CN115796040A

CN115796040A - Facility greenhouse temperature prediction method based on small samples

Info

Publication number: CN115796040A
Application number: CN202211552061.XA
Authority: CN
Inventors: 胡瑾; 雷文晔; 刘行行; 卢有琦; 魏子朝; 杨永霞
Original assignee: Northwest A&F University
Current assignee: Northwest A&F University
Priority date: 2022-12-05
Filing date: 2022-12-05
Publication date: 2023-03-14

Abstract

A facility greenhouse temperature prediction method based on small samples collects environmental factors as features and constructs a data set, wherein the environmental factors comprise: indoor temperature, indoor air relative humidity, indoor soil temperature, indoor illumination intensity, outdoor air temperature, outdoor air relative humidity and outdoor soil temperature; constructing a universal greenhouse temperature prediction model based on the 1D CNN-GRU deep neural network, training to obtain network parameters corresponding to an optimal prediction result, and obtaining a universal optimal model; the method comprises the steps of adjusting optimal model network parameters in a small sample data set in a pre-training and fine-tuning mode based on transfer learning, and constructing greenhouse temperature prediction models under facility greenhouse conditions facing different places and/or different climates.

Description

Facility greenhouse temperature prediction method based on small samples

Technical Field

The invention belongs to the technical field of agricultural Internet of things and agricultural service, relates to a greenhouse technology, and particularly relates to a facility greenhouse temperature prediction method based on small samples.

Background

Temperature is one of the important parameters affecting the yield and quality of crops in a greenhouse, and over-high or under-low can cause damage to the growth and development of crops. Researches prove that the effective temperature control has important significance on the yield increase of facility agriculture. At present, most of production regulation and control of facility agriculture depends on relevant experience of a producer, regulation and control decision is carried out only by data acquisition and analysis, and the method can cause hysteresis in feedback control and influence crop growth. Therefore, the construction of the facility temperature time sequence prediction model by combining multiple environmental factors is an important premise for the accurate and efficient control of the facility greenhouse temperature.

The research finds that the temperature of the facility greenhouse not only has a continuous time sequence with characteristics of time variation, nonlinearity, periodicity and the like, but also has a complex coupling relation with various environmental factors inside and outside the facility greenhouse. Factors affecting the indoor and outdoor temperature include indoor and outdoor air temperature, indoor and outdoor air relative humidity, outdoor wind speed, outdoor wind direction, indoor and outdoor soil temperature, indoor and outdoor illumination intensity, indoor and outdoor CO ₂ The concentration, the indoor and outdoor soil moisture and the like are independent of each other, and the non-time-sequence characteristics such as the environment distribution characteristic, the ventilation characteristic, the humidification characteristic and the like are adopted. Therefore, in order to couple the characteristic information influencing the greenhouse temperature and fully mine the internal potential relation of the characteristics, the facility greenhouse temperature prediction model must comprehensively consider the influence of various environmental factors on the greenhouse temperature, which is the key for accurately predicting the temperature of the facility greenhouse.

In recent years, a large amount of research on facility greenhouse temperature prediction is carried out by domestic and foreign scholars by using various methods. Aiming at the problem that the existing greenhouse environment control system cannot accurately predict the temperature of the greenhouse in the next period, a method for establishing a temperature prediction model by adopting a time sequence analysis method is provided by the Zuoshu and the like (2010); liu Shipeng and the like (2017) select related meteorological factors influencing temperature change in the greenhouse, and a greenhouse temperature prediction model is established by a stepwise regression analysis method; li X, etc. (2020) provides a three-dimensional sunlight greenhouse temperature prediction model based on a BP neural network algorithm; tiandong et al (2020) use ARIMAI and SVR combined methods to improve the ability of approximate simulation of greenhouse temperatures of edible fungi under different environmental conditions. With the development of artificial intelligence technology, more and more scholars integrate the deep learning algorithm into the prediction of the greenhouse temperature of the facility. The special network structure of the convolutional neural network can extract high-dimensional features reflecting complex dynamic changes of temperature, and the extracted feature vectors are constructed into a time sequence form to be used as the input of the cyclic neural network to learn the dynamic change rule in the features. And a Recurrent Neural Network (RNN), a gated recurrent unit neural network (GRU), and a long-short term neural network (LSTM) are also applied to the field of temperature prediction by combining a convolutional neural network. For example, zhao Quanming et al (2020) propose a CNN-GRU-based mushroom house multi-point temperature and humidity prediction method, and deep effective information is mined, so that a GRU model can better learn the high-dimensional time sequence characteristics extracted by CNN; elmaz F et al (2021) propose a CNN-LSTM architecture for modeling of predicted indoor temperature that solves the temperature non-linearity and hysteresis problems and achieves good results. The method improves the prediction precision, also considers the operation efficiency, improves the performance of the whole model, and can provide a theoretical basis for accurate prediction of the temperature.

However, at present, the relevant research focuses on temperature modeling in the same place or same climate greenhouse, and the temperature modeling condition facing to various different greenhouse conditions is rarely considered, so that the following two defects always exist: firstly, the model has no universality under the condition of orienting to various different greenhouses. As most researches only aim at temperature modeling of the same place or the same climate greenhouse, one type of greenhouse model cannot be transferred to another type of greenhouse model, different models need to be trained according to different conditions of the greenhouse, and the calculation cost is increased. Secondly, in computer science, sufficient training data are usually needed to avoid the phenomenon of overfitting in deep learning, however, a large number of greenhouses which are newly built, expanded and reconstructed are available every year, and a high-precision prediction model is difficult to establish due to short running time and insufficient accumulated historical data, so that a rapid and effective modeling method suitable for the greenhouse needs to be explored urgently.

Disclosure of Invention

In order to overcome the defects of the prior art, the invention aims to provide a facility greenhouse temperature prediction method based on small samples, which aims to quickly and effectively construct various prediction models through a transfer learning method for a greenhouse temperature prediction model with a large amount of data under the condition of orienting to various different facility greenhouses, so as to continuously, efficiently and accurately predict the temperature.

In order to achieve the purpose, the invention adopts the technical scheme that:

a facility greenhouse temperature prediction method based on small samples comprises the following steps:

step 1, collecting environmental factors as characteristics to construct a data set, wherein the environmental factors comprise: indoor temperature, indoor air relative humidity, indoor soil temperature, indoor illumination intensity, outdoor air temperature, outdoor air relative humidity and outdoor soil temperature;

step 2, constructing a universal greenhouse temperature prediction model based on the 1D CNN-GRU deep neural network;

the 1D CNN-GRU deep neural network consists of an input layer, a convolutional layer, a circulating layer and an output layer in sequence; wherein, the input layer X _t+j,i Is a two-dimensional matrix; x _t,i～ X _t+j,i I-th feature data representing input samples from time t to time t + j; x _t,1 ～X _t,i The total number of the sample data at the time t is 1-i; extracting two-dimensional features contained in the data by the convolutional layer by adopting a one-dimensional convolution filter; two-layer GRU network is selected by the circulation layer, the extracted two-dimensional features are constructed into a time sequence and input into the first layer GRU network, and the first layer GRU network returns to each time stepThe second layer GRU network returns the final output of each input sequence; the output layer is a full connection layer and outputs a predicted value of a target moment;

step 3, training the universal greenhouse temperature prediction model to obtain network parameters corresponding to the optimal prediction result and obtain a universal optimal model;

and 4, adjusting network parameters in the small sample data set in a pre-training and fine-tuning mode based on transfer learning according to the optimal model, and constructing a greenhouse temperature prediction model facing different places and/or different climates under the facility greenhouse condition, wherein the type of the environmental factors in the small sample data set is consistent with the environmental factors in the step 1, but the data volume is far smaller than that of the data set constructed in the step 1.

In one embodiment, in step 1, data in the data set are collected continuously daily according to a set interval time T, and the continuous time is not less than one month.

In one embodiment, in step 2, the model takes the indoor temperature, the indoor air relative humidity, the indoor soil temperature, the indoor illumination intensity, the outdoor air temperature, the outdoor air relative humidity and the outdoor soil temperature as input, the time step is 6, that is, the size of the input matrix is 6 × 7, the temperature after T time is taken as output, and the output step d is set to be used for predicting the temperature change within T × d time.

In one embodiment, step 3, all data are first normalized and scaled to between [0,1 ]; and then dividing the data subjected to normalization processing into a training set and a test set, training the model by using the training set data, and verifying the generalization ability and the precision of the model by using the test set data.

In one embodiment, in step 3, the size of the convolutional layer filter is set to 3, the step size is 1, the number of convolutional kernels is 64, the number of output nodes of the cyclic layer is 128, and the number of output nodes of the output layer is 1; the convolution layer inputs sample data s at time t _t After performing convolution operation with one-dimensional convolution kernel, applying bias b _cnn Entering a loop layer by activating a function, wherein sample data x of the loop layer _t By resettingAnd the gate r, the updating gate z and the time t provide a temporary unit state c for dot product operation, and output to the full connection layer.

In one embodiment, in step 2, the input layer, the convolutional layer, the cyclic layer, and the output layer all use a Linear rectification function (reduced Linear Unit, relu) as an activation function, a Mean Square Error (MSE) as a loss function, and an Adam optimization algorithm is used to update the weights of the neural network, so that the performance of the model is optimal; after the model completes the specified number of training rounds, judging whether the output result meets the model precision, if not, readjusting the parameter value, and continuing training until the model precision is reached; and if so, outputting the result.

In one embodiment, in step 3, the temperature of the test set is used as the actual value, the output value of the model is used as the predicted value, the temperature and the model are reversely normalized to recover the original dimensional grade, and then the coefficient R is determined ² And the mean square error MSE is used as an evaluation index, and the model precision and the generalization ability are evaluated.

In an embodiment, in the step 4, the network of the general optimal model is used as a pre-training network, the source domain network is sequentially frozen into the convolutional layer, the convolutional layer and the top circulation layer, and the convolutional layer and the two circulation layers, and when network parameters of the rest part are retrained, target domain data is selected for verification, so that a greenhouse temperature prediction model under facility greenhouse conditions facing different places and/or different climates is constructed, and accurate prediction of the facility greenhouse temperature is realized.

Compared with the prior art, the method comprehensively considers the influence of various internal and external environment factors of the greenhouse on the temperature in the greenhouse, takes accurate and efficient temperature prediction as a target under various different types of greenhouse conditions, provides a facility greenhouse temperature prediction model based on deep learning, and verifies the prediction effect of the model at different time (namely different weather conditions) and different places. The invention has the following beneficial effects:

(1) And constructing a universal facility greenhouse temperature model based on the recurrent neural network. The characteristics of time sequence, nonlinearity, high volatility and strong coupling of temperature are fully considered, and effective information contained in environmental data is mined through a convolutional neural network; and reducing the dimension of the extracted high-dimensional features, converting the extracted deep-level abstract features into global feature vectors as the input of a cycle layer, and performing effective dynamic time series data modeling. The operation efficiency and the prediction performance of the model are considered, and a theoretical basis is provided for accurately and efficiently predicting the temperature under various different types of greenhouse conditions.

(2) And constructing a temperature prediction model under the small sample data based on the transfer learning. Under the same type condition and different types of conditions, pre-training is carried out based on a type of greenhouse temperature prediction model with a large amount of data, network parameters are adjusted in a small data set in a fine adjustment mode, and multiple prediction models are quickly and effectively constructed. The technology reduces the calculation cost as much as possible while preventing the model from being over-fitted, and lays a foundation for the accurate regulation and control of the greenhouse environment of the facility.

Drawings

FIG. 1 is a schematic view of the structure of a monitoring platform inside and outside a greenhouse.

Fig. 2 is a schematic diagram of the working principle of 1D CNN.

FIG. 3 is a schematic diagram of a gated cycle cell configuration.

Fig. 4 is a schematic diagram of a 1D CNN-GRU network structure.

FIG. 5 is a flow chart of temperature prediction model building.

Fig. 6 is a schematic diagram of a fitting effect of the winter prediction model according to the embodiment of the present invention.

FIG. 7 is a schematic diagram of a model migration process.

Detailed Description

The embodiments of the present invention will be described in detail below with reference to the drawings and examples.

The invention provides a facility greenhouse temperature prediction method based on small samples, which comprises the steps of firstly, acquiring a large number of data sets as source domains according to current greenhouse internal and external environment information, and mining and extracting two-dimensional features contained in source domain data through a convolutional neural network; secondly, constructing the extracted characteristic vector as a time sequence, inputting the time sequence into a recurrent neural network to extract time characteristics of the time sequence, predicting future greenhouse temperature, and establishing a universal optimal model for greenhouse temperature prediction; finally, network parameters are adjusted in a small data set in a pre-training and fine-tuning mode based on transfer learning, a greenhouse temperature prediction model facing different places and/or different climates under the greenhouse facility conditions is constructed, and a foundation is laid for realizing efficient and accurate prediction and environment accurate regulation and control of the greenhouse temperatures of various facilities.

The prediction method of the present invention can be specifically described as follows:

step 1, collecting environmental factors as characteristics and constructing a data set, wherein the environmental factors comprise: indoor temperature, indoor air relative humidity, indoor soil temperature, indoor illumination intensity, outdoor air temperature, outdoor air relative humidity and outdoor soil temperature.

In one embodiment of the invention, the tests are carried out in Jingyang vegetable test demonstration station No. 5 and No. 3 of northwest agriculture and forestry science and technology university from 8 and 13 days in 2019 to 11 days in 2019 (summer), 12 and 2 days in 2019 to 31 days in 2019 (winter) and from 11 days in 12 and 11 days in 2020 to 12 and 30 days in 2020, the sites are 34 degrees and 53 degrees in northern latitude, 108 degrees and 84 degrees in east longitude and 404m, the test belongs to continental monsoon climate in a warm zone, and the annual average temperature is 13 ℃. The source greenhouse is of a single-span single-slope type, a framework is of a light steel structure, a back slope is covered by a PC plate and silk cotton quilt, and the east side, the west side and the north side are wall parts and are all built by clay. The direction of the utility model is that the utility model is facing south, the length is 50m, the width is 10m, the height of the back wall is 3m, the ridge is 1.6m, and the utility model is provided with a skylight, a fan, a roller shutter machine, a sunshade net, a wet curtain, a sprayer and other equipment. The fan adjusts the temperature of the air in the greenhouse by accelerating the air flow in the greenhouse, so that the temperature values at all positions of the greenhouse are uniformly changed.

During the test period, from 12/month and 2/2019 to 12/month and 31/2019, six weather conditions of sunny weather, sunny weather changing to cloudy weather, haze weather and haze weather changing to cloudy weather are included. For further analysis of the predicted effect of temperature under different weather conditions, the clarity index K is used _T The weather is classified. K _T The method is used for representing the attenuation degree of solar radiation passing through the atmosphere and is closely related to the current weather condition. As shown in table 1, the winter weather is classified into three categories according to the weather phenomenon corresponding to the clarity index. Wherein, K _T More than or equal to 0.5 day and more than 0.5K _T More than or equal to 0.2 day in 5 days,0.2＞K _T More than or equal to 0.1 day, K _T And the total day is 0.1.

TABLE 1 weather phenomena corresponding to clarity index

Crops are not cultivated in a greenhouse in summer, and sterilization and greenhouse closing treatment are carried out during the test period; the greenhouse indoor cultivation crop in winter is tomato, the tomato is cultivated in a drip irrigation mode, the cover is uncovered and the heat preservation quilt is uncovered at the ratio of 8-00 to 8 _T Not less than 0.5 and 0.5 > K _T When the temperature is more than or equal to 0.2, the greenhouse is ventilated at 10-00; when 0.2 > K _T And when the window opening time is more than or equal to 0.1, properly delaying the greenhouse window opening time according to actual conditions.

Referring to fig. 1, in order to realize the collection of environmental factors, a greenhouse internal and external monitoring platform is constructed, which mainly comprises an information monitoring module, an information transmission module and an agricultural production internet of things monitoring platform. And continuously collecting data in the data set every day according to a set interval time T, wherein the continuous time is not less than one month.

In the embodiment, continuous collection is set at 0-00-24 days per day, the sampling interval time T is 20 minutes, the collected environmental factors comprise indoor illumination intensity (QY-150A illumination intensity sensor, measuring range: 0-200 KLux, precision: +/-1 Lux), indoor temperature and outdoor air temperature (PT 1000 temperature sensor, measuring range: 200 ℃ to +200 ℃, precision: +/-0.1 ℃), indoor and outdoor air relative humidity (DHT 11 humidity sensor, measuring range: 20-90%, precision: +/-5%), indoor and outdoor soil temperature (PT 1000 temperature sensor, measuring range: 200 ℃ to +200 ℃, precision: +/-0.1 ℃), and the test is carried out by placing the sensor in the middle of a greenhouse, sample data is 1.2m higher than the ground, collecting for 30 days, and 2160 groups are obtained in total every season. The information transmission module is used for transmitting the sensing information to the gateway node through the ZigBee network for data processing, then transmitting the sensing information to the mobile base station through the GPRS network, realizing remote connection and data interaction by utilizing the Web server, and uploading the acquired data to the agricultural production Internet of things monitoring platform.

And 2, constructing a universal greenhouse temperature prediction model.

On the basis of the obtained test data, a universal optimal model for greenhouse temperature prediction needs to be established for quickly and effectively constructing various prediction models for greenhouse facilities facing different places and/or different climates, and a foundation is laid for realizing efficient and accurate prediction and environment accurate regulation and control of the greenhouse temperatures of various facilities.

The invention uses an improved one-dimensional convolutional network (1D CNN) and a gated cyclic unit (GRU) algorithm to construct a temperature prediction model. The 1D CNN is an effective learning frame in deep learning, and the features of the human brain are extracted from a local image through convolution operation simulation, so that a good effect is achieved in the aspect of computer vision. Meanwhile, in the sequence problem, the 1D CNN can capture the correlation between the input data through the filter of the convolutional layer, and combine it into a shorter characteristic sequence to be input into the next layer network for prediction. For example, when the convolution sliding window is 5, the 1D CNN sequentially extracts local one-dimensional sequence segments from the sequence by window sliding and makes dot products with weights, as shown in fig. 2.

The gated cyclic unit GRU is an optimization network based on a cyclic neural network, optimizes the internal structure of a neuron by introducing a 'gate' concept, selectively transmits information, and solves the problem that the cyclic neural network is difficult to learn long-distance dependency in a sequence. In the hidden layer, each neuron is connected with each other, and the input is related to not only the input of the input layer but also the output of the hidden layer at the previous moment. As shown in fig. 3 a, x is the total input, h is the hidden layer output, and o is the total output. As shown in FIG. 3 b, the GRU "gate" includes a reset gate r that fuses the inputs x and an update gate z _t History data h from previous time _t-1 Discarding historical data irrelevant to time, and keeping the dependency relationship among the short sequence data; update gate z memorizes h by storing last moment _t-1 The information is selectively memorized and its value is continuously updated to capture the dependency between long sequence data.

Based on the network structure, the invention builds a 1D CNN-GRU deep neural network, the concrete structure of which is shown in figure 4, and the whole structure is connected in a one-by-one stacking mode and consists of an input layer, a convolution layer, a circulation layer and an output layer in sequence. Wherein, the input layer X _t+j,i Is a two-dimensional matrix, i is the number of features, and j is the time step; x _t,i～ X _t+j,i I-th feature data representing input samples from time t to time t + j; x _t,1 ～X _t,i The sample data at the time t is shown to have 1 to i characteristics. The convolution layer adopts a one-dimensional convolution filter to extract two-dimensional characteristics contained in the data, and no pooling layer is added because the dimension of the sample data is less. And the circulation layer selects two layers of GRU networks, constructs the extracted two-dimensional features into time sequences and inputs the time sequences into the first layer of GRU network, the first layer of GRU network returns a complete sequence output at each time step, and the second layer of GRU network returns the final output of each input sequence. The output layer is a full connection layer and outputs a predicted value of the target moment.

In the embodiment of the present invention, the winter model takes the indoor temperature, the indoor air relative humidity, the indoor soil temperature, the indoor illumination intensity, the outdoor air temperature, the outdoor air relative humidity and the outdoor soil temperature as input, and the time step is 6, that is, the input matrix size is 6 × 7. The temperature after T time is taken as output, the output step is d, and the temperature change within T × d time is predicted, and a specific flowchart thereof is shown in fig. 5. In this embodiment, T is 20 minutes, and the model is used to predict the temperature change within three hours by setting the output step size 9.

And 3, training the universal greenhouse temperature prediction model to obtain network parameters corresponding to the optimal prediction result and obtain the universal optimal model.

The training process of the invention is as follows:

(1) Because the numerical value ranges of different sample data are greatly different and the dimensions are different, in order to accelerate the fitting speed of the neural network and improve the prediction accuracy of the model, all the sample data are scaled to be between [0,1] by adopting normalization processing, and the used calculation formula is

In the formula x _min Is the minimum value, x, in each dimension of data _max Is the maximum value, x, in each dimension of data _nor Is normalized data.

(2) Dividing the normalized data into a training set and a test set, training the model by using the training set data, and verifying the generalization ability and precision of the model by using the test set data.

In this embodiment, 1512 groups of data (accounting for 70% of the total data of the sample) are selected as the training set, and the remaining 648 groups of data (accounting for 30% of the total data of the sample) are selected as the test set from the normalized 2160 groups of data.

(3) In order to keep the network stable and fully mine local effective information among data, the size of the convolutional layer filter is set to be 3, the step length is 1, the number of convolutional kernels is 64, the number of output nodes of a circulating layer is 128, and the number of output nodes of the output layer is 1.

(4) The convolution layer converts the input sample data s at time t _t After performing a convolution operation with a one-dimensional convolution kernel, applying an offset b _cnn The cyclic layer is entered by the activation function as shown below.

y _t ＝f(W _cnn *s _t +b _cnn )

In the loop layer, its sample data x _t And providing a temporary unit state c at the moment of the reset gate r, the updating gate z and the moment t for carrying out dot product operation, and outputting to a full connection layer.

z＝sigmoid(W _z h _t-1 +U _z x _t )

r＝sigmoid(W _r h _t-1 +U _r x _t )

Wherein f is an activation function, W _cnn For the weight matrix, s, corresponding to the input layer to the convolutional layer _t Input sample data at time t, b _cnn For the input layer to convolutional layer corresponding bias matrix, y _t Output data for convolutional layer, at y _t Extracting a two-dimensional matrix on the basis to obtain x _t 。W _z 、W _r 、W _c 、U _z 、U _r 、U _c To update the gates, reset the gates, and temporarily output the weight matrix.

(5) Each layer of network (i.e. input layer, convolutional layer, cyclic layer and output layer) adopts a Linear rectification function (Relu) as an activation function, a Mean Square Error (MSE) as a loss function, and an Adam optimization algorithm is used for updating the neural network weights in order to update the neural network weights W, U and V, so that the model performance is optimal. After the model completes the specified number of training rounds, judging whether the output result meets the model precision, if not, readjusting the parameter value, and continuing training until the model precision is reached; and if the result is satisfied, outputting the result.

(6) The temperature of the test set is used as an actual value, the output value of the model is used as a predicted value, the temperature and the model are subjected to inverse normalization, the original dimensional grade is recovered, and then a coefficient R is determined ² And mean square error MSE is used as an evaluation index, and model precision and generalization capability are evaluated.

In order to verify the generalization ability and the prediction performance of the winter sunlight greenhouse temperature prediction model constructed in the step, 648 groups of sample data of a temperature test set after 20 minutes are used for verifying the generalization ability and the prediction performance. The verification result shows that the decision coefficient of the model training set is 0.998, the mean square error is 0.101 ℃, and the average error is 0.199 ℃; the coefficient of the test set was 0.998, the mean square error was 0.085 ℃ and the mean error was 0.185 ℃. As shown in fig. 6, the fitting coefficient of the fitted straight line is 1.001 and almost approaches to 1, and the intercept is-0.009 and almost approaches to 0, which shows that the predicted value and the actual value of the model are almost on both sides of the straight line y = x, and the predicted value and the actual value are strongly correlated, so that the model achieves the ideal prediction accuracy.

On the basis of this, it isFurther verifying the temperature prediction accuracy change after winter model prediction for 20 minutes, 40 minutes, 1 hour 20 minutes, 1 hour 40 minutes, 2 hours 20 minutes, 2 hours 40 minutes and 3 hours, to determine the coefficient R ² The three terms of mean square error MSE and mean absolute error MAE are evaluation indexes, and specific results are shown in table 4.

TABLE 4 winter model multistep prediction duration results

As can be seen from table 4, the accuracy of the winter model gradually decreases as the step size increases. The model can accurately predict the temperature of the first 9 steps, namely the temperature after 20 minutes to 3 hours, the model determining coefficient change range of the first 9 steps is 0.964-0.998, the mean square error change range is 0.085-2.276 ℃, the average absolute error change range is 0.185-0.927 ℃, the requirement of agricultural production regulation and control precision is met, and farmers can select proper steps to predict the change trend of the future greenhouse temperature according to self requirements.

To verify the prediction performance of the algorithm of the present invention, the model accuracy after 20 minutes is taken as an example, and the decision coefficient R is adopted ² The mean square error MSE and the mean absolute error MAE are used as evaluation indexes, a multiple linear regression Method (MLR), a time sequence analysis method (AR), a BP neural network algorithm, a support vector machine regression (SVR) and a one-dimensional convolution gated cyclic unit network (1D CNN-GRU) are selected for modeling comparison after parameter optimization, and the prediction results of 5 models are shown in Table 5.

TABLE 5 comparison of different algorithms

As can be seen from table 5, the one-dimensional convolution gated cyclic unit algorithm is significantly superior to the statistical method represented by the multiple linear regression method and the time sequence analysis method, and the shallow neural network method represented by the BP neural network algorithm and the support vector machine regression. This is because it is difficult to establish an accurate mathematical model based on a linear mathematical equation in a statistical method; the shallow neural network method has limited capability of processing multidimensional time series data and is difficult to accurately predict large-scale sample data. In the one-dimensional convolution gating circulation unit, the one-dimensional convolution can obtain the effective characteristics which are mutually connected in the multidimensional data, and meanwhile, the gating circulation unit network can better learn the time sequence characteristics, so that the algorithm can greatly improve the temperature prediction precision.

In conclusion, models built by the one-dimensional convolution gating circulation unit network are all remarkably superior to the 4 groups of comparison models in the aspect of temperature time sequence data prediction, and therefore the superiority of the algorithm is verified.

At the same time, to verify that the winter model is at different K _T Next applicability, the effect was predicted by three types of weather check models in 2019, 12-23 to 2019, 12-31. Wherein K is given at 26 days (sunny) at 12 months, 27 days (cloudy at sunny days) at 12 months, 28 days (sunny) at 12 months, 30 days (sunny) at 12 months, and 31 days (sunny) at 12 months _T Not less than 0.5;12 months and 25 days (cloudy) is 0.5 & gtK _T Not less than 0.2;12 months and 23 days (haze), 12 months and 24 days (haze) and 12 months and 29 days (haze is changed into cloudy), and the K is more than 0.2 _T Not less than 0.1. According to the constructed model, the MLR model and the BP model are taken as examples to be compared with the 1D CNN-GRU model, and the three models are different in K _T The following predicted results are shown in table 6.

TABLE 6 winter various K _T The next multiple model verification results

As shown in Table 6, in three classes K _T Lower 1D CNN-GRU moduleThe prediction effect of the model is the best, and the prediction effect of the MLR model is the worst. When K is _T More than or equal to 0.5, the indoor environment is hardly influenced by cloud layers due to less clouds in sunny days in Shaanxi areas in winter, and the fluctuation of the illumination intensity value is small, so that the prediction effect of the 1D CNN-GRU model is more than K than 0.5 _T Not less than 0.2 and 0.2 > K _T More preferably not less than 0.1. When 0.5 > K _T When the prediction error fluctuation of the 3 models is smaller than or equal to 0.2, the fitting degree of the predicted value and the actual value is lower, because the cloudy greenhouse is influenced by the movement of a cloudy layer, the moving speed is high, the thickness is continuously changed, the indoor illumination intensity is greatly fluctuated, and the prediction error of the models occurs to a certain degree. When 0.2 > K _T When the haze weather is more than or equal to 0.1, the particles in the haze weather can influence the indoor illumination intensity, so the 1D CNN-GRU model has poor prediction effect on sunny days. Comprehensively considered, a winter sunlight greenhouse temperature prediction model is K _T The prediction effect is best when the weather is not less than 0.5 (clear), and the prediction effect is best when the weather is in three types K _T The lower model prediction effects are different, but the determination coefficients are all above 0.984, the mean square error is below 0.115 ℃, the average absolute error is below 0.218 ℃, and the K values are all _T The lower model shows higher accuracy, which shows that the method has applicability and can be applied to actual production, and when indoor solar radiation is weak in winter, the temperature can be predicted in advance through the model, and the heating equipment is used for supplementing and heating the greenhouse, so that a reliable basis is provided for high-quality production of crops.

And 4, adjusting network parameters of the small sample data set in a pre-training and fine-tuning mode based on transfer learning according to the obtained optimal model, and constructing a greenhouse temperature prediction model under facility greenhouse conditions facing different places and/or different climates. Wherein the type of environmental factor in the small sample dataset is consistent with the environmental factor of step 1, but the data volume is often much smaller than that of the dataset constructed by step 1.

Transfer learning is a machine learning method that reuses a pre-trained model in another task. Compared with the traditional method, the method can effectively avoid the problem of insufficient training samples, and improves the generalization capability of the model, thereby improving the prediction precision of the model with insufficient training set. Hu Q and the like (2016) and Huang X and the like (2020) apply a transfer learning method to the processing of wind speed prediction and weather forecast respectively, and by pre-learning a large number of samples and carrying out knowledge transfer on a small number of samples, the problem of overfitting of the small number of samples is solved, and the prediction precision is improved. The basic assumption of transfer learning differs from the traditional machine learning approach in that training and test data are extracted from the same feature space and satisfy similar data distributions. Its goal is to apply learned knowledge or patterns on a certain domain or task to a different but related domain or problem, which is an effective strategy for small sample prediction, including instance-based migration, feature-based migration, and model-sharing parameter-based migration, etc. The model sharing parameter-based transfer learning method can embed a network model pre-trained in a large-scale data set into other task models to serve as a feature extractor, so that the depth feature information of small sample data is effectively extracted for prediction.

The transfer learning mainly comprises two concepts of a domain and a task, wherein the domain is the type and the composition of data and represents basic attributes; tasks are attributes that are completed using data, and in the present invention, are predicted tasks. Typically a domain consists of two parts, which can be represented as D = { Z, P (Z) }, i.e. feature space Z and the edge distribution of feature space P (Z). Where the feature space Z can be represented as Z = { Z | Z _i E.g. Z, i =1,2. Given a domain D = { Z, P (Z) }, a task may be represented as T = { Y, f (·) }, i.e., a label control Y and an objective prediction function f (·).

Existing transfer learning mostly only considers that one source domain D exists _s And a target domain D _t The case (2) is called single-source domain transfer learning. Source domain D _s ＝{(z _s,1 ,y _s,1 ),(z _s,2 ,y _s,2 ),...,(z _s,n ,y _s,n )},z _(s,j) ∈Z _s An observation sample representing the source domain, y _(s,j) ∈Y _s Representing a source domain observation sample z _(s,j) A corresponding label. Target domain D _t ＝{(z _t,1 ,y _t,1 ),(z _t,2 ,y _t,2 ),...,(z _t,n ,y _t,n )},z _(t,j) ∈Z _t Representing target domain observationsSample, y _(t,j) ∈Y _t Representing target domain samples z _(t,j) And outputting correspondingly. Based on the above symbolic definition, the purpose of the transfer learning is: in a given source domain D _s And source domain learning task T _s Target domain D _t And target domain task T _t And satisfy D _t ≠D _s And T _t ≠T _s By using the source domain D _s And T _s To promote the target domain D _t Learning effect of the medium objective prediction function f (·).

On the basis, in order to solve the problems of low greenhouse temperature prediction accuracy and overfitting under the condition of insufficient data, the invention constructs various greenhouse temperature prediction models of different types based on transfer learning, and adjusts network parameters in small data sets by obtaining the universal optimal model for greenhouse temperature prediction established by the invention and adopting a fine adjustment mode. Specifically, a network of an optimal model is used as a pre-training network, a source domain network is sequentially frozen with a convolutional layer, the frozen convolutional layer and a top circulation layer, and the frozen convolutional layer and two circulation layers, and when the network parameters of the rest part are retrained, target domain data are used for verification to construct a greenhouse temperature prediction model facing different places and/or different climates under facility greenhouse conditions, so that the facility greenhouse temperature can be accurately predicted. The migration learning and the feature extraction are complementary, for the frozen model base used for the feature extraction, the top layers are unfrozen, and the unfrozen layers and the newly added part are jointly trained, and the specific migration process is shown in fig. 7.

Because the temperature variation trends of the same type of greenhouse are different due to geographical positions, crop growth and other factors, for the same type of greenhouse, if the trained greenhouse temperature prediction model at a certain place can be applied to the greenhouses at other places, the time cost can be greatly saved. Therefore, data of the east 5 solar greenhouse from 2019, 12 months and 2 days to 2019, 12 months and 31 days are selected as an initial source field of an example; data of east 3 solar greenhouse 2020, 12/11/2020, and 2020, 12/30/2020 were selected as target domains.

The network of the east 5 winter temperature prediction model established by the invention is used as a pre-training network, when the source domain network freezes a convolution layer and other network parameters are retrained, data (total 14 days) of the east 3 solar greenhouse from 12 month and 11 days of 2020 and 12 month and 24 days of 2020 and data (total 9 days) of 11 months and 11 days of 2020 and 12 months and 19 days of 2020 are selected as training set training models in the target domain; when the source domain network freezes the convolutional layer and the top circulation layer, and other network parameters are retrained, the target domain still selects data (total 14 days) of the east 3 solar greenhouse from 2020 12 month by 11 days to 2020 12 month by 24 days and data (total 9 days) from 2020 12 month by 11 days to 2020 12 month by 19 days as training set training models; when the source domain network freezes the convolution layer and the two circulation layers, and other network parameters are retrained, the target domain selects data (total 6 days) of the east 3 solar greenhouse from 12/11/2020/12/16/2020 and data (total 3 days) from 12/11/2020/12/13/2020/as training set training models. The data (total 6 days) of 25 days in 12 months and 30 days in 2020 to 12 months and 30 days in 2020 are used as the test set in the models to verify the generalization capability and the accuracy of the models. The model specific results are shown in table 7.

TABLE 7 prediction duration of 20 minutes

As can be seen from table 7, the models built with the number of frozen layers from 1 layer to 3 layers all perform better in each evaluation index, and the model built with the smaller number of frozen layers has higher accuracy. Meanwhile, when the number of the freezing layers is 1, the prediction effect of the model with the training set of 14 days is better than that of the model with the testing set of 9 days, the MSE of the training set is 0.265 ℃, the MSE of the testing set is 0.172 ℃, and the difference of the results between the two sets is smaller, which indicates that the generalization capability is better than that of the model with the testing set of 9 days. In addition, when the number of freezing layers is 2 to 3, the difference between the training set days of 14 days and the testing set days of 9 days and between the training set days of 6 days and the testing set days of 3 days is not obvious, which indicates that when the number of freezing layers is large, the prediction effect of twice small sample data can be achieved by constructing a prediction model based on the small sample data.

Also, to verify the applicability of the method at multiple steps, the predicted time period was 40 minutes compared to 1 hour for verification as shown in tables 8 and 9.

TABLE 8 prediction of duration of 40 minutes

TABLE 9 duration of prediction 1 hour

The results in tables 8 and 9 show that the three freezing effects also show similar characteristics in modeling of different prediction durations, when the network freezes one layer, the model has the highest precision and the best fitting effect, and the precision of the model established in 14 days is higher than that in 9 days, which shows that the more sample data, the better the model shows in the test set; when the number of the frozen layers of the network is increased, the model is reduced layer by layer with the increased number of the layers, and the sample data amount has no obvious advantage. Meanwhile, the three tables show that the accuracy of the established model is reduced along with the increase of the time step, and the requirement of agricultural actual production is still met.

Similarly, because the greenhouse temperature has different variation trends in different seasons, two temperature models are respectively constructed for winter and summer. However, for the same type of greenhouse, if the trained greenhouse temperature prediction model in one season can be applied to the models in other seasons, the time cost can be greatly saved. Therefore, when the winter model is selected as the source domain model, part of the data of the east 5 solar greenhouse from 2019, 8, 13 and 2019, 9, 11 (summer) is used as the target domain.

The network of the winter temperature prediction model established by the invention is used as a pre-training network, when the convolution layer is frozen in the source domain network and other network parameters are retrained again, data (total 14 days) from 13 days in 8 months in 2019 to 26 days in 8 months in 2019 and data (total 9 days) from 18 days in 8 months in 2019 to 26 days in 8 months in 2019 are selected as training set training models in the target domain; when the source domain network freezes the convolution layer and the top circulation layer, and other network parameters are retrained, the target domain still selects data from 13 days in 8 months in 2019 to 26 days in 8 months in 2019 (total 14 days) and data from 18 days in 8 months in 2019 to 26 days in 8 months in 2019 (total 9 days) as training set training models; when the source domain network freezes the convolution layer and the two circulation layers, and other network parameters are retrained, the target domain selects data from 21 days 8 and 26

days

8 and 8 months 2019 in summer (total 6 days) and data from 24 days 8 and 26 months 2019 in summer (total 3 days) as training set training models. The data (total 6 days) from 27 th in 2019 to 1 th in 2019 from 9 th in month 27 are used as test sets to verify the generalization ability and accuracy of the model, and the three predicted durations are used as comparison verification, as shown in tables 10, 11 and 12.

TABLE 10 prediction duration of 20 minutes

TABLE 11 predicted duration of 40 minutes

TABLE 12 duration of prediction 1 hour

From the table, it can be seen that the model built with 1 to 3 frozen layers shows better performance in each evaluation index, and the model built with a smaller number of frozen layers has higher accuracy. When the prediction time is 20 minutes, the test set fitting degree is 0.964-0.997, the MSE is 0.499-6.117 ℃, and the MAE is 0.351-1.761 ℃; when the prediction time is 40 minutes, the test set fitting degree is 0.945 to 0.997, the MSE is 0.460 to 10.021 ℃, and the MAE is 0.423 to 2.224 ℃; when the prediction time is 1 hour, the test set has the fitting degree of 0.931-0.997, the MSE of 0.464-11.688 ℃ and the MAE of 0.485-2.539 ℃. The model precision is not much different from the prediction model precision when the second section of training data is sufficient, even when partial prediction precision can exceed the traditional method when the training data is sufficient. Therefore, the summer temperature prediction model constructed by the pre-training and fine-tuning method is high in precision, and continuous and accurate prediction of temperatures in different step lengths can be realized under the condition of small samples.

In conclusion, the conventional method mostly adopts a statistical method and a shallow neural network method, and only can realize accurate prediction of short-term greenhouse temperature, but as the prediction time is prolonged, the model prediction accuracy is rapidly reduced due to the difficulty in mining the time sequence characteristics contained in the data, so that the actual production requirement is difficult to meet. The facility greenhouse temperature model construction method based on the convolutional neural network and the gated cyclic network can be used for long-term prediction. The characteristics of time sequence, nonlinearity, high volatility and strong coupling of temperature are fully considered, and effective information contained in environmental data is mined through a convolutional neural network; and reducing the dimension of the extracted high-dimensional features, converting the extracted deep-level abstract features into global feature vectors as the input of a cycle layer, and performing effective dynamic time series data modeling. The method gives consideration to the operation efficiency and the prediction performance of the model, and provides a theoretical basis for accurately and efficiently predicting the temperature under various different types of greenhouse conditions. Meanwhile, the transfer learning method adopted by the invention adjusts network parameters in a small data set in a fine adjustment mode, and quickly and effectively constructs various prediction models. The technology reduces the calculation cost as much as possible while preventing the model from being over-fitted, and lays a foundation for the accurate regulation and control of the greenhouse environment of the facility.

Claims

1. A facility greenhouse temperature prediction method based on small samples is characterized by comprising the following steps:

the 1D CNN-GRU deep neural network comprises an input layer, a convolutional layer and a loop in sequenceA layer and an output layer; wherein, the input layer X _t+j,i Is a two-dimensional matrix; x _t,i～ X _t+j,i The ith characteristic data representing input samples from the t th time to the t + j th time; x _t,1 ～X _t,i The sample data at the time t has 1-i characteristics; extracting two-dimensional features contained in the data by the convolutional layer by adopting a one-dimensional convolution filter; the circulation layer selects two layers of GRU networks, the extracted two-dimensional features are constructed into time sequences and input into the first layer of GRU network, the first layer of GRU network returns a complete sequence output at each time step, and the second layer of GRU network returns the final output of each input sequence; the output layer is a full connection layer and outputs a predicted value of a target moment;

and 4, adjusting network parameters in the small sample data set in a pre-training and fine-tuning mode based on transfer learning according to the optimal model, and constructing a greenhouse temperature prediction model facing different places and/or different climates under the facility greenhouse condition, wherein the type of the environmental factor in the small sample data set is consistent with the environmental factor in the step 1, but the data volume is far smaller than that of the data set constructed in the step 1.

2. The small-sample-based greenhouse temperature prediction method for facilities according to claim 1, wherein in the step 1, data in the data set are continuously collected every day according to a set interval time T, and the continuous time is not less than one month.

3. The facility greenhouse temperature prediction method based on small samples as claimed in claim 2, wherein in the step 2, the model takes the indoor temperature, the indoor air relative humidity, the indoor soil temperature, the indoor illumination intensity, the outdoor air temperature, the outdoor air relative humidity and the outdoor soil temperature as input, the time step is 6, that is, the input matrix size is 6 x 7, the temperature after T time is taken as output, and the output step d is set for predicting the temperature change within T x d time.

4. The small-sample-based facility greenhouse temperature prediction method according to claim 1, wherein in the step 3, all data are normalized and scaled to [0,1 ]; and then dividing the data subjected to normalization processing into a training set and a test set, training the model by using the training set data, and verifying the generalization ability and the precision of the model by using the test set data.

5. The facility greenhouse temperature prediction method based on small samples according to claim 1, wherein in the step 3, the size of the convolution layer filter is set to 3, the step size is 1, the number of convolution kernels is 64, the number of output nodes of the circulation layer is 128, and the number of output nodes of the output layer is 1; the convolution layer inputs sample data s at time t _t After performing convolution operation with one-dimensional convolution kernel, applying bias b _cnn Entering a loop layer by activating a function, wherein sample data x of the loop layer _t And providing a temporary unit state c through the reset gate r, the update gate z and the time t for dot product operation, and outputting to the full connection layer.

6. The facility greenhouse temperature prediction method based on the small samples as claimed in claim 5, wherein in the step 2, the input layer, the convolution layer, the circulation layer and the output layer all adopt a Linear rectification function (Relu) as an activation function, a Mean Square Error (MSE) is a loss function, and simultaneously, an Adam optimization algorithm is used for updating the weight of the neural network, so that the model performance is optimal; after the model finishes the specified number of training rounds, judging whether the output result meets the model precision, if not, readjusting the parameter value, and continuing training until the model precision is reached; and if so, outputting the result.

7. The facility greenhouse temperature prediction method based on small samples as claimed in claim 6, wherein in the step 3, the temperature of the test set is used as an actual value, the output value of the model is used as a predicted value, the temperature and the model are reversely normalized, the original dimension grade is recovered, and then the solution is used for determiningConstant coefficient R ² And the mean square error MSE is used as an evaluation index, and the model precision and the generalization ability are evaluated.

8. The facility greenhouse temperature prediction method based on small samples as claimed in claim 1, wherein in the step 4, the network of the general optimal model is used as a pre-training network, the source domain network is sequentially frozen into the convolutional layer, the frozen convolutional layer and the top cyclic layer, and the frozen convolutional layer and the two cyclic layers, and when network parameters of the rest part are retrained, target domain data selection is used for verification, a greenhouse temperature prediction model facing different places and/or different climates under facility greenhouse conditions is constructed, and facility greenhouse temperature accurate prediction is achieved.