Transportation price prediction method of multi-combination neural network model
Through multi-combination neural network model, the problem of difficult to capture market uncertainty and price fluctuations in traditional methods is solved, and higher prediction accuracy and stability are achieved.
Patent Information
- Application Number
- CN202411243639.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-05
- Publication Date
- 2024-12-31
AI Technical Summary
Traditional commodity transportation price prediction methods are difficult to accurately capture the uncertainty of the market and the nonlinear and non-stationary characteristics of price fluctuations, resulting in insufficient accuracy and stability of the prediction results.
Using a multi-combination neural network model, the historical transportation price data is decomposed by empirical modal decomposition method, the inherent modal functions and residual signals are extracted, and a two-dimensional array is formed, and the convolutional neural network and long-term memory network model are used for Training to form a prediction model to improve prediction accuracy.
It improves the accuracy and stability of commodity transportation price predictions, allows more comprehensive consideration of multiple influencing factors and uncertainty of price fluctuations, and avoids the limitations of a single model.
Smart Images
Figure CN119228255A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of transportation price prediction, and in particular to a transportation price prediction method of a multi-combination neural network model. Background Art
[0002] The fluctuation of commodity transportation prices is often affected by a large number of factors, such as economic cycles, industry policies, market supply and demand, transportation costs, etc. Traditional commodity transportation price forecasting uses a single forecasting model, which will homogenize the information of different characteristics in the original freight rate time series, thereby affecting the accuracy of the forecast results, and has limited effect on the forecast of nonlinear and dynamically changing price trends. Summary of the invention
[0003] The purpose of the present invention is to provide a transportation price prediction method of a multi-combination neural network model to solve the above technical problems;
[0004] The technical problem solved by the present invention can be achieved by adopting the following technical solutions:
[0005] A method for predicting transportation prices based on a multi-combination neural network model, comprising:
[0006] Step S1, collecting historical transportation price data of bulk commodities;
[0007] Step S2, preprocessing the collected historical transportation price data;
[0008] Step S3, using an empirical mode decomposition method to perform multi-scale decomposition on the historical transportation price data, and expressing it as a number of intrinsic mode functions of different frequencies and a residual signal;
[0009] Step S4, forming a two-dimensional array of the intrinsic mode function and the residual signal, and dividing the two into a training data set and a test data set;
[0010] Step S5, using the training data set to train a neural network model to form a prediction model;
[0011] Step S6, inputting the test data set into the prediction model to obtain the transportation price prediction result.
[0012] Preferably, step S3 comprises,
[0013] Step S31, finding the local maximum and local minimum on the time series;
[0014] Step S32, constructing an upper envelope and a lower envelope;
[0015] Step S33, calculating the average envelope;
[0016] Step S34, subtracting the average envelope from the original signal of the historical transportation price data to extract the intrinsic mode function;
[0017] Step S35, repeatedly decomposing the obtained residual signal, and continuously obtaining the next intrinsic mode function, until the fluctuation of the residual signal is less than a preset value or becomes monotonic;
[0018] Step S36, representing the original signal of the historical transportation price data as the sum of all the obtained intrinsic mode functions plus the finally obtained residual signal.
[0019] Preferably, in step S32, the upper envelope and the lower envelope are constructed by spline interpolation method.
[0020] Preferably, in step S33, the average envelope is the average value of the upper envelope and the lower envelope constructed in step S32, and the calculation formula is:
[0021]
[0022] Among them, E max (t) represents the upper envelope, E min (t) represents the lower envelope, and m(t) represents the average envelope.
[0023] Preferably, step S34 includes,
[0024] Step S341, subtracting the average envelope from the original signal of the historical transportation price data to obtain an intermediate signal;
[0025] Step S342, determine whether the intermediate signal satisfies the condition of zero average in the entire time range, and the number of local maximum and minimum points is equal or differs by at most one. If so, record the intermediate signal as an intrinsic mode function; if not, return to execute step S31 to re-screen the intermediate signal.
[0026] Preferably, step S35 includes,
[0027] Step S351, subtracting the current intrinsic mode function from the original signal of the historical transportation price data to obtain the residual signal;
[0028] Step S352, determine whether the current fluctuation of the residual signal is less than the preset value or whether the residual signal becomes monotonic. If it is less than the preset value or becomes monotonic, execute step S36; otherwise, bring in the time series of the residual signal and return to execute step S31 to extract the next inherent mode function.
[0029] Preferably, in step S5, the neural network model adopts a network structure that is a mixture of a convolutional neural network model and a long short-term memory network model. The training data set is input into the neural network model for training, and an unbiased estimate of the neural network model fitting is performed using a validation data set included in the training data set, and the hyperparameters of the neural network model are adjusted to form the prediction model.
[0030] Preferably, in step S5, the two-dimensional array is used as the input of the convolutional neural network model and the long short-term memory network model, the convolutional neural network model uses multiple convolutional neural networks in parallel to extract features of different scales or types, the features extracted by the convolutional neural network model and the long short-term memory network model are fused, the predicted values are synthesized through a fully connected layer, and the batch gradient descent method is used to train the weights and bias of the neural network model to obtain the optimal prediction model.
[0031] Preferably, in step S2, the method for preprocessing the historical transportation price data is to scale the interval of the historical transportation price data to between [0, 1] using maximum and minimum standardization.
[0032] Preferably, in step S1, the historical transportation price data of bulk commodities at least includes road freight rates of iron ore, steel and coal.
[0033] Beneficial effects of the present invention: Due to the adoption of the above technical scheme, the present invention proposes a multi-dimensional bulk commodity transportation price prediction method based on decomposition and integration, which improves the accuracy and stability of the prediction by comprehensively considering multiple influencing factors, using empirical mode decomposition to process the uncertainty of time series data, and using a neural network model to comprehensively consider the uncertainty of bulk commodity transportation price fluctuations. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 A schematic diagram of the steps of a method for predicting transportation prices using a multi-combination neural network model in an embodiment of the present invention;
[0035] Figure 2 It is a flow chart of a method for predicting transportation prices using a multi-combination neural network model in an embodiment of the present invention;
[0036] Figure 3 This is a schematic diagram of step S3 in an embodiment of the present invention;
[0037] Figure 4 This is a schematic diagram of step S34 in an embodiment of the present invention;
[0038] Figure 5 This is a schematic diagram of step S35 in an embodiment of the present invention;
[0039] Figure 6It is a comparison chart of prediction results in the embodiment of the present invention. DETAILED DESCRIPTION
[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0041] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.
[0042] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.
[0043] A method for transport price prediction based on multiple combined neural network models, such as Figure 1 , Figure 2 As shown, including,
[0044] Step S1, collecting historical transportation price data of bulk commodities;
[0045] Step S2, preprocessing the collected historical transportation price data;
[0046] Step S3, using the empirical mode decomposition method to perform multi-scale decomposition on the historical transportation price data, and expressing it as a number of intrinsic mode functions of different frequencies and a residual signal;
[0047] Step S4, forming a two-dimensional array of the intrinsic mode function and the residual signal, and dividing the two into a training data set and a test data set;
[0048] Step S5, using the training data set to train the neural network model to form a prediction model;
[0049] Step S6, input the test data set into the prediction model to obtain the transportation price prediction result.
[0050] Specifically, the purpose of the present invention is to use a multi-combination neural network prediction model based on the decomposition-prediction-integration concept to circumvent the limitations of a single model, thereby greatly improving the accuracy of commodity transportation price prediction.
[0051] Traditional bulk commodity transportation price forecasting methods are often difficult to accurately capture market uncertainties. The multidimensional bulk commodity transportation price forecasting method based on decomposition and integration proposed in the present invention improves the accuracy and stability of the forecast by comprehensively considering multiple influencing factors, using empirical mode decomposition (EMD) to process the uncertainty of time series data, and using a neural network model to comprehensively consider the uncertainty of bulk commodity transportation price fluctuations.
[0052] In a specific embodiment, the present invention includes the following steps:
[0053] Step 1: Collect historical commodity transportation price data.
[0054] Step 2: Clean and standardize the collected data.
[0055] Step 3: Use the empirical mode decomposition (EMD) method to perform multi-scale decomposition. The decomposed sequence data is divided into intrinsic mode functions (IMFs) of different frequencies and residual signals.
[0056] Step 4: For the decomposed intrinsic mode function and residual signal, a two-dimensional array is formed and divided into a training data set and a test data set.
[0057] Step 5, model training: input the training data set into the neural network model of the convolutional neural network model and the long short-term memory network model (CNN-LSTM) for training and learning, and use the validation data set included in the training data set to make an unbiased estimate of the neural network model fitting, so as to adjust the hyperparameters of the neural network model. The optimal hyperparameters of the neural network model are obtained through training and evaluation to form a prediction model.
[0058] In a preferred embodiment, Figure 3 As shown, step S3 includes,
[0059] Step S31, finding the local maximum and local minimum on the time series;
[0060] Step S32, constructing an upper envelope and a lower envelope;
[0061] Step S33, calculating the average envelope;
[0062] Step S34, subtracting the average envelope from the original signal of the historical transportation price data to extract the intrinsic mode function;
[0063] Step S35, repeatedly decomposing the obtained residual signal to continuously obtain the next intrinsic mode function until the fluctuation of the residual signal is less than a preset value or becomes monotonic;
[0064] Step S36, the original signal of the historical transportation price data is represented as the sum of all the obtained intrinsic mode functions plus the finally obtained residual signal.
[0065] Price time series data often has nonlinear and non-stationary characteristics, and traditional linear models are difficult to accurately fit this type of data. The present invention decomposes time series data through the EMD method, which can decompose nonlinear and non-stationary time series data into multiple intrinsic mode functions, so that the model can better adapt to the characteristics of time series data and improve the accuracy of prediction.
[0066] In a preferred embodiment, in step S32, the upper envelope and the lower envelope are constructed by spline interpolation.
[0067] In a preferred embodiment, in step S33, the average envelope is the average value of the upper envelope and the lower envelope constructed in step S32, and the calculation formula is:
[0068]
[0069] Among them, E max (t) represents the upper envelope, E min (t) represents the lower envelope, and m(t) represents the average envelope.
[0070] Specifically, in step S3, local maximum and local minimum of the time series are found, and the upper envelope E is constructed through the local maximum points using cubic spline interpolation. max (t), also use cubic spline interpolation to construct the lower envelope E through the local minimum point min (t), and then calculate the average envelope m(t) through the upper envelope and the lower envelope. The formula is:
[0071]
[0072] In a preferred embodiment, Figure 4 As shown, step S34 includes,
[0073] Step S341, subtracting the average envelope from the original signal of the historical transportation price data to obtain an intermediate signal;
[0074] Step S342, determine whether the intermediate signal satisfies the condition of zero average in the entire time range, and the number of local maximum and minimum points is equal or differs by at most one. If so, record the intermediate signal as an intrinsic mode function; if not, return to execute step S31 to re-screen the intermediate signal.
[0075] For the standardized data, find the local maximum value E of the time series max(t) and the local minimum E min (t). After finding the extreme points, these points are used to perform curve fitting through cubic spline interpolation to form the upper envelope and the lower envelope.
[0076] After obtaining the upper and lower envelopes, calculate the average value between them to obtain the average envelope m(t), which will serve as the local trend of the signal:
[0077]
[0078] Subtracting the mean envelope m(t) from the original signal x(t) yields the intermediate signal h(t):
[0079] h(t)=x(t)-m(t)
[0080] If the intermediate signal h(t) still has local maxima and minima, it means that this is not an intrinsic mode function (IMF) and the screening process needs to be continued until the IMF conditions are met. This process requires multiple iterations until the intermediate signal meets the definition of IMF.
[0081] The first intermediate signal that meets the conditions is the first IMF component of the original signal, denoted as c1(t). After that, this IMF component is subtracted from the original signal to obtain the residual signal r1(t) for the first time.
[0082] r1(t)=x(t)-c1(t)
[0083] Then repeat the above process for the residual signal to obtain other IMF components. Repeat the decomposition until the residual signal r n The fluctuation of (t) is less than a preset threshold or the residual signal becomes monotonous and no longer contains any valuable information. Finally, the EMD decomposition result is obtained.
[0084] The original signal x(t) can be expressed as the sum of all IMFs plus the final residual signal r n (t), that is:
[0085]
[0086] In a preferred embodiment, Figure 5 As shown, step S35 includes,
[0087] Step S351, subtracting the current intrinsic mode function from the original signal of the historical transportation price data to obtain a residual signal;
[0088] Step S352, determine whether the fluctuation of the current residual signal is less than the preset value or whether the residual signal becomes monotonic. If it is less than the preset value or becomes monotonic, execute step S36; otherwise, bring in the time series of the residual signal and return to execute step S31 to extract the next intrinsic mode function.
[0089] Specifically, in step S34, the method for extracting the first intrinsic mode function is:
[0090] Subtracting the average envelope m(t) from the original signal x(t) yields the intermediate signal h(t), expressed as:
[0091] h(t)=x(t)-m(t)
[0092] If the intermediate signal h(t) satisfies the IMF conditions at this time: zero average over the entire time range, and the number of local maximum and minimum points is equal or differs by at most one, then it is determined that the intermediate signal h(t) is the first intrinsic mode function, which is denoted as c1(t).
[0093] In step S35, the method for decomposing the residual signal is to subtract the first intrinsic mode function c1(t) from the original signal x(t) to obtain the residual signal r1(t) for the first time, which is expressed as:
[0094] r1(t)=x(t)-c1(t)
[0095] Then, steps S31 to S34 are repeated for the residual signal r1(t) obtained for the first time, and the next intrinsic mode function is continuously extracted. The intrinsic mode function extracted for the i-th time is recorded as c i (t);
[0096] Repeatedly decompose the residual signal until the residual signal r obtained for the nth time n (t) The fluctuation is smaller than a preset threshold or the residual signal becomes monotonic and no longer contains any valuable information.
[0097] The EMD decomposition result is obtained. The final original signal x(t) can be expressed as the sum of all intrinsic mode functions plus the final residual signal r n (t), expressed as:
[0098]
[0099] In a preferred embodiment, in step S5, the neural network model adopts a network structure that is a mixture of a convolutional neural network model and a long short-term memory network model, and the data set includes a training data set and a test data set. The training data set is input into the neural network model for training, and the neural network model fitting is unbiasedly estimated through the verification data set included in the training data set, and the hyperparameters of the neural network model are adjusted to form a prediction model.
[0100] In a preferred embodiment, in step S5, a two-dimensional array is used as the input of a convolutional neural network model and a long short-term memory network model, the convolutional neural network model uses multiple convolutional neural networks in parallel to extract features of different scales or types, the features extracted by the convolutional neural network model and the long short-term memory network model are fused, the predicted values are synthesized through a fully connected layer, and the batch gradient descent method is used to train the weights and bias of the neural network model to obtain the optimal prediction model.
[0101] Specifically, step S5 divides the data set into 80% as a training data set and 20% as a test data set. Step S5 also includes using a two-dimensional matrix of EMD component combination as the input of the convolutional neural network (CNN) and the long short-term memory network (LSTM). Multiple convolutional neural networks (CNNs) are used in parallel to extract features of different scales or types to improve the model's ability to understand the data and enhance its generalization performance. After the extraction is completed, the two parts of the features are fused, and the predicted values are synthesized through a fully connected layer, and the weights and biases of the neural network are trained using the batch gradient descent method.
[0102] The multi-CNN structure in the present invention is: 32 output channels, and the convolution kernels are (1, 20), (2, 20), ..., (32, 20) in sequence.
[0103] The single LSTM structure is: 2 hidden layers and 192 units.
[0104] After the extraction is completed, the two parts of the features are fused and the prediction values are synthesized through the fully connected layer.
[0105] The fully connected layer is processed with two layers of Liner, and the activation function ReLu is used after the CNN and Liner layers.
[0106] The evaluation indicators used in the evaluation include mean absolute error MAE, root mean square error RMSE and mean square error MSE, expressed as follows:
[0107]
[0108] in represents the i-th predicted value, y i represents the i-th actual value, and n represents the total number of samples.
[0109] Furthermore, the batch gradient descent method was used to train the weights and biases of the neural network. The batch size was 30, the learning rate was 0.0001, the dropout rate was 0.6, and the Adam optimizer was selected as the optimizer. The training was performed for 300 to 500 epochs.
[0110] The latest data to be predicted is predicted using the prediction model, and finally the prediction result is obtained.
[0111] The present invention solves the problem that the forecast of bulk commodity transportation prices faces great challenges due to the influence of complex factors such as uncertainty, nonlinearity, non-stationarity, irregularity and high noise. Traditional bulk commodity transportation price forecasting methods often fail to accurately capture the uncertainty of the market. The multi-dimensional bulk commodity transportation price forecasting method based on decomposition integration proposed by the present invention can more comprehensively consider the uncertainty of bulk commodity transportation price fluctuations and improve the accuracy and stability of forecasting by comprehensively considering multiple influencing factors, using empirical mode decomposition (EMD) to process the uncertainty of time series data, using convolutional neural network (CNN) to directly learn the internal representation of data to avoid the limitation of the human subjective level of manual feature extraction, and using long short-term memory network (LSTM) to more accurately capture the long-term dependency in time series data.
[0112] Price time series data often has nonlinear and non-stationary characteristics, and traditional linear models are difficult to accurately fit this type of data. By decomposing time series data using the EMD method, nonlinear and non-stationary time series data can be decomposed into multiple intrinsic mode functions, so that the model can better adapt to the characteristics of time series data and improve the accuracy of prediction. The traditional method of manually extracting features is limited by personal cognition and cannot fundamentally explore the intrinsic characteristics of time series data. The two-dimensional matrix of the EMD component combination is used as the input of the CNN and LSTM parts to extract the spatiotemporal characteristics of the EMD components respectively. After the extraction is completed, the two parts of the features are fused and the predicted values are synthesized through the fully connected layer. The use of parallel structure enables this strategy structure to take into account more data feature details, thereby having better adaptability, more accurate prediction results, and improving the accuracy of transportation cost control.
[0113] In a preferred embodiment, in step S2, the preprocessing method of the historical transportation price data is to scale the interval of the historical transportation price data to between [0, 1] using maximum and minimum standardization.
[0114] Specifically, step S2 uses Min-Max Normalization to scale the data to between [0, 1], and the formula is:
[0115]
[0116] Where min{x} represents the minimum value, max{x} represents the maximum value, and x norm Represents the scaled value.
[0117] Due to different standards, the features contained in the indicators such as the number of waybills and freight rates have different dimensions and dimensional units. If they are directly regressed, the accuracy and fit of the regression results will be affected, and the efficiency of the model operation will also be affected. Due to the differences in the units of different indicators, it is difficult to achieve the expected results if compared directly. Therefore, it is necessary to normalize the collected basic data sources and scale the data to a similar scale to eliminate the dimensional effects between different features.
[0118] In a preferred embodiment, in step S1, the historical transportation price data of bulk commodities at least includes road freight rates of iron ore, steel and coal.
[0119] Specifically, the transportation price prediction method of the multi-combination neural network model of the present invention is integrated into a bulk commodity service platform, and the domestic bulk commodities on the platform include highway freight rates of iron ore, steel, coal, etc.
[0120] In a specific embodiment, please refer to Figure 2 As shown, the system for implementing the transportation price prediction method of the multi-combination neural network model of the present invention mainly includes a mode decomposition module 1 and a training prediction module 2. The mode decomposition module 1 is used to perform multi-scale decomposition using the empirical mode decomposition method, and the decomposed sequence data is divided into intrinsic mode functions (IMFs) of different frequencies and residual signals; the training prediction module 2 is used to train the neural network prediction model and output the prediction result through the prediction model.
[0121] An embodiment of the method for predicting carbon futures prices in this embodiment includes the following steps:
[0122] Step 1: Collect historical commodity transportation price data.
[0123] In this embodiment, the premise of predicting the commodity transportation price by setting up a prediction model is to collect the basic data set. In order to verify the effectiveness of the model, the commodity transportation freight rate data of various transportation routes in China are obtained through the Wubo bulk transportation platform, and Shandong Rizhao to Shanxi Yuncheng is selected as a case sample, including different dimensions and dimensional units such as the waybill quantity and freight rate of iron ore, steel and coal.
[0124] Step 2: Clean and standardize the collected data.
[0125] Due to different standards, the features contained in the indicators such as the number of waybills and freight rates have different dimensions and dimensional units. If they are directly regressed, the accuracy and fit of the regression results will be affected, and the efficiency of the model operation will also be affected. Due to the differences in the units of different indicators, it is difficult to achieve the expected results if compared directly. Therefore, it is necessary to normalize the collected basic data sources and scale the data to a similar scale to eliminate the dimensional effects between different features.
[0126] This embodiment uses the Min-Max standardization method to preprocess the data to obtain x norm , the sample area can be locked to [0, 1]. The specific expression is:
[0127]
[0128] Step 3: Use the empirical mode decomposition (EMD) method to perform multi-scale decomposition. The decomposed sequence data is divided into intrinsic mode functions (IMFs) of different frequencies and residual signals.
[0129] For the standardized data, find the local maximum value E of the time series max (t) and the local minimum E min (t). After finding the extreme points, these points are used to perform curve fitting through cubic spline interpolation to form the upper envelope and the lower envelope.
[0130] After obtaining the upper and lower envelopes, calculate the average value between them to obtain the average envelope m(t), which will serve as the local trend of the signal:
[0131]
[0132] Subtracting the mean envelope m(t) from the original signal x(t) yields the intermediate signal h(t):
[0133] h(t)=x(t)-m(t)
[0134] If the intermediate signal h(t) still has local maxima and minima, it means that this is not an intrinsic mode function (IMF) and the screening process needs to be continued until the IMF conditions are met. This process requires multiple iterations until the intermediate signal meets the definition of IMF.
[0135] The first intermediate signal that meets the conditions is the first IMF component of the original signal, denoted as c1(t). After that, this IMF component is subtracted from the original signal to obtain the residual signal r1(t) for the first time.
[0136] r1(t)=x(t)-c1(t)
[0137] Then repeat the above process for the residual signal to obtain other IMF components. Repeat the decomposition until the residual signal r n The fluctuation of (t) is less than a preset threshold or the residual signal becomes monotonous and no longer contains any valuable information. Finally, the EMD decomposition result is obtained.
[0138] The original signal x(t) can be expressed as the sum of all IMFs plus the final residual signal r n (t), that is:
[0139]
[0140] Step 4: For the decomposed IMF and residual model, a two-dimensional array is formed and divided into a training data set and a test data set.
[0141] The holdout method is used to randomly divide the data set into 80% training data set and 20% test data set according to a certain ratio in this embodiment, so as to ensure the consistency of data distribution during the division process and avoid introducing additional deviations.
[0142] Step 5: Model training. The training data set is input into the neural network model of the convolutional neural network model and the long short-term memory network model (CNN-LSTM) for training and learning, and the CNN-LSTM neural network model fitting is unbiasedly estimated through the validation data set to adjust the model hyperparameters. The optimal hyperparameters of the CNN-LSTM neural network are obtained through training and evaluation to form a prediction model.
[0143] The two-dimensional matrix of EMD component combination is used as the input of CNN and LSTM. Multiple convolutional neural networks (CNN) are used in parallel to extract features of different scales or types to improve the model's ability to understand the data and enhance its generalization performance.
[0144] In this embodiment, the multi-CNN structure is as follows: 32 output channels, and the convolution kernels are (1, 20), (2, 20), ..., (32, 20) in sequence.
[0145] The single LSTM structure is: 2 hidden layers and 192 units.
[0146] After the extraction is completed, the two parts of the features are fused and the prediction values are synthesized through the fully connected layer.
[0147] The fully connected layer is processed with two layers of Liner, and the activation function ReLu is used after the CNN and Liner layers.
[0148] The evaluation indicators used in the evaluation include mean absolute error MAE, root mean square error RMSE and mean square error MSE, expressed as follows:
[0149]
[0150] in represents the i-th predicted value, y i represents the i-th actual value, and n represents the total number of samples.
[0151] Furthermore, the batch gradient descent method was used to train the weights and biases of the neural network. The batch size was 30, the learning rate was 0.0001, the dropout rate was 0.6, and the Adam optimizer was selected as the optimizer. The training was performed for 300 to 500 epochs.
[0152] Step 6: Input the data to be predicted, use the prediction model to predict, and finally get the prediction result. The prediction result is as follows: Figure 6 As shown, it can be clearly observed that the predicted value output by the prediction model of the present invention is close to the actual value, and the prediction result is accurate and reliable.
[0153] The above description is only a preferred embodiment of the present invention, and does not limit the implementation mode and protection scope of the present invention. For those skilled in the art, it should be aware that all solutions obtained by equivalent substitutions and obvious changes made using the description and illustrations of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for predicting transportation prices using a multi-combination neural network model, characterized in that: include, Step S1, collecting historical transportation price data of bulk commodities; Step S2, preprocessing the collected historical transportation price data; Step S3, using an empirical mode decomposition method to perform multi-scale decomposition on the historical transportation price data, and expressing it as a number of intrinsic mode functions of different frequencies and a residual signal; Step S4, forming a two-dimensional array of the intrinsic mode function and the residual signal, and dividing the two into a training data set and a test data set; Step S5, using the training data set to train a neural network model to form a prediction model; Step S6, inputting the test data set into the prediction model to obtain the transportation price prediction result.
2. The method for predicting transportation prices using a multi-combination neural network model according to claim 1, characterized in that: Step S3 comprises, Step S31, finding the local maximum and local minimum on the time series; Step S32, constructing an upper envelope and a lower envelope; Step S33, calculating the average envelope; Step S34, subtracting the average envelope from the original signal of the historical transportation price data to extract the intrinsic mode function; Step S35, repeatedly decomposing the obtained residual signal, and continuously obtaining the next intrinsic mode function, until the fluctuation of the residual signal is less than a preset value or becomes monotonic; Step S36, representing the original signal of the historical transportation price data as the sum of all the obtained intrinsic mode functions plus the finally obtained residual signal.
3. The method for predicting transportation prices using a multi-combination neural network model according to claim 2, characterized in that: In step S32, the upper envelope and the lower envelope are constructed by spline interpolation method.
4. The method for predicting transportation prices using a multi-combination neural network model according to claim 2, characterized in that: In step S33, the average envelope is the average value of the upper envelope and the lower envelope constructed in step S32, and the calculation formula is: Among them, E max (t) represents the upper envelope, E min (t) represents the lower envelope, and m(t) represents the average envelope.
5. The method for predicting transportation prices using a multi-combination neural network model according to claim 2, characterized in that: Step S34 includes, Step S341, subtracting the average envelope from the original signal of the historical transportation price data to obtain an intermediate signal; Step S342, determine whether the intermediate signal satisfies the condition of zero average in the entire time range, and the number of local maximum and minimum points is equal or differs by at most one. If so, record the intermediate signal as an intrinsic mode function; if not, return to execute step S31 to re-screen the intermediate signal.
6. The method for predicting transportation prices using a multi-combination neural network model according to claim 2, characterized in that: Step S35 includes, Step S351, subtracting the current intrinsic mode function from the original signal of the historical transportation price data to obtain the residual signal; Step S352, determine whether the current fluctuation of the residual signal is less than the preset value or whether the residual signal becomes monotonic. If it is less than the preset value or becomes monotonic, execute step S36; otherwise, bring in the time series of the residual signal and return to execute step S31 to extract the next inherent mode function.
7. The method for predicting transportation prices using a multi-combination neural network model according to claim 2, characterized in that: In step S5, the neural network model adopts a network structure that is a mixture of a convolutional neural network model and a long short-term memory network model. The training data set is input into the neural network model for training, and an unbiased estimate of the neural network model fitting is performed using the verification data set included in the training data set. The hyperparameters of the neural network model are adjusted to form the prediction model.
8. The method for predicting transportation prices using a multi-combination neural network model according to claim 6, characterized in that: In step S5, the two-dimensional array is used as the input of the convolutional neural network model and the long short-term memory network model. The convolutional neural network model uses multiple convolutional neural networks in parallel to extract features of different scales or types, and the features extracted by the convolutional neural network model and the long short-term memory network model are fused. The predicted values are synthesized through a fully connected layer, and the weights and biases of the neural network model are trained using the batch gradient descent method to obtain the optimal prediction model.
9. The method for predicting transportation prices using a multi-combination neural network model according to claim 1, characterized in that: In step S2, the preprocessing method of the historical transportation price data is to scale the interval of the historical transportation price data to between [0, 1] using maximum and minimum standardization.
10. The method for predicting transportation prices using a multi-combination neural network model according to claim 1, characterized in that: In step S1, the historical transportation price data of bulk commodities at least includes road freight rates of iron ore, steel and coal.