Coke demand prediction method based on multi-feature fusion deep neural network

By using a multi-feature fusion deep neural network, the shortcomings of data integration and feature extraction in coke demand forecasting are addressed, achieving higher accuracy and faster adaptability in forecasting.

CN118940986APending Publication Date: 2024-11-12HEFEI UNIV OF TECH +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410621434.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-05-20
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

Existing methods for forecasting coke demand lack a deep understanding of market dynamics, have limited model feature extraction capabilities, and struggle to effectively integrate multiple data sources and process real-time unstructured information, resulting in insufficient forecast accuracy and reliability.

Method used

A multi-feature fusion deep neural network is adopted. By preprocessing numerical and textual data, combining Transformer, GRU, BiGRU and BiLSTM models for feature extraction, and using self-attention mechanism and fully connected layer for feature fusion and prediction, an efficient coke demand prediction model is constructed.

Benefits of technology

It enhances the ability to comprehensively analyze data, improves the accuracy and comprehensiveness of feature extraction, and can better capture the nonlinear correlation and long-term dependence between demand features, thereby improving the accuracy of prediction and the ability to adapt to market changes.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118940986A_ABST
    Figure CN118940986A_ABST
Patent Text Reader

Abstract

The invention discloses a coke demand prediction method based on a multi-feature fusion deep neural network. The method comprises the following steps: 1, preprocessing a numerical data set of coke demands; 2, preprocessing the text type data set of the coke demand; 3, carrying out the pre-extraction of the features through a Transform; 4, carrying out feature extraction by using the GRU, the BiGRU and the BiLSTM; and 5, carrying out demand prediction after fusing the extracted features. According to the method, accurate prediction of the coke demand can be realized, powerful decision support is provided for production and purchase planning of purchasers, and effective use of production resources is improved to the maximum extent, so that an enterprise is helped to reduce the cost and improve the efficiency at the same time.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of computer integrated forecasting, and specifically is a coke demand forecasting method based on multi-feature fusion deep neural network. Background Art

[0002] As a key material in the blast furnace ironmaking process, coke not only acts as a reducing agent but also is the main source of energy. Accurately predicting coke demand has a decisive impact on ensuring the cost-effectiveness of smart coking plants and their environmental sustainability. The accuracy of demand forecasting is directly related to whether producers and consumers can effectively plan production, inventory, and procurement strategies to reduce unnecessary costs and improve the overall efficiency and responsiveness of the supply chain. Against the backdrop of growing global environmental awareness and sustainable development requirements, optimizing coke demand forecasting models can help reduce carbon emissions and promote environmentally friendly production. Despite this, research on coke demand forecasting is still insufficient and faces many challenges. The reasons for these challenges are multifaceted. First, the professionalism and scale of the coke market are relatively small, and researchers and companies have limited attention. Second, the market is highly dependent on the economic cycle, and demand is affected by a variety of economic and non-economic factors, which increases the complexity of forecasting. In addition, the availability of market data and data quality issues also limit the development of forecasting methods.

[0003] In terms of forecasting methods, traditional statistical and econometric methods, such as the autoregressive integrated moving average model (ARIMA), the least absolute shrinkage and selection operator (Lasso), and the generalized autoregressive conditional heteroskedasticity (GARCH), although they have certain advantages in simplicity, often lack the ability to adapt to the complexity of market dynamics, especially in dealing with nonlinear patterns and capturing market sentiment. These methods are also not powerful enough in the face of multi-dimensional and unstructured features of data. At the same time, artificial intelligence methods, such as random forests, Bayesian neural networks, and extreme gradient boosting (XGBoost), although they have potential in processing large data sets and revealing market dynamics, are still limited by the high requirements for data comprehensiveness and effectiveness, as well as the limitations of these models themselves in feature extraction. The emergence of deep learning technology has brought new opportunities in this field. Long short-term memory networks (LSTM), gated recurrent units (GRU), and more advanced models such as bidirectional gated recurrent units (BiGRU) and bidirectional long short-term memory networks (BiLSTM) improve the accuracy of forecasts by automatically learning complex patterns in data. However, even so, deep learning models may not be able to fully utilize all relevant data in specific situations, and the accuracy and reliability of their prediction results still have room for improvement.

[0004] In general, the shortcomings of coke demand forecasting research are the lack of in-depth understanding of market dynamics, limited feature extraction capabilities of the model, and insufficient processing of real-time data and unstructured information. Current demand forecasting methods lack the focus on improving the dynamic adaptability of the model, integrating more types of data sources, and developing more advanced algorithms to integrate this information. Summary of the invention

[0005] In order to solve the shortcomings of the above-mentioned prior art, the present invention proposes a coke demand forecasting method based on a multi-feature fusion deep neural network, in order to improve the comprehensive analysis capability of data, optimize the efficiency of feature extraction, and deeply mine feature information from a multi-dimensional perspective to achieve high-precision prediction of coke demand.

[0006] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme:

[0007] The coke demand forecasting method based on multi-feature fusion deep neural network of the present invention is characterized in that it is carried out according to the following steps:

[0008] Step 1: Preprocess the numerical indicator set P of coke demand to obtain the modal feature set IMF;

[0009] Step 2: Preprocess the text data set of coke demand to obtain the total word vector set X”;

[0010] Step 3: After concatenating the numerical indicator set P, the modal feature set IMF and the total word vector set X” after weight assignment, the coke demand input data set R = {P, IMF, X”} is obtained;

[0011] Step 4: Construct a multi-feature fusion deep neural network consisting of a Transformer network, a feature extraction model, and a fully connected layer;

[0012] Step 4.1, the Transformer network performs feature fusion pre-extraction processing on R to obtain a numerical-modality-text fusion feature R';

[0013] Step 4.2: Build a feature extraction model consisting of a GRU model, a BiGRU model, and a BiLSTM model and perform feature extraction:

[0014] Step 4.2.1: Extract features of R' based on the GRU model to obtain the dynamic characteristics of coke demand in, represents the sth dynamic characteristic of coke demand, s∈[1,S]; S represents the total number of dynamic characteristics of coke demand;

[0015] Step 4.2.2: Extract features of R' based on the BiGRU model to obtain the panoramic features of coke demand in, represents the panoramic characteristics of the s-th coke demand;

[0016] Step 4.2.3: Extract features of R' based on the BiLSTM model to obtain the deep sequence features of coke demand in, represents the sth coke demand depth sequence characteristics;

[0017] Step 4.3: Y GRU , Y BiGRU and Y BiLSTM After splicing, the data is input into the fully connected layer for demand forecasting to obtain the predicted coke demand Y;

[0018] Step 5. Calculate the mean square error based on the prediction result Y and use it as the objective function J, so as to train the multi-feature fusion deep neural network using back propagation and gradient descent method to minimize the objective function J. When the number of iterations reaches the maximum number of iterations, stop training to obtain the trained coke demand prediction model for the prediction of coke demand.

[0019] The coke demand forecasting method based on multi-feature fusion deep neural network described in the present invention is also characterized in that step 1 is performed as follows:

[0020] Step 1.1: Construct a numerical index set P of coke demand = {P1, P2, ..., P n ,…,P N}, where P n represents the nth index, n∈[1,N], N represents the total number of indexes, and P n ={p n,1 ,p n,2 ,…,p n,m ,…,p n,M}, where p n,m Represents the nth index P n The characteristic parameter of the mth time, m∈[1,M], M represents the total number of time;

[0021] Step 1.2, let the current number of empirical mode decomposition be k, and initialize k = 1;

[0022] Step 1.3: Perform Pearson correlation analysis on each indicator in the coke demand indicator set P with other indicators to find the indicator P with the highest correlation. u , so that in P u The k-1th amplitude is ε when the i-thk-1 Gaussian white noise v i After that, we get the i-th characteristic signal i∈[1,I]; I represents the total number of joins;

[0023] Step 1.4: Perform empirical mode decomposition to obtain the kth eigenmode component after empirical mode decomposition Then, the kth intrinsic mode component IMF is obtained using formula (2): k , and use formula (3) to get the kth residual r k ;

[0024]

[0025] r k =P u -IMF k (3)

[0026] Step 1.5: At the kth residual r k The kth amplitude is ε when the i-th addition is added k Gaussian white noise v i , get the i-th new characteristic signal r1+ε1v i , and then use formula (4) to get the k+1th intrinsic mode component IMF k+1 ;

[0027]

[0028] In formula (4), E represents the empirical mode decomposition operation;

[0029] Step 1.6: Calculate the k+1th residual r using formula (5) k+1 ;

[0030] r k+1 =r k -IMF k+1 (5)

[0031] Step 1.7, after copying k+1 to k, return to step 1.5 and execute sequentially until the remaining residual is a monotonic function, thereby obtaining the modal feature set IMF = {IMF1, IMF2, ..., IMF k ,…,IMF K}, K represents the total number of modal features, and IMF k ={imf k,1 ,imf k,2 ,…,imf k,m ,…,imf k,M}, imf k,m IMF kThe mth time corresponds to the modal parameter.

[0032] Described step 2 is carried out as follows:

[0033] Step 2.1: Collect multi-source text sets of coke demand, including: investor comment text set T1 = {T 1,1 ,T 1,2 ,…,T 1,m ,…,T 1,M}、Research report text set T2={T 2,1 ,T 2,2 ,…,T 2,m ,…,T 2,M} and news text set T3 = {T 3,1 ,T 3,2 ,…,T 3,m ,…,T 3,M}, where T 1,m represents the investor comment text data at time m, and Indicates T 1,m The ath investor comment word in T, a∈[1,A], A represents T 1,m Total number of words in investor comments; 2,m represents the text data of the research report at time m, and Indicates T 2,m The bth research report word in T, b∈[1,B], B represents T 2,m Total number of words in the research report; 3,m represents the news text data of the mth date, and in, Indicates T 3,m The cth news word in T, c∈[1,C], C represents T 3,m Total number of news words;

[0034] Step 2.2: T 1,m 、T 2,m and T 3,m After preprocessing, D words are extracted respectively, and then the pre-trained model BERT is used to embed the D words respectively, and the corresponding word vector of the investor comment at the mth time is obtained. The word vector set of the research report at the mth time And the news word vector set at the mth time Thus, we get the investor comment word vector set X1 = {X 1,1 ,X 1,2,…,X 1,m ,…,X 1,M}、Research report word vector set X2={X 2,1 ,X 2,2 ,…,X 2,m ,…,X 2,M} and news word vector set X3 = {X 3,1 ,X 3,2 ,…,X 3,m ,…,X 3,M};in, represents the dth investor comment word vector, represents the d-th research report word vector, represents the d-th news word vector, where d∈[1,D], and D represents the length of the word vector;

[0035] Step 2.3: Transpose the investor comment word vector X1, the research report word vector X2, and the news word vector X3 and concatenate them to obtain the total word vector set Where T represents transpose;

[0036] Step 2.4: Use the fully connected layer to reduce the dimension and exchange information of the total word vector set X to obtain the word vector set X'={X1,X2,…,X e ,…,X E}, where X e Represents the word vector feature of the e-th dimension in the total word vector set X, e∈[1,E]; E represents the total dimension of the word vector feature;

[0037] Step 2.5: Use equations (7) and (8) to calculate the weight value α of feature assignment, and then use equation (9) to obtain the total word vector set X" after weight assignment;

[0038] G = tanh(X′) (7)

[0039] α=softmax(wG) (8)

[0040]

[0041] In formula (7)-formula (8), tanh refers to the hyperbolic tangent function, softmax is the activation function, and w is a parameter.

[0042] The step 4.1 is performed as follows:

[0043] Step 4.1.1, use equations (10) to (12) to obtain the value-modality-text information interaction feature MultiHead;

[0044]

[0045]

[0046] MultiHead=concat(head1,head2,…,head l ,…,head L ) 0 (12)

[0047] In formula (10) to formula (12), Q l represents the query matrix of the lth attention head, K l represents the key-value matrix of the lth attention head, V l Represents the value matrix of the lth attention head; Corresponding to Q l , K l 、V l The coefficient matrix of head, L represents the total number of attention heads, l∈[1,L]; head l represents the lth attention head, d k K l Dimension of W 0 Represents the weight matrix of linear transformation; concat represents concatenation;

[0048] Step 4.1.2, the Transformer network uses formula (13) to obtain formula (15) to obtain the value-modality-text fusion feature R':

[0049] MultiHead′=Norm(R+MultiHead) (13)

[0050] FFN(MultiHead')=Relu(MultiHead'W 1 +f1)W 2 +f2 (14)

[0051] R′=Norm(MultiHead′+FFN(MultiHead)) (15)

[0052] In formula (13), Norm represents the normalization operation; + represents the residual connection; in formula (14), FFN represents the feedforward neural network, Relu represents the activation function, and W 1 and W 2 Represents two weight matrices in the feedforward neural network, f1 and f2 represent two bias terms in the feedforward neural network; Relu represents the activation function.

[0053] An electronic device of the present invention includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the coke demand forecasting method, and the processor is configured to execute the program stored in the memory.

[0054] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program executes the steps of the coke demand forecasting method when the computer program is executed by a processor.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] 1. Enhanced data integration capabilities: The multi-source and multi-modal dataset constructed by the present invention can more comprehensively reflect the factors affecting coke demand, and integrates different types of data, such as numerical demand indicators, text-based investor comments, research reports and news, thereby providing richer information support.

[0057] 2. Improve feature extraction performance: The present invention can more accurately extract the most critical features under each mode by pre-extracting demand index data and text data separately. First, the fully adaptive noise ensemble empirical mode decomposition (CEEMDAN) method is selected as the signal decomposition technology to decompose the index features with the highest correlation. This method can provide clear and accurate modal features with its excellent denoising and signal decomposition capabilities. In terms of text data pre-extraction, the present invention will first use BERT to embed words in multi-source text data, and after dimensionality reduction and information fusion in the fully connected layer, the self-attention mechanism is used to further improve the accuracy and comprehensiveness of text feature extraction.

[0058] 3. Model innovation: This invention uses a fusion model based on Transformer and multiple recurrent neural networks (such as GRU, BiGRU, BiLSTM), which can more effectively process time series data and fully capture the nonlinear associations and long-term dependencies between demand features. The use of the attention mechanism further optimizes the weight distribution of features, allowing the model to focus more on the key information that affects the prediction results.

[0059] 4. Feature fusion and prediction effect: The multi-feature fusion strategy of the present invention helps to integrate the features of each modality and use the advantages of each feature to improve the generalization ability of the prediction model. This fusion prediction may be more accurate and reliable than the prediction results of a single modality.

[0060] 5. Responding to dynamic market changes: The method proposed in this invention takes into account the dynamic nature of the market and enables the forecasting model to adapt to market changes more quickly through real-time data processing capabilities and efficient use of unstructured information, as well as comprehensive analysis of multimodal data. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 A flow chart of the coke demand forecasting method based on a multi-feature fusion deep neural network provided by the present invention;

[0062] Figure 2 A schematic diagram of specific steps of fully adaptive noise ensemble empirical mode decomposition implemented in the present invention;

[0063] Figure 3 A flowchart of preprocessing a text data set of coke demand implemented by the present invention;

[0064] Figure 4 The structural diagram of the GRU used in the present invention;

[0065] Figure 5 A structural diagram of the BiGRU used in the present invention;

[0066] Figure 6 This is a structural diagram of the BiLSTM used in the present invention. DETAILED DESCRIPTION

[0067] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0068] In this embodiment, Figure 1 As shown, a coke demand prediction method based on multi-feature fusion deep neural network includes the following steps:

[0069] Step 1: Preprocess the numerical indicator set P of coke demand to obtain the modal feature set IMF:

[0070] Step 1.1: Construct a numerical index set P of coke demand = {P1, P2, ..., P n ,…,P N}, where P n represents the nth index, n∈[1,N], N represents the total number of indexes, and P n ={p n,1 ,p n,2 ,…,p n,m ,…,p n,M}, where p n,m Represents the nth index P n The characteristic parameter of the mth time, m∈[1,M], M represents the total number of time;

[0071] This indicator should be numerical data of factors affecting coke demand, including but not limited to the following aspects: indicators related to coke itself, such as coke demand, coke production, coke futures price, coke import and export volume, coke import and export amount, and coke inventory; indicators related to upstream and downstream products, such as steel production, steel price, steel import and export volume, steel import and export amount, iron ore price, coking coal production, coking coal demand, coking coal import and export volume, coking coal import and export amount, and coking coal transportation cost; indicators related to substitute products, such as natural gas price, natural gas production, methanol price, methanol production, gasoline price, fuel oil price, crude oil price, and direct reduced iron (DRI) price; macroeconomic indicators, such as global and regional GDP growth rate, industrial production index, and fixed asset investment growth rate; environmental policy and technology development indicators, such as carbon emission price and clean energy policy indicators.

[0072] Step 1.2, let the current number of empirical mode decomposition be k, and initialize k = 1;

[0073] Step 1.3: To prevent data leakage caused by directly performing modal decomposition on the predicted demand indicators, perform Pearson correlation analysis on each indicator in the coke demand indicator set P with other indicators to find the indicator with the highest correlation P. u , thus P u Perform CEEMDAN decomposition, such as Figure 2 As shown, first in P u The k-1th amplitude is ε when the i-th k-1 Gaussian white noise v i After that, we get the i-th characteristic signal I represents the total number of additions.

[0074] Step 1.4: Perform empirical mode decomposition to obtain the kth eigenmode component after empirical mode decomposition Then, the kth intrinsic mode component IMF is obtained using formula (2): k , and use formula (3) to get the kth residual r k ;

[0075]

[0076] r k =P u -IMF k (3)

[0077] Step 1.5: At the kth residual r k The kth amplitude is ε when the i-th addition is added k Gaussian white noise v i , get the i-th new characteristic signal r1+ε1vi , and then use formula (4) to get the k+1th intrinsic mode component IMF k+1 ;

[0078]

[0079] In formula (4), E represents the empirical mode decomposition operation.

[0080] Step 1.6: Calculate the k+1th residual r using formula (5) k+1 ;

[0081] r k+1 =r k -IMF k+1 (5)

[0082] Step 1.7, after copying k+1 to k, return to step 1.5 and execute sequentially until the remaining residual is a monotonic function, thereby obtaining the modal feature set IMF = {IMF1, IMF2, ..., IMF k ,…,IMF K}, K represents the total number of modal features, and IMF k ={imf k,1 ,imf k,2 ,…,imf k,m ,…,imf k,M}, imf k,m IMF k The mth time corresponds to the modal parameter.

[0083] Step 2: Preprocess the text dataset of coke demand to obtain the total word vector set X. The specific process is as follows: Figure 3 As shown:

[0084] Step 2.1: Collect multi-source text sets of coke demand, including: investor comment text set T1 = {T 1,1 ,T 1,2 ,…,T 1,m ,…,T 1,M}、Research report text set T2={T 2,1 ,T 2,2 ,…,T 2,m ,…,T 2,M} and news text set T3 = {T 3,1 ,T 3,2 ,…,T 3,m ,…,T 3,M}, where T 1,m represents the investor comment text data at time m, and Indicates T 1,mThe ath investor comment word in T, a∈[1,A], A represents T 1,m Total number of words in investor comments; 2,m represents the text data of the research report at time m, and Indicates T 2,m The bth research report word in T, b∈[1,B], B represents T 2,m Total number of words in the research report; 3,m represents the news text data of the mth date, and in, Indicates T 3,m The cth news word in T, c∈[1,C], C represents T 3,m Total number of news words;

[0085] The multi-source text should be text-based information on factors affecting coke demand, including but not limited to the following aspects: investor comments in economic forums, research reports in Eastmoney.com, and information on news websites.

[0086] Step 2.2: T 1,m 、T 2,m and T 3,m After preprocessing, D words are extracted respectively, and then the pre-trained model BERT is used to embed the extracted D words respectively, and the corresponding word vector of the investor comment at the mth time is obtained The word vector set of the research report at the mth time And the news word vector set at the mth time Thus, we get the investor comment word vector set X1 = {X 1,1 ,X 1,2 ,…,X 1,m ,…,X 1,M}、Research report word vector set X2={X 2,1 ,X 2,2 ,…,X 2,m ,…,X 2,M} and news word vector set X3 = {X 3,1 ,X 3,2 ,…,X 3,m ,…,X 3,M};in, represents the dth investor comment word vector, represents the d-th research report word vector, represents the d-th news word vector, where d∈[1,D] and D represents the length of the word vector.

[0087] Step 2.3: Transpose the investor comment word vector X1, the research report word vector X2, and the news word vector X3 and concatenate them to obtain the total word vector set in, represents transpose;

[0088] In the text processing flow of this embodiment, the focus is on the text information collected on a daily basis, so transposition and alignment with other features are performed. Such a data structure arrangement is intended to ensure that different types of features can be effectively compared and integrated under the same analysis framework.

[0089] Step 2.4: Use the fully connected layer to reduce the dimension and exchange information of the total word vector set X to obtain the word vector set X'={X1,X2,…,X e ,…,X E}, where X e Represents the word vector feature of the e-th dimension in the total word vector set X, e∈[1,E]; E represents the total dimension of the word vector feature;

[0090] This is because the present invention performs word embedding processing on multi-source text data separately, generating high-dimensional vectors. These high-dimensional vectors are not very beneficial for subsequent feature extraction when spliced ​​together. Therefore, in this embodiment, a dimensionality reduction strategy using a fully connected layer is adopted to balance the influence of text features and other types of features. This step not only reduces the dimension of the data, but also promotes the integration of information from different text sources to a certain extent. Afterwards, in order to more deeply integrate and coordinate text data from different sources, the present invention applies a self-attention mechanism. The use of this strategy further improves the inherent consistency of the data and the efficiency of extracting relevant information.

[0091] Step 2.5: Use equations (7) and (8) to calculate the weight value α of feature assignment, and then use equation (9) to obtain the total word vector set X" after weight assignment;

[0092] G = tanh(X′) (7)

[0093] α=softmax(wG) (8)

[0094]

[0095] In formula (7)-formula (8), tanh refers to the hyperbolic tangent function, softmax is the activation function, and w is a parameter.

[0096] Step 3: After concatenating the numerical indicator set P, the modal feature set IMF and the total word vector set X” after weight assignment, the coke demand input data set R = {P, IMF, X”} is obtained;

[0097] Step 4: Construct a multi-feature fusion deep neural network consisting of a Transformer network, a feature extraction model, and a fully connected layer;

[0098] Step 4.1: The Transformer network performs feature fusion pre-extraction on R to obtain the numerical-modal-text fusion feature R'. The multi-head attention mechanism of the Transformer model can better handle long-term dependencies, which can help solve the gradient vanishing or exploding problems that occur when extracting features from subsequent RNN models. In addition, the Transformer model can globally focus on all input elements, capture richer global semantic information, rather than just being limited to local context, and improve the overall feature extraction effect and prediction accuracy.

[0099] Step 4.1.1, use equations (10) to (12) to obtain the value-modality-text information interaction feature MultiHead;

[0100]

[0101]

[0102] MultiHead=concat(head1,head2,…,head l ,…,head L )W 0 (12)

[0103] In formula (10) to formula (12), Q l represents the query matrix of the lth attention head, K l represents the key-value matrix of the lth attention head, V l Represents the value matrix of the lth attention head; Corresponding to Q l , K l 、V l The coefficient matrix of head, L represents the total number of attention heads, l∈[1,L]; head l represents the lth attention head, d k K l Dimension of W 0 Represents the weight matrix of linear transformation; concat means concatenation.

[0104] Step 4.1.2, the Transformer network uses formula (13) to obtain formula (15) to obtain the value-modality-text fusion feature R':

[0105] MultiHead′=Norm(R+MultiHead) (13)

[0106] FFN(MultiHead')=Relu(MultiHead'W 1 +f1)W 2 +f2 (14)

[0107] R′=Norm(MultiHead′+FFN(MultiHead′)) (15)

[0108] In formula (13), Norm represents the normalization operation; + represents the residual connection; in formula (14), FFN represents the feedforward neural network, Relu represents the activation function, and W 1 and W 2 Represents two weight matrices in the feedforward neural network, f1 and f2 represent two bias terms in the feedforward neural network; Relu represents the activation function.

[0109] Step 4.2: Construct a feature extraction model consisting of a GRU model, a BiGRU model, and a BiLSTM model and perform feature extraction. The structural diagrams of the three models are as follows: Figure 4 , Figure 5 , Figure 6 As shown, by fusing the features extracted by the three methods, the present embodiment can more deeply mine and refine the feature information. Their unique advantages enable the present invention to capture the complex features of sequence data from multiple dimensions, thereby significantly improving the prediction accuracy. At the same time, the use of multiple neural network structures for feature extraction also helps to solve the overfitting problem. Since different network structures have different fitting methods, by integrating their prediction results, the possible deviations of a single model can be effectively reduced, and the robustness of the overall prediction can be improved:

[0110] Step 4.2.1: Extract features of R' based on the GRU model to obtain the dynamic characteristics of coke demand in, represents the sth dynamic characteristic of coke demand, s∈[1,S]; S represents the total number of dynamic characteristics of coke demand;

[0111] The Gated Recurrent Unit (GRU) is an efficient recurrent neural network architecture for processing sequence data, especially suitable for feature extraction in demand forecasting. GRU regulates the flow of information through a unique update gate and reset gate mechanism, which allows the network to more effectively capture the temporal dependencies of data at different time steps, while simplifying the parameters in the traditional long short-term memory network (LSTM), optimizing computational efficiency, while still maintaining the ability to capture long-term dependencies. The update gate helps the model decide how much previous information to retain in the current state, while the reset gate determines how much past information needs to be forgotten, which enables GRU to integrate information at different time scales and dynamically encode time series data.

[0112] Step 4.2.2: Extract features of R' based on the BiGRU model to obtain the panoramic features of coke demand in, represents the panoramic characteristics of the s-th coke demand;

[0113] The Bidirectional Gated Recurrent Unit (BiGRU) combines the efficient gating mechanism of the GRU with the comprehensive temporal understanding of the bidirectional network to improve the feature extraction performance in demand forecasting. By processing the forward and reverse flows of time series data in parallel, the BiGRU can capture richer historical and future information at each time point, thereby providing more refined contextual analysis. This structure not only improves the ability to capture complex dependencies in demand patterns, such as seasonal and trend changes in demand, but also maintains computational efficiency. It dynamically adjusts the model's absorption and forgetting of information at each time step, making the BiGRU more flexible and robust in the face of a changing demand environment, thereby accurately grasping demand characteristics and greatly improving the accuracy and reliability of predictions. In this process, the BiGRU not only optimizes the feature extraction process, but also ensures that the demand forecasting model can respond quickly and accurately to various influencing factors.

[0114] Step 4.2.3: Extract features of R' based on the BiLSTM model to obtain the deep sequence features of coke demand in, represents the sth coke demand depth sequence characteristics;

[0115] The bidirectional long short-term memory network (BiLSTM) combines two LSTMs, one responsible for parsing the positive time order of the data (from the past to the future) and the other processing the reverse time order of the data (from the future to the past), so that the network can fully capture the time correlation in the data. This two-way processing mechanism allows BiLSTM to not only use past information when predicting future demand, but also consider possible future influencing factors. In addition, thanks to the long-term memory capacity of the LSTM part, BiLSTM can identify and maintain long-term data dependencies, which is crucial for identifying and understanding long-term trends and cyclical changes in demand patterns.

[0116] Step 4.3: Y GRU , Y BiGRU and Y BiLSTM After splicing, the data is input into the fully connected layer for demand forecasting to obtain the predicted coke demand Y;

[0117] Step 5. Calculate the mean square error based on the prediction result Y and use it as the objective function J, so as to train the multi-feature fusion deep neural network using back propagation and gradient descent method to minimize the objective function J. When the number of iterations reaches the maximum number of iterations, stop training to obtain the trained coke demand prediction model for the prediction of coke demand.

[0118] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0119] In this embodiment, a computer-readable storage medium stores a computer program on the computer-readable storage medium, and the computer program executes the steps of the above method when executed by a processor.

Claims

1. A coke demand forecasting method based on multi-feature fusion deep neural network, It is characterized in that The steps are as follows: Step 1: Preprocess the numerical indicator set P of coke demand to obtain the modal feature set IMF; Step 2: Preprocess the text data set of coke demand to obtain the total word vector set X”; Step 3: After concatenating the numerical indicator set P, the modal feature set IMF and the total word vector set X” after weight assignment, the coke demand input data set R = {P, IMF, X”} is obtained; Step 4: Construct a multi-feature fusion deep neural network consisting of a Transformer network, a feature extraction model, and a fully connected layer; Step 4.1, the Transformer network performs feature fusion pre-extraction processing on R to obtain a numerical-modality-text fusion feature R'; Step 4.2: Build a feature extraction model consisting of a GRU model, a BiGRU model, and a BiLSTM model and perform feature extraction: Step 4.2.1: Extract features of R' based on the GRU model to obtain the dynamic characteristics of coke demand in, represents the sth dynamic characteristic of coke demand, s∈[1,S]; S represents the total number of dynamic characteristics of coke demand; Step 4.2.2: Extract features of R' based on the BiGRU model to obtain the panoramic features of coke demand in, represents the panoramic characteristics of the s-th coke demand; Step 4.2.3: Extract features of R' based on the BiLSTM model to obtain the deep sequence features of coke demand in, represents the sth coke demand depth sequence characteristics; Step 4.3: Y GRU , Y BiGRU and Y BiLSTM After splicing, the data is input into the fully connected layer for demand forecasting to obtain the predicted coke demand Y; Step 5. Calculate the mean square error based on the prediction result Y and use it as the objective function J, so as to train the multi-feature fusion deep neural network using back propagation and gradient descent method to minimize the objective function J. When the number of iterations reaches the maximum number of iterations, stop training to obtain the trained coke demand prediction model for the prediction of coke demand.

2. According to claim 1, a method for predicting coke demand based on a multi-feature fusion deep neural network, It is characterized in that The step 1 is carried out as follows: Step 1.1: Construct a numerical index set P of coke demand = {P 1 ,P 2 ,…,P n ,…,P N }, where P n represents the nth index, n∈[1,N], N represents the total number of indexes, and P n ={p n,1 ,p n,2 ,…,p n,m ,…,p n,M }, where p n,m Represents the nth index P n The characteristic parameter of the mth time, m∈[1,M], M represents the total number of time; Step 1.2, let the current number of empirical mode decomposition be k, and initialize k = 1; Step 1.3: Perform Pearson correlation analysis on each indicator in the coke demand indicator set P with other indicators to find the indicator P with the highest correlation. u , so that in P u The k-1th amplitude is ε when the i-th k-1 Gaussian white noise v i After that, we get the i-th characteristic signal i∈[1,I]; I represents the total number of joins; Step 1.4: Perform empirical mode decomposition to obtain the kth eigenmode component after empirical mode decomposition Then, the kth intrinsic mode component IMF is obtained using formula (2): k , and use formula (3) to get the kth residual r k ; r k =P u -IMF k (3) Step 1.5: At the kth residual r k The kth amplitude is ε when the i-th addition is added k Gaussian white noise v i , get the i-th new feature signal r 1 +ε 1 v i , and then use formula (4) to get the k+1th intrinsic mode component IMF k+1 ; In formula (4), E represents the empirical mode decomposition operation; Step 1.6: Calculate the k+1th residual r using formula (5) k+1 ; r k+1 =r k -IMF k+1 (5) Step 1.7, after copying k+1 to k, return to step 1.5 and execute sequentially until the remaining residual is a monotonic function, thereby obtaining the modal feature set IMF = {IMF 1 ,IMF 2 ,…,IMF k ,…,IMF K }, K represents the total number of modal features, and IMF k ={imf k,1 ,imf k,2 ,…,imf k,m ,…,imf k,M }, imf k,m IMF k The mth time corresponds to the modal parameter.

3. A method for predicting coke demand based on a multi-feature fusion deep neural network according to claim 2, It is characterized in that Described step 2 is carried out as follows: Step 2.1: Collect multi-source text collections of coke demand, including: investor comment text collection T 1 ={T 1,1 ,T 1,2 ,…,T 1,m ,…,T 1,M }、Research report text collection T 2 ={T 2,1 ,T 2,2 ,…,T 2,m ,…,T 2,M } and news text collection T 3 ={T 3,1 ,T 3,2 ,…,T 3,m ,…,T 3,M }, where T 1,m represents the investor comment text data at time m, and Indicates T 1,m The ath investor comment word in T, a∈[1,A], A represents T 1,m Total number of words in investor comments; 2,m represents the text data of the research report at time m, and Indicates T 2,m The bth research report word in T, b∈[1,B], B represents T 2,m Total number of words in the research report; 3,m represents the news text data of the mth date, and in, Indicates T 3,m The cth news word in T, c∈[1,C], C represents T 3,m Total number of news words; Step 2.2: T 1,m , T 2,m and T 3,m After preprocessing, D words are extracted respectively, and then the pre-trained model BERT is used to embed the D words respectively, and the corresponding word vector of the investor comment at the mth time is obtained. The word vector set of the research report at the mth time And the news word vector set at the mth time Thus, we get the investor comment word vector set X 1 ={X 1,1 ,X 1,2 ,…,X 1,m ,…,X 1,M }、Research report word vector set X 2 ={X 2,1 ,X 2,2 ,…,X 2,m ,…,X 2,M } and news word vector set X 3 ={X 3,1 ,X 3,2 ,…,X 3,m ,…,X 3,M };in, represents the dth investor comment word vector, represents the d-th research report word vector, represents the d-th news word vector, where d∈[1,D], and D represents the length of the word vector; Step 2.3: Transform investor comment word vector X 1 、Research report word vector X 2 and news word vector X 3 After transposition, concatenate to get the total word vector set in, represents transpose; Step 2.4: Use the fully connected layer to reduce the dimension and exchange information of the total word vector set X to obtain the word vector set X'={X 1 ,X 2 ,…,X e ,…,X E }, where X e Represents the word vector feature of the e-th dimension in the total word vector set X, e∈[1,E]; E represents the total dimension of the word vector feature; Step 2.5: Use equations (7) and (8) to calculate the weight value α of feature assignment, and then use equation (9) to obtain the total word vector set X" after weight assignment; G = tanh(X) (7) α=softmax(wG) (8) In formula (7)-formula (8), tanh refers to the hyperbolic tangent function, softmax is the activation function, and w is a parameter.

4. A method for predicting coke demand based on a multi-feature fusion deep neural network according to claim 3, It is characterized in that The step 4.1 is performed as follows: Step 4.1.1, use equations (10) to (12) to obtain the value-modality-text information interaction feature MultiHead; MultiHead=concat(head 1 ,head 2 ,…,head l ,…,head L )W 0 (12) In formula (10) to formula (12), Q l represents the query matrix of the lth attention head, K l represents the key-value matrix of the lth attention head, V l Represents the value matrix of the lth attention head; Corresponding to Q l , K l 、V l The coefficient matrix of head, L represents the total number of attention heads, l∈[1,L]; head l represents the lth attention head, d k K l Dimension of W 0 Represents the weight matrix of linear transformation; concat represents concatenation; Step 4.1.2: The Transformer network uses equations (13) to (15) to obtain the value-modality-text fusion feature R': MultiHead′=Norm(R+MultiHead) (13) FFN(MultiHead')=Relu(MultiHed'W 1 +f 1 )W 2 +f 2 (14) R′=Norm(MultiHead′+FFN(MultiHead′)) (15) In formula (13), Norm represents the normalization operation; + represents the residual connection; in formula (14), FFN represents the feedforward neural network, Relu represents the activation function, and W 1 and W 2 Represents the two weight matrices in the feedforward neural network, f 1 and f 2 Represents two bias terms in a feedforward neural network; Relu represents the activation function.

5. An electronic device comprising a memory and a processor, It is characterized in that The memory is used to store a program that supports the processor to execute the coke demand forecasting method described in any one of claims 1 to 4, and the processor is configured to execute the program stored in the memory.

6. A computer-readable storage medium having a computer program stored thereon, It is characterized in that When the computer program is executed by a processor, the steps of the coke demand forecasting method in claims 1-4 are executed.