A method and system for light stabilizer price forecasting

Abstract

The application belongs to the technical field of artificial intelligence, data processing and fine chemical supply chain management, and discloses a light stabilizer price prediction method and system, and the method steps comprise: collecting structured digital data and unstructured text data; performing NLP sentiment quantification on the unstructured text data to generate a market sentiment index; performing dynamic time window lag alignment on the structured digital data; fusing the aligned structured digital data and the market sentiment index into a feature matrix, inputting the feature matrix into a double-layer LSTM prediction model based on an attention mechanism to perform price prediction; and outputting a light stabilizer price prediction result. The application can solve the problems of traditional prediction methods, such as dependence on manual judgment, neglect of industry chain transmission lag and incapability of quantifying market sentiment.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary fields of artificial intelligence, data processing, and fine chemical supply chain management, and in particular relates to a method and system for predicting the price of light stabilizers. Background Technology

[0002] Light stabilizers are important additives in agricultural films, automotive coatings, engineering plastics and other fields. Their market price is affected by a combination of factors, including upstream raw material costs, downstream industry demand, industry policy regulation and supply chain emergencies. Price fluctuations exhibit highly non-linear and seasonal characteristics. Accurate price forecasting is crucial for inventory management and cost control for plastic additive companies.

[0003] However, current methods for predicting the price of light stabilizers still have many shortcomings in practical industrial applications and are difficult to adapt to the actual transmission patterns of the chemical industry chain and the complex changes in the market.

[0004] Traditional methods of judging the market for light stabilizers rely heavily on the "market feel" and implicit experience of senior product managers or industry experts. They are easily influenced by subjective judgments, and the relevant judgment logic is difficult to form a market analysis capability that can be inherited and replicated, resulting in inconsistent efficiency and accuracy of price judgments among companies.

[0005] Meanwhile, existing quantitative forecasting models typically align upstream raw material prices with downstream finished product prices directly as input. However, in real chemical production, there is a fixed physical time cycle from raw material procurement and synthesis reaction to finished product inventory. This neglect of "dynamic time lag" prevents the models from accurately capturing the lead time of price inflection points.

[0006] Furthermore, existing forecasting methods are mostly limited to structured digital indicators such as spot and futures prices, and cannot effectively handle unstructured text information that causes sharp market fluctuations, such as "environmental protection production restriction notices," "chemical industrial park shutdowns and rectifications," and "major supplier equipment failures." As a result, the predictive performance of the models drops significantly when faced with sudden market events, making it difficult to meet the actual needs of enterprises for price forecasting. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for predicting the price of light stabilizers, so as to solve the above-mentioned problems.

[0008] To achieve the above objectives, in a first aspect of the present invention, a method for predicting the price of light stabilizers is provided, comprising: The data collection includes both structured digital data and unstructured text data; The unstructured text data is subjected to NLP sentiment quantification to generate a market sentiment index; Dynamic time window lag alignment is performed on the structured digital data; The aligned structured digital data is fused with the market sentiment index to form a feature matrix, which is then input into a two-layer LSTM prediction model based on an attention mechanism for price prediction. Price forecast results for output light stabilizers.

[0009] Furthermore, in the aforementioned method for predicting the price of light stabilizers, the structured digital data includes upstream raw material prices, downstream industry operating rates, capacity utilization rates, inventory data, and macroeconomic data. The upstream raw material prices include the prices of o-nitro-p-chloroaniline, tert-butylphenol, and phenol, and the macroeconomic data include crude oil prices and exchange rate data.

[0010] Furthermore, in the aforementioned method for predicting the price of light stabilizers, the unstructured text data includes environmental inspection notices, production restriction announcements for chemical industrial parks, maintenance announcements for production facilities, and news related to supply and demand in the chemical industry.

[0011] Furthermore, in the aforementioned light stabilizer price prediction method, the NLP sentiment quantification uses the FinBERT or BERT model to perform sentiment polarity analysis on the unstructured text data, normalizing the market sentiment index to the range of [-1, 1], where 1 represents an extremely bullish market with shortages, and -1 represents an extremely bearish market with surpluses.

[0012] Furthermore, in the aforementioned method for predicting the price of light stabilizers, the dynamic time window lag alignment includes: calculating the correlation coefficient between each of the structured digital data and the price of light stabilizers under different time shifts using Granger causality tests and Pearson correlation coefficients, determining the conduction lag period corresponding to each of the structured digital data, and then performing time axis misalignment shifting on the structured digital data according to the conduction lag period.

[0013] Furthermore, in the light stabilizer price prediction method, before fusing the aligned structured digital data with the market sentiment index into a feature matrix, all numerical data need to be normalized to the [0,1] interval using Min-Max.

[0014] Furthermore, in the aforementioned method for predicting the price of light stabilizers, the step of inputting the feature matrix into a two-layer LSTM prediction model based on an attention mechanism for price prediction includes: The feature matrix is ​​input into the first LSTM layer to capture the long-term temporal dependence of light stabilizer prices and industry cycle patterns. Then input the second LSTM layer to extract the temporal nonlinear characteristics of light stabilizer prices; Through the attention mechanism layer, feature weights are dynamically assigned to the structured digital data and the market sentiment index based on the degree of influence of each feature on the current price of light stabilizers. The weighted features are processed by a fully connected output layer to output the predicted price of light stabilizers for the next cycle.

[0015] Furthermore, in the aforementioned light stabilizer price prediction method, the training and iterative optimization process of the attention-based two-layer LSTM prediction model includes: The mean square error or root mean square error is used as the loss function to calculate the deviation between the model's predicted price and the actual spot price of light stabilizer. An L2 regularization term is introduced into the loss function to prevent the model from overfitting. The Adam optimization algorithm is used to perform backpropagation with a preset initial learning rate to update the weight parameters and bias terms of the two-layer LSTM prediction model until the loss function converges to below the preset threshold or the number of iterations reaches the maximum number of iterations. Residual analysis is performed on the model prediction results. When the absolute value of the deviation between the predicted value and the actual value exceeds the set threshold in multiple consecutive prediction periods, a concept drift alarm is triggered, and the transmission lag period corresponding to the input features or the structured digital data is readjusted.

[0016] Furthermore, the light stabilizer price prediction method further includes generating business decision instructions based on the output light stabilizer price prediction results, including: The normalized light stabilizer price predictions output by the model are denormalized. The inversely normalized predicted price is smoothed using the exponential moving average method. Based on the fluctuation threshold between the smoothed predicted price and the current actual price of the light stabilizer, corresponding business decision instructions are generated.

[0017] In a second aspect of the invention, a light stabilizer price prediction system is also proposed, characterized in that it includes a data acquisition module, an NLP sentiment quantification module, a dynamic time window alignment module, an LSTM prediction module, and a result output module. The data acquisition module is used to acquire structured digital data and unstructured text data; The NLP sentiment quantification module is used to perform sentiment polarity analysis on the unstructured text data and generate a market sentiment index. The dynamic time window alignment module is used to determine the transmission lag period and time axis misalignment of the structured digital data, and to merge the aligned structured digital data with the market sentiment index into a feature matrix; The LSTM prediction module uses a two-layer LSTM prediction model based on an attention mechanism to predict prices from the feature matrix. It also performs loss assessment, iterative optimization, and residual analysis based on the model, triggers concept drift alarms, and performs readjustment of feature or propagation lag cycles. The result output module is used to output the price prediction results of light stabilizers.

[0018] Compared with the prior art, the present invention has at least the following technical effects: This invention overcomes the limitations of traditional quantitative models that rely solely on structured indicators by performing NLP sentiment quantification on unstructured text data, effectively quantifying market sentiment factors. By dynamically aligning structured digital data with time windows, it can reconstruct the true physical production and price transmission cycles of the chemical industry chain, thus compensating for the shortcomings of existing models that ignore time lags and cannot accurately capture price inflection points. A two-layer LSTM prediction model based on an attention mechanism is used to achieve price prediction based on multi-source fusion data, replacing the traditional subjective judgment method that relies on human experience. This makes the price prediction logic more objective and replicable, thereby improving the efficiency and accuracy of judgment. Attached Figure Description

[0019] Figure 1 This is a flowchart of a method for predicting the price of light stabilizers in one embodiment of the present invention; Figure 2 This is a diagram showing the overall architecture of the prediction system corresponding to the method in one embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the principle of the data preprocessing and dynamic time window lag alignment mechanism in one embodiment of the present invention; Figure 4 This is a flowchart illustrating the feature construction process that integrates NLP sentiment analysis in one embodiment of the present invention. Figure 5 This is a schematic diagram of a two-layer LSTM prediction model structure based on an attention mechanism in one embodiment of the present invention. Detailed Implementation

[0020] The following is a more detailed description of a method and system for predicting the price of light stabilizers according to the present invention, with reference to schematic diagrams illustrating preferred embodiments of the invention. It should be understood that those skilled in the art can modify the invention described herein while still achieving its advantageous effects. Therefore, the following description should be understood as being of general knowledge to those skilled in the art and is not intended to limit the invention.

[0021] For clarity, not all features of the actual embodiments are described. In the following description, well-known functions and structures are not detailed in detail, as they would obscure the invention with unnecessary detail. It should be understood that in the development of any actual embodiment, numerous implementation details must be made to achieve the developer's specific objectives, such as changes from one embodiment to another according to limitations related to the system or business. Furthermore, it should be understood that such development work may be complex and time-consuming, but is merely routine work for those skilled in the art.

[0022] The invention is described more specifically by way of example in the following paragraphs with reference to the accompanying drawings. The advantages and features of the invention will become clearer from the following description. It should be noted that the drawings are in a very simplified form and use non-precise proportions, and are only used to facilitate and clarify the illustration of the embodiments of the invention.

[0023] Based on the teachings of this specification, those skilled in the art can form new technical solutions through cross-combination of different implementation methods without creating technical contradictions. Such variations should all be considered to fall within the protection scope of this invention.

[0024] This embodiment discloses a method for predicting the price of light stabilizers. This method effectively solves the technical problems of traditional prediction methods, such as reliance on manual judgment, neglect of supply chain lags, and inability to quantify market sentiment. It achieves high-precision price prediction for all categories of light stabilizers, including benzotriazoles and hindered amines, and outputs standardized business decision-making instructions. (Refer to...) Figure 1 The specific steps of this method are as follows: S1: The collection includes structured digital data and unstructured text data; S2: Perform NLP sentiment quantification on the unstructured text data to generate a market sentiment index; S3: Perform dynamic time window lag alignment on the structured digital data; S4: The aligned structured digital data is fused with the market sentiment index to form a feature matrix, which is then input into a two-layer LSTM prediction model based on an attention mechanism for price prediction. S5: Price forecast results for output light stabilizers.

[0025] It should be noted that this embodiment uses the benzotriazole light stabilizer UV326 as a preferred example for illustration. Other types of light stabilizers can be quickly reused by modifying the adaptation parameters of this method. In this embodiment, the prediction period is set to weekly, that is, predicting the spot price of the target light stabilizer for the following week.

[0026] For step S1, the collected structured digital data consists of directly quantifiable numerical data, including upstream raw material prices, downstream industry operating rates, capacity utilization rates, inventory data, and macroeconomic data. The specific data collected is as follows: The system collects upstream raw material prices by gathering daily spot prices and near-month main futures contract prices for o-nitro-p-chloroaniline, tert-butylphenol, and phenol. It also collects weekly spot prices for UV326 production auxiliary materials such as liquid caustic soda and toluene. The data is connected to commodity databases such as Wind and Zhuochuang Information via API interfaces to achieve real-time synchronous updates.

[0027] Downstream industry data is collected by collecting weekly operating rates of UV326 downstream application fields such as domestic agricultural film, automotive coatings, and engineering plastics, as well as the apparent consumption of major domestic polyolefin (PP, PE) enterprises. The data is obtained by crawling the statistical weekly or monthly reports released by industry associations such as the China Plastics Processing Industry Association at regular intervals, and supplemented by customs import and export trade data.

[0028] Capacity utilization and inventory data were collected by collecting weekly capacity utilization and actual inventory levels of major domestic UV326 manufacturers. The data came from industry surveys and capacity statistics reports from chemical industrial parks.

[0029] Macroeconomic data is collected by monitoring daily spot prices of Brent crude oil and the USD / CNY exchange rate. The data is captured in real time through a financial data platform. Crude oil prices reflect fluctuations in the cost of basic chemical raw materials, while exchange rates affect the transmission of imported raw material prices.

[0030] In this embodiment, all structured digital data are collected weekly to form a time series dataset, with the time window set to the past 60 months to ensure that the data has sufficient historical cyclical patterns for reference.

[0031] Furthermore, the unstructured text data consists of textual information, including environmental inspection notices, production restriction announcements from chemical industrial parks, maintenance announcements for production facilities, and news related to supply and demand in the chemical industry. This text data is crawled using distributed web crawlers built with Scrapy or Selenium, which monitor news updates on specific chemical industry portals. The text data is then aggregated and deduplicated weekly, retaining only the valid information.

[0032] For step S2, refer to Figure 4 The NLP sentiment quantification includes using FinBERT (a pre-trained language model in the financial field) or BERT model to perform sentiment polarity analysis on the unstructured text data and normalize the market sentiment index to the range of [-1,1].

[0033] As a preferred embodiment, NLP sentiment quantification employs the FinBERT model. This model has been specifically fine-tuned for sentiment analysis of financial and industry information, and it achieves higher accuracy in identifying the sentiment polarity of chemical industry policies and supply-related texts. The specific processing steps are as follows: Text preprocessing: Clean the unstructured text captured by the crawler, remove irrelevant advertisements, duplicate content, and garbled characters, and complete text segmentation, word segmentation, and stop word removal to form a standardized text library; Sentiment polarity analysis: The preprocessed text is input into the fine-tuned FinBERT model. The model performs semantic parsing on the text, determines its tendency to influence the UV326 market price (bullish, bearish, neutral), and outputs an initial sentiment score. Sentiment Index Normalization: The initial sentiment score output by the model is normalized and mapped to the interval [-1, 1]. A normalized value of 1 indicates that the market is extremely bullish and products are in short supply; a value of -1 indicates that the market is extremely bearish and products are in oversupply; and a value of 0 indicates that the market sentiment is neutral and supply and demand are stable.

[0034] In this preferred embodiment, if the text information "A UV326 manufacturer in East China temporarily suspended production due to environmental inspection requirements, which is expected to affect 30% of its capacity" is captured, the model outputs an initial sentiment score of 0.92, and after normalization, the market sentiment index for that week is 0.85. If the text information "A new UV326 production line in China officially started production, adding 10,000 tons of annual capacity" is captured, the model outputs an initial sentiment score of -0.75, and after normalization, the market sentiment index for that week is -0.60. One market sentiment index value is generated each week, consistent with the time dimension of the structured digital data, forming a market sentiment index time series.

[0035] Step S3 addresses the industry characteristic of a physical cycle lag in the transmission of raw material price changes to UV326 spot prices within the chemical industry chain. It involves aligning the structured digital data along a timeline to reconstruct the true price transmission logic. The specific process is as follows: Figure 3 As shown, where, Figure 3 -A represents the existing vertical alignment method, which directly aligns upstream raw material prices and market sentiment with downstream finished product prices in the same period, ignoring the transmission lag. Figure 3 -B represents the misalignment alignment method of this approach. Based on the physical transmission cycles T1 and T2, the prices of upstream raw materials and intermediates are shifted along the time axis to achieve precise alignment of the price transmission logic.

[0036] In this embodiment, the dynamic time window lag alignment includes: calculating the correlation coefficient between each structured digital data and the price of light stabilizer under different time shifts using Granger causality test and Pearson correlation coefficient, determining the conduction lag period corresponding to each structured digital data, and then performing time axis misalignment shift on the structured digital data according to the conduction lag period.

[0037] Specifically, the structured digital data collected in step S1 are used as independent variables, and the historical spot price of UV326 is used as the dependent variable. The Granger causality relationship and Pearson correlation coefficient between each independent variable and the dependent variable under different time shifts (such as shifts of 1 day, 1 week, 2 weeks, and 3 weeks) are calculated respectively. The time shift value corresponding to the largest correlation coefficient is selected and determined as the transmission lag period corresponding to the structured digital data.

[0038] In this embodiment, statistical calculations show that the transmission lag period for the prices of core upstream raw materials such as phenol and tert-butylphenol is 2 weeks, the transmission lag period for crude oil prices is 2 weeks, the transmission lag period for the operating rate of downstream agricultural film is 3 weeks, and the transmission lag period for the capacity utilization rate of UV326 is 1 week.

[0039] Furthermore, based on the transmission lag periods of the structured digital data obtained above, the structured digital data is shifted backward along the time axis. If this embodiment aims to predict the UV326 price in week N, then the phenol price in week N-2, the crude oil price in week N-2, the downstream agricultural film operating rate in week N-3, and the capacity utilization rate in week N-1 are used as model input data to achieve accurate alignment between the structured digital data and the UV326 price in the transmission logic.

[0040] Structured digital data aligned by dynamic time window lag can eliminate spurious correlations caused by temporal misalignment, ensuring that subsequent model training learns the true causal transmission logic of the industrial chain, rather than superficial data fitting.

[0041] For step S4, before fusing the aligned structured digital data with the market sentiment index into a feature matrix, all numerical data need to be normalized to the [0,1] interval using Min-Max normalization to eliminate the dimensional differences between different data (e.g., raw material prices are in yuan / ton, operating rates are in %, and the sentiment index is dimensionless), thus avoiding gradient explosion during model training due to dimensional differences. The formula for calculating Min-Max normalization is: Among them, X norm The data is normalized, and X is the original data. min X max These are the minimum and maximum values ​​of the historical data for this feature, respectively.

[0042] After normalization, all aligned structured digital data features and market sentiment index features are concatenated in the format of "column = data feature, row = prediction sample" to construct a fused feature matrix X. t .

[0043] In this embodiment, the column features of the fusion feature matrix include: phenol price, tert-butylphenol price, o-nitro-p-chloroaniline price, downstream agricultural film operating rate, capacity utilization rate, UV326 inventory data, crude oil price, exchange rate, and market sentiment index, totaling 9 feature dimensions; each row represents a complete feature set within a prediction period.

[0044] Furthermore, the steps of inputting the feature matrix into a two-layer LSTM prediction model based on an attention mechanism for price prediction include: inputting the feature matrix into the first LSTM layer to capture the long-term temporal dependence of light stabilizer prices and industry cycle patterns; then inputting it into the second LSTM layer to extract the temporal nonlinear features of light stabilizer prices; through the attention mechanism layer, dynamically assigning feature weights to the structured digital data and the market sentiment index according to the degree of influence of each feature on the current light stabilizer price; and processing the weighted features through a fully connected output layer to output the predicted value of light stabilizer prices for the next period.

[0045] The model structure is as follows: Figure 5 As shown, from bottom to top, the layers are: input layer, LSTM layer 1, LSTM layer 2, attention mechanism layer, and fully connected output layer, ultimately outputting the predicted price value Y. t+1 The specific steps are as follows: Input layer: fusing the feature matrix X t Input the model as a time series; The first LSTM layer has 128 neurons. Time-series data with a fused feature matrix is ​​input into this layer, and the LSTM passes the data through its internal forgetting gate. t Input gate i t And output gate ot, update cell C t and hidden layer state h t The specific calculation formula is as follows: in, The Sigmoid activation function maps the output to the [0,1] interval, controlling the opening and closing degree of each gate; tanh is the hyperbolic tangent activation function, used to generate candidate cell states; W f W i W C W o These are the weight matrices for the forget gate, input gate, cell state, and output gate, respectively; b f b i b C b o These are the corresponding bias terms; h t-1 Let x be the hidden layer state at time t-1. t Let t be the input features at time t.

[0046] The LSTM layer compresses historical trends and periodic patterns into a high-dimensional hidden state vector h. t Meanwhile, by using the forget gate mechanism, short-term and occasional price noise data is filtered out, such as the price increase pattern during the peak season of agricultural film demand each year and the long-term impact of the cyclical fluctuations of upstream raw materials on the price of UV326. This filters out short-term and occasional price noise and outputs the feature vector sequence of the first stage.

[0047] The second LSTM layer, with 64 neurons, takes the feature vector sequence output from the first layer as input to further refine the temporal nonlinear features of UV326 prices. Its cell state and hidden layer state update logic is consistent with the first LSTM layer. Through input and output gate mechanisms, it enhances the extraction of key price change features. For example, it captures nonlinear price fluctuations caused by the combined effects of upstream raw material prices and downstream demand, as well as the short-term nonlinear impact of sudden factors on prices, outputting a more representative high-dimensional feature vector. Attention Mechanism Layer: An attention mechanism layer is introduced after the two-layer LSTM layer. This layer calculates the influence weights of each input feature on the UV326 price based on the current market situation, dynamically assigning feature weights to the structured digital data and market sentiment index, as detailed below: The attention weights α are calculated above the LSTM layer. t,i This determines the contribution of the i-th input factor to future price prediction at time t, calculated as follows: in, Let be the attention score of the i-th input factor at time t; , , For the trainable weight parameters of the attention mechanism, Here, k is the bias term; k is the total dimension of the input features; exp is the exponential function used to normalize the attention scores, such that all... The sum is 1. Here... This is equivalent to the dynamically changing price correlation factor coefficient in traditional price forecasting formulas. When the market is in a stable state, the corresponding price of raw materials... The market sentiment index has a higher weighting; when unexpected events occur in the market (such as environmental protection-related production restrictions or plant maintenance), the market sentiment index corresponds to... The weights are automatically amplified, enabling the model to respond quickly to unexpected factors.

[0048] Fully Connected Output Layer (Dense Layer): The weighted feature vector from the attention mechanism layer is input into the fully connected output layer. After processing by an activation function (e.g., ReLU), it is mapped to a scalar value, and the output is the UV326 normalized predicted price for the next period (e.g., week t+1). The calculation formula is as follows: in, This is the weight matrix of the fully connected layer. For bias terms of fully connected layers; The price forecast value is the price after Min-Max normalization, and the value range is [0,1].

[0049] Furthermore, before actual prediction, the aforementioned attention-based two-layer LSTM model needs to undergo training and iterative optimization, including: using mean squared error or root mean square error as the loss function to calculate the deviation between the model's predicted price and the actual spot price of the light stabilizer; introducing an L2 regularization term into the loss function to prevent model overfitting; employing the Adam optimization algorithm to perform backpropagation with a preset initial learning rate to update the weight parameters and bias terms of the two-layer LSTM prediction model until the loss function converges to below a preset threshold, or the number of iterations reaches the maximum number of iterations; performing residual analysis on the model's prediction results; and triggering a concept drift alarm when the absolute value of the deviation between the predicted value and the true value exceeds a set threshold within multiple consecutive prediction periods, and readjusting the propagation lag period corresponding to the input features or the structured digital data.

[0050] Specifically, in this embodiment, mean squared error (MSE) or root mean squared error (RMSE) is used as the model's loss function to calculate the deviation between the model's predicted price and the actual spot price of UV326. Since RMSE better reflects the business scenario of price prediction and can directly reflect the absolute deviation between the predicted and actual values, this embodiment prefers RMSE. Simultaneously, an L2 regularization term is introduced into the loss function to prevent the model from overfitting to abnormal noise in historical data. The formula for the loss function with L2 regularization is: in, The actual UV326 price is given by λ, where λ is the regularization penalty coefficient. This is the sum of squares of all weight parameters in the model.

[0051] Furthermore, this embodiment uses the Adam optimization algorithm for model training, sets the initial learning rate to 0.001, and continuously updates the weight parameters and bias terms of the two-layer LSTM model through the backpropagation algorithm, repeatedly iterating until the loss function converges to below the preset threshold (RMSE≤500 yuan / ton in this embodiment), or the number of model iterations reaches the maximum number of iterations (100 times in this embodiment), at which point training stops.

[0052] Furthermore, after the model is put into practical use, residual analysis is continuously performed on the prediction results to calculate the ratio of the absolute value of the deviation between the predicted value and the actual value to the actual value. When the ratio of the absolute value of the deviation between the predicted value and the actual value to the actual value exceeds a set threshold (5% in this embodiment) within three consecutive prediction periods, the system automatically triggers a concept drift alarm. This indicates that a structural change may have occurred in the light stabilizer industry chain (e.g., the commissioning of a new synthesis process or large-scale deployment of new production capacity). At this point, manual intervention is required to readjust the model's input features (such as the proportion of new process capacity, raw material substitution rate, etc.), or to re-determine the transmission lag period of each structured digital data through Granger causality tests and Pearson correlation coefficients, and then retrain the model to ensure that the model's prediction accuracy always adapts to market changes.

[0053] Step S5 further includes generating a business decision instruction based on the output price prediction result of the light stabilizer, including: performing denormalization processing on the normalized light stabilizer price prediction value output by the model; smoothing the denormalized predicted price using the exponential moving average method; and generating a corresponding business decision instruction based on the rise and fall threshold between the smoothed predicted price and the current actual price of the light stabilizer.

[0054] Specifically, firstly, the normalized UV326 price prediction value output in step S4 is... The price is then converted back to its actual value (yuan / ton) using the Min-Max inverse normalization formula. The inverse normalization formula is as follows: in, This represents the predicted actual price of UV326. For normalized predicted values, , These are the minimum and maximum values ​​of the UV326 price during model training.

[0055] Next, the exponential moving average (EMA) method is used to smooth the inversely normalized predicted price, eliminating minor oscillations in the predicted curve that have no practical significance, making the predicted price more closely reflect the actual market trend. The smoothing calculation formula is as follows: in, The final predicted price after smoothing. The original predicted price is after inverse normalization, and β is the smoothing coefficient (0.7 in this example). This is the smoothed price from the previous forecast period.

[0056] Furthermore, the predicted price output by the model is not a single number. Instead, it is first adjusted for a safety margin according to business rules, and then a range-based strategy is provided to buffer actual purchasing decisions and mitigate risks. For example, if the model calculates an 80% probability of price increases in the next two weeks and a target price of 45,000 yuan / ton, it will not directly make a purchasing decision based on 45,000 yuan. Instead, a 2% safety margin is first subtracted, and 44,100 yuan / ton is used as the reference price for actual decisions. This avoids companies blindly stocking up at excessively high price points due to small deviations in the model's prediction.

[0057] Finally, the smoothed final predicted price is compared with the current actual spot price of UV326 to calculate the price fluctuation. Based on the preset fluctuation threshold, a corresponding standardized business decision instruction is generated. In this embodiment, the fluctuation threshold is set to 3%, and the corresponding instruction is as follows: When the predicted price increase is greater than 3% relative to the current actual price, the system outputs the instruction to "establish a three-month strategic inventory and prepare goods in advance" to guide enterprises to complete strategic warehousing before the price increase and reduce procurement costs. When the predicted price fluctuates within ±3% relative to the current actual price, the system outputs the instruction "Purchase as needed, maintain safety stock" to guide enterprises to purchase according to normal production needs and keep inventory at a reasonable level. When the predicted price is less than -3% relative to the current actual price, the system outputs the instruction "consume existing inventory and delay new order pricing" to guide enterprises to reduce purchases, digest existing inventory, and wait for prices to fall back to a reasonable range before making purchases.

[0058] In this embodiment, if the current actual price of UV326 is 43,000 yuan / ton and the smoothed predicted price is 44,500 yuan / ton, the price increase is 3.49% (>3%), then the system outputs the business decision instruction "Establish a three-month strategic inventory and prepare goods in advance"; if the predicted price is 43,800 yuan / ton and the increase is 1.86% (within the ±3% range), then the system outputs the instruction "Purchase as needed and maintain safety stock".

[0059] It should be noted that the model architecture in this embodiment is not limited to a certain category of light stabilizers. It only requires replacing the upstream raw material collection items according to the synthesis process of the target light stabilizer and separately calculating the upstream raw material transmission lag period. The remaining data collection rules, NLP sentiment quantification methods, model training logic, and business decision rules are all general modules. It can be quickly and horizontally extended to price prediction of light stabilizer categories such as benzotriazoles (UV326, UV327, UV328, etc.) and hindered amines (HS-944, HS-770, etc.) without reconstructing the model. This significantly reduces the technology implementation cost for enterprises and lays the foundation for building intelligent monitoring of prices for all categories of chemical additives.

[0060] In another embodiment, a light stabilizer price prediction system is also disclosed, such as Figure 2 As shown, the system includes a data acquisition module, an NLP sentiment quantification module, a dynamic time window alignment module, an LSTM prediction module, and a result output module.

[0061] The data acquisition module is used to collect structured digital data and unstructured text data related to light stabilizer prices. The structured digital data includes upstream raw material prices, downstream industry operating rates, capacity utilization rates, inventory data, and macroeconomic data such as crude oil prices and exchange rates for the target light stabilizer. The unstructured text data includes environmental inspection notices, chemical industrial park production restriction announcements, production facility maintenance announcements, and news related to the supply and demand of the light stabilizer industry.

[0062] The NLP sentiment quantification module uses FinBERT or BERT models to perform sentiment polarity analysis on unstructured text data, transforming text information into a market sentiment index normalized to the range of [-1, 1], where 1 represents an extremely bullish market with shortages and -1 represents an extremely bearish market with excesses.

[0063] The dynamic time window alignment module determines the transmission lag period corresponding to each structured digital data by using Granger causality test and Pearson correlation coefficient. It performs time axis misalignment shift on the structured digital data according to the period, and then merges the aligned structured digital data with the market sentiment index into a unified feature matrix, eliminating time misalignment and dimensional differences, and providing standardized features for model input.

[0064] The LSTM prediction module employs a two-layer LSTM prediction model based on an attention mechanism. First, the feature matrix is ​​input into the first LSTM layer to capture the long-term temporal dependence of light stabilizer prices and industry cyclical patterns. Then, it is input into the second LSTM layer to extract nonlinear fluctuation features. The attention mechanism layer dynamically allocates the weights of each feature, automatically amplifying the impact of market sentiment indices during sudden market fluctuations. During model training, mean squared error or root mean square error is used as the loss function, combined with the Adam optimizer and L2 regularization for iterative optimization. After deployment, residual analysis is continuously performed. When the prediction deviation percentage exceeds a threshold (e.g., 5%) for three consecutive prediction periods, a concept drift alarm is automatically triggered, prompting manual readjustment of input features or propagation lag periods.

[0065] The output module is used to output the price forecast results for light stabilizers. It also performs inverse normalization on the normalized predicted price output by the model, restoring it to the actual price unit (yuan / ton). Then, it smooths the predicted price using the exponential moving average method to filter out short-term fluctuations. Finally, based on the threshold difference between the smoothed predicted price and the current actual price, it generates and outputs corresponding business decision instructions. For example, when the price fluctuation is >3%, it outputs the instruction "Establish a three-month strategic inventory and prepare stock in advance"; when the fluctuation is within ±3%, it outputs the instruction "Purchase as needed and maintain safety stock"; and when the fluctuation is <-3%, it outputs the instruction "Deplete existing inventory and delay new order pricing". Simultaneously, it outputs the price forecast curve and inventory warning signal, providing intuitive decision-making basis for enterprise supply chain management.

[0066] In summary, this invention overcomes the limitations of traditional quantitative models that rely solely on structured indicators by performing NLP sentiment quantification on unstructured text data, effectively quantifying market sentiment factors. By dynamically aligning structured digital data with time windows, it can reconstruct the true physical production and price transmission cycles of the chemical industry chain, thus compensating for the shortcomings of existing models that ignore time lags and cannot accurately capture price inflection points. A two-layer LSTM prediction model based on an attention mechanism is used to achieve price prediction based on multi-source fusion data, replacing the traditional subjective judgment method that relies on human experience, making the price prediction logic more objective and replicable, thereby improving the efficiency and accuracy of judgment.

[0067] The above are merely preferred embodiments of the present invention and do not constitute any limitation on the present invention. Any equivalent substitutions or modifications made by those skilled in the art to the technical solutions and content disclosed in the present invention without departing from the scope of the present invention shall be deemed to have remained within the protection scope of the present invention.

Claims

1. A method for predicting the price of light stabilizers, characterized in that, include: The data collection includes both structured digital data and unstructured text data; The unstructured text data is subjected to NLP sentiment quantification to generate a market sentiment index; Dynamic time window lag alignment is performed on the structured digital data; The aligned structured digital data is fused with the market sentiment index to form a feature matrix, which is then input into a two-layer LSTM prediction model based on an attention mechanism for price prediction. Price forecast results for output light stabilizers.

2. The method according to claim 1, characterized in that, The structured digital data includes upstream raw material prices, downstream industry operating rates, capacity utilization rates, inventory data, and macroeconomic data.

3. The method according to claim 1, characterized in that, The unstructured text data includes environmental inspection notices, production restriction announcements for chemical industrial parks, maintenance announcements for production facilities, and news related to supply and demand in the chemical industry.

4. The method according to claim 1, characterized in that, The NLP sentiment quantification includes using a FinBERT or BERT model to perform sentiment polarity analysis on the unstructured text data and normalize the market sentiment index to the [-1,1] interval.

5. The method according to claim 1, characterized in that, The dynamic time window lag alignment includes: calculating the correlation coefficient between each structured digital data and the price of light stabilizer under different time shifts through Granger causality test and Pearson correlation coefficient, determining the conduction lag period corresponding to each structured digital data, and then performing time axis misalignment shift on the structured digital data according to the conduction lag period.

6. The method according to claim 1, characterized in that, Before merging the aligned structured digital data with the market sentiment index into a feature matrix, all numerical data need to be normalized to the [0,1] interval using Min-Max.

7. The method according to claim 1, characterized in that, The steps for inputting the feature matrix into a two-layer LSTM prediction model based on an attention mechanism to predict prices include: The feature matrix is ​​input into the first LSTM layer to capture the long-term temporal dependence of light stabilizer prices and industry cycle patterns. Then input the second LSTM layer to extract the temporal nonlinear characteristics of light stabilizer prices; Through the attention mechanism layer, feature weights are dynamically assigned to the structured digital data and the market sentiment index based on the degree of influence of each feature on the current price of light stabilizers. The weighted features are processed by a fully connected output layer to output the predicted price of light stabilizers for the next cycle.

8. The method according to claim 1, characterized in that, The training and iterative optimization process of the attention-based two-layer LSTM prediction model includes: The mean square error or root mean square error is used as the loss function to calculate the deviation between the model's predicted price and the actual spot price of light stabilizer. An L2 regularization term is introduced into the loss function to prevent the model from overfitting. The Adam optimization algorithm is used to perform backpropagation with a preset initial learning rate to update the weight parameters and bias terms of the two-layer LSTM prediction model until the loss function converges to below the preset threshold or the number of iterations reaches the maximum number of iterations. Residual analysis is performed on the model prediction results. When the absolute value of the deviation between the predicted value and the actual value exceeds the set threshold in multiple consecutive prediction periods, a concept drift alarm is triggered, and the transmission lag period corresponding to the input features or the structured digital data is readjusted.

9. The method according to claim 1, characterized in that, It also includes generating business decision instructions based on the price forecast results of the output light stabilizer, including: The normalized light stabilizer price predictions output by the model are denormalized. The inversely normalized predicted price is smoothed using the exponential moving average method. Based on the fluctuation threshold between the smoothed predicted price and the current actual price of the light stabilizer, corresponding business decision instructions are generated.

10. A light stabilizer price prediction system, characterized in that, It includes a data acquisition module, an NLP sentiment quantification module, a dynamic time window alignment module, an LSTM prediction module, and a result output module; The data acquisition module is used to acquire structured digital data and unstructured text data; The NLP sentiment quantification module is used to perform sentiment polarity analysis on the unstructured text data and generate a market sentiment index. The dynamic time window alignment module is used to determine the transmission lag period and time axis misalignment of the structured digital data, and to merge the aligned structured digital data with the market sentiment index into a feature matrix; The LSTM prediction module uses a two-layer LSTM prediction model based on an attention mechanism to predict prices from the feature matrix. It also performs loss assessment, iterative optimization, and residual analysis based on the model, triggers concept drift alarms, and performs readjustment of feature or propagation lag cycles. The result output module is used to output the price prediction results of light stabilizers.

Images