A coke demand prediction method and device based on multi-source multi-modal data collaborative driving and a medium

Abstract

The application discloses a coke demand prediction method and device based on multi-source multi-modal data collaborative driving and a medium, and steps of the method comprise the following steps: 1, constructing a multi-modal coke demand dataset; 2, constructing a feature extraction model for extracting features of the multi-modal coke demand; 3, constructing a cross-modal attention mechanism model for fusing the features of the multi-modal coke demand; and 4, predicting demand by using the fused features. The application can assist purchasing personnel in formulating production and purchasing plans by means of a deep learning method, fully utilize production resources, and thus achieve the goal of reducing cost and increasing efficiency of enterprises.

Description

Technical Field

[0001] This invention belongs to the field of computer integrated forecasting technology, specifically a method, equipment, and medium for predicting coke demand based on multi-source, multi-modal data collaborative driving. Background Technology

[0002] As a key raw material for the steel industry, the coking coal industry plays a vital supporting role in national economic development and major engineering projects. Its complex industrial chain, numerous participating enterprises, and wide-ranging impact, coupled with strong demand from enterprises for risk aversion and investment, give coking coal significant strategic and economic value. Coking coal is also one of the main products of coking enterprises. Coking coal demand forecasting provides crucial information for coking enterprises to formulate strategic plans, production arrangements, and sales plans, which is essential for their survival and development. Currently, there is a lack of research on coking coal demand forecasting, mainly due to two reasons: First, coking coal production and application are concentrated in heavy industrial enterprises, resulting in severe environmental pollution, complex processes, significant difficulties in financing, slow automation, and limited application scenarios; Second, the factors influencing coking coal demand are highly complex, exhibiting non-stationary, non-linear, and multi-source characteristics, with various factors intertwined and difficult to fully grasp. Therefore, research on coking coal demand forecasting has high practical value and theoretical significance.

[0003] Demand forecasting methods are mainly divided into two categories: statistical forecasting methods and artificial intelligence methods. Statistical forecasting methods primarily rely on statistical models, such as the Autoregressive Integrated Moving Average (ARIMA) model, the Generalized Autoregressive Conditional Heteroskedasticity (GARCH) model, and the Heterogeneous Autoregressive (HAR) model. Statistical methods must be used based on time-series data with specific linear relationships and constant variance, establishing a correlation between the dependent and independent variables to predict the future trend of the dependent variable. Therefore, statistical forecasting methods are difficult to use for forecasting non-stationary and non-linear coking coal demand time series.

[0004] Coke, as an important energy resource, plays a crucial role in the energy industry. Accurately predicting coke demand to optimize the scheduling of coking production workshops directly impacts a company's production efficiency and costs, thus affecting its decision-making and management. Furthermore, due to the complex coke production process and the intricate upstream and downstream industries, few scholars have researched the problem of coke demand forecasting. Summary of the Invention

[0005] The present invention aims to address the shortcomings of the existing technology by proposing a method, equipment, and medium for predicting coke demand based on multi-source and multi-modal data collaborative driving. It is intended to assist procurement personnel in formulating production and procurement plans through deep learning methods, making full use of production resources, thereby achieving the goal of cost reduction and efficiency improvement for enterprises.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: The present invention provides a method for predicting coke demand based on multi-source, multi-modal data collaborative driving, characterized by the following steps: Step 1: Construct a dataset of multimodal coke demand: Step 1.1: Construct a set of indicators Q for multimodal coke demand. ,in, Let represent the i-th indicator, and ,in, Represents the i-th indicator The nth parameter in , express The number of parameters, , N The number of indicators representing a set of indicators; Step 1.2: Construct a text set T of multimodal coke demand. , This represents the i-th text message, and ,in, express The first in One word, , express Vocabulary size; Step 1.3, for Perform word embedding to obtain word vectors. ,in, This represents the i-th text information. The k-th word vector in the middle, , This represents the word vector length, thus obtaining the set of word vectors X for multimodal coke demand. ; Step 2: Construct a feature extraction model to extract features of multimodal coke demand: Step 2.1: Extract features from the index set Q using the BiLSTM model to obtain the index feature representation. ;in, This represents the i-th indicator feature; Step 2.2: Extract features from the word vector set X based on the ConvGRU model to obtain the text feature representation. ; in, Represents the j-th text feature; Step 3: Construct a cross-modal attention mechanism model to fuse features of multimodal coke demand: Step 3.1: Calculate the i-th indicator feature using equations (1) and (2). With the j-th text feature correlation between and their weights :

[0007]

[0008] Step 3.2: Set a weight threshold. ,like > ,but =1, otherwise, let =0, thus obtaining the threshold matrix. ,Will The inverse threshold matrix is ​​denoted as ; Step 3.3: Calculate the common characteristics of the index using equations (3) and (4). and the unique characteristics of indicators :

[0009]

[0010] Step 3.4: Identify the common characteristics of the indicators. After being input into the Transformer model, the Q-vector is obtained. K vector , V vector ; The unique characteristics of the indicators After being input into the Transformer model, the Q-vector is obtained. K vector , V vector ; Step 3.5: Use equation (5) to obtain the attention score A for the cross-fusion of the two modalities:

[0011] Step 3.6: Assign attention scores Common characteristics of indicators Perform a dot product operation to obtain the fusion features of the indicator text. ; Step 4: Predict demand using the fused features: Step 4.1: Merge features Demand prediction is performed in the fully connected layer to obtain the prediction results. ; Step 4.2: A coke demand prediction model is constructed from a feature extraction model, an attention mechanism model, and a fully connected layer, using the mean square logarithmic error as the objective function. The objective function J is trained using backpropagation and gradient descent to minimize it. Training stops when the maximum number of iterations is reached, thus obtaining the trained coke demand prediction model for predicting coke demand.

[0012] The present invention provides an electronic device, comprising a memory and a processor, wherein the memory is used to store a program that supports the processor in executing the coke demand forecasting method, and the processor is configured to execute the program stored in the memory.

[0013] The present invention discloses a computer-readable storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, performs the steps of the coke demand forecasting method.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Based on the actual production needs of coking enterprises, this invention proposes a coke demand forecasting method driven by multi-source and multi-modal data collaboration for enterprise procurement decision-making. It is applied to actual procurement and production scenarios, which not only broadens the application field of deep learning algorithms, but also solves the defects of traditional procurement processes lacking intelligent methods, ensures that errors caused by human factors in procurement are avoided, allows for flexible adjustments to procurement plans, brings competitive advantages to enterprises, and realizes the long-term development of enterprises.

[0015] 2. The attention mechanism of this invention enhances the network's ability to autonomously select key information by simulating the human brain's attention allocation mechanism, and is therefore widely used in multimodal fusion operations. Standard attention mechanisms assign different weights to different modal information input into the model to measure the importance of each feature to the final prediction result, but fail to consider the cross-relationships between modalities. Therefore, based on the consideration of using an attention mechanism, common and unique features in the modal information are extracted, and then a cross-modal attention mechanism is used to calculate the attention score for cross-modal fusion, obtaining the fusion features of different modalities, thereby improving the accuracy and robustness of coke demand forecasting. Attached Figure Description

[0016] Figure 1 The flowchart of the coke demand forecasting method based on multi-source multimodal data collaborative driving provided by the present invention is shown below. Figure 2 The flowchart for multimodal feature fusion of coke demand based on cross-modal attention mechanism provided by the present invention is as follows. Detailed Implementation

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0018] In this embodiment, as Figure 1 As shown, a coke demand forecasting method based on multi-source, multi-modal data collaborative driving includes the following steps: Step 1: Construct a dataset of multimodal coke demand: Step 1.1: Construct a set of indicators Q for multimodal coke demand. ,in, Let i represent the i-th indicator, and ,in, Represents the i-th indicator The nth parameter in , express The number of parameters, , N The number of indicators representing a set of indicators; This indicator should be numerical information among the factors influencing coke demand, including but not limited to the following: indicators related to coke itself, such as coke demand, coke production, coke futures prices, coke import and export volume, and coke import and export value; indicators related to upstream and downstream products, such as steel production, steel prices, steel import and export volume, and steel import and export value; coal production, coal demand, coal import and export volume, and coal import and export value; indicators related to substitute products, such as coking coal prices, coking coal production, methanol prices, methanol production, gasoline prices, fuel oil prices, natural gas prices, crude oil inventories, and crude oil prices; and macroeconomic indicators, such as inflation rate, exchange rate, and international income and expenditure.

[0019] Step 1.2: Construct a text set T of multimodal coke demand. , This represents the i-th text message, and ,in, express The first in One word, , express Vocabulary size; The text should be textual information among the factors influencing coke demand, including but not limited to the following: comments from netizens on economic forums, news information on news websites, and textual information related to the coke industry and its upstream and downstream industries in policy documents on the Chinese government website.

[0020] Step 1.3, for Perform word embedding to obtain word vectors ,in, Represents the i-th text information The k-th word vector in the middle, , This represents the word vector length, thus obtaining the set of word vectors X for multimodal coke demand. This enables the symbolization and mathematization of text information, transforming it into a format that computers can recognize. Step 2: Construct a feature extraction model to extract features of multimodal coke demand: Step 2.1: Extract features from the index set Q using the BiLSTM model to obtain the index feature representation. ;in, This represents the i-th indicator feature; BiLSTM consists of a forward LSTM neural network and a backward LSTM network. The forward LSTM network reads the indicator information sequentially, while the backward LSTM network reads it in reverse order. It calculates two separate hidden states for the forward and backward hidden sequences, and the final output is obtained by concatenating these two hidden states. BiLSTM can extract data feature information from two perspectives at the current time point, better capturing bidirectional dependencies. It outperforms LSTM in processing time-series data with correlations between preceding and following times, solving the problem that LSTM can only process data forward and ignores subsequent states. Furthermore, BiLSTM neural networks have stronger non-linear expressive power, making them highly effective for sequence-based tasks and better able to handle the complexity and uncertainty of coke demand time series.

[0021] Step 2.2: Extract features from the word vector set X based on the ConvGRU model to obtain the text feature representation. ; in, Represents the j-th text feature; The core idea of ​​ConvGRU is to add convolutional operations to the matrix operations of traditional GRU networks, thus combining the local feature extraction capabilities of CNNs and the temporal modeling capabilities of GRUs. The convolutional layers at the bottom of ConvGRU can better extract spatiotemporal features from coke demand text information. Through convolutional operations between the kernel and the data, the kernel parameters are shared among the input data, so the convolutional operation can also better extract generalized features from coke demand text information.

[0022] Step 3: Classical attention mechanisms, when performing modality fusion, only assign different weights based on the importance of different features within a modality, failing to consider the interactions between different modalities. Therefore, two solutions are proposed to address this deficiency: 1. Calculate the common and unique features of the indicators, using the unique features as supplementary features during multimodal fusion. This fully utilizes the relationships between different modalities and achieves the goal of fusing low-level and high-level features, significantly improving the fusion effect of multimodal features. 2. Calculate cross-modal attention to obtain a cross-modal representation that includes complementary information from both modalities as well as unique indicator features, reducing information loss during cross-modal fusion and ultimately obtaining cross-modal fused features. For example... Figure 2 As shown, a cross-modal attention mechanism model is constructed to fuse features of multimodal coke demand: Step 3.1: Calculate the i-th indicator feature using equations (1) and (2). With the j-th text feature correlation between and their weights :

[0023]

[0024] Step 3.2: Set a weight threshold. ,like > ,but =1, otherwise, let =0, thus obtaining the threshold matrix. ,Will The inverse threshold matrix is ​​denoted as ; Step 3.3: Calculate the common characteristics of the index using equations (3) and (4). and the unique characteristics of indicators :

[0025]

[0026] Step 3.4: Identify the common characteristics of the indicators. After being input into the Transformer model, the Q-vector is obtained. K vector , V vector ; The unique characteristics of the indicators After being input into the Transformer model, the Q-vector is obtained. K vector , V vector ; Step 3.5: Use equation (5) to obtain the attention score A for the cross-fusion of the two modalities:

[0027] Step 3.6: Assign attention scores Common characteristics of indicators Perform a dot product operation to obtain the fusion features of the indicator text. ; Step 4: Predict demand using the fused features: Step 4.1: Merge features Demand prediction is performed in the fully connected layer to obtain the prediction results. ; Step 4.2: A coke demand prediction model is constructed from a feature extraction model, an attention mechanism model, and a fully connected layer, using the mean square logarithmic error as the objective function. The objective function J is trained using backpropagation and gradient descent to minimize it. Training stops when the maximum number of iterations is reached, thus obtaining the trained coke demand prediction model for predicting coke demand.

[0028] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0029] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

Claims

1. A method for predicting coke demand based on multi-source, multi-modal data collaborative driving, characterized in that, The procedure is as follows: Step 1: Construct a dataset of multimodal coke demand: Step 1.1: Construct a set of indicators Q for multimodal coke demand. ,in, Let represent the i-th indicator, and ,in, Represents the i-th indicator The nth parameter in , express The number of parameters, N represents the number of indicators in the indicator set; Step 1.2: Construct a text set T of multimodal coke demand. , Let i represent the i-th text message, and ,in, express The first in One word, , express Vocabulary size; Step 1.3, for Perform word embedding to obtain word vectors. ,in, This represents the i-th text information. The k-th word vector in the middle, , This represents the word vector length, thus obtaining the set of word vectors X for multimodal coke demand. ; Step 2: Construct a feature extraction model to extract features of multimodal coke demand: Step 2.1: Extract features from the index set Q using the BiLSTM model to obtain the index feature representation. ;in, This represents the i-th indicator feature; Step 2.2: Extract features from the word vector set X based on the ConvGRU model to obtain the text feature representation. ; in, Represents the j-th text feature; Step 3: Construct a cross-modal attention mechanism model to fuse features of multimodal coke demand: Step 3.1: Calculate the i-th indicator feature using equations (1) and (2). With the j-th text feature correlation between and their weights : (1) (2) Step 3.2: Set a weight threshold. ,like > ,but =1, otherwise, let =0, thus obtaining the threshold matrix. ,Will The inverse threshold matrix is ​​denoted as ; Step 3.3: Calculate the common characteristics of the index using equations (3) and (4). and the unique characteristics of indicators : (3) (4) Step 3.4: Identify the common characteristics of the indicators. After being input into the Transformer model, the Q-vector is obtained. K vector , V vector ; The unique characteristics of the indicators After being input into the Transformer model, the Q-vector is obtained. K vector , V vector ; Step 3.5: Use equation (5) to obtain the attention score A for the cross-fusion of the two modalities: (5) Step 3.6: Assign attention scores Common characteristics of indicators Perform a dot product operation to obtain the fusion features of the indicator text. ; Step 4: Predict demand using the fused features: Step 4.1: Merge features Demand prediction is performed in the fully connected layer to obtain the prediction results. ; Step 4.2: A coke demand prediction model is constructed from a feature extraction model, an attention mechanism model, and a fully connected layer, using the mean square logarithmic error as the objective function. The objective function J is trained using backpropagation and gradient descent to minimize it. Training stops when the maximum number of iterations is reached, thus obtaining the trained coke demand prediction model for predicting coke demand.

2. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store programs that support the processor in executing the coke demand forecasting method of claim 1, and the processor is configured to execute the programs stored in the memory.

3. A computer-readable storage medium storing a computer program, characterized in that, The computer program is executed by the processor to perform the steps of the coke demand forecasting method of claim 1.

Images