Environmental Prediction in Cold Chain Transportation of Agricultural Products Based on K-Means++ and LSTM Neural Network

Abstract

Experiments have proven that traditional prediction research methods have limitations in practice. Proposing countermeasures for environmental changes is the key to optimal control of the cold chain environment and reducing the lag of control effects. In this paper, a cold chain transportation environment prediction method, combining k-means++ and a long short-term memory (LSTM) neural network, is proposed according to the characteristics of the cold chain transportation environment of agricultural products. The proposed prediction model can predict the trend of cold chain environment changes in the next ten minutes, which allows cold chain vehicle managers to issue control instructions to the environmental control equipment in advance. The fusion process for temperature and humidity data measured by multiple data sensors is performed with the k-means++ algorithm, and then the fused data are fed into an LSTM neural network for prediction based on time series. The prediction error of the prediction model proposed in this paper is very satisfactory, with a root-mean-square error (RMSE), mean absolute error (MAE), mean squared error (MSE), mean absolute percentage error (MAPE) and R-squared of 0.5707, 0.2484, 0.3258, 0.0312 and 0.9660, respectively, for temperature prediction, and with an RMSE, MAE, MSE, mean absolute percentage error and R-squared of 1.6015, 1.1770, 2.5648, 0.2736 and 0.9702, respectively, for humidity prediction. Finally, the LSTM neural network and back propagation (BP) neural network are compared in order to enhance the reliability of the results. In terms of the prediction effect of the temperature and humidity in cold chain vehicles transporting agricultural products, the proposed model has a higher prediction accuracy than that of existing models and can provide strategic support for the fine management and regulation of the cold chain transportation environment.

1. Introduction

The scale of cold chain logistics has been rapidly developing and growing in recent years due to the increase in people’s levels of consumption. Because of the strong time-dependency of agricultural products, the quality of the cold chain transportation environment is particularly important for the preservation of agricultural products. Among many environmental factors, the internal temperature and humidity of compartments are key factors affecting the quality of agricultural products. Abnormal changes in temperature and humidity in the cold chain transportation environment lead to the reproduction and proliferation of bacteria, which is very detrimental to the quality of agricultural products. Therefore, the study of efficient cold chain transport environments, with the aim of obtaining accurate temperature and humidity predictions, can help the cold chain department better control temperature and humidity, which is important for ensuring the safety and quality of agricultural products.

To date, there has been a great deal of research in the area of cold chain monitoring [1,2,3,4]. Experts have been studying temperature and humidity forecasting for years [5,6,7,8,9,10]. One study in the literature [11] proposed a new method for fruit freshness prediction based on multi-sensing technology and a machine learning algorithm. The results showed prediction accuracies of 90.87% (BP), 92.24% (RBF), 94.01% (SVM) and 91.31% (ELM). Maintaining and monitoring the quality of eggs during cold chain storage and transportation is a major concern due to the variation in external environmental factors, such as temperature or humidity [12]. The convolutional neural network (CNN) and long short-term memory (LSTM) algorithms have been stacked to make one deep learning model with hyperparameter optimization to increase HU value prediction performance. Experiments have proven that traditional prediction research methods have limitations in practice. Therefore, in order to grasp the change pattern of each variable in the refrigerated compartment, new artificial intelligence prediction algorithms are used to accurately predict temperature and humidity in the cold chain environment.

Commonly used forecasting methods include time series analysis and machine learning. In recent years, deep learning methods, represented by recurrent neural networks and convolutional neural networks in machine learning, have achieved good results in cold chain prediction. Taking BP as an example, Zhang et al. [13] proposed an improved bacterial chemotaxis optimization (IBCO), which was then integrated into the back propagation (BP) artificial neural network to develop an efficient forecasting model for the prediction of various stock indices. Experiments showed its performance to be superior to that of other methods in terms of learning ability and generalization. Qiao et al. [14] constructed a shelf-life prediction model based on a back propagation (BP) neural network using four kinds of environmental parameters, including temperature, humidity, oxygen and carbon dioxide, to perceive the quality of post-harvest strawberries, also building a cold chain transportation quality perception system (CCT-QPS) with the help of LabVIEW software for constantly monitoring the cold chain environment and commodity quality. The results showed that the proposed method could precisely predict the remaining shelf-life of post-harvest strawberries. A long short-term memory (LSTM) network is a special kind of RNN that is widely used in various fields [15,16,17,18,19,20,21,22]; among them, Duan [23] used LSTM neural networks for travel time prediction. The evaluation results showed that deep learning models considering sequence relations are promising in traffic series data prediction. Park et al. [24] used LSTM neural networks to predict the end-of-life of lithium-ion batteries to reduce the risk of battery failure; the results showed that LSTM is very effective in predicting battery lifetime. Nelson et al. [25] predicted future trends in stock prices based on LSTM neural network forecasts. The results that were obtained were promising, reaching an average of 55.9% accuracy when predicting whether the price of a particular stock is going to go up in the near future. Tsironi et al. [26] collected a large amount of data through shelf-life studies under isothermal and variable conditions, as well as in the actual cold chain. They developed predictive models for the effect of temperature on ready-to-eat green leafy salads. The LSTM prediction model used in this paper predicted the temperature and humidity data in the cold chain environment, and the error obtained was obviously smaller than those of the prediction results of other methods.

However, these studies have relied only on sensor output for analysis and processing. Moreover, there may have been lags in regulation, leading to decreases in the quality of produce. Proposing countermeasures for environmental changes is the key to optimal control of the cold chain environment and reducing the lag of control effects. The relevant data for this study were collected from distinct distributional spaces. The spatial data fusion of temperature and humidity series was performed by using the k-means++ algorithm. This action achieves data dimensionality compression and data redundancy reduction and improves the accuracy of the prediction. Since its inception, LSTM has been employed to address the pervasive long-term dependency issue in general recursive neural networks. Utilizing LSTM can effectively transmit and express information in long-term sequences without neglecting valuable information from the distant past and hence is deemed particularly efficacious for time series prediction. Furthermore, LSTM can resolve the vanishing gradient problem in RNNs. The temperature and humidity data in cold chain environments constitute a long-term sequence. This dataset is highly suitable for time series prediction using LSTM. This paper proposes a cold chain transport prediction model based on a combination of k-means++ and LSTM. It is based on the idea of the fusion of temperature and humidity data using the k-means++ algorithm. The LSTM prediction model is later used to predict the future temperature and humidity data. Experiments have proven that this can effectively improve prediction accuracy.

Section 2 describes the system framework and hardware design. Section 3 presents the details of the forecasting principles and methods. Section 4 gives the different experimental results, together with the analysis and discussion. Section 5 gives practical applications. Finally, our conclusions are given in Section 6.

2. System Framework and Hardware Design

First, we designed an environmental prediction platform for the cold chain transport of agricultural products based on IoT architecture. A schematic diagram of the system is shown in Figure 1. The Stm32f405 chip (32-bit processor) was used for the main control unit. We used it as the control unit of our system for the following reasons: the power consumption of Stm32f405 is lower, and the development of Stm32f405 is relatively easy, which can shorten the development time. In this study, we directly programmed the Stm32f405 chip by connecting it to a PC and utilizing the interfaces of programmer devices (In-System Programming, ISP). The ISP programmer that we utilized to program the chip is PEmicro CYCLONE-LC-UNIV. The MQ Telemetry Transport (MQTT) protocol is transplanted to the master control unit and controls the ESP8266 communication module [27,28,29]. The temperature and humidity concentrations in the cold chain vehicle were transmitted to AliCloud in real time. A WEB console was then further developed based on the IoT architecture to receive and process these real-time data from the master control unit.

The overall block diagram of the hardware system is shown in Figure 2a. It mainly consists of the main control unit Stm32f405 module, a DHT11 sensor module, an ADT7320 temperature sensor module, an ESP8266 communication module and a Li-ion battery power supply module. In this study, the DHT11 is only responsible for collecting the humidity in the cold chain vehicle, and, to ensure the response time, the proposed system uses the ADT7320 temperature module as the temperature acquisition source for the cold chain vehicle. The ESP8266 communication module is responsible for realizing the data interaction between the main control unit and the cloud server. The lithium battery power supply module is responsible for providing the driving voltage (3.3 V) required by the main control unit and the previous modules to ensure the normal operation of the hardware system in the cold chain vehicle. A picture of the employed hardware system to implement the idea is shown in Figure 2b.

3. Forecasting Principles and Methods

3.1. K-Means++ Data Fusion

The k-means++ algorithm is an improved clustering version of the k-means algorithm [30,31,32]. It is based on the principle of the k-means clustering algorithm. This algorithm is obtained by optimizing the initial clustering centers. For example, a complex spectral–spatial classification scheme for hyperspectral images was proposed and explored in the literature [33]. The k-means++ algorithm is used for image clustering. The proposed method provides improved precision and speed of hyperspectral data classification. Spatial data fusion of temperature and humidity series was performed by using the k-means++ algorithm. This action achieves data dimensionality compression and data redundancy reduction and improves the accuracy of the data. We used the fused temperature and humidity time series to represent the temperature and humidity state of the entire refrigerated compartment and then used it to construct the input time series samples for the prediction model. The process is shown in Figure 3.

The k-means++ data fusion algorithm has the following main steps:

(1) Randomly select one sample point on data set γ as the first initial clustering center, α.

(2) Count the shortest distance between each sample to the existing clustering center and express it as $D\left(x\right)$.

(3) Count the probability, $P\left(x\right)$, of each sample point being selected as the next clustering center and select the sample point corresponding to the maximum probability value as the next clustering center.

$$P\left(x\right)=\frac{D{\left(x\right)}^2}{{{\displaystyle \sum }}_{x\in \mathsf{\gamma }}D{\left(x\right)}^2}$$

(1)

(4) Repeat steps (2) and (3) until the kth cluster center is selected.

(5) Count the k clustering centers obtained and run the k-means clustering algorithm.

3.2. LSTM Neural Network

A long short-term memory neural network is a special type of recurrent neural network (RNN) [34,35,36,37,38]. To implement this mechanism, the maintenance of information is controlled by three gates: an input gate, a forget gate and an output gate. It is a better solution to the long-time dependence problem, the gradient disappearance problem and the explosion problem of RNNs [39,40,41,42].

Figure 4 shows the internal structure of an LSTM cell with a peephole. It is particularly popular for predicting time series, where fixed length windows of time series are generated and fed into an LSTM network.

The input sequence of the LSTM can be displayed as $\left\{x_1,x_2,\dots ,x_t\right\}$, where the subsequent $x_t$∈$R^K$ is the K-dimensional vector associated with the tth time interval. The LSTM structure comprises an input gate, $i_t$; an output gate, $O_t$; and a forget gate, $f_t$, $W_f, W_i, \mathrm{and} W_o$ are the forget gate weights, input gate weights and output gate weights, respectively. $W_c$ is the candidate vector weight. $b_c$ is the candidate vector bias. $C_t$ is the candidate vector at moment $t$. ${\tilde{C}}_t$ is the updated value of the candidate vector at time t. $h_t \mathrm{and} h_{t-1}$ are all outputs of the model at times t and t − 1. The descriptions [45] are as follows:

(1) Forget gate

$$f_t=\sigma \left(W_f\left[h_{t-1},x_t\right]+b_f\right)$$

(2)

(2) Input gate

$$i_t=\sigma \left(W_i\left[h_{t-1},x_t\right]+b_i\right)$$

(3)

(3) Output gate

$$O_t=\sigma \left(W_o\left[h_{t-1},x_t\right]+b_o\right)$$

(4)

(4) Unit

$${\tilde{C}}_t=\tanh \left(W_c\left[h_{t-1},x_t\right]+b_c\right)$$

(5)

$$C_t=f_t{C}_{t-1}+i_t{\tilde{C}}_t$$

(6)

(5) Final output

$$h_t=O_{t}tanh\left({C^{\prime }}_t\right)$$

(7)

3.3. Combination of K-Means++ Algorithm and LSTM Neural Network

This experiment constructed a k-means++-based LSTM prediction model. It is based on the idea of the fusion of temperature and humidity data using the k-means++ algorithm. The LSTM prediction model was later used to predict future temperature and humidity data. Experiments have proven that this can effectively improve prediction accuracy. The prediction process is shown in Figure 5.

The specific steps in the cold chain prediction are as follows:

(1) Remote collection of temperature and humidity data from each sensor in the cold vehicle.

(2) Cleaning of the data.

(3) Data fusion using the k-means++ algorithm.

(4) Dividing the fused data into a training set and a test set.

(5) Performing parameter tuning on the model to obtain the optimal parameters.

(6) Importing the test set data into the model for prediction and obtaining the prediction results.

3.4. Model Evaluation Indicators

In order to assess the prediction effect of the model, this paper chose four indicators to evaluate it.

(1) Root-mean-square error (RMSE). The order of magnitude is more intuitive, and the range is [0, +∞). When the true value is the same as the predicted value, it is equal to zero, which is a perfect model.

$$e_{RMSE}=\sqrt{\frac{1}{m}{\displaystyle \sum }_{i=1}^n{\left(x\left(i\right)-y\left(i\right)\right)}^2}$$

(8)

(2) Mean absolute error (MAE). When the real value is the same as the predicted value, the value is zero, which is a perfect model.

$$e_{MAE}=\frac{1}{m}{\displaystyle \sum }_{i=1}^n\left|y\left(i\right)-x\left(i\right)\right|$$

(9)

(3) Mean absolute percentage error (MAPE). The range is [0,+∞). A result of 0% indicates a perfect model, and a result greater than 100% indicates an inferior model.

$$e_{MAPE}=\frac{100\%}{m}{\displaystyle \sum }_{i=1}^n\left|\frac{x\left(i\right)-y\left(i\right)}{y\left(i\right)}\right|$$

(10)

(4) Mean squared error (MSE). As the error becomes larger, the value becomes larger.

$$e_{MSE}=\frac{1}{m}{\displaystyle \sum }_{i=1}^n{\left(x\left(i\right)-y\left(i\right)\right)}^2$$

(11)

(5) R-squared coefficient of determination (where R is the correlation coefficient, and the square of the correlation coefficient is the coefficient of determination). As the R-squared coefficient of determination comes closer to 1, the model’s prediction accuracy becomes higher.

4. Results and Discussion

4.1. Experimental Preparation

The main products transported in the cold chain were vegetables and other agricultural products. The transport route was from Guangzhou Baiyun to Guangzhou Haizhu. The dimensions of the box contained in the cold chain truck were 4075 × 2060 × 1860 (mm). In combination with the size of the cold chain truck and the location of the agricultural products, four hardware systems (i.e., four temperature collection sources and four humidity collection sources), designed for this study, were placed inside the boxes.

The experimental data collection period for this study was from 1 March 2022 to 6 March 2022. A total of approximately 1500 temperature and humidity data points were collected by the system. The system was set to collect temperature and humidity data every 10 min. Some of the raw data are shown in

Table 1. Selected raw collected data.

Node	Time	Temperature/°C	Relative Humidity/%
1	07:05:04	−20.2	79.3
4	07:15:02	−20.3	79.2
2	07:25:56	−20	79.8
3	07:35:23	−20.2	78.8
2	07:45:12	−20.1	70.8
…	…	…	…
4	15:22:28	−10	67.84
2	15:32:22	−19.3	70.01
3	15:42:34	−19.6	72.21
1	15:52:01	−19.9	74.41
2	15:02:46	−20.2	75.01

. To ensure the accuracy of the prediction model, this study used linear interpolation for data loss repair.

4.2. Forecast Results

4.2.1. K-Means++ Data Fusion Results

This paper used one time period from the raw data for analysis. The raw temperature data are illustrated in Figure 6a. The temperature changes were cyclical in nature, with a cycle time between 70 and 80 min. The relative humidity data are illustrated in Figure 6b. The temperature variation was also cyclical, with a cycle time between 65 and 75 min.

The k-means++ algorithm performs clustering based on the Euclidean distance of the data from the initial points. It is characterized by its high computational speed and efficiency. The algorithm abandons the k-means random selection of clustering centers. The fused data are shown in Figure 7a,b. This data fusion improved the accuracy to a certain extent and presented a more accurate periodicity. In this paper, we used the fused temperature and humidity time series representation. Thus, the input time series samples were constructed.

4.2.2. LSTM Neural Network Prediction Results

The data were processed using the k-means++ algorithm. The data were then normalized and divided into a training set and a test set according to a 7:3 division ratio. The data were fed into the LSTM model and an LSTM prediction model based on the k-means++ algorithm that was constructed. In this paper, the models were trained separately for temperature and humidity, and the predictions were more accurate.

The LSTM model for temperature was trained using the Adam algorithm optimizer. The number of iterations was set to 82, and the number of neurons was ten. Four experiments were conducted separately. The number of nodes in the hidden layer and the time steps are shown in

Table 2. Comparison of temperature prediction parameters for cold chain transportation.

Node	Number of Nodes in the Hidden Layer	Time Steps
1	5	5
2	25	1
3	35	2
4	45	3

.

Figure 8 represents the trend in the predicted and actual values of temperature for the four groups of the LSTM neural network. As can be seen from the figures, the temperature prediction curve fits well for the four sets of experiments. The predictions at the extremes were relatively less accurate overall, which is where errors mainly occurred.

The LSTM model for humidity was also trained using the Adam algorithm optimizer. The number of iterations was set to 95, and the number of neurons was 27. Four experiments were conducted separately. The number of nodes in the hidden layer and the time steps are shown in

Table 3. Comparison of humidity prediction parameters for cold chain transportation.

Node	Number of Nodes in the Hidden Layer	Time Steps
1	5	3
2	5	1
3	10	1
4	15	2

.

Figure 9 represents the trend of the predicted and actual values of humidity for the four groups of the LSTM neural network. The main errors occurred where the data set fluctuates considerably.

4.2.3. Evaluation of Forecast Results

It can be seen from Figure 8 that Test one was able to fit the approximate trend of the actual temperature profile better than the other three sets of tests. The remaining three trials did not have as good of a fit as Test one because of the number of nodes in the hidden layer and the time steps. As can be seen from the curves in Figure 9, Test one could better fit the actual humidity curve trend. Test four had too many nodes in the implied layer and appeared to be over-fitted. The curve appears to fluctuate, which reduces the prediction effect. The experiments show that the settings of the number of nodes in the hidden layer and the time steps hada relatively large degree of influence on the curve-fitting effect.

In order to reflect the error values of the four prediction models more directly, we made a prediction error comparison with different parameters.

Table 4. Prediction error of temperature and humidity in cold chain transportation environment with different parameters.

Node	Temperature/°C					Relative Humidity/%
	RMSE	MAE	MSE	MAPE	R-Squared	RMSE	MAE	MSE	MAPE	R-Squared
1	0.5707	0.4284	0.3258	0.0312	0.9724	1.6015	1.1770	2.5648	0.2736	0.9702
2	0.5770	0.4118	0.3329	0.0356	0.9652	2.4152	1.3956	5.8333	0.6237	0.9322
3	0.6189	0.4896	0.3831	0.0408	0.9600	2.4035	1.3885	5.7772	0.5886	0.9329
4	0.5826	0.4431	0.3394	0.0348	0.9646	1.7535	1.0510	3.0748	0.3547	0.9643
Average value	0.5873	0.4432	0.3453	0.0356	0.9655	2.0434	1.2530	4.3125	1.8406	0.9499

shows the RMSE, MAE, MSE and R-squared of the four prediction models. The first group of experiments achieved the best results for both the humidity and temperature prediction models, as listed in Table 4. The RMSE, MAE, MSE, MAPE and R-squared for temperature reached 0.5707, 0.4284, 0.3258, 0.0312 and 0.9660, respectively. The RMSE, MAE, MSE, MAPE and R-squared for relative humidity reached 1.6015, 1.1770, 2.5648, 0.2736 and 0.9702, respectively. This illustrates that the k-means++-based LSTM prediction model is more satisfactory for cold chain transportation environment prediction.

4.2.4. Comparison

In order to verify the proposed prediction models, a comparison experiment was set up in this paper. Using k-means++–BP as the reference model, the temperature and humidity prediction effects of the two different prediction models were examined and compared. The prediction results of the two groups of models are shown in Figure 10. The errors of the two groups of models are shown in

Table 5. Prediction error of the two groups of prediction models.

Model	Temperature/°C					Relative Humidity/%
	RMSE	MAE	MSE	MAPE	R-Squared	RMSE	MAE	MSE	MAPE	R-Squared
k-means++–BP	0.7681	0.5943	0.5901	0.0634	0.9384	2.3917	1.5816	5.7206	0.5543	0.9335
k-means++–LSTM	0.5707	0.4284	0.3258	0.0312	0.9660	1.6015	1.1770	2.5648	0.2736	0.9702

. It can be seen from Figure 10 that both sets of prediction models could predict the temperature and humidity in the cold chain transport carriage to different degrees. The results show that the prediction results of the two groups of models are in line with the trend of the actual data changes, but the prediction effects were different.

It can be concluded that the prediction results of the k-means++ and LSTM neural network are better than those of the k-means++ and BP neural network. The prediction accuracy of the k-means++ and LSTM neural network prediction model established in this experiment was relatively high. It could accurately predict the change pattern of temperature and humidity in the cold chain transport carriage in the next 10 min.

The results in Table 5 show that the RMSE, MAE, MSE and MAPE of the k-means++-LSTM temperature predictions were reduced by 25.69%, 27.91%, 44.78% and 50.19%, respectively, relative to the k-means++–BP predictions. The RMSE, MAE, MSE and MAPE for relative humidity were reduced by 33.03%, 25.58%, 55.16% and 50.64%, respectively.

As can be seen from Table 5, the k-means++–LSTM prediction model had less error volatility and better results with a single time series prediction compared to those of the k-means++–BP prediction model. The LSTM neural network could accurately predict the temperature and humidity in the cold chain environment 10 min in advance, which could provide a decision basis for the regulation of temperature and humidity in the cold chain transportation environment.

5. Practical Applications

The predicted results from this experiment were applied to the control of the temperature and humidity environment in a cold chain vehicle in practice. The predicted data from the present study were applied to actual cold chain transport in the same vehicle used for the experiment. Again, the cold chain truck was used to transport mainly agricultural products, such as vegetables, following the same driving route.

The actual data collected are shown in Figure 11. The refrigeration system and humidity regulation systems were adjusted in advance inside the cold chain vehicle with reference to the predictions of this paper. The aim of this experimental cold chain prediction was to achieve early command of the refrigeration system and the humidity regulation system. The results show that this application achieved the desired objective and provided good assurance on the quality of the produce.

6. Conclusions

For the characteristics of cold chain data, this paper proposes a new cold chain transport prediction method, which effectively improves the prediction effect. The results indicate that, in a cold chain transportation environment, the RMSE, MAE, MSE and R-squared of the temperature prediction by the k-means++-based LSTM prediction model are 0.5707, 0.4284, 0.3258 and 0.9660, respectively. The RMSE, MAE, MSE and R-squared of the relative humidity prediction are 1.6015, 1.1770, 2.5648 and 0.9702, respectively. The prediction effect is obvious and can accurately reflect the actual cold chain environment.

The cold chain temperature and humidity prediction model proposed in this paper accurately reflects the temperature and humidity trend in the cold chain transportation environment of agricultural products. The proposed prediction model can predict the trend of cold chain environment changes in the next ten minutes, which allows the cold chain vehicle managers to issue control instructions to the environmental control equipment in advance. This greatly alleviates the problem of lagging control effects.

In the future, we will test other algorithms, such as the Kalman filter with a state transition matrix taking into account the cyclicity (Fourier series), and we will study a novel prediction approach based on multi-model combination.