CN110414115B - Wavelet neural network tomato yield prediction method based on genetic algorithm - Google Patents

Abstract

The invention discloses a genetic algorithm-based wavelet neural network tomato yield prediction method, which belongs to the field of agricultural information technology. This method first selects a certain amount of parameters as input variables, and obtains a suitable data population by classifying, processing and screening these parameters and applying genetic algorithm to optimize them. Then determine the number of nodes in the input layer, hidden layer and output layer of the network, apply the wavelet basis function to replace the activation function of the hidden layer nodes of the BP neural network, and introduce translation factors and scale factors to continuously adjust the weights, after many training and iterations , so that the error of the model prediction is continuously reduced, the output result is constantly close to the measured value, the error is controlled within a reasonable range, the prediction accuracy and function convergence performance of the model are improved, and a more reliable theoretical support is provided for predicting the crop yield in the greenhouse.

Description

A Method of Tomato Yield Prediction Based on Wavelet Neural Network Based on Genetic Algorithm

technical field

The invention belongs to the technical field of agricultural information, in particular to a method for predicting tomato yield based on a genetic algorithm based on a wavelet neural network.

Background technique

Tomatoes are one of the main vegetables cultivated in protected areas in northern China. Tomatoes in solar greenhouses in northern areas mostly adopt high-efficiency crop rotation cultivation patterns. The environmental factors of the greenhouse, the nutrient content of the soil, and the physiological characteristics of tomatoes directly determine the yield of tomatoes. Predicting the tomato yield in the greenhouse can determine the appropriate planting plan, reasonably arrange the amount of fertilization and irrigation, take measures such as insecticide and weeding in time, and regularly take measures such as ventilation, temperature increase, and supplementary light to create healthy growth for tomatoes. Good environment, and lay the foundation for high-quality and high-yield greenhouse tomatoes. Previous studies on tomato yield prediction are very extensive, but most of them focus on applying crop mechanism models or crop growth and development models to predict tomato yield in greenhouses, involving relatively few parameters, and calculation methods have certain limitations. The combination of genetic algorithm and wavelet neural network is used to predict the yield of greenhouse tomatoes, which can be based on historical data over the years, and through continuous correction of various parameter variables in the model to improve and optimize the model, improve the prediction accuracy of the model, and make it more suitable for The actual environment of the solar greenhouse in the north has a broad application prospect.

Contents of the invention

The invention provides a genetic algorithm-based wavelet neural network tomato yield prediction method, which can predict the tomato yield in the northern energy-saving solar greenhouse and can determine a suitable planting plan.

A wavelet neural network tomato yield forecasting model based on genetic algorithm, it is characterized in that, the input variable of this model comprises following feature parameter: ambient temperature; Environmental humidity; Irrigation amount; Nitrogen fertilizer input amount; Phosphate fertilizer input amount; ₂ concentration; light intensity;

The model is built by the following steps:

Randomly generate an initial population X _m×n :

n=s ₁ ×s ₂ +s ₂ ×s ₃ +s ₂ +s ₃ (1)

Among them: m is the initial population size, n is the individual length, and the individual length not only represents the number of gene values of each individual, but also represents the initial weight value of a neural network; s ₁ is the number of nodes in the input layer; s ₂ is the hidden layer The number of nodes; s ₃ is the number of nodes in the output layer;

The genetic algorithm calculates the output error value E _i and the fitness value f _i of each individual in the initial population, and evaluates it according to the value of the individual fitness value, and selects the individuals whose fitness value is within the preset range A in the initial population to enter The subpopulation continues with optimization training:

In the subpopulation, the probability that the i-th individual performs crossover or mutation operation is p _i , according to the adaptive function of crossover rate p _c and mutation rate p _m , it is judged whether the individual needs crossover or genetic operation:

分别为种群中最大适应度值和平均适应度值，

为种群的收敛程度；In the formula: k _c and k _m are real numbers less than 1, f _c is the individual fitness value to be crossed, f _m is the individual fitness value to be mutated, f _max ,

are the maximum fitness value and the average fitness value in the population, respectively,

is the degree of convergence of the population;

Eliminate the data whose deviation exceeds ±10% of the annual average value in the input variables, apply genetic algorithm to optimize the data, and the data dynamic range ≤ ±5% is considered to have a good processing effect, and the filtered and optimized data is used as a wavelet neural network input data and perform subsequent calculations;

The characteristic parameters after the above processing are used as the input of the wavelet neural network model, and the greenhouse tomato output is used as the output, which includes the following steps:

The mother wavelet function selected for this model is:

g(x)=cos(1.75x)exp(-x ² /2) (7)

Scale and translation transform the mother wavelet function to construct the wavelet basis function:

Among them, a _j and b _j are the scale factor and translation factor of the jth hidden layer node respectively;

y _k is the output of the model:

Among them, x _i (i=1, 2, ..., I) is the input of the i-th node of the input layer, y _k (j = 1, 2, ..., k) is the input of the j-th node of the output layer output;

Define the error function E as:

Among them, y _k is the actual output of the kth node of the output layer, and t _k is the target output of the kth node of the output layer;

Adjustments to weights, scale factors, and translation factors include:

Among them, learning rate: η (η>0), momentum factor: μ (0<μ<1).

More preferably, the error value is represented by a percentage, then the fitness value is in the range of 0-1, and A=f _i >0.667.

The invention provides a genetic algorithm-based wavelet neural network tomato yield prediction method, through classifying, processing and screening data and applying the genetic algorithm to optimize training to obtain suitable data populations. Then determine the number of nodes in the input layer, hidden layer and output layer of the network, apply the wavelet basis function to replace the activation function of the hidden layer nodes of the BP neural network, and introduce translation factors and scale factors to continuously adjust the weights. After multiple trainings and iterations , so that the error of the model prediction is continuously reduced, and the output result is continuously close to the measured value, and the error can be controlled within a reasonable range.

Description of drawings

Fig. 1 is the wavelet neural network structural diagram of a kind of wavelet neural network tomato yield prediction method based on genetic algorithm provided by the present invention;

Figure 2 is the prediction effect diagram of the GA-WNN model;

Figure 3 is the error percentage curve of the GA-WNN model;

Figure 4 is a WNN model prediction effect diagram;

Fig. 5 is the WNN model error percentage curve;

Fig. 6 is the prediction result of BP neural network model;

Fig. 7 is BP neural network model error percentage curve;

Fig. 8 is a flow chart of a method for predicting tomato yield based on genetic algorithm based on wavelet neural network provided by the present invention.

Detailed ways

A specific embodiment of the present invention will be described in detail below in conjunction with the accompanying drawings, but it should be understood that the protection scope of the present invention is not limited by the specific embodiment.

The establishment process of the tomato yield prediction model: firstly, classify, process and screen the data, and then apply the genetic algorithm to optimize the training to obtain a suitable data population. Then determine the number of nodes in the input layer, hidden layer and output layer of the network, apply the wavelet basis function to replace the activation function of the hidden layer nodes of the BP neural network, and introduce translation factors and scale factors to continuously adjust the weights. After multiple trainings and iterations , so that the error of the model prediction is continuously reduced, the output result is continuously close to the measured value, and the error is controlled within a reasonable range. The specific process is shown in Figure 8.

The specific execution steps of the GA-WNN algorithm:

This model selects the following parameters as input variables: ①Environmental temperature; ②Environmental humidity; ③Irrigation amount; ④Nitrogen fertilizer input; ⑤Phosphate fertilizer input; output variable. First, the genetic algorithm is used as the pre-means of the WNN algorithm to optimize the original data to obtain data with a small dynamic range (data dynamic range ≤ ±5%), and use the filtered and optimized data as the subsequent WNN algorithm initial value of . Then, the wavelet function is used to improve the traditional BP neural network, and the characteristics of multi-resolution analysis of the wavelet function are combined with the adaptive ability of the neural network to construct the wavelet neural network model, and it is constantly revised and perfected. Dadi has improved the prediction accuracy and function convergence performance of the model, providing more reliable theoretical support for predicting crop yields in greenhouses.

The specific operation process is as follows:

(1) The genetic algorithm is used to screen and optimize the selected 8 parameters. First, randomly generate an initial population X _m×n to optimize the original data, where m is the initial population size (because 8 variables are selected, so m=8), and the individual length n represents the gene value of each individual The number also represents the initial weight value of a neural network (the number of hidden layer nodes will change with the network training, and the number of weight values will also change accordingly. Initially, 8 input layer nodes, 5 hidden layer nodes, and 1 output layer are set. Layer nodes, the number of initial weights is 46), and the gene values in the individual correspond to the initial weights of the neural network one by one. In this study, the gene value is encoded by the real number encoding method, which can avoid the decoding process and improve the training efficiency.

n=s ₁ ×s ₂ +s ₂ ×s ₃ +s ₂ +s ₃ (1)

In the formula: n is the individual length; s ₁ is the number of input layer nodes; s ₂ is the number of hidden layer nodes; s ₃ is the number of output layer nodes.

The genetic algorithm calculates the output error value E _i and fitness value f _i of each individual in the initial population, and evaluates it according to the value of the individual fitness value, and selects the larger fitness value in the initial population (the error value is expressed as a percentage, Then the fitness value is in the range of 0 to 1, and the value of f _i >0.667 is regarded as a value with better fitness) individuals enter the subpopulation to continue optimization training.

In the subpopulation, the probability that the ith individual is selected for crossover or mutation operation is p _i , and according to the adaptive function of crossover rate p _c and mutation rate p _m to judge whether the individual needs crossover or genetic operation, p The values of _c and p _m will change the degree of adaptation according to the value of the individual fitness value _fi , so as to keep the population always diverse.

分别为种群中最大适应度值和平均适应度值，

为种群的收敛程度。依据以上遗传算法，对原始数据进行筛选与优化处理，剔除偏差较大(某种变量超过其年平均值±10％均视为偏差较大的值，并以此上下限约束其值，将其规定为边界值进行后续计算)，应用遗传算法对数据优化，能够使数据的动态范围减小，数据动态范围≤±5％视为处理效果较好，将经过筛选与优化处理后的数据作为小波神经网络的输入数据并进行后续运算。In the formula: k _c and k _m are real numbers less than 1, f _c is the individual fitness value to be crossed, f _m is the individual fitness value to be mutated, f _max ,

are the maximum fitness value and the average fitness value in the population, respectively,

is the convergence degree of the population. According to the genetic algorithm above, the original data is screened and optimized, and the large deviation is eliminated (a variable that exceeds its annual average value by ±10% is regarded as a value with a large deviation, and its value is constrained by this upper and lower limit, and its It is specified as the boundary value for subsequent calculation), and the genetic algorithm is used to optimize the data, which can reduce the dynamic range of the data, and the data dynamic range ≤ ± 5% is regarded as a good processing effect, and the data after screening and optimization are used as wavelet Input data to the neural network and perform subsequent operations.

(2) The 8 characteristic parameters processed and optimized in the previous step are used as the input of the wavelet neural network (WNN) model, and the greenhouse tomato output is used as the output. The structure of the network is very important for the prediction accuracy and stability of the model. When building a model, the structure of the model must first be considered so that it can meet the test requirements and ensure the stability and reliability of the model. Then, the error function is used to analyze the existing problems, and the model is continuously optimized to improve the prediction accuracy.

The mother wavelet function selected for this model is:

g(x)=cos(1.75x)exp(-x ² /2) (7)

Scale and translation transform the mother wavelet function to construct the wavelet basis function:

Among them, a _j and b _j are the scale factor and translation factor of the jth hidden layer node respectively. y _k is the output of the model:

x _i (i=1, 2, . . . , I)—the input of the i-th node in the input layer, y _k (j=1, 2, . . . , k)—the output of the j-th node in the output layer.

By continuously adjusting the weights, scale factors, and translation factors in this model, the stability and prediction accuracy of the model are improved, making this model more suitable for the actual on-site environment, and the model is continuously improved and optimized. Define the error function E as:

Among them, y _k is the actual output of the kth node in the output layer, and t _k is the target output of the kth node in the output layer. Adjustments to weights, scale factors, and translation factors include:

Among them, learning rate: η (η>0), momentum factor: μ (0<μ<1). By continuously adjusting the parameters in the formula, the prediction accuracy of the model is improved, and MATLAB software is used for simulation to verify the prediction effect of the model.

Then, by continuously adjusting parameters such as GA parameters, BP parameters, weights, scale factors, and translation factors in this model, the stability and prediction accuracy of the model are improved, making this model more suitable for the actual on-site environment, and it is constantly improved. and optimize the model.

In this experiment, the data from 2010 to 2015 were used to train the model, and the data from 2016 to 2018 was used to test the actual prediction effect of the GA-WNN model. Use 6 years of data to train this network, constantly adjust the model parameters such as weight, scale factor, translation factor, learning rate, momentum factor and iteration number, compare the test results many times and calculate the error, and constantly improve it Through improvement and optimization, a model with better prediction effect has been obtained, and its error is within a reasonable range (the error percentage is within 5%). The difference between the predicted value and the measured value is very small, and the accurate prediction of greenhouse tomato production has basically been realized. GA - The prediction effect of the WNN model is shown in the figure below. The results show that the predicted value of this model has a better tracking effect on the measured value, and the stability of the model is higher, which can effectively predict the greenhouse tomato yield. The comparison chart of experimental data and test results is as follows:

In order to verify the prediction effect of the model, the GA-WNN model is compared with the WNN model and the BP neural network model. The comparison of the prediction effect is shown in the table below. According to the analysis, there are certain errors between the predicted value and the actual value of the two models. The average relative error of the GA-WNN model is 0.66%, the average relative error of the WNN model is 1.02%, and the average relative error of the BP neural network model is 2.42%. , the prediction effect of GA-WNN model is ideal. The convergence speed of the GA-WNN model is better than that of the WNN and BP neural network models, and the prediction effect has reached the optimum after 208 steps. In summary, the GA-WNN model constructed by combining genetic algorithm, wavelet analysis and BP neural network has a faster convergence speed and higher prediction accuracy, and realizes accurate prediction of tomato yield, which can provide a reasonable arrangement for the greenhouse. Provide a basis for crop types, irrigation and fertilization and other management plans.

In this study, the GA-WNN model was used to predict the tomato yield in the northern solar greenhouse. The factors affecting the greenhouse tomato yield were comprehensively considered. The parameters were processed and screened by means of genetic algorithm, and then the feature vector of the data was extracted as the model input. A greenhouse tomato yield prediction model was constructed, and the scale factor and translation factor in the wavelet neural network were used to adjust the weights to improve the prediction accuracy of the model. The simulation results show that the average relative error of the GA-WNN model is 0.66%, the average relative error of the WNN model is 1.02%, and the average relative error of the BP neural network model is 2.42%. The prediction effect of the GA-WNN model is better than that of the WNN and BP neural network models , after 208 steps, the prediction effect has reached the optimum, which proves that the GA-WNN model has good practical application value. Compared with the traditional WNN and BP neural network models, the average relative error of the GA-WNN model is smaller, the accuracy is higher, and the stability is stronger. It is more in line with the forecast of greenhouse tomato production and can provide a basis for the decision-making of tomato planting in the greenhouse. , and provided some theoretical support for the irrigation, fertilization, supplementary light and ventilation measures taken during tomato planting, and played a role in promoting the improvement and optimization of the greenhouse system.

The above disclosures are only a few specific embodiments of the present invention, however, the embodiments of the present invention are not limited thereto, and any changes conceivable by those skilled in the art shall fall within the protection scope of the present invention.

Claims (2)

1. A wavelet neural network tomato yield prediction model based on genetic algorithm, it is characterized in that, the input variable of this model comprises following feature parameter: ambient temperature; Environmental humidity; Irrigation amount; Nitrogen fertilizer input amount; Phosphate fertilizer input amount; Potassium fertilizer input amount ; _CO2 concentration; light intensity; The model is built by the following steps: Randomly generate an initial population X _m×n : n=s ₁ ×s ₂ +s ₂ ×s ₃ +s ₂ +s ₃ (1) Among them: m is the initial population size, n is the individual length, and the individual length not only represents the number of gene values of each individual, but also represents the initial weight value of a neural network; s ₁ is the number of nodes in the input layer; s ₂ is the hidden layer The number of nodes; s ₃ is the number of nodes in the output layer; The genetic algorithm calculates the output error value E _i and the fitness value f _i of each individual in the initial population, and evaluates it according to the value of the individual fitness value, and selects the individuals whose fitness value is within the preset range A in the initial population to enter The subpopulation continues with optimization training:

In the subpopulation, the probability that the i-th individual performs crossover or mutation operation is p _i , according to the adaptive function of crossover rate p _c and mutation rate p _m , it is judged whether the individual needs crossover or genetic operation:

In the formula: k _c and k _m are real numbers less than 1, f _i is the individual fitness value; f _c is the individual fitness value to be crossed, f _m is the individual fitness value to be mutated, f _max ,

are the maximum fitness value and the average fitness value in the population, respectively,

is the degree of convergence of the population; Eliminate the data whose deviation exceeds ±10% of the annual average value in the input variables, apply the genetic algorithm to optimize the data, and the absolute value of the dynamic range of the data is within 5%, which is regarded as a good processing effect, and the data after screening and optimization are used as Input data of the wavelet neural network and perform subsequent operations; The characteristic parameters after the above processing are used as the input of the wavelet neural network model, and the greenhouse tomato output is used as the output, which includes the following steps: The mother wavelet function selected for this model is: g(x)=cos(1.75x)exp(-x ² /2) (7) x is the independent variable of the mother wavelet function, which represents the input variable of the output forecasting model; Scale and translation transform the mother wavelet function to construct the wavelet basis function:

Among them, a _j and b _j are the scale factor and translation factor of the jth hidden layer node respectively; y _k is the output of the model:

Among them, x _i (i=1, 2, ..., I) is the input of the i-th node of the input layer, y _k (j = 1, 2, ..., k) is the input of the j-th node of the output layer output; According to the formula (9), the output prediction value can be obtained and compared with the actual production value, so as to calculate the prediction error of the network, and the error function E is defined as:

Among them, y _k is the actual output of the kth node of the output layer, and t _k is the target output of the kth node of the output layer; According to the gradient descent method, adjust the weight W _ji from the input layer to the hidden layer and the weight W _kj from the hidden layer to the output layer, scale factor a _j (n+1), translation factor b _j (n+1), the formula as follows:

Among them, Wkj(n+1) and Wji(n+1) in formulas (11)-(14) represent the weights at the current moment, Wkj(n) and Wji(n) represent the weights at the previous moment, learning rate: η (η>0), momentum factor: μ (0<μ<1). 2. a kind of wavelet neural network tomato yield prediction method based on genetic algorithm as claimed in claim 1, is characterized in that, error value represents with percentage, and then fitness value is in the scope of 0～1, A=f _i >0.667 .

Images