Tobacco transportation quantity prediction method based on one-dimensional convolutional neural network

Abstract

The invention discloses a tobacco shipping quantity prediction method based on a one-dimensional convolutional neural network, and belongs to the technical field of tobacco shipping quantity prediction, and the method comprises the steps: collecting the data of the tobacco shipping quantity, and the data comprises the shipping quantity, the delivery customer quantity, the delivery time and the delivery route; standardizing the acquired data of the tobacco shipping amount; a tobacco shipping quantity prediction model is constructed, the tobacco shipping quantity prediction model selects a one-dimensional convolutional neural network as a main structure to perform feature extraction, and a full connection layer is introduced to output a final result; the pre-processed training data set is adopted to train the constructed tobacco shipping quantity prediction model until the tobacco shipping quantity prediction model is trained; and inputting the pre-processed to-be-predicted data into the trained tobacco shipping quantity prediction model for prediction to obtain the tobacco shipping quantity. According to the method, accurate prediction of the freight volume of a new area can be realized, the transportation resource configuration can be optimized, the logistics management efficiency of an enterprise is improved, and a more reasonable production and distribution strategy is formulated.

Description

Technical Field

[0001] The present invention relates to the technical field of tobacco freight volume prediction, and in particular to a tobacco freight volume prediction method based on a one-dimensional convolutional neural network. Background Art

[0002] As one of the world's largest tobacco production and consumption markets, China's tobacco industry has been a vital component of domestic economic development for decades, maintaining its scale and market share. Under the current tobacco industry model, tobacco transportation volume planning relies primarily on companies' historical experience and market feedback, relying heavily on manual judgment and subjective decision-making. This traditional approach leads to information asymmetry, hindering the accuracy of transportation scheduling and making it difficult to achieve the desired efficiency and accuracy levels within the tobacco supply chain. This not only results in unsold or out-of-stock cigarettes for some brands, but can also lead to wasted resources and increased costs in supply chain management.

[0003] In recent years, the rapid development of artificial intelligence (AI), big data analytics, the Internet of Things (IoT), and smart hardware devices has provided new development opportunities for the tobacco industry. Advanced AI algorithms can intelligently predict future transportation volumes based on a variety of data sources, including historical transportation data, retailer order information, market trends, and environmental factors. This not only helps optimize tobacco companies' production plans and improve the accuracy of logistics scheduling, but also effectively reduces inventory backlogs and supply-demand imbalances. Furthermore, the introduction of intelligent logistics systems can enhance the automation level of tobacco transportation, for example, by optimizing delivery routes through intelligent scheduling algorithms, reducing transportation costs and carbon emissions. Combined with big data analysis, companies can respond more quickly to market demand, thereby improving the operational efficiency of the entire supply chain.

[0004] At present, most of the prediction research on tobacco focuses on the prediction of sales volume. The literature "Li Mingzhao, Li Yixu, Wang Jia. Tobacco yield prediction method based on gradient boosting regression tree model [J]. Yunnan Chemical Industry, 2023, 50(9): 109-111" established a tobacco yield prediction method based on gradient boosting regression tree (GBRT) model, using historical production data for trend analysis to optimize tobacco planting and production planning. The literature "Zhu Junjiang, He Xiangzhu, Wang Jianshu, et al. Cigarette sales volume prediction based on hybrid model [J]. Journal of Chinese Tobacco Society, 2016, 22(5): 120-125" combined wavelet transform, regression analysis and neural network algorithm to predict cigarette sales volume, and attempted to improve the prediction accuracy through multi-scale feature extraction method. The paper "Liu Lu, Ding Fuli, Sun Limin. Tobacco Sales Volume Forecasting Based on SVM [J]. Software Guide, 2016, 15(11): 134-137" proposes a sales volume forecasting method based on support vector machines (SVMs), which can improve the generalization ability of forecasts when the amount of data is limited. However, tobacco sales volume forecasting is mainly based on the purchase quantity of a fixed retailer over a period of time. Although these methods have improved the accuracy of sales forecasts to a certain extent, their limitation is that they only consider historical sales data and ignore other external factors that may affect sales.

[0005] Tobacco transportation volume forecasting differs from tobacco sales volume forecasting. Tobacco sales volume forecasting is a statistical analysis of tobacco sales data across all regions over a period of time. Different data sets have a certain temporal correlation, so tobacco sales volume forecasting can be considered a time series problem. Tobacco transportation data is a statistical analysis of tobacco transportation data across different regions over a period of time. Because transportation volume data is aggregated based on local conditions, transportation data forecasting is not time-sensitive. Furthermore, because tobacco transportation volume data is affected by distance, it exhibits strong nonlinearity. Furthermore, tobacco transportation data is the foundation of tobacco sales data, and only by ensuring sufficient transportation volume can the accuracy of sales data be guaranteed. However, most research focuses on tobacco sales data forecasting, while rarely considering the forecasting of transportation volume.

[0006] Currently, few studies have comprehensively considered the impact of multiple factors, such as delivery time, number of customers, and delivery distance, on tobacco shipment volume. Existing statistical methods typically rely on a single factor for forecasting. While these methods can capture regularity to a certain extent, they lack direct correlation between shipment volume forecasts across different regions, limiting their accuracy. Summary of the Invention

[0007] The technical problem to be solved by the present invention is: a tobacco freight volume prediction method based on a one-dimensional convolutional neural network can achieve accurate prediction of freight volume in new areas, optimize transportation resource allocation, improve the company's logistics management efficiency, and ultimately help companies formulate more reasonable production and distribution strategies to adapt to market dynamics and improve overall operational efficiency.

[0008] The technical solution adopted by the present invention is: a tobacco shipment volume prediction method based on a one-dimensional convolutional neural network, the method comprising the following steps:

[0009] Step 1: Collect tobacco shipment data, including shipment quantity, number of delivery customers, delivery time, and delivery distance;

[0010] Step 2: Data preprocessing: standardize the collected tobacco shipment data to make it conform to a distribution with an expected value of 1 and a variance of 0;

[0011] Step 3: Build a tobacco shipment volume prediction model. This model uses a one-dimensional convolutional neural network as the main structure for feature extraction and introduces a fully connected layer to output the final result.

[0012] Step 4: Using the preprocessed training data set to train the constructed tobacco freight volume prediction model until the tobacco freight volume prediction model is trained;

[0013] Step 5: Use the pre-processed data to be predicted to input the trained tobacco shipment volume prediction model to perform prediction and obtain the tobacco shipment volume.

[0014] Furthermore, the data normalization method in step 2 above is: assuming that the input sequence data is X T =[X1,X2,…,X T ], the standardization formula is as follows:

[0015]

[0016] Where, X T Indicates the value at the T serial number location, that is, X T The original data corresponding to the four dimensions of the number of goods shipped, the number of customers delivered, the delivery time, and the delivery distance, for example, X1 = |142, 14003, 2.93, 70|); μ T represents the mean of the input sequence (the number of goods shipped, the number of customers delivered, the delivery time, and the delivery distance); σ T Represents the variance of the input sequence (the number of goods shipped, the number of customers delivered, the delivery time, and the delivery distance); Represents the standardized input sequence (four values: quantity of goods shipped, number of customers delivered, delivery time, and delivery distance).

[0017] Furthermore, the method for constructing the tobacco shipment volume prediction model in step 3 above is as follows:

[0018] The one-dimensional convolutional neural network directly calculates the input one-dimensional data without transforming the original data. At the same time, it establishes the mapping relationship between different parameters through the convolution kernel. Its calculation formula is as follows:

[0019]

[0020] Where, χ T×i It represents the quantity of goods delivered at location T, the number of customers delivered, the delivery time and the delivery distance, and Z (T,f) represents the eigenvalue after calculation by the one-dimensional convolutional neural network; T represents the order of the input sequence; f represents different parameters; k represents the convolution kernel size, w i×1 Represents the convolution matrix corresponding to the convolution kernel; b represents the bias parameter; there are two layers of one-dimensional convolutional neural network. The second layer convolution calculation is the same as the calculation process of formula (4). Therefore, the characteristics after passing through the one-dimensional convolutional neural network layer are:

[0021]

[0022] In order to better handle the nonlinear relationship in the input sequence, a fully connected layer is used to output the final result in the final output. The calculation formula is as follows:

[0023]

[0024] Where Y represents the output result, W f×1 represents the weight matrix and b represents the bias parameter.

[0025] Furthermore, the above model uses back propagation to continuously update the weight parameters of the matrix, so that the loss function value becomes smaller and smaller, gradually making the model converge. The training loss function is:

[0026]

[0027] Where loss represents the value of the loss function, and Y represents the tobacco shipment volume; Represents the input of the model.

[0028] The beneficial effects of the present invention are as follows: compared with the prior art, the tobacco freight volume prediction method based on a one-dimensional convolutional neural network of the present invention combines key factors such as freight time, number of customers, and transportation distance, and constructs a tobacco freight volume prediction model based on machine learning technology. The model can not only achieve accurate prediction of freight volume in new areas, but also optimize the allocation of transportation resources, improve the logistics management efficiency of enterprises, and ultimately help enterprises formulate more reasonable production and distribution strategies to adapt to market dynamics and improve overall operational efficiency; choosing to standardize data can solve the problem that different data ranges between different parameters and parameters with a large scale range may dominate model training, resulting in reduced model performance; in the final output, a fully connected layer is used to output the final result, which can better handle the nonlinear relationship in the input sequence; the weight parameters of the matrix are continuously updated using back propagation, so that the loss function value becomes smaller and smaller, and the model gradually converges. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 For the prediction flow chart;

[0030] Figure 2 is the model convergence diagram;

[0031] Figure 3 This is the model prediction result graph. DETAILED DESCRIPTION

[0032] Example 1: Figures 1 to 3 As shown, a tobacco shipment volume prediction method based on a one-dimensional convolutional neural network includes the following steps:

[0033] Step 1: Collect tobacco shipment data, including shipment quantity, number of delivery customers, delivery time, and delivery distance;

[0034] Step 2: Data preprocessing: standardize the collected tobacco shipment data to make it conform to a distribution with an expected value of 1 and a variance of 0;

[0035] Since the data range varies between different parameters, parameters with large scale ranges may dominate model training and reduce model performance. Therefore, we choose to standardize the data to conform to a distribution with an expectation of 1 and a variance of 0.

[0036] The data standardization method is: Assume that the input sequence data is X T =[X1,X2,…,X T ], the standardization formula is as follows:

[0037]

[0038] Where, X T Indicates the value at location T (T represents a different delivery area or sample number, XT The original data corresponding to the four dimensions of the number of goods shipped, the number of customers delivered, the delivery time, and the delivery distance, for example, X1 = |142, 14003, 2.93, 70|); μ T represents the mean of the input sequence; σ T represents the variance of the input sequence; Represents the normalized input sequence;

[0039] Step 3: Build a tobacco shipment volume prediction model. Since the input data for tobacco shipment volume prediction does not include time relationships, the tobacco shipment volume prediction model uses a one-dimensional convolutional neural network (1DCNN) as the main structure for feature extraction and introduces a fully connected layer to output the final results.

[0040] The method for constructing the tobacco shipment volume prediction model is as follows:

[0041] The one-dimensional convolutional neural network directly calculates the input one-dimensional data and establishes the mapping relationship between different parameters through the convolution kernel. Its calculation formula is as follows:

[0042]

[0043] Where, χ T×i It represents the quantity of goods shipped at location T, the number of customers delivered, the delivery time, and the delivery distance. (T,f) represents the eigenvalue after calculation by the one-dimensional convolutional neural network; T represents the order of the input sequence; f represents different parameters; k represents the convolution kernel size, w i×1 Represents the convolution matrix corresponding to the convolution kernel; b represents the bias parameter; there are two layers of one-dimensional convolutional neural network, and the network structure is as follows:

[0044] 1. Input data: Each sample is a one-dimensional sequence of length T. Each position contains four features (number of shipments, number of customers, delivery time, and delivery distance). The input shape is (T, 4).

[0045] 2. First layer 1DCNN:

[0046] The convolution kernel size k = 3, the number of convolution kernels f = 64, the step size = 1, the activation function ReLU, and the output feature map shape are (T-2, 64).

[0047] 3. Second layer 1DCNN:

[0048] The convolution kernel size k = 3, the number of convolution kernels f = 32, the step size = 1, the activation function ReLU, and the output feature map shape are (T-4, 32).

[0049] 4. Fully connected layer: After flattening the second layer output, it is mapped to the final output (shipping volume) through the fully connected layer. In formula (6), W f×1 The dimension of must match the number of input features.

[0050] The calculation process of the second convolution layer is the same as that of formula (4). Therefore, the features after passing through the one-dimensional convolutional neural network layer are:

[0051]

[0052] In order to better handle the nonlinear relationship in the input sequence, a fully connected layer is used to output the final result in the final output. The calculation formula is as follows:

[0053]

[0054] Where Y represents the output result, W f×1 represents the weight matrix, b represents the bias parameter;

[0055] The model uses back propagation to continuously update the weight parameters of the matrix, so that the loss function value becomes smaller and smaller, and the model gradually converges. The loss function of training is:

[0056]

[0057] Where loss represents the value of the loss function, and Y represents the tobacco shipment volume; Represents the input of the model;

[0058] Step 4: Using the preprocessed training data set to train the constructed tobacco freight volume prediction model until the tobacco freight volume prediction model is trained;

[0059] Step 5: Use the pre-processed data to be predicted to input the trained tobacco shipment volume prediction model to perform prediction and obtain the tobacco shipment volume.

[0060] The present invention first standardizes the data. Since the data ranges of various parameters vary greatly, the data needs to be unified into a similar range to enable the model to converge more quickly. Secondly, the data is divided into a training set and a test set. The training data is then input into the model so that the model learns the features in the data. Finally, the trained model is used to predict freight volume.

[0061] In order to illustrate the effect of the present invention, the following simulation verification is performed:

[0062] 1 Experimental data

[0063] The data used in this experimental simulation was provided by a tobacco company in Guizhou Province. This data includes five parameters: delivery area, number of delivery customers, delivery time, delivery route distance, and number of shipments—a total of five parameters. Because the number of shipments is related to the number of delivery customers, delivery time, and delivery route distance, these four data sets were selected as input into the model. The data provided by the tobacco company in Guizhou Province contains 3,500 data points. To ensure sufficient data for training, 3,000 data points were selected as the training set and 500 data points as the test set. Table 1 shows a sample of the dataset used.

[0064] Table 1: Examples of some tobacco transportation data

[0065]

[0066] 2 Experimental Environment

[0067] In this experiment, the experimental environment and experimental parameters used are as follows: Windows 11 system, Intel(R) Core(TM) i5-10400 2.90GHz CPU, 16GB RAM, RTX3080 graphics card, and the PyTorch package used. The learning rate is 0.00008, the number of epochs is 100, the loss function is MSE, and the optimizer is Adam.

[0068] 3 Experimental indicators

[0069] In order to better evaluate the effectiveness of the model in tobacco transportation volume prediction, MAE and MSE were selected as evaluation indicators. MAE is an indicator obtained by averaging the absolute errors between the predicted value and the actual value. MSE is an indicator obtained by averaging the squared residuals between the predicted value and the actual observed value, and is more sensitive to certain errors that are larger. Smaller MAE and MSE values ​​represent smaller prediction errors of the model, indicating higher model performance. Larger values ​​represent larger prediction errors of the model, indicating lower model performance. The calculation formulas for MAE and MSE are as follows:

[0070]

[0071] Among them, n represents the length of the output sequence, y i represents the i-th value in the sequence, Represents the output corresponding to the i-th value in the sequence.

[0072] 4 Experimental results

[0073] Figure 2The figure shows the convergence of the model. As can be seen from the figure, the loss value decreases rapidly during the first 20 iterations, indicating that the model is able to effectively learn data features and quickly optimize parameters in the initial stage. However, as the number of iterations increases to around 100, the loss value stabilizes and further converges during subsequent training, indicating that the model has reached a relatively good state and avoided overfitting.

[0074] Because tobacco shipment data lacks time series features, we employed a 1DCNN for feature extraction and introduced fully connected layers to enhance the ability to model nonlinear relationships between data. Furthermore, to further validate the effectiveness of the model structure, we selected 1DCNN and fully connected layers as comparison methods. We maintained consistent hyperparameters across all methods, adjusting only the network structure to ensure comparability of the experimental results.

[0075] Table 2 shows the experimental results. It can be seen that the method proposed in the present invention performs better than the method using 1DCNN or the fully connected layer alone in terms of MAE and MSE indicators. Among them, the method of the present invention reaches 0.2533 in MAE indicator and 0.1079 in MSE indicator, showing high prediction accuracy. In contrast, the method using only 1DCNN has a MAE of 0.2690 and an MSE of 0.1160. This is because although 1DCNN can extract local features, it fails to effectively capture the nonlinear relationship between data. The method using only the fully connected layer has a MAE of 0.2834 and an MSE of 0.1311, which is relatively poor. This is because the fully connected layer can only model the nonlinear relationship of the data, but cannot effectively extract deep features.

[0076] Table 2 Experimental results

[0077]

[0078] Figure 3 The prediction results of the freight volume are shown. It can be observed from the figure that the method proposed by the present invention can accurately predict the tobacco freight volume in most areas. However, in some specific areas, there is a large difference between the prediction results of the model and the actual value. This is mainly because the freight volume in these areas is extremely high or low, making it difficult for the model to effectively learn its features, thereby affecting the prediction accuracy. The data in these special areas may be relatively scarce, resulting in the model being unable to fully capture its inherent patterns during the training process. Overall, the method of the present invention can still fit the freight volume data well and achieve relatively stable prediction results in most areas. This further verifies the effectiveness of the model, that is, the feature extraction method based on factors such as the number of retailers, freight time, and freight distance can more accurately portray the changing trend of tobacco freight volume.

[0079] Simulation Conclusion: This paper proposes a tobacco shipment volume forecasting method based on 1DCNN to address the difficulty of traditional statistical methods in accurately predicting retailer demand in new regions. This method incorporates factors such as the number of retailers, shipment time, and shipment distance to construct a deep learning model to improve the accuracy of shipment volume forecasts. Experimental validation in a specific area of ​​Guiyang demonstrates that this method has good forecasting results in some regions.

[0080] This study not only demonstrates the effectiveness of deep learning methods in tobacco shipment volume forecasting but also provides an intelligent transportation management solution for the tobacco industry, enabling optimized production planning and improved transportation efficiency. However, the model's predictions still exhibit some deviations in certain specific regions. Future research could further incorporate more parameters related to tobacco shipment volume and utilize richer data to achieve more accurate forecasts.

Claims

1. A tobacco shipment volume prediction method based on a one-dimensional convolutional neural network, characterized in that: The method comprises the following steps: Step 1: Collect tobacco shipment data, including shipment quantity, number of delivery customers, delivery time, and delivery distance. Step 2: Data preprocessing: standardize the collected tobacco shipment data to make it conform to a distribution with an expected value of 1 and a variance of 0; Step 3: Build a tobacco shipment volume prediction model. This model uses a one-dimensional convolutional neural network as the main structure for feature extraction and introduces a fully connected layer to output the final result. Step 4: Using the preprocessed training data set to train the constructed tobacco freight volume prediction model until the tobacco freight volume prediction model is trained; Step 5: Use the pre-processed data to be predicted to input the trained tobacco shipment volume prediction model to perform prediction and obtain the tobacco shipment volume.

2. The tobacco shipment volume prediction method based on a one-dimensional convolutional neural network according to claim 1, characterized in that: The data standardization method in step 2 is: Assume that the input sequence data is X T =[X1,X2,…,X T ], the standardization formula is as follows: Where, X T Indicates the value at the T serial number location, that is, X T The original data corresponding to the four dimensions of shipment quantity, number of delivery customers, delivery time, and delivery distance; μ T represents the mean of the input sequence; σ T represents the variance of the input sequence; Represents the normalized input sequence.

3. The tobacco shipment volume prediction method based on a one-dimensional convolutional neural network according to claim 2, characterized in that: The method for constructing the tobacco shipment volume prediction model in step 3 is: The one-dimensional convolutional neural network directly calculates the input one-dimensional data and establishes the mapping relationship between different parameters through the convolution kernel. Its calculation formula is as follows: Where, χ T×k Indicates the number of goods shipped at location T, the number of customers delivered, the delivery time and the delivery distance (e.g. 1×1 =142), Z (T,f) represents the eigenvalue after calculation by the one-dimensional convolutional neural network; T represents the order of the input sequence; f represents different parameters; k represents the convolution kernel size, w k×1 Represents the convolution matrix corresponding to the convolution kernel; b represents the bias parameter; there are two layers of one-dimensional convolutional neural network. The second layer convolution calculation is the same as the calculation process of formula (4). Therefore, the characteristics after passing through the one-dimensional convolutional neural network layer are: In the final output, the fully connected layer is used to output the final result, and the calculation formula is as follows: Where Y represents the output result, W f×1 represents the weight matrix and b represents the bias parameter.

4. The tobacco shipment volume prediction method based on a one-dimensional convolutional neural network according to claim 3, characterized in that: The model uses back propagation to continuously update the weight parameters of the matrix, and the training loss function is: Where loss represents the value of the loss function, and Y represents the tobacco shipment volume; Represents the input of the model.

Images