Construction of an AHP-based Evaluation Index System for the Operational Status of Artificial Wetlands

Artificial wetlands refer to comprehensive ecosystems that simulate the structure and function of natural wetlands. They achieve water quality purification and ecological improvement through the synergistic effects of physics, chemistry, and biology. They play an important role in improving the regional water ecological environment, ensuring water safety, and promoting water resource recycling. The operating system of artificial wetlands is shown in Fig. 1.

In Fig. 1, when the artificial wetland is well managed and maintained, it can effectively purify water quality and maintain ecological balance. At this point, plants, microorganisms, and substrates in the wetland are in good condition and can work together to remove pollutants from the wastewater. In some cases, artificial wetlands may be overloaded due to excessive wastewater treatment or pollutants. This may lead to the disruption of ecological balance within wetlands, a decrease in water purification function, and even the problem of effluent indicators being inferior to inflow indicators. Therefore, to improve the operational status and effectiveness of artificial wetlands, research is needed to establish an environmental status assessment model for artificial wetlands, monitor the wetland status in real-time, and promote the sustainable development of artificial wetlands. The AHP decomposes complex problems into various constituent elements based on data and evaluation criteria, grouping them into ordered hierarchical structures according to their dominant relationships. Next, based on a certain ratio calibration, the judgment is quantified through pairwise comparison, forming a comparative judgment matrix. Ultimately, this method boils down the system analysis to determine the relative importance weights of the lowest level relative to the overall goal or the sorting problem of the relative optimal order, thereby providing a basis for selecting decision-making options [20-22]. The architecture of the AHP is shown in Fig. 2.

Fig. 2 presents the architecture of the AHP, which is according to the problem’s nature and the overall goal. The problem is broken down into individual factors, which are then combined and aggregated at various levels based on their interrelationships and membership relationships, creating a multi-level analytical structure model. The AHP method finds the eigenvector corresponding to its maximum eigenvalue, as indicated in Equation (1).

Fig. 1.

Operation system of constructed wetland.

Fig. 2.

Analytic hierarchy process architecture.

In Equation (1), λ_max represents the eigenvalues, A represents the corresponding eigenvectors, and P represents the matrix. Next, it normalizes A and uses the normalized values of A to represent the relative coefficients of different indicator weights. This study uses the sum product method to calculate the eigenvectors corresponding to the maximum eigenvalue of P, as indicated in Equation (2).

In Equation (2), (AP)_i means the i-th element in the AP result. Consistency testing is an important method for evaluating the rationality of judgment matrices. The judgment matrix is usually used to decide the relative importance of each indicator and is an important component of decision analysis methods such as AHP. To determine whether the matrix is reasonable, a consistency test is conducted as shown in Equation (3).

In Equation (3), C₁ represents the consistency index and n represents the order of the judgment matrix P. Research determines the weights of each evaluation indicator through AHP. Based on expert opinions and data analysis, a hierarchical structure is constructed, followed by the construction of pairwise judgment matrices, and the weights of each indicator are worked out using the feature vector method of AHP. The evaluation index system for the operational status of artificial wetlands is a comprehensive evaluation framework that comprehensively measures the operational efficiency and management effectiveness of artificial wetlands. This system integrates multiple key element layers for detailed analysis and evaluation of the overall performance of wetlands [23, 24].

The basic principles of constructing the evaluation index system of constructed wetlands include scientificity, systematicness, operability, dynamics, and comparability. The index selection has been carefully considered when constructing the wetland evaluation index system. First, a series of key indicators are identified based on wetland ecosystems’ key functions and values, such as water purification, biodiversity conservation, flood regulation, etc. Secondly, through expert consultation, literature review, and field research, the existing knowledge and practical experience on wetland assessment were collected to ensure the rationality and validity of the selected indicators. In addition, the sensitivity and responsiveness of the indicators are also considered, that is, whether they can reflect the change of wetland environmental state in a timely manner. In order to ensure the scientific rationality of indicator selection, correlation analysis of indicators is also carried out to avoid selecting highly correlated indicators to reduce redundancy and improve the efficiency of the indicator system. At the same time, through pre-evaluation and pilot study, the preliminary selected indicators are tested and validated to evaluate their effectiveness and reliability in practical applications. Finally, according to the test results and feedback, the index system was optimized and adjusted to ensure its applicability and accuracy in the environmental state assessment of constructed wetlands. The specific evaluation index system for the operation status of artificial wetlands is shown in Fig. 3.

In Fig. 3, wetland indicators mainly focus on wetlands’ physical and biological characteristics. Among them, the ponds’ design involves the wetlands’ layout, size, and shape, directly affecting the wetlands’ water flow pattern and purification capacity. Filler design focuses on wetland materials, significantly impacting the wetland’s biodiversity and water purification efficiency. Plant design involves wetland plants’ types, distribution, and density, which play important roles in wetland ecosystems, including providing habitat and absorbing pollutants. Operational indicators focus on the wetlands’ various parameters and conditions during operation. Hydraulic load refers to the ability of wetlands to process water, which directly affects the purification effect of wetlands. The hydraulic retention time determines the flow rate and retention time of water in wetlands, which is crucial for the removal efficiency of pollutants. Temperature and pH are important factors affecting wetland biological activity and chemical reactions, while dissolved oxygen is related to the respiration and metabolism of microorganisms in wetlands. In addition, water quality stability is a critical indicator for evaluating wetland operational performance. The effluent quality indicators directly reflect the removal efficiency of pollutants by wetlands.

Fig. 3.

Evaluation index system of operation state of constructed wetland.

COD is a crucial indicator for measuring the degree of organic pollution in water, while ammonia nitrogen, total nitrogen, and total phosphorus are key parameters for assessing the risk of eutrophication in water bodies. Management indicators focus on the daily management and maintenance of wetlands. Online monitoring can provide real-time monitoring of the operational status of wetlands, enabling timely detection and resolution of issues. Plant management involves pruning and replanting wetland plants to ensure the health and stability of wetland ecosystems. Congestion management aims to prevent and address potential blockage issues in wetlands, ensuring their normal operation [25, 26]. The single-level ranking of indices for every level can be obtained through the judging matrix. This ranking is then used to calculate the overall ranking of the relative target layer. Table 1 shows the weight of index layer C relative to target layer A.

In Table 1, the weight values of 8 indicators, including filler design, plant design, hydraulic load, hydraulic retention time, water quality stability, effluent COD, TN, and TP, are above 0.05, which are relatively critical and have a significant influence on the operation status of artificial wetlands. When conducting hierarchical analysis, the dimensions and degree of variation of different indicators may lead to bias in the results. Therefore, standardization or dimensionless processing of data is crucial. Standardization can not only eliminate the influence of dimensionality but also make different variables numerically comparable, thus more accurately reflecting the relative importance between indicators. For qualitative data, it is usually necessary to assign values based on scoring criteria or expert scoring methods to convert them into numerical values that can be used for mathematical operations. For quantitative data, actual measured values can be directly used. A common method in data standardization is “minimum-maximum normalization”, also called dispersion normalization. This method requires performing a linear transformation on the initial data, mapping the resulting values to a range between 0 and 25. For the case where a larger indicator value is better, the conversion function is shown in Equation (4).

In Equation (4), X_i indicates the raw data, X ⁱ_min and X ⁱ_max are the minimum and max values in the data, and X̂ ' is the standardized data. For the case where a smaller indicator value is better, the conversion function is shown in Equation (5).

When conducting hierarchical analysis, the standardized data can be directly utilized to build a judgment matrix and then calculate the weights of each indicator. This method can eliminate the impact of dimensionality and degree of variation, making the analysis results more accurate and reliable.

Construction of Artificial Wetland Operation Status Model Based on BPNN

The study uses the evaluation results obtained from the previous section of the AHP method as the learning and training samples for the BPNN, takes the artificial wetland data reflecting various indicators as the inputting of the BPNN, and uses the corresponding evaluation results obtained from the AHP method as the outputting of the BPNN. In order to enhance the prediction ability of the model, more dimensions of environmental data are introduced. Integrating climate, land-use change, biodiversity, hydrology, water quality, and biological activity, data enables models to more accurately capture the complexity and dynamic changes in the environmental state of constructed wetlands, improve prediction accuracy, and provide scientific and real-time decision support for wetland management. Data preprocessing is a key step in building efficient and accurate models. In order to improve the performance of the model, wavelet transform technology is introduced, which is excellent in processing non-stationary signals and can effectively separate the trend and periodic components from the original data to reduce noise interference. In addition, using principal component analysis to reduce dimensionality, PCA helps simplify model complexity by extracting the major variation factors in the data while preserving the most critical information. This method not only reduces the computational burden of the model but also improves the robustness to outliers and noise. In order to further enhance the generalization ability of the model, the data enhancement technique is used to expand the training set. By applying various transformations, such as scaling, rotating, and adding slight noise, data enhancement techniques are able to generate new training samples without actually collecting more data. Combining these advanced data processing methods, the model can extract useful information from complex environmental data more effectively and maintain stable predictive performance under diverse environmental conditions.

BPNN is a supervised learning algorithm that utilizes gradient descent. It is a multi-layer feedforward neural network that typically includes inputting, hidden, and outputting layers. The training process of BPNN can be broken down into two sections: forward and backward propagation. During forward propagation, inputting samples are passed through the inputting layer of the network to the hidden layer (HL) and outputting layer, and the outputting value of each neuron is calculated. The outputting value is mapped nonlinearly through an activation function. Back propagation reduces errors by adjusting network parameters based on the error between the outputting and the actual values. Firstly, it calculates the outputting error and then uses the chain rule to calculate the error contribution of each neuron in the order from the outputting layer to the inputting layer and updates the weights and biases. The training of BPNN is an iterative process, which continuously adjusts weights and biases to make the network’s outputting approximate the actual value [27-29]. It assumes that there are n nodes in the inputting layer, m nodes in the outputting layer, and l nodes in the HL. The excitation function is the functional relationship between the inputting and outputting of hidden and outputting layer nodes in a neural network, as shown in Equation (6).

In Equation (6), g(x) represents the excitation function. The outputting of the HL is shown in Equation (7).

In Equation (7), w_ij indicates the weight from the inputting layer to the HL, H_j represents the HL’s outputting, and a_j represents the biased term of the j-th neuron. The outputting of the outputting layer is indicated in Equation (8).

In Equation (8), O_k represents the output value of the k-th output neuron, b_k represents the offset term of the k-th output neuron, w_jk represents the weight from the outputting layer to the HL. The error back propagation process of BPNN is actually a process in which the outputting error of the network is reversed layer by layer through the HL to the inputting layer in a certain form, and the error is distributed to the neurons and neural units of each layer. This process is carried out in a loop, and the weights are constantly adjusted until the error in the network outputting is reduced to an acceptable level or until the pre-set number of learning iterations is reached. In BPNNs, the error is determined by the difference between the actual outputting of the outputting layer and the expected outputting. To reduce this difference, the network will adjust the weights of each neuron layer by layer based on the path of error back propagation, to minimize the total error. This adjustment process is grounded on the gradient descent method, which calculates the gradient of the error function on the weight and adjusts the weight along the gradient descent direction to gradually reduce the error. In training, the network continuously learns the mapping relationship between inputting data and expected outputting and updates the weights through back propagation, gradually making the network’s outputting approach the expected outputting. When the network reaches a certain number of training iterations or an acceptable level of error, the training process ends, and the network weights are fixed, which can be used for predicting new data [30]. The flowchart of the BPNN is shown in Fig. 4.

In Fig. 4, the inputting data is propagated through the network, and the outputting values of the HL and outputting layer are calculated. It calculates the error by comparing the predicted and actual values of the outputting layer and updates the weights and bias values using gradient descent. Forward propagation and backward propagation are repeated until the stop condition is arrived. The number of input layer nodes depends on the dimension of the inputting vector, and the optimal number of HL nodes can be determined, as denoted in Equation (9).

In Equation (9), m', n', and l represent the amount of hidden, inputting, and outputting layer nodes, respectively. a is a constant from 1 to 10. However, the BPNN has some obvious drawbacks. The optimization of the BPNN relies on the gradient descent method, which is susceptible to becoming trapped in local minima while searching for the global optimal solution. This can cause the network to fail to find the true optimal solution. Therefore, the study adopts an adaptive mutation genetic algorithm (GA) to raise the BPNN model. GA is an optimization algorithm that simulates natural selection and genetic mechanisms. It can perform a global search in the search space and has a certain degree of adaptability and robustness. The adaptive mutation GA can dynamically adjust the mutation rate based on the problem’s characteristics and the search’s progress, thereby improving the algorithm’s performance. In the search for the optimal global solution using GAs, the adaptability of each living organism is dynamically changing. Therefore, it is recommended that adaptive mutation probability (MP) be adopted. When the population’s fitness is poor, MP should be enhanced to increase the population’s diversity and increase the number of excellent individuals. Conversely, the MP should be lowered when the fitness is good, i.e., close to the globally optimal solution [31]. The calculation of adaptive MP P' is denoted in Equation (10).

In Equation (10), P₀ is the assumed initial MP, P_min is the MP range’s min value, and F̅ and max F X( _K )

X_K∈Ω are the population’s average and max fitness values, respectively. M is the max evolution generation and m'' is the current evolution generation. The research uses adaptive mutation GA to improve the BPNN model. There is a positive correlation between the amount of experimental samples and the accuracy of reaction results. However, when the sample size reaches a certain level, the accuracy will remain stable within a range, and no significant changes will occur. The larger the network size, the more complex the mapping relationship of the network. There are two common methods for selecting initial weights: one is to choose a sufficiently small initial weight, and the other is to make the number of initial weights equal to +1 and -1. Typically, several networks can be trained, and the most appropriate one can be chosen based on the analysis findings. In this study, a single HL neural network vector model was used in MATLAB, as depicted in Fig. 5.

Fig. 4.

BPNN flowchart.

In Fig. 5, P̂ is the inputting vector with a size of R × 1, as denoted in Equation (11).

b₁ is the threshold vector of the inputting layer neurons, with a size of S₁ × 1, as denoted in Equation (12).

IW₁ means the connection weight vector between the inputting layer neurons and the inputting vector, with a size of S₁ × R, as denoted in Equation (13).

n₁ is the intermediate calculation outcome of the first layer neuron, which is the weighted sum of the connection weight vector and threshold vector, with a size of S₁ × 1, as denoted in Equation (14).

a₁ means the outputting direction of the first layer neuron, with a size of S₁ × 1, as denoted in Equation (15).

The objective of the GA is to minimize the sum of squared errors in all evolutionary generations by obtaining the network weight and threshold. GA progresses towards higher fitness function values. Therefore, the fitness function is defined as the inverse of the individual’s learning error, as shown in Equation (16). Therefore, the fitness function is defined as the inverse of the individual’s learning error, as shown in Equation (16).

In Equation (16), f_fitness stands for fitness and E represents the learning error. The model uses an adaptive MP mutation operation to strengthen the diversity of the GA population, enabling it to jump out of local optimal solutions and search for global optimal solutions in a timely manner, thereby avoiding the occurrence of premature convergence [32]. The algorithm model is shown in Fig. 6.

Fig. 6 illustrates that the optimal solution is searched by a GA and inputted into the BPNN as the initial weights and thresholds of the network. The model’s data flow is as follows: Starting from the BPNN section, learning samples are determined using the data collected from the survey questionnaire, and the topological structure of the neural network is identified, including the number of network layers and neurons. Afterward, the initial population of the adaptive mutant GA is obtained. The adaptive mutation process in GA involves encoding, fitness calculation, and genetic operations to determine the optimal weight and threshold that satisfy the stopping conditions. The initial weights and thresholds obtained can then be utilized by the BPNN

Fig. 5.

Vector model of multilayer neural network.

BPNN part

Adaptive mutation GA part to shorten the time required to find the optimal weights and thresholds, thereby speeding up the convergence speed of the network. After meeting the learning error or iteration requirements, an optimal GA-BP algorithm model can be generated. In order to further improve the training efficiency and performance of the model, L1 and L2 regularization techniques were first introduced to effectively control the complexity of the model and reduce the risk of overfitting. L1 regularization makes some unimportant weights zero by punishing the absolute value of weights, thus realizing the effect of feature selection. L2 regularization makes the weights tend to smaller values by punishing the sum of squares of the weights, increasing the model’s stability. At the same time, Adam and RMSprop advanced optimization algorithms have been adopted to optimize the model. The Adam algorithm combines AdaGrad and RMSProp advantages, adaptively adjusts each parameter’s learning rate, and makes the model converge more efficiently in the training process. The RMSprop algorithm provides a separate learning rate for each parameter by adjusting the cumulative average gradient square, which is especially suitable for dealing with non-stationary targets. Applying these optimization algorithms not only speeds up the convergence speed of the model but also improves the model’s ability to find the optimal solution in the complex data distribution.

Fig. 6.

Model of GA-BP algorithm.

GA-BP Algorithm Model Performance

Research collected environmental status data of artificial wetlands, including indicators such as water quality, biodiversity, and soil conditions, and performed preprocessing and normalization. The optimized BPNN was used for training and testing, and a simulation experiment based on an improved GA-optimized BPNN for evaluating the environmental status of artificial wetlands was carried out using Matlab2013b. Combining GA-BPNN with BPNN, Extreme Learning Machine (ELM), Support Vector Machine (SVM), Random Forest (Random Forest RF), and Gradient Boosting Machine (GBM) for comparative analysis, the accuracy rate, recall rate, and F1 values of several algorithms are shown in Fig. 7.

Fig. 7.

The accuracy rate, recall rate, and F1 value of the six algorithms.

In Fig. 7a), GA-BPNN and BPNN rise faster and quickly approach higher accuracy rates. With the increase in the number of iterations, the accuracy of GA-BPNN tends to be stable. It reaches a high accuracy of close to 1.0 after about 100 iterations, which shows its fast convergence characteristic during training. In contrast, the accuracy of other algorithms, such as ELM, SVM, RF, and GBM, increases gently, fluctuates greatly in the iterative process, and finally stabilizes between 0.6 and 0.8. This shows that GA-BPNN has better learning and generalization abilities when evaluating environmental states. In Fig. 7b), GA-BPNN also performs well in the recall rate, and its curve rises rapidly in the early iterations and stabilizes above 0.8 after about 100 iterations, showing its high efficiency in identifying positive samples. BPNN also had a higher recall rate but was slightly lower than GA-BPNN. The advantages of GA-BPNN in processing unbalanced data sets are confirmed, especially in improving the recognition rate of positive samples. In Fig. 7c), GA-BPNN also had an outstanding performance in the F1 score. Its curve rose rapidly in the early stages of iteration. It stabilized above 0.8 after about 100 iterations, showing its superiority in its comprehensive consideration of accuracy and recall rate, as well as its comprehensiveness and stability. To further prove the model’s high accuracy of the raised model, the study analyzed and compared the training and learning results, as denoted in Fig. 8.

In Fig. 8a), the error of the BPNN model was relatively large, with some data falling outside the error range of plus or minus 0.2. In Fig. 8b), 80% of the data fell within the error range of plus or minus 0.2, indicating that the learning speed of the model has been greatly improved. That is to say, the model learned most of the data patterns in relatively few iterations. The study predicted the evaluation results of the last 20 sample data sets, as shown in Fig. 9.

In Fig. 9a), the prediction error of the BPNN model was relatively large. In Fig. 9b), the predicted results were basically consistent with the actual results, and the fitting degree reached over 90%. This indicated that GA optimization has played a critical role in improving the performance of BPNN models. The regression results could reflect the overall goodness of fit of the evaluation model, as indicated in Fig. 10.

In Fig. 10, regardless of the overall sample, training sample, or test sample, the regression results of the data showed excellent fit and avoided overfitting issues. This fully demonstrated that the model had an excellent ability to capture the inherent patterns and information of data, thereby being able to predict and interpret the dynamic behavior of data more accurately. The calculated R² value was as high as 0.99163, and the MSE was also approaching zero. The R² value was extremely close to 1, proving the effectiveness of the model-fitting effect.

Fig. 8.

Model training learning results.

Fig. 9.

Neural network model prediction results.

Fig. 10.

Model regression result.

Example Analysis of Artificial Wetland Environmental Status Assessment Model

The study preprocessed qualitative and quantitative indicator data to conduct a comprehensive evaluation of the environmental status of artificial wetlands. It multiplied the processed data with the weights of various influencing factors to obtain comprehensive evaluation outcomes, as indicated in Table 2.

In Table 2, the comprehensive score of the established wetland was low, indicating that the overall environmental status of the wetland is not ideal and there may be a series of problems. The wetland may suffer from pollution from surrounding areas, and comprehensive measures need to be taken to raise the environmental status of the constructed wetland. Therefore, further analysis of the COD purification effect of pollutants was conducted, as shown in Fig. 11.

In Fig. 11a), the range of COD concentration in the influent of artificial wetland A was large, ranging from 25 mg/L to 142 mg/L, indicating significant fluctuations in influent water quality. The COD concentration range in the effluent was relatively narrow, ranging from 8 mg/L to 34 mg/L, and overall lower than that in the influent, indicating that the wetland system had a certain removal effect on COD. The average COD removal rate was 39%; although there was some removal effect, the removal rate was not high. The COD value of the effluent was lower than that of the inflow, indicating that the wetland system could play a certain role in COD removal most of the time. However, due to significant fluctuations in influent water quality and low removal rates, wetland systems may not fully meet water quality requirements in certain situations. In Fig. 11b), the COD concentration range of the influent in constructed wetland B ranged from 40 mg/L to 160 mg/L, with a slightly smaller fluctuation range compared to wetland A. The COD concentration range in the effluent was relatively large, ranging from 16 mg/L to 122 mg/L, indicating that the wetland system had an unstable COD removal effect. The average COD removal rate was 21%, which was lower than wetland A, indicating that wetland B had poor performance in COD removal. Further analysis was conducted on the purification effect of pollutants COD along the process, as shown in Fig. 12.

Fig. 11.

Purification effect of pollutant COD.

In Fig. 12a), artificial wetland A showed a slight increase in COD content in spring, mainly due to incomplete harvesting of plants in winter, resulting in residual plants starting to rot in spring. The decay of plant residues released organic matter into wetlands, thereby increasing the content of COD. In Fig. 12b), the secondary subsurface flow tank and the primary surface flow had poor performance in COD removal. Especially in winter and spring, the COD content in the effluent was even higher than that in the inflow, indicating that the water quality of the constructed wetland in this section has been polluted during these two seasons. The ecological and functional indicators of the constructed wetland environmental status evaluation model (Model 1) were compared with the wetland monitoring and evaluation model based on remote sensing technology (Model 2) and the wetland health evaluation model based on the comprehensive index method (Model 3), as shown in Fig. 13.

Fig. 12.

Purifying effect of pollutant COD along the way.

Fig. 13.

Comparison of ecological index and functional index of three models.

In Fig. 13a), Model 1 had the highest ecological index, averaging over 93%. In Fig. 13b), Model 1 had the highest functional indicators, averaging over 92%. Model 1 provides strong support for wetland protection and management and provides a reference for improving and optimizing other wetland models. A comprehensive performance comparison of different algorithms in the environmental state assessment of constructed wetlands is shown in Table 3.

In Table 3, from the stability index, the GA-BPNN algorithm leads with a score of 0.915, showing that it has the best stability under different test conditions. In terms of MSE, GA-BPNN also performs best with a low error value of 0.004, indicating its high prediction accuracy. The GA-BPNN algorithm performs well on several key performance indicators, especially in stability, prediction accuracy, anti-noise ability, and cross-domain generalization ability. This shows that GA-BPNN is a powerful tool for environmental assessment tasks that require high precision and efficiency. Despite its high model complexity, high scores on other performance metrics suggest that it is a comprehensive and effective choice.