A method for predicting and controlling temperature of cold chain logistics

Abstract

The present invention discloses a cold chain logistics temperature prediction method and a temperature control method. The collected data are first normalized; then the data set is input into the ELM model to calculate the model evaluation value. The input weight w and the hidden layer bias b of the ELM model are optimized using the mayfly algorithm, the positions of male and female mayflies are continuously updated, and a crossover operation is performed to obtain new w and b, and a new model evaluation value is calculated based on the new w and b, and the minimum model evaluation value is updated to achieve the minimum error of the prediction result. It is determined whether the mayfly algorithm has reached the maximum number of iterations, and the optimal solution is output if the condition is met, otherwise iterative optimization is continued until the stop condition is reached. The optimal parameters of the mayfly algorithm are output, the results are brought into the extreme learning machine model, the predicted values ​​are output, and the performance of the model is evaluated. This method can obtain more accurate temperature prediction results and achieve effective temperature control based on the results.

Description

Technical Field

[0001] The present invention relates to an improvement of a cold chain logistics temperature prediction method, and specifically to a cold chain logistics temperature prediction method and a temperature control method based on an ephemeral-extreme learning machine, and belongs to the technical field of cold chain logistics. Background Art

[0002] With the improvement of living standards, people's demand for fresh foods such as fruits and vegetables is increasing. Compared with ordinary goods, fresh foods such as fruits and vegetables still have vital signs during transportation, and they will respire and consume organic matter, thereby reducing the nutritional content of fresh foods. In addition to respiration, the intrinsic factors of fresh foods (heat, moisture, climate factors) and transportation environment (vibration, temperature, oxygen, carbon dioxide) will affect the quality. Among these factors, temperature has the greatest impact. For some temperature-sensitive fresh foods, if they are not transported within the required temperature range, harmful substances may even be produced, threatening the health of consumers. Therefore, using cold chain logistics to transport fresh products is an important means to reduce food waste and ensure food safety. Since fresh foods need to be in a low-temperature environment at all times, this has triggered research on temperature control of cold chain logistics by scholars in many fields including engineering, food, and supply chain.

[0003] At present, the research on the temperature of cold chain logistics is still focused on temperature monitoring. In the research on temperature monitoring of cold chain logistics, most scholars reflect the temperature monitoring of cold chain logistics through the medium of temperature monitoring of refrigerated trucks. Their work focuses on establishing a cold chain logistics temperature monitoring system and focuses on technical implementation. In earlier studies, wireless sensor networks (WSN) were widely used in cold chain logistics temperature monitoring systems. However, these studies all monitor the current temperature of cold chain logistics, and stakeholders can only be notified after the temperature deviates. However, for temperature-sensitive fresh foods, temperature deviations in a short period of time are sufficient to cause changes in the quality of fresh foods. Therefore, it is necessary to predict the temperature change trend and actively identify potential risks. As far as we know, at present, the research on temperature prediction of cold chain logistics is still very limited. In the limited research, it is found that the current research has the following deficiencies: On the one hand, the factors considered are not comprehensive. Existing studies do not consider the impact of spatial position on temperature, and do not consider the impact of factors such as load on temperature. Second, the prediction needs to be improved. Existing studies mainly use neural network methods for prediction, but neural networks converge slowly and are prone to overfitting and underfitting. The prediction accuracy of the model needs to be improved. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the purpose of the present invention is to propose a cold chain logistics temperature prediction method and a temperature control method based on the mayfly-extreme learning machine. This method can obtain more accurate temperature prediction results and achieve effective temperature control based on the results.

[0005] The technical solution of the present invention is achieved in this way:

[0006] A cold chain logistics temperature prediction method is carried out in the following steps:

[0007] 1) First, establish the extreme learning machine prediction model and optimize its input weight w and hidden layer bias b;

[0008] 1.1) Collect temperature-related data in cold chain transport vehicles or storage equipment, and collect temperature values ​​at different locations in cold chain transport vehicles or storage equipment at the same time, filter out the highest and lowest temperatures in the temperature values, and take the temperature-related data and the highest and lowest temperatures at the corresponding time as a group of data; repeat the collection at a certain interval to obtain several groups of data; normalize all the data in the above several groups of data collected;

[0009] 1.2) The extreme learning machine consists of an input layer, a hidden layer, and an output layer. The number of input neurons of the extreme learning machine is set to correspond to the number of temperature-related data, and the number of output neurons is two, corresponding to the highest temperature prediction value and the lowest temperature prediction value; the initial parameters of the extreme learning machine are randomly set, including the input weight w and the hidden layer bias b;

[0010] 1.3) Optimize the input weight w and hidden layer bias b of the extreme learning machine using the mayfly algorithm, and set the input weight w and hidden layer bias b of the extreme learning machine as candidate solution individuals of the mayfly algorithm, that is, the input weight w and hidden layer bias b of the extreme learning machine are expressed by the position of each mayfly. At a certain moment, all mayfly positions represent feasible solutions, but there is only one relatively optimal solution; initialize the mayfly population at the same time, including setting the population size N, the maximum number of iterations, the attraction constants P1 and P2, the visibility constant α, the random walk coefficient w, the dance coefficient v and the crossover coefficient;

[0011] 1.4) Input the temperature-related data in each group of data collected in step 1.1) into the extreme learning machine model under the input weight w corresponding to one of the mayflies and the hidden layer bias b for training, and obtain two outputs of the output layer, namely, the maximum temperature prediction value and the minimum temperature prediction value of each group of data, thereby obtaining the maximum temperature prediction value and the minimum temperature prediction value of all groups of data, and calculating the corresponding extreme learning machine model evaluation value according to the following formula to evaluate the prediction effect of extreme learning;

[0012]

[0013] Among them, X c Represents the actual maximum temperature value, X cz Refers to the actual maximum temperature value X c The corresponding maximum temperature prediction value, X d Represents the actual value of the lowest temperature, X dz Refers to the actual value of the lowest temperature X d The corresponding minimum temperature prediction value; h refers to the number of data groups;

[0014] 1.5) changing the input weights w and hidden layer biases b represented by different mayflies and training according to step 1.4) to obtain the evaluation values ​​corresponding to the input weights w and hidden layer biases b represented by the different mayflies; until the input weights w and hidden layer biases b represented by all mayflies are rotated once, the evaluation values ​​corresponding to the input weights w and hidden layer biases b corresponding to all mayflies are obtained, and the input weights w and hidden layer biases b corresponding to the mayfly with the smallest evaluation value are selected as the input weights w and hidden layer biases b after this round of optimization;

[0015] 1.6) According to the movement rule of the mayfly algorithm, one mayfly iteration is completed, that is, the position of the male mayfly, the position of the female mayfly and the mating of the mayfly are updated, so as to obtain all the mayflies and their positions after this iteration; thereby, the input weights w and the hidden layer bias b corresponding to the multiple mayflies after this iteration are obtained;

[0016] 1.7) Repeat steps 1.4)-1.5) to calculate the minimum evaluation value after this iteration, compare the minimum evaluation value after this iteration with the minimum evaluation value before the iteration, and use the minimum evaluation value between the two as the new minimum evaluation value, and use the input weight w and hidden layer bias b of the mayfly representation corresponding to the new minimum evaluation value as the newly optimized input weight w and hidden layer bias b;

[0017] 1.8) Determine whether the maximum number of iterations of the Mayfly algorithm has been reached. If so, output the newly optimized input weight w and hidden layer bias b obtained in step 1.7), which are the final optimized input weight w and hidden layer bias b; otherwise, jump to step 1.6) and repeat steps 1.6)-1.7) until the stop condition is reached;

[0018] 2) Collecting actual data related to temperature prediction in cold chain transport vehicles or storage equipment that require temperature prediction, the number and type of the actual data are consistent with the number and type of temperature-related data in step 1.1);

[0019] 3) Input the actual data collected into the extreme learning machine prediction model under the input weight w and hidden layer bias b finally optimized in step 1) to obtain two outputs of the output layer, which are the highest temperature prediction value and the lowest temperature prediction value corresponding to the actual data.

[0020] Among them, there are five temperature-related data in step 1.1), namely, outside temperature, air conditioning air outlet speed, air conditioning air outlet temperature, cargo capacity and whether the car door is open; the input layer corresponds to five inputs, corresponding to outside temperature, air conditioning air outlet speed, air conditioning air outlet temperature, cargo capacity and whether the car door is open.

[0021] Furthermore, in step 1.1), when collecting temperature values ​​at different locations in a cold chain transport vehicle or storage equipment at the same time, temperature values ​​at five different locations are collected. The five locations are the centers of the other five inner surfaces of the vehicle compartment except the inner surface of the vehicle compartment where the door is located. A temperature sensor is set at each of the five locations to measure the temperature of the refrigerated truck.

[0022] The present invention also proposes a temperature control method for cold chain logistics. The above method is first used to predict the temperature to be controlled. If the maximum temperature prediction value and the minimum temperature prediction value are both within the set range, the current air-conditioning outlet temperature and air-outlet speed are maintained unchanged; if one of the maximum temperature prediction value and the minimum temperature prediction value exceeds the set range, the expected maximum temperature and minimum temperature as well as the outside temperature, cargo volume, and whether the door is open are input into the extreme learning machine prediction model through the input layer, and the air-conditioning outlet temperature prediction value and air-outlet speed prediction value are obtained through the two outputs of the output layer of the extreme learning machine prediction model. The air-conditioning outlet temperature and air-outlet speed are adjusted to the outlet temperature prediction value and the outlet speed prediction value to achieve temperature control.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] The cold chain logistics temperature prediction method based on the Mayfly-extreme learning machine proposed in the present invention uses the Mayfly algorithm (MA) to optimize the ELM parameters (i.e., input weight w and hidden layer bias b) that have the most important impact on the output result, and the mean relative error (MRE), mean absolute error (MAE), root mean square error (RMSE), mean square error (MSE) and correlation coefficient (R 2 ) and other five indicators to evaluate the prediction results. The results show that the biological heuristic algorithm can improve the prediction accuracy of the ELM model, and the MA algorithm is better than the particle swarm algorithm and genetic algorithm in optimizing the ELM parameters.

[0025] Based on the predicted temperature, if the predicted value is within the set range, the current air-conditioning outlet temperature and air-outlet speed are maintained unchanged; if the predicted value exceeds the set range, the expected temperature value is input into the extreme learning machine prediction model to obtain the air-conditioning outlet temperature predicted value and the air-outlet speed predicted value, and the air-conditioning outlet temperature and air-outlet speed are adjusted to the air-outlet temperature predicted value and the air-outlet speed predicted value to achieve temperature control, thereby realizing timely temperature regulation.

[0026] Compared with GA and PSO, MA algorithm has better global search ability. The above analysis shows that MA-ELM is an effective method to predict temperature and air conditioning status.

[0027] By accurately predicting the temperature of cold chain logistics, the management level of cold chain logistics can be improved. Accurate temperature prediction and regulation can greatly reduce food and ingredient waste and ensure food quality. At the same time, accurate temperature control can avoid temperature regulation too low, reduce energy waste and save energy. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is the network structure diagram of extreme learning machine.

[0029] Figure 2 It is a flow chart of the MA-ELM solution steps proposed in the present invention.

[0030] Figure 3 This is a graph showing how the training set MSE of model 1 and model 2 changes as the number of hidden layer nodes increases.

[0031] Figure 4 This is a comparison chart of the actual and predicted values ​​of low and high temperatures under different algorithms.

[0032] Figure 5 It is the low temperature and high temperature prediction error diagram under different algorithms.

[0033] Figure 6 It is the fitness convergence diagram of MA, GA, and PSO algorithms optimizing ELM parameters in the maximum and minimum temperature prediction.

[0034] Figure 7 This is a comparison chart of the actual and predicted values ​​of the air outlet speed and air outlet temperature under different algorithms.

[0035] Figure 8 It is the prediction error diagram of air outlet speed and air outlet temperature under different algorithms.

[0036] Fig. 9 It is the fitness convergence diagram of MA, GA, and PSO algorithms optimizing ELM parameters in the prediction of outlet air speed and outlet air temperature. DETAILED DESCRIPTION

[0037] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0038] In order to better understand the present method, the extreme learning machine and the mayfly algorithm related to the present invention are first introduced.

[0039] 1. Extreme Learning Machine

[0040] The extreme learning machine is an advanced single hidden layer feedforward neural network algorithm. It overcomes the problem that the results of traditional feedforward neural networks are prone to fall into local optimality and slow convergence due to the use of gradient descent algorithm. The extreme learning machine consists of an input layer, a hidden layer, and an output layer. Its network structure is as follows: Figure 1 As shown. Use the extreme learning machine to train N arbitrary samples (x i ,y i ), where x i Represents input, y i Represents the corresponding x i The expected output of the extreme learning machine is o j The expression of is as follows (1).

[0041]

[0042] Among them, β i represents the output weight, g(x) represents the activation function, and w i represents the input weight, b i represents the bias of the hidden layer, and L represents the number of nodes in the hidden layer.

[0043] The goal of the extreme learning machine is to minimize the output error, which can be expressed as formula (2), that is, there exists w i , b i , β i So that formula (3) is valid. The matrix form of formula (3) is Hβ=T, where H represents the output matrix of the hidden layer, as shown in formula (4), β is the weight matrix, and T is the output matrix, as shown in formula (5).

[0044]

[0045]

[0046]

[0047]

[0048] During the training process of the extreme learning machine, w and b are randomly generated by the algorithm, and their numerical values ​​do not change. Therefore, according to formula (4), the output matrix H of the hidden layer is determined. When the number of neurons in the hidden layer is the same as the number of samples in the training set, for any w and b, the extreme learning machine can approximate the training samples with zero error. However, in most cases, the number of training samples is much larger than the number of neurons in the hidden layer, which makes it impossible for the extreme learning machine to approximate the training samples with zero error. Therefore, the approximate output weight β can be obtained according to formula (6): * , where H + is the Moore-Penrose generalized inverse of the matrix H.

[0049] β * =H+T (6)

[0050] 2. Mayfly Algorithm

[0051] It is known from the above that the two parameters that affect the performance of the extreme learning machine include the input weight w and the hidden layer bias b. The present invention proposes to use the mayfly algorithm to optimize the above two parameters. The optimization of the mayfly algorithm includes three stages, namely the movement of male mayflies, the movement of female mayflies, and the mating of males and females.

[0052] Stage 1. Movement of male mayflies

[0053] Male mayflies gather together, and each male mayfly will adjust its position based on its own experience and the experience of surrounding mayflies. The position update formula is as follows (7), where x i (t) represents the position of mayfly i at time t, x i (t+1) and v i (t+1) represent the position and velocity of mayfly i at time t+1. i (0)∈[x min , x max ]. In this algorithm, it is assumed that the male mayfly performs the nuptial dance at a constant speed, and the speed calculation formula is as follows (8). id (t) and x id (t) represent the velocity and position of mayfly i in dimension d (d = 1, 2...n) at time t, p1 and p2 are positive attraction constants, and α represents the visibility range of mayflies. i represents the historical best position of mayfly i. When solving the minimization problem, the calculation formula of the historical best position of mayfly i at the next time step is as shown in (9), and f(·) represents the objective function used to measure the quality of the solution, that is, the fitness function. gbest represents the global best position of the mayfly, which is defined as in formula (10), and N represents the total number of male mayflies in the population. r mRepresents the distance between the current position and pbest, calculated as formula (11), r n Represents the distance between the current position and gbest, calculated as formula (12).

[0054] x i (t+1)=x i (t)+v i (t+1) (7)

[0055]

[0056]

[0057] gbest∈min{f(pbest1), f(pbest2),…, f(pbest N )} (10)

[0058]

[0059]

[0060] In order to obtain the optimal solution, the best male mayfly in the population will constantly change its speed to perform a continuous up and down dance, which introduces a random element into the algorithm. In this case, the speed calculation formula at time t+1 is as follows (13), where v is the dance coefficient and r is a random number between [-1, 1].

[0061] v id (t+1)=v id (t)+vr (13)

[0062] Stage 2. Movement of female mayflies

[0063] Female mayflies will not gather together, instead they will fly towards males to mate. The position of female mayfly i at time t+1 is calculated as (14), y i (t) represents the position of female mayfly i at time t, v i (t+1) represents the speed of female mayfly i at time t+1. i (0)∈[y min ,y max ].

[0064] The Mayfly algorithm sets attraction as a fixed process, that is, selecting mating partners based on performance, which means that the best performing females are attracted to the best performing males. When solving the minimization problem, the speed of the female is calculated as shown in Equation (15). Where y i (t) and v id(t) represent the position and velocity of female mayfly i in dimension d (d = 1, 2...n) at time t, p2 is a positive attraction constant, α is a visibility constant, r0 represents the distance between male and female. w represents the random walk coefficient of female when she is not attracted by male, and r is a random number between [-1, 1].

[0065] y i (t+1)=y i (t)+v i (t+1) (14)

[0066]

[0067] Stage 3. Mayflies mating

[0068] The crossover operator represents the mating process of male and female mayflies. According to the principle of mayfly attraction, one male mayfly is generated as the father and one female mayfly is generated as the mother. The generation of the father and the mother can be random or based on the fitness function. The two offspring generated are as shown in equations (16) and (17), where s1 and s2 refer to offspring 1 and offspring 2 respectively, μ represents a random value within a specific range, male represents the father, female represents the mother, and the initial velocity of the offspring is 0.

[0069] s1=μ×male+(1-μ)×female (16)

[0070] s2=μ×female+(1-μ)×male (17)

[0071] 3. The cold chain logistics temperature prediction method based on the Ephemerida-extreme learning machine of the present invention

[0072] 1) First, establish the extreme learning machine prediction model and optimize its input weight w and hidden layer bias b;

[0073] 1.1) Collect temperature-related data in cold chain transport vehicles or storage equipment, and collect temperature values ​​at different locations in cold chain transport vehicles or storage equipment at the same time, filter out the highest and lowest temperatures in the temperature values, and take the temperature-related data and the highest and lowest temperatures at the corresponding time as a group of data; repeat the collection at a certain interval to obtain several groups of data; normalize all the data in the above several groups of data collected;

[0074] 1.2) The extreme learning machine consists of an input layer, a hidden layer, and an output layer. The number of input neurons of the extreme learning machine is set to correspond to the number of temperature-related data, and the number of output neurons is two, corresponding to the highest temperature prediction value and the lowest temperature prediction value; the initial parameters of the extreme learning machine are randomly set, including the input weight w and the hidden layer bias b;

[0075] 1.3) Optimize the input weight w and hidden layer bias b of the extreme learning machine using the mayfly algorithm, and set the input weight w and hidden layer bias b of the extreme learning machine as candidate solution individuals of the mayfly algorithm, that is, the input weight w and hidden layer bias b of the extreme learning machine are expressed by the position of each mayfly. At a certain moment, all mayfly positions represent feasible solutions, but there is only one relatively optimal solution; initialize the mayfly population at the same time, including setting the population size N, the maximum number of iterations, the attraction constants P1 and P2, the visibility constant α, the random walk coefficient w, the dance coefficient v and the crossover coefficient;

[0076] 1.4) The temperature-related data in each group of data collected in step 1.1) is input into the extreme learning machine model under the input weight w corresponding to one of the mayflies and the hidden layer bias b for training, and two outputs of the output layer are obtained, namely, the maximum temperature prediction value and the minimum temperature prediction value of each group of data, thereby obtaining the maximum temperature prediction value and the minimum temperature prediction value of all groups of data, and calculating the corresponding extreme learning machine model evaluation value according to the following formula to evaluate the prediction effect of extreme learning; the smaller the evaluation value, the better the prediction effect.

[0077]

[0078] Among them, X c Represents the actual maximum temperature value, X cz Refers to the actual maximum temperature value X c The corresponding maximum temperature prediction value, X d Represents the actual value of the lowest temperature, X dz Refers to the actual value of the lowest temperature X d The corresponding minimum temperature prediction value; h refers to the number of data groups;

[0079] 1.5) changing the input weights w and hidden layer biases b represented by different mayflies and training according to step 1.4) to obtain the evaluation values ​​corresponding to the input weights w and hidden layer biases b represented by the different mayflies; until the input weights w and hidden layer biases b represented by all mayflies are rotated once, the evaluation values ​​corresponding to the input weights w and hidden layer biases b corresponding to all mayflies are obtained, and the input weights w and hidden layer biases b corresponding to the mayfly with the smallest evaluation value are selected as the input weights w and hidden layer biases b after this round of optimization;

[0080] 1.6) According to the movement rule of the mayfly algorithm, one mayfly iteration is completed, that is, the position of the male mayfly, the position of the female mayfly and the mating of the mayfly are updated, so as to obtain all the mayflies and their positions after this iteration; thereby, the input weights w and the hidden layer bias b corresponding to the multiple mayflies after this iteration are obtained;

[0081] 1.7) Repeat steps 1.4)-1.5) to calculate the minimum evaluation value after this iteration, compare the minimum evaluation value after this iteration with the minimum evaluation value before the iteration, and use the minimum evaluation value between the two as the new minimum evaluation value, and use the input weight w and hidden layer bias b of the mayfly representation corresponding to the new minimum evaluation value as the newly optimized input weight w and hidden layer bias b;

[0082] 1.8) Determine whether the maximum number of iterations of the Mayfly algorithm has been reached. If so, output the newly optimized input weight w and hidden layer bias b obtained in step 1.7), which are the final optimized input weight w and hidden layer bias b; otherwise, jump to step 1.6) and repeat steps 1.6)-1.7) until the stop condition is reached;

[0083] 2) Collecting actual data related to temperature prediction in cold chain transport vehicles or storage equipment that require temperature prediction, the number and type of the actual data are consistent with the number and type of temperature-related data in step 1.1);

[0084] 3) Input the actual data collected into the extreme learning machine prediction model under the input weight w and hidden layer bias b finally optimized in step 1) to obtain two outputs of the output layer, which are the highest temperature prediction value and the lowest temperature prediction value corresponding to the actual data.

[0085] In step 1.1), there are five data related to temperature, namely, outside temperature, air conditioning air outlet speed, air conditioning air outlet temperature, cargo capacity and whether the door is open; the input layer corresponds to five inputs, corresponding to outside temperature, air conditioning air outlet speed, air conditioning air outlet temperature, cargo capacity and whether the door is open.

[0086] Step 1.1) When collecting temperature values ​​at different locations in a cold chain transport vehicle or storage equipment at the same time, collect temperature values ​​at five different locations. The five locations are the centers of the other five inner surfaces of the compartment except the inner surface where the door is located. A temperature sensor is set at each of the five locations to measure the temperature of the refrigerated truck.

[0087] A cold chain logistics temperature control method, firstly uses the above method to predict the temperature to be controlled, if the maximum temperature prediction value and the minimum temperature prediction value are both within the set range, the current air-conditioning outlet temperature and air-conditioning outlet speed are maintained unchanged; if one of the maximum temperature prediction value and the minimum temperature prediction value exceeds the set range, the expected maximum temperature and minimum temperature as well as the outside temperature, the cargo volume, and whether the door is open are input into an extreme learning machine prediction model through an input layer, and the air-conditioning outlet temperature prediction value and the air-conditioning outlet speed prediction value are obtained through two outputs of the output layer of the extreme learning machine prediction model, and the air-conditioning outlet temperature and the air-conditioning outlet speed are adjusted to the outlet temperature prediction value and the outlet speed prediction value to achieve temperature control.

[0088] The same extreme learning machine prediction model, when used to predict temperature, is a temperature prediction model; when used to predict the air outlet temperature and air outlet speed of the air conditioner, it is an air conditioning prediction model.

[0089] 4. Case Studies

[0090] This study takes Company A (anonymous) as the research object. Company A is a third-party logistics company specializing in cold chain logistics and distribution.

[0091] 4.1 Data Source

[0092] This section introduces the data sources for temperature prediction and control. Since the temperature field distribution of refrigerated trucks is irregular, the temperature at different locations in the refrigerated trucks will be different. This study combines the internal temperature measurement method of refrigerated trucks in the performance test method of insulated trucks and refrigerated trucks in the automobile industry standard QC / T 449-2000 of the People's Republic of China and the actual situation. In addition to the inner surface of the compartment where the door is located, a temperature measurement point is set at the center of each of the other five compartments to measure the temperature of the refrigerated truck. In addition to temperature, data collection is also required for five variables, including the outside temperature, air conditioning outlet speed, air conditioning outlet temperature, cargo capacity, and whether the door is open.

[0093] This study samples data every 15 minutes, where the outside temperature, air conditioning air speed, air conditioning air temperature, cargo capacity, and temperature inside the refrigerated truck are all actual measured values. Whether the door is open uses a 0-1 variable, where 0 represents the door closed and 1 represents the door open. The data sampling period is 20 days (2020.04.12-2020.05.01), and a total of 850 valid data were obtained, of which 800 data were randomly selected as training sets and the remaining 50 data were used as test sets. The variables that affect the temperature inside the refrigerated truck have different units. In order to eliminate the impact of different units on the results, the data needs to be normalized and all variables need to be converted into dimensionless values ​​that fall within (0, 1), as shown in the following formula, where x a represents the ath input, xai Represents the input x a The converted value, x max , x min Represents x a The maximum and minimum values ​​of .

[0094]

[0095] 4.2 Parameter settings

[0096] The number of hidden nodes has an important impact on the performance of the extreme learning machine. Too few hidden nodes will result in insufficient model accuracy, while too many hidden nodes will cause the model to overfit. Based on experience, the optimization range of the initial number of hidden nodes of the two prediction models (temperature prediction model and air conditioning prediction model) involved in the present invention is set to [8, 35], and the mean square error of the training set is used as the evaluation index to find the optimal number of hidden layer nodes. The image of the mean square error of the training set of the two models is shown in Figure 3 (a) and (b) in Figure 3 (a)(b) It can be seen that when the number of hidden nodes is 25 and 32 respectively, the extreme learning machine model works best.

[0097] Secondly, the number of input neurons and output neurons needs to be set according to the research purpose. As mentioned above, the temperatures of the five points where the temperature is measured in the refrigerated truck are different. These temperatures need to be within the specified temperature range. Therefore, it is only necessary to require that the maximum temperature and the minimum temperature are within the specified range to ensure that the temperatures of all the temperature measurement points are within the specified range. Therefore, the seven variables that need to be input into the model are determined, including the outside temperature, the air-conditioning outlet speed, the air-conditioning outlet temperature, the cargo volume, whether the door is open, the maximum temperature, and the minimum temperature. In the temperature prediction model, there are five input neuron nodes, including the outside temperature, the air-conditioning outlet speed, the air-conditioning outlet temperature, whether the door is open, and the cargo volume. The output neurons are set to two, namely the maximum temperature and the minimum temperature. In the temperature control model, there are five input neuron nodes, including the outside temperature, the maximum temperature, the minimum temperature, the cargo volume, and whether the door is open. There are two output neuron nodes, the air-conditioning outlet speed and the air-conditioning outlet temperature. The specific parameter settings of the extreme learning machine are shown in Table 1.

[0098] Table 1. Parameters of extreme learning machine

[0099] Model System Level Input Neuron Hidden Node Output Neuron Activation Function Temperature prediction model 1 3 5 25 2 Sigmoid Temperature Control Model 2 3 5 32 2 Sigmoid

[0100] In order to verify the effectiveness of the Mayfly algorithm for optimizing the parameters of the extreme learning machine, the present invention compares the Mayfly algorithm with the genetic algorithm and the particle swarm algorithm. The parameters of the relevant algorithms are set as follows: (1) Mayfly algorithm: population size n = 30, attraction constant P1 = 1, P2 = 0.5, visibility coefficient α = 2, random walk coefficient t = 0.1, dance coefficient w = 0.1, crossover rate crossover = 0.95. (2) Genetic algorithm: population size n = 30, crossover probability 0.8, mutation probability 0.2. (3) Particle swarm algorithm: population size 30, learning factor C1 is 2, learning factor C2 is 2, inertia factor 0.9. The maximum number of iterations of the three algorithms is set to 1000.

[0101] 4.3 Determination of evaluation indicators

[0102] Before analyzing the prediction results, it is necessary to determine the evaluation indicators. The present invention uses the mean relative error (MRE), mean absolute error (MAE), root mean square error (RMSE), mean square error (MSE) and correlation coefficient (R 2 ) is used to evaluate the prediction results, as shown in Table 2, where Y i is the true value, Y ia is the predicted value, Y ib is the average of the true value, i is the sample i=1, 2, ..., n.

[0103] The closer the values ​​of MRE, MAE, RMSE and MSE are to 0, the better the model effect is and the higher the prediction accuracy is. 2 The closer the value is to 1, the better the model fitting effect is.

[0104] Table 2. Evaluation indicators of prediction results

[0105]

[0106]

[0107] 5. Analysis of prediction results

[0108] In order to verify the effectiveness of the Ephemerida-ELM model proposed in this paper, several experiments were carried out. Section 5.1 analyzes the temperature prediction results in cold chain logistics distribution, and Section 5.2 analyzes the prediction of air volume and air temperature of air conditioners.

[0109] 5.1 Temperature prediction

[0110] This section uses four extreme learning machine models (ELM, GA-ELM, PSO-ELM, MA-ELM) to predict the temperature in the cold chain logistics distribution process. Figure 4The distribution diagrams of the actual and predicted values ​​of low and high temperatures under the four models are shown. Figure 4 It can be seen that the predicted value of MA-ELM is closer to the target value regardless of high or low temperature. Figure 5 The low and high temperature prediction error plots are confirmed. Figure 5 It can be found that compared with the other three models, the error fluctuation of MA-ELM prediction is closer to the zero line. Furthermore, in order to increase the accuracy of the analysis, the five indicators MRE, MAE, RMSE, MSE and R introduced in Section 4.3 are used. 2 The prediction results were evaluated and Table 3 was obtained.

[0111] It can be seen from Table 3 that when predicting the minimum temperature, the prediction effect of MA-ELM (MRE=0.057, MAE=0.169RMSE=0.218, MSE=0.047, R2=0.909) is better than that of GA-ELM (MRE=0.063, MAE=0.185RMSE=0.246, MSE=0.060, R2=0.909), PSO-ELM (MRE=0.069, MAE=0.25, RMSE=0.246, MSE=0.060, R2=0.909), and ELM (MRE=0.109, MAE=0.295, RMSE=0.381, MSE=0.145, R2=0.526). Specifically, in terms of the calculation of the four indicators of MRE, MAE, RMSE, and MSE, compared with ELM, the genetic algorithm reduced by 0.046, 0.11, 0.135, and 0.085, the particle swarm algorithm reduced by 0.04, 0.09, 0.126, and 0.08, and the mayfly algorithm reduced by 0.052, 0.126, 0.163, and 0.098. As for the R2 indicator, the three optimization algorithms increased by 0.239, 0.158, and 0.383, respectively. The above results show that compared with the standard ELM, the use of optimization algorithms (genetic, particle swarm, and mayfly algorithms) improves the prediction accuracy of the extreme learning machine. Moreover, among these optimization algorithms, the mayfly algorithm has the best effect, which shows the positive role of the mayfly algorithm in optimizing the extreme learning machine in improving the prediction accuracy. Similarly, the results of high temperature prediction were analyzed, and it was found that the four models were ranked from best to worst as follows: MA-ELM (MRE=0.037, MAE=0.132RMSE=0.173, MSE=0.030, R2=0.924)>GA-ELM (MRE=0.042, MAE=0.147RMSE=0.185, MSE=0.034, R2=0.797)>PSO-ELM (MRE=0.047, MAE=0.169, RMSE=0.222, MSE=0.049, R2=0.704)>ELM (MRE=0.086, MAE=0.299, RMSE=0.373, MSE=0.139, R2=0.543). Therefore, the Mayfly-Extreme Learning Machine model has good performance in both high and low temperature prediction. However, it is worth noting that there are certain differences in the prediction results of low temperature (MRE=0.057, MAE=0.169, RMSE=0.218, MSE=0.047, R2=0.909) and high temperature (MRE=0.037, MAE=0.132, RMSE=0.173, MSE=0.030, R2=0.924) under MA-ELM optimization, and the prediction performance of high temperature is better than that of low temperature.However, in general, the model can accurately predict the highest and lowest temperatures in the cold chain logistics distribution process.

[0112] Table 3. Low temperature and high temperature prediction evaluation results

[0113]

[0114] pass Figure 6 The fitness convergence curves of GA, PSO and MA in the process of optimizing the parameters of the extreme learning machine can further illustrate the superiority of the mayfly algorithm. The fitness function convergence curves of the three algorithms follow the same trend, that is, the fitness values ​​are first greatly reduced, and then the decreasing trend continues to decrease until it approaches stability. In this process, it can be seen that the number of iterations when the MA algorithm reaches a stable state is less, and the fitness value when the MA algorithm reaches a stable state is lower. This result shows the effectiveness of using the mayfly algorithm proposed in the present invention to optimize the parameters of the extreme learning machine.

[0115] 5.2 Air Conditioning Status Prediction

[0116] The previous section predicted the maximum and minimum temperatures in the car. This section is about timely adjustment of the prediction results. That is, when the predicted maximum and minimum temperatures are not within the range required by cold chain logistics, the expected temperature can be input into the model to obtain the adjustment value of the air conditioning state (air outlet speed and air outlet temperature), thereby achieving precise temperature control.

[0117] This section uses MA-ELM to predict the outlet air velocity and outlet air temperature, and uses ELM, GA-ELM, and PSO-ELM as comparison models. The distribution of the prediction results of the four models and the true values ​​is shown in the figure. Figure 7 As shown in the figure, it can be seen that MA-ELM has the best prediction effect, which is because its predicted value is closest to the true value. Figure 8 The forecast errors of the four models are shown, among which the MA-ELM model has the smallest fluctuation range of forecast errors, which also shows that its forecast effect is the best. Fig. 9 From the algorithm evolution curve, we can also see that the optimization of the extreme learning machine by the MA algorithm can maximize the accuracy of the model. Compared with GA and PSO optimization, the MA algorithm can find a solution with a lower fitness value and reach a stable state first.

[0118] From the above graphical perspective, it is concluded that MA-ELM has the best prediction effect. The following is a further verification of this conclusion through numerical methods. Table 4 lists the five indicators MRE, MAE, RMSE, MSE and R for evaluating the prediction results of outlet air temperature and outlet air volume. 2In predicting wind speed, compared with the standard ELM, MA-ELM (MRE = 0.053, MAE = 0.186, RMSE = 0.217, MSE = 0.047, R2 = 0.881) can improve the prediction accuracy to the greatest extent, followed by GA-ELM (MRE = 0.082, MAE = 0.292, RMSE = 0.355, MSE = 0.126, R2 = 0.843), and finally PSO-ELM (MRE = 0.890, MAE = 0.292, RMSE = 0.355, MSE = 0.126, R2 = 0.843). In predicting the outlet air temperature, MA-ELM (MRE=0.197, MAE=0.142, RMSE=0.169, MSE=0.289, R2=0.920) also produced the best performance, ELM (MRE=0.833, MAE=0.705, RMSE=0.793, MSE=0.629, R2=0.731) had the worst prediction performance, and GA-ELM (MRE=0.336, MAE=0.219, RMSE=0.273, MSE=0.074, R2=0.834) had better prediction performance than PSO-ELM (MRE=0.453, MAE=0.443, RMSE=0.503, MSE=0.253, R2=0.795). The above results show that the use of optimization algorithms can improve the prediction accuracy of the ELM model, and the accuracy improved by the optimization algorithms is ranked from high to low as mayfly algorithm, genetic algorithm, and particle swarm algorithm.

[0119] Therefore, the present invention explores the effects of four prediction models, namely, ELM, GA-ELM, PSO-ELM and MA-ELM. In temperature prediction, for the five evaluation indicators of MRE, MAE, RMSE, MSE and R2, the prediction effects of the models are ranked from best to worst as MA-ELM, GA-ELM, PSO-ELM and ELM. In the study of air conditioning status (air outlet speed and air outlet temperature) prediction, there is a conclusion consistent with the above, that is, the prediction effects of the models are also ranked from best to worst as MA-ELM, GA-ELM, PSO-ELM and ELM. The above research results show that: (1) Using the biological intelligence optimization algorithm to optimize the extreme learning machine model can improve the accuracy of the prediction. (2) Compared with GA and PSO, the MA algorithm has better global search capabilities. The above analysis shows that MA-ELM is an effective method for predicting temperature and air conditioning status.

[0120] Table 4. Evaluation of the prediction results of air velocity and air volume.

[0121]

[0122] The application of machine learning algorithms avoids the establishment of complex mathematical models. It does not require understanding the physical relationship between various variables, and can predict temperature only through data-driven methods. Among machine learning algorithms, ELM is widely used in the prediction field due to its fast learning rate and strong generalization ability. However, the parameters (weights and biases) of the extreme learning machine have a great influence on the prediction results, so how to optimize the parameters has become a key issue.

[0123] The present invention uses ELM to predict the temperature of the refrigerated truck, and proposes to use the mayfly algorithm to optimize the extreme learning machine parameters (weights and biases), and compares it with the standard ELM and the optimized GA-ELM and PSO-ELM models to evaluate the superiority of the proposed MA-ELM model. Furthermore, the obtained temperature prediction value is used as the input of the model, and the air conditioning status (air outlet temperature and air outlet speed) of the refrigerated truck is predicted to achieve timely control of the temperature of the refrigerated truck. The present invention realizes two aspects of prediction, one is the temperature of the refrigerated truck, and the other is the air conditioning status (air outlet temperature and air outlet speed) of the refrigerated truck. By accurately predicting the temperature of the refrigerated truck, the management level of cold chain logistics can be improved. Accurate temperature prediction can reduce food waste and ensure food quality. At the same time, accurate temperature control can reduce energy waste and save energy.

[0124] In summary, the invention is innovative in the following aspects: first, the temperature of different spatial positions in the refrigerated truck is taken into account; second, five factors affecting the temperature in the refrigerated truck are determined, including the outside temperature, air conditioning outlet speed, air conditioning outlet temperature, cargo volume, and whether the door is open; third, two aspects of prediction are achieved, one is the temperature in the car, and the other is the air conditioning outlet speed and outlet temperature. Fourth, the prediction is made using the Mayfly-Extreme Learning Machine algorithm model.

[0125] The above embodiments of the present invention are merely examples for illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes and modifications can be made based on the above description. It is impossible to list all the embodiments here. Any obvious changes or modifications derived from the technical solution of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for predicting temperature of cold chain logistics, characterized in that: Follow the steps below, 1) First, establish the extreme learning machine prediction model and optimize its input weight w and hidden layer bias b; 1.1) Collect temperature-related data in cold chain transport vehicles or storage equipment, and collect temperature values ​​at different locations in cold chain transport vehicles or storage equipment at the same time, filter out the highest and lowest temperatures in the temperature values, and take the temperature-related data and the highest and lowest temperatures at the corresponding time as a group of data; repeat the collection at a certain interval to obtain several groups of data; normalize all the data in the above several groups of data collected; 1.2) The extreme learning machine consists of an input layer, a hidden layer, and an output layer. The number of input neurons of the extreme learning machine is set to correspond to the number of temperature-related data, and the number of output neurons is two, corresponding to the highest temperature prediction value and the lowest temperature prediction value; the initial parameters of the extreme learning machine are randomly set, including the input weight w and the hidden layer bias b; 1.3) Optimize the input weight w and hidden layer bias b of the extreme learning machine using the mayfly algorithm, and set the input weight w and hidden layer bias b of the extreme learning machine as candidate solution individuals of the mayfly algorithm, that is, the input weight w and hidden layer bias b of the extreme learning machine are expressed by the position of each mayfly. At a certain moment, all mayfly positions represent feasible solutions, but there is only one relatively optimal solution; initialize the mayfly population at the same time, including setting the population size N, the maximum number of iterations, the attraction constants P1 and P2, the visibility constant α, the random walk coefficient w, the dance coefficient v and the crossover coefficient; 1.4) Input the temperature-related data in each group of data collected in step 1.1) into the extreme learning machine model under the input weight w corresponding to one of the mayflies and the hidden layer bias b for training, and obtain two outputs of the output layer, namely, the maximum temperature prediction value and the minimum temperature prediction value of each group of data, thereby obtaining the maximum temperature prediction value and the minimum temperature prediction value of all groups of data, and calculating the corresponding extreme learning machine model evaluation value according to the following formula to evaluate the prediction effect of extreme learning; Among them, X c Represents the actual maximum temperature value, X cz Refers to the actual maximum temperature value X c The corresponding maximum temperature prediction value, X d Represents the actual value of the lowest temperature, X dz Refers to the actual value of the lowest temperature X d The corresponding minimum temperature prediction value; h refers to the number of data groups; 1.5) changing the input weights w and hidden layer biases b represented by different mayflies and training according to step 1.4) to obtain the evaluation values ​​corresponding to the input weights w and hidden layer biases b represented by the different mayflies; until the input weights w and hidden layer biases b represented by all mayflies are rotated once, the evaluation values ​​corresponding to the input weights w and hidden layer biases b corresponding to all mayflies are obtained, and the input weights w and hidden layer biases b corresponding to the mayfly with the smallest evaluation value are selected as the input weights w and hidden layer biases b after this round of optimization; 1.6) According to the movement rule of the mayfly algorithm, one mayfly iteration is completed, that is, the position of the male mayfly, the position of the female mayfly and the mating of the mayfly are updated, so as to obtain all the mayflies and their positions after this iteration; thereby, the input weights w and the hidden layer bias b corresponding to the multiple mayflies after this iteration are obtained; 1.7) Repeat steps 1.4)-1.5) to calculate the minimum evaluation value after this iteration, compare the minimum evaluation value after this iteration with the minimum evaluation value before the iteration, and use the minimum evaluation value between the two as the new minimum evaluation value, and use the input weight w and hidden layer bias b of the mayfly representation corresponding to the new minimum evaluation value as the newly optimized input weight w and hidden layer bias b; 1.8) Determine whether the maximum number of iterations of the Mayfly algorithm has been reached. If so, output the newly optimized input weight w and hidden layer bias b obtained in step 1.7), which are the final optimized input weight w and hidden layer bias b; otherwise, jump to step 1.6) and repeat steps 1.6)-1.7) until the stop condition is reached; 2) Collecting actual data related to temperature prediction in cold chain transport vehicles or storage equipment that require temperature prediction, the number and type of the actual data are consistent with the number and type of temperature-related data in step 1.1); 3) Input the actual data collected into the extreme learning machine prediction model under the input weight w and hidden layer bias b finally optimized in step 1) to obtain two outputs of the output layer, which are the highest temperature prediction value and the lowest temperature prediction value corresponding to the actual data.

2. The cold chain logistics temperature prediction method according to claim 1, characterized in that: In step 1.1), there are five data related to temperature, namely, outside temperature, air conditioning air outlet speed, air conditioning air outlet temperature, cargo capacity and whether the door is open; the input layer corresponds to five inputs, corresponding to outside temperature, air conditioning air outlet speed, air conditioning air outlet temperature, cargo capacity and whether the door is open.

3. The cold chain logistics temperature prediction method according to claim 1, characterized in that: Step 1.1) When collecting temperature values ​​at different locations in a cold chain transport vehicle or storage equipment at the same time, collect temperature values ​​at five different locations. The five locations are the centers of the other five inner surfaces of the compartment except the inner surface where the door is located. A temperature sensor is set at each of the five locations to measure the temperature of the refrigerated truck.

4. A method for controlling the temperature of cold chain logistics, characterized in that: First, the method described in claim 1 is used to predict the temperature to be controlled. If the maximum temperature prediction value and the minimum temperature prediction value are both within the set range, the current air outlet temperature and air outlet speed of the air conditioner are maintained unchanged; If one of the maximum temperature prediction value and the minimum temperature prediction value exceeds the set range, the expected maximum temperature and minimum temperature as well as the outside temperature, cargo volume, and whether the door is open are input into the extreme learning machine prediction model through the input layer. The air conditioning outlet temperature prediction value and the air outlet speed prediction value are obtained through the two outputs of the output layer of the extreme learning machine prediction model. The air conditioning outlet temperature and the air outlet speed are adjusted to the outlet temperature prediction value and the outlet speed prediction value to achieve temperature control.

Images