CN111994970A

CN111994970A - LSTM-based dosing prediction method and dosing system for efficient sewage sedimentation tank

Info

Publication number: CN111994970A
Application number: CN202010757643.6A
Authority: CN
Inventors: 杨志科; 蒋秋明; 王兴荣; 董孔益; 林汉涛; 李益; 周吉; 俞晓
Original assignee: Shanghai Shangshi Longchuang Intelligent Technology Co Ltd
Current assignee: Shanghai Shangshi Longchuang Intelligent Technology Co Ltd
Priority date: 2020-07-31
Filing date: 2020-07-31
Publication date: 2020-11-27
Anticipated expiration: 2040-07-31
Also published as: CN111994970B

Abstract

The invention relates to an LSTM-based intelligent dosing prediction method for a high-efficiency sewage sedimentation tank, which comprises the following steps of: s1: acquiring prediction related data and preprocessing the data; s2: respectively constructing a feedforward characteristic component and a feedback characteristic component by using the prediction related data; s3: acquiring prediction input data and inputting a trained prediction model; s4: the prediction model outputs a predicted value of the adding amount of the phosphorus removal drug PAC; s5: intelligently adding chemicals into the efficient sewage sedimentation tank according to the predicted value of the adding amount of the phosphorus removal chemical PAC. Compared with the prior art, the method has the advantages of implementing accurate prediction, reducing medicine consumption, ensuring stable and standard effluent and the like.

Description

LSTM-based dosing prediction method and dosing system for efficient sewage sedimentation tank

Technical Field

The invention relates to the field of sewage treatment, in particular to an LSTM-based dosing prediction method and dosing system for a sewage high-efficiency sedimentation tank.

Background

In the modern sewage treatment industry, the removal of total phosphorus depends on chemical phosphorus removal to a great extent, the variability of sewage inflow easily causes large fluctuation of effluent, increases the phenomena such as medicine consumption and the like, and causes serious waste, most chemical phosphorus removal equipment and devices (mainly high-efficiency sedimentation tanks) in the market at present only use an online instrument for feedback regulation, and the online instrument lags behind the detection principle, can only relieve the fluctuation of inflow load to a certain extent, and the medicine consumption is waste, so that the problem cannot be fundamentally solved.

Disclosure of Invention

The invention aims to overcome the defects in the prior art and provide an LSTM-based efficient sewage sedimentation tank dosing prediction method and dosing system for ensuring the effluent to stably reach the standard.

The purpose of the invention can be realized by the following technical scheme:

an LSTM-based intelligent dosing prediction method for a high-efficiency sewage sedimentation tank comprises the following steps:

s1: acquiring prediction related data and preprocessing the data;

s2: respectively constructing a feedforward characteristic component and a feedback characteristic component by using the prediction related data;

s3: acquiring prediction input data and inputting a trained prediction model;

s4: the prediction model outputs a predicted value of the adding amount of the phosphorus removal drug PAC;

s5: intelligently adding chemicals into the efficient sewage sedimentation tank according to the predicted value of the adding amount of the phosphorus removal chemical PAC.

Furthermore, the prediction related data comprise the inflow water flow Q and the inflow water suspended solid concentration SS_inTotal phosphorus TP in the feed water_inTotal phosphorus TP in effluent_outThe sewage temperature T, the sewage pH value and the phosphorus removal agent PAC concentration.

Furthermore, the prediction input data comprises feed-forward data, feedback data and dosage data, wherein the feed-forward data comprises inflow Q and inflow total phosphorus TP_inWater influent suspended solids concentration SS_inThe sewage treatment system comprises a sewage temperature T, a sewage pH value and a feedforward characteristic component, wherein feedback data comprise a feedback characteristic component, and dosage data comprise the concentration of a phosphorus removal drug PAC.

Further preferably, the prediction related data further comprise the concentration of suspended solid SS of the effluent_outAnd the concentration of the PAM adsorbent, and the feedback data also comprises the concentration SS of the solid suspended matter in the effluent_outThe dosage data also comprises the concentration of the PAM adsorbent, the concentration of the effluent solid suspended matter and the concentration of the PAM adsorbent have certain influence on the effect of the phosphorus removal agent PAC, and the prediction accuracy can be further improved by considering the two factors.

Further, the feedforward characteristic component is constructed by statistical characteristics, including the first 4t_oMaximum total phosphorus TP fed over time_{in_max}First 4t_oMinimum influent total phosphorus TP over time_{in_min}And the first 4t_oAverage total phosphorus TP fed over time_{in_mean}Wherein t is_oFor a set data acquisition frequency. In view of large fluctuation of sewage, long treatment time and water retention time of corresponding processes, the current water quality and water quantity are considered, the water quality condition of the water tank in the past period theoretically is also comprehensively considered, and the corresponding statistical indexes can reflect the real condition of continuous time sequence variables to increase the accuracy of the model.

Furthermore, the data acquisition frequency t_oIs set to 0And 5h, the setting of the data acquisition frequency accords with the flow rate and the change speed of the content of the internal chemical substances of the general efficient sewage sedimentation tank, the prediction accuracy can be improved, and meanwhile, redundant and useless data cannot be generated due to too frequent acquisition.

Further, the feedback characteristic component is constructed by PID service logic and comprises a proportional feedback characteristic DATA_KCIntegral feedback characteristic DATA_KIAnd differential feedback characteristic DATA_KD. Further, the expression of the feedback characteristic component is as follows:

DATA_KC(t)＝E(t)-E(t-1)

DATA_KI(t)＝E(t)

DATA_KD(t)＝E(t)-2E(t-1)+E(t-2)

E(t)＝TP_out(t)-TP_set

wherein t is the current prediction time, E (t) is the dosage deviation at the time t, and TP_out(t) is the total phosphorus sampling value of the effluent at the time t, TP_setThe set value of the total phosphorus of the effluent is. The feedback characteristic variable can reflect the actual sedimentation process of the efficient sedimentation tank, ensure the effluent quality condition, enhance the effluent safety, form closed-loop control, increase the reliability and generalization capability of the model, and the linear and nonlinear relations between the water quality variable and the dosing quantity can be further reflected by different property variables of proportion, integral and differentiation, so that the nonlinear expression of the model is enhanced.

Furthermore, the neural network structure of the prediction model comprises a first LSTM neural network, a first Dropout layer, a second LSTM neural network, a second Dropout layer and a third LSTM neural network which are sequentially connected, training is completed through an Adam optimization algorithm, a loss function adopts a root mean square error function, the LSTM neural network comprises an input layer, three LSTM cell layers and an output layer, and the LSTM cell layer comprises a forgetting gate F_tInput gate I_tAnd an output gate O_t。

The utility model provides a high-efficient sedimentation tank of sewage intelligence medicine system based on LSTM, includes data acquisition module, control module and dosing pump, data acquisition module set up in the high-efficient sedimentation tank of sewage for gather the prediction relevant data in the high-efficient sedimentation tank of sewage, the dosing pump be used for throwing the phosphorus removal medicine PAC to the high-efficient sedimentation tank of sewage, control module be connected with data acquisition module and dosing pump respectively, including memory, treater and controller, the memory in store have computer program, this program realizes if when being executed by the treater the medicine prediction method that adds, the treater pass through the controller and be connected with the dosing pump, the phosphorus removal medicine PAC of output control signal control dosing pump throws the volume. The data acquisition module comprises a flow meter, a water inlet turbidimeter, a water inlet phosphate meter, a control module, a dosing pump, a water outlet phosphate meter, a water outlet turbidimeter, a PH meter and a thermometer, wherein the flow meter, the water inlet turbidimeter, the water inlet phosphate meter, the water outlet turbidimeter, the PH meter and the thermometer.

Compared with the prior art, the invention has the following advantages:

1) according to the invention, through collecting various data of the sewage high-efficiency sedimentation tank, feedforward data, feedback data and dosage data are obtained, and meanwhile, the data are used as a basis for predicting the PAC dosage of the phosphorus removing medicine, and an LSTM neural network is adopted for prediction, so that the PAC dosage of the phosphorus removing medicine can be rapidly controlled in real time, the problem of lag of the traditional online instrument on the detection principle is solved, the medicine consumption is reduced, the fluctuation of the water inlet load is effectively relieved, and the stable standard reaching of the outlet water is ensured;

2) the feedforward data of the invention comprises feedforward characteristic components, namely total phosphorus TP of the inflow water_inThe feedforward characteristic component is formed by adopting a statistical characteristic structure, in view of large fluctuation of sewage, long treatment time and water retention time of corresponding processes, the current water quality and water quantity are considered, the water quality condition of the water in a pool in the past period of time is also considered comprehensively, and the corresponding statistical index can reflect the real condition of continuous time sequence variable to increase the accuracy of the model.

3) The feedback data of the invention comprises feedback characteristic components and utilizes the total phosphorus TP of effluent_outAnd the total phosphorus setting value TP of the effluent_outThe PID business logic structure is adopted to form a feedback characteristic component, and the feedback characteristic variable can reflect the actual high-efficiency sedimentation tank sedimentationAnd in the precipitation process, the water quality condition of the effluent is ensured, the effluent safety is enhanced, closed-loop control is formed, the reliability and the generalization capability of the model can be increased, and the linear and nonlinear relations between the water quality variable and the dosing quantity can be better reflected by different property variables of proportion, integral and differentiation, so that the nonlinear expression of the model is enhanced.

4) According to the invention, the prediction model is constructed by adopting the LSTM neural network, the Dropout layer is arranged, overfitting is prevented in backward propagation, the generalization capability of the model is improved, the prediction accuracy is improved, the dosage is predicted in real time, the control is accurate, and the stable standard of effluent is realized.

Drawings

FIG. 1 is a schematic flow diagram of the present invention;

FIG. 2 is a schematic diagram of a neural network structure of a prediction model;

FIG. 3 is a diagram of the structure of the LSTM neural unit;

FIG. 4 is a schematic diagram of the system of the present invention;

FIG. 5 is a diagram illustrating the prediction result.

Detailed Description

The invention is described in detail below with reference to the figures and specific embodiments. It is to be understood that the embodiments described are only a few embodiments of the present invention, and not all embodiments. All other embodiments, which can be obtained by a person skilled in the art without any inventive step based on the embodiments of the present invention, shall fall within the scope of protection of the present invention.

As shown in FIG. 1, the invention provides an LSTM-based intelligent dosing prediction method for a high-efficiency sewage sedimentation tank, which comprises the following steps:

s1: acquiring prediction related data and preprocessing the data;

s3: acquiring prediction input data and inputting a trained prediction model;

The specific process is as follows:

influence factor of PAC dosage of phosphorus removal drug

In the sewage treatment process, the high-efficiency sedimentation tank is arranged behind the biochemical tank, and the removal of the total phosphorus in the sewage is realized by adding a proper amount of phosphorus removal medicine, and the set value TP of the total phosphorus in the medicine types, the components and the effluent_setUnder the relatively fixed condition, the influencing factors influencing the PAC dosage of the phosphorus removal medicine in the high-efficiency sedimentation tank comprise: inflow Q and total phosphorus TP_inWater influent suspended solids concentration SS_inWater inlet turbidity, sewage temperature T, sewage pH value and water inlet chemical oxygen demand COD_inPAM (polyacrylamide) dosage and suspended solid concentration SS (suspended solid concentration) of effluent_outTotal phosphorus TP in effluent_outAnd chemical oxygen demand COD of the effluent_outAnd the like. The influence factors can be simplified through normalization processing and principal component analysis, and a basis is provided for input variable selection of a prediction model, and the specific steps are as follows:

1. data cleaning: abnormal value detection is carried out by adopting an isolated Forest (Isolation Forest) algorithm, and missing value filling is carried out through a near value;

2. and (3) normalization processing of data: data normalization before data principal component analysis is beneficial to dimensionless weighting of data and rapid convergence of a loss function of a following algorithm, and a data normalization formula is as follows:

wherein the content of the first and second substances,

is normalized sample data, i is the number of samples, j is the sample component, D_ijIs the jth sample component of the ith sample, m is the number of samples of the jth sample component, and σ j is the variation of the jth sample component.

3. Principal component analysis of data: the normalized data can reduce the number of input variables through principal component analysis, and finally, the influence factors selected and considered by the method comprise: inflow Q and total phosphorus TP_inWater influent suspended solids concentration SS_inTotal phosphorus TP in effluent_outThe concentration SS of suspended solid in the effluent_outSewage temperature T and sewage pH.

(II) input data feature processing

In the present invention, the input data of the prediction model includes feedback input data and feedforward input data, which are specifically as follows:

1. feedback data

The data used for the feedback input of the prediction model comprises total phosphorus TP in the effluent_outThe concentration SS of suspended solid in the effluent_outWherein the effluent total phosphorus TP_outAfter data processing is needed, feedback characteristic components are constructed and then serve as feedback data to be input into the prediction model, and other data can be directly used as feedback data to be input into the prediction model.

The feedback characteristic components are specifically constructed as follows:

the total phosphorus TP is calculated according to the time t and the effluent_outThe dosage deviation E (t) can be calculated by the feedback of the total phosphorus in the effluent as a variable, and the specific formula is as follows:

E(t)＝TP_out(t)-TP_set

wherein E (t) is a sampling value TP of total phosphorus in the effluent at the time t_out(t) and the total phosphorus setting value TP of the effluent_setThe difference of (a).

Through proportional-integral-derivative control, the change D (t) of the PAC dosage of the phosphorus removing medicine under feedback control can be obtained:

D(t)＝K_c[E(t)-E(t-1)]+K_iE(t)+K_d[E(t)-2E(t-1)+E(t-2)]

in the formula K_c、K_i、K_dProportional coefficient, integral coefficient and differential coefficient respectively, and the feedback characteristic component includes: DATA_KC、DATA_KI、DATA_KDThe expressions are respectively:

DATA_KC(t)＝E(t)-E(t-1)

DATA_KI(t)＝E(t)

DATA_KD(t)＝E(t)-2E(t-1)+E(t-2)

2. feed forward data

The data adopted by the feedforward input of the prediction model comprises inflow water flow Q and inflow water total phosphorus TP_inWater influent suspended solids concentration SS_inSewage temperature T and sewage pH value, wherein the total phosphorus TP of the inlet water_inAfter data processing is needed, feedforward characteristic components are constructed and then serve as feedforward data input prediction models, and other data can be directly used as feedforward data input prediction models.

The feedforward characteristic component is specifically constructed as follows:

setting the feedforward characteristic component includes: 4t before time t_oMaximum influent total phosphorus TP_{in_max}4t before time t_oMinimum influent total phosphorus TP_{in_min}And 4t before time t_oAverage influent total phosphorus TP of_{in_mean}Wherein t is_oFor the set data collection frequency, t is preferably set in this embodiment_o0.5h, namely the total phosphorus TP of the inlet water input into the prediction model_inData needs to be collected within two hours before the current time.

3. Dose data

The data input by the prediction model also comprises the medicine amount data in the efficient sewage sedimentation tank, including the concentration of the phosphorus removal medicine PAC and the concentration of the adsorbent PAM.

(III) prediction model building

The PAC dosage of the phosphorus removal medicine is predicted by adopting an LSTM neural network based on a python platform, and a long-term and short-term memory network (LSTM) is a special RNN network, can learn and remember a longer sequence and does not depend on a pre-specified window lag observation value as input; the value t can be predicted through t-1 value before a sequence, and the previous information can be memorized to understand the current content, so that the problem of gradient extinction easily occurring in the RNN network is solved.

The LSTM neural network model constructed on the python platform comprises the following components: one input layer, three layers of LSTM cells, and one output layer. Wherein, as shown in fig. 3, the LSTM cell layer is internally provided with a plurality of thresholds including a forgetting gate F_tInput gate I_tAnd an output gate O_tIn this embodiment, the forward propagation function of the LSTM recurrent neural network is set as:

I_t＝σ(W_xiX_t+W_hih_t-1+b_i)

F_t＝σ(W_xfX_t+W_hfh_t-1+b_f)

c_t＝F_t*c_t-1+tanh(W_xcX_t+W_hch_t-1+b_c)

O_t＝σ(W_xoX_t+W_hoh_t-1+b_o)

wherein, W_xiAs a weight between input layers to input gates, W_hiIs the weight between the hidden layer and the input gate at the previous time, W_xfFor weights between input layers to forget gate, W_hfIs the weight between the hidden layer and the forgetting gate at the last moment, W_xcAs weights between input layers to state cells, W_hcWeight between hidden layer and state cell at last moment, W_xoAs weights between input layers to output gates, W_hoFor the weights between the hidden layer and the output gate at the previous time, σ () is the sigma function, X_tFor input, h_t-1To represent the output of the hidden layer at the last moment, b_iFor input gate biasing, b_fTo forget the door bias, c_tIs state cell export, c_t-1For cell export in the last state, b_cState of cell bias, b_oIs an output gate bias.

As shown in FIG. 2, the neural network structure of the prediction model in the present invention includes a first LSTM neural network, a first Dropout layer, a second LSTM neural network, a second Dropout layer and a third LSTM neural network, which are connected in sequence. In the embodiment, the Keras is used as a modeling environment to model and learn and train the LSTM neural network; during training, the network layer of the LSTM neural network maintains state between fixed line numbers of data in the training data set in which the LSTM neural network operates before updating the network weights. Determining the time for emptying the state of the LSTM neural network layer by adopting a reset _ states function, compiling the LSTM neural network by using a mean _ squared _ error loss function by adopting a root mean square error as a loss function, and finishing the training of the LSTM neural network by using an adam optimization algorithm, wherein the specific implementation code and the hyper-parameter are set as follows:

model＝Sequential()

model.add(LSTM(50,input_shape＝(12,10),return_sequences＝True))

model.add(Dropout(0.3))

model.add(LSTM(100,return_sequences＝True))

model.add(Dropout(0.3))

model.add(LSTM(50,return_sequences＝False))

model.add(Dense(train_Y.shape[1]))

model.compile(loss＝”mean_squared_error”,optimizer＝”adam”)

the Dropout layer forwards propagates a random value (0-1) of each neuron, sets the neuron value of which the random value is smaller than a set threshold value and corresponding to the neuron to be 0, sets the neuron value of which the random value is larger than the set threshold value and corresponding to the neuron to be 2 multiplied by an original value, stores the random values corresponding to all the neurons, and uses the random values in backward propagation, so that overfitting can be prevented, and the generalization capability of the model is improved.

The following table shows some of the training data used in this example:

(IV) model prediction

And respectively inputting the feedforward data and the feedback data into the trained prediction model in the third part, and finally outputting the predicted PAC dosage of the phosphorus removal medicine by the prediction model.

Fig. 5 shows the prediction case using the method of the present invention.

As shown in fig. 4, the invention further provides an LSTM-based dosing system for a high-efficiency sedimentation tank for sewage, which comprises a flow meter, a water inlet turbidity meter, a water inlet phosphate meter, a control module, a dosing pump, a water outlet phosphate meter, a water outlet turbidity meter, a PH meter and a thermometer, wherein the flow meter, the water inlet turbidity meter, the water inlet phosphate meter, the water outlet turbidity meter, the PH meter and the thermometer are respectively arranged in the high-efficiency sedimentation tank and are respectively connected with the control module for collecting relevant data in the high-efficiency sedimentation tank to serve as a basis for predicting data. The dosing pump is connected with the control module and used for executing dosing sleeving. The control module comprises a memory, a processor and a controller, wherein a computer program is stored in the memory, the processor acquires data input by each data acquisition meter and executes the program to realize the LSTM-based dosing prediction method for the efficient sewage sedimentation tank, and finally the controller is connected with the dosing pump to control the dosing pump to dose the phosphorus removal pesticide PAC of the efficient sewage sedimentation tank.

While the invention has been described with reference to specific embodiments, the invention is not limited thereto, and those skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope of the invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. An LSTM-based intelligent dosing prediction method for a high-efficiency sewage sedimentation tank is characterized by comprising the following steps:

s1: acquiring prediction related data and preprocessing the data;

s3: acquiring prediction input data and inputting a trained prediction model;

2. The LSTM-based intelligent dosing prediction method for the efficient sewage sedimentation tank according to claim 1, wherein the relevant prediction data comprises inflow Q and inflow suspended solid concentration SS_inTotal phosphorus TP in the feed water_inTotal phosphorus TP in effluent_outThe sewage temperature T, the sewage pH value and the phosphorus removal agent PAC concentration.

3. The LSTM-based intelligent dosing prediction method for the efficient wastewater sedimentation tank according to claim 2, wherein the prediction input data comprises feed-forward data, feedback data and dosage data, and the feed-forward data comprises feed water flow Q and total phosphorus TP of the feed water_inWater influent suspended solids concentration SS_inThe sewage treatment system comprises a sewage temperature T, a sewage pH value and a feedforward characteristic component, wherein feedback data comprise a feedback characteristic component, and dosage data comprise the concentration of a phosphorus removal drug PAC.

4. The LSTM-based intelligent dosing prediction method for the efficient sewage sedimentation tank according to claim 3, wherein the prediction related data further comprises an effluent suspended solid concentration SS_outAnd the concentration of the PAM adsorbent, and the feedback data also comprises the concentration SS of the solid suspended matter in the effluent_outAnd the medicine quantity data also comprises the concentration of an adsorbent PAM.

5. The LSTM-based intelligent dosing prediction method for efficient sewage sedimentation tank according to claim 3Wherein the feedforward characteristic component is constructed by statistical characteristics, including the first 4t_oMaximum total phosphorus TP fed over time_{in_max}First 4t_oMinimum influent total phosphorus TP over time_{in_min}And the first 4t_oAverage total phosphorus TP fed over time_{in_mean}Wherein t is_oFor a set data acquisition frequency.

6. The LSTM-based intelligent dosing prediction method for efficient sewage sedimentation tank according to claim 5, wherein the data acquisition frequency t is_oThe value of (2) was set to 0.5 h.

7. The LSTM-based intelligent dosing prediction method for the efficient sewage sedimentation tank according to claim 3, wherein the feedback feature component is constructed through PID business logic and comprises a proportional feedback feature DATA_KCIntegral feedback characteristic DATA_KIAnd differential feedback characteristic DATA_KD。

8. The LSTM-based intelligent dosing prediction method for the efficient sewage sedimentation tank according to claim 7, wherein the expression of the feedback characteristic component is as follows:

DATA_KC(t)＝E(t)-E(t-1)

DATA_KI(t)＝E(t)

DATA_KD(t)＝E(t)-2E(t-1)+E(t-2)

E(t)＝TP_out(t)-TP_set

wherein t is the current prediction time, E (t) is the dosage deviation at the time t, and TP_out(t) is the total phosphorus sampling value of the effluent at the time t, TP_setThe set value of the total phosphorus of the effluent is.

9. The LSTM-based intelligent dosing prediction method for the efficient wastewater sedimentation tank according to claim 1, wherein the neural network structure of the prediction model comprises a first LSTM neural network, a first Dropout layer, a second LSTM neural network and a second Dro layer which are sequentially connectedThe training of the pout layer and the third LSTM neural network is completed through an Adam optimization algorithm, the loss function adopts a root mean square error function, the LSTM neural network comprises an input layer, three LSTM cell layers and an output layer, and the LSTM cell layers comprise forgetting gates F_tInput gate I_tAnd an output gate O_t。

10. An LSTM-based intelligent dosing system for a high-efficiency sewage sedimentation tank is characterized by comprising a data acquisition module, a control module and a dosing pump, wherein the data acquisition module is arranged in the high-efficiency sewage sedimentation tank and used for acquiring prediction related data in the high-efficiency sewage sedimentation tank, the dosing pump is used for dosing a phosphorus removal drug PAC into the high-efficiency sewage sedimentation tank, the control module is respectively connected with the data acquisition module and the dosing pump and comprises a memory, a processor and a controller, a computer program is stored in the memory, when the program is executed by the processor, the dosing prediction method of the phosphorus removal drug PAC in any one of claims 1-9 is realized, the processor is connected with the dosing pump through the controller, and a control signal is output to control the dosing amount of the phosphorus removal drug PAC of the dosing pump.