Advanced early warning method for comprehensive monitoring of industrial pollution discharge

Abstract

The invention provides an advanced early warning method for comprehensive monitoring of industrial pollution discharge, and aims to solve the problems that an online monitoring method in the prior art generally adopts tail end discharge monitoring and only can monitor a few key indexes; the method mainly comprises the following four steps: establishing a learning model; training to obtain an enhanced model; generating a comprehensive model according to a prediction accuracy weight matrix; and performing clustering algorithm calculation. Compared with the prior art, the method has the advantages that the learning model is established based on the LSTM, the matching degree of the model to an actual scene is improved by gradually enhancing a single factor, a collaborative factor and traffic data, and finally, the missing data aggregation degree is extracted through a clustering algorithm to make up for the limitation that the model cannot learn missing data and cannot be predicted; the purpose of predicting the online monitoring data in the next 24 hours by inputting the online monitoring data of pollutants in the previous 24 hours is achieved, and therefore a comprehensive online monitoring early warning method is provided for productive enterprises.

Description

Technical Field

[0001] The present invention relates to the technical field of sewage discharge monitoring, and in particular to an advance warning method for comprehensive monitoring of industrial sewage discharge. Background Art

[0002] At present, the situation of ecological environmental protection is becoming more and more severe, and the results of environmental protection are gradually emerging. The natural environment in which people live is getting better and better. In particular, productive enterprises have the greatest impact on the environment. Therefore, for key polluting productive enterprises, according to relevant regulations, key enterprises must install online monitoring to monitor the fluctuations of pollution factors in real time.

[0003] The existing problem is that the traditional online monitoring method uses the monitoring of discharged water and discharged gas, and its monitoring method is usually end-of-pipe emission monitoring. Therefore, only a few indicators can be monitored in real time, and the concentration of monitoring factors cannot be known in advance, resulting in a lack of decision-making basis and data support for the improvement of industrial pollution production and treatment processes. Summary of the invention

[0004] In response to the shortcomings in the above-mentioned prior art, the present application provides an advance warning method for comprehensive monitoring of industrial pollution discharge. A learning model is established based on LSTM. By gradually enhancing single factors, synergistic factors and flow data, the matching degree of the model to the actual scenario is improved. Finally, the clustering algorithm is used to refine the clustering degree of missing data to make up for the limitation of the model that cannot learn missing data and cannot predict. Finally, the goal of predicting the online monitoring data of the next 24 hours by inputting the online monitoring data of pollutants in the previous 24 hours is achieved, providing a comprehensive online monitoring and early warning method for productive enterprises.

[0005] The technical solution adopted by the present invention is as follows:

[0006] An advance warning method for comprehensive monitoring of industrial sewage discharge comprises the following steps:

[0007] Step S1: Based on the LSTM model, at least three learning models containing different numbers of factors are established, namely a basic learning model, a first learning model and a second learning model;

[0008] Step S2: using the online monitoring hourly data of pollution sources in the most recent year as learning data, the three learning models are respectively subjected to input, calculation, memory and output processes, and three enhanced models are obtained by training, which correspond to the basic model, the first enhanced model and the second enhanced model respectively;

[0009] Step S3: randomly select 100 groups of data and input them into the three enhanced models in step S2 for prediction to obtain a prediction accuracy weight matrix, and generate a comprehensive model, the prediction value of the comprehensive model is the average of the prediction values ​​of the three models;

[0010] Step S4: Finally, the learning data is calculated using a clustering algorithm to achieve the integrity of the continuous cycle of the predicted data.

[0011] Furthermore, the number of factors in the basic learning model is single, and a pollution source emission factor sequence is constructed based on the LSTM model, where the factor x t For smoke or ammonia nitrogen, the sequence X = (x1, x2, ..., x t ,…)T, the input of each time step can be expressed as x t , use sliding window to divide the sequence into several windows of size L, step size step, when the data is divided to the end, if it is less than L and cannot form a window, the missing data is filled with 0; the pollution source that continuously emits 24 hours a day has 24 hours of time series data input, that is Set the window size to 12 and the step size to 10. We divide the data into two windows, namely

[0012]

[0013] Assuming that a year has 365 days, this model prepares the online monitoring hourly data of pollution sources in the most recent year as learning data, and finally forms X=(x1,x2,…,x 730 )T.

[0014] Furthermore, the input training steps of the basic learning model are as follows: the input of the model unit includes the input at the current moment The output state ht-1 at the previous moment and the unit state ct-1 at the previous moment are initialized to h0 and c0 when the first layer LSTM unit is operated. In the subsequent LSTM units, ht-1 and ct-1 are obtained from the previous LSTM unit, where xt, ht-1, and ct-1 represent the input information at the current moment, the output information at the previous moment, and the memory information at the previous moment, respectively; m is the window size (length) after processing the input sequence, h t-1 is the output state at the previous moment, with a shape of h t-1∈R d×1, d is the hidden state size of the LSTM unit, and ct-1 is the unit state at the previous moment, with a shape of ct-1∈R d×1, which has the same shape as h t-1.

[0015] h t-1 and Concatenate to form longer vectors Right now As shown in the following formula, then It will be passed into each gate, and the formula is as follows:

[0016]

[0017] Furthermore, the computational training steps of the basic learning model are: the input of each gate is Will Multiply the weight matrix W in the gate and add the bias value b to get the intermediate result M, then take Sigmoid (loss function value range; [0,1]) for M to get the output gt of the gate, which has the same shape as the unit state ct, that is, gt∈R d×1

[0018]

[0019] in, W f , W i , W o ∈R d×(d+m), b f 、b i , b0∈R d×1,f t 、i t , o t ∈R d×1.

[0020] Furthermore, the memory training steps of the basic learning model are: calculate the current input unit state The value of indicates the amount of current input to be retained in memory. and the weight matrix W of the current state unit c Multiply and add the bias value bc to get the intermediate result Mc, then take tanh on Mc to get the output ct, the formula is as follows:

[0021]

[0022] in, W c ∈R d×(d+m), b c ∈R d×1,

[0023] Next, the current cell state c t The calculation of the forget gate and the input gate obtains the result f t 、i t and the unit state c at the previous moment t-1 To calculate the current cell state c t , respectively f t 、c t-1 Element-wise multiplication, i t and Multiply them element by element, and then add them together to get the current cell state c t , the formula is as follows:

[0024]

[0025] Among them, f t ∈R d×1 when the forget gate output, i t ∈R d×1 is the input gate output, is the current input state unit, c t-1 ∈R d×1 is the state unit at the previous moment, and ° represents element-wise multiplication.

[0026] Furthermore, the output training steps of the basic learning model are as follows: the output of the model is h t and the current cell state c t , and h t The current cell state c t The output o of the output gate t OK, set the current cell state c t Take tanh to get d t , then d t With o t Multiply element by element to get the final h t , the formula is as follows:

[0027]

[0028] where h t ∈R d×1 is the hidden state of the current layer, o t ∈R d×1 is the output of the output gate, c t ∈R d×1 is the state unit at the current moment.

[0029] Furthermore, the training learning method of the first enhanced model is the same as that of the basic learning model, where X = (x1, x2, ..., x 730 )x in T t It is [smoke, flow].

[0030] Furthermore, the training learning method of the second enhanced model is the same as that of the basic learning model, where X = (x1, x2, ..., x 730 )x in T t It is [smoke, flow, sulfur dioxide].

[0031] Furthermore, the clustering algorithm in step S4 is K-Means clustering calculation, and its calculation formula is:

[0032]

[0033] The beneficial effects of the present invention are as follows:

[0034] Compared with the prior art, the present invention relies on LSTM to establish a learning model, and improves the matching degree of the model to the actual scenario by gradually enhancing single factors, synergistic factors and flow data. Finally, the clustering algorithm is used to refine the clustering degree of missing data to make up for the limitation of the model that cannot learn missing data and cannot predict. Finally, the goal of predicting the online monitoring data of the next 24 hours by inputting the online monitoring data of pollutants in the previous 24 hours is achieved, providing a comprehensive online monitoring early warning method for productive enterprises. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is a flow chart of the present invention;

[0036] Figure 2 A flowchart of constructing a learning model and training an enhanced model in the present invention;

[0037] Figure 3 The figure is a flow chart for calculating the prediction data in the present invention. DETAILED DESCRIPTION

[0038] The specific implementation of the present invention will be described below in conjunction with the accompanying drawings.

[0039] The present invention provides an advance warning method for comprehensive monitoring of industrial pollution discharge, aiming to solve the problem that the online monitoring method in the prior art is usually terminal emission monitoring, and the concentration of the monitoring factor cannot be known in advance, resulting in a lack of decision-making basis and data support for the improvement of industrial pollution production and treatment processes. The present invention can effectively improve the compliance rate of enterprise pollution emission data and promote the green development of enterprises.

[0040] Specifically, the preferred embodiments of the present invention are:

[0041] like Figures 1 to 3 As shown, first, a single-factor basic learning model is generated, and the model input data sequence is prepared: Based on the LSTM model, a pollution source emission factor sequence (exhaust gas: smoke) X = (x1, x2, ..., x t ,…)T, the input of each time step can be expressed as x t , use sliding window to divide the sequence into several windows of size L, step size step, when the data is divided to the end, if it is less than L and cannot form a window, the missing data is filled with 0. A pollution source that continuously emits 24 hours a day has 24 hours of time series data input, that is T, set the window size to 12 and the step size to 10 to divide the data into two windows, namely:

[0042]

[0043] Assuming that a year has 365 days, this model prepares the online monitoring hourly data of pollution sources in the most recent year as learning data, and finally forms X=(x1,x2,…,x 730 )T

[0044] like Figure 1 The figure shows the input, calculation, memory and output process of the smoke and dust data in the atmosphere in the recent year;

[0045] In one embodiment of the present invention, the input of the model unit includes the input at the current moment The output state ht-1 of the previous moment and the unit state ct-1 of the previous moment. When operating the first layer of LSTM units, h0 and c0 are initialized, and ht-1 and ct-1 in the subsequent LSTM units are obtained from the previous LSTM unit. xt, ht-1, and ct-1 represent the input information of the current moment, the output information of the previous moment, and the memory information of the previous moment, respectively. Among them, ∈Rm×1, m is the window size (length) after input sequence processing, ht-1 is the output state of the previous moment, the shape is ht-1∈Rd×1, d is the hidden state size of the LSTM unit, ct-1 is the unit state of the previous moment, the shape is ct-1∈Rd×1, and it has the same shape as ht-1.

[0046] Furthermore, ht-1 and Concatenate to form longer vectors Right now ×1, as shown in the formula below, then will be passed to each gate. When multiple batches are used, ×n.

[0047]

[0048] In one embodiment of the invention, the output of each gate is then calculated, and the input of each gate is Will Multiply the weight matrix W in the gate and add the bias value b to get the intermediate result M. Then take Sigmoid (loss function value range; [0,1]) for M to get the output gt of the gate, which has the same shape as the unit state ct, that is, gt∈R d×1;

[0049]

[0050] in, W f , W i , W o ∈R d×(d+m), b f 、b i, b0∈R d×1,f t 、i t , o t ∈R d×1.

[0051] In one embodiment of the present invention, the current input unit state is calculated The value of indicates the amount of current input to be retained in memory. and the weight matrix W of the current state unit c Multiply and add the bias value bc to get the intermediate result Mc, then take tanh on Mc to get the output ct.

[0052]

[0053] in, W c ∈R d×(d+m), b c ∈R d×1,

[0054] In one embodiment of the present invention, the current unit state c is then performed. t Calculation of the result f using the forget gate and input gate t 、i t and the unit state c at the previous moment t-1 To calculate the current cell state c t , respectively f t 、c t-1 Element-wise multiplication, i t and Multiply them element by element, and then add them together to get the current cell state c t .

[0055]

[0056] Among them, f t ∈R d×1 when the forget gate output, i t ∈R d×1 is the input gate output, is the current input state unit, c t-1 ∈R d×1 is the state unit at the previous moment, and ° represents element-wise multiplication.

[0057] In one embodiment of the present invention, the output of the final model is h t and the current cell state c t , and h t The current cell state c t The output o of the output gate t , determine the current cell state ct Take tanh to get d t , then d t With o t Multiply element by element to get the final h t

[0058]

[0059] where h t ∈R d×1 is the hidden state of the current layer, o t ∈R d×1 is the output of the output gate, c t ∈R d×1 is the state unit at the current moment.

[0060] In particular, after the enhanced model is trained, the model needs to be evaluated and solidified. The above two steps can be implemented by relatively mature conventional evaluation and solidification methods in the field, and will not be repeated here.

[0061] Furthermore, X=(x1,x2,…,x 730 )T's x t Converted into [smoke, flow], the learning is completed in the above way to form the corresponding enhanced model 1 in the figure.

[0062] Furthermore, X=(x1,x2,…,x 730 )T's x t Converted into [smoke, flow, sulfur dioxide], and completed the learning in the above way to form the corresponding enhanced model 2 in the figure.

[0063] Finally, 100 sets of data were randomly selected and input into the above three models for prediction to obtain the prediction accuracy weight matrix W[a,a1,a2]. The budget value of the comprehensive model is the average value of the prediction value*W of the three models.

[0064] Finally, X=(x1,x2,…,x 730 )T performs K-Means clustering calculation and obtains the time interval where the input parameter is 0. The specific calculation formula is as follows:

[0065]

[0066] By obtaining the interval where the data is 0, the problem of missing data not being able to be budgeted is solved, and the integrity of the continuous cycle of the predicted data is achieved.

[0067] In summary, the present invention relies on LSTM to establish a learning model, and improves the matching degree of the model to the actual scenario by gradually enhancing single factors, synergistic factors and flow data. Finally, the clustering algorithm is used to refine the clustering degree of missing data to make up for the limitation that the model cannot learn missing data and cannot predict. Finally, the goal of predicting the online monitoring data of the next 24 hours by inputting the online monitoring data of pollutants in the previous 24 hours is achieved, providing a comprehensive online monitoring early warning method for productive enterprises.

[0068] The above description is an explanation of the present invention, not a limitation of the present invention. The scope of the present invention is defined in the claims. Any form of modification may be made within the scope of protection of the present invention.

Claims

1. An advance warning method for comprehensive monitoring of industrial sewage discharge, characterized by: The steps include: Step S1: Based on the LSTM model, at least three learning models containing different numbers of factors are established, namely a basic learning model, a first learning model and a second learning model; Step S2: using the online monitoring hourly data of pollution sources in the most recent year as learning data, the three learning models are respectively subjected to input, calculation, memory and output processes, and three enhanced models are obtained by training, which correspond to the basic model, the first enhanced model and the second enhanced model respectively; Step S3: randomly select 100 groups of data and input them into the three enhanced models in step S2 for prediction to obtain a prediction accuracy weight matrix, and generate a comprehensive model, the prediction value of the comprehensive model is the average of the prediction values ​​of the three models; Step S4: Finally, the learning data is calculated using a clustering algorithm to achieve the integrity of the continuous cycle of the predicted data.

2. The method for advanced warning of comprehensive monitoring of industrial sewage discharge according to claim 1, characterized in that: The number of factors in the basic learning model is single. Based on the LSTM model, a pollution source emission factor sequence is constructed, where the factor x t For smoke or ammonia nitrogen, the sequence X = (x1, x2, ..., x t ,…) T , the input at each time step can be expressed as x t , use sliding window to divide the sequence into several windows with window size L and step size step. When the data is divided to the end, if it is less than L and cannot form a window, the missing data is filled with 0; if the pollution source that continuously emits 24 hours a day has 24 hours of time series data input, that is Set the window size to 12 and the step size to 10, then divide the data into two windows, Assuming that a year has 365 days, this model prepares the online monitoring hourly data of pollution sources in the most recent year as learning data, and finally forms X=(x1,x2,…,x 730 ) T .

3. The advance warning method for comprehensive monitoring of industrial sewage discharge according to claim 1 is characterized in that: The input training steps of the basic learning model are as follows: the input of the model unit includes the input at the current moment The output state ht-1 at the previous moment and the unit state ct-1 at the previous moment are initialized to h0 and c0 when the first layer LSTM unit is operated. In the subsequent LSTM units, ht-1 and ct-1 are obtained from the previous LSTM unit, where xt, ht-1, and ct-1 represent the input information at the current moment, the output information at the previous moment, and the memory information at the previous moment, respectively; m is the window size (length) after the input sequence is processed, h t-1 is the output state at the previous moment, the shape is h t-1∈R d×1, d is the hidden state size of the LSTM unit, ct-1 is the unit state at the previous moment, the shape is ct-1∈R d×1, which has the same shape as h t-1; h t-1 and Concatenate to form longer vectors Right now As shown in the following formula, then It will be passed into each gate, and the formula is as follows:

4. The method for advanced warning of comprehensive monitoring of industrial sewage discharge according to claim 3, characterized in that: The computational training steps of the basic learning model are: the input of each gate is Will Multiply the weight matrix W in the gate and add the bias value b to get the intermediate result M, then take Sigmoid (loss function value range; [0,1]) for M to get the output gt of the gate, which has the same shape as the unit state ct, that is, gt∈R d×1 Among them, W f 、W i 、W o ∈R d×(d+m), b f 、b i 、b0∈R d×1, f t 、i t 、o t ∈Rd×1。 5. The method for advanced warning of comprehensive monitoring of industrial sewage discharge according to claim 4, characterized in that: The memory training steps of the basic learning model are: calculating the current input unit state The value of indicates the amount of current input to be retained in memory. and the weight matrix W of the current state unit c Multiply and add the bias value bc to get the intermediate result Mc, then take tanh on Mc to get the output ct, the formula is as follows: in, W c ∈R d×(d+m), b c ∈R d×1, Next, the current cell state c t The calculation of the forget gate and the input gate obtains the result f t 、i t and the unit state c at the previous moment t-1 To calculate the current cell state c t , respectively f t 、c t-1 Element-wise multiplication, i t and Multiply them element by element, and then add them together to get the current cell state c t , the formula is as follows: Among them, f t ∈R d×1 when the forget gate output, i t ∈R d×1 is the input gate output, ∈R d×1 is the current input state unit, c t-1 ∈R d×1 is the state unit at the previous moment, and ° represents element-wise multiplication.

6. The method for advanced warning of comprehensive monitoring of industrial sewage discharge according to claim 5, characterized in that: The output training steps of the basic learning model are: the output of the model is h t and the current cell state c t , and h t The current cell state c t The output o of the output gate t OK, set the current cell state c t Take tanh to get d t , then d t With o t Multiply element by element to get the final h t , the formula is as follows: where h t ∈R d×1 is the hidden state of the current layer, o t ∈R d×1 is the output of the output gate, c t ∈R d×1 is the state unit at the current moment.

7. The method for advanced warning of comprehensive monitoring of industrial sewage discharge according to claim 6, characterized in that: The training learning method of the first enhanced model is the same as that of the basic learning model, where X = (x1, x2, ..., x 730 )x in T t It is [smoke, flow] or [ammonia nitrogen, flow].

8. The method for advanced warning of comprehensive monitoring of industrial sewage discharge according to claim 6, characterized in that: The training and learning method of the second enhanced model is the same as that of the basic learning model, where X = (x1, x2, ..., x 730 )x in T t It is [smoke, flow, sulfur dioxide], [ammonia nitrogen, COD, flow] or [ammonia nitrogen, total phosphorus, flow].

9. The method for advanced warning of comprehensive monitoring of industrial sewage discharge according to claim 1, characterized in that: The clustering algorithm in step S4 is K-Means clustering calculation, and its calculation formula is:

Images