CN106651012A - Crop transpiration prediction method based on improved extreme learning machine - Google Patents

Abstract

The invention relates to a crop transpiration prediction method based on an improved extreme learning machine. First, soil environment data and meteorological data of a farmland are collected and normalized to obtain a training set; then, the training set is adopted to train an extreme learning machine network and improve an extreme learning machine; last, normalized data collected again is input into the improved extreme learning machine, and the improved extreme learning machine outputs crop transpiration obtained through prediction. Extreme learning machine improvement mainly comprises the steps that 1, a function based on waveform superposition is adopted to serve as an activation function for a hidden layer of the extreme learning machine; 2, a particle swarm optimization algorithm is adopted to optimize an input weight value and a threshold value between a network input layer and the hidden layer of the extreme learning machine. Through the prediction method, the prediction precision of crop transpiration is improved, prediction time loss is reduced, and meanwhile generalization performance and prediction stability of the traditional extreme learning machine network are improved.

Description

A kind of crop transpirstion amount Forecasting Methodology based on improvement extreme learning machine

Technical field

The invention belongs to agriculture Internet of Things intelligent irrigation research field, is related to a kind of based on the crop for improving extreme learning machine Transpiration rate Forecasting Methodology.

Background technology

Agriculture Internet of Things is highly integrated and integrated use of the generation information technology in agriculture field, and China's agricultural is believed Breathization development has important leading action, changes conventional agriculture, promote agricultural to it is intelligent, become more meticulous direction Transformation.Production estimation environment real time information is gathered using substantial amounts of sensor node, monitoring system is constituted by network technology System, helps peasant to pinpoint the problems in time, and accurately determines the position of generation problem.Make to depend on originally the life for isolating machinery Product pattern turns to the intelligent production model centered on information and software, so as to reach the purpose of agricultural high-effiency production.

In intelligent irrigation field exactly calculate and predict the water demand of crop, understand Study on Crop Water Requirement Rules, be formulation science, Rational irrigation program, determines lrrigation Area amount, implements the fine basis irrigated；It is to reach water saving, highly efficient and productive mesh , realize the effective means and basic guarantee of Water Resources Irrigation sustainable development；Be formulate river basin planning, regional planning for water resources development, Water resource arranges the basic foundation in the fields such as planning, design, the management of engineering using planning and filling.Calculate exactly and predict and make Thing water requirement improves efficiency of water application for the water consumption for reducing Crop growing stage, and Developing Water-saving Agriculture has particularly significant Meaning.

Accurately substantial amounts of achievement in research is had in prediction water demand of crop field in research at present, to being reached for irrigation pipe Reason layer and policymaker provide intuitively Visual decision-making foundation, instruct irrigated area to accomplish that timely and appropriate discovery is irrigated, and improve Irrigation Project Design water The utilization rate and utilization ratio of resource.Patent of invention " a kind of Water Consumption of Winter Wheat amount Forecasting Methodology based on weather forecast information " (application number：201410117085.1).Based on by weather forecast information, consider crop own growth development condition, The impact of the environmental factor such as meteorological condition and soil regime is estimating the forecast model of winter wheat crop reference evapotranspiration, the party Method needs the information content of collection more, and for specific crop, appropriate is difficult to promote." field intelligently fills patent of invention Irrigate On-line Control management method " (application number：201410655632.1), it is proposed that a kind of field intelligent irrigation On-line Control management Method, by the soil moisture content of soil-water environment monitoring device the real time measure crop root zone, adopts in real time to data collecting system The data of collection such as are analyzed, conclude, processing at the secondary operation, calculate the real-time Evapotranspiration of crop, and the method is steamed to crop The amount of rising precision of prediction is limited, additionally, adopted special equipment monitoring soil moisture content, price is costly.

The content of the invention

It is an object of the invention to provide it is a kind of based on the crop transpirstion amount Forecasting Methodology for improving extreme learning machine, so as to improve The precision of prediction of crop transpirstion amount, improves the Generalization Capability and prediction stability of conventional limit learning machine network.

To reach above-mentioned purpose, the technical solution used in the present invention is：

A kind of crop transpirstion amount Forecasting Methodology based on improvement extreme learning machine, crop transpirstion amount (ET₀) refer to the United Nations Food and agricultural organization in March nineteen ninety in the Methods of Reference Crop Evapotranspiration computational methods international symposium that Rome, ITA is held, to ginseng Examine Methods of Reference Crop Evapotranspiration and give latest definition i.e.：Reference crop evapo-transpiration is a kind of imaginary reference crop canopy Evapotranspiration speed, and the plant height is assumed for 2m, fixed crop blade face resistance is 70s/m, and reflectivity is 0.23, very Similar to surface it is open, highly consistent, grow it is vigorous, be completely covered ground and the evapotranspiration speed on the green meadow of non-water shortage Rate, can pass through first to calculate the ET of each growing stage with meteorologic factor₀, being then multiplied by crop coefficient can be each in the hope of the crop The actual water demand of crop in stage；

The principle of extreme learning machine (ELM) primal algorithm is：Huang Guangbin et al. proposed a kind of new study in 2006 Algorithm is called extreme learning machine, and it is applied to single hidden layer feedforward neural network (SLFNs), and it is following two bold to be based primarily upon Conclusion：(1) N arbitrary sample { (x are given_i,t_i)}(x_i∈Rⁿ,t_i∈R^m) and an arbitrary intervals on infinitely can be micro- activation letter Number g (x), R → R, in RⁿRespectively input weights W is generated at random according to arbitrary continuation probability-distribution function with R spaces_iAnd threshold Value b_i, the hidden layer output matrix H of SLFN is necessarily reversible, and | | H β-T | |=0；(2) the positive number ε, N of any one very little are given Arbitrary sample { (x_i,t_i)}(x_i∈Rⁿ,t_i∈R^m) and an arbitrary intervals on infinitely can be micro- activation primitive g (x), R → R, For in RⁿRespectively input weights W is generated at random according to arbitrary continuation probability-distribution function with R spaces_iWith threshold value b_i, certainly existMake formulaSet up；

It is a kind of based on the crop transpirstion amount Forecasting Methodology for improving extreme learning machine, the crop transpirstion amount that the present invention is provided Forecasting Methodology is that, based on the crop transpirstion amount Forecasting Methodology for improving extreme learning machine, concrete prediction steps are as follows：

1) the soil environment data and meteorological data in farmland are gathered；

2) to step 1) in all data of collection be normalized and obtain training set, normalized purpose be overcome due to The dimension of input variable is inconsistent, if directly inputting network, will become to make a big difference by accumulator, ultimately results in Whole network is difficult the shortcoming for converging to optimal value；

3) extreme learning machine network is trained using training set and improves extreme learning machine, the improvement extreme learning machine is referred to Using activation primitive of the function as extreme learning machine hidden layer for being based on addition of waveforms, while excellent using particle swarm optimization algorithm Change the input weights and threshold value between extreme learning machine network input layer and hidden layer；

4) the soil environment data and meteorological data in farmland are gathered again, and all data to gathering are normalized, By the data input after normalization to improving in extreme learning machine, the Crop transpirstion that prediction is obtained is exported by improvement extreme learning machine Amount；

The improvement step of the extreme learning machine is as follows：

(1) extreme learning machine network structure |input paramete is initialized, activation primitive g (x) of hidden layer, the limit is set The input layer of learning machine network is soil environment data and meteorological data, and output layer is crop transpirstion amount, the extreme learning machine The |input paramete of network includes the input number of plies, hidden layer nodeWith the output number of plies, the input number of plies is by soil environment data Determine with meteorological data, the hidden layer nodeFor stochastic inputs, the output number of plies is 1；Extreme learning machine network implies The model of the activation primitive of layer plays very crucial effect to the performance of whole network, and original extreme learning machine network has three Plant activation primitive：Sigmoid functions, sine functions and signal number functions, these three function pair crop transpirstion amounts it is pre- The limited accuracy of survey, the present invention selects to improve original single ripple based on the new activation primitive of addition of waveforms using a kind of Shape activation primitive, two kinds of addition of waveforms of inverse hyperbolic SIN function and wavelet function that waveform is adopted, because wavelet function then has Low-and high-frequency signal fitting ability is emphasized, inverse hyperbolic SIN function accelerates to a certain extent the convergence rate of model, this Planting two kinds of function superpositions improves the network structure of original very limited learning machine so that the network of hidden layer has higher dynamic Energy disposal ability, activation primitive g (x) is used to calculate the output weights between hidden layer and output layer, the activation primitive g X the equation of () is as follows：

In formula, w₀For frequency, w₀=1, j are imaginary number, and x is the data after normalization；

(2) particle swarm parameter is initialized, with input weights and threshold between extreme learning machine network input layer and hidden layer It is worth for particle, all of input weights and sets of threshold values are into population, the limit after the |input paramete initialization of extreme learning machine network Input weights and threshold number between learning machine network input layer and hidden layer determine, according to extreme learning machine network input layer Input weights and threshold number random initializtion particle vector dimension and scope between hidden layer, the initialisation range of particle For [- 0.5,0.5], maximum iteration time k of the particle swarm parameter including population_max, Population Size popsize, particle rapidity Undated parameter c₁And c₂, the span of the speed of each particle is [- 0.5,0.5], the span of position for [- 0.5, 0.5]；

(3) extreme learning machine network and the adaptation in particle swarm optimization algorithm are trained as input data using training set Degree function calculates the fitness of particle, and in particle swarm optimization algorithm, fitness function is used to judge grain during Evolution of Population The root-mean-square error (RMSE) of predicted value and desired value is defined as particle fitness letter by the quality of sub- position, the present invention Number, wherein RMSE is less, shows that precision of prediction is higher, and the concrete formula of fitness function RMSE is：

In formula, ET_0-PM56I () is the calculating by FAO-56PM (Peng Man models) equations to the i-th particle crop transpirstion amount Value, unit is mmday^-1, ET_0-predictedI () is training predicted value of the extreme learning machine to the i-th particle crop transpirstion amount, single Position is mmday^-1, N is the group number of input sample；

(4) according to the fitness size of particle in population, the personal best particle P of particle is obtained_idWith global optimum position Put P_gd, personal best particle P_idWith global optimum position P_gdReplacement criteria be：

In formula, P_iFor the positional information of the i-th particle, RMSE (P_i) be the i-th particle fitness value, RMSE (P_gd) for it is global most The fitness value of excellent particle, RMSE (P_id) it is individual optimal particle fitness value；

(5) according to particle swarm optimization algorithm dynamic tracking personal best particle P_idWith global optimum position P_gdTo carry out not The disconnected speed for updating all particles and position, and particle fitness, more new individual optimum position are recalculated according to fitness function Put P_idWith global optimum position P_gd；The speed for updating all particles and position are specially：Particle swarm optimization algorithm is feasible The random particle of a group is initialized in solution space, each particle is likely to become a potential optimal solution, and each particle has There are three index features：Position, speed and fitness, wherein fitness are tried to achieve by the fitness function for defining, for representing The fine or not degree of particle, it is assumed that one population X=(X is constituted by n particles in the search space of D dimensions₁,X₂,...,X_n), X_i= (x_i1,x_i2,...,x_iD)^TRepresent the position of the i-th particle, V_i=(v_i1,v_i2,...,v_iD)^TRepresent the speed of the i-th particle, P_id= (p_i1,p_i2,...,p_iD)^TRepresent personal best particle, P_g=(p_g1,p_g2,...,p_gD)^TRepresent global optimum position, each iteration During, particle updates oneself speed itself and position by following two formula：

In formula, ω is inertia weight, c₁、c₂It is two nonnegative constants, is defined as acceleration factor, r₁、r₂It is two random Number, span is [0,1], and k is current iterations, v_idFor the speed of the i-th particle d dimensions, d=1,2 ..., D, x_id For the position of the i-th particle d dimensions, personal best particle P_id=(p_i1,p_i2,...,p_iD)^T, global optimum position P_gd=(p_g1, p_g2,...,p_gD)^T, v_id ^k+1For the speed of kth the i-th particle of+1 iteration d dimensions, v_id ^kFor kth time the i-th particle of iteration d dimensions Speed, P_id ^kFor kth time iteration personal best particle, x_id ^k+1For the position of kth the i-th particle of+1 iteration d dimensions, P_gd ^kFor kth Secondary iteration global optimum position, x_id ^kFor the position of kth time the i-th particle of iteration d dimensions；

(6) judge whether particle swarm optimization algorithm reaches maximum iteration time k_max, if it is preserve the grain of current iteration Subgroup, the population of current iteration is between the extreme learning machine network input layer of particle swarm optimization algorithm optimization and hidden layer Input weights and threshold value, otherwise step (5) continue iteration；

(7) hidden layer output matrix H is calculated according to activation primitive g (x) of extreme learning machine network hidden layer, specially： N group sample { (X are given in extreme learning machine network_s,t_s), s=1 ... N, hidden layer node isThen extreme learning machine net Network structure is as follows：

In formula, X_s=[X_s1,X_s2,...,X_sn]^T∈Rⁿ, t_s=[t_s1,t_s2,...,t_sm]^T∈R^m, X_sAnd t_sRepresent respectively defeated Enter variable and corresponding output variable, W_v=[W_v1,W_v2,...,W_vn]^TIt is between connection v hidden layer nodes and n input layer Input weight vector, b_vIt is the threshold vector of v hidden layers, β_v=[β_v1,β_v2,...,β_vm]^TIt is connection v hidden layer nodes With the weight vector between m output layer, W_v·X_sInner product is sought in expression；

Above-mentioned equation transform is obtained：

H β=T；

In formula, H is hidden layer output matrix, and β is output weight matrix, and T to expect output matrix, serve as reasons by the element in T The calculated Crop transpirstion value of FAO-56PM (Peng Man models) equation, wherein：

Because the input weights and threshold value between the input layer and hidden layer of particle swarm optimization algorithm optimization are determination value, then W_vAnd b_vValue determination, can calculate hidden layer output matrix H；

(8) calculating limit learning machine output weight matrix β, by input hidden layer nodeActivation primitive g (x), population Input weights and threshold value, hidden layer output matrix between the extreme learning machine network input layer and hidden layer of optimized algorithm optimization H and output weight matrix β obtain improving extreme learning machine, preserve and improve extreme learning machine, the calculating limit learning machine output The formula of weight matrix β is as follows：

In formula,The Moore-Penrose generalized inverses of matrix H are asked in expression.

As preferred technical scheme：

It is as above a kind of based on the crop transpirstion amount Forecasting Methodology for improving extreme learning machine, the soil environment data It is specially with meteorological data：The per day humidity of soil mean daily temperature, soil, air mean daily temperature, the per day humidity of air, Per day total solar radiation, the wind speed of 2m eminences and atmospheric pressure, the unit of temperature for DEG C, the unit of humidity is %, the day The unit of average total solar radiation is MJm^-2day^-1, the unit of the wind speed is ms^-1, the unit of the pressure is KPa.

It is as above a kind of based on the crop transpirstion amount Forecasting Methodology for improving extreme learning machine, step 2) and step 4) in Normalization refer to will collection all data normalizations to [- 1,1], normalize formula as follows：

In formula, x is the data after normalization, and x* is to be currently needed for normalized data, x_minTo need normalized data In minimum of a value, x_maxMaximum in need normalized data.

It is as above a kind of based on the crop transpirstion amount Forecasting Methodology for improving extreme learning machine, it is described using training training Training set has 1080～1800 groups when practicing extreme learning machine network.

It is as above a kind of based on the crop transpirstion amount Forecasting Methodology for improving extreme learning machine, realize in limited soil Under conditions of environmental data and meteorological data, the target output of crop transpirstion amount is predicted using FAO-56PM formula computation model Value, the ET_0-PMF56Computing formula it is as follows：

In formula, Δ is the function relation curve slope of saturation vapour pressure-temperature, and unit is KPa DEG C^-1, R_nFor input hat The net radiation of layer, unit is MJm^-2day^-1, G is soil heat flux, is ignored here, and unit is MJm^-2day^-1, e_sFor saturation Vapour pressure, unit is KPa, e_aFor actual water vapor pressure, unit is KPa, and γ represents thermometer constant, and unit is KPa DEG C^-1, Q is Daily mean temperature, unit for DEG C, u is the wind speed of 2 meters of eminences, and unit is ms^-1；

The computing formula of Δ is as follows：

It is as above a kind of based on the crop transpirstion amount Forecasting Methodology for improving extreme learning machine, by coefficient of determination R²Meter Calculate the precision of prediction of the Forecasting Methodology, coefficient of determination R²Closer to 1, precision of prediction is higher, the coefficient of determination R²Calculating it is public Formula is：

Beneficial effect：

(1) present invention predicts crop transpirstion amount using intelligent algorithm extreme learning machine, it is proposed that improves extreme learning machine and calculates Method, the algorithm is better than standard limit learning machine algorithm Generalization Capability and prediction stability；

(2) improvement extreme learning machine algorithm of the invention not only pace of learning quickly, time loss is low, and crop is steamed The precision of prediction of the amount of rising is high；

(3) present invention makes existing internet of things sensors technology in combination with field of agricultural irrigation, combined with intelligent algorithm, real In the case of now limited gathered data, accurate prediction crop transpirstion amount as inexpensive as possible makes the invention applicability Increase.

Description of the drawings

Fig. 1 is the flow chart based on the crop transpirstion amount Forecasting Methodology for improving extreme learning machine；

Fig. 2 is the sigmoid activation primitives of standard limit learning machine algorithm；

Fig. 3 is the sine activation primitives of standard limit learning machine algorithm；

Fig. 4 is the signal number activation primitives of standard limit learning machine algorithm；

Fig. 5 is that standard limit learning machine algorithm is used to predict the schematic network structure of crop transpirstion amount；

Fig. 6 improves RMSE and R that extreme learning machine (PSO_SW extreme learning machines) predicts crop transpirstion amount for the present invention²'s Final result；

Fig. 7 is the RMSE and R that algorithm BP predicts crop transpirstion amount²Final result；

Fig. 8 is the RMSE and R that algorithm PSO_BP predicts crop transpirstion amount²Final result；

Fig. 9 is the RMSE and R that algorithm SVM predicts crop transpirstion amount²Final result；

Figure 10 is the RMSE and R that algorithm extreme learning machine predicts crop transpirstion amount²Final result；

Figure 11 is the RMSE and R that algorithm PSO_ extreme learning machines predict crop transpirstion amount²Final result.

Specific embodiment

With reference to specific embodiment, the present invention is expanded on further.It should be understood that these embodiments are merely to illustrate this Bright rather than restriction the scope of the present invention.In addition, it is to be understood that after the content for having read instruction of the present invention, art technology Personnel can make various changes or modifications to the present invention, and these equivalent form of values equally fall within the application appended claims and limited Fixed scope.

A kind of crop transpirstion amount Forecasting Methodology based on improvement extreme learning machine that the present invention is provided is as shown in figure 1, concrete Prediction steps are as follows：

1) the soil environment data and meteorological data in farmland are gathered, the soil environment data of collection and meteorological data are from upper The agriculture internet of things equipment of information centre of sea market agricultural commission research and development, including put down soil mean daily temperature and humidity, air day Temperature and humidity, per day total solar radiation, the wind speed of 2m eminences, atmospheric pressure, the unit of temperature for DEG C, the unit of humidity For %, the unit of per day total solar radiation is MJm^-2day^-1, the unit of wind speed is ms^-1, the unit of pressure is KPa；

2) to step 1) in all data of collection be normalized and obtain training set；Normalization is referred to all of collection Data normalization normalizes formula as follows to [- 1,1]：

In formula, x is the data after normalization, and x* is to be currently needed for normalized data, x_minTo need normalized data In minimum of a value, x_maxMaximum in need normalized data.

3) train extreme learning machine network using 1080 groups of training sets and improve extreme learning machine, improving extreme learning machine is Refer to the activation primitive as extreme learning machine hidden layer using the function based on addition of waveforms, while using particle swarm optimization algorithm Input weights and threshold value between optimization extreme learning machine network input layer and hidden layer；

4) the soil environment data and meteorological data in farmland are gathered again, and all data to gathering are normalized, By the data input after normalization to improving in extreme learning machine, the Crop transpirstion that prediction is obtained is exported by improvement extreme learning machine Amount；

The improvement step of extreme learning machine is as follows：

(1) extreme learning machine network structure |input paramete is initialized, activation primitive g (x) of hidden layer, standard limit is set Learning machine algorithm be used for predict crop transpirstion amount schematic network structure as shown in figure 5, respectively include input layer, hidden layer and Output layer；Input layer is the soil environment data and meteorological data of collection, and output layer is crop transpirstion amount, and network structure is input The number of plies is 7, and the output number of plies is 1, through test of many times, here in the middle of the implicit number of plies be set to 27 layers of relatively reasonable, selection limit Activation primitive g (x) of habit machine network hidden layer, the selection of activation primitive g (x) of extreme learning machine (ELM) hidden layer is to whole The performance of network plays very crucial effect, and original ELM networks have three kinds of activation primitives：Sigmoid activation primitives are (as schemed Shown in 2), sine activation primitives (as shown in Figure 3) and signal number activation primitives (as shown in Figure 4), these three function pairs ET₀Prediction limited accuracy, the present invention select improve original based on the new activation primitive of addition of waveforms using a kind of Single waveform activation primitive, waveform adopt two kinds of addition of waveforms of inverse hyperbolic SIN function and wavelet function, due to small echo Function then has emphasizes low-and high-frequency signal fitting ability, and inverse hyperbolic SIN function accelerates to a certain extent the convergence of model Speed, this two kinds of function superpositions improve the network structure of original EL M so that the network of hidden layer has higher dynamic Energy disposal ability, activation primitive g (x) is used to calculate the output weights between hidden layer and output layer, the side of activation primitive g (x) Formula is as follows：

In formula, w₀For frequency, w₀=1, j are imaginary number, and x is the data after normalization；

(2) particle swarm parameter is initialized, with input weights and threshold between extreme learning machine network input layer and hidden layer It is worth for particle, all of input weights and sets of threshold values are into population, the limit after the |input paramete initialization of extreme learning machine network Input weights and threshold number between learning machine network input layer and hidden layer determine, according to extreme learning machine network input layer Input weights and threshold number random initializtion particle vector dimension and scope between hidden layer, the initialisation range of particle For [- 0.5,0.5], maximum iteration time k of the particle swarm parameter including population_max=200, Population Size popsize is 30, grain Sub- speed undated parameter c₁=1.49445, c₂=1.49445, the span of the speed of each particle is [- 0.5,0.5], position The span put is [- 0.5,0.5]；

(3) extreme learning machine network and the adaptation in particle swarm optimization algorithm are trained as input data using training set Degree function calculates the fitness of particle, and the root-mean-square error (RMSE) of predicted value and desired value is defined as particle and is adapted to by the present invention Degree function, wherein RMSE is less, shows that precision of prediction is higher, and the concrete formula of fitness function RMSE is：

In formula, ET_0-PM56I () is the calculating by FAO-56PM (Peng Man models) equations to the i-th particle crop transpirstion amount Value, unit is mmday^-1, ET_0-predictedI () is training predicted value of the extreme learning machine to the i-th particle crop transpirstion amount, single Position is mmday^-1, N is the group number of input sample；

ET_0-PMF56Computing formula it is as follows：

In formula, Δ is the function relation curve slope of saturation vapour pressure-temperature, and unit is KPa DEG C^-1, R_nFor input hat The net radiation of layer, unit is MJm^-2day^-1, G is soil heat flux, is ignored here, and unit is MJm^-2day^-1, e_sFor saturation Vapour pressure, unit is KPa, e_aFor actual water vapor pressure, unit is KPa, and γ represents thermometer constant, and unit is KPa DEG C^-1, Q is Daily mean temperature, unit for DEG C, u is the wind speed of 2 meters of eminences, and unit is ms^-1；

The computing formula of Δ is as follows：

(4) according to the fitness size of particle in population, the personal best particle P of particle is obtained_idWith global optimum position Put P_gd, personal best particle P_idWith global optimum position P_gdReplacement criteria be：

In formula, P_iFor the positional information of the i-th particle, RMSE (P_i) be the i-th particle fitness value, RMSE (P_gd) for it is global most The fitness value of excellent particle, RMSE (P_id) it is individual optimal particle fitness value；

(5) according to particle swarm optimization algorithm dynamic tracking personal best particle P_idWith global optimum position P_gdTo carry out not The disconnected speed for updating all particles and position, and particle fitness, more new individual optimum position are recalculated according to fitness function Put P_idWith global optimum position P_gd；Every time in iterative process, particle updates oneself speed of itself by following two formula And position：

In formula, ω is inertia weight, c₁、c₂It is two nonnegative constants, is defined as acceleration factor, r₁、r₂It is two random Number, span is [0,1], and k is current iterations, v_idFor the speed of the i-th particle d dimensions, d=1,2 ..., D, x_id For the position of the i-th particle d dimensions, personal best particle P_id=(p_i1,p_i2,...,p_iD)^T, global optimum position P_gd=(p_g1, p_g2,...,p_gD)^T, v_id ^k+1For the speed of kth the i-th particle of+1 iteration d dimensions, v_id ^kFor kth time the i-th particle of iteration d dimensions Speed, P_id ^kFor kth time iteration personal best particle, x_id ^k+1For the position of kth the i-th particle of+1 iteration d dimensions, P_gd ^kFor kth Secondary iteration global optimum position, x_id ^kFor the position of kth time the i-th particle of iteration d dimensions；

(6) judge whether particle swarm optimization algorithm reaches maximum iteration time k_max, if it is preserve the grain of current iteration Subgroup, the population of current iteration is between the extreme learning machine network input layer of particle swarm optimization algorithm optimization and hidden layer Input weights and threshold value, otherwise step (5) continue iteration；

(7) hidden layer output matrix H is calculated according to activation primitive g (x) of extreme learning machine network hidden layer, specially： N group sample { (X are given in extreme learning machine network_s,t_s), s=1 ... N, hidden layer node isThen extreme learning machine net Network structure is as follows：

In formula, X_s=[X_s1,X_s2,...,X_sn]^T∈Rⁿ, t_s=[t_s1,t_s2,...,t_sm]^T∈R^m, X_sAnd t_sRepresent respectively defeated Enter variable and corresponding output variable, W_v=[W_v1,W_v2,...,W_vn]^TIt is between connection v hidden layer nodes and n input layer Input weight vector, b_vIt is the threshold vector of v hidden layers, β_v=[β_v1,β_v2,...,β_vm]^TIt is connection v hidden layer nodes With the weight vector between m output layer, W_v·X_sInner product is sought in expression；

Above-mentioned equation transform is obtained：

H β=T；

In formula, H is hidden layer output matrix, and β is output weight matrix, and T to expect output matrix, serve as reasons by the element in T The calculated Crop transpirstion value of FAO-56PM (Peng Man models) equation, wherein：

Because the input weights and threshold value between the input layer and hidden layer of particle swarm optimization algorithm optimization are determination value, then W_vAnd b_vValue determination, can calculate hidden layer output matrix H；

(8) calculating limit learning machine output weight matrix β, by input hidden layer nodeActivation primitive g (x), population Input weights and threshold value, hidden layer output matrix between the extreme learning machine network input layer and hidden layer of optimized algorithm optimization H and output weight matrix β obtain improving extreme learning machine, preserve and improve extreme learning machine, calculating limit learning machine output weights The formula of matrix β is as follows：

In formula,The Moore-Penrose generalized inverses of matrix H are asked in expression.

The present invention can also pass through coefficient of determination R except predicted value with the root-mean-square error (RMSE) of desired value²Calculate prediction side The precision of prediction of method, coefficient of determination R²Closer to 1, precision of prediction is higher, coefficient of determination R²Computing formula be：

The present invention predicts the RMSE and R of crop transpirstion amount using extreme learning machine (PSO_SW extreme learning machines) is improved²'s Final result is as shown in fig. 6, R²=0.99158, RMSE=0.19982；

In the case of training set and the data identical being input in improvement extreme learning machine, steamed using BP prediction crops The RMSE and R of the amount of rising²Final result as shown in fig. 7, R²=0.94876, RMSE=0.53117；

In the case of training set and the data identical being input in improvement extreme learning machine, crop is predicted using SVM The RMSE and R of transpiration rate²Final result as shown in figure 8, R²=0.97831, RMSE=0.33566；

In the case of training set, the data being input in improvement extreme learning machine and population initiation parameter identical, The RMSE and R of crop transpirstion amount are predicted using PSO_BP²Final result as shown in figure 9, R²=0.96716, RMSE= 0.50478；

Identical with the data being input in improvement extreme learning machine in training set, the initialization of extreme learning machine network parameter is same In the case of etc. condition, using standard limit learning machine, here activation primitive using sigmoid activation primitives, steam for prediction by crop The RMSE and R of the amount of rising²Final result as shown in Figure 10, R²=0.95514, RMSE=0.46256；

, the population initiation parameter equal conditions identical with the data being input in improvement extreme learning machine in training set, And activation primitive predicts the RMSE and R of crop transpirstion amount using in the case of sigmoid activation primitives using PSO_ELM²'s Final result is as shown in figure 11, R²=0.98006, RMSE=0.31261.

When contrast can be seen that the improvement extreme learning machine prediction crop transpirstion amount using the present invention, coefficient of determination R²More Close 1, predicted value is less with the root-mean-square error of desired value (RMSE), illustrates the improvement extreme learning machine of the present invention relative to one As algorithm have higher precision of prediction.

Claims (6)

1. it is a kind of to it is characterized in that based on the crop transpirstion amount Forecasting Methodology for improving extreme learning machine, the crop transpirstion amount prediction Method is that, based on the crop transpirstion amount Forecasting Methodology for improving extreme learning machine, concrete prediction steps are as follows：

1) the soil environment data and meteorological data in farmland are gathered；

2) to step 1) in all data of collection be normalized and obtain training set；

3) extreme learning machine network is trained using training set and improves extreme learning machine, the improvement extreme learning machine refers to employing Function based on addition of waveforms as extreme learning machine hidden layer activation primitive, while using particle swarm optimization algorithm optimize pole Input weights and threshold value between limit learning machine network input layer and hidden layer；

4) the soil environment data and meteorological data in farmland are gathered again, and all data to gathering are normalized, and will return Data input after one change exports the crop transpirstion amount that prediction is obtained to improving in extreme learning machine by improvement extreme learning machine；

The improvement step of the extreme learning machine is as follows：

(1) extreme learning machine network structure |input paramete is initialized, activation primitive g (x) of hidden layer, the limit study is set The input layer of machine network is soil environment data and meteorological data, and output layer is crop transpirstion amount, the extreme learning machine network |input paramete include input the number of plies, hidden layer nodeWith the output number of plies, the input number of plies is gentle by soil environment data Image data determination, the hidden layer nodeFor stochastic inputs, the output number of plies is 1, and activation primitive g (x) is based on The output weights between hidden layer and output layer are calculated, the equation of activation primitive g (x) is as follows：

$g\left(x\right)=\frac{1}{\sqrt[4]{\mathrm{\π}}}{e}^{{\mathrm{jw}}_{0}x}{e}^{-x^2}+\arcsin h\left(x\right)\≈cos\left(w_{0}x\right)e^{-x^2}+\arcsin h\left(x\right);$

In formula, w₀For frequency, w₀=1, j are imaginary number, and x is the data after normalization；

(2) particle swarm parameter is initialized, is with the input weights and threshold value between extreme learning machine network input layer and hidden layer Particle, all of input weights and sets of threshold values are into population, limit study after the |input paramete initialization of extreme learning machine network Input weights between machine network input layer and hidden layer and threshold number determine, according to extreme learning machine network input layer with it is hidden Containing the input weights between layer and threshold number random initializtion particle vector dimension and scope, the initialisation range of particle for [- 0.5,0.5], the particle swarm parameter includes maximum iteration time k of population_max, Population Size popsize, particle rapidity update Parameter c₁And c₂, the span of the speed of each particle is [- 0.5,0.5], and the span of position is [- 0.5,0.5]；

(3) extreme learning machine network and the fitness letter in particle swarm optimization algorithm are trained as input data using training set Number calculates the fitness of particle, and the concrete formula of fitness function RMSE is：

$RMSE=\sqrt{\frac{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\left({\mathrm{ET}}_{0-PM56}\right(i)-{\mathrm{ET}}_{0-predicted}(i\left)\right)}^2}{N}};$

In formula, ET_0-PM56I () is the calculated value by FAO-56PM equations to the i-th particle crop transpirstion amount, unit is mmday^-1, ET_0-predictedI () is training predicted value of the extreme learning machine to the i-th particle crop transpirstion amount, unit is mmday^-1, N is The group number of input sample；

(4) according to the fitness size of particle in population, the personal best particle P of particle is obtained_idWith global optimum position P_gd, Personal best particle P_idWith global optimum position P_gdReplacement criteria be：

$P_{id}=\{\begin{array}{cc}{P}_i& \left(RMSE\right(P_{id})>RMSE(P_i\left)\right)\\ P_{id}& else\end{array};$

$P_{gd}=\{\begin{array}{cc}{P}_i& \left(RMSE\right(P_{gd})>RMSE(P_i\left)\right)\\ P_{gd}& else\end{array};$

In formula, P_iFor the positional information of the i-th particle, RMSE (P_i) be the i-th particle fitness value, RMSE (P_gd) it is global optimum's grain The fitness value of son, RMSE (P_id) it is individual optimal particle fitness value；

(5) according to particle swarm optimization algorithm dynamic tracking personal best particle P_idWith global optimum position P_gdTo be constantly updated The speed of all particles and position, and particle fitness is recalculated according to fitness function, update personal best particle P_idWith Global optimum position P_gd；Every time in iterative process, particle updates oneself speed itself and position by following two formula：

$v_{id}^{k+1}={\mathrm{\ωv}}_{id}^k+c_1\·r_1\·(p_{id}^k-x_{id}^k)+c_2\·r_2\·(p_{gd}^k-x_{id}^k);$

$x_{id}^{k+1}=x_{id}^k+v_{id}^{k+1};$

In formula, ω is inertia weight, c₁、c₂It is two nonnegative constants, is defined as acceleration factor, r₁、r₂It is two random numbers, takes Value scope is [0,1], and k is current iterations, v_idFor the speed of the i-th particle d dimensions, d=1,2 ..., D, x_idFor i-th The position of particle d dimensions, personal best particle P_id=(p_i1,p_i2,...,p_iD)^T, global optimum position P_gd=(p_g1,p_g2,..., p_gD)^T, v_id ^k+1For the speed of kth the i-th particle of+1 iteration d dimensions, v_id ^kFor the speed of kth time the i-th particle of iteration d dimensions, P_id ^k For kth time iteration personal best particle, x_id ^k+1For the position of kth the i-th particle of+1 iteration d dimensions, P_gd ^kIt is complete for kth time iteration Office's optimal location, x_id ^kFor the position of kth time the i-th particle of iteration d dimensions；

(6) judge whether particle swarm optimization algorithm reaches maximum iteration time k_max, if it is preserve the particle of current iteration Group, the population of current iteration is between the extreme learning machine network input layer of particle swarm optimization algorithm optimization and hidden layer Input weights and threshold value, otherwise step (5) continue iteration；

(7) hidden layer output matrix H is calculated according to activation primitive g (x) of extreme learning machine network hidden layer, specially：In pole N group sample { (X are given in limit learning machine network_s,t_s), s=1 ... N, hidden layer node isThen extreme learning machine network is tied Structure is as follows：

$\underset{v=1}{\overset{\tilde{N}}{\Σ}}{\mathrm{\β}}_{V}g(W_V\·X_S+b_V)=t_S$

In formula, X_s=[X_s1,X_s2,...,X_sn]^T∈Rⁿ, t_s=[t_s1,t_s2,...,t_sm]^T∈R^m, X_sAnd t_sRepresent input respectively to become Amount and corresponding output variable, W_v=[W_v1,W_v2,...,W_vn]^TIt is to connect defeated between v hidden layer nodes and n input layer Enter weight vector, b_vIt is the threshold vector of v hidden layers, β_v=[β_v1,β_v2,...,β_vm]^TIt is connection v hidden layer nodes and m Weight vector between individual output layer, W_v·X_sInner product is sought in expression；

Above-mentioned equation transform is obtained：

H β=T；

In formula, H is hidden layer output matrix, and β is output weight matrix, and to expect output matrix, the element in T is by FAO- to T The calculated Crop transpirstion value of 56PM equations, wherein：

$\begin{array}{c}H(W_1,...,W_{\tilde{N}},b_1,...,b_{\tilde{N}},X_1,...,X_N)\\ ={\left[\begin{array}{ccc}g(W_1\·X_1+b_1)& ...& g(W_{\tilde{N}}\·X_1+b_{\tilde{N}})\\ \·& & \·\\ \·& ...& \·\\ \·& & \·\\ g(W_1\·X_N+b_1)& ...& g(W_{\tilde{N}}\·X_N+b_{\tilde{N}})\end{array}\right]}_{N*\tilde{N}},\mathrm{\β}={\left[\begin{array}{c}{\mathrm{\β}}_1^T\\ {\mathrm{\β}}_2^T\\ \·\\ \·\\ \·\\ {\mathrm{\β}}_{\tilde{N}}^T\end{array}\right]}_{\tilde{N}*m},T={\left[\begin{array}{c}{t}_1^T\\ t_2^T\\ \·\\ \·\\ \·\\ t_L^T\end{array}\right]}_{N*m}\end{array};$

Input weights and threshold value between the input layer optimized due to particle swarm optimization algorithm and hidden layer are determination value, then W_vWith b_vValue determination, can calculate hidden layer output matrix H；

(8) calculating limit learning machine output weight matrix β, by input hidden layer nodeActivation primitive g (x), particle group optimizing Input weights and threshold value, hidden layer output matrix H between the extreme learning machine network input layer of algorithm optimization and hidden layer and Output weight matrix β obtains improving extreme learning machine, preserves and improves extreme learning machine, and the calculating limit learning machine exports weights The formula of matrix β is as follows：

In formula,The Moore-Penrose generalized inverses of matrix H are asked in expression.

2. according to claim 1 a kind of based on the crop transpirstion amount Forecasting Methodology for improving extreme learning machine, its feature exists It is specially in, soil environment data and meteorological data：The per day humidity of soil mean daily temperature, soil, air are per day The per day humidity of temperature, air, per day total solar radiation, the wind speed of 2m eminences and atmospheric pressure, the unit of temperature for DEG C, it is wet The unit of degree is %, and the unit of the per day total solar radiation is MJm^-2day^-1, the unit of the wind speed is ms^-1, it is described The unit of pressure is KPa.

3. according to claim 1 a kind of based on the crop transpirstion amount Forecasting Methodology for improving extreme learning machine, its feature exists In step 2) and step 4) in normalization refer to will collection all data normalizations to [- 1,1], normalize formula as follows：

$x=2\frac{(x*-x_{min})}{(x_{\max }-x_{min})}-1$

In formula, x is the data after normalization, and x* is to be currently needed for normalized data, x_minIn to need normalized data Minimum of a value, x_maxMaximum in need normalized data.

4. according to claim 1 a kind of based on the crop transpirstion amount Forecasting Methodology for improving extreme learning machine, its feature exists In training set has 1080～1800 groups during the employing training set training extreme learning machine network.

5. according to claim 1 a kind of based on the crop transpirstion amount Forecasting Methodology for improving extreme learning machine, its feature exists In the ET_0-PMF56Computing formula it is as follows：

${\mathrm{ET}}_{0-PMF56}=\frac{0.408\mathrm{\Δ}(R_n-G)+\mathrm{\γ}\frac{900}{Q+273}u(e_s-e_a)}{\mathrm{\Δ}+\mathrm{\γ}(1+0.34u)};$

In formula, Δ is the function relation curve slope of saturation vapour pressure-temperature, and unit is KPa DEG C^-1, R_nTo be input into canopy Net radiation, unit is MJm^-2day^-1, G is soil heat flux, is ignored here, and unit is MJm^-2day^-1, e_sFor saturation vapour Pressure, unit is KPa, e_aFor actual water vapor pressure, unit is KPa, and γ represents thermometer constant, and unit is KPa DEG C^-1, Q is flat for day Equal temperature, unit for DEG C, u is the wind speed of 2 meters of eminences, and unit is ms^-1；

The computing formula of Δ is as follows：

$\mathrm{\Δ}=\frac{4.098\left(0.6108e^{\frac{17.27Q}{Q+237.3}}\right)}{{(Q+237.3)}^2}.$

6. according to claim 1 a kind of based on the crop transpirstion amount Forecasting Methodology for improving extreme learning machine, its feature exists In by coefficient of determination R²Calculate the precision of prediction of the Forecasting Methodology, coefficient of determination R²Closer to 1, precision of prediction is higher, institute State coefficient of determination R²Computing formula be：

$R^2=1-\frac{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\left({\mathrm{ET}}_{0-predicted}\right(i)-{\mathrm{ET}}_{0-PM56}(i\left)\right)}^2}{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\left({\mathrm{ET}}_{0-PM56}\right(i)-{\mathrm{meanET}}_{0-predicted}(i\left)\right)}^2}.$