Coastal river network pollutant evidence fusion early warning method based on adaptive sliding window LSTM

Abstract

The invention discloses a coastal river network pollutant evidence fusion early warning method based on an adaptive sliding window LSTM, and the method comprises the steps: firstly obtaining a feature sequence of a pollutant index through a sensor, constructing an attention mechanism long-short term memory network model AM-LSTMi related to the pollutant index, and obtaining a feature prediction value of the pollutant index according to the feature sequence; secondly, an adaptive sliding window adjustment model is constructed, the sliding window size of AM-LSTMi at the t + 1 moment is obtained, and iterative optimization is carried out on parameters of the sliding window adjustment model and the AM-LSTMi model; and finally, constructing a pollution level fuzzy membership model related to the pollutant indexes, obtaining reliability distribution of pollutant levels according to the feature prediction values of the pollutant indexes, carrying out weighted fusion, and carrying out over-limit alarm according to the reliability distribution obtained after fusion. According to the method, the accuracy of feature sequence prediction is improved, and river network pollutant early warning is accurately and efficiently completed.

Description

Technical Field

[0001] The present invention belongs to the field of river network water pollution control, and relates to a coastal river network pollutant evidence fusion early warning method based on an adaptive sliding window LSTM. Background Art

[0002] If environmental pollution exceeds the environmental capacity, it will cause a devastating blow to the local ecosystem, and with the circulation of materials and energy, it will pose a huge threat to people's lives, which poses a great challenge to environmental law enforcement departments. Traditionally, it relies on manual monitoring and experience judgment, which has problems such as poor timeliness and low accuracy.

[0003] Therefore, it is particularly important to predict the concentration of pollutants in river networks in real time and accurately assess their pollution levels to prevent pollutants from causing further damage to the river network ecosystem. With the continuous development of artificial intelligence technology, the use of big data analysis and advanced artificial intelligence algorithms to predict the concentration of pollutants in river networks has important significance and value. Summary of the invention

[0004] In view of the shortcomings of the existing technology, the present invention proposes a pollutant evidence fusion early warning method for coastal river networks based on adaptive sliding window LSTM.

[0005] The present invention comprises the following steps:

[0006] (1) In the coordinated monitoring system of pollutants in the coastal river network, the COD concentration (x) at the monitoring point at time t is obtained online through sensors. 1 ), soil pH value (x 2 ) and chlorophyll a concentration (x 3 ) three pollutant index data, pre-process the collected data, and obtain the characteristic sequence X of the i-th (i=1,2,3) pollutant index respectively. i .

[0007] (2) Constructing the pollutant index x i Attention mechanism long short-term memory network model (AM-LSTM i ), whose input is the feature sequence X i , η i The statistical characteristics of the (t) characteristic values ​​and the characteristic values ​​at the current time t are output as the characteristic prediction value at the future time t+1, based on which the pollutant index x can be obtained. i The predicted value of the feature.

[0008] (3) Constructing AM-LSTM i Adaptive sliding window adjustment model (ER i ), based on the feature prediction value at time t, the output is the AM-LSTM at time t+1 i The sliding window size ηi (t+1) and used in step (2).

[0009] (4) Using historical feature samples to analyze ER i Model and AM-LSTM i The parameters of the model are iteratively optimized to make AM-LSTM i The predicted results are close to the true value.

[0010] (5) Constructing the pollutant index x i The fuzzy membership model of pollution level (FL i ), the iterated AM-LSTM i For x i The feature prediction value is brought into the model FL i , obtain the confidence distribution about the pollutant level.

[0011] (6) The pollutant index COD concentration (x 1 ), soil pH value (x 2 ) and chlorophyll a concentration (x 3 )’s pollutant level confidence distribution is weightedly fused, and an over-limit alarm is issued based on the confidence distribution obtained after fusion.

[0012] The pollutant evidence fusion early warning method for coastal river network based on adaptive sliding window LSTM proposed in this paper first obtains the pollutant index data of the monitoring point in real time from the collaborative monitoring system of pollutants in coastal river network, and obtains the pollutant index x after preprocessing. i Feature sequence. AM-LSTM is constructed for each pollutant indicator. i Model, and introduce ER i , to dynamically adjust AM-LSTM i The sliding window size of the model; then construct FL i The obtained pollutant index prediction sequence is converted into the confidence distribution of pollution level. Finally, the confidence distribution of pollution level of different pollutant indicators is weighted and fused, and timely alarm is realized for the concentration exceeding the limit of pollutant indicators based on the fused confidence distribution.

[0013] The present invention provides a coastal river network pollutant evidence fusion early warning method based on adaptive sliding window LSTM, which has the following beneficial effects:

[0014] 1. The present invention integrates the attention mechanism with the LSTM model to enhance the LSTM model's ability to identify key information in feature sequences. Through this integration, the LSTM model can more accurately mine important historical information in feature sequences and use this information to improve the accuracy of feature sequence prediction.

[0015] 2. The present invention dynamically infers the input sliding window size suitable for the AM-LSTM model through the ER model. This process helps the model to adapt to different data features more flexibly, thereby improving the prediction accuracy.

[0016] 3. Based on the historical characteristic sample data, the present invention obtains the initial reference evidence matrix table about the input and output of the ER model to describe the nonlinear relationship between the input and output, and uses the sequential linear programming method to optimize and update the relevant parameters in the model, so that the prediction accuracy of AM-LSTM is further improved, and the pollutant early warning of the coastal river network is completed accurately and efficiently. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. By referring to the drawings, the features and advantages of the present invention will be more clearly understood. The drawings are schematic and should not be understood as limiting the present invention in any way. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them:

[0018] Figure 1 is a flowchart of the method of the present invention;

[0019] Figure 2 It is a pollution level map after the pollutant indicators are integrated. DETAILED DESCRIPTION

[0020] The embodiment of the present application proposes a coastal river network pollutant evidence fusion early warning method based on an adaptive sliding window LSTM, and the method flow chart is as follows: Figure 1 As shown, it includes the following steps:

[0021] (1) In the coordinated monitoring system of pollutants in the coastal river network, the COD concentration (x) at the monitoring point at time t is obtained online from the sensor. 1 ), soil pH value (x 2 ) and chlorophyll a concentration (x 3 ) three pollutant index data, pre-process the collected data, and obtain the characteristic sequence X of the i-th (i=1,2,3) pollutant index respectively. i ;

[0022] (2) Constructing the pollutant index x i AM-LSTM i The model takes the sequence X as input i The statistical characteristics of the feature and the feature value at the current time t are output as the feature prediction value at the future time t+1;

[0023] (3) Constructing AM-LSTM i ERi The model takes the sequence X as input i The statistical characteristics of the feature prediction value and the error between the true value at time t, the output is the AM-LSTM at time t+1 i The sliding window size η i (t+1);

[0024] (4) Using historical feature samples to analyze ER i Model and AM-LSTM i The parameters of the model are iteratively optimized to make AM-LSTM i The prediction result is close to the true value;

[0025] (5) Constructing the pollutant index x i FL i Model, x i The feature prediction value is brought into FL i Model, obtain the confidence distribution about the pollutant level;

[0026] (6) x 1 、x 2 、x 3 The confidence distribution of the pollutant levels is weightedly fused, and an over-limit alarm is issued based on the confidence distribution obtained after the fusion.

[0027] The specific steps of step (1) are as follows:

[0028] (1-1) Receive raw data from monitoring points, including COD concentration in river water (x 1 ), soil pH value (x 2 ) and chlorophyll a concentration (x 3 ), identify and mark outliers and missing values ​​in the data, correct or delete outliers according to preset thresholds, and use interpolation methods to fill in missing values.

[0029] (1-2) The processed data is used as the characteristic data of the pollutant index, and the characteristic sequence X of the pollutant index is obtained according to the sampling time. i ={x i (1),...,x i (t),...,x i (T)|i=1,2,3}.

[0030] The specific steps of step (2) are as follows:

[0031] (2-1) AM-LSTM constructed by attention mechanism AM and LSTM i Model, set the parameters, set the number of hidden layers of the model to 1, the number of hidden nodes to 64, the learning rate to 0.008, and the number of iterations to 300.

[0032] (2-2) Obtain the feature sequence X according to step (1) i ={x i (1),...,x i (t),...,x i (T)|i=1,2,3}, AM-LSTM can be determined i The input vector L of the model i (t) = [L i,1 (t),L i,2 (t),L i,3 (t)], the formula is as follows:

[0033] L i,1 (t) = x i (t) (1)

[0034]

[0035]

[0036] Among them, η i (t) is the adaptive sliding window size at time t, which is a positive integer greater than or equal to 3, s is the sample standard deviation, and T is the length of the sequence.

[0037] (2-3) Using attention mechanism for LSTM i The output of the layer is weighted to extract the most relevant features, LSTM i The output of the layer is h i,s =[h i,1 ,...,h i,t ,...,h i,T ], where h i,t is the hidden state vector at time t, h i,s is the hidden state at each sampling moment of the input sequence.

[0038] (2-4) Calculate the attention weight at each sampling moment, the formula is as follows:

[0039]

[0040] Among them, h i,t is the hidden state at the target time step and W is the learnable weight matrix.

[0041] (2-5) The calculated attention weights are normalized and the features output by the LSTMi layer are weighted and combined using the attention weights. The formula is as follows:

[0042] α i,t =softmax(score(hi,t ,h i,s )) (5)

[0043]

[0044] Among them, α i,t is the attention weight at time t, and softmax is the normalized exponential function.

[0045] (2-6) o i Used for full connection layer processing and predicting the output feature prediction value at time t+1

[0046] For ease of understanding, steps (2-2) to (2-6) are described here as an example. The process is as follows:

[0047] AM-LSTM i An input sample of the model is the vector of the pollutant index feature sequence at time t in step (1) and the AM-LSTM in step (3). i Adaptive sliding window adjustment model ER i The inferred adaptive sliding window size η at time t i (t) is a feature vector. The sliding window size η is set i The possible values ​​are 3, 4, and 5, which are used to construct AM-LSTM. i The first input sample of the model needs to be at least at time t = 6, setting η i (6) = 5. From formulas (1), (2) and (3), we can get L i,1 (6) = x i (6) Therefore, we can get AM-LSTM i The first set of input vectors L of the model i (6) = [L i,1 (6),L i,2 (6),L i,3 (6)], according to this step, all input vectors can be obtained; the obtained input vectors are input into the LSTM model, and all hidden states h can be obtained in the LSTM layer. i,s =[h i,6 ,...,h i,t ,...,h i,T ], and the attention mechanism is used to perform weighted processing, and we can get Normalizing the obtained attention weights, we can get o i For subsequent LSTM i After fully connected layer processing, the sequence prediction value is obtained.

[0048] The specific steps of step (3) are as follows:

[0049] (3-1) Constructing AM-LSTM i ER i The model has an input vector E i (t) = [E i,1 (t),E i,2 (t),E i,3 (t)], the calculation formula is as follows:

[0050]

[0051]

[0052]

[0053] in, It is AM-LSTM i The model's feature prediction value at time t+1; and according to ER i The input of the model can get the AM-LSTM at time t+1 i The sliding window size η i (t+1).

[0054] (3-2) Setting ER based on historical feature samples i Model input vector E i The corresponding reference vector is

[0055]

[0056] At the same time, set ER i Model output η i The reference value is D = {D k |k=1,2,...,K},D k The value is a positive integer greater than or equal to 3, and D 1 <D 2 ...<... <D K , K≥3 is the number of reference values ​​output.

[0057] (3-3) Obtain the corresponding initial reference evidence matrix based on historical feature sample data to describe ER i Model input vector E i and output η i The nonlinear mapping relationship between them. The initial reference evidence matrix is ​​shown in Table 1, where Represents the input reference vector corresponding evidence; Indicates that when the model input vector E i for When the model outputs ηi D k The reliability, satisfaction And there is

[0058] Table 1 Input vector E i Initial reference evidence matrix

[0059]

[0060]

[0061] (3-4) Obtain the input vector E i The corresponding L=J 1 ×J 2 ×J 3 Reference vector The Euclidean distance between them is normalized and used as the activation weight ω of the reference evidence. l ,l=1,2,...,L, the calculation is as follows:

[0062]

[0063] (3-5) Obtain the activation weight ω of the reference evidence l Finally, the L reference evidences in the reference evidence matrix table are fused using the evidence reasoning rules, and the fusion result is:

[0064]

[0065]

[0066] In formulas (11) and (12): Reference evidence Indicates that when the model input vector E i When it is the lth reference vector, the model outputs η i D k The confidence of the kth reference vector.

[0067] (3-6) Estimate the ER at time t+1 based on the fusion result obtained in step (3-5) i The output of the model is i (t+1),

[0068] The calculation formula is as follows:

[0069]

[0070] In formula (13): D k Output η for the model i The reference value of is the confidence in the model output.

[0071] For ease of understanding, steps (3-2) to (3-6) are described here as an example. The process is as follows:

[0072] Here, the pollutant index x 1 Take ER as an example. 1 Model Input E 1,1 , E 1,2 and E 1,3 The reference value sets are and Set ER at the same time 1 The output of the model is 1 The reference value is D = {3, 4, 5}, and the initial reference evidence matrix can be obtained based on the historical feature sample data. At t = 6, ER 1 Model Input E 1,1 (6), E 1,2 (6) and E 1,3 (6), we can get the model input vector E 1 (6) = [0.0010, -0.1621, 0.0002], input vector E 1 (6) Activate all the evidence in the reference evidence matrix. According to steps (3-4), the activation weights of the model input activation of each evidence can be obtained. The activated evidence is fused according to formulas (11) and (12). The fusion result is [0.2974, 0.3470, 0.3556]. The model output η is calculated by formula (13): 1 (7)=4.

[0073] The specific steps of step (4) are as follows:

[0074] (4-1) Use historical characteristic samples to build a parameter optimization model, and compare the actual observed values ​​of pollutant index data with the ER-based i AM-LSTM i The mean absolute percentage error (MAPE) between the predicted values ​​generated by the model is used as the objective function, which is as follows:

[0075]

[0076] (4-2) Based on the parameter optimization model established in step (4-1), ER i The model parameters are optimized in real time, and the optimized parameters are uploaded to achieve dynamic updating of the reference evidence matrix table, and the updated reference evidence matrix table is used as the initial parameters of the model at time t+1.

[0077] (4-3) Repeat the above steps, iterate in sequence, and implement AM-LSTM by online optimization and updating of model parameters.i The model's adaptive sliding window adjustment improves the model's predictive ability.

[0078] For ease of understanding, the above pollutant index x is used. 1 , the specific optimization process of the optimization model is explained at time t = 6. The optimization model will ER 1 The elements in the belief matrix table in the inference model are used as parameter optimization objects, and have the characteristics of online optimization and real-time parameter updating.

[0079] For ER at t = 6 1 Reasoning model input, activating ER 1 All the evidence in the confidence matrix table in the model. At this time, the optimization model optimizes the confidence of all activated evidence. After the optimization is completed, upload the ER 1 The elements in the reliability matrix table in the model are updated accordingly to obtain a new reliability matrix table, as shown in Table 2, and it is used as the ER at the next moment 1 The reliability matrix of the model.

[0080] Table 2 Input vector E 1 Reference Evidence Matrix

[0081]

[0082]

[0083] The specific steps of step (5) are as follows:

[0084] (5-1) AM-LSTM after iterating through step 4-3 i Obtain characteristic prediction sequence of pollutant indicators

[0085]

[0086] (5-2) Constructing the pollutant index x i FL i Model, river water pollution level set V i ={I,II,III}, “I” represents the low level of river pollution, “II” represents the medium level of river pollution, and “III” represents the high level of river pollution.

[0087] Establish pollutant index x i The membership function for level “I” is denoted as And set the upper limit of the fuzzy threshold to b i , the lower limit of the fuzzy threshold is a i , and a i i , the formula is as follows:

[0088] ​

[0089] Establish pollutant index x i The membership function for level “II” is denoted as And set the upper limit of the fuzzy threshold to c i , the fuzzy threshold value is b i , the lower limit of the fuzzy threshold is a i , and a i i <c i , the formula is as follows:

[0090]

[0091] Similarly, establish the pollutant index x i The membership function for level “III” is denoted as And set the upper limit of the fuzzy threshold to c i , the lower limit of the fuzzy threshold is b i , and b i <c i , the formula is as follows:

[0092]

[0093] (5-3) x i The feature prediction value is brought into the model FL i , the confidence distribution of pollutant level is obtained as

[0094]

[0095] For ease of understanding, steps (5-1) to (5-3) are described here by way of example. The process is as follows:

[0096] Pollutant index x 1 Take t=7 as an example to get the predicted value The predicted value Bring in the pollution level fuzzy membership model FL we constructed 1 Among them, a 1 =8, b 1 =8.8, c 1 =9.6, we can get the confidence distribution of the pollutant level at time t=7

[0097] The specific steps of step (6) are as follows:

[0098] The confidence distribution of the three pollutant levels is obtained and the pollution level is determined by weighted fusion, with the weights r 1 、r 2 、r 3 , and satisfies r​1 +r 2 +r 3 =1, the confidence distribution of the three pollutant indicators is weighted and integrated according to their weights, and the formula is as follows:

[0099] B o =r 1 ×B(x 1 )+r 2 ×B(x 2 )+r 3 ×B(x 3 ) (18)

[0100] According to the fused reliability distribution B o , determine the final pollutant level, and issue an alarm if the pollutant level is "III".

[0101] The following is a detailed description of the embodiments of the method of the present invention with reference to the accompanying drawings:

[0102] 1. In the coordinated monitoring system of pollutants in the coastal river network, the COD concentration (x 1 ), soil pH value (x 2 ) and chlorophyll a concentration (x 3 ) three pollutant index data, pre-process the collected data, and obtain the characteristic sequence X of the i-th (i=1,2,3) pollutant index respectively. i ={x i (1),...,x i (t),...,x i (T)|i=1,2,3}, T=48.

[0103] 2. Here, the pollutant index x 1 As an example, first of all, AM-LSTM 1 The model parameters are set, the number of hidden layers is set to 1, the number of hidden nodes is set to 64, the learning rate is set to 0.008, and the number of iterations is set to 300; and according to the feature sequence X 1 , get AM-LSTM 1 The input vector of the model, in this embodiment, the sliding window size η is set i The possible values ​​are 3, 4, and 5. Set η 1 (6) = 5, AM-LSTM 1 The first input sample of the model, at time t = 6, can be obtained from formulas (1), (2) and (3) 1,1 (6) = x 1 (6) Get AM-LSTM 1 The first set of input vectors L of the model 1(6) = [7.3536, 0.0005, -0.2183], and the predicted value at t = 7 is obtained

[0104]

[0105] 3. ER can be obtained through step (3) 1 The model inputs vector E at time t=6 1 (6)

[0106] E 1 (6) = [0.0005, -0.2183, 0.0838]. Through the initial evidence inference table, the input vector E 1 (6) Activate all the evidence in the reference evidence matrix. According to steps (3-4), the activation weights of the model input activation of each evidence can be obtained. The activated evidence is fused according to formulas (11) and (12). The fusion result is [0.2974, 0.3470, 0.3556]. The model output η is calculated by formula (13): 1 (7)=4.

[0107] 4. Build a parameter optimization model to convert ER 1 Elements in the belief matrix table and AM-LSTM in the inference model 1 The other parameters of the model are optimized and have the characteristics of online optimization and real-time parameter update. 1 Reasoning model input, activating ER 1 All the evidence in the confidence matrix table in the model. At this time, the optimization model optimizes the confidence of all activated evidence. After the optimization is completed, upload the ER 1 The elements in the credibility matrix table in the model are updated accordingly to obtain a new credibility matrix table, as shown in Table 3, and it is used as the ER at the next moment 1 The reliability matrix of the model;

[0108] Table 3 Input vector E 1 Reference Evidence Matrix

[0109]

[0110]

[0111] For pollutant index x 2 , x 3 Following the above steps, its feature prediction sequence can be obtained.

[0112] 5. According to the above steps, the characteristic prediction sequence of pollutant indicators can be obtained And the FL constructed by step (5) i The model obtains the confidence distribution about the pollutant level, as shown in Table 4:

[0113] Table 4 Pollutant level reliability distribution table

[0114]

[0115]

[0116] 6. The confidence distribution of the pollutant level obtained above is used to determine the pollution level through weighted fusion, and the weights are set as r 1 、r 2 、r 3 , and satisfies r 1 +r 2 +r 3 =1, the confidence distribution of the three pollutant indicators is weighted and integrated according to their weights, and the formula is as follows:

[0117]

[0118] According to the fused reliability distribution B o , determine the final pollutant level, if the pollutant level is "III" an alarm is issued, an alarm is issued at t = 36, the result is as follows Figure 2 shown.

Claims

1. A pollutant evidence fusion early warning method for coastal river network based on adaptive sliding window LSTM, characterized in that: The following steps are involved: Step 1: Obtain pollutant index data of coastal river network through sensors and preprocess them to obtain the characteristic sequence X of the i-th pollutant index i ; Step 2: Construct the pollutant index x i Attention mechanism long short-term memory network model AM-LSTM i , according to the characteristic sequence X i , get the pollutant index x i The feature prediction value of Step 3: Build AM-LSTM i Adaptive sliding window adjustment model ER i , based on the feature prediction value at time t, we get the AM-LSTM at time t+1 i The sliding window size η i (t+1); Step 4: Use historical feature samples to analyze ER i Model and AM-LSTM i The parameters of the model are iteratively optimized to make AM-LSTM i The prediction result is close to the true value; Step 5: Construct the pollutant index x i Fuzzy membership model of pollution level FL i , the iterated AM-LSTM i For x i The feature prediction value is brought into the model FL i , obtain the confidence distribution of pollutant levels; Step 6: Perform weighted fusion on the pollutant level confidence distribution of each pollutant indicator, and issue an over-limit alarm based on the confidence distribution obtained after fusion.

2. The method for fusion early warning of pollutants in coastal river networks based on adaptive sliding window LSTM according to claim 1 is characterized in that: The specific implementation process of step 1 is as follows: Step 1-1, receiving the original data from the monitoring points, including COD concentration in river water x1, soil pH value x2 and chlorophyll a concentration x3, identifying and marking the outliers and missing values ​​in the data, correcting or deleting the outliers according to the preset threshold, and using the interpolation method to complete the missing values; Step 1-2: Use the processed data as the characteristic data of the pollutant index, and obtain the characteristic sequence X of the pollutant index according to the sampling time. i ={x i (1),...,x i (t),...,x i (T)|i=1,2,3}.

3. The method for early warning of pollutant evidence fusion in coastal river network based on adaptive sliding window LSTM according to claim 2 is characterized in that: The specific implementation process of step 2 is as follows: Step 2-1: AM-LSTM constructed by attention mechanism AM and LSTM i Model, set parameters; Step 2-2: Obtain feature sequence X i ={x i (1),...,x i (t),...,x i (T)|i=1,2,3}, determine AM-LSTM i The input vector L of the model i (t) = [L i,1 (t),L i,2 (t),L i,3 (t)], as follows: L i,1 (t)=x i (t) Where, is the adaptive sliding window size at time t, which is a positive integer greater than or equal to 3, s is the sample standard deviation, T is the length of the sequence, and η i (t) is the number of statistical characteristics of eigenvalues; Step 2-3: Use attention mechanism for LSTM i The output of the layer is weighted, LSTM i The output of the layer is h i,s =[h i,1 ,...,h i,t ,...,h i,T ], where h i,t is the hidden state vector at time t, h i,s is the hidden state at each sampling moment of the input sequence; Step 2-4: Calculate the attention weight at each sampling moment. The formula is as follows: Among them, h i,t is the hidden state at the target time step, and W is the learnable weight matrix; Step 2-5: Normalize the attention weights and use the attention weights to weightedly combine the features output by the LSTMi layer to obtain feature o i ; Step 2-6, o i Used for full connection layer processing and predicting the output feature prediction value at time t+1 4. The method for fusion early warning of pollutants in coastal river networks based on adaptive sliding window LSTM according to claim 3 is characterized in that: The specific implementation process of step 3 is as follows: Step 3-1: Build AM-LSTM i ER i The model has an input vector E i (t) = [E i,1 (t),E i,2 (t),E i,3 (t)], the calculation formula is as follows: in, It is AM-LSTM i The feature prediction value of the model at time t+1; Step 3-2: Set ER based on historical feature samples i Model input vector E i The corresponding reference vector is At the same time, set ER i Model output η i The reference value is D = {D k |k=1,2,...,K},D k The value is a positive integer greater than or equal to 3, and D1 <D2...<...<D K , K≥3 is the number of reference values ​​output; Step 3-3: Obtain the corresponding initial reference evidence based on historical feature sample data and describe ER i Model input vector E i and output η i The nonlinear mapping relationship between them; Step 3-4: Obtain the input vector E i The corresponding L=J1×J2×J3 reference vectors The Euclidean distance between them is normalized and used as the activation weight ω of the reference evidence. l ,l=1,2,...,L; Step 3-5: Get the activation weight ω of the reference evidence l After that, the Lth reference evidence is fused using the evidence reasoning rules, and the fusion result is: In the formula, reference evidence Indicates that when the model input vector E i When it is the lth reference vector, the model outputs η i D k The reliability of the kth reference vector; Step 3-6: Estimate the ER at time t+1 based on the fusion results obtained in step 3-5 i The output of the model is i (t+1), calculated as follows: Where D k For ER i Model output η i reference value.

5. The method for fusion early warning of pollutants in coastal river networks based on adaptive sliding window LSTM according to claim 4 is characterized in that: The specific implementation process of step 4 is as follows: Step 4-1: Use historical characteristic samples to build a parameter optimization model, and compare the actual observed values ​​of pollutant index data with the ER-based i AM-LSTM i The mean absolute percentage error MAPE between the predicted values ​​generated by the model is used as the objective function; Step 4-2: Optimize the model based on the parameters and i The model parameters are optimized in real time, and the optimized parameters are uploaded to achieve dynamic update of reference evidence, and the updated reference evidence is used as the initial parameters of the model at time t+1; Step 4-3: Repeat steps 4-1 and 4-2, iterate in sequence, and implement AM-LSTM by online optimization and updating of model parameters. i The model's adaptive sliding window adjustment improves the model's predictive ability.

6. The method for fusion early warning of pollutants in coastal river networks based on adaptive sliding window LSTM according to claim 5 is characterized in that: The specific implementation process of step 5 is as follows: Step 5-1: AM-LSTM after iterating through step 4-3 i Obtain characteristic prediction sequence of pollutant indicators Step 5-2: Construct the pollutant index x i FL i Model, river water pollution level set V i ={I,II,III}, where I, II and III represent the river water pollution levels of low, medium and high, respectively; Establish pollutant index x i The membership function of level I is denoted as And set the upper limit of the fuzzy threshold to b i , the lower limit of the fuzzy threshold is a i , and a i i , the formula is as follows:​ Establish pollutant index x i The membership function of level II is denoted as And set the upper limit of the fuzzy threshold to c i , the fuzzy threshold value is b i , the lower limit of the fuzzy threshold is a i , and a i i <c i , the formula is as follows:​ Similarly, establish the pollutant index x i The membership function of level III is denoted as And set the upper limit of the fuzzy threshold to c i , the lower limit of the fuzzy threshold is b i , and b i <c i , the formula is as follows: Step 5-3, x i The feature prediction value is brought into the model FL i , the confidence distribution of pollutant level is obtained as 7. The method for early warning of pollutant evidence fusion in coastal river network based on adaptive sliding window LSTM according to claim 6 is characterized in that: The specific implementation process of step 6 is as follows: The reliability distribution of the three pollutant levels is set with weights r1, r2, and r3, and r1+r2+r3=1. The reliability distribution of the three pollutant indicators is weighted and fused according to their weights. According to the fused reliability distribution B o Determine the final contamination level and issue an alarm if the contamination level is III.

Images