A method for predicting the lethality of sediment-laden water flow on fish

Abstract

This invention relates to a method for predicting the lethal effects of sediment-laden water flow on fish, comprising: data acquisition; model experiments; construction of a backpropagation neural network model; training of the backpropagation neural network model; model testing; and prediction application. This invention utilizes a backpropagation neural network model and an improved particle swarm optimization algorithm to predict the lethal effects of sediment-laden water flow on fish. By adjusting the backpropagation neural network model through multiple error calculations, the prediction accuracy is improved, keeping the prediction error within ±6%. This invention can be used to assess the impact of high-sediment-laden water flow processes on fish during reservoir sediment discharge and dam removal. Compared to existing assessment methods such as SI and SEV, this invention can comprehensively consider the influence of various environmental factors, significantly improving prediction accuracy and providing a new means and basis for assessing the aquatic ecological impact of reservoir sediment discharge.

Description

Technical Field

[0001] This invention relates to a method for predicting the lethal effects of sediment-laden water flow on fish, which is an ecological assessment method and an assessment method for evaluating the ecological impact of water conservancy project construction and operation on rivers. Background Technology

[0002] During the operation of a reservoir, hydraulic sediment removal is often necessary to reduce sedimentation and extend its lifespan. During decommissioning and dismantling, it's crucial to remove the dam and restore river continuity. The high-sediment-laden flow generated during sediment removal can cause drastic fluctuations in sediment concentration and dissolved oxygen levels in downstream channels within a short period, potentially leading to mass mortality of fish and other aquatic plants and animals. Fish occupy a top position in the river's aquatic ecosystem, making significant contributions to the river's material and energy cycles and serving as indicator organisms reflecting the river's ecology.

[0003] Currently, there are two main types of methods for assessing the impact of suspended sediment on fish: The first type estimates fish mortality based on the possible average or maximum sediment concentration discharged from the reservoir. However, using a constant sediment concentration threshold to determine the degree of fish impact obviously introduces many uncertainties. The second type is a statistical analysis method that establishes the correlation between evaluation factors and target factors through sample data, including the Stress Index (SI) assessment model and the Severity of Ill Effects (SEV) assessment model. These two models use suspended sediment concentration and fish exposure time as evaluation factors, which can more accurately assess the degree of impact on fish than using sediment concentration alone. However, these two models have significant drawbacks: because they are too simplistic, only considering sediment concentration and its duration, they fail to consider factors such as dissolved oxygen and water temperature, which play a crucial role in fish survival. This makes it difficult for these models to accurately reflect the survival status of fish, and their use is subject to many limitations. Since both models assess the impact of high sediment concentration water bodies on fish based on measured data, they cannot predict the impact of high sediment concentration water bodies on fish under different possible concentration conditions, i.e., they lack predictive assessment capabilities.

[0004] How to accurately assess the impact of the complex nonlinear relationships between multiple factors in sediment flow on the fish ecosystem is a problem that needs to be solved. Summary of the Invention

[0005] To overcome the problems of existing technologies, this invention proposes a method for predicting the lethal effects of sediment-laden water flow on fish. The method proposes a prediction approach for the impact of high-sediment-laden water bodies on fish based on a backpropagation neural network model. This method can consider complex nonlinear problems between multiple factors. By learning from sample data using the backpropagation neural network model, it can not only assess the impact on fish based on measured data, but also predict the degree of impact on fish under different possible sediment concentration scenarios.

[0006] The objective of this invention is achieved as follows: a method for predicting the lethality of fish caused by sandy water flow, the steps of which are as follows:

[0007] Step 1, Data Collection: Collect various data of the research subjects, including: fish species, sand content, median particle size of sediment, water temperature, oxygen solubility, and time, to form the original dataset;

[0008] Step 2, Model Experiment: Construct an experimental model simulating the real environment based on the original dataset. Simulate the living environment of the research subjects in the experimental model and collect various experimental data. The experimental data should include at least: fish species and size, duration, sand content and sand particle size distribution, water temperature, and dissolved oxygen. Construct an experimental dataset of the mortality rate of the research subjects in sandy water bodies based on the experimental data.

[0009] Step 3, Construct the backpropagation neural network model: Construct a 3-layer backpropagation neural network containing an input layer, a hidden layer, and an output layer. The training algorithm is the Bayes regularization algorithm, the maximum number of iterations is 20,000, and the learning rate is set to 0.01 based on experience.

[0010] 1) Input layer: The input parameters include six experimental observation data points: fish species, duration, sediment content, median particle size of sediment, water temperature, and dissolved oxygen. The hyperbolic tangent function Tansig is used as the transfer function between the input layer and the hidden layer. The original experimental information is recorded using the six experimental observation data points for each sample and then passed to the neural network.

[0011] 2) Hidden Layer: Through trial and error, the number of neurons in the hidden layer was determined to be p = 12, meaning there are a total of 12 neurons in the hidden layer; The trial calculation formula is as follows:

[0012]

[0013] In the formula: p is the number of neurons in the hidden layer; m is the number of parameters in the input layer, which is taken as m=6; n is the number of parameters in the output layer, which is one parameter for fish mortality, so n=1; A is a constant between [1, 10].

[0014] 3) Output layer: The output value of the output layer needs to be denormalized, and the output result is the fish mortality rate; the linear function Purelin is selected as the transfer function between the hidden layer and the output layer;

[0015] 4) Assessing the predictive ability of the backpropagation neural network model: The predictive ability of the backpropagation neural network model is evaluated using the root mean square error, the absolute value of the mean relative error, and the Nash efficiency coefficient. The calculation formulas are as follows:

[0016] Root mean square error:

[0017]

[0018] Mean relative absolute value:

[0019]

[0020] Nash efficiency coefficient:

[0021]

[0022] In the formula: y i These are the output layer values ​​of the neural network and the measured values ​​of the experimental dataset, respectively, representing the simulated and measured values ​​of fish mortality rates. is the average of the expected values, i.e., the average of the measured fish mortality rates; i is the number of sample data groups; n is the total number of sample data.

[0023] Step 4, train the backpropagation neural network model:

[0024] 1) Select array data from the experimental dataset, perform normalization processing, and form a training sample dataset;

[0025] 2) Import the training sample dataset into the backpropagation neural network model and initialize the particle swarm optimization within it;

[0026] 3) Utilize an improved particle swarm optimization algorithm to call the backpropagation neural network model, using the output error of the backpropagation neural network model as the fitness function; Fitness function F:

[0027]

[0028] In the formula: y ki , respectively, represent the output value of the backpropagation neural network and the expected value in the training sample dataset, i.e., the predicted and measured values ​​of fish mortality; k is the number of iterations in the improved particle swarm optimization algorithm;

[0029] The improved particle swarm optimization algorithm and backpropagation neural network couple information through particle information and fitness value. By iteratively seeking the particle swarm with the optimal fitness, the parameters that minimize the error of the backpropagation neural network model are obtained.

[0030] 4) Determine if the maximum number of iterations or the preset precision has been reached; otherwise, return to step 3).

[0031] 5) Assign the optimized weights and thresholds to the backpropagation neural network, and then train the backpropagation neural network using gradient descent.

[0032] Step 5, Model Testing: Select another array of data from the experimental dataset to form a test sample dataset. Input the test sample dataset into the backpropagation neural network model trained in Step 4 to test the backpropagation neural network and observe whether it meets the requirements. If it does not meet the requirements, adjust the backpropagation neural network parameters and return to Step 4 to continue training.

[0033] Step 6, Prediction Application: Based on the prediction requirements, multiple sets of environmental variable data are generated in conjunction with the reservoir sediment discharge process to be predicted. These sets of data are then normalized and input into the trained backpropagation neural network model to output the corresponding predicted fish mortality rate.

[0034] The advantages and beneficial effects of this invention are as follows: This invention utilizes a backpropagation neural network model and an improved particle swarm optimization algorithm to predict the lethal impact of sediment-laden water flow on fish. Furthermore, by adjusting the backpropagation neural network model through multiple error calculations, the prediction accuracy is improved, keeping the prediction error within ±6%. This invention can be used to assess the impact of high-sediment-laden water flow processes on fish during reservoir sediment discharge and dam removal. Compared to existing assessment methods such as SI and SEV, this invention comprehensively considers the influence of various environmental factors, significantly improving prediction accuracy and providing a new means and basis for assessing the aquatic ecological impact of reservoir sediment discharge. Attached Figure Description

[0035] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0036] Figure 1 This is a flowchart of the method described in the embodiments of the present invention;

[0037] Figure 2 This is the experimental sand particle size distribution of the application example described in the embodiments of the present invention;

[0038] Figure 3 This is the test condition setting table for the application example described in the embodiments of the present invention;

[0039] Figure 4The application example described in this embodiment of the invention uses a comparison table of fitting equations between experimental environmental factors and fish mortality rates;

[0040] Figure 5 This is a comparison curve of partial sample data used for training and prediction results in the application example described in the embodiments of the present invention;

[0041] Figure 6 This is a training error analysis diagram of the improved particle swarm optimization algorithm-backpropagation neural network model in the application example described in the embodiments of the present invention;

[0042] Figure 7 Table 3, which is an application example described in the embodiments of the present invention, is a comparison table of experimental data and predicted values ​​of the improved particle swarm algorithm-backpropagation neural network model. Detailed Implementation

[0043] Example:

[0044] This embodiment describes a method for predicting the lethality of fish caused by sediment-laden water flow. The steps of the method are as follows, and the flowchart is shown in Figure 1:

[0045] Step 1, Data Collection: Collect various data of the research subjects, including: fish species, sand content, median particle size of sediment, water temperature, oxygen solubility, and duration, to form the original dataset.

[0046] This step is used to collect data on the living environment of the research subjects (target fish), especially the survival status of the target fish in the high sediment content water body formed during reservoir sediment discharge, i.e., the lethal process of high sediment content water flow, in order to facilitate indoor experiments.

[0047] Application Example: To quantitatively assess the impact of high sediment content water bodies on fish, this application example uses two representative fish species from the Yellow River main stream, the upper Yellow River naked carp (Gymnocypris Eckloni) and the middle and lower Yellow River carp (Cyprinus Carpio), as the research subjects to study their survival characteristics in high sediment content water bodies.

[0048] Duration refers to the period of time during which the target fish species, as the subject of the study, is exposed to water with high sediment content and experiences a certain mortality rate. For example, the spotted naked carp exposed to 64.17 kg / m³ of sediment... 3 In 140 minutes in the water, 10% of the spotted naked carp died; this 140 minutes is referred to as the "duration".

[0049] Step 2, Model Experiment: Construct an experimental model simulating the real environment based on the original dataset. Simulate the living environment of the research subjects in the experimental model and collect various experimental data. The experimental data should include at least: fish species and size, duration, sand content and sand particle size distribution, water temperature, and dissolved oxygen. Construct an experimental dataset of the mortality rate of the research subjects in sandy water bodies based on the experimental data.

[0050] The experimental model was a 1m × 1m × 0.8m (length × width × height) acrylic glass experimental tank. Two groups of fish were divided into two experimental groups, and both experiments were conducted in this tank. A variable-speed motor-driven propeller was installed at the bottom of the tank to ensure the suspension of sediment. The sand used in the experiment was taken from the capture location of the corresponding fish species. The particle size distribution curve is shown below. Figure 2 As shown. A total of 16 groups of two types of fish were tested. The fish used in the experiment weighed 0.5-1 kg, and 10 fish were used in each experiment. The conditions of the experimental tanks were kept consistent with the natural conditions of the fish's habitat. The initial water temperature, dissolved oxygen, and sand used in the experiment were based on the conditions of the upper and middle and lower reaches of the Yellow River, respectively. The experimental setup is shown in [reference needed]. Figure 3 Table 1 shows the results. Water temperature and dissolved oxygen were measured every 15–30 minutes during the experiment. An additional set of measurements was taken when the fish died.

[0051] Data from 16 experimental scenarios were used to determine the relationships between mortality rates of Yellow River carp and spotted naked carp and various environmental factors, including the duration of exposure to high-sediment-content water, sediment concentration, dissolved oxygen, and water temperature. Samples from a single experimental scenario and samples from different scenarios showed the same affinity, equivalent to conducting 439 separate physical experiments, recording fish mortality rates each time the tank parameters were identical to the corresponding sample data.

[0052] The linear correlation equation between the mortality rate of the experimental fish and the sand content can be obtained from the experimental data, such as... Figure 4 Table 2 shows the equations fitted using each environmental factor, with R0... 2 The values ​​are all less than 0.5, indicating that the linear combination of environmental factors has poor explanatory power for mortality and cannot fully account for the mechanism by which mortality is constrained by multiple environmental factors. Therefore, it is not appropriate to use linear fitting equations to estimate fish mortality.

[0053] Step 3, Construct the backpropagation neural network model: Construct a 3-layer backpropagation neural network containing an input layer, a hidden layer, and an output layer. The training algorithm is the Bayes regularization algorithm, the maximum number of iterations is 20,000, and the learning rate is set to 0.01 based on experience.

[0054] To accurately assess the impact of various factors such as fish species, duration, sediment content, dissolved oxygen, and water temperature on the mortality rate of experimental fish, an application example uses an artificial neural network to predict the lethal effects on fish. Based on the measurement data of 16 sets of experiments, 439 sample data points were determined for training, validating, and testing the neural network. Each measurement data point consists of 7 variables: fish species, duration, sediment content, median particle size of sediment, dissolved oxygen, water temperature, and fish mortality rate.

[0055] 1) Input layer: The input parameters include six experimental observation data: fish species, duration, sediment content, median sediment particle size, water temperature, and dissolved oxygen. The hyperbolic tangent function Tansig is used as the transfer function between the input layer and the hidden layer. The original experimental information is recorded and transmitted to the neural network through the six variables of each sample (i.e., fish species, duration, sediment content, median sediment particle size, water temperature, and dissolved oxygen).

[0056] 2) Hidden Layer: Through trial and error, the number of neurons in the hidden layer was determined to be p = 12, meaning there are a total of 12 neurons in the hidden layer; The trial calculation formula is as follows:

[0057]

[0058] In the formula: p is the number of neurons in the hidden layer; m is the number of parameters in the input layer, which is taken as m=6; n is the number of parameters in the output layer, which is one parameter for fish mortality, so n=1; A is a constant between [1, 10].

[0059] 3) Output layer: The output value of the output layer needs to be denormalized, and the output result is the fish mortality rate; the linear function Purelin is selected as the transfer function between the hidden layer and the output layer.

[0060] 4) Assessing the predictive ability of the backpropagation neural network model: The predictive ability of the backpropagation neural network model is evaluated using the root mean square error, the absolute value of the mean relative error, and the Nash efficiency coefficient. The calculation formulas are as follows:

[0061] Root mean square error:

[0062]

[0063] Mean relative absolute value:

[0064]

[0065] Nash efficiency coefficient:

[0066]

[0067] In the formula: y iThese are the output layer values ​​of the neural network and the measured values ​​of the experimental dataset, respectively, representing the simulated and measured values ​​of fish mortality rates. is the average of the expected values, i.e., the average of the measured fish mortality rates; i is the number of sample data groups; n is the total number of sample data.

[0068] Backpropagation neural network models exhibit strong randomness and uncertainty, resulting in slow learning convergence and an inability to guarantee convergence to the global minimum. Currently, there are two main methods to improve the prediction accuracy of backpropagation neural networks: one is to improve the calculation methods of weights and thresholds; the other is to improve the initialization methods of weights and thresholds. This embodiment adopts the second method, improving the accuracy of the neural network output by refining the initialization methods of weights and thresholds. Specifically, it uses an improved particle swarm optimization algorithm, which has fast convergence and powerful global search capabilities, to address the aforementioned shortcomings of backpropagation neural networks.

[0069] Step 4, train the backpropagation neural network model:

[0070] 1) Select array data from the indoor experimental dataset, perform normalization processing, and form a training sample dataset.

[0071] 2) Import the training sample dataset into the backpropagation neural network model and initialize the particle swarm.

[0072] 3) Utilize an improved particle swarm optimization algorithm to call the backpropagation neural network model, using the output error of the backpropagation neural network model as the fitness function; Fitness function F:

[0073]

[0074] In the formula: y ki , respectively, represent the output value of the backpropagation neural network and the expected value in the training sample dataset, i.e., the predicted and measured values ​​of fish mortality; k is the number of iterations in the improved particle swarm optimization algorithm.

[0075] The improved particle swarm optimization algorithm and backpropagation neural network are coupled by particle information and fitness value. By iteratively seeking the particle swarm with the optimal fitness, the parameters that minimize the error of the improved particle swarm optimization algorithm-backpropagation neural network model are obtained.

[0076] 4) Determine if the maximum number of iterations or the preset precision has been reached; otherwise, return to step 3.

[0077] 5) Assign the optimized weights and thresholds to the backpropagation neural network, and then train the backpropagation neural network using gradient descent.

[0078] Step 5, Model Testing: Select another array of data from the experimental dataset to form a test sample dataset. Input the test sample dataset into the backpropagation neural network model trained in Step 4 to test the model and observe whether it meets the requirements. If it does not meet the requirements, adjust the model parameters and return to Step 4 to continue training.

[0079] For the practical testing of the improved particle swarm optimization algorithm-backpropagation neural network model, additional sample data from the experimental data can be input into the model for prediction. The measured values ​​are then compared with the predicted values. If the difference between the two is within a certain range, such as ±5%, it is considered to have achieved a satisfactory level. This test can be conducted using multiple sets of samples. If all samples fall within the error range, or occasionally exceed the range, it can also be considered to have achieved a satisfactory level.

[0080] For the 439 measurement data points obtained from the experiment, 419 data points were used to form a training sample dataset to train and validate the improved particle swarm optimization algorithm-backpropagation neural network model. The training set consisted of 80% randomly selected data points, and the validation set comprised 20%. Twenty data points were used as test data. The six variable values ​​in the test sample dataset largely covered the range of experimental measurements and were highly representative. The predicted and measured mortality rates of some training sample data are shown below. Figure 5 As shown, by Figure 5 It can be seen that the measured values ​​and predicted values ​​are quite similar and have the same trend when the different types of fish and sand content are tested.

[0081] The data points for the mortality rate in the training sample are all concentrated near the auxiliary line of y = x, with a correlation coefficient R = 0.995, which meets the training requirements. (See...) Figure 6 Statistical analysis of the prediction error distribution shows that the prediction error of the training samples is generally within ±5%, with only one set of data having an error greater than 10%. Based on the above analysis, the improved particle swarm optimization algorithm-backpropagation neural network model is considered to have high accuracy in predicting the mortality rate of spotted naked carp and common carp in water bodies with high sediment content.

[0082] After the improved particle swarm optimization algorithm-backpropagation neural network model was trained, in order to eliminate the influence of the randomness of the backpropagation neural network model on the prediction results, the calculation was repeated 10 times on a test sample dataset of 20 groups of samples, and the average value was taken for analysis. The experimental measurement data and prediction results are as follows: Figure 7 As shown in Table 3, the test sample dataset was not used for neural network training and was equivalent to the ordinary dataset to be predicted. Table 3 shows that the predicted mortality rate was basically consistent with the measured value, with an average error of 1.81% and a maximum error of 5.48%.

[0083] Step 6, Prediction Application: Based on the prediction requirements, multiple sets of environmental variable data are generated in conjunction with the reservoir sediment discharge process to be predicted. These sets of data are then normalized and input into the trained backpropagation neural network model to output the corresponding predicted fish mortality rate.

[0084] Based on the sediment discharge process of a reservoir in the lower reaches of the Yellow River, the parameters of the high sediment-laden flow are: duration = 8h, sediment concentration in the downstream channel = 40kg / m³. 3 The sediment median particle size is 0.01 mm, dissolved oxygen is 3.5 mg / L, water temperature is 22℃, and the target fish species is carp. After normalizing the parameters to be predicted, they are input into an improved particle swarm optimization algorithm-backpropagation neural network model. The predicted carp mortality rate downstream of the reservoir due to this sediment discharge process is 11.7%.

[0085] Finally, it should be noted that the above is only used to illustrate the technical solution of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred arrangements, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solution of the present invention (such as the experimental water tank and the living environment of the simulated research object, neural networks, the application of various formulas, the order of steps, etc.) without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A method of predicting the lethal effects of a sediment-laden flow on fish, characterized in that, The steps of the method are as follows: Step 1, data collection: collect various data of the research object, including: fish species, sediment concentration, median particle size of sediment, water temperature, oxygen solubility, duration, to form an original data set; Step 2, model experiment: according to the original data set, an experimental model simulating the real environment is constructed, the living environment of the research object is simulated in the experimental model, and various experimental data are collected; the experimental data at least include: fish species and size, duration, sediment concentration and sand particle size distribution, water temperature, dissolved oxygen; the experimental data set of the mortality of the research object in the sediment-laden water body is constructed from the experimental data; Step 3, constructing a back propagation neural network model: a three-layer back propagation neural network containing an input layer, a hidden layer and an output layer is constructed, the training algorithm is Bayes regularization algorithm, the maximum number of iterations is 20000, and the learning rate is set to 0.01 according to experience; 1) input layer: the input parameters include fish species, duration, sediment concentration, median particle size of sediment, water temperature and dissolved oxygen, a total of 6 experimental observation data; the hyperbolic tangent function Tansig is selected as the transfer function between the input layer and the hidden layer; the original test information of each sample is recorded through 6 variables and is transmitted to the neural network; 2) hidden layer: the number of hidden layer neurons p=12 is obtained by trial algorithm, that is, there are 12 neurons in the hidden layer; the trial formula is: In the formula: p is the number of hidden layer neurons; m is the number of input layer parameters, m=6; n is the number of output layer parameters, the output layer is fish mortality, n=1; A is a constant between 1 and 10; 3) output layer: the output value of the output layer needs to be processed by inverse normalization, and the output result is the mortality of fish; the linear function Purelin is selected as the transfer function between the hidden layer and the output layer; 4) prediction ability judgment of back propagation neural network model: the prediction ability of back propagation neural network model is evaluated by using root mean square error, average relative error absolute value and Nash efficiency coefficient, and the calculation formulas are respectively: Root mean square error: Average relative error absolute value: Nash efficiency coefficient: wherein: y i are the neural network output layer values and the measured values of the experimental data set, i.e. the simulated and measured values of the fish mortality, respectively; is the average of the expected values, i.e. the average of the measured values of the fish mortality; i is the number of sample data sets; n is the total number of sample data. Step 4, training of back propagation neural network model: 1) according to the selected array data in the indoor experiment data set, normalization processing is carried out to form a training sample data set; 2) import the training sample data set into the back propagation neural network model, and initialize the particles therein; 3) call the back propagation neural network model by using the improved particle swarm algorithm, and take the output error of the back propagation neural network model as the fitness function; the fitness function F is: In the formula: y ki respectively, the back propagation neural network output value and the expected value in the training sample data set, that is, the predicted value and the measured value of fish mortality; k is the iteration number in the improved particle swarm algorithm. The improved particle swarm algorithm and the back propagation neural network are coupled through particle information and fitness value, and the particle swarm with the optimal fitness is obtained through continuous iteration, that is, the parameters that make the error of the back propagation neural network model minimum are obtained; 4) judge whether the maximum number of iterations or the preset precision is reached, otherwise return to step 3); 5) assign the optimized weights and thresholds to the back propagation neural network, and then train the back propagation neural network by using gradient descent method; Step 5, model testing: alternatively, a test sample data set is formed from the experimental data set, the test sample data set is input into the back propagation neural network model trained in step 4, the back propagation neural network is tested, and it is observed whether the requirements are met; if the requirements are not met, the parameters of the back propagation neural network are adjusted and the training is continued in step 4; Step 6, prediction application: according to the prediction requirement, a plurality of groups of environmental variable data are generated in combination with the reservoir desilting process to be predicted, the groups of data are normalized, input into the trained back propagation neural network model, and corresponding fish mortality prediction values are output.

Images