New energy power key raw material demand prediction method fusing generative AI emotion reasoning features

Abstract

The invention discloses a new energy power key raw material demand prediction method fusing generative AI emotional reasoning features, and the method comprises the steps: 1, obtaining new energy power key raw material demand data from a material industry website, and obtaining text data through a large language model; 2, acquiring an emotion score for the text data by utilizing a large language model; 3, respectively extracting demand features and text features of new energy power key raw materials by using the two models; 4, constructing a demand prediction system fusing the generative AI emotion reasoning features; and 6, inputting the fusion features into a full connection layer for prediction to obtain a new energy power key raw material demand prediction value. According to the new energy power key raw material demand prediction method, the multi-modal data of the new energy power key raw materials and the generative AI emotion reasoning features are fused, multiple deep learning models are innovatively combined, multi-dimensional feature extraction and emotion reasoning of text type data and numerical type data are achieved, and the new energy power key raw material demand prediction precision is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of computer integrated prediction technology, and in particular provides a demand prediction method for key raw materials of new energy power that integrates generative AI emotion inference features. Background Art

[0002] With the booming development of the new energy vehicle industry and the continuous progress of battery technology, the market shows a relatively high demand for new energy power raw materials. As key materials, ternary precursors and electrolytic copper are widely used in the materials of lithium-ion batteries, especially showing great application potential in fields such as electric vehicles, energy storage systems, portable electronic devices, and aerospace. In the production process of ternary precursors, salts / oxides of nickel, cobalt, and manganese are used as raw materials, mixed in proportion and then subjected to calcination or ball milling pretreatment, and then a synthesis reaction is carried out in a reaction kettle through solvents and reactants under specific temperature and pressure conditions. The entire process requires strict control of process parameters to ensure the product purity, and finally, after post-treatment such as washing, drying, and screening, a product with a specific crystal structure and chemical composition is obtained to meet the performance requirements of the cathode material of lithium-ion batteries. The substance produced in this way is called "ternary precursor". The crude copper anode and pure copper cathode sheet are immersed in the copper sulfate electrolysis system. After power-on, the anode copper undergoes ionization migration, and electron exchange and directional deposition are completed at the cathode interface, and finally a high-purity metal copper plate is formed, called "electrolytic copper".

[0003] In order to ensure the high accuracy and reliability of the demand prediction of new energy power raw materials, the prediction work must comprehensively consider its complex influencing factors. However, the demand for new energy power raw materials is affected by various factors, including the development trend of the new energy vehicle industry, the progress of battery technology, etc. These factors are intertwined, making the demand fluctuations of new energy power raw materials full of uncertainties. In addition, the collection and analysis of market data for new energy power raw materials also face many challenges, because the data sources are extensive and complex, involving multiple stakeholders such as battery manufacturers, automobile manufacturers, and raw material suppliers. This makes it difficult to guarantee the accuracy and timeliness of the data, further increasing the difficulty of demand prediction. At the same time, the time series data of the demand for new energy power raw materials also shows large fluctuations, which undoubtedly brings greater challenges to the demand prediction work.

[0004] In the field of demand forecasting, although traditional statistical methods can construct prediction models with a certain degree of interpretability, they have limitations in terms of prediction accuracy and are difficult to meet the high-standard prediction requirements. This is mainly attributed to two aspects: on the one hand, these methods often rely on specific assumptions, which usually do not fully conform to the characteristics of actual data, resulting in deviations between the prediction results and the actual situation; on the other hand, these methods have difficulties in dealing with non-linear data and are unable to effectively capture the dynamic changes in the market, leading to biases in the prediction results. At the same time, machine learning methods also have some limitations in the field of demand forecasting. Although machine learning methods are good at processing large-scale data sets, they are sensitive to abnormal data, have limited generalization ability, and the constructed models have weak interpretability. In addition, existing machine learning methods often ignore emotional factors when making demand forecasts, making it difficult for the models to accurately capture changes in subjective factors such as consumer sentiment, market trends, or brand reputation, resulting in deviations between the prediction results and the actual demand. Therefore, both traditional statistical methods and machine learning methods have certain limitations in the demand forecasting of new energy power raw materials. Summary of the Invention

[0005] The present invention is proposed to solve the above-mentioned deficiencies of the existing technologies, and provides a method for predicting the demand for key raw materials of new energy power by integrating generative AI emotion inference features, aiming to achieve multi-dimensional feature extraction and emotion inference for text data and numerical data, and significantly improve the accuracy of predicting the demand for key raw materials of new energy power.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A method for predicting the demand for key raw materials of new energy power by integrating generative AI emotion inference features is characterized by the following steps:

[0008] Step 1: Construct a data set for the demand of key raw materials of new energy power;

[0009] Step 1.1: According to the factors affecting the demand for key raw materials of new energy power, use a large language model to generate text data X = {X1, X2,..., X i ,..., X N} for the demand of key raw materials of new energy power, where X i represents the i-th text information, and N represents the number of text information;

[0010] Step 1.2: Perform word embedding on X i to obtain the i-th word vector Y i = {y i,1 , y i,2 , …, y i,p , …yi,P}, where y i,p represents the p-th word vector in Y, p ∈ [1, P], and P represents the length of the word vector, thus obtaining the word vector set Y = {Y1, Y2,..., Y i},..., Y i ,..., Y N};

[0011] Step 1.3: Obtain the numerical data Q = {Q1, Q2,..., Q i ,..., Q N} of the key raw materials for new energy power, where Q i represents the numerical data of the i-th key raw material demand for new energy power, i ∈ [1, N];

[0012] Step 1.4: Use a large language model to perform fusion generative AI sentiment reasoning on the text data X to obtain the sentiment reasoning feature score set = {K1, K2,..., K i ,.., K N}, where K i corresponds to the sentiment feature score of X i ;

[0013] Step 1.5: According to the text data X and the numerical data Q, construct the index system T = {t1, t2,..., t i ,..., t N} of the key raw materials demand for new energy power, where t i represents the i-th index, and ;

[0014] Step 2: Extract the index feature set S of the index system T T , the text feature set S of the word vector set Y Y , and input them together with the sentiment reasoning feature score set into the multi-head attention mechanism model for fusion processing to obtain the index-text-sentiment reasoning fusion feature ;

[0015] Step 3: Construct a prediction network for the key raw materials demand of new energy power that fuses generative AI sentiment reasoning features, including: a feature extraction module, a Transformer module, and a fully connected layer, and input S F into the fully connected layer for demand prediction to obtain the predicted value D of the key raw materials demand for new energy power;

[0016] Step 4: Calculate the mean variance based on the predicted values D and Q and use it as the loss function J of the new energy power key raw material demand prediction network, so as to train the new energy power key raw material demand prediction network, minimize the objective function J to update the network parameters, and obtain the new energy power raw material demand prediction model with optimal parameters for predicting the new energy power key raw material demand.

[0017] The characteristics of a new energy power key raw material demand prediction method integrating generative AI emotion reasoning features according to the present invention also lie in that the step 2 includes the following steps:

[0018] Step 2.1: Construct a BiGRU model to extract features from the index system T to obtain the index feature set S of the new energy power key raw materials T ={S T,1 ,S T,2 ,…,S T,i ,…,S T,M}, where S T,i represents the i-th index feature, i ∈ [1, M];

[0019] Based on the TextCNN model, extract features from the word vector set Y to obtain the text feature set S of the new energy power key raw materials Y ={S Y,1 ,S Y,2 ,…,S Y,j ,…,S Y,N}, where S Y,j represents the j-th text feature, j ∈ [1, N];

[0020] Step 2.2: After splicing the index feature set , the text feature set and the emotion reasoning feature score set , input them into the multi-head attention mechanism model for fusion processing to obtain the index-text-emotion reasoning fusion feature .

[0021] Furthermore, the step 2.2 includes:

[0022] Step 2.2.1: Concatenate to form the sequence set input for the new energy power key raw material demand, and multiply them with the query coefficient matrix of the k-th attention head, the key-value coefficient matrix of the k-th attention head, and the value coefficient matrix of the k-th attention head respectively, and correspondingly obtain the k-th attention head multiplication result, and correspondingly obtain the k-th Query matrix of one attention head , key-value matrix and value matrix ;

[0023] Step 2.2.2: Obtain the attention weight features of the key raw materials for new energy power output by the th attention head using Equation (2) ;

[0024] (2)

[0025] In Equation (2), represents transpose, represents the activation function; represents 's dimension;

[0026] Step 2.2.3: Obtain the information interaction features of index-text-sentiment reasoning using Equation (3) ultiHead;

[0027] (3)

[0028] In Equations (1)-(3), H represents the total number of attention heads, ; concat represents concatenation; represents a learnable linear weight matrix;

[0029] Step 2.2.4: Obtain the normalized information interaction features using Equation (4) :

[0030] (4)

[0031] In Equation (4), Norm represents the normalization operation, represents the residual connection;

[0032] Step 2.2.5: Nonlinearly enhance using Equation (5):

[0033] + (5)

[0034] In Equation (5), FFN represents the feed-forward neural network, Gelu represents the activation function, and represent two weight matrices in the feed-forward neural network; and represent two bias terms in the feed-forward neural network;

[0035] Step 2.2.6. Obtain the index-text-sentiment inference fusion feature using Equation (6). :

[0036] (6)

[0037] In Equation (4) - Equation (5), represents the residual connection.

[0038] Furthermore, the said Step 3 includes:

[0039] Step 3.1. The feature extraction module processes S F using Equation (7) and Equation (8), and outputs the temporal-semantic fusion feature S FF ;

[0040] (7)

[0041] (8)

[0042] In Equation (7) and Equation (8), W c is the weight of the convolution kernel, b c is the bias term, is the one-dimensional convolution operation, is the feature tensor after convolution, is the bidirectional gated recurrent unit model, W g is the weight matrix of BiGRU, b g is the bias of the model;

[0043] Step 3.2. The Transformer module processes S FF using Equation (9), Equation (10), and Equation (11) to obtain the global dependence feature S TR :

[0044] EncMultiHead = concat(head1,…,head h ,…,head H )W O,enc (9)

[0045] Z = LayerNorm( + EncMultiHead) (10)

[0046] S TR = LayerNorm(Z + FFN enc (Z)) (11)

[0047] In Equation (9), Equation (10), and Equation (11), W O,encis the output projection matrix, EncMultiHead is the feature generated by the encoder multi-head attention mechanism, and W 1,enc and W 2,enc are the two weights of the feed-forward network, Z represents the normalized intermediate feature, and LayerNorm represents the layer normalization operation.

[0048] Step 3.3, the fully connected layer processes S TR using Equation (12) to obtain the predicted value D of the demand for key raw materials for new energy power:

[0049] (12)

[0050] In Equation (12), W fc1 and W fc2 are the two weights of the fully connected layer, b fc1 and b fc2 are the two bias terms of the fully connected layer, and GELU is the activation function.

[0051] An electronic device according to the present invention includes a memory and a processor, characterized in that the memory is used to store a program for supporting the processor to execute the method for predicting the demand for key raw materials for new energy power, and the processor is configured to execute the program stored in the memory.

[0052] A computer-readable storage medium according to the present invention, characterized in that when the computer program stored on the computer-readable storage medium is run by a processor, it executes the steps of the method for predicting the demand for key raw materials for new energy power.

[0053] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0054] 1. The present invention uses a cutting-edge cognitive large model to obtain text data and sentiment feature scores, ensuring data coherence while avoiding data defects and data quality problems caused by traditional data acquisition methods, such as web crawling.

[0055] 2. The present invention incorporates an attention mechanism to obtain fused features of different modalities, thereby improving the accuracy and robustness of the prediction of the demand for new energy power raw materials.

[0056] 3. The present invention combines deep learning and, through an efficient feature extraction technique, deeply obtains various features related to the demand for key raw materials for new energy power, thereby improving the prediction accuracy of the model for the demand for key raw materials for new energy power.

[0057] 4. The present invention incorporates emotional reasoning feature scores and innovatively includes them in the feature layer of the prediction model, making the structure of the new energy power key raw material demand prediction model more reasonable. While bringing higher prediction accuracy for the new energy power key raw material demand, it can reflect the impact of various environmental factors on demand fluctuations. Compared with the traditional method of constructing a mathematical model for prediction, the present invention can better grasp the fluctuations in demand and greatly improve the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a flowchart of the specific steps of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0059] In this embodiment, a method for predicting the demand for key raw materials of new energy power that integrates generative AI emotional reasoning features is as Figure 1 shown and is carried out according to the following steps:

[0060] Step 1: Construct a dataset for the demand of key raw materials for new energy power;

[0061] Step 1.1: According to the factors affecting the demand for key raw materials for new energy power, use a large language model to generate text data X = {X1, X2,..., X i ,..., X N} for the demand of key raw materials for new energy power, where X i represents the i-th text information and N represents the number of text information;

[0062] Step 1.2: Perform word embedding on X i to obtain the i-th word vector Y i = {y i,1 , y i,2 , …, y i,p , … y i,P}, where y i,p represents the p-th word vector in Y i , p ∈ [1, P], and P represents the length of the word vector, thereby obtaining the word vector set Y = {Y1, Y2,..., Y i ,..., Y N} of X;

[0063] Step 1.3: Obtain numerical data Q = {Q1, Q2,..., Q i ,..., Q N} for the demand of key raw materials for new energy power, where Q i represents the numerical data of the demand for the i-th key raw material for new energy power, i ∈ [1, N];

[0064] Step 1.4: Use a large language model to perform fusion generative AI sentiment reasoning on the text data X to obtain a set of sentiment reasoning feature scores ={K1, K2,..., K i ,.., K N}, where K i corresponds to the sentiment feature score of X i ;

[0065] Step 1.5: Based on the text data X and the numerical data Q, construct an index system T = {t1, t2,..., t i ,..., t N} for the demand of key raw materials for new energy power, where t i represents the i-th index, and .

[0066] Step 2: Extract the index feature set S T of the index system T, the text feature set S Y of the word vector set Y, and input them together with the sentiment reasoning feature score set into the multi-head attention mechanism model for fusion processing to obtain the index-text-sentiment reasoning fusion feature ;

[0067] Step 2.1: Construct a BiGRU model to extract features from the index system T to obtain the index feature set S T ={S T,1 , S T,2 , …, S T,i , …, S T,M} of the key raw materials for new energy power, where S T,i represents the i-th index feature, i ∈ [1, M];

[0068] Extract features from the word vector set Y based on the TextCNN model to obtain the text feature set S Y ={S Y,1 , S Y,2 , …, S Y,j , …, S Y,N} where S Y,j represents the j-th text feature, j ∈ [1, N];

[0069] BiGRU (Bidirectional Gated Recurrent Unit) is a recurrent neural network that introduces a bidirectional structure based on GRU (Gated Recurrent Unit). It consists of two independent GRU units, one processing data forward in time series and the other processing data backward in time series. The BiGRU model can capture both forward and backward information of sequential data through the bidirectional structure, improving the accuracy and stability of prediction. In addition, BiGRU reduces the number of gated units, lowering the model complexity and computational cost.

[0070] TextCNN is a convolutional neural network model for text classification. It extracts text features through multiple convolutional kernels of different sizes and combines pooling layers to reduce the feature dimension. Its advantages lie in its simple structure, few parameters, fast training speed, and the ability to effectively capture local features, showing good performance in tasks such as text classification.

[0071] Step 2.2, concatenate the metric feature set , the text feature set and the sentiment inference feature score set , and then input them into the multi-head attention mechanism model for fusion processing to obtain the metric-text-sentiment inference fusion feature .

[0072] Step 2.2.1, concatenate into the sequence set of new energy power key raw material demand inputs , and multiply them respectively with the query coefficient matrix of the th attention head, the key-value coefficient matrix of the th attention head, and the value coefficient matrix of the th attention head, respectively obtaining the query matrix , key-value matrix and value matrix of the th attention head;

[0073] Step 2.2.2, use Equation (2) to obtain the attention weight feature of the new energy power key raw materials output by the th attention head;

[0074] (2)

[0075] In Equation (2), represents transpose, represents the activation function; represents 's dimension;

[0076] Step 2.2.3: Obtain the information interaction features of index-text-sentiment inference using Equation (3) ultiHead;

[0077] (3)

[0078] In Equations (1)-(3), H represents the total number of attention heads, ; concat represents concatenation; represents a learnable linear weight matrix;

[0079] Step 2.2.4: Obtain the normalized information interaction features using Equation (4) :

[0080] (4)

[0081] In Equation (4), Norm represents the normalization operation, represents the residual connection;

[0082] Step 2.2.5: Nonlinearly enhance using Equation (5):

[0083] + (5)

[0084] In Equation (5), FFN represents the feed-forward neural network, Gelu represents the activation function, and represent two weight matrices in the feed-forward neural network; and represent two bias terms in the feed-forward neural network;

[0085] Step 2.2.6: Obtain the index-text-sentiment inference fusion features using Equation (6) :

[0086] (6)

[0087] In Equations (4)-(5), Norm represents the normalization operation, which helps to stabilize the training process and accelerate convergence; + represents the addition operation of the residual connection, and the residual connection helps to alleviate the vanishing gradient problem in deep neural networks; FFN represents the feed-forward neural network, which receives MultiHead’ as input and outputs a transformed feature vector, Gelu represents the activation function, which is used to increase the non-linear expression ability of the network; and 、 and respectively represent two weight matrices and two bias terms in the network.

[0088] Step 3: Construct a new energy - powered key raw material demand prediction network that fuses and generates AI emotion - inference features, including: a feature extraction module, a Transformer module, and a fully - connected layer, and input S F into the fully - connected layer for demand prediction to obtain the predicted value D of the new energy - powered key raw material demand;

[0089] Step 3.1: The feature extraction module processes S F using equations (7) and (8), and outputs the temporal - semantic fusion feature S FF ;

[0090] (7)

[0091] (8)

[0092] In equations (7) and (8), W c is the weight of the convolution kernel, b c is the bias term, is a one - dimensional convolution operation, is the feature tensor after convolution, is a bidirectional gated recurrent unit model, W g is the weight matrix of BiGRU, b g is the bias of the model.

[0093] Step 3.2: The Transformer module processes S FF using equations (9), (10), and (11) to obtain the global - dependency feature S TR :

[0094] EncMultiHead = concat(head1,…,head h ,…,head H )W O,enc (9)

[0095] Z = LayerNorm( + EncMultiHead) (10)

[0096] S TR = LayerNorm(Z + FFN enc (Z)) (11)

[0097] In equations (9), (10), and (11), W O,encis the output projection matrix, EncMultiHead is the feature generated by the encoder multi-head attention mechanism, and W 1,enc , W 2,enc are the two weights of the feed-forward network, Z represents the normalized intermediate feature, and LayerNorm represents the layer normalization operation. By eliminating the scale difference between features through LayerNorm, the problem of gradient disappearance in deep networks can be alleviated, and the convergence stability of the model can be improved.

[0098] Step 3.3, the fully connected layer processes S TR using Equation (12) to obtain the predicted value D of the demand for key raw materials for new energy power:

[0099] (12)

[0100] In Equation (12), W fc1 , W fc2 are the two weights of the fully connected layer, b fc1 , b fc2 are the two bias terms of the fully connected layer, and GELU is the activation function. The fully connected layer constructs a non-linear mapping space from abstract features to specific demand values. As shown in Equation (12), first, the feature S fc1 is projected into a high-dimensional hidden space through the weight matrix W fc1 and the bias term b TR , and a non-linear transformation is performed using the GELU activation function. Compared with the traditional ReLU function, the probability gating mechanism of the GELU activation function can better model the asymmetric noise distribution c existing in the prediction of raw material demand.

[0101] Step 4, according to the predicted value D and Q, calculate the mean and variance and use them as the loss function J of the new energy power key raw material demand prediction network, so as to train the new energy 2 power key raw material demand prediction network, minimize the objective function J to update the network parameters, and obtain the new energy power raw material demand prediction model with optimal parameters, which is used to process the input fusion features and predict the demand for new energy power key raw materials.

[0102] In this embodiment, an electronic device includes a memory and a processor. 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.

[0103] In this embodiment, a computer-readable storage medium stores a computer program on the computer-readable storage medium. When the computer program is run by the processor, it executes the steps of the above method.

Claims

1. A method for predicting the demand for key raw materials of new energy power that integrates generative AI emotion inference features, characterized in that, It is carried out according to the following steps: Step 1, construct a dataset of the demand for key raw materials of new energy power; Step 1.

1. According to the factors affecting the demand for key raw materials of new energy power, use the large language model to generate text data X = {X1, X2,..., X i ,..., X N} of the demand for key raw materials of new energy power, where X i represents the i-th text information, and N represents the number of text information; Step 1.

2. Embed X i to obtain the i-th word vector Y i ={y i,1 , y i,2 , …, y i,p , … y i,P}, where y i,p represents the p-th word vector in Y i , p ∈ [1, P], and P represents the length of the word vector, thus obtaining the word vector set Y = {Y1, Y2, ..., Y i , ..., Y N} of X; Step 1.

3. Obtain the numerical data Q = {Q1, Q2,..., Q i ,..., Q N}, where Q i represents the numerical data of the i-th key raw material requirement for new energy power, and i ∈ [1, N]; Step 1.4: Use a large language model to perform fusion generative AI sentiment reasoning on the text data X, and obtain a set of sentiment reasoning feature scores ={K1, K2,..., K i ,.., K N}, where K i corresponds to the sentiment feature score of X i ; Step 1.

5. Based on the text-type data X and the numerical-type data Q, construct an index system T = {t1, t2,..., t i ,..., t N} for the demand of key raw materials for new energy power, where t i represents the i-th index, and ; Step 2: Extract the index feature set S of the index system T T , the text feature set S of the word vector set Y Y , and input them together with the sentiment inference feature score set into the multi-head attention mechanism model for fusion processing to obtain the index-text-sentiment inference fusion feature ; Step 3: Construct a new energy power key raw material demand prediction network that integrates generative AI emotion inference features, including: a feature extraction module, a Transformer module, and a fully connected layer, and input S F into the fully connected layer for demand prediction to obtain the predicted value D of the new energy power key raw material demand; Step 4, according to the predicted values D and Q, calculate the mean variance and use it as the loss function J of the prediction network for the demand of key raw materials of new energy power, so as to train the prediction network for the demand of key raw materials of new energy power, minimize the objective function J to update the network parameters, and obtain a prediction model of the demand for new energy power raw materials with optimal parameters for predicting the demand for key raw materials of new energy power.

2. The new energy power key raw material demand prediction method integrating generative AI emotion inference features according to claim 1, wherein, The said Step 2 includes the following steps: Step 2.1: Construct a BiGRU model to extract features from the indicator system T, and obtain the indicator feature set S of key raw materials for new energy power T ={S T,1 , S T,2 , …, S T,i , …, S T,M}, where S T,i represents the i-th indicator feature, i ∈ [1, M]; Extract features from the word vector set Y based on the TextCNN model to obtain the text feature set S of key raw materials for new energy power Y ={S Y,1 ,S Y,2 ,…,S Y,j ,…,S Y,N} Among them, S Y,j represents the j-th text feature, j ∈ [1, N]; Step 2.

2. Concatenate the index feature set , the text feature set , and the sentiment inference feature score set , then input them into the multi-head attention mechanism model for fusion processing to obtain the index-text-sentiment inference fusion feature .

3. A new energy power key raw material demand prediction method integrating generative AI emotion inference features according to claim 1, characterized in that, The said Step 2.2 includes: Step 2.2.1: From to splice into a sequence set of new energy power key raw material demand inputs , and respectively multiply with the query coefficient matrix of the th attention head , the key-value coefficient matrix of the th attention head , and the value coefficient matrix of the th attention head to correspondingly obtain the query matrix of the th attention head, the key-value matrix , and the value matrix ; Step 2.2.

2. Obtain the attention weight features of the key raw materials for new energy power output by the th attention head using Equation (2); ​ (2) In formula (2), represents the transpose,[[]] represents the activation function; represents the dimension of; Step 2.2.3: Obtain the information interaction features of index-text-sentiment inference using Equation (3) ultiHead; (3) In formulas (1)-(3), H represents the total number of attention heads, ; concat represents concatenation; represents a learnable linear weight matrix; Step 2.2.4: Obtain the normalized information interaction features using Equation (4). : (4) In Equation (4), Norm represents the normalization operation, represents the residual connection; Step 2.2.5, use Equation (5) to perform non-linear enhancement: + (5) In Equation (5), FFN represents a feed - forward neural network, and Gelu represents an activation function. and represent two weight matrices in the feed - forward neural network; and represent two bias terms in the feed - forward neural network; Step 2.2.6: Obtain the index-text-sentiment inference fusion feature using Equation (6) : (6) In Formula (4) - Formula (5), represents a residual connection.

4. A new energy power key raw material demand prediction method integrating generative AI emotion inference features according to claim 3, characterized in that, The said Step 3 includes: Step 3.

1. The feature extraction module processes S using Equations (7) and (8) F and outputs the temporal-semantic fusion feature S FF . (7) (8) In equations (7) and (8), W c is the weight of the convolutional kernel, and b c is the bias term. is the one-dimensional convolution operation, is the feature tensor after convolution, is the bidirectional gated recurrent unit model, and W g is the weight matrix of the BiGRU, and b g is the bias of the model. Step 3.2: The Transformer module processes S using equations (9), (10), and (11). FF to obtain the global dependence feature S TR : EncMultiHead = concat(head1,…,head h ,…,head H )W O,enc (9) Z = LayerNorm( + EncMultiHead) (10) S TR = LayerNorm(Z + FFN enc (Z)) (11) In formulas (9), (10), and (11), W O,enc is the output projection matrix, EncMultiHead is the feature generated by the encoder multi-head attention mechanism, and W 1,enc , W 2,enc are the two weights of the feed-forward network, Z represents the normalized intermediate feature, and LayerNorm represents the layer normalization operation.

5. Step 3.3: The fully connected layer processes S using Equation (12). TR to obtain the predicted value D of the demand for key raw materials for new energy power. (12) In formula (12), W fc1 , W fc2 are two weights of the fully connected layer, b fc1 , b fc2 are two bias terms of the fully connected layer, and GELU is the activation function.

6. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program for supporting the processor to execute any one of the new energy power key raw material demand prediction methods in claims 1-4, and the processor is configured to execute the program stored in the memory.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it executes the steps of any one of the new energy power key raw material demand prediction methods in claims 1-4.

Images