A power plant coal matching and procurement decision method and system

Abstract

This invention discloses a method and system for coal matching and procurement decision-making in power plants, relating to the field of power plant fuel management technology. It addresses the problems of low matching degree between coal sources and unit demand, and high procurement costs and low decision-making efficiency caused by reliance on manual experience in existing power plant coal procurement decision-making methods. The method includes: acquiring multi-source heterogeneous data; processing the multi-source heterogeneous data to generate a standardized feature dataset; inputting the standardized feature dataset into an AI fusion analysis model that integrates gradient boosting trees and attention mechanisms, outputting cost prediction results and key indicator weights; and employing an improved hybrid particle swarm optimization algorithm to iteratively determine the optimal solution of the four-dimensional variables of the multi-objective function through a dynamic adjustment strategy of inertia weights, thereby obtaining the target procurement plan. The coal matching and procurement decision-making method provided by this invention improves the matching degree between coal sources and unit demand in coal procurement decisions, reduces procurement costs, and increases decision-making efficiency.

Description

Technical Field

[0001] This invention relates to the field of power plant fuel management technology, and in particular to a method and system for power plant coal matching and procurement decision-making. Background Technology

[0002] In the thermal power generation industry, coal costs account for a large proportion of the total costs of thermal power plants. The scientific nature of coal procurement decisions directly affects the safe operation, environmental compliance, and economic benefits of power plants. Some power plants overemphasize the price per ton of coal during procurement, neglecting the matching of coal quality parameters with boiler design conditions and the treatment capacity of environmental protection devices. The key indicators of the selected coal, such as calorific value, volatile matter, and sulfur content, deviate from the allowable range for the unit, which not only increases the unit's operation and maintenance costs but also increases the risk of unplanned shutdowns. In addition, most power plants rely on manual experience for procurement decisions, which easily leads to high procurement costs and low decision-making efficiency.

[0003] Therefore, there is an urgent need for a method and system for power plant coal matching and procurement decisions. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for matching and purchasing coal in power plants, which solves the problems of low matching degree between coal source and unit demand, high procurement cost and low decision-making efficiency caused by reliance on manual experience in existing coal purchasing decision-making methods for power plants.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] In a first aspect, the present invention provides a method for power plant coal matching and procurement decision-making, comprising:

[0007] Acquire multi-source heterogeneous data, including power plant-side data, coal source-side data, supply chain-side data, and market-side data;

[0008] The multi-source heterogeneous data is processed to generate a standardized feature dataset;

[0009] A standardized feature dataset is input into an AI fusion analysis model that integrates gradient boosting trees and attention mechanisms, and the model outputs cost prediction results and key indicator weights. The cost prediction results include predicted values ​​for power generation coal consumption, environmental treatment costs, and transportation loss costs.

[0010] Based on the cost prediction results and the weights of the key indicators, an improved hybrid particle swarm optimization algorithm is adopted. The optimal solution of the four-dimensional variables of the multi-objective function is determined iteratively through a dynamic adjustment strategy of inertia weight, and the target procurement plan is obtained. The four-dimensional variables include coal source, procurement quantity, transportation mode and delivery time.

[0011] Optionally, the step of using an improved hybrid particle swarm optimization algorithm, based on the cost prediction results and the weights of the key indicators, to iteratively determine the optimal solution of the four-dimensional variables of the multi-objective function through a dynamic adjustment strategy of inertia weights, and obtaining the target procurement plan includes:

[0012] A four-dimensional objective function for coal procurement is constructed by setting multi-dimensional objectives and constraints. The multi-dimensional objectives include at least the following: the overall procurement cost meets a preset cost threshold, the supply chain risk meets a preset risk threshold, the environmental compliance rate meets a preset environmental threshold, and the delivery time is within a preset delivery period. The constraints include coal quality constraints, supplier constraints, inventory constraints, and policy constraints.

[0013] Particle initialization is performed according to the feature weighting guidance strategy to obtain the target initial particle swarm;

[0014] Based on the dynamic adjustment strategy of inertial weight and the genetic crossover mutation operator, starting from the target initial particle swarm, the position and velocity of each particle in the target initial particle swarm are iteratively updated until the number of iterations reaches the preset number, and the optimal solution of the four-dimensional objective function corresponding to the particle is determined as the target procurement plan.

[0015] Optionally, the dynamic adjustment strategy based on inertial weights and the genetic crossover and mutation operator, starting from the target initial particle swarm, iteratively updates the position and velocity of each particle in the target initial particle swarm until the number of iterations reaches a preset number, and determines the optimal solution of the four-dimensional objective function corresponding to the particle as the target procurement plan, including:

[0016] A linearly decreasing first inertia weight dynamic adjustment strategy and a genetic crossover mutation operator are adopted to iteratively update each particle of the target initial particle swarm according to the coal source selection dimension and transportation mode dimension until the number of iterations reaches a first preset number, thus obtaining a first particle swarm; the four-dimensional objective function value corresponding to the particle position in the first particle swarm is less than a first threshold; the four-dimensional objective function value is calculated based on the cost prediction result and the key indicator weights;

[0017] Using fixed inertia weights and genetic crossover mutation operators, the particles in the first particle group are iteratively updated according to the procurement quantity allocation dimension until the number of iterations reaches a second preset number, thus obtaining a second particle group; the comprehensive procurement cost value corresponding to each particle in the second particle group is less than the comprehensive procurement cost value in the first particle group, and the supply chain risk value corresponding to each particle in the second particle group is less than or equal to the supply chain risk value corresponding to each particle in the first particle group.

[0018] A dynamic adjustment strategy of linearly decreasing second inertial weights and a genetic crossover mutation operator are adopted to iteratively update each particle of the second particle swarm according to the procurement quantity allocation dimension and the delivery time dimension until the number of iterations reaches the first preset number, thus obtaining a third particle swarm; the position vector of the particle with the smallest four-dimensional objective function value in the third particle swarm is determined as the target procurement scheme.

[0019] Optionally, the step of inputting the standardized feature dataset into the AI ​​fusion analysis model that integrates gradient boosting trees and attention mechanisms, and outputting cost prediction results and key indicator weights, includes:

[0020] A standardized feature dataset is input into a basic prediction model to predict the cost and output a feature activation matrix. The basic prediction model includes a feature attention enhancement layer, a LightGBM model, and a residual correction layer.

[0021] The SHAP value is used to quantify the feature contribution value of each feature in the standardized feature dataset, and the feature contribution value is optimized according to the feature activation matrix to obtain the target feature weight of each feature. The target feature weights are sorted in descending order, and the top N target feature weights are identified as key indicator weights; N is a positive integer.

[0022] The update interval is monitored. When the update interval reaches a preset time, the parameters of the basic prediction model are adjusted according to the actual procurement data and the cost prediction results, and cost prediction is performed based on the adjusted basic prediction model.

[0023] Optionally, the standardized feature dataset includes coal quality feature groups, price feature groups, supply chain feature groups, and unit operation feature groups; the step of inputting the standardized feature dataset into the basic prediction model to predict the cost prediction result includes:

[0024] The standardized feature dataset is input into the basic prediction model. The feature attention enhancement layer uses a multi-head attention mechanism to weight and enhance different feature groups, constructing a feature group weight allocation network. The LightGBM model uses a gradient boosting tree to predict costs based on the weights allocated by the feature group weight allocation network, and obtains the output result.

[0025] An error correction layer is used to correct the output results to obtain the cost prediction results.

[0026] Optionally, the step of initializing particles according to a feature weight-guided strategy to obtain the target initial particle swarm includes:

[0027] Generate a preset number of first initial particle groups; the first initial particle groups include at least a preset proportion of particles whose coal source selection dimension meets the coal quality standards.

[0028] The procurement quantity allocation dimension of each particle in the first initial particle swarm is adjusted to the preset procurement quantity to obtain the second initial particle swarm.

[0029] The arrival time dimension of each particle in the second initial particle swarm is adjusted so that the arrival deviation of the adjusted particles is within a preset deviation threshold, thus obtaining the target initial particle swarm; the arrival deviation is the difference between the current arrival time and the optimal arrival time.

[0030] Optionally, the step of quantifying the feature contribution value of each feature in the standardized feature dataset using SHAP values, and optimizing the feature contribution value according to the feature activation matrix to obtain the target feature weight of each feature includes:

[0031] Formula used:

[0032] ;

[0033] Calculate the feature contribution value of each feature in the standardized feature dataset;

[0034] in, This represents the complete value after performing SHAP decomposition on the i-th sample. As the global baseline value, is the feature contribution value of the j-th feature to the i-th sample; m is the total number of core features involved in the comprehensive cost prediction, that is, the number of feature dimensions in the standardized feature dataset, corresponding to the dimension of the standardized feature dataset.

[0035] Formula used:

[0036] ;

[0037] The feature contribution values ​​are optimized to obtain the target feature weights for each feature;

[0038] in, The target feature weight for the j-th feature is... Let j be the average activation intensity of the j-th feature. Let be the average activation intensity of the k-th feature, where k is the index variable of the feature. is the feature activation matrix.

[0039] Optionally, processing the multi-source heterogeneous data to generate a standardized feature dataset includes:

[0040] Box plots are used to filter out suspected outliers from the multi-source heterogeneous dataset; the isolated forest algorithm is used to identify the suspected outliers, determine the target outliers, and delete the target outliers from the multi-source heterogeneous dataset to obtain the first dataset;

[0041] The first dataset is divided into clusters using a clustering algorithm, and missing values ​​are imputed within each cluster using a weighted k-nearest neighbor algorithm to obtain the second dataset.

[0042] The data in the second dataset are divided according to their feature attributes to obtain multiple sets of feature data. The Pearson correlation coefficient is used to select target data that is strongly correlated with the overall cost from the multiple sets of feature data, and the target data is standardized to obtain a standardized feature dataset.

[0043] Compared with the prior art, the present invention provides a method for power plant coal matching and procurement decision-making, comprising:

[0044] This invention acquires heterogeneous data from multiple sources; processes the heterogeneous data to generate a standardized feature dataset; inputs the standardized feature dataset into an AI fusion analysis model that integrates gradient boosting trees and attention mechanisms, outputting cost prediction results and key indicator weights; based on the cost prediction results and key indicator weights, an improved hybrid particle swarm optimization algorithm is used to iteratively determine the optimal solution of the four-dimensional variables of the multi-objective function through a dynamic adjustment strategy of inertia weights, thus obtaining the target procurement plan. This invention integrates four types of heterogeneous data from the power plant side, coal source side, supply chain side, and market side to construct a comprehensive data foundation for procurement decisions, effectively avoiding the limitations of traditional methods due to single data sources and delayed market response. The AI ​​fusion analysis model that integrates gradient boosting trees and attention mechanisms can quantify the matching relationship between coal quality and generating units, improving the accuracy of coal and unit matching. By dynamically optimizing the multi-objective function through an improved particle swarm optimization algorithm, the optimal matching solution is output, which can reduce the overall procurement cost, including procurement cost, transportation cost, and environmental cost. This method reduces reliance on human intervention, avoids subjective errors caused by decision-making experience, improves the scientific nature and stability of procurement decisions, and increases decision-making efficiency.

[0045] Secondly, the present invention also provides a power plant coal matching and procurement decision-making system, comprising:

[0046] A multi-source data acquisition module is used to acquire multi-source heterogeneous data, including power plant-side data, coal source-side data, supply chain-side data, and market-side data.

[0047] The big data preprocessing module is used to process the multi-source heterogeneous data and generate a standardized feature dataset;

[0048] The AI ​​fusion analysis module is used to input standardized feature datasets into an AI fusion analysis model that integrates gradient boosting trees and attention mechanisms, and outputs cost prediction results and key indicator weights; the cost prediction results include predicted values ​​for power generation coal consumption, environmental treatment costs, and transportation loss costs.

[0049] The multi-objective optimization decision module is used to determine the optimal solution of the four-dimensional variables of the multi-objective function based on the cost prediction results and the weights of key indicators, using an improved hybrid particle swarm optimization algorithm and a dynamic adjustment strategy for inertial weights, thereby obtaining the target procurement plan; the four-dimensional variables include coal source, procurement quantity, transportation method, and delivery time.

[0050] Optionally, the AI ​​fusion analysis module includes:

[0051] The cost prediction submodule is used to input the standardized feature dataset into the basic prediction model, predict the cost prediction result, and output the feature activation matrix; the basic prediction model includes a feature attention enhancement layer, a LightGBM model, and a residual correction layer;

[0052] The impact factor quantification submodule is used to quantify the feature contribution value of each feature in the standardized feature dataset using SHAP values, optimize the feature contribution value according to the feature activation matrix to obtain the target feature weight of each feature, sort the target feature weights in descending order, and confirm the top N target feature weights as key indicator weights; N is a positive integer.

[0053] The model self-optimization submodule is used to monitor the update interval. When the update interval reaches a preset time, the parameters of the basic prediction model are adjusted according to the actual procurement data and the cost prediction results, and cost prediction is performed based on the adjusted basic prediction model.

[0054] Compared with the prior art, the beneficial effects of the power plant coal matching and procurement decision-making system provided by the present invention are the same as the beneficial effects of the power plant coal matching and procurement decision-making method described in the above technical solution, and will not be repeated here. Attached Figure Description

[0055] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0056] Figure 1 A flowchart of a power plant coal matching and procurement decision-making method provided by the present invention;

[0057] Figure 2This is a schematic diagram of the structure of a power plant coal matching and procurement decision-making system provided by the present invention. Detailed Implementation

[0058] To facilitate a clear description of the technical solutions in the embodiments of the present invention, the terms "first" and "second" are used to distinguish identical or similar items with essentially the same function and effect. For example, the first threshold and the second threshold are merely used to distinguish different thresholds and do not limit their order. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that the terms "first" and "second" are not necessarily different.

[0059] It should be noted that in this invention, the terms "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0060] In this invention, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can represent: a, b, c, a combination of a and b, a combination of a and c, a combination of b and c, or a, b, and c, where a, b, and c can be single or multiple.

[0061] Before introducing the embodiments of the present invention, the relevant terms involved in the embodiments of the present invention are first defined as follows:

[0062] Leaf-wise is a decision tree growth strategy employed by LightGBM. It achieves more efficient error reduction and accuracy improvement by prioritizing the splitting of the node with the largest splitting gain among all current leaf nodes.

[0063] Dropout is a regularization technique used when training neural networks to prevent overfitting.

[0064] The WebSocket protocol is an application layer communication protocol based on the TCP protocol. It solves the pain point of real-time communication in the "request-response" mode of the HTTP protocol, and realizes full-duplex bidirectional real-time data transmission between the client and the server.

[0065] A RESTful API is an API design specification that follows the REST architectural style. REST stands for Representational State Transfer.

[0066] The Weighted k-Nearest Neighbors (KNN) algorithm is an improved version of the k-nearest neighbor algorithm. It improves the accuracy of classification or regression by introducing a distance weighting mechanism. Its core idea is to assign higher weights to neighbors closer to the query point and lower weights to neighbors farther away in k-nearest neighbor classification or regression, thereby more reasonably reflecting the influence of neighbors on the query point.

[0067] To address the problems in existing power plant coal procurement, such as poor matching between coal sources and unit demand, high procurement costs due to reliance on manual experience, low decision-making efficiency, and lagging market response, this invention provides a method and system for power plant coal matching and procurement decision-making. Through multi-source data integration, intelligent algorithm modeling, and dynamic optimization, it achieves precise coal source matching and optimal procurement plan decisions, significantly reducing overall procurement costs and improving the scientific rigor and timeliness of procurement decisions. The following description, in conjunction with the accompanying drawings, will illustrate this approach.

[0068] See Figure 1 The present invention provides a method for power plant coal matching and procurement decision-making, comprising the following steps:

[0069] Step 100: Obtain multi-source heterogeneous data.

[0070] Multi-source heterogeneous data includes power plant-side data, coal source-side data, supply chain-side data, and market-side data. Power plant-side data includes rated load of generating units, boiler design coal type parameters, historical coal consumption data for power generation, and operating cost data of power plant operating systems. Coal source-side data includes industrial analysis indicators, elemental analysis indicators, pithead prices, and port prices of major coal sources. Supply chain-side data includes unit freight costs, transportation cycles, historical on-time delivery rates of suppliers, and coal quality compliance rates for different transportation modes. Market-side data includes industry environmental protection policies and updated energy consumption standards.

[0071] Multi-source heterogeneous data is collected through designated real-time and periodic acquisition channels. The real-time channel uses the OPC UA protocol to connect to the power plant's SIS system and fuel management system, setting up a data cache queue to synchronize 28 key parameters, including unit operating parameters, real-time power generation, real-time coal inventory, boiler operating status data, and real-time coal quality monitoring data, at preset intervals. These preset intervals can be set according to needs, such as 15 minutes or 20 minutes. The periodic channel connects to the coal trading platform, supplier management system, logistics monitoring platform, and industry policy database via a RESTful API interface, synchronizing 92 data items daily, including the previous day's coal price, industrial analysis indicators such as calorific value, ash content, and sulfur content, elemental analysis indicators, supplier inventory, unit freight costs for different transportation methods, transportation cycle, supplier historical on-time delivery rate, and coal quality compliance rate. Web crawling technology is used to capture updated industry environmental policies and energy consumption standards to supplement the market-side data.

[0072] Step 200: Process the multi-source heterogeneous data to generate a standardized feature dataset.

[0073] The processing steps include outlier detection: using box plots to filter out suspected outliers from the multi-source heterogeneous dataset; using the isolated forest algorithm to identify suspected outliers, determine target outliers, and delete target outliers from the multi-source heterogeneous dataset to obtain the first dataset; the target outliers include coal quality detection outliers and transportation cost outliers. Missing value imputation: Based on attributes such as coal source origin, coal type, and mining method, a clustering algorithm is used to divide the first dataset into clusters, grouping similar coal sources into the same cluster. Within each cluster, a weighted k-nearest neighbor algorithm is used to impute missing values ​​for coal quality indicators. The weights are dynamically allocated based on coal source similarity and data reliability, resulting in the second dataset. Feature engineering: The data in the second dataset is divided according to feature attributes, resulting in multiple sets of feature data, including basic features, derived features, and correlated features. Basic features are the originally collected coal quality, price, and transportation data; derived features include price fluctuation coefficient, coal quality stability score, and supplier reputation rating; correlated features include coal quality and unit compatibility coefficient, and the matching degree between transportation cycle and coal inventory. Pearson correlation coefficient is used to screen target data strongly correlated with comprehensive cost from the multiple sets of feature data, and the target data is standardized to obtain a standardized feature dataset.

[0074] Step 300: Input the standardized feature dataset into the AI ​​fusion analysis model that integrates gradient boosting tree and attention mechanism, and output the cost prediction results and key indicator weights.

[0075] The cost forecast results include predicted values ​​for power generation coal consumption, environmental treatment costs, and transportation loss costs.

[0076] The AI ​​fusion analysis model is based on the LightGBM algorithm with a fusion attention mechanism to build the basic prediction model. The basic prediction model adopts a dual-branch hybrid architecture, containing two core components: a feature attention extractor and a LightGBM predictor. Traditional methods place the attention mechanism before the input; this method allows each regression tree to calculate an attention score on the current sample set at each split node, and then use that score as a multiplier for the split gain, achieving white-box attention. The original dynamic sample weights are sorted only by gradient, and additionally, samples exceeding environmental standards are forcibly retained to prevent rare but high-cost scenarios from being sampled out.

[0077] Step 400: Based on the cost prediction results and the weights of the key indicators, an improved hybrid particle swarm optimization algorithm is used to iteratively determine the optimal solution of the four-dimensional variables of the multi-objective function through a dynamic adjustment strategy of inertia weight, thereby obtaining the target procurement plan.

[0078] The four-dimensional variables include coal source, purchase volume, transportation method, and delivery time.

[0079] An improved hybrid particle swarm optimization algorithm is proposed, incorporating the genetic crossover and mutation operator from a genetic algorithm. In the initial iteration, the inertia weight is set to 0.9-1.0 to ensure global search capability; in the middle iteration, the inertia weight linearly decreases to 0.5-0.7 to balance global and local searches; in the later iteration, the inertia weight is set to 0.3-0.5 to focus on local optima. The algorithm optimizes four variables—coal source selection, procurement allocation, transportation mode, and delivery time—and outputs the optimal solution.

[0080] This application integrates four types of heterogeneous data from power plants, coal sources, supply chains, and markets, transforming unstructured procedural texts into interpretable embeddings to build a comprehensive data foundation for procurement decisions. This effectively avoids the limitations of traditional methods, such as relying on a single data source and experiencing delayed market response. An AI fusion analysis model employing gradient boosting and attention mechanisms can quantify the matching relationship between coal quality and generating units, improving the accuracy of coal-unit matching. By improving the particle swarm optimization algorithm for dynamic optimization of multi-objective functions, the optimal matching solution is output, reducing overall procurement costs, including procurement, transportation, and environmental costs. This method reduces reliance on human intervention, avoids subjective errors stemming from decision-making experience, enhances the scientific rigor and stability of procurement decisions, and improves decision-making efficiency.

[0081] based on Figure 1 In addition to the method described herein, this specification also provides some specific implementation methods of this method, which will be described below.

[0082] As an optional approach, step 300 can be implemented based on S1-S3:

[0083] S1: Input the standardized feature dataset into the basic prediction model to predict the cost prediction result and output the feature activation matrix; the basic prediction model includes a feature attention enhancement layer, a LightGBM model and a residual correction layer; the standardized feature dataset includes coal quality feature group, price feature group, supply chain feature group and unit operation feature group;

[0084] The feature activation matrix is ​​the feature activation matrix generated for each sample during the prediction process by the basic prediction model. It has a dimension of n×m, where n is the total number of samples and m is the total number of core features involved in the comprehensive cost prediction.

[0085] Specifically, S1 can be implemented based on S11-S12:

[0086] S11: The standardized feature dataset is input into the basic prediction model. The feature attention enhancement layer uses a multi-head attention mechanism to weight and enhance different feature groups, and constructs a feature group weight allocation network. The LightGBM model uses a gradient boosting tree to predict costs based on the weights allocated by the feature group weight allocation network, and obtains the output result.

[0087] The basic prediction model adopts a three-stage architecture consisting of a feature attention enhancement layer, a LightGBM model, and a residual correction layer, as detailed below:

[0088] Input layer: Receives a standardized feature dataset containing more than 60 core features, including 12 coal quality features such as calorific value, ash content, and sulfur content; 8 price features such as pithead price and transportation price; 15 supply chain features such as transportation cycle and on-time delivery rate; and 25 unit operation features such as rated load and boiler status.

[0089] The attention calculation formula for the feature attention enhancement layer is shown in formula (1):

[0090] (1)

[0091] in, For attention mechanisms, For querying the matrix, The key matrix, For value matrices, , , Each feature group's data is generated through a linear transformation. is the dimension of the key matrix.

[0092] The feature group weight allocation network is shown in Equation (2):

[0093] (2)

[0094] in, , Highlight the impact of core feature groups; For attention output, The attention results are for the coal quality characteristic groups. The attention results are for the price feature group. The attention results are for the supply chain feature groups. The attention results are for the unit operation feature groups. , , as well as All are coefficients.

[0095] The LightGBM model is constructed using a tree structure with cross-feature awareness, with the following specific settings:

[0096] Decision tree growth strategy: A depth-limited leaf-wise growth method is adopted, with the maximum tree depth set to 8-12, which is adaptively adjusted based on the feature dimension.

[0097] Classification criteria: A custom cost-sensitive split gain function is defined, as shown in formula (3):

[0098] (3)

[0099] in, The cost-sensitive split gain function. Let be the cost loss value of node t. This is the depth penalty coefficient. The node depth is used as a criterion to balance model accuracy and generalization ability. Let be the total number of samples contained in the node t to be split in the decision tree, and n be the total number of samples. Let be the left child of the node t to be split. The number of samples in the left child node. Let be the right child of the node t to be split. This represents the number of samples in the right child node.

[0100] Regularization mechanism: L1 regularization is introduced to suppress feature redundancy, with a coefficient of 0.01-0.05. Combined with the Dropout layer to prevent overfitting, the dropout rate is 0.2.

[0101] S12: The residual correction layer uses a linear residual compensation model to correct the error in the output result, thereby obtaining the cost prediction result.

[0102] The residual compensation model is shown in formula (4):

[0103] (4)

[0104] in, For cost forecast results, For the output results, The original feature vector, This is the weight matrix. For bias terms, This represents the residual weight.

[0105] Optionally, the basic prediction model can be optimized as follows: using the time series partitioning method, it is divided into training set, validation set, and test set in a 7:2:1 ratio; the loss function adopts a weighted mixed loss function, as shown in formula (5):

[0106] (5)

[0107] in, For the comprehensive error, Mean square error, The predicted values ​​of the basic prediction model, This is the actual value. The mean absolute error, The mean absolute percentage error.

[0108] The optimizer uses an adaptive momentum estimation optimizer with an initial learning rate of 0.03. It is dynamically adjusted using a cosine annealing strategy and has an iteration cycle of 100 rounds.

[0109] Early stopping mechanism: When the validation set loss does not decrease for 15 consecutive rounds, training is automatically stopped and the optimal model parameters are saved.

[0110] S2: The SHAP value is used to quantify the feature contribution value of each feature in the standardized feature dataset, and the feature contribution value is optimized according to the feature activation matrix to obtain the target feature weight of each feature. The target feature weights are sorted in descending order, and the top N target feature weights are identified as key indicator weights; N is a positive integer.

[0111] Specifically, formula (6) is used:

[0112] (6)

[0113] Calculate the feature contribution value of each feature in the standardized feature dataset;

[0114] in, This represents the complete value after performing SHAP decomposition on the i-th sample. As the global baseline value, is the feature contribution value of the j-th feature to the i-th sample; m is the total number of core features involved in the comprehensive cost prediction, that is, the number of feature dimensions in the standardized feature dataset, corresponding to the dimension of the standardized feature dataset.

[0115] Formula (7) is used:

[0116] (7)

[0117] The feature contribution values ​​are optimized to obtain the target feature weights for each feature;

[0118] in, The target feature weight for the j-th feature is... Let be the average activation intensity of the j-th feature, and k be the index variable of the feature. is the average activation intensity of the k-th feature; A is the feature activation matrix with dimensions n×m, where n is the total number of samples.

[0119] The total weight of calorific value, sulfur content, and ash content among all characteristics is greater than or equal to 75%. Key indicators include three coal quality indicators: calorific value, sulfur content, and ash content.

[0120] S3: Monitor the update interval of the basic prediction model. When the update interval reaches a preset time, adjust the parameters of the basic prediction model according to the actual procurement data and the cost prediction results, and perform cost prediction based on the adjusted basic prediction model.

[0121] Specifically, the parameter adjustment steps are as follows;

[0122] Formula (9) is used:

[0123] (9)

[0124] Calculate the overall error;

[0125] in, This is the total error. This represents the mean square error between the predicted and actual coal consumption for power generation. The mean square error between the predicted and actual environmental treatment costs. This represents the mean square error between the predicted and actual transportation loss costs.

[0126] The calculation formula is shown in formula (10):

[0127] (10)

[0128] in, Let i be the predicted coal consumption for power generation corresponding to the i-th sample. , where n is the actual coal consumption for power generation, and n is the total number of samples, where n is greater than 0.

[0129] and Calculation and The calculation principle is the same, so I will not repeat it here.

[0130] like The top 10 key features are selected based on the target feature weights of each feature. The hyperparameters of the LightGBM model, such as tree depth, learning rate, and number of leaf nodes, are adjusted by the Bayesian optimization algorithm. At the same time, the weight parameters of the feature attention enhancement layer are updated.

[0131] By establishing the aforementioned model self-iterative optimization mechanism and real-time data update mechanism, the procurement plan is updated every 6 hours and the model parameters are optimized every 30 days, ensuring that the decision-making plan always adapts to market changes and unit operating status, thereby improving the dynamic adaptability of decision-making.

[0132] Step 300, with its multi-dimensional data fusion and acquisition mechanism, integrates four heterogeneous data sources—power plant, coal source, supply chain, and market—for the first time. It converts unstructured procedural texts such as operating procedures, expert experience, and accident cases into interpretable vectors for model use, building a comprehensive data foundation for procurement decisions and addressing the issues of single data sources and delayed market response in traditional methods. Attention weights are explicitly embedded into the leaf node splitting criterion of LightGBM, achieving white-box attention and highlighting the impact of key coal quality indicators such as calorific value, sulfur content, and ash content, as well as real-time coal prices, on costs. This attention reduces cost prediction errors to within 3%, improving prediction accuracy by 15% compared to traditional models. By decomposing the contribution of each feature to the overall cost using SHAP values, it clarifies that the weight of calorific value, sulfur content, and ash content should not be less than 75%. Combined with global sensitivity analysis, it identifies the threshold impact of price fluctuations and transportation cycle changes on costs, providing a basis for optimized decision-making. This step can significantly improve the accuracy of coal source matching. By using an AI fusion analysis model to quantify the compatibility between coal quality and generating units, the matching accuracy is more than 40% higher than that of traditional methods. By monitoring and updating the update interval of the basic prediction model, it can adapt to cost changes.

[0133] As an alternative approach, step 400 can be implemented based on the following steps:

[0134] Step 1: Use formula (11):

[0135] (11)

[0136] Calculate the total cost forecast;

[0137] in, This is the overall cost forecast. This is the predicted value of coal consumption for power generation. This is a projected cost for environmental treatment. This is a predicted value for transportation loss costs; , , As a cost weight, , The price is per unit of electricity.

[0138] Step 2: Set multi-dimensional objectives and constraints to construct a four-dimensional objective function for coal procurement;

[0139] The multi-dimensional objectives include at least meeting the preset cost threshold for comprehensive procurement costs, the preset risk threshold for supply chain risks, the preset environmental compliance rate for environmental protection, and delivery time within the preset delivery period. The core of the multi-dimensional objectives is to optimize comprehensive benefits. The comprehensive procurement cost is the sum of procurement costs, transportation costs, environmental protection costs, and loss costs. Supply chain risks include supplier default risks and transportation interruption risks. The preset environmental compliance threshold is 100%. The preset cost threshold, preset risk threshold, and preset delivery period can all be set as needed.

[0140] The constraints include coal quality constraints, supplier constraints, inventory constraints, and policy constraints. Coal quality constraints include a calorific value greater than or equal to the unit's design requirements and a sulfur content less than or equal to the environmental standard's upper limit (e.g., sulfur content less than or equal to 0.6%). Supplier supply capacity must be greater than or equal to the purchase quantity, and the historical compliance rate must be greater than or equal to 95% (e.g., supplier A's maximum daily supply is 5000 tons). Inventory constraints include a transportation cycle less than or equal to the coal inventory warning line cycle minus 3 days (e.g., transportation time range: 3 to 5 days by rail, 102 days by road, and 7 to 10 days by sea). Policy constraints include compliance with the latest environmental and energy consumption policies.

[0141] Specifically, the four-dimensional objective function is shown in formula (12):

[0142] (12)

[0143] in, The four-dimensional objective function value. The total procurement cost is represented by a projected value, adjusted according to the weight of key indicators, such as sulfur content exceeding the standard. Adding a 20% environmental penalty cost, This is the supply chain risk coefficient, ranging from 0 to 1, calculated as follows: , This figure represents the risk of supplier default, calculated in reverse based on historical compliance rates. For example, if the compliance rate is 95%, then... It is 0.05. For example, during the rainy season, roads are subject to the risk of transportation disruptions. It is 0.2; This is the delivery time deviation coefficient, with a value ranging from 0 to 1. , The value is the coal inventory warning line period minus 3 days. The transportation cycle is, for example, a 4-day cycle for rail transportation. =0; To ensure environmental compliance, only when the coal quality indicators meet the following criteria are considered: calorific value ≥ 4500 kcal / kg, sulfur content ≤ 0.6%, and ash content ≤ 20%. It is 1 if it is true, otherwise it is 0. The weight for the overall procurement cost is set to 0.4. The weight for the supply chain risk coefficient is 0.2. The weight for the environmental compliance rate is set to 0.3. The weight of the delivery time deviation coefficient is 0.1. , , as well as The sum of is 1.

[0144] Step 3: Initialize particles according to the feature weighting guidance strategy to obtain the target initial particle swarm;

[0145] Particles are carriers of four-dimensional variables. The position vector of each particle corresponds to a complete set of procurement plans, the mathematical expression of which is: , Let be the position vector of the particle. The component of the particle's position vector in the coal source selection dimension takes the value of the available coal source number, such as 1-20, corresponding to 20 candidate coal sources; Assign a component to the particle's position vector in the dimension of the purchase quantity, with values ​​ranging from [0, 1] to [0]. ], This corresponds to the maximum supply capacity of the coal source, such as 0-5000 tons; The component of the particle's position vector in the mode of transport dimension, with values ​​of 1, 2, and 3 corresponding to rail, road, and sea transport, respectively; Let be the component of the particle's position vector in the arrival time dimension, with values ​​ranging from . , To achieve the shortest transportation cycle, This is equal to the coal storage warning line period minus 3 days, such as 1-4 days. The expression for the particle's velocity vector is: , Let V be the particle's velocity vector. Select the velocity component of the dimension for the coal source; Assign a speed component to the dimension of the purchase volume; The velocity component represents the mode of transport. As the velocity component of the delivery time dimension, the particle's velocity vector is used to update the particle's position during the iteration process, thereby optimizing the decision-making scheme.

[0146] The feature weighting guidance strategy is to initialize particles according to priority levels.

[0147] The first priority is coal quality compliance: Specifically, a preset number of initial particle groups are generated; the first initial particle group includes at least a preset proportion of particles whose coal source selection dimensions meet the coal quality standards; the environmental compliance rate of coal sources meeting the coal quality standards is 100%; the coal quality standards can be set according to needs, for example, coal sources with a calorific value ≥4500kcal / kg and sulfur content ≤0.6% meet the coal quality standards. The preset proportion can be 80% or 85%, etc. The remaining proportion of particles is randomly generated to explore potential high-quality coal sources.

[0148] The second priority is cost advantage: Specifically, the procurement quantity allocation dimension of each particle in the first initial particle swarm is adjusted to the preset procurement quantity to obtain the second initial particle swarm.

[0149] In practical applications, for the first initial particle, the particles are sorted from smallest to largest according to the predicted comprehensive cost of procurement. The initial positions of the particles in the top 50% of coal sources are shifted toward the upper limit of procurement allocation. The preset procurement quantity is a preset proportion of the upper limit of procurement allocation. For example, coal source A has the lowest cost, and the initial procurement quantity is set to 70% of its maximum supply capacity.

[0150] The third priority is timeliness matching: the arrival time dimension of each particle in the second initial particle swarm is adjusted so that the arrival deviation of the adjusted particles is within a preset deviation threshold, thus obtaining the target initial particle swarm; the arrival deviation is the difference between the current arrival time and the optimal arrival time.

[0151] In practical applications, the preset deviation threshold is: The values ​​in the vicinity. The adjustment process for each particle in the second initial particle swarm is as follows:

[0152] Adjustment of velocity direction: If the particle is currently If the delivery time is later than the optimal time, there is a risk of coal shortage, then the velocity vector will be... The direction is set to the negative direction, that is, let Decrease during iteration, towards Approaching.

[0153] If the particle is currently If the delivery time is earlier than the optimal time, there is a risk of inventory backlog. Set the direction to positive, that is, let Increase during iteration, towards convergence.

[0154] For other dimensions ( , , The speed and direction need to be combined. Adjustment requirements are linked to settings, such as If it needs to be reduced, then... Shift towards "shorter transportation cycles," such as optimizing from sea to road transport.

[0155] Adjustment of speed magnitude: using formulas (13) and (14):

[0156] (13)

[0157] (14)

[0158] Adjustment The amplitude is adjusted proportionally, while the speed of the associated dimension is also adjusted proportionally.

[0159] in: This is the current arrival time. This is the deviation correction factor. The greater the deviation upon arrival, The higher the value, the better. hour, ; hour, ; The speed reference value is set by the initial parameters of the algorithm, such as 0.2; the coefficient 0.6 in formula (14) can be set based on the strong correlation between the transportation mode and the delivery time to ensure that the adjustment of the transportation mode can match the optimization requirements of the delivery time.

[0160] For example, a certain particle is currently sky, If the deviation is 50%, then And the direction is negative, meaning in the next iteration It will decrease by 0.54 days, moving closer to 4 days, while The direction is to adjust towards "shorter transportation cycles" to reduce overall delivery deviations.

[0161] The above method uses a feature weight-guided strategy for particle initialization, which can avoid the random particle initialization of the traditional PSO algorithm, which easily generates a large number of invalid solutions that do not meet the coal quality requirements, such as in the case of low calorific value coal sources, and improve the accuracy of the output results.

[0162] Step 4: Based on the dynamic adjustment strategy of inertia weight and the genetic crossover mutation operator, starting from the target initial particle swarm, iteratively update the position and velocity of each particle in the target initial particle swarm until the number of iterations reaches the preset number, and determine the optimal solution of the four-dimensional objective function corresponding to the particle as the target procurement plan.

[0163] The iterative algorithm consists of three phases: global exploration, local optimization, and precise convergence.

[0164] For example, the number of iterations during the global exploration period is set to 1-50 generations, and the dynamic adjustment strategy of inertia weight is as follows: the inertia weight adjustment range is 0.9-1.0; the adjustment method is linear decrease; the crossover probability is 0.8-0.9, the mutation probability is 0.05-0.08, the core objective is: to traverse all coal sources and transportation mode combinations to avoid missing high-quality solutions; scenario adaptability: adaptable to situations where coal prices fluctuate greatly, it is necessary to quickly screen low-priced and qualified coal sources.

[0165] The number of iterations during the local optimization period is set to 51-150 generations. The dynamic adjustment strategy for inertia weight is as follows: the inertia weight adjustment range is 0.5-0.7; the adjustment method is fixed; the crossover probability is 0.6-0.7, and the mutation probability is 0.03-0.05. The core objective is to fine-tune the procurement volume allocation dimension based on the key indicator weights, such as increasing the proportion of high-calorific-value coal sources. Scenario adaptability: adaptable to situations where coal reserves are sufficient, it is necessary to balance cost and coal quality stability.

[0166] The precise convergence period consists of iterations from generation 151 to 200. The dynamic adjustment strategy for inertia weight is as follows: the inertia weight adjustment range is 0.3-0.5; the adjustment method is linear decrease; the crossover probability is 0.4-0.5, and the mutation probability is 0.01-0.02. The core objective is to lock in the optimal combination and optimize the delivery time window, such as adjusting transportation batches. Scenario adaptability: adaptable to situations where coal reserves are tight, ensuring that the transportation cycle does not exceed the warning line.

[0167] Specifically, step 4 can be achieved through the following steps:

[0168] Step 41: Using a linearly decreasing first inertial weight dynamic adjustment strategy and a genetic crossover mutation operator, the particles of the target initial particle swarm are iteratively updated according to the coal source selection dimension and transportation mode dimension until the number of iterations reaches the first preset number, thus obtaining the first particle swarm; the four-dimensional objective function value corresponding to the particle position in the first particle swarm is less than the first threshold.

[0169] This step is an iterative process during the global exploration period. The first inertia weight dynamic adjustment strategy can be adopted using formula (15):

[0170] (15)

[0171] Calculate the inertia weight for each iteration;

[0172] in, For inertial weights, The initial value for the inertia weight is set to 1.0. The inertial weight termination value is set to 0.9. This represents the total number of iterations in this phase. This represents the current iteration number.

[0173] The iterative steps in this stage are as follows: Velocity update: Update the particle velocity according to the PSO velocity formula, while incorporating the direction guidance of the genetic crossover operator, as shown in formula (16):

[0174] (16)

[0175] in, , As a learning factor, , and The result is a random number, with a value range of [0,1]. The current iteration number particle velocity, To initialize the optimal position of each particle, that is, to initialize it to its own position. For the number of iterations The corresponding particle position at that time The velocity increment brought about by the crossover operator is taken as the position difference between the crossed particles, such as when particle p crosses with particle q. .

[0176] Dimensional variable iteration: The discrete dimension in this stage is Value replacement is achieved through genetic crossover and mutation operators. For example, if particle p has a coal source selection dimension of 3, it crosses with particle q, which has a coal source selection dimension of 5, to generate a new coal source selection dimension of 3 or 5. The better solution is retained according to the four-dimensional objective function value; continuous dimension Adjust linearly according to the speed formula, such as ,and ≤Maximum coal supply capacity; ,and .

[0177] During the iteration process, the four-dimensional objective function value of all particles is calculated in each iteration. If the particle's new position... The four-dimensional objective function value Smaller than itself If the four-dimensional objective function value is obtained, then update... If particles exist in the population of Value less than of If the value is not updated, then update. The position of the particle; remove Particles whose values ​​exceed the threshold, such as Particles are directly eliminated and replaced with new particles to ensure that the population direction meets the rigid requirements of environmental protection.

[0178] Step 42: Using fixed inertia weights and genetic crossover mutation operators, iteratively update each particle in the first particle group according to the procurement quantity allocation dimension until the number of iterations reaches the second preset number to obtain the second particle group; the comprehensive procurement cost value corresponding to each particle in the second particle group is less than the comprehensive procurement cost value in the first particle group, and the supply chain risk value corresponding to each particle in the second particle group is less than or equal to the supply chain risk value corresponding to each particle in the first particle group.

[0179] This step is an iterative process during the local optimization period: the inertia weight in this stage is fixed at 0.5-0.7, and the intermediate value of 0.6 can be taken. The large-scale search is abandoned, and the focus is on the local fine-tuning of high-quality particles.

[0180] The iterative steps in this stage are as follows: Speed ​​update: reduce the influence weight of the global optimum and strengthen the fine-tuning of the individual optimum, as shown in formula (17):

[0181] (17)

[0182] The above formula is based on formula (16), but with an increase of... Reduced to 2.2 Version 1.8 focuses on optimizing high-quality individual solutions.

[0183] Dimensional variable iteration: For coal source selection, the key indicator weights of the AI ​​fusion analysis model are combined, such as a calorific value of 40% and a sulfur content of 25%. Adjustments are only made between coal sources with high weights that meet the quality standards. For example, from… The calorific value is 4800 kcal / kg. The calorific value is adjusted to 4900 kcal / kg; the procurement quantity allocation is fine-tuned based on the principle of minimizing total cost, such as adjusting the particle size distribution. The capacity has been adjusted from 3,000 tons to 3,500 tons; adjustments have been made in conjunction with changes to the transportation method and delivery time to ensure... Towards To move closer.

[0184] During the iteration process, the focus is on the four-dimensional objective function value. Comprehensive procurement cost and supply chain risk coefficient , if the particle In the value The proportion decreased, and If there is no increase, it is considered a valid optimization and updated. ;like If the overall decrease is ≥3%, then update. This ensures that local optimization does not deviate from the global optimal direction.

[0185] Step 43: Using a linearly decreasing second inertial weight dynamic adjustment strategy and a genetic crossover mutation operator, iteratively update each particle of the second particle swarm according to the procurement quantity allocation dimension and the delivery time dimension until the number of iterations reaches the first preset number, to obtain the third particle swarm; determine the position vector of the particle with the smallest four-dimensional objective function value in the third particle swarm as the target procurement plan.

[0186] This step is the iterative process of the precise convergent. In this stage, the inertia weight is adjusted using a linear decreasing strategy, reducing the inertia weight from 0.5 to 0.3.

[0187] The iterative steps in this stage are as follows: Speed ​​update: Further reduce the speed amplitude, making only minor adjustments, as shown in formula (18):

[0188] (18)

[0189] Among them, the speed amplitude is controlled at To avoid large fluctuations.

[0190] Dimensional variable iteration: Only the purchase quantity allocation dimension and the delivery time dimension are fine-tuned by ±5%, such as... The amount was adjusted from 3,500 tons to 3,325 tons. The timeframe has been adjusted from 4 days to 3.8 days to ensure the feasibility of the plan.

[0191] During the iteration process, the four-dimensional objective function value is used. The final criterion is to select the population with the smallest number of cells and all dimensions satisfying the constraints. The 10 smallest particles; verifying the practical feasibility of these particles, such as supplier capacity and the adequacy of transportation resources; finally determining the global optimal solution. This refers to the target procurement plan. The corresponding particle position is converted into an executable procurement plan, which includes a recommended coal source list, the procurement quantity of each coal source, the transportation method and the delivery time, and is synchronized to the power plant procurement terminal.

[0192] This application utilizes the LightGBM algorithm with embedded attention fusion to predict results, combining the advantages of particle swarm optimization and genetic algorithms to construct a multi-objective optimization model encompassing cost, risk, environmental protection, and delivery time. The innovative algorithm handles high-dimensional, nonlinear, and multi-objective real-time decision-making. The two are seamlessly coupled through "attention weights," forming a practical, scalable, and self-evolving intelligent optimization method for coal-fired power generation. This overcomes the limitations of traditional single-objective optimization that focuses solely on cost, achieving comprehensive optimization of the procurement plan. This application can significantly reduce overall procurement costs; multi-objective optimization decision-making reduces the sum of procurement, transportation, and environmental costs by 10%-15%. It improves decision-making efficiency and dynamic response capabilities, shortening the procurement decision-making cycle from several days to several hours, and enabling real-time response to market changes such as coal price fluctuations and policy adjustments.

[0193] The embodiments of the present invention can divide functional modules according to the above method examples. For example, each function can be divided into its own functional module, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware or as a software functional module. It should be noted that the module division in the embodiments of the present invention is illustrative and only represents one logical functional division; other division methods may be used in actual implementation.

[0194] When dividing each function into modules according to its corresponding function. Figure 2 A schematic diagram of the structure of a power plant coal matching and procurement decision-making system provided by the present invention is shown. Figure 2 As shown, the device includes:

[0195] The multi-source data acquisition module 201 is used to acquire multi-source heterogeneous data, including power plant-side data, coal source-side data, supply chain-side data, and market-side data.

[0196] The big data preprocessing module 202 is used to process the multi-source heterogeneous data to generate a standardized feature dataset;

[0197] AI fusion analysis module 203 is used to input standardized feature datasets into an AI fusion analysis model that integrates gradient boosting trees and attention mechanisms, and output cost prediction results and key indicator weights; the cost prediction results include predicted values ​​for power generation coal consumption, environmental treatment costs, and transportation loss costs.

[0198] The multi-objective optimization decision module 204 is used to determine the optimal solution of the four-dimensional variables of the multi-objective function by using an improved hybrid particle swarm optimization algorithm and a dynamic adjustment strategy of inertia weight, based on the cost prediction results and the weights of key indicators, so as to obtain the target procurement plan; the four-dimensional variables include coal source, procurement quantity, transportation mode and delivery time.

[0199] The visualization output module 205 is used to construct a matching heat map using a visualization library, with coal source as the horizontal axis and unit as the vertical axis, marking the matching score and cost composition ratio of each coal source and unit; it displays the cost, risk, and suitability index comparison of different procurement schemes through radar charts; it generates an editable Excel format procurement decision report, including a recommended coal source list, optimal procurement quantity, transportation method, delivery time, and cost details, and pushes it to the power plant procurement management system terminal and the mobile terminal of the management personnel in real time via the WebSocket protocol.

[0200] Optionally, the AI ​​fusion analysis module 203 may specifically include:

[0201] The cost prediction submodule is used to input the standardized feature dataset into the basic prediction model, predict the cost prediction result, and output the feature activation matrix; the basic prediction model includes a feature attention enhancement layer, a LightGBM model, and a residual correction layer;

[0202] The impact factor quantification submodule is used to quantify the feature contribution value of each feature in the standardized feature dataset using SHAP values, optimize the feature contribution value according to the feature activation matrix to obtain the target feature weight of each feature, sort the target feature weights in descending order, and confirm the top N target feature weights as key indicator weights; N is a positive integer.

[0203] The model self-optimization submodule is used to monitor the update interval. When the update interval reaches a preset time, the parameters of the basic prediction model are adjusted according to the actual procurement data and the cost prediction results, and according to the preset time period.

[0204] Specifically, after the system has been running for 20 days, the model self-optimization submodule automatically extracts the actual procurement data from the past 20 days, such as coal source, procurement quantity, and actual cost, and compares it with the model's predicted data to calculate the mean squared error. If the mean squared error is greater than 3%, the hyperparameters of the LightGBM model, such as tree depth, learning rate, and number of leaf nodes, are adjusted using the Bayesian optimization algorithm to update the model parameter library.

[0205] In practical applications, the cost prediction submodule is responsible for core numerical prediction, the influence factor quantification submodule provides the basis for feature importance, and the model self-optimization submodule realizes dynamic error correction. The three form a closed-loop collaboration of "prediction, analysis and optimization", and finally output accurate comprehensive cost prediction value and feature weight ratio, thus fully realizing the goal of AI fusion analysis.

[0206] Optionally, the standardized feature dataset includes coal quality feature groups, price feature groups, supply chain feature groups, and unit operation feature groups; the cost prediction submodule can specifically be used for:

[0207] The standardized feature dataset is input into the basic prediction model. The feature attention enhancement layer uses a multi-head attention mechanism to weight and enhance different feature groups, constructing a feature group weight allocation network. The LightGBM model uses a gradient boosting tree to predict costs based on the weights allocated by the feature group weight allocation network, and obtains the output result.

[0208] An error correction layer is used to correct the output results to obtain the cost prediction results.

[0209] Optionally, the multi-objective optimization decision module 204 may specifically include:

[0210] The four-dimensional objective function construction submodule is used to set multi-dimensional objectives and constraints to construct a four-dimensional objective function for coal procurement. The multi-dimensional objectives include at least the following: the overall procurement cost meets a preset cost threshold, the supply chain risk meets a preset risk threshold, the environmental compliance rate meets a preset environmental threshold, and the delivery time is within a preset delivery period. The constraints include coal quality constraints, supplier constraints, inventory constraints, and policy constraints.

[0211] The particle swarm initialization submodule is used to initialize particles according to the feature weight guidance strategy to obtain the target initial particle swarm.

[0212] The iterative calculation submodule is used to dynamically adjust the position and velocity of each particle in the target initial particle swarm based on the inertial weight dynamic adjustment strategy and the genetic crossover mutation operator, starting from the target initial particle swarm, until the number of iterations reaches the preset number, and determine the optimal solution of the four-dimensional objective function corresponding to the particle as the target procurement plan.

[0213] Optionally, the iterative calculation submodule can be specifically used for:

[0214] A linearly decreasing first inertia weight dynamic adjustment strategy and a genetic crossover mutation operator are adopted to iteratively update each particle of the target initial particle swarm according to the coal source selection dimension and transportation mode dimension until the number of iterations reaches a first preset number, thus obtaining a first particle swarm; the four-dimensional objective function value corresponding to the particle position in the first particle swarm is less than a first threshold; the four-dimensional objective function value is calculated based on the cost prediction result and the key indicator weights;

[0215] Using fixed inertia weights and genetic crossover mutation operators, the particles in the first particle group are iteratively updated according to the procurement quantity allocation dimension until the number of iterations reaches a second preset number, thus obtaining a second particle group; the comprehensive procurement cost value corresponding to each particle in the second particle group is less than the comprehensive procurement cost value in the first particle group, and the supply chain risk value corresponding to each particle in the second particle group is less than or equal to the supply chain risk value corresponding to each particle in the first particle group.

[0216] A dynamic adjustment strategy of linearly decreasing second inertial weights and a genetic crossover mutation operator are adopted to iteratively update each particle of the second particle swarm according to the procurement quantity allocation dimension and the delivery time dimension until the number of iterations reaches the first preset number, thus obtaining a third particle swarm; the position vector of the particle with the smallest four-dimensional objective function value in the third particle swarm is determined as the target procurement scheme.

[0217] Optionally, the particle swarm initialization submodule can be specifically used for:

[0218] Generate a preset number of first initial particle groups; the first initial particle groups include at least a preset proportion of particles whose coal source selection dimension meets the coal quality standards.

[0219] The procurement quantity allocation dimension of each particle in the first initial particle swarm is adjusted to a preset procurement quantity to obtain a second initial particle swarm; the delivery time dimension of each particle in the second initial particle swarm is adjusted so that the delivery deviation corresponding to the adjusted particle is within a preset deviation threshold to obtain a target initial particle swarm; the delivery deviation is the difference between the current delivery time and the optimal delivery time.

[0220] Optionally, the impact factor quantification submodule can be specifically used for:

[0221] Formula used:

[0222] ;

[0223] Calculate the feature contribution value of each feature in the standardized feature dataset;

[0224] in, This represents the complete value after performing SHAP decomposition on the i-th sample. As the global baseline value, is the feature contribution value of the j-th feature to the i-th sample; m is the total number of core features involved in the comprehensive cost prediction, that is, the number of feature dimensions in the standardized feature dataset, corresponding to the dimension of the standardized feature dataset.

[0225] Formula used:

[0226] ;

[0227] The feature contribution values ​​are optimized to obtain the target feature weights for each feature;

[0228] in, The target feature weight for the j-th feature is... Let j be the average activation intensity of the j-th feature. Let be the average activation intensity of the k-th feature, where k is the index variable of the feature. is the feature activation matrix.

[0229] Optionally, the big data preprocessing module 202 may specifically include:

[0230] The outlier identification submodule is used to filter out suspected outliers from the multi-source heterogeneous dataset using box plots; to identify the suspected outliers using the isolated forest algorithm, to determine the target outlier, and to delete the target outlier from the multi-source heterogeneous dataset to obtain the first dataset;

[0231] The missing value imputation submodule is used to divide the first dataset into clusters using a clustering algorithm, and to impute missing values ​​within each cluster using a weighted k-nearest neighbor algorithm to obtain the second dataset.

[0232] The feature engineering submodule is used to divide the data in the second dataset according to feature attributes to obtain multiple sets of feature data; the Pearson correlation coefficient is used to filter out target data that is strongly correlated with the overall cost from the multiple sets of feature data, and the target data is standardized to obtain a standardized feature dataset.

[0233] The above mainly describes the solutions provided by the embodiments of the present invention from the perspective of the interaction between various modules. It is understood that, in order to achieve the above functions, it includes corresponding hardware structures and / or software modules for executing each function. Those skilled in the art should readily recognize that, in conjunction with the units and algorithm steps of the various examples described in the embodiments disclosed herein, the present invention can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0234] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer programs or instructions. When the computer program or instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present invention are performed entirely or partially. The computer can be a general-purpose computer, a special-purpose computer, a computer network, a terminal, a user equipment, or other programmable device. The computer program or instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another. For example, the computer program or instructions can be transferred from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium, such as a floppy disk, hard disk, or magnetic tape; it can also be an optical medium, such as a digital video disc (DVD); or it can be a semiconductor medium, such as a solid-state drive (SSD).

[0235] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, the disclosure, and the appended claims in carrying out the claimed invention. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0236] Although the invention has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made therein without departing from the spirit and scope of the invention. Accordingly, this specification and drawings are merely exemplary descriptions of the invention as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if such modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include such modifications and modifications.

Claims

1. A method for coal plant matching and procurement decision making, characterized in that, include: Acquire multi-source heterogeneous data; the multi-source heterogeneous data includes power plant data, coal source data, supply chain data, and market data; The multi-source heterogeneous data is processed to generate a standardized feature dataset; A standardized feature dataset is input into an AI fusion analysis model that integrates gradient boosting trees and attention mechanisms, and the model outputs cost prediction results and key indicator weights. The cost prediction results include predicted values ​​for power generation coal consumption, environmental treatment costs, and transportation loss costs. Based on the cost prediction results and the weights of the key indicators, an improved hybrid particle swarm optimization algorithm is used to iteratively determine the optimal solution of the four-dimensional variables of the multi-objective function through a dynamic adjustment strategy of inertia weight, thereby obtaining the target procurement plan; the four-dimensional variables include coal source, procurement quantity, transportation method and delivery time. The process of inputting the standardized feature dataset into the AI ​​fusion analysis model that integrates gradient boosting trees and attention mechanisms, and outputting cost prediction results and key indicator weights, includes: A standardized feature dataset is input into a basic prediction model to predict the cost and output a feature activation matrix. The basic prediction model includes a feature attention enhancement layer, a LightGBM model, and a residual correction layer. The SHAP value is used to quantify the feature contribution value of each feature in the standardized feature dataset, and the feature contribution value is optimized according to the feature activation matrix to obtain the target feature weight of each feature. The target feature weights are sorted in descending order, and the top N target feature weights are identified as key indicator weights; N is a positive integer. The update interval is monitored. When the update interval reaches a preset time, the parameters of the basic prediction model are adjusted according to the actual procurement data and the cost prediction results, and the cost prediction is performed based on the adjusted basic prediction model. The standardized feature dataset includes coal quality feature groups, price feature groups, supply chain feature groups, and unit operation feature groups; the step of inputting the standardized feature dataset into the basic prediction model to predict cost forecast results includes: The standardized feature dataset is input into the basic prediction model. The feature attention enhancement layer uses a multi-head attention mechanism to weight and strengthen different feature groups, constructing a feature group weight allocation network. The feature group weight allocation network is shown in the formula: （2） wherein, , ; is the attention output, is the attention result corresponding to the coal quality feature group, is the attention result corresponding to the price feature group, is the attention result corresponding to the supply chain feature group, is the attention result corresponding to the unit operation feature group, , , and are coefficients; the LightGBM model adopts gradient boosting tree for cost prediction according to the weights allocated by the feature group weight allocation network, to obtain an output result. A residual correction layer is used to correct the error in the output result to obtain the cost prediction result. Based on the cost prediction results and the weights of the key indicators, an improved hybrid particle swarm optimization algorithm is used to iteratively determine the optimal solution of the four-dimensional variables of the multi-objective function through a dynamic adjustment strategy of inertia weights, resulting in the target procurement plan, including: A four-dimensional objective function for coal procurement is constructed by setting multi-dimensional objectives and constraints. The multi-dimensional objectives include at least the following: the overall procurement cost meets a preset cost threshold, the supply chain risk meets a preset risk threshold, the environmental compliance rate meets a preset environmental threshold, and the delivery time is within a preset delivery period. The constraints include coal quality constraints, supplier constraints, inventory constraints, and policy constraints. Particle initialization is performed according to the feature weighting guidance strategy to obtain the target initial particle swarm; Based on the dynamic adjustment strategy of inertial weight and the genetic crossover mutation operator, starting from the target initial particle swarm, the position and velocity of each particle in the target initial particle swarm are iteratively updated until the number of iterations reaches the preset number, and the optimal solution of the four-dimensional objective function corresponding to the particle is determined as the target procurement plan. The four-dimensional objective function is shown in the formula: ； in, The four-dimensional objective function value. The total procurement cost is the predicted total cost, adjusted according to the weights of key indicators. This is the supply chain risk coefficient, ranging from 0 to 1, and is calculated as follows: , To mitigate the risk of supplier default, To mitigate the risk of transportation disruptions, This is the delivery time deviation coefficient, with a value ranging from 0 to 1. , The value is the coal inventory warning line period minus 3 days. For the transportation cycle; To ensure environmental compliance, only when the coal quality indicators meet the following criteria are considered: calorific value ≥ 4500 kcal / kg, sulfur content ≤ 0.6%, and ash content ≤ 20%. It is 1 if it is true, otherwise it is 0. As a weighting of the overall procurement cost, As the weight of the supply chain risk coefficient, As a weighting factor for the environmental compliance rate, The weight of the delivery time deviation coefficient; , , as well as The sum of is 1.

2. The method for matching and purchasing coal in power plants according to claim 1, characterized in that, The inertial weight dynamic adjustment strategy and genetic crossover mutation operator, starting from the target initial particle swarm, iteratively update the position and velocity of each particle in the target initial particle swarm until the number of iterations reaches a preset number, and determine the optimal solution of the four-dimensional objective function corresponding to the particle as the target procurement plan, includes: A linearly decreasing first inertia weight dynamic adjustment strategy and a genetic crossover mutation operator are adopted to iteratively update each particle of the target initial particle swarm according to the coal source selection dimension and transportation mode dimension until the number of iterations reaches a first preset number, thus obtaining a first particle swarm; the four-dimensional objective function value corresponding to the particle position in the first particle swarm is less than a first threshold; the four-dimensional objective function value is calculated based on the cost prediction result and the key indicator weights; Using fixed inertia weights and genetic crossover mutation operators, the particles in the first particle group are iteratively updated according to the procurement quantity allocation dimension until the number of iterations reaches a second preset number, thus obtaining a second particle group; the comprehensive procurement cost value corresponding to each particle in the second particle group is less than the comprehensive procurement cost value in the first particle group, and the supply chain risk value corresponding to each particle in the second particle group is less than or equal to the supply chain risk value corresponding to each particle in the first particle group. A dynamic adjustment strategy of linearly decreasing second inertial weights and a genetic crossover mutation operator are adopted to iteratively update each particle of the second particle swarm according to the procurement quantity allocation dimension and the delivery time dimension until the number of iterations reaches the first preset number, thus obtaining a third particle swarm; the position vector of the particle with the smallest four-dimensional objective function value in the third particle swarm is determined as the target procurement scheme.

3. The method of claim 1, wherein: The process of initializing particles according to the feature weight guidance strategy to obtain the target initial particle swarm includes: Generate a preset number of first initial particle groups; the first initial particle groups include at least a preset proportion of particles whose coal source selection dimension meets the coal quality standards. The procurement quantity allocation dimension of each particle in the first initial particle swarm is adjusted to the preset procurement quantity to obtain the second initial particle swarm. The arrival time dimension of each particle in the second initial particle swarm is adjusted so that the arrival deviation of the adjusted particles is within a preset deviation threshold, thus obtaining the target initial particle swarm; the arrival deviation is the difference between the current arrival time and the optimal arrival time.

4. The method of claim 1, wherein: The step of quantifying the feature contribution value of each feature in the standardized feature dataset using SHAP values ​​and optimizing the feature contribution value based on the feature activation matrix to obtain the target feature weights for each feature includes: Formula used: ； Calculate the feature contribution value of each feature in the standardized feature dataset; wherein, is the complete value of the SHAP decomposition for the i-th sample, is the global reference value, is the feature contribution value of the j-th feature to the i-th sample; m is the total number of core features participating in the comprehensive cost prediction, that is, the number of feature dimensions in the standardized feature data set, corresponding to the dimension of the standardized feature data set. Formula used: ； The feature contribution values ​​are optimized to obtain the target feature weights for each feature; wherein, is a target feature weight for the jth feature, is an average activation strength for the jth feature, is an average activation strength for the kth feature, k being an index variable for the features, is a feature activation matrix.

5. The method for matching and purchasing coal in power plants according to claim 1, characterized in that, The process of processing the multi-source heterogeneous data to generate a standardized feature dataset includes: Box plots are used to filter out suspected outliers from the multi-source heterogeneous dataset; the isolated forest algorithm is used to identify the suspected outliers, determine the target outliers, and delete the target outliers from the multi-source heterogeneous dataset to obtain the first dataset; The first dataset is divided into clusters using a clustering algorithm, and missing values ​​are imputed within each cluster using a weighted k-nearest neighbor algorithm to obtain the second dataset. The data in the second dataset are divided according to their feature attributes to obtain multiple sets of feature data. The Pearson correlation coefficient is used to select target data that is strongly correlated with the overall cost from the multiple sets of feature data, and the target data is standardized to obtain a standardized feature dataset.

6. A power plant coal matching and procurement decision-making system, characterized in that, include: A multi-source data acquisition module is used to acquire multi-source heterogeneous data, including power plant-side data, coal source-side data, supply chain-side data, and market-side data. The big data preprocessing module is used to process the multi-source heterogeneous data and generate a standardized feature dataset; The AI ​​fusion analysis module is used to input standardized feature datasets into an AI fusion analysis model that integrates gradient boosting trees and attention mechanisms, and outputs cost prediction results and key indicator weights; the cost prediction results include predicted values ​​for power generation coal consumption, environmental treatment costs, and transportation loss costs. The multi-objective optimization decision module is used to determine the optimal solution of the four-dimensional variables of the multi-objective function based on the cost prediction results and the weights of key indicators, using an improved hybrid particle swarm optimization algorithm and a dynamic adjustment strategy for inertia weights, thereby obtaining the target procurement plan. The four-dimensional variables include coal source, purchase quantity, transportation method, and delivery time; The AI ​​fusion analysis module includes: The cost prediction submodule is used to input the standardized feature dataset into the basic prediction model, predict the cost prediction result, and output the feature activation matrix; the basic prediction model includes a feature attention enhancement layer, a LightGBM model, and a residual correction layer; The impact factor quantification submodule is used to quantify the feature contribution value of each feature in the standardized feature dataset using SHAP values, optimize the feature contribution value according to the feature activation matrix to obtain the target feature weight of each feature, sort the target feature weights in descending order, and confirm the top N target feature weights as key indicator weights; N is a positive integer. The model liberalization submodule is used to monitor the update interval. When the update interval reaches a preset time, the parameters of the basic prediction model are adjusted according to the actual procurement data and the cost prediction results, and cost prediction is performed based on the adjusted basic prediction model. The standardized feature dataset includes coal quality feature groups, price feature groups, supply chain feature groups, and unit operation feature groups; the cost prediction submodule is specifically used for: The standardized feature dataset is input into the basic prediction model. The feature attention enhancement layer uses a multi-head attention mechanism to weight and strengthen different feature groups, constructing a feature group weight allocation network. The feature group weight allocation network is shown in the formula: （2） in, , ; For attention output, The attention results are for the coal quality characteristic groups. The attention results are for the price feature group. The attention results are for the supply chain feature groups. The attention results are for the unit operation feature groups. , , as well as All are coefficients; the LightGBM model assigns weights to the network based on the feature group weight allocation network, and uses a gradient boosting tree to predict costs and obtain the output results; A residual correction layer is used to correct the error in the output result to obtain the cost prediction result. The multi-objective optimization decision module includes: The four-dimensional objective function construction submodule is used to set multi-dimensional objectives and constraints to construct a four-dimensional objective function for coal procurement. The multi-dimensional objectives include at least the following: the overall procurement cost meets a preset cost threshold, the supply chain risk meets a preset risk threshold, the environmental compliance rate meets a preset environmental threshold, and the delivery time is within a preset delivery period. The constraints include coal quality constraints, supplier constraints, inventory constraints, and policy constraints. The particle swarm initialization submodule is used to initialize particles according to the feature weight guidance strategy to obtain the target initial particle swarm. The iterative calculation submodule is used to dynamically adjust the position and velocity of each particle in the target initial particle swarm based on the inertial weight dynamic adjustment strategy and the genetic crossover mutation operator, starting from the target initial particle swarm, until the number of iterations reaches the preset number, and determine the optimal solution of the four-dimensional objective function corresponding to the particle as the target procurement plan. The four-dimensional objective function is shown in the formula: ； in, The four-dimensional objective function value. The total procurement cost is the predicted total cost, adjusted according to the weights of key indicators. This is the supply chain risk coefficient, ranging from 0 to 1, and is calculated as follows: , To mitigate the risk of supplier default, To mitigate the risk of transportation disruptions, This is the delivery time deviation coefficient, with a value ranging from 0 to 1. , The value is the coal inventory warning line period minus 3 days. For the transportation cycle; To ensure environmental compliance, only when the coal quality indicators meet the following criteria are considered: calorific value ≥ 4500 kcal / kg, sulfur content ≤ 0.6%, and ash content ≤ 20%. It is 1 if it is true, otherwise it is 0. As a weighting of the overall procurement cost, As the weight of the supply chain risk coefficient, As a weighting factor for the environmental compliance rate, The weight of the delivery time deviation coefficient; , , as well as The sum of is 1.

Images