Thermal power plant fuel supply intelligent optimization method and system based on multi-source information fusion

By using a multi-source information fusion method and leveraging a multi-layer LSTM network and a Bayesian risk network, risks are dynamically quantified and Pareto optimal scheduling instructions are generated. This solves the problems of lagging fuel calorific value measurement and insufficient risk early warning in traditional thermal power plant fuel supply, and realizes intelligent optimization and safety improvement of fuel supply.

CN121724191APending Publication Date: 2026-03-24HUANENG ZUOQUAN COAL&POWER CO LTD
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202511697713.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-11-19
Publication Date
2026-03-24

AI Technical Summary

Technical Problem

Traditional coal-fired power plant fuel supply methods lack the ability to perform real-time soft measurement of fuel calorific value, making it impossible to accurately quantify cost and environmental deviations. This results in lagging coal blending strategies, crude inventory control, and insufficient early warning of supply disruption risks.

Method used

By employing a multi-source information fusion method, a soft measurement model of the calorific value of coal fed into the furnace is constructed through a multi-layer LSTM network with physical constraints correction. Combined with a Bayesian risk network and a deep deterministic policy gradient network, the risk is dynamically quantified and Pareto optimal fuel scheduling instructions are generated, thereby achieving intelligent optimization of fuel supply.

Benefits of technology

It significantly improves the accuracy of coal calorific value prediction, dynamically responds to the combustion process, quantifies various risks, and generates Pareto optimal scheduling instructions. It solves the problems of reliance on human experience and response lag in traditional methods, and improves the safety and economy of thermal power plants.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121724191A_ABST
    Figure CN121724191A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of intelligent fuel supply, in particular to an intelligent optimization method and system for fuel supply of a thermal power plant based on multi-source information fusion. The method comprises the following steps: collecting multi-source heterogeneous data and carrying out space-time alignment and normalization processing; constructing a coal as fired calorific value soft measurement model by adopting a physical constraint corrected LSTM network; generating a transportation delay predicted value and probability distribution thereof through an XGBoost regression model; calculating a comprehensive risk score in combination with a Bayesian risk network; solving a Pareto optimal solution set under the constraint of an inventory safety threshold by adopting a depth deterministic strategy gradient network, and generating a fuel scheduling instruction; and updating the as-fired coal calorific value soft measurement model on line. Through combination of the physical constraint corrected LSTM network and multi-source time sequence data fusion and regularization training, the fire coal calorific value prediction precision is remarkably improved, and the problems that a traditional method depends on off-line assay, response is lagged, and the real-time dynamic state of the combustion process is difficult to reflect are solved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent fuel supply technology, and more specifically, to an intelligent optimization method and system for fuel supply in thermal power plants based on multi-source information fusion. Background Technology

[0002] With the deepening of energy transition and power market reform, thermal power plants face multiple pressures related to environmental protection, economics, and dispatch flexibility. Fuel supply, as a core component of thermal power plant operation, directly impacts the safe operation and overall efficiency of the units due to its stability, economics, and environmental friendliness. Currently, power plant fuel management relies heavily on manual experience, making it difficult to dynamically integrate heterogeneous data from multiple sources, such as boiler combustion, fuel inventory, transportation and meteorological data, and grid dispatch. This results in lagging coal blending strategies, inadequate inventory control, and insufficient early warning of supply disruption risks.

[0003] Traditional methods lack the ability to perform real-time soft measurement of fuel calorific value, making it impossible to accurately quantify cost and environmental deviations. Therefore, this paper proposes an intelligent optimization method and system for fuel supply in thermal power plants based on multi-source information fusion. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for intelligent optimization of fuel supply in thermal power plants based on multi-source information fusion, so as to solve the problem that the traditional methods mentioned in the background art lack the ability to perform real-time soft measurement of fuel calorific value and cannot accurately quantify cost and environmental deviations.

[0005] To achieve the above objectives, the present invention aims to provide a smart optimization method for fuel supply in thermal power plants based on multi-source information fusion, comprising the following steps: S1. Collect boiler combustion data, fuel inventory data, transportation and meteorological data, and power grid dispatch command data; perform spatiotemporal alignment and normalization processing on the above-collected multi-source heterogeneous data to construct a unified feature vector; S2. Based on a unified feature vector, a soft measurement model for the calorific value of coal entering the furnace is constructed using a multi-layer LSTM network with physical constraint correction, to obtain the predicted calorific value of coal. And calculate cost deviation and environmental deviation based on the predicted calorific value of coal; S3. Process transportation and meteorological data using an XGBoost regression model to generate predicted transportation delays and their probability distributions; use predicted transportation delays, cost deviations, and environmental deviations as inputs, and combine this with a Bayesian risk network to quantify the probability of supply disruptions. Probability of exceeding environmental protection standards and the probability of cost overrun The risk weights are dynamically adjusted according to the power grid dispatch instructions, and the comprehensive risk score is calculated. ; S4. An optimization model is constructed with the goal of minimizing the comprehensive risk score. A deep deterministic strategy gradient network is used to solve the Pareto optimal solution set under the inventory safety threshold constraint. Fuel scheduling instructions are generated based on the Pareto optimal solution set. S5. Execute the fuel dispatching command and dynamically adjust the coal blending strategy, and update the soft measurement model of the calorific value of the coal fed into the furnace online.

[0006] As a further improvement to this technical solution, in step S1, the specific steps involved in performing spatiotemporal alignment and normalization processing on the collected multi-source heterogeneous data to construct a unified feature vector are as follows: The collected multi-source heterogeneous data were interpolated and aligned using a unified timestamp, and the spatial dimension was unified to the unit level. The Z-score method is used to standardize numerical features and one-hot encoding is used for categorical features; The aligned and standardized multi-source heterogeneous data are concatenated column-wise into a unified feature vector. .

[0007] As a further improvement to this technical solution, in step S2, the specific steps involved in constructing a soft measurement model of the calorific value of the coal fed into the furnace using a multi-layer LSTM network with physical constraint correction to obtain the predicted calorific value of the coal are as follows: The unified feature vector under historical operating conditions As input, a multi-layer LSTM network is constructed, and the preliminary prediction result of the calorific value of the coal entering the furnace is output. ; Based on the law of conservation of mass, a mass balance constraint is constructed. The mass balance constraint is used to limit the difference between the total input material flow rate and the total output material flow rate of the combustion system to approach 0 within a specific time step. Based on the law of conservation of energy, an energy balance constraint is constructed. This constraint is used to limit the difference between the total energy of the input materials and the total energy of the output materials of the combustion system, as well as the system heat loss, to approach zero within a specific time period. Physical constraints are constructed based on mass balance constraints and energy balance constraints, and mass deviation and net energy deviation are obtained. The mean square error loss term is calculated based on the preliminary prediction results and the measured calorific value of coal, and the physical constraint loss term is constructed based on the quality deviation and net energy deviation. The physical constraint loss term is incorporated as a regularization term into the mean squared error loss term of the multilayer LSTM network to form the total loss function; A soft measurement model for the calorific value of coal entering the furnace is constructed by iteratively optimizing the weight matrix and bias vector of the multilayer LSTM network using the total loss function. The unified feature vector at the time to be predicted is then input into the trained soft measurement model to obtain the predicted calorific value of the coal. .

[0008] As a further improvement to this technical solution, the specific steps involved in calculating cost deviation and environmental deviation based on the predicted calorific value of coal in step S2 are as follows: Based on power grid dispatch command data, the target load of generating units contained therein is analyzed. Combined with the preset boiler efficiency and standard coal consumption rate for power generation The target calorific value was calculated. ; Based on measured coal feeder speed data from boiler combustion data The current coal feed rate is obtained by converting the coal feeder speed versus coal feed rate calibration curve function. ; Based on predicted calorific value of coal Target calorific value Current coal supply Combined with the preset coal unit price Calculate cost deviation ; Based on predicted calorific value of coal The measured sulfur content in the boiler combustion data Combined with preset emission limits Environmental penalty coefficient Calorific value and environmental impact coefficient and the optimal environmental calorific value threshold Calculate environmental deviation .

[0009] As a further improvement to this technical solution, in step S3, the specific steps involved in processing transportation and meteorological data using the XGBoost regression model to generate predicted transportation delay values ​​and their probability distributions are as follows: All features from the preprocessed transportation and meteorological data are concatenated into a transportation and meteorological feature vector. ; Using transportation and meteorological feature vectors Using the input as input, an XGBoost regression model is constructed by integrating multiple CART regression decision trees to obtain preliminary predictions of transportation delays; The loss function term uses the squared loss function to quantify the difference between the predicted and actual transport delay values, and the regularization term uses the cumulative leaf nodes to quantify the complexity of the tree structure. The XGBoost regression model is trained by minimizing the objective function containing the loss function term and the regularization term. Using the same transportation and meteorological feature vectors The quantile regression method was used to train the system with specific quantiles. Multiple XGBoost regression models targeting a given goal; each XGBoost regression model independently outputs its corresponding quantile. Predicted value These predicted values ​​collectively constitute the probability distribution of the predicted transportation delay values. Based on the obtained set of quantile predictions A linear interpolation method is used to fit the probability distribution of the complete transport delay prediction value, and the transport delay prediction value is obtained from the probability distribution. This allows for the evaluation of arbitrary delay time thresholds. Risk probability .

[0010] As a further improvement to this technical solution, the specific steps involved in S3, which combine Bayesian risk networks to quantify the probability of supply disruption risk, the probability of exceeding environmental standards risk, and the probability of cost overrun risk, are as follows: Based on the risk impact mechanism of fuel supply to thermal power plants, risk nodes are set as supply interruption risk nodes, environmental protection exceeding risk nodes, and cost overrun risk nodes. The parent node of the nodes at risk of supply disruption is the predicted value of transportation delay. The parent node of the environmental protection exceedance risk node is the environmental protection deviation node. And the parent node of the cost overrun risk node is the cost deviation node. Construct a Bayesian risk network structure; Based on historical data statistics, a prior probability distribution is set for each parent node, and a conditional probability table is defined to describe the dependency relationship between child nodes and parent nodes. Introducing a preset supply interruption time threshold, based on an arbitrary delay time threshold. Risk probability Calculate the probability that the predicted transportation delay exceeds the supply disruption time threshold, thus obtaining the preliminary probability of supply disruption risk. ; By introducing a preset environmental allowable deviation threshold, the probability of environmental deviation exceeding the environmental allowable deviation range is calculated based on the prior probability distribution of environmental deviation, thus obtaining the preliminary probability of environmental exceedance risk. ; By introducing a preset cost tolerance threshold, the probability of cost deviation exceeding the cost tolerance threshold is calculated based on the prior probability distribution of cost deviation, thus obtaining the preliminary probability of cost overrun risk. ; Predicted transportation delays Environmental deviations and cost deviation The evidence is input into the Bayesian risk network, and the probability distribution of each node in the Bayesian risk network is updated through the Junction Tree algorithm, ultimately outputting the probability of supply disruption. Probability of exceeding environmental protection standards and the probability of cost overrun .

[0011] As a further improvement to this technical solution, in step S3, the specific steps involved in dynamically adjusting the risk weights according to the power grid dispatch instructions and calculating the comprehensive risk score are as follows: Based on the unit's target load By combining the preset load-risk weight mapping table, the risk weight of power outage is dynamically determined. Environmental protection exceeding risk weight and cost overrun risk weight ; Based on the probability of supply disruption Probability of exceeding environmental protection standards and the probability of cost overrun Combined with dynamically adjusted supply disruption risk weights Environmental protection exceeding risk weight and cost overrun risk weight The comprehensive risk score is calculated using a weighted summation method. ; Comprehensive risk score Perform max-min normalization.

[0012] As a further improvement to this technical solution, in step S4, a deep deterministic strategy gradient network is used to solve for the Pareto optimal solution set under the inventory safety threshold constraint. The specific steps involved in generating fuel scheduling instructions based on this solution set are as follows: To minimize the overall risk score To establish an optimization model with the core objective as the objective. The optimized model must meet the inventory safety constraint, that is, the inventory level does not exceed the upper and lower bounds of the inventory safety threshold. Based on the unified feature vector at the current time and comprehensive risk score To obtain the environmental state ; Set Action It is a continuous action vector, and the actions include coal blending bin switching commands, fuel blending ratio vectors, and coal feed control commands for each coal feeder; Based on the objectives of minimizing overall risk and satisfying inventory safety constraints, a reward function is obtained through a weighted penalty mechanism; An initial Actor network is constructed using a multi-layer fully connected neural network. The input of the initial Actor network is the environment state, and the output is the action. Similarly, a multi-layer fully connected neural network is used to construct the initial Critic network. The input of the initial Critic network is the environment state-action pair, and the output is the value assessment. Simultaneously, a target Actor network with the same structure as the initial Actor network but different parameters, and a target Critic network with the same structure as the initial Critic network but different parameters are established. A deep deterministic policy gradient network is constructed based on the initial Actor network, the initial Critic network, the target Actor network, and the target Critic network. The transition sample, consisting of the current environment state, action, reward function, and the next environment state, is stored in the experience replay buffer. The initial Critic network parameters are updated by minimizing the mean square error between the target value assessment and the current value assessment. The initial Critic network parameters This includes the weight matrix and bias terms of the initial Critic network; To maximize the evaluation value of the initial Critic network, the parameters of the initial Actor network are updated using the policy gradient method. The initial Actor network parameters This includes the weight matrix and bias terms of the initial Actor network; The parameters of the target Actor network and the target Critic network are updated using a soft update method, and the training is repeated until the constructed deep deterministic policy gradient network converges. Under the constraint of inventory safety, the trained deep deterministic policy gradient network is run to scan the action space and obtain a series of non-dominated solutions, which constitute the Pareto optimal solution set. Preferences are determined from the Pareto optimal solution set based on the load-risk weight mapping table. Based on the fuel inventory and unit coal consumption rate in the screening scheme, the coal blending bin switching sequence is obtained by setting the operating bin / standby bin and inventory switching threshold; Based on the calorific value and sulfur content in the screening scheme, combined with the basic properties of the fuel, the optimal blending ratio of the fuel is calculated. Based on the coal consumption demand corresponding to the blending ratio and the target load of the unit, the specific coal feeder speed command is converted into a calibration curve function of coal feeder speed-coal feed rate. Based on the current fuel inventory and the predicted calorific value of coal, and the predicted transportation delay, a coal replenishment dispatch order is calculated. The coal replenishment dispatch order includes the type of coal to be replenished, the amount of coal to be replenished, and the departure time. The decision module outputs the coal blending bin switching sequence, optimal fuel blending ratio, coal feeder speed command, and coal replenishment scheduling command as fuel scheduling commands.

[0013] As a further improvement to this technical solution, the specific steps in S5 for updating the soft measurement model of the calorific value of the coal fed into the furnace online and generating an emergency dispatch plan are as follows: Based on the current measured calorific value of coal Compared with the current predicted calorific value of coal Calculate the prediction error at the current time. If the prediction error Exceeding the preset error threshold This triggers the online model update mechanism; Use a sliding window mechanism to retain the most recent The unified feature vectors and measured heat values ​​at each time step constitute the incremental training dataset. ; Based on the incremental training dataset, the parameters of the multilayer LSTM network are fine-tuned by minimizing the total loss function, thereby realizing the online update of the soft measurement model of the calorific value of coal entering the furnace.

[0014] On the other hand, the present invention provides an intelligent optimization system for fuel supply in thermal power plants based on multi-source information fusion, used to implement the intelligent optimization method for fuel supply in thermal power plants based on multi-source information fusion described in any one of the above-mentioned methods, characterized in that it includes: The data acquisition module collects boiler combustion data, fuel inventory data, transportation and meteorological data, and power grid dispatch command data; it performs spatiotemporal alignment and normalization processing on the above-mentioned multi-source heterogeneous data to construct a unified feature vector; The calorific value soft measurement module, based on a unified feature vector, uses a multi-layer LSTM network with physical constraint correction to construct a soft measurement model of the calorific value of the coal entering the furnace, thereby obtaining the predicted calorific value of the coal. And calculate cost deviation and environmental deviation based on the predicted calorific value of coal; The comprehensive risk assessment module processes transportation and meteorological data using an XGBoost regression model to generate predicted transportation delays and their probability distributions. Using predicted transportation delays, cost deviations, and environmental deviations as inputs, it combines a Bayesian risk network to quantify the probability of supply disruptions. Probability of exceeding environmental protection standards and the probability of cost overrun The risk weights are dynamically adjusted according to the power grid dispatch instructions, and the comprehensive risk score is calculated. ; The decision-making module constructs an optimization model with the goal of minimizing the comprehensive risk score, and uses a deep deterministic strategy gradient network to solve for the Pareto optimal solution set under the inventory safety threshold constraint. Fuel scheduling instructions are generated based on the Pareto optimal solution set. The update module executes the fuel scheduling instructions and dynamically adjusts the coal blending strategy, updating the soft measurement model of the calorific value of the coal fed into the furnace online.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The intelligent optimization method and system for fuel supply in thermal power plants based on multi-source information fusion adopts a multi-layer LSTM network with physical constraint correction to construct a soft measurement model of the calorific value of coal entering the furnace. It combines the law of conservation of mass and energy to construct a physical constraint loss term. Through multi-source time series data fusion and regularized training, it significantly improves the accuracy of coal calorific value prediction and solves the problems of traditional methods relying on offline testing, slow response, and difficulty in reflecting the real-time dynamics of the combustion process.

[0016] 2. The intelligent optimization method and system for fuel supply in thermal power plants based on multi-source information fusion is based on Bayesian risk network and multi-objective deep reinforcement learning. It integrates transportation delay prediction, cost deviation and environmental deviation, dynamically quantifies the risks of supply interruption, environmental exceedance and cost overrun, and generates Pareto optimal scheduling instructions under inventory safety constraints. It solves the problems of traditional methods relying on human experience and lacking multi-risk collaborative optimization and dynamic response capabilities. Attached Figure Description

[0017] Figure 1 This is a flowchart of the overall method of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention. Example 1:

[0019] Please see Figure 1 As shown, this embodiment provides a smart optimization method for fuel supply in thermal power plants based on multi-source information fusion, including the following steps: S1. Collect boiler combustion data, fuel inventory data, transportation and meteorological data, and power grid dispatch command data; perform spatiotemporal alignment and normalization processing on the above-collected multi-source heterogeneous data to construct a unified feature vector; The boiler combustion data includes measured coal calorific value, sulfur content, coal feeder speed, and air-to-coal ratio. Fuel inventory data includes the type of coal received and the amount in stock; Transportation and meteorological data include weather forecasts and road condition information; Power grid dispatch instructions include target loads for generating units; Specifically, in this embodiment, the specific steps involved in performing spatiotemporal alignment and normalization processing on the collected multi-source heterogeneous data to construct a unified feature vector are as follows: The collected multi-source heterogeneous data were interpolated and aligned using a unified timestamp, and the spatial dimension was unified to the unit level. The Z-score method is used to standardize numerical features and one-hot encoding is used for categorical features; The aligned and standardized multi-source heterogeneous data are concatenated column-wise into a unified feature vector. .

[0020] S2. Based on a unified feature vector, a soft measurement model for the calorific value of coal entering the furnace is constructed using a multi-layer LSTM network with physical constraint correction, to obtain the predicted calorific value of coal. And calculate cost deviation and environmental deviation based on the predicted calorific value of coal; Specifically, in this embodiment, the specific steps involved in constructing a soft measurement model of the calorific value of the coal fed into the furnace using a multi-layer LSTM network with physical constraint correction to obtain the predicted calorific value of the coal are as follows: The unified feature vector under historical operating conditions As input, a multi-layer LSTM network is constructed, and the preliminary prediction result of the calorific value of the coal entering the furnace is output. ; The core computation process of the LSTM unit is as follows: Input Gate: ; Forgotten Gate: ; Output gate: ; Candidate memory units: ; Memory unit: ; Hidden state: ; in, This represents the sigmoid activation function; The input gate weight matrix, Here is the forget gate weight matrix. This is the output gate weight matrix. The candidate memory unit weight matrix; This is the input gate bias vector. Forget gate bias vector, This is the output gate bias vector. The bias vector for candidate memory cells; This is the hidden state of the previous time step, used to pass on historical time sequence information; This represents element-wise multiplication (Hadamard product), which enables bit-by-bit filtering of the gating signal and the memory information. The hidden state at the current time step; The top-level output of a multi-layer LSTM network (composed of multiple LSTM units) As input to the prediction layer, the prediction layer passes through a fully connected layer. The high-dimensional features implied are mapped to the quantified value of the calorific value of the coal fed into the furnace, i.e., the preliminary prediction result. ; Based on the law of conservation of mass, a mass balance constraint is constructed. The mass balance constraint is used to limit the difference between the total input material flow rate and the total output material flow rate of the combustion system to approach 0 within a specific time step. Based on the law of conservation of energy, an energy balance constraint is constructed. This constraint is used to limit the difference between the total energy of the input materials and the total energy of the output materials of the combustion system, as well as the system heat loss, to approach zero within a specific time period. Physical constraints are constructed based on mass balance constraints and energy balance constraints, and mass deviation and net energy deviation are obtained. Mass balance constraints: ; Energy balance constraints: ; in, Indicates the first Input material flow rate (such as coal feed rate). Indicates the first The flow rate of various output materials (such as ash, slag, and flue gas). and These are the corresponding calorific values ​​(unit: kJ / kg) of the input and output materials, respectively. Δt represents the heat loss of the combustion system (including heat dissipation loss, flue gas heat loss, etc., unit: kJ), and Δt represents the time step (unit: h). For quality deviation (under ideal conditions) (This reflects the degree of material conservation). Net energy deviation (under ideal conditions) (This reflects the degree of energy conservation). The mean square error loss term is calculated based on the preliminary prediction results and the measured calorific value of coal, and the physical constraint loss term is constructed based on the quality deviation and net energy deviation. Mean squared error loss: ; Physical constraint loss: ; in, For the sample size, To measure the calorific value of coal, Preliminary prediction results for the output of a multilayer LSTM network ; The physical constraint loss is amplified by the squared term; The mean squared error loss of the multilayer LSTM network; The physical constraint loss term is incorporated as a regularization term into the mean squared error loss term of the multilayer LSTM network to form the total loss function; Total loss function: ; in, The physical constraint weight coefficients (dimensionless, optimized through 5-fold cross-validation, with values ​​ranging from 0.1 to 0.5). This represents the total loss in the soft measurement model of the calorific value of the coal fed into the furnace; A soft measurement model for the calorific value of coal entering the furnace is constructed by iteratively optimizing the weight matrix and bias vector of the multilayer LSTM network using the total loss function. The unified feature vector at the time to be predicted is then input into the trained soft measurement model to obtain the predicted calorific value of the coal. (Unit: kJ / kg); In this embodiment, the specific steps involved in calculating cost deviation and environmental deviation based on the predicted calorific value of coal are as follows: Based on power grid dispatch command data, the target load of generating units contained therein is analyzed. Combined with the preset boiler efficiency and standard coal consumption rate for power generation The target calorific value was calculated. ; ; in, The boiler efficiency is obtained by regression analysis based on historical boiler combustion data. The standard coal consumption rate for power generation is denoted as , and the inherent economic parameters of the unit are denoted as . The target calorific value (unit: kJ / kg) is the theoretical calorific value level of the fuel required to achieve the unit's target load. The target load of generating units included in the power grid dispatch instruction data; Based on measured coal feeder speed data from boiler combustion data The current coal feed rate is obtained by converting the coal feeder speed versus coal feed rate calibration curve function. ; Coal feeder speed-coal feed rate calibration curve function: ; in, This is the output coefficient of the coal feeder. The zero-point offset (unit: tons) is determined by the coal feeder equipment characteristic calibration test; A function representing the coal feeder speed versus coal feed rate calibration curve; This represents the current coal supply (unit: tons). This refers to the measured speed data of the coal feeder in the boiler combustion data. Based on predicted calorific value of coal Target calorific value Current coal supply Combined with the preset coal unit price Calculate cost deviation ; Cost deviation: ; in, Price of coal (unit: yuan / ton); Cost deviation is used to quantify the increase in fuel costs caused by the actual calorific value of the coal entering the furnace deviating from the scheduling target value; Based on predicted calorific value of coal The measured sulfur content in the boiler combustion data Combined with preset emission limits Environmental penalty coefficient Calorific value and environmental impact coefficient and the optimal environmental calorific value threshold Calculate environmental deviation ; Environmental deviations: ; in, To predict sulfur content (unit: %). Emission limits (unit: %) Environmental penalty coefficient (unit: yuan / unit exceeding the standard); The environmental impact coefficient is the calorific value. The optimal environmental calorific value threshold (unit: kJ / kg, obtained by regression analysis of historical environmental compliance data); This environmental deviation reflects the risk of environmental penalties that may result from excessive sulfur content.

[0021] S3. Process transportation and meteorological data using an XGBoost regression model to generate predicted transportation delays and their probability distributions; use predicted transportation delays, cost deviations, and environmental deviations as inputs, and combine this with a Bayesian risk network to quantify the probability of supply disruptions. Probability of exceeding environmental protection standards and the probability of cost overrun The risk weights are dynamically adjusted according to the power grid dispatch instructions, and the comprehensive risk score is calculated. ; In this embodiment, the specific steps involved in processing transportation and meteorological data using the XGBoost regression model to generate predicted transportation delays and their probability distributions are as follows: All features from the preprocessed transportation and meteorological data are concatenated into a transportation and meteorological feature vector. ; The transportation and meteorological data include precipitation, wind speed, visibility, and temperature from weather forecasts, as well as traffic congestion index, road type (such as highways, national highways, and provincial highways, processed by unique thermal coding), and historical average transportation time from road condition information. Using transportation and meteorological feature vectors Using the input as input, an XGBoost regression model is constructed by integrating multiple CART regression decision trees to obtain preliminary predictions of transportation delays; The mathematical expression of the XGBoost regression model: ; in, It is the first Predicted transport delay values ​​for each sample; Representing the Each CART regression decision tree is responsible for learning the preceding data. The residual between the tree combination prediction and the actual transportation delay gradually approximates the actual delay time. This represents the total number of decision trees in the model (a hyperparameter, to be determined through subsequent tuning). It is the function space of all possible CART trees; It is the first Transportation and meteorological feature vectors for each sample; The loss function term uses the squared loss function to quantify the difference between the predicted and actual transport delay values, and the regularization term uses the cumulative leaf nodes to quantify the complexity of the tree structure. The XGBoost regression model is trained by minimizing the objective function containing the loss function term and the regularization term. The objective function of the XGBoost regression model is: ; in, This represents the actual value of the transportation delay (obtained from historical data statistics). This is the predicted value for transportation delays. For the first The complexity regularization term for trees; The total number of training samples, This includes all parameters of the model (including the structure of each tree and the weights of the leaf nodes). The objective function value of the XGBoost regression model; Using the same transportation and meteorological feature vectors The quantile regression method is used to train the system with specific quantiles. (like Multiple XGBoost regression models with the objective of , and each XGBoost regression model independently outputs its corresponding quantile. Predicted value These predicted values ​​collectively constitute the probability distribution of the predicted transportation delay values. Based on the obtained set of quantile predictions A linear interpolation method is used to fit the probability distribution of the complete transport delay prediction value, and the transport delay prediction value is obtained from the probability distribution. This allows for the evaluation of arbitrary delay time thresholds. Risk probability ; The predicted transport delays are estimated using the quantile regression method. At different quantiles (e.g.) The values ​​under () are then fitted to obtain the probability distribution of the predicted transport delay values; ; in, This is the cumulative distribution function (usually assumed to be normally distributed). This is the predicted mean of the transport delay forecast. The forecast standard deviation of the predicted transportation delay values; Forecasted transportation delays (in hours); quantiles; For any delay time threshold The probability of risk; This is a probability calculation function; Furthermore, quantile regression directly estimates the conditional distribution of the dependent variable at a given quantile by minimizing the weighted absolute error loss function, thereby capturing the global influence of the predictor variable rather than just the central tendency; its core is the quantile. The weights are asymmetrically distributed, with error weights above the quantile being... The weight of the lower is ; In this embodiment, the specific steps involved in quantifying the probability of supply disruption risk, the probability of exceeding environmental standards risk, and the probability of cost overrun risk using a Bayesian risk network are as follows: Based on the risk impact mechanism of fuel supply to thermal power plants, risk nodes are set as supply interruption risk nodes, environmental protection exceeding risk nodes, and cost overrun risk nodes. The parent node (risk influencing factor) of the node with supply disruption risk is used as the predicted value for transportation delay. The parent node of the environmental protection exceedance risk node is the environmental protection deviation node. And the parent node of the cost overrun risk node is the cost deviation node. Construct a Bayesian risk network structure; Based on historical data statistics, a prior probability distribution is set for each parent node, and a conditional probability table (CPT) is defined to describe the dependency relationship between child nodes and parent nodes; The parent node includes the predicted transportation delay value. Environmental deviations and cost deviation ; Sub-nodes are risk nodes, including supply disruption risk nodes, environmental protection exceeding risk nodes, and cost overrun risk nodes; The Conditional Probability Table (CPT) is based on the historical fuel supply data of thermal power plants (including historical parent node values ​​and corresponding child node risk event records). First, the parent node is divided into different state intervals according to preset thresholds (such as the interruption time threshold, environmental allowable deviation threshold, and cost tolerance threshold). Then, the frequency of the occurrence and non-occurrence of child node risks in each interval is counted and converted into conditional probabilities to form the CPT. Introducing a preset supply interruption time threshold, based on an arbitrary delay time threshold. Risk probability Calculate the probability that the predicted transportation delay exceeds the supply disruption time threshold, thus obtaining the preliminary probability of supply disruption risk. ; ; in, The supply disruption time threshold is dynamically calculated based on the current inventory level and safety margin, and is usually 12 to 72 hours. This is a probability calculation function; The initial probability of supply disruption risk is calculated based on the probability distribution of predicted transportation delays (obtained from step S3). By introducing a preset environmental allowable deviation threshold, the probability of environmental deviation exceeding the environmental allowable deviation range is calculated based on the prior probability distribution of environmental deviation, thus obtaining the preliminary probability of environmental exceedance risk. ; ; in, The pre-set environmental allowable deviation threshold is set based on emission standards and historical environmental data, and is usually within a deviation range of ±5% to 10%. The preliminary probability of environmental pollution exceeding standards is calculated based on the prior probability distribution of environmental deviations (set according to historical data); By introducing a preset cost tolerance threshold, the probability of cost deviation exceeding the cost tolerance threshold is calculated based on the prior probability distribution of cost deviation, thus obtaining the preliminary probability of cost overrun risk. ; ; in, A pre-defined cost tolerance threshold, set based on budget control targets, is typically 5% to 10% of daily fuel costs. The initial probability of cost overrun risk is calculated based on the prior probability distribution of cost deviation (set according to historical data). Predicted transportation delays Environmental deviations and cost deviation The evidence is input into the Bayesian risk network, and the probability distribution of each node in the Bayesian risk network is updated through the Junction Tree algorithm, ultimately outputting the probability of supply disruption. Probability of exceeding environmental protection standards and the probability of cost overrun ; The posterior probability of each risk node is calculated using the Junction Tree algorithm; The specific formula is as follows: ; in, The observed state of the parent node. The type of risk to be assessed (i.e., risk node); This is the posterior probability, i.e., the observed state of the parent node ( When a certain type of risk ( The probability of occurrence (the core probability that needs to be evaluated in the end); Likelihood probability, i.e., a certain type of risk ( When this occurs, the current parent node state is observed. The probability of ( ) can be found in the Conditional Probability Table (CPT). Prior probability, that is, the probability of a certain type of risk when there is no observational evidence. The initial probability of occurrence (based on historical data); The evidence probability, i.e., the observed state of the current parent node ( The marginal probability of ) (used for probability normalization to ensure that the sum of the posterior probabilities is 1); The Junction Tree algorithm is an accurate algorithm for probabilistic inference in Bayesian networks. It transforms the Bayesian network into a clique tree structure, passes messages (marginal probabilities) between clique nodes, and finally efficiently calculates the posterior probability of the target node. This algorithm ensures the accuracy and operability of probabilistic queries in complex networks. In this embodiment, the specific steps involved in dynamically adjusting the risk weights according to the power grid dispatch instructions and calculating the comprehensive risk score are as follows: Based on the unit's target load By combining the preset load-risk weight mapping table, the risk weight of power outage is dynamically determined. Environmental protection exceeding risk weight and cost overrun risk weight ; The load-risk weight mapping table is constructed based on historical operating strategies and scheduling priority experience. A typical mapping relationship is shown below (not a fixed value, and needs to be calibrated according to actual operating conditions): High load ( )hour, , , Prioritize ensuring power supply safety and reduce the risk of power outages; Medium load ( )hour, , , A balance between power supply, environmental protection, and cost; Low load ( )hour, , , It focuses on environmental compliance and cost control; in, This is the rated load of the unit. For the risk weight of supply disruption, For environmental protection exceeding risk weight, As a risk weight for cost overrun, and satisfying ; Based on the probability of supply disruption Probability of exceeding environmental protection standards and the probability of cost overrun Combined with dynamically adjusted supply disruption risk weights Environmental protection exceeding risk weight and cost overrun risk weight The comprehensive risk score is calculated using a weighted summation method. ; ; in, The overall risk score is used to quantify the overall risk level of the current fuel supply strategy. Comprehensive risk score Perform max-min normalization to make it fall within the interval [0,1].

[0022] S4. An optimization model is constructed with the goal of minimizing the comprehensive risk score. A deep deterministic strategy gradient network is used to solve the Pareto optimal solution set under the inventory safety threshold constraint. Fuel scheduling instructions are generated based on the Pareto optimal solution set. Specifically, in this embodiment, the goal is to minimize the overall risk score. To establish an optimization model with the core objective as the objective. The optimized model must meet the inventory safety constraint, that is, the inventory level does not exceed the upper and lower bounds of the inventory safety threshold. Inventory safety constraints: ; in, The current inventory level (included in the unified feature vector) middle), This is the lower bound of the inventory safety threshold. This is the upper bound of the inventory safety threshold; Based on the unified feature vector at the current time and comprehensive risk score To obtain the environmental state ; Specifically, environmental conditions: ; in, For environmental conditions, This is the unified feature vector at the current moment. This represents the overall risk score at the current moment. Set Action It is a continuous action vector, and the actions include coal blending bin switching commands, fuel blending ratio vectors, and coal feed control commands for each coal feeder; Based on the objectives of minimizing overall risk and satisfying inventory safety constraints, a reward function is obtained through a weighted penalty mechanism; Define reward function : ; in, The penalty coefficient for inventory exceeding the lower bound. The penalty coefficient for inventory exceeding the upper limit; The reward function aims to minimize overall risk and strictly meet inventory safety constraints; This is a function to find the maximum value. Furthermore, the weighted penalty mechanism applies a negative reward to inventory out-of-bounds behavior through a reward function. If the inventory falls below the safe lower limit, the penalty intensity is the inventory out-of-bounds penalty coefficient. The product of the quantity shortage; if the inventory exceeds the safety limit, the penalty intensity is the inventory exceeding the upper limit penalty coefficient. The product of excess and the product of excess; by adjusting the inventory lower limit penalty coefficient. And the penalty coefficient for exceeding the upper limit of inventory Control the strength of constraints to force the strategy to meet inventory safety constraints; An initial Actor network (policy network) is constructed using a multi-layer fully connected neural network. The input of the initial Actor network is the environment state, and the output is the action. The input layer dimension is consistent with the environment state dimension, and the output layer dimension is consistent with the action dimension. The Tanh activation function is used to constrain the range of output action values. ; in, These are the parameters of the initial Actor network. The core function of the initial Actor network is to generate the optimal action that fits the current state. Similarly, a multi-layer fully connected neural network is used to construct the initial Critic network (value network). The input of the initial Critic network is the environment state-action pair, and the output is the value assessment. The input layer dimension of the initial Critic network is the sum of the environment state dimension and the action dimension, and the output layer dimension is 1. ; The core function of the initial Critic network is to evaluate the quality of the actions generated by the initial Actor network. These are the parameters for the initial Critic network. For the value assessment of the output; Simultaneously, a target Actor network with the same structure as the initial Actor network but different parameters, and a target Critic network with the same structure as the initial Critic network but different parameters are established. A deep deterministic policy gradient (DDPG) network is constructed based on the initial Actor network, the initial Critic network, the target Actor network, and the target Critic network. The transition samples, consisting of the current environment state, action, reward function, and the next environment state, are stored in the experience replay buffer. The initial Critic network parameters are updated by minimizing the mean squared error (MSE) between the target value assessment and the current value assessment. The initial Critic network parameters This includes the weight matrix and bias terms of the initial Critic network; Among them, the transferred samples are ; The initial loss function of the Critic network is: ; ; in, For the first Target value assessment of individual samples (based on target network calculation, reflecting future value). For the parameters of the target Actor network, The parameters of the target Critic network; For the target Actor network, For the target Critic network; Discount factor (dimensionless, This is used to weigh the importance of current rewards against future rewards; For the first The reward function for each sample; To find the mean function, This represents the initial loss value of the Critic network; To maximize the evaluation value of the initial Critic network, the parameters of the initial Actor network are updated using the policy gradient method. The initial Actor network parameters This includes the weight matrix and bias terms of the initial Actor network; The gradient formula is as follows: ; Make the actions generated by the initial Actor network as high as possible to maximize the evaluation value of the initial Critic network output; The parameters of the target Actor network and the target Critic network are updated using a soft update method, and the training is repeated until the constructed deep deterministic policy gradient network converges. Soft update method: ; ; in, This is the soft update coefficient (dimensionless, typically ranging from 0.001 to 0.01, used to control the smoothness of the target network parameter updates). Under the constraint of inventory safety, the trained deep deterministic policy gradient network is run to scan the action space and obtain a series of non-dominated solutions, which constitute the Pareto optimal solution set. From the Pareto optimal solution set, preferences (such as priority supply, priority economic protection, or equilibrium control) are determined according to the load-risk weight mapping table. Preferences are determined based on the load-risk weighting table: High load ( When [the situation is such that] supply is prioritized, priority is given to ensuring supply and screening for the probability of supply disruptions. The lowest-cost option; Medium load ( When balanced control is in place, the probability of supply disruption is considered. Probability of exceeding environmental protection standards and the probability of cost overrun All proposals are within a reasonable range; Low load ( When considering environmental protection and economic factors, prioritize environmental protection and economic considerations, and screen for the probability of exceeding environmental standards. and the probability of cost overrun The lowest-cost option; Based on the fuel inventory and unit coal consumption rate in the screening scheme, the coal blending bin switching sequence is obtained by setting the operating bin / standby bin and inventory switching threshold (such as the operating bin inventory being lower than 1 / 3 or 1 / 2 of the average daily coal consumption). Based on the calorific value and sulfur content in the screening scheme, combined with the basic properties of the fuel, the optimal blending ratio of the fuel is calculated. Based on the coal consumption demand corresponding to the blending ratio and the target load of the unit, the specific coal feeder speed command is converted into a calibration curve function of coal feeder speed-coal feed rate. Based on the current fuel inventory and the predicted calorific value of coal, and the predicted transportation delay, a coal replenishment dispatch order is calculated. The coal replenishment dispatch order includes the type of coal to be replenished, the amount of coal to be replenished, and the departure time. The decision module outputs the coal blending bin switching sequence, optimal fuel blending ratio, coal feeder speed command, and coal replenishment scheduling command as fuel scheduling commands.

[0023] S5. Execute the fuel scheduling command and dynamically adjust the coal blending strategy, and update the soft measurement model of the calorific value of the coal fed into the furnace online; Specifically, in this embodiment, based on the currently measured calorific value of coal... Compared with the current predicted calorific value of coal Calculate the prediction error at the current time. If the prediction error Exceeding the preset error threshold This triggers the online model update mechanism; ; in, For prediction error, The preset error threshold is used; Use a sliding window mechanism to retain the most recent The unified feature vectors and measured heat values ​​at each time step constitute the incremental training dataset. ; Based on the incremental training dataset, the parameters of the multilayer LSTM network are fine-tuned by minimizing the total loss function, thereby realizing the online update of the soft measurement model of the calorific value of coal entering the furnace. Example 2:

[0024] This embodiment provides an intelligent optimization system for fuel supply in thermal power plants based on multi-source information fusion, used to implement the intelligent optimization method for fuel supply in thermal power plants based on multi-source information fusion as described in Embodiment 1, characterized by: including... The data acquisition module collects boiler combustion data, fuel inventory data, transportation and meteorological data, and power grid dispatch command data; it performs spatiotemporal alignment and normalization processing on the above-mentioned multi-source heterogeneous data to construct a unified feature vector; The calorific value soft measurement module, based on a unified feature vector, uses a multi-layer LSTM network with physical constraint correction to construct a soft measurement model of the calorific value of the coal entering the furnace, thereby obtaining the predicted calorific value of the coal. And calculate cost deviation and environmental deviation based on the predicted calorific value of coal; The comprehensive risk assessment module processes transportation and meteorological data using an XGBoost regression model to generate predicted transportation delays and their probability distributions. Using predicted transportation delays, cost deviations, and environmental deviations as inputs, it combines a Bayesian risk network to quantify the probability of supply disruptions. Probability of exceeding environmental protection standards and the probability of cost overrun The risk weights are dynamically adjusted according to the power grid dispatch instructions, and the comprehensive risk score is calculated. ; The decision-making module constructs an optimization model with the goal of minimizing the comprehensive risk score, and uses a deep deterministic strategy gradient network to solve for the Pareto optimal solution set under the inventory safety threshold constraint. Fuel scheduling instructions are generated based on the Pareto optimal solution set. The update module executes the fuel scheduling instructions and dynamically adjusts the coal blending strategy, updating the soft measurement model of the calorific value of the coal fed into the furnace online.

[0025] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent optimization of fuel supply in thermal power plants based on multi-source information fusion, characterized in that, Includes the following steps: S1. Collect boiler combustion data, fuel inventory data, transportation and meteorological data, and power grid dispatch command data; perform spatiotemporal alignment and normalization processing on the above-collected multi-source heterogeneous data to construct a unified feature vector; S2. Based on a unified feature vector, a soft measurement model for the calorific value of coal entering the furnace is constructed using a multi-layer LSTM network with physical constraint correction, to obtain the predicted calorific value of coal. And calculate cost deviation and environmental deviation based on the predicted calorific value of coal; S3. Process transportation and meteorological data using an XGBoost regression model to generate predicted transportation delays and their probability distributions; use predicted transportation delays, cost deviations, and environmental deviations as inputs, and combine this with a Bayesian risk network to quantify the probability of supply disruptions. Probability of exceeding environmental protection standards and the probability of cost overrun The risk weights are dynamically adjusted according to the power grid dispatch instructions, and the comprehensive risk score is calculated. ; S4. An optimization model is constructed with the goal of minimizing the comprehensive risk score. A deep deterministic strategy gradient network is used to solve the Pareto optimal solution set under the inventory safety threshold constraint. Fuel scheduling instructions are generated based on the Pareto optimal solution set. S5. Execute the fuel dispatching command and dynamically adjust the coal blending strategy, and update the soft measurement model of the calorific value of the coal fed into the furnace online.

2. The intelligent optimization method for fuel supply in thermal power plants based on multi-source information fusion as described in claim 1, characterized in that: In step S1, the specific steps involved in performing spatiotemporal alignment and normalization processing on the collected multi-source heterogeneous data to construct a unified feature vector are as follows: The collected multi-source heterogeneous data were interpolated and aligned using a unified timestamp, and the spatial dimension was unified to the unit level. The Z-score method is used to standardize numerical features and one-hot encoding is used for categorical features; The aligned and standardized multi-source heterogeneous data are concatenated column-wise into a unified feature vector. .

3. The intelligent optimization method for fuel supply in thermal power plants based on multi-source information fusion as described in claim 1, characterized in that: In step S2, the specific steps involved in constructing a soft measurement model for the calorific value of the coal fed into the furnace using a multi-layer LSTM network with physical constraint correction to obtain the predicted calorific value of the coal are as follows: The unified feature vector under historical operating conditions As input, a multi-layer LSTM network is constructed, and the preliminary prediction result of the calorific value of the coal entering the furnace is output. ; Based on the law of conservation of mass, a mass balance constraint is constructed. The mass balance constraint is used to limit the difference between the total input material flow rate and the total output material flow rate of the combustion system to approach 0 within a specific time step. An energy balance constraint is constructed based on the law of conservation of energy. This constraint is used to limit the difference between the total energy of the input materials and the total energy of the output materials of the combustion system, as well as the system heat loss, to approach zero within a specific time period. Physical constraints are constructed based on mass balance constraints and energy balance constraints, and mass deviation and net energy deviation are obtained. The mean square error loss term is calculated based on the preliminary prediction results and the measured calorific value of coal, and the physical constraint loss term is constructed based on the quality deviation and net energy deviation. The physical constraint loss term is incorporated as a regularization term into the mean squared error loss term of the multilayer LSTM network to form the total loss function; A soft measurement model for the calorific value of coal entering the furnace is constructed by iteratively optimizing the weight matrix and bias vector of the multilayer LSTM network using the total loss function. The unified feature vector at the time to be predicted is then input into the trained soft measurement model to obtain the predicted calorific value of the coal. .

4. The intelligent optimization method for fuel supply in thermal power plants based on multi-source information fusion according to claim 3, characterized in that: In step S2, the specific steps involved in calculating cost deviation and environmental deviation based on the predicted calorific value of coal are as follows: Based on power grid dispatch command data, the target load of generating units contained therein is analyzed. Combined with the preset boiler efficiency and standard coal consumption rate for power generation The target calorific value was calculated. ; Based on measured coal feeder speed data from boiler combustion data The current coal feed rate is obtained by converting the coal feeder speed versus coal feed rate calibration curve function. ; Based on predicted calorific value of coal Target calorific value Current coal feed rate Combined with the preset coal unit price Calculate cost deviation ; Based on predicted calorific value of coal The measured sulfur content in the boiler combustion data Combined with preset emission limits Environmental penalty coefficient Calorific value and environmental impact coefficient and the optimal environmental calorific value threshold Calculate environmental deviations .

5. The intelligent optimization method for fuel supply in thermal power plants based on multi-source information fusion according to claim 1, characterized in that: In step S3, the specific steps involved in processing transportation and meteorological data using the XGBoost regression model to generate predicted transportation delays and their probability distributions are as follows: All features from the preprocessed transportation and meteorological data are concatenated into a transportation and meteorological feature vector. ; Using transportation and meteorological feature vectors Using the input as input, an XGBoost regression model is constructed by integrating multiple CART regression decision trees to obtain preliminary predictions of transportation delays; The loss function term uses the squared loss function to quantify the difference between the predicted and actual transport delay values, and the regularization term uses the cumulative leaf nodes to quantify the complexity of the tree structure. The XGBoost regression model is trained by minimizing the objective function that includes the loss function term and the regularization term. Using the same transportation and meteorological feature vectors The quantile regression method was used to train the system with specific quantiles. Multiple XGBoost regression models targeting a given goal; each XGBoost regression model independently outputs its corresponding quantile. Predicted value These predicted values ​​collectively constitute the probability distribution of the predicted transportation delay values. Based on the obtained set of quantile predictions A linear interpolation method is used to fit the probability distribution of the complete transport delay prediction value, and the transport delay prediction value is obtained from the probability distribution. This allows for the evaluation of arbitrary delay time thresholds. Risk probability .

6. The intelligent optimization method for fuel supply in thermal power plants based on multi-source information fusion according to claim 5, characterized in that: In step S3, the specific steps involved in quantifying the probabilities of supply disruption, environmental compliance, and cost overrun using a Bayesian risk network are as follows: Based on the risk impact mechanism of fuel supply to thermal power plants, risk nodes are set as supply interruption risk nodes, environmental protection exceeding risk nodes, and cost overrun risk nodes. The parent node of the nodes at risk of supply disruption is the predicted value of transportation delay. The parent node of the environmental protection exceedance risk node is the environmental protection deviation node. And the parent node of the cost overrun risk node is the cost deviation node. Construct a Bayesian risk network structure; Based on historical data statistics, a prior probability distribution is set for each parent node, and a conditional probability table is defined to describe the dependency relationship between child nodes and parent nodes. Introducing a preset supply interruption time threshold, based on an arbitrary delay time threshold. Risk probability Calculate the probability that the predicted transportation delay exceeds the supply disruption time threshold, thus obtaining the preliminary probability of supply disruption risk. ; By introducing a preset environmental allowable deviation threshold, the probability of environmental deviation exceeding the environmental allowable deviation range is calculated based on the prior probability distribution of environmental deviation, thus obtaining the preliminary probability of environmental exceedance risk. ; By introducing a preset cost tolerance threshold, the probability of cost deviation exceeding the cost tolerance threshold is calculated based on the prior probability distribution of cost deviation, thus obtaining the preliminary probability of cost overrun risk. ; Predicted transportation delays Environmental deviations and cost deviation The evidence is input into the Bayesian risk network, and the probability distribution of each node in the Bayesian risk network is updated through the Junction Tree algorithm, ultimately outputting the probability of supply disruption. Probability of exceeding environmental protection standards and the probability of cost overrun .

7. The intelligent optimization method for fuel supply in thermal power plants based on multi-source information fusion according to claim 6, characterized in that: In step S3, the specific steps involved in dynamically adjusting the risk weights according to the power grid dispatch instructions and calculating the comprehensive risk score are as follows: Based on the unit's target load By combining the preset load-risk weight mapping table, the power outage risk weight is dynamically determined. Environmental protection exceeding risk weight and cost overrun risk weight ; Based on the probability of supply disruption Probability of exceeding environmental protection standards and the probability of cost overrun Combined with dynamically adjusted supply disruption risk weights Environmental protection exceeding risk weight and cost overrun risk weight The comprehensive risk score is calculated using a weighted summation method. ; Comprehensive risk score Perform max-min normalization.

8. The intelligent optimization method for fuel supply in thermal power plants based on multi-source information fusion according to claim 1, characterized in that: In step S4, a deep deterministic strategy gradient network is used to solve for the Pareto optimal solution set under the inventory safety threshold constraint. The specific steps involved in generating fuel scheduling instructions based on this solution set are as follows: To minimize the overall risk score To establish an optimization model with the core objective as the objective. The optimized model must meet the inventory safety constraint, that is, the inventory level does not exceed the upper and lower bounds of the inventory safety threshold. Based on the unified feature vector at the current time and comprehensive risk score To obtain the environmental state ; Set Action It is a continuous action vector, and the actions include coal blending bin switching commands, fuel blending ratio vectors, and coal feed control commands for each coal feeder; Based on the objectives of minimizing overall risk and satisfying inventory safety constraints, a reward function is obtained through a weighted penalty mechanism; A multi-layer fully connected neural network is used to construct the Actor network. The input of the Actor network is the environment state, and the output is the action. Similarly, a multi-layer fully connected neural network is used to construct the initial Critic network. The input of the initial Critic network is the environment state-action pair, and the output is the value assessment. Simultaneously, a target Actor network with the same structure as the initial Actor network but different parameters, and a target Critic network with the same structure as the initial Critic network but different parameters are established. A deep deterministic policy gradient network is constructed based on the initial Actor network, the initial Critic network, the target Actor network, and the target Critic network. The transition sample, consisting of the current environment state, action, reward function, and the next environment state, is stored in the experience replay buffer. The initial Critic network parameters are updated by minimizing the mean square error between the target value assessment and the current value assessment. The initial Critic network parameters This includes the weight matrix and bias terms of the initial Critic network; To maximize the evaluation value of the initial Critic network, the parameters of the initial Actor network are updated using the policy gradient method. The initial Actor network parameters This includes the weight matrix and bias terms of the initial Actor network; The parameters of the target Actor network and the target Critic network are updated using a soft update method, and the training is repeated until the constructed deep deterministic policy gradient network converges. Under the constraint of inventory safety, the trained deep deterministic policy gradient network is run to scan the action space and obtain a series of non-dominated solutions, which constitute the Pareto optimal solution set. Preferences are determined from the Pareto optimal solution set based on the load-risk weight mapping table. Based on the fuel inventory and unit coal consumption rate in the screening scheme, the coal blending bin switching sequence is obtained by setting the operating bin / standby bin and inventory switching threshold; Based on the calorific value and sulfur content in the screening scheme, combined with the basic properties of the fuel, the optimal blending ratio of the fuel is calculated. Based on the coal consumption demand corresponding to the blending ratio and the target load of the unit, the specific coal feeder speed command is converted into a calibration curve function of coal feeder speed-coal feed rate. Based on the current fuel inventory and the predicted calorific value of coal, and the predicted transportation delay, a coal replenishment dispatch order is calculated. The coal replenishment dispatch order includes the type of coal to be replenished, the amount of coal to be replenished, and the departure time. The decision module outputs the coal blending bin switching sequence, optimal fuel blending ratio, coal feeder speed command, and coal replenishment scheduling command as fuel scheduling commands.

9. The intelligent optimization method for fuel supply in thermal power plants based on multi-source information fusion according to claim 1, characterized in that: In step S5, the specific steps for updating the soft measurement model of the calorific value of the coal fed into the furnace online and generating an emergency dispatch plan are as follows: Based on the current measured calorific value of coal Compared with the current predicted calorific value of coal Calculate the prediction error at the current time. If the prediction error Exceeding the preset error threshold This triggers the online model update mechanism; Use a sliding window mechanism to retain the most recent The unified feature vectors and measured heat values ​​at each time step constitute the incremental training dataset. ; Based on the incremental training dataset, the parameters of the multilayer LSTM network are fine-tuned by minimizing the total loss function, thereby realizing the online update of the soft measurement model of the calorific value of coal entering the furnace.

10. A smart optimization system for fuel supply in thermal power plants based on multi-source information fusion, used to implement the smart optimization method for fuel supply in thermal power plants based on multi-source information fusion as described in any one of claims 1-9, characterized in that: include The data acquisition module collects boiler combustion data, fuel inventory data, transportation and meteorological data, and power grid dispatch command data; it performs spatiotemporal alignment and normalization processing on the above-mentioned multi-source heterogeneous data to construct a unified feature vector; The calorific value soft measurement module, based on a unified feature vector, uses a multi-layer LSTM network with physical constraint correction to construct a soft measurement model of the calorific value of the coal entering the furnace, thereby obtaining the predicted calorific value of the coal. And calculate cost deviation and environmental deviation based on the predicted calorific value of coal; The comprehensive risk assessment module processes transportation and meteorological data using an XGBoost regression model to generate predicted transportation delays and their probability distributions. Using predicted transportation delays, cost deviations, and environmental deviations as inputs, it combines a Bayesian risk network to quantify the probability of supply disruptions. Probability of exceeding environmental protection standards and the probability of cost overrun The risk weights are dynamically adjusted according to the power grid dispatch instructions, and the comprehensive risk score is calculated. ; The decision-making module constructs an optimization model with the goal of minimizing the comprehensive risk score, and uses a deep deterministic strategy gradient network to solve for the Pareto optimal solution set under the inventory safety threshold constraint. Fuel scheduling instructions are generated based on the Pareto optimal solution set. The update module executes the fuel scheduling instructions and dynamically adjusts the coal blending strategy, updating the soft measurement model of the calorific value of the coal fed into the furnace online.

Citation Information

Cited By

  • A coal consumption prediction management system for a thermal power plant

    CN122198262A