutant region to be analyzed comprises:

[0030] The geographical boundary of the pollutant region to be analyzed is determined, and a GPS positioning system is used to mark the sampling point positions, with a sampling point spacing controlled within a range of 50 m to 200 m.

[0031] According to the water depth and flow conditions of the pollutant region to be analyzed, a sampling depth is selected, with a sampling depth of 0.3 m to 0.5 m in shallow water areas and a sampling depth of 1.0 m to 3.0 m in deep water areas; when the water flow velocity exceeds 0.5 m / s, an anti-disturbance sampler is used to ensure the integrity of the sample;

[0032] The volume range of the collected water sample is 500 mL to 2 L, the material of the sampling bottle is polyethylene or glass, the bottle body needs to be pre-acid washed and cleaned with deionized water to avoid secondary pollution, and the pre-acid washing specifically comprises soaking in a 1:1 HNO3 solution for 30 minutes;

[0033] The environmental condition data, including water temperature, pH value and dissolved oxygen concentration, are recorded during the sampling process and are attached with a sampling time stamp to obtain the regional water body sample.

[0034] Preferably, the processing of the regional water body sample in step S2 to obtain nitrite concentration data, ammonia nitrogen concentration data and nitrate concentration data respectively comprises:

[0035] When the N-1-naphthyl ethylenediamine spectrophotometric method is used to determine the nitrite content, the detection wavelength is set to 540 nm, if the measured concentration is 0.02 mg / L-1.0 mg / L, and the color reaction time is not more than 10 minutes, it is considered as stable nitrite concentration data; if color drift or interference peak shift is greater than ± 5 nm, the detection is invalid;

[0036] When the Nash reagent colorimetric method is used to detect ammonia nitrogen concentration, under the condition of determination wavelength 425 nm, if the ammonia nitrogen concentration is stable between 0.1 mg / L and 10 mg / L, and the correlation coefficient R² of the colorimetric absorbance curve and the standard curve is greater than 0.98, it is determined that the detected ammonia nitrogen concentration data is valid; if the concentration exceeds 15 mg / L or the reagent coloration is incomplete, it needs to be diluted and measured again;

[0037] When the ultraviolet spectrometry or ion chromatography method is used to determine the nitrate concentration, the concentration range is 0.5 mg / L-20 mg / L, the absorbance ratio between 220 nm and 275 nm is greater than 1.5, and the background correction value should not exceed 0.05, which can be confirmed as valid nitrate concentration data; if the deviation of the results of two consecutive detections is greater than ± 1.0 mg / L, an abnormality flag is triggered.

[0038] Preferably, the training of the TCN model in step S3 using the standardized model training set and the model test set comprises:

[0039] The TCN model structure is constructed, wherein the TCN model structure comprises an input layer, a causal convolution layer, a dilated convolution layer and an output layer, wherein the input layer receives the time series in the standardized model training set, and the input tensor has a shape of (B, T, F), wherein B is the batch size, T is the time step number, and F is the feature number;

[0040] Based on the causal convolution layer, (K-1) zeros are filled in the left side of the time dimension of the input tensor by zero padding to obtain first model training data, wherein K is the kernel size;

[0041] Based on the dilated convolution layer, a dilated factor D is introduced to the first model training data to capture the time-dependent relationship to obtain second model training data, wherein the dilated factor D represents the interval of the convolution kernel when processing the first model training data;

[0042] The second model training data is subjected to multi-layer convolution and activation transformation by an activation function, and the task type of the transformation result is determined, when the task type is a classification task, a Softmax function is used after the output layer to generate a probability distribution, and a probability distribution result is obtained, when the task type is a regression task, a prediction value is directly output by the output layer, wherein the shape of the output tensor in the output layer is (B, T, C), wherein C is the number of output channels;

[0043] The model precision of the probability distribution result or the prediction value is evaluated according to the model test set, so as to obtain the trained TCN model.

[0044] Preferably, the causal convolution layer further comprises:

[0045] Each residual block comprises two causal convolution layers, each of which is followed by an activation function and a regularization layer;

[0046] In the first layer of causal convolution, the convolution kernel size is K, the expansion factor is D, and the number of output channels is C; after the convolution output, a non-linear feature is introduced through a ReLU activation function, and the calculation formula is:

[0047] ;

[0048] In the formula, is the input value, is the maximum value function;

[0049] A Dropout regularization layer is set after the activation function, and the random discard ratio is 0.2;

[0050] The convolution kernel size and the expansion factor of the second layer of causal convolution are the same as those of the first layer, that is, the convolution kernel size K and the expansion factor D do not change;

[0051] The input tensor of the residual block is directly added to the output of the second layer of causal convolution through a jump connection, wherein the output of the residual block is defined by the following formula:

[0052] ;

[0053] In the formula, is the final output of the residual block, is the input tensor, is the first layer convolution operation, is the application of a nonlinear activation function, is the second layer convolution operation.

[0054] Preferably, the task type of the transformation result is determined by:

[0055] A classification task is determined when any of the following conditions are met: the label format of the training data of the second model is a discrete category index, and the number of categories is greater than 1; the number of channels C in the last dimension of the output tensor matches the number of categories, and the value range is mainly between 0 and 1; the tensor features after convolution and activation transformation satisfy the probability distribution characteristics.

[0056] A task is considered a regression task when the following conditions are met simultaneously: the labels of the second model's training data are continuous real values, and the last dimension of the output tensor C=1; the output values ​​after convolution and activation transformation are not limited to 0,1, specifically... Fluctuations within a certain range; the output tensor does not satisfy the probability normalization characteristic.

[0057] Preferably, step S4, which calculates the predicted pollutant flux into the sea based on water flow prediction data and pollutant concentration prediction data, includes:

[0058] The predicted pollutant inflow flux into the sea is calculated based on the predicted water flow and pollutant concentration data using the inflow flux calculation formula, which is shown below:

[0059] ;

[0060] in for Instantaneous flux of pollutants into the sea at any given time. for Real-time pollutant concentration prediction data, for Real-time water flow prediction data For a moment.

[0061] The present invention has the following beneficial effects:

[0062] Firstly, this invention constructs a training set based on multi-source time-series hydrological and water quality data (flow rate, nitrite, ammonia nitrogen, and nitrate). Compared with single concentration or flow rate prediction methods, it can fully explore the mutual influence and coupling relationship between various elements, effectively improving the overall prediction accuracy and stability of the model for pollutant flux into the sea. Especially under conditions of drastic fluctuations in pollutant concentration or abnormal flow rate, it can still maintain high prediction robustness.

[0063] Secondly, in step S1, by combining historical extreme value site location with Pearson correlation analysis, the area to be analyzed for pollutants is dynamically screened and finely delineated. This can accurately pinpoint the sensitive area for pollutant migration, avoiding the blind spots and redundant areas of traditional methods based on a single fixed monitoring range, thereby significantly improving the economy and spatiotemporal resolution of regional sampling and prediction.

[0064] Third aspect: In step S2, a stratified sampling scheme is designed according to water depth, flow rate and disturbance resistance characteristics of the sampler, and strict detection deviation and quality control threshold values are set for each chemical analysis method (ammonium molybdate spectrophotometry, naphthyl ethylenediamine spectrophotometry, Nash reagent colorimetry, ultraviolet spectrophotometry or ion chromatography), which can effectively guarantee the accuracy and consistency of the data, provide high-quality input for model training, and reduce the prediction error caused by abnormal samples.

[0065] Fourth aspect: The one-dimensional causal convolution network (TCN) used in step S3 combines dilated convolution and residual structure, which can capture long-term and short-term time series dependence and avoid the gradient disappearance problem of the cyclic network; according to the label type, the classification / regression output is automatically switched, so that the model has universality and adaptability, which greatly improves the operation efficiency and accuracy in real-time online prediction scene.

[0066] Fifth aspect: In step S4, the instantaneous flux into the sea is directly calculated by using the predicted flow and concentration data, and the formula is simple and clear (F = Q * C) ), without additional parameter fitting in the calculation process, which is convenient for rapid deployment in monitoring and early warning system to realize the "prediction-calculation-alarm-feedback" closed loop, significantly shorten the response time, and improve the timeliness and operability of pollution prevention and control. BRIEF DESCRIPTION OF DRAWINGS

[0067] Figure 1 It is a step flowchart of a pollution flux into the sea monitoring and prediction method based on deep learning;

[0068] Figure 2 It is a step flowchart of a pollution flux into the sea monitoring and prediction method based on deep learning provided by an embodiment of the present application;

[0069] Figure 3 It is a structure example diagram of a TCN model of a pollution flux into the sea monitoring and prediction method based on deep learning provided by an embodiment of the present application;

[0070] Figure 4 It is a comparison diagram of nitrite instantaneous flux into the sea of a pollution flux into the sea monitoring and prediction method based on deep learning provided by an embodiment of the present application;

[0071] Figure 5 It is a comparison diagram of nitrate instantaneous flux into the sea of a pollution flux into the sea monitoring and prediction method based on deep learning provided by an embodiment of the present application;

[0072] Figure 6 It is a comparison diagram of ammonia nitrogen instantaneous flux into the sea of a pollution flux into the sea monitoring and prediction method based on deep learning provided by an embodiment of the present application;

[0073] The objectives, functional characteristics and advantages of the present application will be further described with reference to the embodiments in conjunction with the drawings. DETAILED DESCRIPTION

[0074] The technical method of the present application will be described clearly and completely below in conjunction with the drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0075] In addition, the drawings are only schematic illustrations of the present application and are not necessarily drawn to scale. Identical reference numerals in the drawings represent identical or similar parts, and thus repeated descriptions thereof will be omitted. Some of the block diagrams shown in the drawings are functional entities, which do not necessarily have to correspond to physically or logically independent entities. The functional entities can be implemented in the form of software, or in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0076] It should be understood that although the terms "first", "second" and the like can be used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another. For example, without departing from the scope of the exemplary embodiments, a first element can be referred to as a second element, and similarly a second element can be referred to as a first element. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0077] To achieve the above-mentioned object, please refer to Figures 1 to 6 A pollution flux into the sea monitoring and prediction method based on deep learning, the method comprising the following steps:

[0078] Step S1: collecting historical water flow data of the target area based on a first time interval; collecting historical pollutant concentration data of the target area based on a second time interval; determining a pollutant analysis area according to the historical water flow data and the historical pollutant concentration data;

[0079] In an embodiment, the first time interval can be set to 15 minutes. A high-frequency continuous observation station is arranged on the coastline. From 00:00 on February 1, 2024, the instantaneous flow value at the intertidal zone section is automatically recorded at 15-minute intervals. After the instantaneous flow value collected at each observation station is calibrated by UTC time, it is sequentially stored in the database table WaterFlowHistory to generate a water flow time series data set from 00:00 on February 1, 2024 to the cutoff time.

[0080] The second time interval can also be set to 4 hours, and through the automatic water quality analyzer installed in the same observation station network, for example, since November 1, 2022, 00:00, the concentrations of pollutants such as suspended particulate matter (TSS), chemical oxygen demand (COD), and heavy metals are sampled and determined online at 4-hour intervals. The pollutant concentration values at each time point are stored in the water quality management system in the form of PollutionConcHistory together with the corresponding sampling time, sampling depth, instrument model, and other metadata.

[0081] The maximum value MaxFlow and the minimum value MinFlow of the flow value at each time point in the water body flow time series data set are calculated, and the maximum value MaxConc and the minimum value MinConc of each pollutant index in the pollutant concentration time series data set are calculated. Respectively, (MaxFlow, MinFlow), (MaxConc, MinConc) as the center, search for the nearest monitoring station position from the four center values respectively, determine the first monitoring station A and the second monitoring station B, take the coastal section where the stations A and B are located as the boundary of the pollutant analysis area, draw a polygon monitoring area, and mark all sub-section observation points in the domain as the pollutant analysis area.

[0082] Step S2: Obtain regional water samples based on the pollutant analysis area; process the regional water samples to obtain nitrite concentration data, ammonia nitrogen concentration data, and nitrate concentration data, respectively, and combine historical water flow data as a model training set;

[0083] In an embodiment, the sampling interval is determined to be every six hours, and from the date of the start of sample collection work, it continues until the total amount of required samples is collected. A number of equidistant sampling points are pre-selected within the delineated polygon monitoring area. The sampling boat or unmanned sampling platform operates according to the predetermined route, reaches each sampling point, extracts water samples from half a meter below the water surface, fills each collected raw water into a special sampling bottle, and labels the bottle with the sampling date, specific time, and sampling point number. On-site low-temperature preservation is carried out using a refrigerator or ice bag, and the samples are kept below four degrees Celsius when transported back to the laboratory.

[0084] The classical colorimetric analysis method can be used to react the water sample with nitrite color reagent, and the spectrophotometer is used to read the absorbance at the corresponding wavelength to obtain the concentration of nitrite. The Kjeldahl method can be used to treat the water sample, and the generated ammonia gas is measured by titration or automatic nitrogen determination instrument after heating digestion, and the ammonia nitrogen content is converted. The ion chromatography method can be used to inject the water sample into the chromatography system, separate and detect the peak area through the ion exchange column, and directly read the concentration of nitrate. Each sample item is measured at least three times in parallel, and the average value of the three results is taken as the final concentration data. At the same time, blank samples and standard samples are used to verify the accuracy of the analysis method.

[0085] The four nutrient salt concentrations obtained at each sampling time point are corresponded to the flow monitoring values at the same time point and the same position one by one. The flow value and the four nutrient salt concentrations obtained at each sampling time are arranged in chronological order, and the flow value and the four nutrient salt concentrations obtained at each sampling time are arranged in chronological order. The complete record is arranged into an electronic table or a database, and all records are arranged in sequence to form a training set table containing flow data and nutrient salt concentration data, which can be directly used for subsequent machine learning model training or statistical analysis.

[0086] Step S3: preprocessing the model training set to obtain a standardized model training set, and collecting the water flow data in real time as a model test set; using the standardized model training set and the model test set to train the TCN model, and using the trained TCN model to predict the flow and pollutant concentration of the analyzed area, to obtain water flow prediction data and pollutant concentration prediction data;

[0087] In an embodiment, it is checked whether there are data points that have not been collected or failed in the training set. The average value of the adjacent two measurement values before and after the same monitoring point is used to fill in these missing values, and the upper and lower limits of each record are judged. If a value exceeds the normal range of flow or concentration history change (for example, the flow exceeds the local tidal extreme value, or the ammonia nitrogen concentration exceeds the upper limit of environmental regulations), it is rejected or replaced with a nearby reasonable value. The "min-max normalization" method is used to map all flow and concentration data to 0 to 1: find the historical maximum and minimum value of each monitoring index; subtract the minimum value from the original value, and then divide by "maximum value minus minimum value" to obtain the normalized value. After processing, the data of different indicators have the same measurement scale, which facilitates fast convergence during model training.

[0088] All the standardized data obtained are divided into 80% for model training, and the remaining 20% of historical data and the latest flow data collected in real time are used as input for model performance testing and online prediction after training. The flow data collected in real time needs to be normalized by the same min-max normalization process to ensure consistency with the training set scale.

[0089] Temporal Convolutional Network (TCN) is adopted, which uses one-dimensional convolution and dilated convolution to capture long-time dependencies, and is more efficient than traditional recurrent networks. The input layer accepts the traffic and four concentration normalized values of "several hours" or "several days", and stacks 3 layers of dilated convolution, each with a kernel size of 3, and the dilation coefficients (i.e. the receptive field expansion factor) are 1, 2, and 4 in turn. After each convolution, a batch normalization and a ReLU activation are connected, and a fully connected layer is connected at the end to map the network output to the traffic and concentration prediction values of "future time".

[0090] In the Python environment, TensorFlow or PyTorch framework can be selected, the loss function uses mean square error (MSE), and the optimizer can use Adam with an initial learning rate of 0.001; during training, the batch size is set to 32 for each iteration, and a maximum of 100 training cycles is performed, and the performance is evaluated on the test set every 10 cycles; if the test error does not decrease significantly in three consecutive evaluations, the training is stopped in advance.

[0091] The latest traffic and concentration values of several time periods collected and normalized are input into the trained network, and the network can output the traffic prediction values and four nutrient salt concentration prediction values in the future period. The standardized prediction values output by the network are inversely mapped according to the recorded maximum and minimum values to restore the actual physical quantities, and the prediction data that can be directly applied to environmental monitoring and early warning are obtained.

[0092] Step S4: Calculate the pollutant flux prediction value according to the water flow prediction data and the pollutant concentration prediction data to perform the pollutant flux monitoring and prediction operation.

[0093] In an embodiment, the flow prediction value of each sampling time in the future period is obtained from the model output, with a unit of cubic meters per second, and the concentration prediction value of each pollutant (such as nitrite, ammonia nitrogen, nitrate) in the same period is also obtained, with a unit of milligrams per liter.

[0094] Convert the flow of "cubic meters per second" to "liters per second" (one cubic meter is equal to one thousand liters) to match the concentration unit. For each time, multiply the flow (liters per second) at that time by the concentration (milligrams per liter) of each pollutant to obtain the instantaneous flux of that pollutant at that time, with a unit of milligrams per second. If kilogram or ton-level flux is required, first convert milligrams to kilograms (one million milligrams is equal to one kilogram), and then accumulate or average.

[0095] Set a monitoring period (for example, one day or one week), add the instantaneous fluxes of each time point in the period according to the time interval to obtain the cumulative flux of the period, in milligrams or kilograms. If the average flux is needed, the cumulative flux is divided by the total number of seconds in the period to obtain the average value in milligrams per second or kilograms per second.

[0096] Arrange the cumulative flux and average flux of each pollutant into a table, list the time period, flow prediction value, concentration prediction value, instantaneous flux, cumulative flux, etc. columns, and export as an Excel table or database record.

[0097] Preferably, the pollutant analysis area to be analyzed in step S1 is determined according to historical water flow data and historical pollutant concentration data, including:

[0098] Step S11: Centering on the highest value and the lowest value of the historical pollutant concentration data, respectively determine the first monitoring station position and the second monitoring station position closest to the highest value and the lowest value;

[0099] Step S12: According to the first monitoring station position and the second monitoring station position, analyze the monitoring radius of the monitoring station to obtain the first monitoring station monitoring radius and the second monitoring station monitoring radius;

[0100] Step S13: Determine the initial pollutant analysis area between the first monitoring station position and the second monitoring station position based on the first monitoring station monitoring radius and the second monitoring station monitoring radius;

[0101] Step S14: Calculate the Pearson correlation coefficient between the historical water flow data and the historical pollutant concentration data;

[0102] Step S15: Compare the Pearson correlation coefficient with the preset correlation threshold value, and when the Pearson correlation coefficient is greater than or equal to the preset correlation threshold value, determine the third monitoring station position based on the Pearson correlation coefficient, and reselect the initial pollutant analysis area according to the third monitoring station position to obtain the pollutant analysis area.

[0103] In an embodiment, please refer to Figure 2 In the historical concentration records of all monitoring stations, find out the monitoring station where the highest value of the pollutant concentration is located and the monitoring station where the lowest value is located. If there are multiple stations with the same concentration, select the station closest to the coastline as the first monitoring station, and the second closest as the second monitoring station. The concentration extreme value must be based on at least one year of data (daily sampling frequency is not less than once a day), otherwise it may be distorted due to seasonal or occasional events, and if the extreme value station is distributed in both estuary and sea areas, the station in the estuary area should be selected first to ensure the maximum correlation between the prediction area and the inflow.

[0104] According to the design specification of the monitoring instrument, determine the effective coverage radius of the single station water quality monitoring (recommended range 1km~5km), record the effective coverage radius of the first monitoring station and the second monitoring station as "first monitoring radius" and "second monitoring radius" respectively, wherein the coverage radius shall not exceed half of the distance between adjacent monitoring stations, so as to avoid excessive data redundancy caused by excessive area overlap.

[0105] In another embodiment, draw two circles with the first monitoring station and the second monitoring station as the center and the respective monitoring radius as the radius, and take the common intersection area of the two circles as the "initial analysis area", if there is no intersection between the two circles, connect the center point of the shortest line between the two circles, take the center point as the new center, and take half of the sum of the two original radii as the radius to form a third circle, and then take the intersection, the shape of the area can be directly generated by using the "circle intersection" tool in GIS software (such as ArcGIS, QGIS).

[0106] In another embodiment, in the initial analysis area, the number of synchronous sampling of historical water flow and historical pollutant concentration is counted respectively (it is recommended to be not less than 100 pairs), and the Pearson correlation coefficient is calculated by using common statistical software (such as R, Python Pandas+SciPy) to measure the linear correlation degree of flow change and concentration change, and the calculated correlation coefficient is compared with the threshold value: if the correlation coefficient is less than the threshold value, the initial analysis area is retained and no new monitoring station is added; if the correlation coefficient is greater than or equal to the threshold value, in all candidate monitoring stations, the monitoring station with the second highest correlation with the area flow-concentration is selected as the third monitoring station, the monitoring radius of the third monitoring station is determined in the same way as the above radius, and then the intersection with the initial analysis area is taken to obtain the final "pollutant analysis area", wherein if there is no second highest correlation station (i.e. only two station data), the intersection of the two stations is directly taken as the final area.

[0107] Preferably, the third monitoring station position is determined based on the Pearson correlation coefficient, and the initial pollutant analysis area is re-screened according to the third monitoring station position, including:

[0108] Grid the initial pollutant analysis area, and determine the Pearson correlation coefficient of each grid cell in the initial pollutant analysis area;

[0109] Screen the grid cells with a Pearson correlation coefficient greater than or equal to a preset correlation threshold to obtain a pollutant transport sensitive area;

[0110] Based on the edge of the pollutant transport sensitive area, determine the position of the third monitoring station closest to the edge;

[0111] According to the position of the third monitoring station, obtain the cross-section position collected by the third monitoring station, and combine the historical water flow to calculate the flow velocity gradient of the river section to determine the pollutant migration space range;

[0112] The intersection of the pollutant migration spatial range and the initial pollutant analysis region is calculated to obtain the pollutant analysis region.

[0113] In an embodiment, the initial analysis region is divided into small grids each having an area of about 0.1 to 0.5 square kilometers, and the Pearson correlation coefficient is calculated at the center of each grid based on historical flow and concentration data. The sensitive grids having a correlation coefficient greater than or equal to a preset threshold (recommended value: 0.7) are retained to form a pollutant migration sensitive region. A third monitoring station is selected at the edge of the sensitive region and at the shortest distance from the monitoring station, the cross-section position of the third monitoring station is read, and the flow velocity gradient is calculated using the flow and width data of adjacent cross-sections. The pollutant migration range is estimated in the upstream and downstream directions. The intersection of the migration range and the initial analysis region is calculated to obtain the final pollutant analysis region.

[0114] In another embodiment, the extension distance can be determined by multiplying the 95th percentile flow velocity gradient by the longest prediction period (e.g., 24 hours) to avoid excessive extension. The intersection calculation can be completed by the "polygon intersection" tool in GIS software, and the result must be checked for surface shape to ensure that there are no broken or redundant small pieces.

[0115] Preferably, when the Pearson correlation coefficient is less than the preset correlation threshold, the initial pollutant analysis region determined based on the first monitoring station position and the second monitoring station position is divided into a plurality of grid cells of equal size.

[0116] When the Pearson correlation coefficient is less than the preset correlation threshold, the initial pollutant analysis region determined based on the first monitoring station position and the second monitoring station position is divided into a plurality of grid cells of equal size.

[0117] For each grid cell, the difference between the average historical pollutant concentration of the corresponding monitoring station and the global average historical concentration of the target region is calculated to obtain the concentration deviation value of the grid cell.

[0118] The grid cells having a concentration deviation value greater than a preset deviation threshold are selected as high-concentration candidate cells, wherein the preset deviation threshold is greater than the global average concentration + 2 times the standard deviation.

[0119] The high-concentration candidate cells are subjected to spatial clustering analysis to extract pollutant aggregation sub-regions that are adjacent to each other and have consistent deviations.

[0120] Based on the extracted pollutant aggregation sub-regions, the outer polygon boundary of the pollutant aggregation sub-regions is used as the pollutant analysis region.

[0121] In an embodiment, when the Pearson correlation coefficient is lower than a preset correlation threshold, the region to be analyzed is preferably redefined in the following manner: the initial region with the first and second monitoring stations as the center and the respective monitoring radius as the radius is equally divided into a plurality of grids (it is recommended that the area of each grid be controlled within 0.2 to 0.5 square kilometers to ensure that at least 50 historical concentration records are contained in each grid); the difference between the average concentration of each grid and the historical average concentration of the entire region is calculated, and high-concentration candidate grids with a difference greater than the "global average concentration plus twice the standard deviation" (if the standard deviation is significantly high due to abnormal values, the abnormal points can be removed or the data window can be extended) are screened out; then, spatial clustering (DBSCAN algorithm can be used, and the minimum cluster size is not less than 3 adjacent grids to avoid the interference of isolated points) is performed on the candidate grids to extract the pollution aggregation sub-region with adjacent and consistent concentration deviation; finally, the outer polygon boundary of the sub-region is taken as the final pollution region to be analyzed. In this method, the grid size, deviation threshold and clustering parameters should be adjusted appropriately in combination with the actual data density and the site layout to ensure that each step has sufficient sample support and can reflect the local pollution characteristics.

[0122] Preferably, the step S2 of obtaining the regional water sample based on the pollution region to be analyzed comprises:

[0123] The geographical boundary of the pollution region to be analyzed is determined, and the sampling point positions are marked by a GPS positioning system, with a sampling point spacing controlled within 50m to 200m;

[0124] The sampling depth is selected according to the water depth and flow conditions of the pollution region to be analyzed, with a sampling depth of 0.3m to 0.5m in shallow water and a sampling depth of 1.0m to 3.0m in deep water; when the water flow speed exceeds 0.5m / s, a disturbance-resistant sampler is used to ensure the integrity of the sample;

[0125] The volume of the water sample collected is 500mL to 2L, the material of the sampling bottle is polyethylene or glass, the bottle body needs to be pre-acid washed and washed clean with deionized water to avoid secondary pollution, and the pre-acid washing specifically comprises soaking in a 1:1 hydrochloric acid solution for 30min;

[0126] The environmental condition data, including water temperature, pH value and dissolved oxygen concentration, are recorded during the sampling process, and a sampling timestamp is added to obtain the regional water sample.

[0127] ​In an embodiment, in actual operation, the following points should be paid special attention to: during GPS calibration, the signal should be stable and accurate to ensure that the position error of each sampling point is not more than ±5 m; the distance between sampling points should be between 50 m and 200 m, while the representativeness of the region and the work efficiency are taken into account; secondly, the water depth and flow velocity should be accurately determined, and in shallow water areas (≤1 m), sampling can be performed at 0.3 m to 0.5 m, and in deep water areas (>1 m), multiple depths in the range of 1.0 m to 3.0 m should be selected to reflect the vertical variation of the water body; when the flow velocity exceeds 0.5 m / s, a sampler with a disturbance prevention device must be used to prevent the sample from being turbid or lost due to rapid flow; in addition, the material of the sampling bottle should be matched with the analysis project, and polyethylene or glass bottles should be soaked in 1:1 nitric acid for 30 minutes, thoroughly washed with deionized water and naturally dried to prevent metal or organic residues; the sampling volume should be controlled at 500 mL to 2 L to meet the analysis requirements and facilitate carrying; the environmental parameters (such as water temperature, pH, and dissolved oxygen) and accurate time should be recorded during each sampling, the instrument should be calibrated daily, and the sampling personnel should wear powder-free gloves and avoid touching the bottle opening with any contaminants to ensure the quality of the sample and the comparability of the data.

[0128] Preferably, the processing of the water sample in the treatment area in step S2 to obtain nitrite concentration, ammonia nitrogen concentration and nitrate concentration data respectively includes:

[0129] When the N-1-naphthyl ethylenediamine spectrophotometric method is used to determine the content of nitrite, the detection wavelength is set to 540 nm, and if the measured concentration is 0.02 mg / L to 1.0 mg / L and the color reaction time is not more than 10 minutes, it is considered as stable nitrite concentration data; if color drift or interference peak shift is greater than ±5 nm, the detection is invalid;

[0130] When the Nash reagent colorimetric method is used to detect the ammonia nitrogen concentration, under the condition of a detection wavelength of 425 nm, if the ammonia nitrogen concentration is stable between 0.1 mg / L and 10 mg / L, and the correlation coefficient R² of the colorimetric absorbance curve and the standard curve is greater than 0.98, it is determined that the detected ammonia nitrogen concentration data is valid; if the concentration exceeds 15 mg / L or the reagent color development is incomplete, it needs to be diluted and measured again;

[0131] When the ultraviolet spectrometry or ion chromatography method is used to determine the nitrate concentration, the concentration range is 0.5 mg / L to 20 mg / L, the absorbance ratio between 220 nm and 275 nm should be greater than 1.5, and the background correction value should not exceed 0.05, which can be confirmed as valid nitrate concentration data; if the deviation of the results of two consecutive detections is greater than ±1.0 mg / L, an abnormality flag is triggered.

[0132] In an embodiment, in actual determination, the concentration range, wavelength and deviation limit value should be strictly controlled according to the method specification: nitrite colorimetry at 540 nm wavelength 0.02-1.0 mg / L, color development <10 min, if the color drift or peak shift is >±5 nm, it is judged as invalid; ammonia nitrogen determination at 425 nm should be maintained at 0.1-10 mg / L, the standard curve correlation coefficient is >0.98, if >15 mg / L or color development is incomplete, dilution and retest are required; when nitrate is quantified, if UV method or ion chromatography is used, the measurement wavelength is 220 nm / 275 nm absorbance ratio must be >1.5, background correction is ≤0.05, and the deviation of two consecutive results is >±1.0 mg / L, which needs to be marked as abnormal and rechecked. All reagents and instruments should be kept clean, calibrated in time, and the retest or dilution record of samples below the detection limit or abnormal value should be well kept.

[0133] Preferably, the training of the TCN model in step S3 includes:

[0134] constructing a TCN model structure, wherein the TCN model structure includes an input layer, a causal convolution layer, a dilated convolution layer, and an output layer, wherein the input layer receives a time series in the standardized model training set, and the shape of the input tensor is (B, T, F), wherein B is the batch size, T is the number of time steps, and F is the number of features;

[0135] based on the causal convolution layer, (K-1) zeros are filled on the left side of the time dimension of the input tensor by zero padding to obtain first model training data, wherein K is the size of the convolution kernel;

[0136] based on the dilated convolution layer, a dilated factor D is introduced to the first model training data to capture the time dependence relationship to obtain second model training data, wherein the dilated factor D represents the interval of the convolution kernel when processing the first model training data;

[0137] the second model training data is subjected to multi-layer convolution and activation transformation through an activation function, and the task type of the transformation result is determined, when the task type is a classification task, a Softmax function is used after the output layer to generate a probability distribution to obtain a probability distribution result, when the task type is a regression task, the output layer directly outputs a predicted value, wherein the shape of the output tensor in the output layer is (B, T, C), wherein C is the number of output channels;

[0138] the probability distribution result or the predicted value is subjected to model precision evaluation according to the model test set to obtain the trained TCN model.

[0139] In an embodiment, referring to Figure 3 , the constructed TCN model structure is as follows:

[0140] 1. Input Layer: The input layer receives time series data, and the shape of the input tensor is (B, T, F), where: B is the batch size, representing the number of time series samples input at once; T is the time steps, representing the length of each time series; F is the number of features, representing the feature dimension of each time step. The input layer converts the original time series data into a multi-dimensional tensor format suitable for convolution operations, ensuring that the model can effectively process features at different time steps in subsequent convolution operations.

[0141] Causal Convolutional Layer: Causal convolution ensures that the model's output at time step t only depends on the current and previous time steps t, t-1, t-2, …, and does not use future time step information. By zero padding (K-1) zeros on the left side of the input tensor's time dimension, where K is the kernel size, the output tensor maintains the same length as the input tensor in the time dimension. In this way, causal convolution can avoid the leakage of future information and effectively learn the historical dependence of time series.

[0142] Dilated Convolutional Layer: Dilated convolution increases the receptive field of the convolution kernel by introducing a dilation factor D, thereby capturing longer temporal dependencies. The dilation factor D represents the interval at which the convolution kernel processes input data, for example, D=1 represents standard convolution, and D=2 represents that the convolution kernel samples every other time step. The dilation factor grows exponentially with network depth, for example: D=1, 2, 4, 8, …, allowing the model to efficiently process long sequence data without sacrificing computational efficiency.

[0143] 2. Output Layer: The output layer is usually a fully connected layer or a convolutional layer that maps the final output of the TCN to the target dimension. The shape of the output tensor is (B, T, C), where C is the number of output channels. For classification tasks, the output layer is followed by a Softmax function to generate a probability distribution; for regression tasks, the predicted value is directly outputted.

[0144] In another embodiment, after data preprocessing is completed, the model training phase follows. The input data is passed into the model (such as Figure 1), and processed through a series of convolutional and dilated convolutional layers. Convolutional layers extract local temporal features, while dilated convolutional layers capture longer-term dependencies by increasing the receptive field. At each layer, convolutional operations are computed over the data by sliding the convolutional kernel, while residual connections are adopted to add the input and output to speed up training and avoid vanishing gradients. Activation functions (such as ReLU) are applied at each layer to help introduce nonlinearity. After multiple layers of convolutional and activation transformations, higher-level features are gradually extracted from the data, and the final result is computed through the output layer. This process enables the model to gradually learn complex patterns and long-term dependencies in time series through the stacking of multiple convolutional layers.

[0145] Preferably, the causal convolutional layer further comprises:

[0146] Each residual block contains two causal convolutional layers, each followed by an activation function and a regularization layer;

[0147] In the first layer of causal convolution, the convolution kernel size is K, the dilation factor is D, and the number of output channels is C; after the convolution output, a ReLU activation function is used to introduce nonlinear features, and its calculation formula is:

[0148] ;

[0149] In the formula, is the input value, is the maximum value function;

[0150] A Dropout regularization layer is set after the activation function with a random dropout ratio of 0.2;

[0151] The convolution kernel size and dilation factor of the second layer of causal convolution are the same as the first layer, i.e., the convolution kernel size K and the dilation factor D do not change;

[0152] The input tensor of the residual block is directly added to the output of the second layer of causal convolution through a skip connection, where the output of the residual block is defined by the following formula:

[0153] ;

[0154] In the formula, is the final output of the residual block, is the input tensor, is the first layer convolution operation, is the application of a nonlinear activation function, is the second layer convolution operation.

[0155] 2. In an embodiment, the residual block: each residual block is composed of two causal convolutional layers, each followed by an activation function (such as ReLU) and a regularization layer (such as Dropout). The residual block directly adds the input to the output through a skip connection, avoiding the problem of gradient disappearance and accelerating model training. The specific structure of the residual block includes the following parts: the input tensor first passes through the first layer of causal convolution, the convolution kernel size is K, the expansion factor is D, and the output channel number is C; then, the ReLU activation function is added after the first layer of convolution to introduce non-linear characteristics, and the formula is: f(x) = max(0, x); Next, the Dropout layer is applied, which randomly discards part of the neurons to prevent overfitting, and in our model, two layers of Dropout = 0.2 are set, next, the input tensor passes through the second layer of causal convolution, and the convolution kernel size and expansion factor are the same as the first layer; finally, the input of the residual block is added to the output of the second convolutional layer to obtain the final output. The output formula of the residual block is: Output = Activation(Conv2(Activation(Conv1(Input)))) + Input.

[0156] Preferably, the task type of the transformed result is determined, including:

[0157] When any of the following conditions occurs, it is determined as a classification task: the label format of the second model training data is discrete category index, and the number of categories is greater than 1; the last dimension channel number C of the output tensor matches the number of categories, and the numerical range is mainly between 0, 1; the tensor characteristics after convolution and activation transformation meet the probability distribution characteristics;

[0158] When the following conditions occur at the same time, it is determined as a regression task: the label of the second model training data is a continuous real number, and the last dimension C of the output tensor is 1; the output value range after convolution and activation transformation is not limited to 0, 1, and specifically fluctuates in the range of ; the output tensor does not meet the probability normalization characteristics.

[0159] In an embodiment, attention should be paid in actual determination: for a classification task, the label must be a discrete index and the number of categories must be greater than one, the output channel number of the model should be consistent with the number of categories, and the output value distribution should be approximately probability (i.e. the sum of each channel is 1 and falls within the range of 0-1); for a regression task, the label is a continuous real number and the output channel is 1, the output value can span a large range (far beyond 0-1), and no probability normalization is performed; in addition, it is necessary to ensure that the convolution and activation layers do not introduce unexpected normalization operations (such as Softmax) to affect the regression scenario, and to verify the consistency of the label before the classification scenario to prevent misjudgment.

[0160] Preferably, the step S4 of calculating the pollutant sea flux prediction value according to the water flow prediction data and the pollutant concentration prediction data comprises:

[0161] The water flux prediction data and the pollutant concentration prediction data are used to calculate the pollutant flux prediction value according to a flux calculation formula, and the flux calculation formula is as follows:

[0162]

[0163] is the instantaneous pollutant flux at time t, is the pollutant concentration prediction data at time t, is the water flux prediction data at time t, is the time t.

[0164] In an embodiment, in the TCN, multiple convolutional layers and dilated convolutional layers are stacked together. The output of each layer is passed to the next layer as input until the last layer. Each layer extracts higher-order features step by step and further expands the receptive field. After all the convolutional layers (including dilated convolution and residual connection), there is usually a fully connected layer or other types of output layer. The role of the output layer is to calculate the final prediction result according to the features learned by the network.

[0165] In another embodiment, referring to Figures 4-6 , the prediction of the flux is obtained by predicting the flux and the pollutant concentration respectively, and then multiplying the prediction results according to the formula: is the instantaneous pollutant flux at time t, is the pollutant concentration prediction data at time t, is the water flux prediction data at time t, is the time t. This method can effectively avoid the complexity of directly predicting the flux, and at the same time, the relationship between the flux and the pollutant concentration is used. The final flux prediction value is obtained by multiplying the predicted flux and the pollutant concentration. In this way, the model captures the change trend of the flux and the pollutant concentration respectively, and then obtains the instantaneous pollutant flux prediction through the multiplication relationship between them.

[0166] Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, and the scope of the present application is defined by the appended claims rather than the above description, and therefore all changes falling within the meaning and scope of the equivalent elements of the application file are intended to be included in the present application.

[0167] ​​​​​​​​​The foregoing is considered as illustrative only of the principles of the application. Numerous modifications and changes will readily occur to those skilled in the art, and it is intended to embrace all such modifications and changes that fall within the scope of the application. Accordingly, the application is not to be restricted in scope to the specific embodiments disclosed herein but is to be accorded the full scope that the principles and novel features request appropriately granted.

Claims

1.A method for monitoring and predicting pollutant flux into the sea based on deep learning, characterized in that, The method comprises the following steps: Step S1: collecting historical water flow data of the target area based on a first time interval; collecting historical pollutant concentration data of the target area based on a second time interval; and determining a pollutant analysis area according to the historical water flow data and the historical pollutant concentration data; Step S2: obtaining a regional water sample based on the pollutant analysis area; processing the regional water sample to obtain nitrite concentration data, ammonia nitrogen concentration data and nitrate concentration data respectively, and combining the historical water flow data as a model training set; wherein the step S2 of obtaining the regional water sample based on the pollutant analysis area comprises: determining the geographical boundary of the pollutant analysis area, marking the sampling point position by using a GPS positioning system, and controlling the sampling point spacing within the range of 50m-200m; selecting the sampling depth according to the water depth and flow conditions of the pollutant analysis area, the sampling depth being 0.3m-0.5m in shallow water area and 1.0m-3.0m in deep water area; when the water flow velocity exceeds 0.5m / s, an anti-disturbance sampler is used to ensure the integrity of the sample; The volume of water sample collection ranges from 500 mL to 2 L. The material of the sampling bottle is polyethylene or glass. The bottle body needs to be pre-acid washed and washed clean with deionized water to avoid secondary pollution. The specific pre-acid washing is to use 1:1 solution to soak for 30 min; recording environmental condition data including water temperature, pH value and dissolved oxygen concentration and adding a sampling time stamp during the sampling process to obtain the regional water sample; Step S3: preprocessing the model training set to obtain a standardized model training set, and taking the real-time collected water flow data as a model test set; training a TCN model by using the standardized model training set and the model test set, and predicting the flow and pollutant concentration of the pollutant analysis area by using the trained TCN model to obtain water flow prediction data and pollutant concentration prediction data; Step S4: calculating the pollutant sea flux prediction value according to the water flow prediction data and the pollutant concentration prediction data to perform the pollutant sea flux monitoring and prediction operation. 2.The deep learning-based monitoring and prediction method for pollutant flux into the sea according to claim 1, characterized in that, The step S1 of determining the pollutant analysis area according to the historical water flow data and the historical pollutant concentration data comprises: taking the maximum value and the minimum value of the historical pollutant concentration data as the center to determine the first monitoring station position and the second monitoring station position respectively which are closest to the maximum value and the minimum value; analyzing the monitoring radius of the monitoring station according to the first monitoring station position and the second monitoring station position to obtain the first monitoring station monitoring radius and the second monitoring station monitoring radius; determining the initial pollutant analysis area between the first monitoring station position and the second monitoring station position based on the first monitoring station monitoring radius and the second monitoring station monitoring radius; calculating the Pearson correlation coefficient between the historical water flow data and the historical pollutant concentration data; comparing the Pearson correlation coefficient with a preset correlation threshold value, and when the Pearson correlation coefficient is greater than or equal to the preset correlation threshold value, determining a third monitoring station position based on the Pearson correlation coefficient, and reselecting the initial pollutant analysis area according to the third monitoring station position to obtain the pollutant analysis area. 3.The deep learning based monitoring and forecasting method of pollutant flux into sea according to claim 2, characterized in that, The step of determining the third monitoring station position based on the Pearson correlation coefficient and reselecting the initial pollutant analysis area according to the third monitoring station position comprises: gridding the initial pollutant analysis area, and determining the Pearson correlation coefficient of each grid unit in the initial pollutant analysis area; Screening the grid cells with Pearson correlation coefficient greater than or equal to a preset correlation threshold to obtain a pollutant transport sensitive area; Determining a third monitoring station position closest to the edge of the pollutant transport sensitive area based on the edge of the pollutant transport sensitive area; Obtaining a section position collected by the third monitoring station according to the third monitoring station position, and calculating a flow velocity gradient of the river section in combination with historical water flow to determine a pollutant migration spatial range; Performing region intersection calculation on the pollutant migration spatial range and the initial pollutant region to be analyzed to obtain the pollutant region to be analyzed. 4.The deep learning-based monitoring and prediction method for pollutant flux into the sea according to claim 2, characterized in that, When the Pearson correlation coefficient is less than the preset correlation threshold, the method further comprises: When the Pearson correlation coefficient is less than the preset correlation threshold, the initial pollutant region to be analyzed determined based on the first monitoring station position and the second monitoring station position is divided into a plurality of grid cells with equal size; For each grid cell, a difference between a historical pollutant concentration average value of the corresponding monitoring station and a global historical average concentration of the target region is calculated to obtain a concentration deviation value of the grid cell; Grid cells with a concentration deviation value greater than a preset deviation threshold in all grid cells are screened as high-concentration candidate cells, wherein the preset deviation threshold is greater than the global average concentration + 2 times the standard deviation; A clustering analysis is performed on the high-concentration candidate cells in space to extract pollutant aggregation sub-regions that are adjacent to each other and have consistent deviations; Based on the extracted pollutant aggregation sub-regions, the outer polygon boundary thereof is used as the pollutant region to be analyzed. 5.The deep learning based monitoring and forecasting method of pollutant flux into sea according to claim 1, characterized in that, The processing of the water samples in the region in step S2 to obtain nitrite concentration, ammonia nitrogen concentration and nitrate concentration data respectively comprises: When N-1-naphthyl ethylenediamine spectrophotometry is used to determine the content of nitrite, the detection wavelength is set to 540 nm, if the measured concentration is 0.02 mg / L-1.0 mg / L, and the color reaction time is not more than 10 minutes, it is considered as stable nitrite concentration data; if color drift or interference peak shift is greater than ± 5 nm, the detection is invalid; When Nash reagent colorimetric method is used to detect ammonia nitrogen concentration, under the condition of determination wavelength 425 nm, if the ammonia nitrogen concentration is stable between 0.1 mg / L and 10 mg / L, and the correlation coefficient R² of the colorimetric absorbance curve and the standard curve is greater than 0.98, it is determined that the effective ammonia nitrogen concentration data is detected; if the concentration exceeds 15 mg / L or the reagent color development is incomplete, it needs to be diluted and measured again; When ultraviolet spectrometry or ion chromatography is used to determine the concentration of nitrate, the concentration range is 0.5 mg / L-20 mg / L, the absorbance ratio between 220 nm and 275 nm is greater than 1.5, and the background correction value should not exceed 0.05, which is considered as valid data of nitrate concentration; if the deviation of the results of two consecutive detections is greater than ± 1.0 mg / L, an abnormality flag is triggered. 6.The deep learning based monitoring and forecasting method of pollutant flux into sea according to claim 1, characterized in that, The training of the TCN model in step S3 using the standardized model training set and the model test set comprises: The TCN model structure is constructed, wherein the TCN model structure comprises an input layer, a causal convolution layer, a dilated convolution layer, and an output layer, wherein the input layer receives a time series in a standardized model training set, and an input tensor has a shape of (B, T, F), wherein B is a batch size, T is a number of time steps, and F is a number of features; Based on the causal convolution layer, K-1 zeros are padded on the left side of the time dimension of the input tensor by zero padding to obtain first model training data, wherein K is a kernel size; Based on the dilated convolution layer, a dilation factor D is introduced to capture time dependence for the first model training data to obtain second model training data, wherein the dilation factor D represents an interval of the convolution kernel when processing the first model training data; The second model training data is subjected to multi-layer convolution and activation transformation through an activation function, and a task type of a transformation result is determined, when the task type is a classification task, a Softmax function is used after the output layer to generate a probability distribution to obtain a probability distribution result, when the task type is a regression task, a predicted value is directly output by the output layer, wherein an output tensor in the output layer has a shape of (B, T, C), wherein C is a number of output channels; The model accuracy of the probability distribution result or the predicted value is evaluated according to a model test set to obtain a trained TCN model. 7.The deep learning based monitoring and forecasting method of pollutant flux into sea according to claim 6, characterized in that, The causal convolution layer further comprises: Each residual block contains two causal convolution layers, each of which is followed by an activation function and a regularization layer; In the first layer causal convolution, the kernel size is K, the dilation factor is D, and the number of output channels is C; after convolution, a ReLU activation function is used to introduce nonlinear features, and the calculation formula is: ； wherein is the input value, is the maximum function; A Dropout regularization layer is set after the activation function, and the random discard ratio is 0.2; The kernel size and the dilation factor of the second layer causal convolution are the same as those of the first layer, that is, the kernel size K and the dilation factor D do not change; The input tensor of the residual block is directly added to the output of the second layer causal convolution through a skip connection, wherein the output of the residual block is defined by the following formula: ； wherein is the final output of the residual block, is the input tensor, is the first layer convolution operation, is the application of a non-linear activation function, is the second layer convolution operation. 8.The deep learning based monitoring and forecasting method of pollutant flux into sea according to claim 6, characterized in that, The task type of the transformation result includes: When any of the following conditions occurs, it is determined to be a classification task: the label format of the second model training data is a discrete category index, the number of categories is greater than 1; the last dimension of the output tensor C matches the number of categories, and the numerical range is between 0 and 1; the tensor features after convolution and activation transformation satisfy the probability distribution features; When the following conditions occur simultaneously, the regression task is determined: the label of the second model training data is a continuous real value, and the last dimension C of the output tensor is 1; the output value range after convolution and activation transformation is not limited to 0, 1, specifically fluctuating in the range of -10 6 ~ 10 6 ; the output tensor does not satisfy the probability normalization feature. 9.The deep learning based monitoring and forecasting method of pollutant flux into sea according to claim 1, characterized in that, The calculation of the pollutant sea flux prediction value from the water flow prediction data and the pollutant concentration prediction data in step S4 includes: The pollutant sea flux prediction value of the water flow prediction data and the pollutant concentration prediction data is calculated according to the sea flux calculation formula, wherein the sea flux calculation formula is as follows: ； wherein is the instantaneous flux of the pollutant into the water body at the time instant, is the pollutant concentration prediction data at the time instant, is the water body at the time instant Flow prediction data, is the time.

Images