Intelligent coal price automatic temporary estimation and dynamic display system

Abstract

The application discloses an intelligent coal price automatic temporary estimation and dynamic display system and relates to the technical field of coal price prediction. The system comprises: a sudden event real-time sensing and classification module, which is used for collecting multi-source open information and converting the multi-source open information into structured event data; a prediction model dynamic correction module, which receives the structured event data, combines historical price data, and generates a temporary estimated price; and a dynamic display and early warning linkage module, which receives the temporary estimated price and characteristic contribution data, performs visual display, and pushes early warning information according to a price fluctuation range. The application realizes accurate temporary estimation of the coal price by sensing multi-source sudden events in real time and converting the multi-source sudden events into structured data and combining a dynamically corrected prediction model, achieves the effect of improving the timeliness and accuracy of price prediction, and provides intuitive information and timely reminders for users through visual display and early warning linkage.

Description

Technical Field

[0001] This invention relates to the field of coal price forecasting technology, specifically to an intelligent automatic provisional estimation and dynamic display system for coal prices. Background Technology

[0002] As a key energy resource, coal price fluctuations have a profound impact on the energy industry, industrial production, and related trade activities. Accurately grasping coal price trends is of great significance for corporate procurement and sales decisions and the stable development of the industry. Currently, various coal price forecasting systems exist in the market. These systems are mostly based on historical price data and employ models such as BP neural networks, genetic algorithms, and the fusion of XGBoost and graph convolutional networks for price prediction.

[0003] However, existing technologies have obvious shortcomings, namely, insufficient adaptability to sudden market changes. When faced with sudden events such as policy adjustments, natural disasters, and supply chain disruptions, models trained solely on historical data cannot respond quickly, resulting in a significant decrease in prediction accuracy. This makes it impossible to provide timely and reliable price references for enterprises and to meet the needs of actual business decision-making. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent automatic provisional estimation and dynamic display system for coal prices, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an intelligent automatic provisional estimation and dynamic display system for coal prices, comprising:

[0006] The real-time perception and classification module for emergencies is used to collect public information from multiple sources and convert it into structured event data. The structured event data includes the event type, impact level weight, GeoHash encoding of the coverage area, and duration.

[0007] The predictive model dynamic correction module receives the structured event data and generates a provisional price by combining it with historical price data. The predictive model dynamic correction module includes a basic predictive model, an event embedding layer, and a model self-updating unit. The event embedding layer performs multi-dimensional feature fusion on the structured event data, and the model self-updating unit adjusts the model parameters based on price deviation data.

[0008] The dynamic display and early warning linkage module receives the estimated price and feature contribution data, displays them visually, pushes early warning information based on price fluctuations, and feeds back the deviation between the actual price and the estimated price to the prediction model dynamic correction module.

[0009] Preferably, the real-time emergency perception and classification module includes:

[0010] The multi-source event acquisition unit obtains publicly available information by connecting to government open data interfaces, industry legal information service platform data interfaces, meteorological department official early warning release channels, and authorized social media data services;

[0011] The keyword extraction and classification unit extracts event keywords from the publicly available information and uses a pre-trained BERT model to classify events into policy, natural, supply chain, market, security, and environmental categories.

[0012] The impact level quantification unit, based on the evaluation matrix trained on historical cases, determines the impact level weight, GeoHash encoding of the coverage area, and duration of the event, and generates the structured event data.

[0013] Preferably, the influence level quantification unit includes:

[0014] The spatial range coding subunit uses GeoHash coding to hierarchically identify the event coverage range, with the mining area level using 8-bit coding and the regional level using 6-bit coding;

[0015] The time span is used to determine the sub-units, and the duration range of the event is determined based on the nature of the event and historical data;

[0016] The weighting sub-unit calculates the impact level weight of an event based on its type, scope of influence, and duration using an evaluation matrix.

[0017] Preferably, the basic prediction model is an LSTM model, which generates an initial price prediction sequence based on historical price data over a preset time period.

[0018] Preferably, the event embedding layer includes:

[0019] The spatial dimension processing unit calculates the percentage of the intersection area between the event coverage area and the target area to obtain the regional correlation coefficient. The calculation formula is as follows:

[0020] ,

[0021] in, Indicates the degree of regional correlation. This represents the intersection area of ​​the event-covered grid and the target grid. The total area of ​​the target grid is represented by the event coverage grid, which is the grid corresponding to the GeoHash code of the coverage area in the structured event data, and the target grid is the grid corresponding to the GeoHash code of the price region to be predicted.

[0022] The time-dimensional processing unit uses different weight decay formulas based on the event duration:

[0023] When the duration is ≤7 days ;

[0024] When the duration is >7 days ;

[0025] in, This represents the weight value at time t. This represents the initial impact level weight determined by the impact level quantification unit, and t represents the number of days the event has lasted.

[0026] The type dimension processing unit uses a sliding window weighted average to calculate the influence coefficient of the current event. The calculation formula is as follows:

[0027] ,

[0028] in, This represents the impact coefficient of the current event. This represents the magnitude of the impact of the i-th similar historical event. Let represent the time decay factor of the i-th historical similar event, and , where n represents the number of months since the i-th historical similar event occurred, and the sliding window size is the last 5 similar events;

[0029] The attention mechanism unit sets up multiple attention heads to focus on different event feature dimensions, and calculates the feature weights using a formula:

[0030] ,

[0031] in, Represents the feature vector of the current event. Represents a vector of historical event feature base. This indicates the feature dimension and has a value of 64. The attention mechanism unit generates a feature contribution heatmap by representing the matrix transpose operation.

[0032] Preferably, the model self-updating unit includes:

[0033] The incremental training subunit uses a sliding window to filter samples with an "event-price" correlation of ≥0.5 within the past 72 hours, freezes the parameters of the first 3 layers of the LSTM model, and unfreezes the parameters of the last 2 layers and the event embedding layer for fine-tuning training. The correlation is calculated using the Pearson coefficient.

[0034] The abnormal event handling subunit, when dealing with a novel event without historical reference, invokes the similarity event retrieval engine. Based on BERT sentence vectors, it calculates the event text similarity, retrieves the top 3 historical events with a similarity ≥ 0.7 as references, and initiates an online learning mode to update the influence coefficient. The update formula is:

[0035] ,

[0036] in, This represents the updated impact coefficient. This represents the influence coefficient before the update. This represents the difference between the actual price and the estimated price. This indicates a provisional price;

[0037] The judgment and triggering subunit calculates the cosine similarity between the feature vector output by the event embedding layer and the historical effective feature set after the feature vector is output. If the similarity is ≥0.6, it is input into the basic prediction model for price correction. If the similarity is <0.6, it is returned to the real-time perception and classification module of the sudden event for reprocessing.

[0038] Preferably, the incremental training subunit uses the AdamW optimizer for parameter optimization, with an initial learning rate of 0.0005, a batch size of 32, and the number of training epochs determined by the formula. The calculation shows that N is the sample size, and the sample size satisfies 500≤N≤1000.

[0039] Preferably, the judgment and triggering subunit further includes a dual threshold verification mechanism: when the price deviation is ≥5% for 3 consecutive times, an absolute deviation or relative deviation verification is initiated. The absolute deviation is |actual price - provisional price| ≥ 10 yuan / ton, and the relative deviation is |actual price - provisional price| / provisional price ≥ 8%. If either condition is met, a full parameter adjustment is triggered, and the learning rate is dynamically increased according to the following formula:

[0040] ,

[0041] in, This represents the adjusted learning rate. This represents the learning rate before adjustment. This indicates the actual price deviation.

[0042] Preferably, the dynamic display and early warning linkage module includes:

[0043] The visualization unit generates an event-price correlation chart, which includes a provisional price series, a price deviation line, and a feature contribution heatmap. The price deviation is the ratio of the difference between the provisional price under the influence of the event and the initial predicted price without the event assumption.

[0044] The early warning push unit pushes early warning information to users when the estimated price fluctuation is ≥3%. The early warning information includes event details, price correction logic, and response suggestions.

[0045] The feedback unit periodically collects actual transaction prices, calculates the deviation from the estimated prices, and feeds it back to the prediction model's dynamic correction module.

[0046] Optionally, the visualization unit supports multi-terminal adaptation, including PC, mobile and large-screen devices, and the event-price correlation chart marks the event duration axis and the weight of core influencing factors.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] By sensing multi-source emergencies in real time and transforming them into structured data, combined with a dynamically corrected prediction model, accurate preliminary estimates of coal prices are achieved, improving the timeliness and accuracy of price forecasts. Through multi-dimensional feature fusion and a model self-updating mechanism, the system can adapt to market changes, enhancing its dynamic response capability to the impact of various events. Visual displays and early warning linkage provide users with intuitive information and timely reminders. Combined with closed-loop feedback to optimize the model, the system assists in decision-making and continuously improves system performance. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of an intelligent coal price automatic provisional estimation and dynamic display system provided in an embodiment of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Please see Figure 1 The present invention provides an intelligent automatic provisional estimate and dynamic display system for coal prices, including a real-time perception and classification module 11 for emergencies, a dynamic correction module 12 for prediction models, and a dynamic display and early warning linkage module 13.

[0052] The real-time perception and classification module 11 for emergencies is used to collect public information from multiple sources and convert it into structured event data. The structured event data includes the event type, impact level weight, GeoHash encoding of the coverage area, and duration.

[0053] Specifically, the real-time event perception and classification module 11 refers to the functional module in the system used to acquire external event information and transform it into structured data. The purpose of this module is to provide comprehensive and accurate event data support for subsequent price estimates. By collecting and processing information from multiple sources, it ensures that the system can promptly capture various emergencies that may affect coal prices, thereby improving the timeliness and accuracy of price estimates. Its implementation involves first collecting publicly available information through various legal channels, then using technologies such as natural language processing to analyze, classify, and quantify the information, ultimately generating structured event data containing key information.

[0054] Among them, multi-source public information refers to information from different channels that is open to the public, including government announcements, industry news platform reports, and meteorological warnings; structured event data refers to data organized according to a preset format that contains key event attributes, facilitating processing and analysis by subsequent modules of the system; event type refers to the classification of the nature of the event, such as policy-related or natural events; impact level weight refers to the numerical value used to measure the degree of impact of the event on coal prices; GeoHash encoding of the coverage area refers to the string encoded by the GeoHash algorithm for the geographical area affected by the event; and duration refers to the length of time from the occurrence to the end of the event.

[0055] In one possible implementation, the collection of publicly available information from multiple sources can employ a combination of scheduled collection and real-time monitoring. Scheduled collection can be set to occur hourly, while real-time monitoring provides immediate responses to updates from important channels such as government websites to ensure the timeliness of information. Furthermore, it should be noted that the fields in the structured event data can be expanded according to actual needs. In addition to including event type, impact level weight, GeoHash encoding of coverage area, and duration, fields such as event occurrence time and event source can be added to make the data more complete.

[0056] The prediction model dynamic correction module 12 receives the structured event data and generates a provisional price by combining it with historical price data. The prediction model dynamic correction module 12 includes a basic prediction model, an event embedding layer, and a model self-updating unit. The event embedding layer performs multi-dimensional feature fusion on the structured event data, and the model self-updating unit adjusts the model parameters based on price deviation data.

[0057] Specifically, the predictive model dynamic correction module 12 refers to the functional module in the system that receives structured event data and historical price data, generates provisional prices through model calculations, and dynamically adjusts them. The purpose of this module is to combine event factors and historical price trends to generate accurate provisional coal prices, and continuously optimize the model based on actual conditions to adapt to dynamic market changes. Its implementation involves first generating an initial prediction based on historical price data using a basic predictive model, then correcting the initial prediction by integrating event factors through an event embedding layer, and finally adjusting model parameters based on price deviations through a model self-updating unit to continuously improve prediction accuracy.

[0058] Here, historical price data refers to the record of coal transaction prices over a past period; provisional price refers to the coal price predicted by the system at a future point in time; the basic prediction model refers to the algorithm model used for price prediction based on historical price data, which is an LSTM model in this embodiment; the event embedding layer refers to the functional unit used to transform structured event data into feature vectors that can be fused with price data; the model self-updating unit refers to the functional unit used to adjust model parameters according to the deviation between the actual price and the provisional price; multi-dimensional feature fusion refers to the process of processing and integrating event features from multiple dimensions such as space, time, and type; and price deviation data refers to the difference between the actual transaction price and the system's provisional price.

[0059] In one possible implementation, the historical price data used by the base prediction model can be selected from daily transaction prices over the past 12 months to balance the timeliness and trend of the data. Additionally, it should be noted that the multi-dimensional feature fusion of the event embedding layer can use a weighted summation method to combine spatial, temporal, and type-dimensional features into a comprehensive feature vector. The weights can be determined through training with historical data based on the importance of each dimension's impact on prices.

[0060] The dynamic display and early warning linkage module 13 receives the estimated price and feature contribution data, performs visualization display, and pushes early warning information according to the price fluctuation range. At the same time, it feeds back the deviation between the actual price and the estimated price to the prediction model dynamic correction module 12.

[0061] Specifically, the dynamic display and early warning linkage module 13 refers to the functional module in the system used to present price information, issue early warnings, and provide feedback on actual price deviations. The purpose of this module is to intuitively display the provisional estimate of coal prices and the impact of related events to users, and to promptly alert users when price fluctuations are significant. Simultaneously, it feeds back actual price data to the model correction module, forming a closed-loop optimization of the system. Its implementation involves receiving provisional price and feature contribution data, displaying them through visualization technology, setting a price fluctuation threshold, triggering an early warning when the provisional price fluctuation exceeds the threshold, and simultaneously collecting actual prices and calculating the deviation to feed back to the prediction model dynamic correction module 12.

[0062] Among them, feature contribution data refers to data used to illustrate the degree of impact of each event feature on the provisional price; visualization refers to the way data is displayed in an intuitive form such as charts; price fluctuation range refers to the percentage change of the provisional price compared with the benchmark price; early warning information refers to the prompt information issued to users when the price fluctuation exceeds the threshold; actual price refers to the transaction price of coal in actual transactions; deviation value refers to the difference between the actual price and the provisional price.

[0063] In one possible implementation, the visualization can be presented via web and mobile applications. Web applications are suitable for displaying overall trends on large screens, while mobile applications allow users to view the information anytime. In addition to the estimated price series, the displayed content may also include a contribution percentage chart for each event characteristic to help users understand the price composition. Furthermore, it should be noted that the push notifications for alerts can support user-customized reception methods, such as receiving only SMS messages, only system pop-ups, or multiple methods simultaneously, to meet the needs of different users.

[0064] In an optional embodiment, the real-time event perception and classification module 11 includes:

[0065] The multi-source event acquisition unit obtains publicly available information by connecting to government open data interfaces, industry legal information service platform data interfaces, meteorological department official early warning release channels, and authorized social media data services;

[0066] The keyword extraction and classification unit extracts event keywords from the publicly available information and uses a pre-trained BERT model to classify events into policy, natural, supply chain, market, security, and environmental categories.

[0067] The impact level quantification unit, based on the evaluation matrix trained on historical cases, determines the impact level weight, GeoHash encoding of the coverage area, and duration of the event, and generates the structured event data.

[0068] It should be noted that in this embodiment, the multi-source event acquisition unit refers to the functional unit that acquires various types of publicly available information through legitimate interfaces. Its purpose is to ensure that the system can comprehensively capture event information that may affect coal prices, and to ensure the authenticity and timeliness of information by connecting with authoritative data from different channels. The keyword extraction and classification unit refers to the functional unit that extracts key information from publicly available information and classifies events. Through the deep semantic understanding of the BERT model, it achieves the accuracy of event classification, providing a foundation for subsequent impact quantification. The impact level quantification unit refers to the functional unit that quantifies the degree, scope, and duration of the event's impact. Through the evaluation matrix trained on historical cases, it transforms qualitative events into computable quantitative data, generating structured event data and providing standardized input for the prediction model. These three units work together to form a complete process from information acquisition to data processing, ensuring the comprehensiveness, accuracy, and usability of event data.

[0069] In one possible implementation, the multi-source event acquisition unit can prioritize information from different channels, with government public data interfaces and meteorological department early warning channels having higher priority than social media data services, thus prioritizing authoritative information. Additionally, it should be noted that the keyword extraction and classification unit can periodically update the BERT model's training data, incorporating the latest coal industry terminology and event examples to improve the adaptability of the classification.

[0070] For example, in one feasible implementation, the multi-source event acquisition unit obtains information such as "a coal mine has experienced a safety accident leading to a production stoppage" by connecting to the official website of the National Development and Reform Commission's Energy Bureau, the coal industry information platform, and the provincial meteorological early warning system; the keyword extraction and classification unit extracts keywords such as "coal mine safety accident" and "production stoppage" from it, and classifies it into a safety event using the BERT model; the impact level quantification unit determines the impact level weight of the event to be 0.6 based on the evaluation matrix, the GeoHash code of the coverage area is an 8-bit code at the mine area level, the estimated duration is 15 days, and finally generates structured event data and transmits it to the prediction model dynamic correction module 12.

[0071] It should be noted that the construction and training process of the evaluation matrix is ​​as follows:

[0072] Training data source: Collection of coal industry emergency cases over the past 5 years (a total of 1200 cases, covering 6 types of events including policy-related and natural events) and coal price fluctuation data (daily price change rate) during the corresponding events.

[0073] Evaluation dimensions are determined as follows: The matrix is ​​set horizontally by event type (6 categories), coverage (divided into mining area level and regional level according to GeoHash encoding accuracy), and duration (divided into ≤7 days, 8-30 days, and >30 days according to the interval), and vertically by impact level weight (0.1-1.0, with an interval of 0.1).

[0074] Initial weight assignment: Using the Analytic Hierarchy Process (AHP), five experts from the coal industry and data analysis field were invited to score the degree of influence of each dimension combination, and the average weight was calculated as the initial value of the matrix.

[0075] Model training optimization: Using the deviation between the weight values ​​output by the evaluation matrix and the actual price fluctuation range as the loss function, the gradient descent method is used to iteratively optimize the matrix parameters. The number of iterations is set to 500 times until the loss function value is ≤3%, thus completing the training of the evaluation matrix.

[0076] For example, for the combination of 'coal mine safety accident' (safety-related event), 'mine area coverage' (8-bit GeoHash encoding), and 'duration of 15 days' (8-30 day range), the evaluation matrix outputs an initial weight of 0.55 through the above training process. After verification with actual price fluctuation data, it is optimized to 0.6, which matches the impact level weight of the event.

[0077] Furthermore, the construction and training process of the BERT model is as follows:

[0078] Base model selection: The open-source BERT-base-uncased version (containing 12 Transformer layers, 768-dimensional feature vectors, and 12 attention heads) is adopted.

[0079] Pre-training data preparation: Collect 100,000 text data related to the coal industry (including government policy documents, industry news reports, corporate event announcements, weather warning notices, etc.), and use them as pre-training corpus after text cleaning (removing special symbols and unifying lowercase).

[0080] Fine-tuning the data annotation: Annotate sample data for 6 types of events, with 5,000 annotations for each type of event (30,000 in total). Each sample contains the correspondence between 'event text' and 'event type label' (e.g., 'environmental protection production restriction notice' corresponds to the 'policy category' label).

[0081] Model training parameters: The Adam optimizer was used in the fine-tuning stage, with the initial learning rate set to 2e-5, the batch size set to 32, the number of training epochs set to 3, the dropout probability set to 0.1, and the training objective set to minimize the cross-entropy loss function. Training was stopped when the validation set accuracy was ≥92%, thus completing the construction and training of the BERT model.

[0082] In this process, the BERT model encodes keywords such as 'coal mine safety accident' and 'production stoppage' using semantic features of coal industry terms learned through pre-training, and outputs a 768-dimensional feature vector. After being calculated by the classification layer (fully connected layer + softmax function), the probability of 'safety event' is 0.95, which is higher than the probability of other categories, thus achieving accurate classification of the event.

[0083] In an optional embodiment, the influence level quantification unit includes:

[0084] The spatial range coding subunit uses GeoHash coding to hierarchically identify the event coverage range, with the mining area level using 8-bit coding and the regional level using 6-bit coding;

[0085] The time span is used to determine the sub-units, and the duration range of the event is determined based on the nature of the event and historical data;

[0086] The weighting sub-unit calculates the impact level weight of an event based on its type, scope of influence, and duration using an evaluation matrix.

[0087] It should be noted that, in this embodiment, the spatial range encoding subunit refers to the functional subunit that performs GeoHash encoding and classification of the event coverage area. Through encoding with different precision (8 bits for mining areas and 6 bits for regional areas), it achieves accurate identification of the geographical areas affected by the event, providing basic data for subsequent spatial dimension feature fusion. The time span determination subunit refers to the functional subunit that determines the duration interval of the event. Combining the nature of the event and historical data, it ensures that the estimation of the event's impact duration is reasonable, providing a basis for the weight decay calculation in the time dimension. The weight assignment subunit refers to the functional subunit that calculates the weight of the event's impact level. Taking into account factors such as event type, impact range, and duration, it derives a quantified weight value through an evaluation matrix, reflecting the potential impact of the event on coal prices. These three subunits quantify the event from three dimensions: space, time, and impact level, collectively constituting the core logic of impact level quantification and ensuring the scientific nature of the structured event data.

[0088] In one possible implementation, the spatial range coding subunit can automatically select the coding precision based on the event type. For example, a nationwide policy event might use a regional-level 6-bit code, while a single mining area accident might use a mining area-level 8-bit code. Additionally, it should be noted that the time span determination subunit can dynamically adjust the estimated duration by incorporating real-time feedback on event progress information; for instance, the estimated duration can be shortened when the accident handling process accelerates.

[0089] For example, in one feasible implementation, a region experiences heavy rainfall that disrupts coal transportation. The spatial range coding subunit uses a regional-level 6-bit GeoHash code for the region; the time span determination subunit determines the event duration range as 7-10 days based on historical cases of rainfall impacting transportation and meteorological forecasts of continuous rainfall; the weight assignment subunit integrates natural event attributes, regional coverage, and estimated duration, calculates the impact level weight as 0.45 through an evaluation matrix, and finally integrates these quantitative data into the structured event data.

[0090] In one optional embodiment, the basic prediction model is an LSTM model, which generates an initial price prediction sequence based on historical price data over a preset time period.

[0091] It should be noted that in this embodiment, the basic prediction model refers to the model in the dynamic correction module 12 of the prediction model that generates the initial price prediction sequence based on historical price data, which in this embodiment is an LSTM model. LSTM stands for Long Short-Term Memory Network, a deep learning model suitable for processing time series data, capable of capturing long-term dependencies and fluctuation patterns in historical price data. The purpose of this model is to provide a basic trend prediction for preliminary price estimates. By training and learning from historical price data over a preset time period (such as the past 12 months), it outputs an initial price sequence for a future period, serving as a benchmark for subsequent corrections based on event data. The implementation idea is to utilize the gating mechanism of LSTM to effectively alleviate the long-term dependency problem, model the historical trend of coal prices, and thus generate initial prediction results with a trend.

[0092] In one possible implementation, the input data for the LSTM model can include historical factors other than prices, such as coal inventory and import volumes for the same period, to enhance the richness of the initial forecast. Additionally, it should be noted that the preset time period can be adjusted according to the cyclical characteristics of coal prices; for example, it can be shortened to the past three months during the peak winter coal consumption period to enhance the model's ability to capture recent trends.

[0093] For example, in one feasible implementation, the basic prediction model uses an LSTM model, inputting daily coal transaction price data from the past 12 months. By adjusting the number of hidden layer nodes and the number of iterations, the model's fitting error to historical prices is controlled within 5%. After training, the model outputs an initial price prediction sequence for the next 7 days, with the predicted price for day 1 being 850 yuan / ton, day 2 being 845 yuan / ton, and decreasing sequentially, providing a basis for subsequent adjustments based on event data.

[0094] In an optional embodiment, the event embedding layer includes:

[0095] The spatial dimension processing unit calculates the percentage of the intersection area between the event coverage area and the target area to obtain the regional correlation coefficient. The calculation formula is as follows:

[0096] ,

[0097] in, Indicates the degree of regional correlation. This represents the intersection area of ​​the event-covered grid and the target grid. This represents the total area of ​​the target grid. The event coverage grid is the grid corresponding to the GeoHash code of the coverage area in the structured event data, and the target grid is the grid corresponding to the GeoHash code of the price region to be predicted. The value range of is [0,1]. The closer the value is to 1, the higher the spatial overlap between the event coverage area and the target area, and the greater the potential impact on coal prices in the target area.

[0098] The time-dimensional processing unit uses different weight decay formulas based on the event duration:

[0099] When the duration is ≤7 days ;

[0100] When the duration is >7 days ;

[0101] in, This represents the weight value at time t. This represents the initial impact level weight determined by the impact level quantification unit, t represents the number of days the event has lasted, and 0.7 is the decay coefficient at the 7-day time point. As the term represents the exponential decay, the impact of an event will rapidly decay exponentially as t increases, effectively reflecting the trend of the gradual weakening of the impact of long-term events.

[0102] The type dimension processing unit uses a sliding window weighted average to calculate the impact coefficient of the current event. By constructing a sliding window weighted average model, it fully explores the impact value of similar historical events on the current event. Using the last five similar events as the analysis window, it comprehensively considers the impact magnitude and time decay factor of each historical event to calculate the impact coefficient of the current event. The calculation formula is as follows:

[0103] ,

[0104] in, This represents the impact coefficient of the current event. This represents the magnitude of the impact of the i-th historical similar event, specifically the impact of similar historical events on coal prices, derived through historical data analysis and quantitative assessment. This represents the time decay factor of the i-th historical event of the same type, reflecting the degree to which the influence of a historical event weakens over time. , where n represents the number of months since the i-th historical similar event occurred, and the sliding window size is the last 5 similar events;

[0105] The attention mechanism unit sets up multiple attention heads to focus on different event feature dimensions, and calculates the feature weights using a formula:

[0106] ,

[0107] in, This represents the feature vector of the current event, which contains multi-dimensional information such as the event's space, time, and type. This represents a vector of historical event feature data, storing a large amount of feature data from historical events. This represents the feature dimension and has a value of 64, used for scale normalization of the calculation results; This represents the matrix transpose operation. Through a multi-head attention mechanism, this unit can mine the correlations between event features from different perspectives and ultimately generate a feature contribution heatmap. This visually demonstrates the importance of each feature dimension to coal prices, providing more comprehensive and accurate feature information support for subsequent price forecasts.

[0108] It should be noted that, in this embodiment, the steps for converting GeoHash encoding to the actual area and calculating the intersection area are as follows:

[0109] GeoHash encoding is parsed into latitude and longitude ranges: The open-source GeoHash decoding algorithm is used to parse the GeoHash encoding of the event coverage grid (such as the mining area-level 8-bit encoding "wtw3sq7z") into the corresponding longitude range [lon1,lon2] and latitude range [lat1,lat2]. The longitude accuracy of the 8-bit encoding is ±0.0061° and the latitude accuracy is ±0.0030°; the longitude accuracy of the 6-bit encoding is ±0.0488° and the latitude accuracy is ±0.0244°.

[0110] Calculation of actual grid area: The actual area of ​​a single grid cell is calculated based on the area formula in spherical coordinates. The formula is as follows:

[0111] ,

[0112] in, The radius of the Earth is taken as 6371 km. , The grid's longitude boundary (in radians). , For grid latitude boundaries (radians);

[0113] Intersection area calculation: Determine the latitude and longitude intersection range between the event coverage grid and the target grid through geometric operations. and The area of ​​the intersection region is calculated using the same area formula described above, which gives the intersection area of ​​the event-covered grid and the target grid. .

[0114] It should be noted that in this embodiment, the spatial dimension processing unit refers to the functional unit that calculates the regional correlation coefficient. By measuring the proportion of the intersection area between the event coverage grid and the target grid, it quantifies the spatial correlation between the event-affected area and the price-to-be-predicted area, reflecting the differences in the impact of geographical factors on prices. The temporal dimension processing unit refers to the functional unit that calculates weight decay based on the event duration. Through different decay formulas, it reflects the changing pattern of event impact over time, making the weight of recent impact higher than that of long-term impact. The type dimension processing unit refers to the functional unit that calculates the impact coefficient of the current event. Based on historical similar event data from a sliding window, combined with a time decay factor, it derives a coefficient value reflecting the inherent impact of the event type. The attention mechanism unit refers to the functional unit that calculates feature weights and generates a heatmap. Through multi-head attention, it focuses on different event feature dimensions, highlighting core influencing factors and improving the model's sensitivity to key features. These four units fuse event data from four dimensions: spatial, temporal, type, and feature importance. This transforms structured event data into high-dimensional features that can be combined with price data, providing accurate quantitative basis for event impact for price correction.

[0115] In one possible implementation, the spatial dimension processing unit can layer the target grid, such as appropriately increasing the weight coefficient of the grid for major coal-producing areas to highlight the influence of the core region. Additionally, it should be noted that the number of attention heads in the attention mechanism unit can be adjusted according to the number of event feature dimensions; increasing the number of attention heads when there are many feature dimensions ensures effective attention to each dimension.

[0116] For example, in one feasible implementation, the spatial dimension processing unit calculates the regional correlation coefficient of a certain environmental protection production restriction event to be 0.7 (the intersection area of ​​the event coverage grid and the target production area grid accounts for 70%); the temporal dimension processing unit, since the event duration is 20 days, uses the long-term event decay formula to calculate the weight of the 5th day as the initial value of 0.8 × 0.7 × The type dimension processing unit calculates the current impact coefficient of policy-related events as 0.09 using a sliding window. The attention mechanism unit sets up 8 attention heads, focusing on features such as "implementation strength" and "covered provinces", calculates the feature weights and generates a heat map, and finally inputs the fused feature vector into the LSTM model for price correction.

[0117] It should be noted that in this embodiment, the number of multi-head attention heads in the attention mechanism unit is fixed at 8, and the feature allocation rules for each attention head are as follows:

[0118] First: Focus on "policy strength" and "scope of implementation" (core characteristics of policy-related events);

[0119] First 2: Focus on "event duration" and "time decay weight" (core features of the time dimension);

[0120] First 3: Focus on "influence level weight" and "regional correlation" (core characteristics of influence degree);

[0121] First 4: Focus on "supply chain node type" and "transportation disruption ratio" (core characteristics of supply chain events);

[0122] First 5: Focus on "Natural Disaster Level" and "Early Warning Response Level" (core characteristics of natural events);

[0123] First 6: Focus on "market supply and demand gap" and "inventory fluctuation range" (core characteristics of market events);

[0124] First 7: Focus on "safety incident level" and "scale of production stoppage" (core characteristics of safety incidents);

[0125] First 8: Focus on "the extent of upgrading environmental standards" and "the intensity of emission restrictions" (core characteristics of environmental incidents);

[0126] After each attention head independently calculates its feature weights, they are merged into the final feature vector through a concatenation operation.

[0127] In an optional embodiment, the model self-updating unit includes:

[0128] The incremental training subunit uses a sliding window to filter samples with an "event-price" correlation of ≥0.5 within the past 72 hours, freezes the parameters of the first 3 layers of the LSTM model, and unfreezes the parameters of the last 2 layers and the event embedding layer for fine-tuning training. The correlation is calculated using the Pearson coefficient.

[0129] The abnormal event handling subunit, when dealing with a novel event without historical reference, invokes the similarity event retrieval engine. Based on BERT sentence vectors, it calculates the event text similarity, retrieves the top 3 historical events with a similarity ≥ 0.7 as references, and initiates an online learning mode to update the influence coefficient. The update formula is:

[0130] ,

[0131] in, This represents the updated impact coefficient, used to quantify the degree of impact of this type of event on coal prices; This represents the influence coefficient before the update. It represents the difference between the actual price and the estimated price, reflecting actual market fluctuations; This indicates a provisional price;

[0132] The judgment and triggering subunit calculates the cosine similarity between the feature vector output by the event embedding layer and the historical effective feature set after the feature vector is output. If the similarity value is ≥0.6, it indicates that the current event has a high similarity with the historical event and can be directly input into the basic prediction model for price correction. If the similarity value is <0.6, it is determined that the current event has special characteristics, and the system returns it to the real-time perception and classification module 11 of sudden events to trigger the secondary feature extraction and deep analysis process to ensure the accuracy and reliability of price prediction.

[0133] It should be noted that, in this embodiment, the specific parameters for calculating the Pearson coefficient are as follows:

[0134] Time window length: A rolling time window is used, with a window size of 14 consecutive days after the event (including the day the event occurs). The correlation between the "event characteristic data sequence" and the "coal price fluctuation sequence" within these 14 days is calculated.

[0135] Data standardization method: Both types of sequences are standardized using Z-score, with the following formula:

[0136] ,

[0137] in, The original data, The mean of the data within the window. The standard deviation of the data within the window;

[0138] Coefficient calculation logic: Pearson coefficient The calculation formula is:

[0139] ,

[0140] in, This refers to standardized event feature data (such as daily impact weights). For the standardized daily price fluctuation value, n=14 (data volume within the window), when When the sample is identified as having high relevance, it is considered a high-relevance sample.

[0141] In addition, the specific implementation of the similar event retrieval engine is as follows:

[0142] The retrieval features include six core dimensions: event type label (e.g., policy category = 01, nature category = 02, etc.), latitude and longitude range after GeoHash encoding of the scope of influence, duration interval, keyword vector (768-dimensional sentence vector generated by the BERT model), historical impact level weight mean, and price fluctuation correlation (Pearson coefficient).

[0143] The matching threshold is determined by training with historical data. Specifically, 1,000 known similar events from the past 3 years are selected, and the correlation between the cosine similarity of their feature vectors and the deviation of actual price impact is calculated. When the similarity is ≥0.7, the adjustment error of the impact coefficient of similar events is ≤5%, so this value is set as the matching threshold.

[0144] Search process: First, perform an initial screening using event type tags (only search for events of the same type), then calculate the weighted cosine similarity of the remaining feature dimensions (keyword vector weight 0.4, latitude and longitude range weight 0.2, duration weight 0.2, historical influence weight 0.1, relevance weight 0.1), and take the top 3 events with similarity ≥ 0.7 as references.

[0145] It should be noted that in this embodiment, the incremental training subunit refers to the functional subunit that fine-tunes some parameters of the model. By selecting highly correlated samples and freezing some network parameters, it ensures model update efficiency while avoiding resource consumption and time delays caused by full training. The abnormal event handling subunit refers to the functional subunit that deals with new events without historical reference. Through similar event retrieval and online learning, it enables the model to quickly adapt to unknown events and make up for the lack of historical data. The judgment and triggering subunit refers to the functional subunit that judges the validity of feature vectors and triggers corresponding operations. Through cosine similarity verification, it ensures that the feature data input to the model is valid and avoids interference from invalid data on the prediction results. These three subunits work together to achieve dynamic optimization of the model. Incremental training ensures stable accuracy in normal scenarios, abnormal event handling improves the adaptability of the system, and the judgment mechanism ensures data quality, thus jointly maintaining the long-term effectiveness of the model.

[0146] In one possible implementation, the incremental training subunit can set a dynamic threshold for sample relevance, lowering the threshold to 0.4 when market fluctuations are severe to include more potentially relevant samples. Additionally, it should be noted that the similarity retrieval engine in the abnormal event handling subunit can set a minimum similarity threshold for search results (e.g., not lower than 0.6). When no matching historical events are found, a default impact coefficient is used, and a manual intervention prompt is issued.

[0147] For example, in one feasible implementation, the incremental training subunit selects 800 samples with an "event-price" correlation of ≥0.5 within the past 72 hours, freezes the parameters of the first 3 layers of the LSTM model, and fine-tunes the last 2 layers and the event embedding layer, completing the training within 2 minutes; when a new "energy policy pilot" event occurs, the abnormal event handling subunit retrieves a historical "new energy subsidy" event with a similarity of 0.75 as a reference, starts online learning, and adjusts the influence coefficient from 0.1 to 0.108 based on the price deviation after 4 hours; the judgment and triggering subunit calculates the cosine similarity between the feature vector and the historical effective feature set, which is 0.65 (≥0.6), judges it as valid, and inputs it into the model for price correction.

[0148] In an optional embodiment, the incremental training subunit uses the AdamW optimizer for parameter optimization, setting the initial learning rate to 0.0005, the batch size to 32, and the number of training epochs according to the formula. The calculation shows that N is the sample size, and the sample size satisfies 500≤N≤1000.

[0149] It should be noted that in this embodiment, the incremental training subunit refers to the functional subunit in the model self-update unit that performs partial parameter fine-tuning of the model. Its purpose is to improve the model update efficiency and adapt to dynamic market changes while ensuring the model's prediction accuracy. This subunit uses a sliding window to filter samples with high correlation (≥0.5) to price within the past 72 hours to ensure that the data used for training is targeted; it freezes the parameters of the first 3 layers of the LSTM model (preserving the basic temporal feature extraction capability) and only unfreezes the parameters of the last 2 layers and the event embedding layer to reduce the amount of training computation; it uses the AdamW optimizer to optimize the parameters, sets a reasonable initial learning rate (0.0005), batch size (32), and training epochs (calculated from the sample size), and controls the sample size to be within the range of 500-1000 samples to balance training effect and time cost. Through these designs, the incremental training subunit can quickly respond to new data, achieve efficient model iteration, and avoid the waste of resources from full retraining.

[0150] In one possible implementation, the incremental training subunit can dynamically adjust the batch size based on the sample size. For example, the batch size could be set to 32 when the sample size is 500, and to 64 when the sample size is 1000, to maintain stable training efficiency. Additionally, it should be noted that an upper limit can be set for parameter adjustments during training to prevent excessively high learning rates from causing model oscillations.

[0151] For example, in one feasible implementation, the incremental training subunit acquires event and price data from the past 72 hours, and selects 650 samples with a correlation coefficient ≥ 0.5 using Pearson coefficient calculation; freezes the parameters of the first 3 layers of the LSTM model, and unfreezes the last 2 layers and the event embedding layer; uses the AdamW optimizer, sets the initial learning rate to 0.0005, the batch size to 32, and calculates the number of training epochs to (650 / 32) = 21; fine-tunes the parameters according to the changes in the loss function during training, and completes training within 1.5 minutes, thereby improving the model's fitting accuracy to new data by 8%.

[0152] In an optional embodiment, the judgment and triggering subunit further includes a dual threshold verification mechanism: when the price deviation is ≥5% for three consecutive times, an absolute deviation or relative deviation verification is initiated. The absolute deviation is |actual price - provisional price| ≥ 10 yuan / ton, and the relative deviation is |actual price - provisional price| / provisional price ≥ 8%. When any verification dimension triggers the threshold condition, the system will immediately initiate a full parameter adjustment program to simultaneously optimize the algorithm model parameters within the system. The learning rate is dynamically increased according to the following formula:

[0153] ,

[0154] in, This represents the adjusted learning rate. This represents the learning rate before adjustment. This represents the actual price deviation. The formula is designed following the principle that "the greater the anomaly, the higher the learning rate increase." By introducing an adjustment coefficient of 0.2, the increase in the learning rate is linearly positively correlated with the degree of price deviation. When the price deviation reaches a critical threshold, the learning rate will increase by a base percentage of 20%, ensuring that the model can quickly adapt to market price changes and continuously optimize the provisional estimation accuracy.

[0155] It should be noted that in this embodiment, the dual-threshold verification mechanism refers to the verification rules used in the judgment and triggering subunit to initiate full parameter adjustment. Its purpose is to promptly trigger deeper model optimization when price prediction shows persistent deviations, ensuring the stability of prediction accuracy. This mechanism sets a condition for the number of consecutive deviations (3 consecutive price deviations ≥ 5%) as a pre-triggered condition, and then verifies it using two thresholds: absolute deviation and relative deviation. Absolute deviation refers to the absolute value of the difference between the actual price and the provisional price reaching or exceeding 10 yuan / ton, applicable to high-value coal types, and directly reflecting the absolute magnitude of price fluctuations. Relative deviation refers to the ratio of the absolute value of the difference between the actual price and the provisional price to the provisional price reaching or exceeding 8%, applicable to low-priced coal types, and better reflecting the relative magnitude of price fluctuations. When either deviation threshold is met, the system triggers full parameter adjustment and dynamically increases the learning rate through a specific formula to accelerate model convergence. This mechanism takes into account the characteristics of coal types at different price levels, avoids the limitations of a single threshold, and ensures that the model can be corrected promptly when prices fluctuate significantly.

[0156] In one possible implementation, the thresholds for absolute and relative deviations can be customized according to the coal type. For example, the absolute deviation threshold can be increased to 15 yuan / ton for scarce coal types, while the relative deviation threshold can be reduced to 6% for common coal types. Additionally, it should be noted that the number of consecutive deviations can be adjusted according to user needs. For instance, in scenarios requiring high price stability, the number of consecutive deviations can be set to 2.

[0157] For example, in one feasible implementation, the provisional price of a certain high-value coal type deviates from the actual price by ≥5% for three consecutive days. On the first day, the actual price is 890 yuan / ton and the provisional price is 870 yuan / ton (absolute deviation ≥ 10 yuan / ton). On the second day, the actual price is 895 yuan / ton and the provisional price is 875 yuan / ton (absolute deviation ≥ 10 yuan / ton). On the third day, the actual price is 900 yuan / ton and the provisional price is 880 yuan / ton (absolute deviation ≥ 10 yuan / ton). The system triggers a double threshold check. Because the absolute deviation threshold is met, a full parameter adjustment is initiated. At this time, the actual deviation is 5.7%, and the learning rate is adjusted from 0.0005 to 0.0005 × [1 + 0.2 × (5.7% - 5%) / 5%] = 0.0005 × 1.0028 = 0.0005014 according to the formula.

[0158] In an optional embodiment, the dynamic display and early warning linkage module 13 includes:

[0159] The visualization unit generates an event-price relationship chart, which includes a 3D interactive interface.

[0160] Provisional Price Series: The time axis is used as the horizontal axis to display the provisional value curve of coal prices affected by the event in real time;

[0161] Price Deviation Line: By comparing the provisional price under the influence of the event with the initial forecast price without the event assumption, a line is drawn with the percentage difference as the vertical axis to intuitively present the degree of price fluctuation deviation;

[0162] Feature contribution heatmap: Based on machine learning algorithms, the contribution of various influencing factors (such as changes in supply and demand, policy adjustments, etc.) to price fluctuations is analyzed, and the influence weight of different regions is marked in the form of a heatmap;

[0163] The early warning push unit automatically activates the early warning program when the estimated price fluctuation is ≥3%.

[0164] Generate a structured report containing event details, including event type, time of occurrence, and scope of impact;

[0165] Demonstrate the price correction logic by explaining the process of how events correct the price prediction model through flowcharts or formula analysis;

[0166] By combining industry knowledge base and historical cases, it intelligently generates response suggestions including inventory adjustments and changes in procurement strategies, and notifies users through multiple channels such as SMS, email, and APP push.

[0167] The feedback unit connects to the trading platform in real time via an API interface, periodically collects actual transaction price data, and calculates the deviation from the provisional price based on the mean squared error (MSE) algorithm.

[0168] Generate a deviation analysis report, including statistical data such as error distribution and fluctuation trends;

[0169] The deviation value and analysis results are fed back to the prediction model dynamic correction module 12, triggering the model adaptive optimization program. The model parameters are adjusted through the gradient descent algorithm to achieve continuous iterative optimization of the prediction model.

[0170] In one possible implementation, the visualization unit can support interactive chart operations. For example, when a user clicks on a point on the price deviation line, the specific event and its impact coefficient at that moment can be displayed. Additionally, it should be noted that the alert information from the alert push unit can be customized according to user roles, such as pushing purchase quantity suggestions to purchasing personnel and pricing suggestions to sales personnel.

[0171] For example, in one feasible implementation, the visualization unit generates a provisional price sequence for thermal coal in a certain region (850 yuan / ton, 865 yuan / ton, 880 yuan / ton, etc. for the next 7 days), a price deviation line (the deviation ratio compared with the assumption of no event), and a feature contribution heatmap (showing the feature weight of "increased transportation costs" at 0.6). When the provisional price on the 3rd day fluctuates by 4% compared with the initial forecast, the early warning push unit pushes an early warning information to the user, explaining the details of the "railway transportation restriction" event, the calculation process of price correction, and the suggestion of "stockpiling 5 days' worth of usage in advance". The feedback unit collects the actual transaction price daily, calculates the deviation value of 5 yuan / ton for the 1st day, and feeds this data back to the prediction model dynamic correction module 12.

[0172] In one optional embodiment, the visualization unit supports multi-terminal adaptation, including PC, mobile and large-screen devices, and the event-price correlation chart labels the event duration axis and the weights of core influencing factors.

[0173] It should be noted that in this embodiment, multi-terminal adaptation refers to the visualization unit's ability to display event-price correlation charts on different devices. This aims to meet users' viewing needs in various scenarios and improve the system's convenience and usability. Specifically, PC refers to personal computer terminals, suitable for detailed data viewing and analysis, displaying complete chart details and data parameters; mobile terminals refer to smartphones, tablets, and other mobile devices, allowing users to view key information anytime, anywhere, with a more concise and focused display; large-screen devices refer to large-size display devices used for monitoring or conference presentations, suitable for displaying overall trends and key indicators, enhancing visual impact. The event duration axis in the event-price correlation chart refers to the axis marking the time interval from the event's occurrence to its end in the chart. The core influencing factor weights refer to the event characteristics that have a significant impact on price and their weight values ​​marked in the chart. These annotations help users quickly understand the relationship between events and prices.

[0174] In one possible implementation, multi-terminal adaptation can automatically adjust the chart layout according to the device screen size, such as displaying a multi-chart split-screen layout on PCs and a single-chart carousel layout on mobile devices. Additionally, it should be noted that the event duration axis can support dynamic interaction; for example, users can drag the timeline to view the price and event details at the corresponding moment.

[0175] For example, in one feasible implementation, the visualization unit displays a split-screen interface for PC users, including a provisional price series, a price deviation line graph, and a feature contribution heatmap, with the duration axis of the "environmental protection production restriction" event (1-3 months) and the feature weight of "implementation strength" at 0.72. For mobile users, it displays a simplified provisional price trend chart, with the event duration axis and core weights labeled. For large-screen devices, it displays a full-screen overlay chart of price trends and event impacts, highlighting price fluctuations and corresponding events at key time points. Users can choose different terminals to view and obtain the information they need based on their own scenarios.

[0176] In this embodiment, by sensing multi-source emergencies in real time and converting them into structured data, combined with a dynamically corrected prediction model, accurate preliminary estimates of coal prices are achieved, thereby improving the timeliness and accuracy of price predictions. Through multi-dimensional feature fusion and a model self-updating mechanism, the system can adapt to market changes, enhancing its dynamic response capability to the impact of various events. Through visualization and early warning linkage, intuitive information and timely reminders are provided to users. Combined with a closed-loop feedback optimization model, the system assists in decision-making and continuously improves system performance.

[0177] Furthermore, it should be noted that the combination of the various technical features in this case is not limited to the combination methods described in the claims of this case or the combination methods described in the specific embodiments. All technical features described in this case can be freely combined or combined in any way, unless they contradict each other.

[0178] It should be noted that the above examples are merely specific embodiments of the present invention, and the present invention is obviously not limited to the above embodiments, with many similar variations. All modifications that can be directly derived or conceived by those skilled in the art from the content disclosed in this invention should fall within the protection scope of this invention.

[0179] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An intelligent automatic provisional estimation and dynamic display system for coal prices, characterized in that, include: The real-time perception and classification module for emergencies is used to collect public information from multiple sources and convert it into structured event data. The structured event data includes the event type, impact level weight, GeoHash encoding of the coverage area, and duration. The predictive model dynamic correction module receives the structured event data and generates a provisional price by combining it with historical price data. The predictive model dynamic correction module includes a basic predictive model, an event embedding layer, and a model self-updating unit. The event embedding layer performs multi-dimensional feature fusion on the structured event data, and the model self-updating unit adjusts the model parameters based on price deviation data. The dynamic display and early warning linkage module receives the estimated price and feature contribution data, performs visualization display, pushes early warning information according to the price fluctuation range, and feeds back the deviation between the actual price and the estimated price to the prediction model dynamic correction module. The event embedding layer includes: The spatial dimension processing unit calculates the percentage of the intersection area between the event coverage area and the target area to obtain the regional correlation coefficient. The calculation formula is as follows: in, Indicates the degree of regional correlation. This represents the intersection area of ​​the event-covered grid and the target grid. The total area of ​​the target grid is represented by the event coverage grid, which is the grid corresponding to the GeoHash code of the coverage area in the structured event data, and the target grid is the grid corresponding to the GeoHash code of the price region to be predicted. The time-dimensional processing unit uses different weight decay formulas based on the event duration: When the duration is ≤7 days ; When the duration is >7 days ; in, This represents the weight value at time t. This represents the initial impact level weight determined by the impact level quantification unit, and t represents the number of days the event has lasted. The type dimension processing unit uses a sliding window weighted average to calculate the influence coefficient of the current event. The calculation formula is as follows: in, This represents the impact coefficient of the current event. This represents the magnitude of the impact of the i-th similar historical event. Let represent the time decay factor of the i-th historical similar event, and , where n represents the number of months since the i-th historical similar event occurred, and the sliding window size is the last 5 similar events; The attention mechanism unit sets up multiple attention heads to focus on different event feature dimensions, and calculates the feature weights using a formula: in, Represents the feature vector of the current event. Represents a vector of historical event feature base. This indicates the feature dimension and has a value of 64. This represents the matrix transpose operation; The model self-updating unit includes: The judgment and triggering subunit calculates the cosine similarity between the feature vector output by the event embedding layer and the historical effective feature set after the feature vector is output. If the similarity is ≥0.6, it is input into the basic prediction model for price correction. If the similarity is <0.6, it is returned to the real-time perception and classification module of the sudden event for reprocessing. The dynamic display and early warning linkage module includes: The visualization unit generates an event-price correlation chart, which includes a provisional price series, a price deviation line, and a feature contribution heatmap. The price deviation is the ratio of the difference between the provisional price under the influence of the event and the initial predicted price without the event assumption.

2. The intelligent coal price automatic provisional estimation and dynamic display system according to claim 1, characterized in that, The real-time emergency perception and classification module includes: The multi-source event acquisition unit obtains publicly available information by connecting to government open data interfaces, industry legal information service platform data interfaces, meteorological department official early warning release channels, and authorized social media data services; The keyword extraction and classification unit extracts event keywords from the publicly available information and uses a pre-trained BERT model to classify events into policy, natural, supply chain, market, security, and environmental categories. The impact level quantification unit, based on the evaluation matrix trained on historical cases, determines the impact level weight, GeoHash encoding of the coverage area, and duration of the event, and generates the structured event data.

3. The intelligent coal price automatic provisional estimation and dynamic display system according to claim 2, characterized in that, The impact level quantification unit includes: The spatial range coding subunit uses GeoHash coding to hierarchically identify the event coverage range, with the mining area level using 8-bit coding and the regional level using 6-bit coding; The time span is used to determine the sub-units, and the duration range of the event is determined based on the nature of the event and historical data; The weighting sub-unit calculates the impact level weight of an event based on its type, scope of influence, and duration using an evaluation matrix.

4. The intelligent coal price automatic provisional estimation and dynamic display system according to claim 1, characterized in that, The basic prediction model is an LSTM model, which generates an initial price prediction sequence based on historical price data over a preset time period.

5. The intelligent coal price automatic provisional estimation and dynamic display system according to claim 1, characterized in that, The model self-updating unit also includes: The incremental training subunit uses a sliding window to filter samples with an "event-price" correlation of ≥0.5 within the past 72 hours, freezes the parameters of the first 3 layers of the LSTM model, and unfreezes the parameters of the last 2 layers and the event embedding layer for fine-tuning training. The correlation is calculated using the Pearson coefficient. The abnormal event handling subunit, when dealing with a novel event without historical reference, invokes the similarity event retrieval engine. Based on BERT sentence vectors, it calculates the event text similarity, retrieves the top 3 historical events with a similarity ≥ 0.7 as references, and initiates an online learning mode to update the influence coefficient. The update formula is: in, This represents the updated impact coefficient. This represents the influence coefficient before the update. This represents the difference between the actual price and the estimated price. This indicates a provisional price.

6. The intelligent coal price automatic provisional estimation and dynamic display system according to claim 5, characterized in that, The incremental training subunit uses the AdamW optimizer for parameter optimization, with an initial learning rate of 0.0005 and a batch size of 32. The number of training epochs is determined by the formula... The calculation shows that N is the sample size, and the sample size satisfies 500≤N≤1000.

7. The intelligent coal price automatic provisional estimation and dynamic display system according to claim 1, characterized in that, The judgment and triggering subunit also includes a dual threshold verification mechanism: when the price deviation is ≥5% for 3 consecutive times, an absolute deviation or relative deviation verification is initiated. The absolute deviation is |actual price - provisional price| ≥ 10 yuan / ton, and the relative deviation is |actual price - provisional price| / provisional price ≥ 8%. If either condition is met, a full parameter adjustment is triggered, and the learning rate is dynamically increased according to the following formula: in, This represents the adjusted learning rate. This represents the learning rate before adjustment. This indicates the actual price deviation.

8. The intelligent coal price automatic provisional estimation and dynamic display system according to claim 1, characterized in that, The dynamic display and early warning linkage module also includes: The early warning push unit pushes early warning information to users when the estimated price fluctuation is ≥3%. The early warning information includes event details, price correction logic, and response suggestions. The feedback unit periodically collects actual transaction prices, calculates the deviation from the estimated prices, and feeds it back to the prediction model's dynamic correction module.

Images