A small sample-based air flotation synergistic oil-water separation effect prediction method and model

Abstract

The application provides a kind of based on small sample's air floatation synergistic oil-water separation effect prediction method and model, obtains the multiple original features of inlet water;According to the original feature generation expansion feature vector, the expansion feature vector at least includes original feature vector and physical interaction feature vector;Build base model, and the expansion feature vector is as the input of the base model;Adaptive optimization algorithm is used to train the base model;Through the prediction result of aggregation multiple base models, obtain the final prediction value and uncertainty estimation.The application still greatly enhances the ability of model to capture complex nonlinear relationship and prediction accuracy on the basis of small sample through polynomial expansion, physical interaction feature construction and deep neural network, effectively controls the model complexity by combining elastic network regularization and weight constraint, prevents overfitting, ensures that the prediction result conforms to the physical and chemical law through various constraints, improves the reliability and credibility of the model.

Description

Technical Field

[0001] This invention relates to the field of oil-water separation in industrial wastewater treatment, and in particular to a method and model for predicting the effect of air flotation-enhanced oil-water separation based on a small sample. Background Technology

[0002] Air flotation (AF) technology is one of the core processes in current oily wastewater treatment. It generates microbubbles that attach to oil droplets, separating them using buoyancy. AF efficiency can be enhanced by using multi-stage AF and varying the AF method. The separation effect (usually measured by indicators such as effluent oil removal rate and suspended solids removal rate) is affected by various factors, such as influent water quality (initial oil content, pH value) and AF operating parameters (dissolved air pressure, reflux ratio, reagent dosage), and there are complex nonlinear relationships among these factors.

[0003] Predictions of the effects of air flotation often rely on mechanistic models based on traditional theories or traditional statistical models. However, mechanistic models require precise physicochemical parameters, which are difficult to obtain in actual industrial scenarios and have poor adaptability; traditional statistical models (such as response surface methodology and linear regression) have limited prediction accuracy because they are difficult to capture highly nonlinear relationships.

[0004] In recent years, machine learning has provided new insights for modeling complex industrial processes. However, deep learning models typically require a large amount of high-quality training data. In actual oilfield wastewater treatment plants, due to high testing costs and long cycles, the amount of effective data under stable operating conditions is very limited (i.e., a small sample problem). Directly applying traditional deep learning models easily leads to overfitting and poor generalization ability, making them unsuitable for actual prediction and production guidance. Therefore, developing a predictive model that can accurately predict the oil-water separation effect enhanced by air flotation even with limited data samples is of great significance for achieving real-time optimization and intelligent control of this process. Summary of the Invention

[0005] To address the problems existing in the background technology, this invention proposes a method and model for predicting the effect of air flotation-enhanced oil-water separation based on small samples.

[0006] A method for predicting the effect of air flotation-enhanced oil-water separation based on small samples, comprising the following steps:

[0007] S100: Obtain multiple original characteristics of the incoming water;

[0008] S200. Generate an extended feature vector based on the original features, wherein the extended feature vector includes at least the original feature vector and the physical interaction feature vector;

[0009] S300. Construct a base model and use the extended feature vector as the input to the base model;

[0010] S400. An adaptive optimization algorithm is used to train the base model;

[0011] S500. By aggregating the prediction results of multiple base models, the final predicted value and uncertainty estimate are obtained.

[0012] Based on the above, in step S100, the original characteristics include at least the initial oil content, pH value, temperature, dissolved gas content, oil layer thickness after separation, diameter of the separation equipment, height of the separation equipment, and height of the inlet position of the water.

[0013] Based on the above, in step S200, the physical interaction feature vector is:

[0014]

[0015] in:

[0016]

[0017] ψ1: Temperature-standardized interaction term between mass and equipment dimensions;

[0018]

[0019] ψ2: Logarithmic interaction term between pH and oil layer thickness, including location offset penalty;

[0020]

[0021] ψ3: The product of dissolved gas ratio and reservoir coverage;

[0022]

[0023] ψ4: Trigonometric interaction term between temperature and pH;

[0024] to These represent the initial oil content of the influent, pH value, temperature, dissolved air content, oil layer thickness after separation, diameter of the separation equipment, height of the separation equipment, and height of the inlet position, respectively.

[0025] Based on the above, in step S300, the base model is a deep resilient network architecture, which includes a feature transformation layer and a resilient network prediction head. The feature transformation layer includes multiple fully connected hidden layers, wherein the output of the l-th layer... Defined by the following formula:

[0026]

[0027] in ; This is the output of the l-th hidden layer; Here is the weight matrix for the l-th layer; Here, represents the bias vector of the l-th layer; ReLU is the activation function of the rectified linear unit.

[0028] An elastic network prediction head is used to output the master prediction value. Its definition is:

[0029]

[0030] For the predicted output; L is the index of the last hidden layer; w is the output layer weight vector; b is the output layer bias.

[0031] Based on the above, in step S300, a multi-objective loss function is designed for the base model. The multi-objective loss function It is the weighted sum of the following loss terms:

[0032]

[0033] ① Main task elastic network loss Its definition is:

[0034]

[0035] N is the number of samples; λ is the regularization strength; This is the regularization ratio;

[0036] ② Weight constraint regularization loss This is used to apply Frobenius norm penalty to the weight matrix of each layer of the feature transformation layer;

[0037] ③ Physical constraint losses, including:

[0038] Monotonicity constraint loss Used to constrain predicted values Oil content in influent Monotonically increasing and the effect on dissolved gas content Monotonically decreasing;

[0039] Boundary constraint loss Used to constrain predicted values Non-negative and not greater than the oil content of the influent ;

[0040] ④ Losses from auxiliary tasks It is achieved by using a hidden layer in the middle. Draw out auxiliary output And calculate its mean square error compared to the true value.

[0041] Based on the above, in step S500, the final predicted value and uncertainty estimation Its definition is:

[0042]

[0043]

[0044] Where M is the number of base models.

[0045] A prediction model for the effect of air flotation-enhanced oil-water separation based on small samples, using any of the above-mentioned methods for predicting the effect of air flotation-enhanced oil-water separation based on small samples.

[0046] This invention possesses significant substantive features and remarkable progress compared to existing technologies. Specifically, through polynomial expansion, physical interaction feature construction, and deep neural networks, it greatly enhances the model's ability to capture complex nonlinear relationships and improves prediction accuracy even with small sample sizes. By combining elastic network regularization with weight constraints, it effectively controls model complexity, prevents overfitting, and ensures stable model performance on unknown data. Through multiple constraints, it guarantees that the prediction results conform to physical and chemical laws, improving the model's reliability and credibility. The ensemble learning framework not only improves prediction accuracy but also provides a quantitative estimate of prediction uncertainty, offering a risk reference for decision-making. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating the process of this invention.

[0048] Figure 2 This is a schematic diagram of the architecture of the basic model of this invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] like Figure 1 As shown, a method for predicting the effect of air flotation-enhanced oil-water separation based on small samples has the following main steps:

[0051] 1. Input variable definition and feature engineering extension

[0052] Define the input independent variable vector ,in to These represent the initial oil content of the influent, pH value, temperature, dissolved air content, oil layer thickness after separation, diameter of the separation equipment, height of the separation equipment, and height of the inlet position, respectively; Strain variable The oil content in the effluent. Feature engineering is performed on the input independent variable vector to generate an extended feature vector f(z), which is composed of the following parts:

[0053] (1) The original eigenvector f(x);

[0054] (2) Polynomial eigenvectors It includes original features Feature interaction items and Multiple items , where ∀i,j∈{1,…,8} and i≤j.

[0055] Among them, the polynomial eigenvectors In this embodiment, the physical interaction feature vector is... :

[0056]

[0057] ψ1: Temperature-standardized interaction term between mass and equipment dimensions;

[0058]

[0059] ψ2: Logarithmic interaction term between pH and oil layer thickness, including location offset penalty;

[0060]

[0061] ψ3: The product of dissolved gas ratio and reservoir coverage;

[0062]

[0063] ψ4: Trigonometric interaction term between temperature and pH.

[0064] 2. Deeply resilient network architecture

[0065] like Figure 2 As shown, the deep resilient network architecture includes a feature transformation layer, which comprises at least three fully connected hidden layers, where the output of the l-th layer... Defined by the following formula:

[0066]

[0067] in ; Output of the lth hidden layer; : The weight matrix of the l-th layer; : The bias vector of the l-th layer; ReLU: Rectified Linear Unit activation function. The formulas for other fully connected hidden layers are the same as those for the l-th layer, the only difference being the parameters, i.e., the indices of each layer. and the corresponding weight matrix Bias vector Different; Input transmission method: The output of the previous layer is used as the input of the next layer.

[0068] An elastic network prediction head is used to output the master prediction value. Its definition is:

[0069]

[0070] : Predicted output; L: Index of the last hidden layer; w: Output layer weight vector; b: Output layer bias.

[0071] 3. Design a multi-objective loss function

[0072] The multi-objective loss function It is the weighted sum of the following loss terms:

[0073]

[0074] ① Main task elastic network loss Its definition

[0075]

[0076] N: Number of samples; λ: Regularization strength; : Regularization ratio.

[0077] ② Weight constraint regularization loss This is used to apply Frobenius norm penalty to the weight matrix of each layer of the feature transformation layer;

[0078] ③ Physical constraint losses, including:

[0079] Monotonicity constraint loss Used to constrain predicted values Oil content in influent Monotonically increasing and the effect on dissolved gas content Monotonically decreasing;

[0080] Boundary constraint loss Used to constrain predicted values Non-negative and not greater than the oil content of the influent ;

[0081] ④ Losses from auxiliary tasks It is achieved by using a hidden layer in the middle. Draw out auxiliary output And the mean square error between it and the true value is calculated.

[0082] 4. Adaptive optimization algorithm is used for model training.

[0083] An adaptive learning rate scheduler that combines learning rate warm-up with cosine annealing is used, and gradient pruning is implemented to prevent gradient explosion during training.

[0084] 5. Construct an integrated prediction framework

[0085] M base models are trained using the Bootstrap sampling method, and the final predicted value is obtained by aggregating the prediction results of each base model. and uncertainty estimation Its definition is:

[0086]

[0087]

[0088] This embodiment uses an air flotation enhancement device in a produced water treatment station of an oilfield as an example:

[0089] Operational data from the device over the past year was collected, and after cleaning and screening, 268 valid samples (N=268) were obtained. Each sample contains 8 input features (initial oil content, pH value, temperature, dissolved air content, oil layer thickness, equipment diameter, equipment height, and inlet height) and 1 output label (oil content in effluent). The specific input data is shown in Table 1 below:

[0090] Feature name symbol numerical values Initial oil content in influent (mg / L) <![CDATA[x1]]> 120 pH value <![CDATA[x2]]> 7.2 Temperature (°C) <![CDATA[x3]]> 28 Dissolved gas volume (L / min) <![CDATA[x4]]> 0.35 Oil layer thickness (cm) <![CDATA[x5]]> 12 Equipment diameter (m) <![CDATA[x6]]> 3 Equipment height (m) <![CDATA[x7]]> 4.5 Import height (m) <![CDATA[x8]]> 1.2

[0091] Table 1

[0092] Based on the input data, the physical interaction characteristics are calculated, and the values ​​of ψ1, ψ2, ψ3, and ψ4 are 248.89, 25.25, 0.933, and 0.465, respectively.

[0093] In this embodiment, four base models are used as an example to calculate the model prediction values. The output of the trained ensemble model (4 sub-networks) is shown in Table 2 below (unit: mg / L):

[0094] Subnetwork Predicted value 1 8.6 2 9.1 3 8.8 4 9.0

[0095] Table 2

[0096] Final predicted value:

[0097]

[0098] The comparison and analysis of the actual values ​​and predicted values ​​are shown in Table 3 below:

[0099] True value Predicted value relative error 9.2 8.88 3.48%

[0100] Table 3

[0101] As can be seen, although this embodiment uses a small sample size (far less than 1,000 groups), the error between its predicted value and the actual value is small, and it can fully guide actual prediction. Therefore, under the premise of significantly reducing detection costs and detection cycle, it still significantly improves the accuracy, robustness and physical consistency of oil content prediction in effluent.

[0102] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for predicting the effect of air flotation-enhanced oil-water separation based on small samples, characterized in that, Including the following steps: S100: Obtain multiple original characteristics of the incoming water; S200. Generate an extended feature vector based on the original features, wherein the extended feature vector includes at least the original feature vector and the physical interaction feature vector; S300. Construct a base model and use the extended feature vector as the input to the base model; S400. An adaptive optimization algorithm is used to train the base model; S500. By aggregating the prediction results of multiple base models, the final predicted value and uncertainty estimate are obtained.

2. The method for predicting the effect of air flotation-enhanced oil-water separation based on small samples according to claim 1, characterized in that: In step S100, the original characteristics include at least the initial oil content, pH value, temperature, dissolved gas content, oil layer thickness after separation, diameter of the separation equipment, height of the separation equipment, and height of the inlet position of the influent.

3. The method for predicting the effect of air flotation-enhanced oil-water separation based on small samples according to claim 1, characterized in that: In step S200, the physical interaction feature vector is: in: ψ1: Temperature-standardized interaction term between mass and equipment dimensions; ψ2: Logarithmic interaction term between pH and oil layer thickness, including location offset penalty; ψ3: The product of dissolved gas ratio and reservoir coverage; ψ4: Trigonometric interaction term between temperature and pH; to These represent the initial oil content of the influent, pH value, temperature, dissolved air content, oil layer thickness after separation, diameter of the separation equipment, height of the separation equipment, and height of the inlet position, respectively.

4. The method for predicting the effect of air flotation-enhanced oil-water separation based on small samples according to claim 1, characterized in that: In step S300, the base model is a deep resilient network architecture, which includes a feature transformation layer and a resilient network prediction head. The feature transformation layer includes multiple fully connected hidden layers, wherein the output of the l-th layer... Defined by the following formula: in ; This is the output of the l-th hidden layer; Here is the weight matrix for the l-th layer; Here, represents the bias vector of the l-th layer; ReLU is the activation function of the rectified linear unit. An elastic network prediction head is used to output the master prediction value. Its definition is: is the predicted output; L is the index of the last hidden layer; w is the output layer weight vector; b is the output layer bias.

5. The method for predicting the effect of air flotation-enhanced oil-water separation based on small samples according to claim 1, characterized in that: In step S300, a multi-objective loss function is designed for the base model. The multi-objective loss function It is the weighted sum of the following loss terms: ① Main task elastic network loss Its definition is: N is the number of samples; λ is the regularization strength; This is the regularization ratio; ② Weight constraint regularization loss This is used to apply Frobenius norm penalty to the weight matrix of each layer of the feature transformation layer; ③ Physical constraint losses, including: Monotonicity constraint loss Used to constrain predicted values Oil content in influent Monotonically increasing and the effect on dissolved gas content Monotonically decreasing; Boundary constraint loss Used to constrain predicted values Non-negative and not greater than the oil content of the influent ; ④ Losses from auxiliary tasks It is achieved by using a hidden layer in the middle. Draw out auxiliary output And the mean square error between it and the true value is calculated.

6. The method for predicting the effect of air flotation-enhanced oil-water separation based on small samples according to claim 1, characterized in that: In step S500, the final predicted value and uncertainty estimation Its definition is: Where M is the number of base models.

7. A prediction model for the enhanced oil-water separation effect of air flotation based on small samples, characterized in that: The method for predicting the effect of air flotation-enhanced oil-water separation based on a small sample, according to any one of claims 1-6.

Images