Lithium carbonate price prediction method based on multi-source multi-modal data cooperative driving

Abstract

The invention discloses a lithium carbonate price prediction method based on multi-source and multi-modal data cooperative driving, and belongs to the technical field of computer integrated prediction, and the method comprises the following steps: S1, constructing a multi-modal lithium carbonate price data set; s2, determining an optimal parameter combination by using a parameter optimization algorithm; s3, constructing a feature extraction model for extracting the price features of the multi-modal lithium carbonate; s4, constructing a cross-modal mixed attention mechanism model for fusing the price characteristics of the multi-modal lithium carbonate; and S5, predicting the price by using the fused features. According to the lithium carbonate price prediction method based on multi-source and multi-modal data cooperative driving, purchasing personnel can be assisted in production and purchase plan making through a deep learning method, production resources are fully utilized, and therefore the purposes of cost reduction and efficiency improvement of enterprises can be achieved.

Description

Technical Field

[0001] The present invention relates to the field of computer integrated forecasting technology, and in particular to a lithium carbonate price forecasting method based on collaborative driving of multi-source and multi-modal data. Background Art

[0002] With the rapid development of the global new energy industry, the market demand for lithium carbonate, as a key raw material for lithium batteries, continues to rise. The fluctuation of lithium carbonate prices not only directly affects the cost control of battery manufacturers, but also has a profound impact on the profitability and market competitiveness of the entire new energy vehicle industry chain. Therefore, accurately predicting the price of lithium carbonate has become the focus of attention within and outside the industry.

[0003] At present, although the forecast of lithium carbonate prices can reflect market trends to a certain extent, it has limitations in data acquisition difficulty, processing capabilities and forecasting models. It is often difficult to capture subtle changes in the market and the immediate impact of emergencies, resulting in greater uncertainty and lag in the forecast results. The main reasons can be summarized as follows: First, the supply and demand relationship in the lithium carbonate market is complex, and is affected by multiple factors such as global lithium mineral resources, new energy vehicle demand, and policy environment. The uncertainty and dynamic changes of these factors make price forecasting particularly difficult; second, the data is scattered and difficult to integrate, and the limitations of data processing technology increase the difficulty of forecasting. Although there are multiple combinations of forecasting models, each method has its limitations and it is difficult to fully adapt to the complexity and variability of the market. Therefore, the study of lithium carbonate price forecasting has high practical value and theoretical significance.

[0004] Accurate lithium carbonate price forecasts have a significant impact on the development of industrial enterprises. On the one hand, it can help enterprises optimize their procurement strategies, conduct better price negotiations with suppliers, and assist enterprises to purchase raw materials and other inputs at the best price at the right time, thereby reducing costs and increasing profit margins; on the other hand, accurate price forecasts can help enterprises effectively manage risks. By predicting price fluctuations, enterprises can take measures to mitigate the impact of fluctuations on their business. In general, accurate price forecasts can help enterprises discover opportunities for revenue growth, enable enterprises to develop more effective pricing strategies, optimize procurement processes and make smarter investment decisions by understanding market price trends.

[0005] As the core raw material of the new energy industry, lithium carbonate occupies a pivotal position in the development of lithium battery manufacturing and new energy vehicles. How to accurately predict the price of lithium carbonate is of vital importance for optimizing production scheduling, improving production efficiency, controlling production costs, and assisting corporate decision-making and management. The fluctuation of lithium carbonate prices is not only affected by multiple factors such as global lithium mineral resource mining, market demand changes, and policy orientation, but also more elusive due to the complexity of its industrial chain and the close connection between upstream and downstream industries. In view of these challenges, coupled with many uncertain factors in the lithium carbonate production process, there is a relative lack of research on lithium carbonate price prediction, and few scholars can fully and deeply explore this issue. Therefore, strengthening the research on lithium carbonate price prediction and exploring more accurate and efficient prediction methods are of great significance to promoting the sustainable development of the new energy industry and enhancing the market competitiveness of enterprises. Summary of the invention

[0006] The purpose of the present invention is to provide a lithium carbonate price prediction method based on the collaborative drive of multi-source and multi-modal data, which can assist purchasing personnel in production, formulate purchasing plans, and make full use of production resources through deep learning methods, thereby achieving the goal of reducing costs and increasing efficiency of enterprises.

[0007] To achieve the above object, the present invention provides a lithium carbonate price prediction method based on multi-source multi-modal data collaborative driving, comprising the following steps:

[0008] S1. Construct a multimodal lithium carbonate price dataset;

[0009] S2, using a parameter optimization algorithm to determine the best parameter combination;

[0010] S3. construct a feature extraction model to extract the features of multimodal lithium carbonate prices;

[0011] S4. Construct a cross-modal hybrid attention mechanism model to integrate the features of multi-modal lithium carbonate prices;

[0012] S5. Use the fused features to predict prices.

[0013] Preferably, the S1 comprises the following steps:

[0014] S1.1. Constructing a multimodal lithium carbonate index set Q = {q1, q2, ..., q i ,…,q N}, where q i represents the i-th index, and Among them, s i,n Represents the i-th index q i The nth parameter in, n∈[1,N i ],N i Indicates qi Parameter quantity, i∈[1,N], N represents the indicator quantity of the indicator set;

[0015] S1.2. Construct a text set T = {t1, t2, ..., t i ,…,t N}, t i represents the i-th text information, and Among them, w i,m Indicates t i The mth word in, m∈[1,M i ],M i Indicates t i The number of words;

[0016] S1.3, for t i Perform word embedding to get word vector X i ={x i,1 ,x i,2 ,…,x i,k ,…,x i,k}, where x i,k Represents the i-th text information t i The kth word vector in , k∈[1,K], K represents the length of the word vector, so as to obtain the word vector set X={X1,X2,…,X i ,…,X N}, X i Represents the i-th word vector.

[0017] Preferably, S2 comprises the following steps:

[0018] S2.1. Use the DMS-PSO algorithm to optimize the set Q and X parameters respectively, divide the entire particle population into multiple sub-populations, and the evolution of each particle is carried out in its own sub-population. The velocity vector v i,j and the position vector x i,j The update method is:

[0019] v i,j =w·v i,j +c1·r 1,j ·(pb i,j -x i,j )+c2·r 2,j ·(gb j -x i,j );

[0020] x i,j =x i,j +v i,j ;

[0021] Among them, w is the inertia weight, c1 and c2 are learning factors, pb i,j and gb j Represent the individual historical optimal solution and the optimal particle position vector of the sub-population where particle i is located, r 1,j and r 2,j is a random number in the interval [0,1], x i,j and v i,j It represents the position vector and velocity vector of the i-th particle in the j-th dimension.

[0022] Preferably, S3 comprises the following steps:

[0023] S3.1. Extract features from the indicator set Q based on the BiGRU model and obtain the indicator feature representation V Q = {V Q,1 ,V Q,2 ,…,V Q,i ,…,V Q,N}; where V Q,i represents the i-th indicator feature;

[0024] S3.2. Extract features from the word vector set X based on the TCN model to obtain the text feature representation V T = {V T,1 ,V T,2 ,…,V T,j ,…,V T,N}; where V T,j represents the jth text feature.

[0025] Preferably, S4 comprises the following steps:

[0026] S4.1. Use channel attention mechanism to process text data, that is, use F to represent V Q ,but represents the activation value of the kth channel (x, y) position in the lth layer, w k represents the attention weight of the kth channel, then the convolutional layer feature map of the (l+1)th layer is expressed as:

[0027]

[0028] S4.2, the prediction branch output based on channel attention is expressed as:

[0029]

[0030] Among them, P() represents mean pooling, and σ() is expressed as:

[0031]

[0032] S4.3. Use the cross entropy loss function to calculate the average error between the prediction branch output S and the true label Y on the entire data set, which is specifically expressed as:

[0033]

[0034] Where N is the total number of samples; S i and Y i are the output and true label of the i-th sample respectively;

[0035] S4.4. Use the autocorrelation attention mechanism to process data-type data, that is, use n to represent the data length and W to represent the data period, then the data dimension can be transformed into The autocorrelation attention A between two period features is expressed as:

[0036]

[0037] Among them, A(l, j) represents the attention weight coefficient between the lth cycle and the jth cycle, F(l, k) and F(j, k) represent the kth element of the lth and jth cycles respectively;

[0038] S4.5. Merge the channel attention and autocorrelation attention mechanisms through matrix multiplication and convolution, using Y － Represents the fused features:

[0039]

[0040] S4.6. Use the cross entropy loss function to calculate the difference between feature anomalies and true labels, expressed as:

[0041]

[0042] Among them, Y －i is the feature label of the i-th sample;

[0043] S4.7. Optimize the entire feature fusion process by minimizing the sum of the two loss function values, expressed as:

[0044] min(λ1L1+λ2L2);

[0045] Among them, λ1 and λ2 are the weights of the two components of the loss function respectively.

[0046] Preferably, S5 comprises the following steps:

[0047] S5.1. Fusion feature Y － Input into the fully connected layer to perform demand forecasting and obtain the forecast results

[0048] S5.2. A lithium carbonate price prediction model is constructed by a feature extraction model, an attention mechanism model and a fully connected layer. The mean square logarithmic error is used as the objective function J. The objective function J is trained by back propagation and gradient descent method to minimize the objective function J. When the number of iterations reaches the maximum number of iterations, the training is stopped to obtain a trained lithium carbonate price prediction model for predicting the price of lithium carbonate.

[0049] Therefore, the present invention adopts the above-mentioned lithium carbonate price prediction method based on the collaborative drive of multi-source and multi-modal data, which can assist purchasing personnel in production, formulate purchasing plans, and make full use of production resources through deep learning methods, thereby achieving the goal of reducing costs and increasing efficiency of enterprises.

[0050] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a flow chart of a lithium carbonate price prediction method based on multi-source multi-modal data collaborative drive of the present invention;

[0052] Figure 2 The present invention discloses a lithium carbonate price prediction method based on the collaborative driving of multi-source and multi-modal data, and constructs a multi-modal lithium carbonate price fusion feature flow chart. DETAILED DESCRIPTION

[0053] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0054] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0055] Embodiment 1

[0056] like Figure 1 As shown, a lithium carbonate price prediction method based on multi-source multi-modal data collaborative driving includes the following steps:

[0057] S1. Construct a dataset of multimodal lithium carbonate prices.

[0058] S1.1. Constructing a multimodal lithium carbonate index set Q = {q1, q2, ..., q i ,…,q N}, where q i represents the i-th index, and Among them, s i,n Represents the i-th index q i The nth parameter in, n∈[1,N i ],N i Indicates qi Parameter quantity, i∈[1,N], N represents the indicator quantity of the indicator set;

[0059] S1.2. Construct a text set T = {t1, t2, ..., t i ,…,t N}, t i represents the i-th text information, and Among them, w i,m Indicates t i The mth word in, m∈[1,M i ],M i Indicates t i The number of words;

[0060] S1.3, for t i Perform word embedding to get word vector X i ={x i,1 ,x i,2 ,…,x i,k ,…,x i,k}, where x i,k Represents the i-th text information t i The kth word vector in , k∈[1,K], K represents the length of the word vector, so as to obtain the word vector set X={X1,X2,…,X i ,…,X N}, X i Represents the i-th word vector.

[0061] S2. Use parameter optimization algorithm to determine the best parameter combination.

[0062] S2.1. Use the DMS-PSO algorithm to optimize the set Q and X parameters respectively, divide the entire particle population into multiple sub-populations, and the evolution of each particle is carried out in its own sub-population. The velocity vector v i,j and the position vector x i,j The update method is:

[0063] v i,j =w·v i,j +c1·r 1,j ·(pb i,j -x i,j )+c2·r 2,j ·(gb j -x i,j ) Formula (1);

[0064] x i,j =x i,j +v i,j Formula (2);

[0065] Among them, w is the inertia weight, c1 and c2 are learning factors, pb i,j and gb j Represent the individual historical optimal solution and the optimal particle position vector of the sub-population where particle i is located, r 1,j and r 2,j is a random number in the interval [0,1], x i,j and v i,j It represents the position vector and velocity vector of the i-th particle in the j-th dimension.

[0066] In the PSO algorithm, each particle represents a candidate solution to the problem, and the flight process of the entire population is regarded as the search process of the algorithm. The intelligence of problem solving is achieved through the simple behavior of the particles themselves and the information exchange between the population. In the DMS-PSO algorithm, the population is divided into multiple sub-populations, and the evolution of each particle is carried out separately in its respective sub-population. The division of sub-populations in the DMS-PSO algorithm is dynamic, and all particles will be in different sub-populations at different stages of evolution. Therefore, this strategy greatly promotes information exchange between particles and avoids the premature convergence problem caused by the excessive size of the PSO algorithm population.

[0067] S3. Construct a feature extraction model to extract the features of multimodal lithium carbonate prices.

[0068] S3.1. Extract features from the indicator set Q based on the BiGRU model and obtain the indicator feature representation V Q = {V Q,1 ,V Q,2 ,…,V Q,i ,…,V Q,N}; where V Q,i represents the i-th indicator feature;

[0069] BiGRU is a bidirectional recurrent network composed of two GRUs with opposite propagation directions. Because the price of lithium carbonate is not only related to factors related to the past, but also affected by other factors in the future, BiGRU is proposed to extract the feature dependency relationship between the price of lithium carbonate before and after. Information mining is performed on the input data in both sequential and reverse order, and the output of GRU units at different times is added to form the output layer. BiGRU can process sequence data in an efficient and comprehensive way, especially in scenarios where it is necessary to capture bidirectional contextual dependencies.

[0070] S3.2. Extract features from the word vector set X based on the TCN model to obtain the text feature representation V T = {V T,1 ,V T,2 ,…,V T,j ,…,V T,N}; where VT,j represents the jth text feature;

[0071] TCN is derived from the convolutional architecture of sequential data. Each convolutional layer has a causal relationship, and any element in the output sequence can only establish a dependency relationship with the previous element in the previous sequence. TCN can convert input time series of any length into a fixed-length sequence, and mainly uses one-dimensional full convolution to process time series, feeding all the information of the input sequence to the model to generate an output sequence of the same length as the input. TCN, with its clever combination of causal convolution and dilated convolution, demonstrates high efficiency, long-distance dependency capture capability, stability, and good trainability in processing sequential data.

[0072] S4. When performing modal fusion, the classic attention mechanism only assigns different weights according to the importance of different features in the modality, but fails to take into account the interaction between different modalities and the weight allocation and relationship modeling of internal features of the data. Therefore, two solutions are proposed to address this defect.

[0073] 1. The channel attention mechanism is used to dynamically adjust the weights of different channels (or feature dimensions) so that the model can focus on the most critical feature combination for the task. It is not limited to a single modality, but can be extended to multimodal scenarios. By calculating the channel weights between different modalities, cross-modal feature selection and fusion can be achieved, thereby further improving the fusion effect of multimodal data.

[0074] Second, the autocorrelation attention mechanism breaks the time step limit of the traditional sequence model and captures the global dependency by directly calculating the correlation between any elements in the sequence. This not only enhances the model's ability to capture long-distance dependencies, but also enables a more comprehensive understanding of the contextual information of the sequence. This helps reduce the loss of information during the fusion process and generates a richer and more comprehensive multimodal representation.

[0075] like Figure 2 As shown, based on the cross-modal attention mechanism, a cross-modal hybrid attention mechanism model is constructed to integrate the characteristics of multi-modal lithium carbonate prices.

[0076] S4.1. Use channel attention mechanism to process text data, that is, use F to represent V Q ,but represents the activation value of the k-th channel (x, y) position of the l-th layer, wk represents the attention weight of the k-th channel, and the convolutional layer feature map of the (l+1)-th layer is expressed as:

[0077]

[0078] S4.2, the prediction branch output based on channel attention is expressed as:

[0079]

[0080] Among them, P() represents mean pooling, and σ() is expressed as:

[0081]

[0082] S4.3. Use the cross entropy loss function to calculate the average error between the prediction branch output S and the true label Y on the entire data set, which is specifically expressed as:

[0083]

[0084] Where N is the total number of samples; S i and Y i are the output and true label of the i-th sample respectively.

[0085] S4.4. Use the autocorrelation attention mechanism to process data-type data, that is, use n to represent the data length and W to represent the data period, then the data dimension can be transformed into The autocorrelation attention A between two period features is expressed as:

[0086]

[0087] Among them, A(l, j) represents the attention weight coefficient between the lth cycle and the jth cycle, and F(l, k) and F(j, k) represent the kth element of the lth and jth cycles respectively.

[0088] S4.5. Merge the channel attention and autocorrelation attention mechanisms through matrix multiplication and convolution, using Y － Represents the fused features:

[0089]

[0090] S4.6. Use the cross entropy loss function to calculate the difference between feature anomalies and true labels, expressed as:

[0091]

[0092] Among them, Y －i is the feature label of the i-th sample

[0093] S4.7. Optimize the entire feature fusion process by minimizing the sum of the two loss function values, expressed as:

[0094] min(λ1L1+λ2L2) Formula (10);

[0095] Among them, λ1 and λ2 are the weights of the two components of the loss function respectively.

[0096] S5. Use the fused features to predict prices:

[0097] S5.1. Fusion feature Y － Input into the fully connected layer to perform demand forecasting and obtain the forecast results

[0098] S5.2. A lithium carbonate price prediction model is constructed by a feature extraction model, an attention mechanism model and a fully connected layer. The mean square logarithmic error is used as the objective function J, and the objective function J is trained by back propagation and gradient descent method to minimize the objective function J. When the number of iterations reaches the maximum number of iterations, the training is stopped to obtain the trained lithium carbonate price prediction model for predicting the price of lithium carbonate.

[0099] Therefore, the present invention adopts the above-mentioned lithium carbonate price prediction method based on the collaborative drive of multi-source and multi-modal data, which can assist purchasing personnel in production, formulate purchasing plans, and make full use of production resources through deep learning methods, thereby achieving the goal of reducing costs and increasing efficiency of enterprises.

[0100] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A lithium carbonate price prediction method based on multi-source multi-modal data collaborative drive, characterized in that: The following steps are involved: S1. Construct a multimodal lithium carbonate price dataset; S2, using a parameter optimization algorithm to determine the best parameter combination; S3. construct a feature extraction model to extract the features of multimodal lithium carbonate prices; S4. Construct a cross-modal hybrid attention mechanism model to integrate the features of multi-modal lithium carbonate prices; S5. Use the fused features to predict prices.

2. A lithium carbonate price prediction method based on multi-source multi-modal data collaborative drive according to claim 1, characterized in that: The S1 comprises the following steps: S1.

1. Constructing a multimodal lithium carbonate index set Q = {q1, q2, ..., q i ,…,q N }, where q i represents the i-th index, and Among them, s i,n Represents the i-th index q i The nth parameter in, n∈[1,N i ],N i Indicates q i Parameter quantity, i∈[1,N], N represents the indicator quantity of the indicator set; S1.

2. Construct a text set T = {t1, t2, ..., t i ,…,t N }, t i represents the i-th text information, and Among them, w i,m Indicates t i The mth word in, m∈[1,M i ],M i Indicates t i The number of words; S1.3, for t i Perform word embedding to get word vector X i ={x i,1 ,x i,2 ,…,x i,k ,…,x i,k }, where x i,k Represents the i-th text information t i The kth word vector in , k∈[1,K], K represents the length of the word vector, so as to obtain the word vector set X={X1,X2,…,X i ,…,X N }, X i Represents the i-th word vector.

3. A lithium carbonate price prediction method based on multi-source multi-modal data collaborative drive according to claim 1, characterized in that: The S2 comprises the following steps: S2.

1. Use the DMS-PSO algorithm to optimize the set Q and X parameters respectively, divide the entire particle population into multiple sub-populations, and the evolution of each particle is carried out in its own sub-population. The velocity vector v i,j and the position vector x i,j The update method is: v i,j =w·v i,j +c1·r 1,j ·(pb i,j -x i,j )+c2·r 2,j ·(gb j -x i,j ); x i,j =x i,j +v i,j ; Among them, w is the inertia weight, c1 and c2 are learning factors, pb i,j and gb j Represent the individual historical optimal solution and the optimal particle position vector of the sub-population where particle i is located, r 1,j and r 2,j is a random number in the interval [0,1], x i,j and v i,j It represents the position vector and velocity vector of the i-th particle in the j-th dimension.

4. A lithium carbonate price prediction method based on multi-source multi-modal data collaborative drive according to claim 1, characterized in that: The S3 comprises the following steps: S3.

1. Extract features from the indicator set Q based on the BiGRU model and obtain the indicator feature representation V Q = {V Q,1 ,V Q,2 ,…,V Q,i ,…,V Q,N }; where V Q,i represents the i-th indicator feature; S3.

2. Extract features from the word vector set X based on the TCN model to obtain the text feature representation V T = {V T,1 ,V T,2 ,…,V T,j ,…,V T,N }; where V T,j represents the jth text feature.

5. The lithium carbonate price prediction method based on multi-source multi-modal data collaborative driving according to claim 1, characterized in that: The S4 comprises the following steps: S4.

1. Use channel attention mechanism to process text data, that is, use F to represent V Q ,but represents the activation value of the kth channel (x, y) position in the lth layer, w k represents the attention weight of the kth channel, then the convolutional layer feature map of the (l+1)th layer is expressed as: S4.2, the prediction branch output based on channel attention is expressed as: Among them, P() represents mean pooling, and σ() is expressed as: S4.

3. Use the cross entropy loss function to calculate the average error between the prediction branch output S and the true label Y on the entire data set, which is specifically expressed as: Where N is the total number of samples; S i and Y i are the output and true label of the i-th sample respectively; S4.

4. Use the autocorrelation attention mechanism to process data-type data, that is, use n to represent the data length and W to represent the data period, then the data dimension can be transformed into The autocorrelation attention A between two period features is expressed as: Among them, A(l, j) represents the attention weight coefficient between the lth cycle and the jth cycle, F(l, k) and F(j, k) represent the kth element of the lth and jth cycles respectively; S4.

5. Merge the channel attention and autocorrelation attention mechanisms through matrix multiplication and convolution, using Y － Represents the fused features: S4.

6. Use the cross entropy loss function to calculate the difference between feature anomalies and true labels, expressed as: Among them, Y －i is the feature label of the i-th sample; S4.

7. Optimize the entire feature fusion process by minimizing the sum of the two loss function values, expressed as: min(λ1L1+λ2L2); Among them, λ1 and λ2 are the weights of the two components of the loss function respectively.

6. The lithium carbonate price prediction method based on multi-source multi-modal data collaborative driving according to claim 1, characterized in that: The S5 comprises the following steps: S5.

1. Fusion feature Y － Input into the fully connected layer to perform demand forecasting and obtain the forecast results S5.

2. A lithium carbonate price prediction model is constructed by a feature extraction model, an attention mechanism model and a fully connected layer. The mean square logarithmic error is used as the objective function J. The objective function J is trained by back propagation and gradient descent method to minimize the objective function J. When the number of iterations reaches the maximum number of iterations, the training is stopped to obtain a trained lithium carbonate price prediction model for predicting the price of lithium carbonate.

Images