Construction method and construction system for real-time prediction model of loading capacity of coal power unit

Abstract

The invention discloses a method for constructing a real-time prediction model for the loading capacity of a coal power unit based on machine learning. The method comprises the following steps: acquiring and preprocessing historical operation data of the coal power unit; constructing the data into time sequence samples, wherein each sample comprises an input sequence formed by a feature vector sequence and a corresponding future unit loading capacity target vector; and constructing a long-short-term memory network model, and training the model by using the sample to accurately predict the future load capacity of the unit. Furthermore, a unit load capacity real-time evaluation model is constructed to determine a current safe operation boundary, and the predicted future load capacity is combined with the boundary to provide decision support for unit operation. The invention further provides a construction system comprising the corresponding function modules. According to the method, the prediction accuracy is improved, the combination of prediction and a real-time security boundary is realized, and an optimal operation strategy can be generated.

Description

Technical Field

[0001] This invention relates to the field of power technology, and more specifically, to a method and system for constructing a real-time prediction model of the load-carrying capacity of coal-fired power units based on machine learning. Background Technology

[0002] With the increasing penetration rate of intermittent renewable energy sources such as wind and solar power in the power grid, coal-fired power units, as the basic support of the power grid, are required to have greater operational flexibility to adapt to large-scale, high-frequency load regulation, i.e., deep peak-shaving operation. Under such complex dynamic operating conditions, ensuring the safe, stable, and economical operation of the units has become a key technical challenge.

[0003] In deep peak shaving, accurately grasping the future load regulation potential of generating units, i.e., their load-carrying capacity, is crucial for grid dispatching and the operation of the units themselves. However, existing technologies often rely on calculations based on static or simplified models when assessing the load-carrying capacity of generating units. This makes it difficult to capture the long-term time dependence of multivariable and strongly coupled thermal power units in dynamic processes, resulting in insufficient accuracy in predicting future load-carrying capacity.

[0004] Furthermore, the information provided to operators by existing technologies is often fragmented. For example, some systems focus on real-time monitoring of the unit's current operating parameters to ensure safety, while others provide some form of trend forecasting. However, there is a lack of technical solutions to effectively combine forward-looking forecasting information with the unit's real-time safe operating boundaries. This results in incomplete decision-making information for operators, making it difficult to fully explore the unit's regulation potential while ensuring safety.

[0005] Furthermore, existing technologies typically stop after providing data monitoring or predictive information, failing to extend the analysis results to the control execution level. Even with assessment information on load capacity, how to specifically adjust the numerous controllable variables in the boiler combustion system (such as the opening degree of each damper) to meet load commands while simultaneously considering multiple optimization objectives such as coal consumption rate and pollutant emissions still heavily relies on the personal experience of operators. This approach struggles to guarantee operational optimality and cannot directly translate the value of predictive analysis into unit operating benefits. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method and system for constructing a real-time prediction model of the load-carrying capacity of coal-fired power units based on machine learning. This solves the problem that existing technologies struggle to accurately predict the load-carrying capacity of coal-fired power units under deep peak-shaving conditions in real time, and are unable to effectively diagnose the causes of load constraints, thereby affecting the flexibility and economy of unit operation.

[0007] To address the aforementioned technical problems, this invention provides a method and system for constructing a real-time prediction model for the load-carrying capacity of coal-fired power units based on machine learning.

[0008] The first aspect of this invention provides a method for constructing a real-time prediction model for the load-carrying capacity of coal-fired power units based on machine learning, comprising:

[0009] S1. Obtain historical operating data of the coal-fired power unit, which includes unit performance parameters, boiler combustion parameters, and key auxiliary equipment parameters;

[0010] S2. Perform data preprocessing on the historical operating data of the coal-fired power unit;

[0011] S3. The historical operating data after the data preprocessing is constructed into time series samples. Each time series sample includes an input sequence consisting of a feature vector sequence within a preset time window length, and a target vector consisting of the load-carrying capacity of the coal-fired power unit at a future preset time corresponding to the input sequence.

[0012] S4. Construct a long short-term memory network model, and train the long short-term memory network model using the time series samples to obtain a long short-term memory network model that can predict the load-carrying capacity of the coal-fired power unit in real time.

[0013] Preferably, step S2 specifically includes:

[0014] The historical operating data of the coal-fired power unit is subjected to max-min normalization processing, linearly mapping the historical operating data to a preset numerical range to obtain the preprocessed historical operating data. The calculation process is shown in the following formula:

[0015]

[0016] In the formula: x norm is the normalized eigenvalue; x is the original observed value of the feature; x min x is the minimum value of this feature in the historical running dataset; max This is the maximum value of this feature in the historical running dataset.

[0017] Preferably, between step S2 and step S3, the following step is also included:

[0018] The historical running data after data preprocessing is dimensionality reduced, and the principal components obtained from the dimensionality reduction are used as elements constituting each feature vector in the feature vector sequence.

[0019] In one specific embodiment, the load-carrying capacity of the coal-fired power unit includes the unit's maximum stable load-carrying capacity and the unit's minimum stable load-carrying capacity.

[0020] Preferably, the long short-term memory network model includes the following core units:

[0021] A time-series state information storage unit used to selectively transmit and maintain information;

[0022] A forget gate is used to determine the information to be discarded from the state information storage unit of the previous moment;

[0023] An input gate is used to determine whether to update the state information storage unit with new information;

[0024] An output gate is used to generate the hidden state at the current moment based on the updated state information storage unit.

[0025] The update process of its core computing unit at time t is as follows:

[0026] The forget gate calculation is shown in the formula:

[0027] f t =σ(W f ·[h t-1 ,x t ]+b f );

[0028] The input gate calculation is shown in the formula:

[0029] i t =σ(W i ·[h t-1 ,x t ]+b i );

[0030] The calculation of the candidate state information storage unit is shown in the formula:

[0031]

[0032] The update calculation for the status information storage unit is shown in the formula:

[0033]

[0034] The output gate is calculated as shown in the formula:

[0035] o t =σ(W o ·[h t-1 ,x t ]+b o );

[0036] The hidden state is calculated as shown in the formula:

[0037] h t =o t⊙tanh(C t );

[0038] In the formula: f t i is the activation vector of the forget gate; t The activation vector of the input gate; o t C is the activation vector of the output gate; t C is the vector of storage units for the current state information; t-1 This is the state information storage unit vector from the previous time step; h is the vector of candidate state information storage units at the current moment; t h is the hidden state vector at the current time step. t-1 x is the hidden state vector from the previous time step; t W is the input feature vector at time t; f W i W C W o These are the weight matrices corresponding to the forget gate, input gate, state information storage unit, and output gate, respectively; b f b i b C b o σ(·) represents the bias vectors corresponding to the forget gate, input gate, state information storage unit, and output gate, respectively; σ(·) is the Sigmoid activation function; tanh(·) is the hyperbolic tangent activation function; [·,·] represents the vector concatenation operation; ⊙ represents the Hadamard product operation.

[0039] In one specific embodiment, the Long Short-Term Memory network model transforms the hidden state of the last time step in the input sequence through a fully connected layer to output the load capacity of the unit at the future preset time corresponding to the target vector format, wherein the total number of time steps in the input sequence is determined by the preset time window length.

[0040] Preferably, step S4, the step of training the long short-term memory network model, includes:

[0041] The mean squared error (MSE) is used as the loss function. The weight parameters of the Long Short-Term Memory (LSTM) network model are adjusted using the backpropagation algorithm until the LSM model converges to the value of the loss function, thus completing the training. The MSE is calculated as shown in the formula:

[0042]

[0043] In the formula: L MSE N represents the mean squared error loss value. sample y is the total number of training samples; j Let j be the true target vector of the j-th sample; Let be the target vector predicted by the model for the j-th sample; This indicates that index j ranges from 1 to N. sample Perform a summation operation.

[0044] Furthermore, the construction method also includes:

[0045] Using the support vector regression method and historical operating data of the coal-fired power unit, a real-time load capacity assessment model is constructed. This model is used to determine the current safe operating boundary of the coal-fired power unit based on its real-time operating parameters.

[0046] The load capacity of the coal-fired power unit at a predetermined future time is predicted by the long short-term memory network model that can predict the load capacity of the coal-fired power unit in real time, and is combined with the safe operation boundary to provide decision support for the operation of the coal-fired power unit.

[0047] Furthermore, the method also includes:

[0048] By defining functions with the coal consumption rate and pollutant emission concentration of the coal-fired power unit as optimization objectives, the adjustable operating parameters of the coal-fired power unit as decision variables, and the safe operating boundary as constraints, a multi-objective optimization model for boiler combustion is constructed. The multi-objective optimization model for boiler combustion is solved online to obtain the optimal operating strategy, which can be used to guide or automatically adjust the operating parameters of the coal-fired power unit.

[0049] A second aspect of the present invention provides a system for constructing a real-time prediction model of the load-carrying capacity of coal-fired power units based on machine learning, applicable to the method described in any of the foregoing embodiments, comprising:

[0050] The data acquisition module is used to acquire historical operating data of coal-fired power units, including unit performance parameters, boiler combustion parameters, and key auxiliary equipment parameters.

[0051] The data preprocessing module is used to preprocess the historical operating data of the coal-fired power unit;

[0052] The sample construction module is used to construct time series samples from the historical operating data after the data preprocessing. Each time series sample includes an input sequence consisting of a feature vector sequence within a preset time window length and a target vector consisting of the unit's load capacity at a future preset time corresponding to the input sequence.

[0053] The model building module is used to build a long short-term memory network model;

[0054] The model training module is used to train the long short-term memory network model using the time series samples to obtain a long short-term memory network model that can predict the load-carrying capacity of the coal-fired power unit in real time.

[0055] In one specific embodiment, the system may further include a processor and a memory, the memory storing computer-executable instructions, the processor executing the computer-executable instructions to implement the method described in any of the foregoing embodiments.

[0056] This invention provides a method and system for constructing a real-time prediction model of the load-carrying capacity of coal-fired power units based on machine learning. It has the following beneficial effects:

[0057] 1. This invention constructs a long short-term memory network model and trains it using time series samples containing feature vector sequences. This effectively captures the long-range dependencies and dynamic characteristics of coal-fired power unit operation data in the time dimension. Therefore, the resulting prediction model can accurately predict the future load capacity of the unit, including the maximum and minimum stable load capacity, providing a reliable forward-looking data foundation for operation scheduling.

[0058] 2. This invention not only constructs a long short-term memory network model for predicting future load capacity, but also constructs a real-time assessment model of unit load capacity using the support vector regression method to determine the current safe operating boundary of the unit. By combining the predicted future load capacity with the real-time assessed safe operating boundary, it can provide a comprehensive decision-making basis that simultaneously considers future adjustment potential and current safety limitations, ensuring the integrity and security of decision support information.

[0059] 3. This invention further constructs a multi-objective optimization model for boiler combustion and uses the determined safe operating boundary as the constraint condition of the optimization model. By solving the function with coal consumption rate and pollutant emission concentration as optimization objectives online, it can generate specific and executable optimal operating strategies. This invention can transform the prediction results into control instructions that can be directly used to guide or automatically adjust operating parameters, realizing the technical extension from data analysis to optimization control. This helps to achieve economic and environmentally friendly operation of the unit while meeting the load capacity requirements. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of the overall process of the construction method according to an embodiment of the present invention;

[0061] Figure 2 This is a schematic diagram of the overall structure of the construction system according to an embodiment of the present invention. Detailed Implementation

[0062] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] Figure 1 This is a schematic diagram of the overall process of the construction method according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the overall structure of the construction system according to an embodiment of the present invention.

[0064] Please see the appendix Figure 1 and attached Figure 2 The method for constructing a real-time prediction model of the load-carrying capacity of coal-fired power units based on machine learning, provided in this embodiment of the invention, includes the following steps:

[0065] S1. Obtain historical operating data of the coal-fired power unit, including unit performance parameters, boiler combustion parameters, and key auxiliary equipment parameters.

[0066] S2. Perform data preprocessing on the acquired historical operating data of coal-fired power units.

[0067] S3. Construct time series samples from the historical operational data after data preprocessing.

[0068] S4. Construct and train a long short-term memory network model using time series samples to obtain a long short-term memory network model that can predict the load-carrying capacity of coal-fired power units in real time.

[0069] Specifically, in step S2, data preprocessing may include performing max-min normalization on the historical operating data, linearly mapping the historical operating data to a preset numerical range. The calculation for this process is as follows:

[0070]

[0071] In the formula: x norm is the normalized eigenvalue; x is the original observed value of the feature; x min x is the minimum value of this feature in the historical running dataset; max This is the maximum value of this feature in the historical running dataset.

[0072] Furthermore, between steps S2 and S3, the process may further include dimensionality reduction of the historical running data after data preprocessing, and using the principal components obtained from the dimensionality reduction as elements constituting the feature vectors in subsequent time series samples.

[0073] In step S3, each time series sample includes an input sequence consisting of a sequence of feature vectors within a preset time window, and a target vector consisting of the load-carrying capacity of the coal-fired power unit at a preset future time corresponding to the input sequence. The load-carrying capacity of the coal-fired power unit may include the unit's maximum stable load-carrying capacity and the unit's minimum stable load-carrying capacity.

[0074] In step S4, the constructed Long Short-Term Memory (LSTM) network model contains multiple core units. Its calculation process at time t is as follows:

[0075] f t =σ(W f ·[h t-1 ,x t ]+b f );

[0076] i t =σ(W i ·[h t-1 ,x t ]+b i );

[0077]

[0078] o t =σ(W o ·[h t-1 ,x t ]+b o );

[0079] h t =o t ⊙tanh(C t );

[0080] In the formula: f t i is the activation vector of the forget gate; t The activation vector of the input gate; o t C is the activation vector of the output gate; t C is the vector of storage units for the current state information; t-1 This is the state information storage unit vector from the previous time step; h is the vector of candidate state information storage units at the current moment; t h is the hidden state vector at the current time step. t-1 x is the hidden state vector from the previous time step; t W is the input feature vector at time t; f W i W C W o These are the weight matrices corresponding to the forget gate, input gate, state information storage unit, and output gate, respectively; b f b ib C b o σ(·) represents the bias vectors corresponding to the forget gate, input gate, state information storage unit, and output gate, respectively; σ(·) is the Sigmoid activation function; tanh(·) is the hyperbolic tangent activation function; [·,·] represents the vector concatenation operation; ⊙ represents the Hadamard product operation.

[0081] In step S4, when training the Long Short-Term Memory (LSTM) network model, the mean squared error can be used as the loss function, and the model's weight parameters can be adjusted through the backpropagation algorithm. The mean squared error is calculated as follows:

[0082]

[0083] In the formula: L MSE N represents the mean squared error loss value. sample y is the total number of training samples; j Let j be the true target vector of the j-th sample; Let be the target vector predicted by the model for the j-th sample; This indicates that index j ranges from 1 to N. sample Perform a summation operation.

[0084] In some embodiments, the method may further include: constructing a real-time load capacity assessment model for the unit using a support vector regression method to determine the current safe operating boundary of the unit. Then, the future load capacity predicted by the trained long short-term memory network model is combined with the safe operating boundary to provide decision support for unit operation.

[0085] In other embodiments, the method may further include: constructing a multi-objective optimization model for boiler combustion, wherein the model uses coal consumption rate and pollutant emission concentration as optimization objectives, adjustable operating parameters as decision variables, and a safe operating boundary as constraint. The optimal operating strategy is obtained by solving the model online, which is used to guide or automatically adjust the unit's operating parameters.

[0086] This invention also provides a construction system for executing the aforementioned construction method. The construction system includes: a data acquisition module 10, a data preprocessing module 20, a sample construction module 30, a model construction module 40, and a model training module 50.

[0087] The data acquisition module 10 is used to acquire historical operating data of coal-fired power units. The data preprocessing module 20 is used to preprocess the historical operating data. The sample construction module 30 is used to construct time series samples from the preprocessed data. The model construction module 40 is used to construct a long short-term memory network model. The model training module 50 is used to train the model using the time series samples to obtain a long short-term memory network model that can predict the load-carrying capacity of coal-fired power units in real time.

[0088] The following describes in detail each step of the method in the embodiments of the present invention.

[0089] In step S1, historical operating data of the coal-fired power unit is acquired. This data comes from the distributed control system (DCS) and monitoring information system (SIS) of the coal-fired power unit, and is collected at a preset time frequency (e.g., once per minute). The collected dataset forms the basis for subsequent model training.

[0090] The specific performance parameters of the unit may include: unit load (generator active power), main steam pressure, main steam temperature, reheat steam pressure, and reheat steam temperature.

[0091] The specific parameters of boiler combustion may include: total coal feed rate, speed of coal feeders at each level, total air volume, primary air volume, opening degree of dampers at each level of secondary air, opening degree of burnout air damper, furnace pressure, and oxygen content at the economizer outlet.

[0092] Key auxiliary parameters may include: primary air fan current, forced draft fan current, induced draft fan current, current and outlet temperature of each coal mill, and feedwater pump current.

[0093] The set of these parameters can comprehensively characterize the operating status and energy conversion process of coal-fired power units under different operating conditions.

[0094] In step S2, the acquired historical running data is preprocessed to eliminate noise and inconsistencies in the original data, preparing a high-quality dataset for subsequent dimensionality reduction and model training. This preprocessing includes data cleaning, data normalization, and other operations.

[0095] The first step in the data cleaning phase is to address missing values ​​in the data. For time series data, interpolation methods based on adjacent time points, such as linear interpolation or forward / backward imputation, can be used to fill in missing data points.

[0096] Subsequently, outliers in the data are identified and processed. The statistically based 3-sigma criterion can be used to identify and process outliers. Data points within a certain range are identified as outliers and replaced with the average or median of their nearest normal values. Here, μ is the mean of the data sequence. The standard deviation is denoted as .

[0097] In the data normalization stage, the max-min normalization method is used to linearly map all cleaned feature data to a unified numerical range. The purpose of this operation is to eliminate the influence of different physical units and numerical ranges on the model training process, ensuring that each feature contributes equally to the model weight update. The calculation process is as follows:

[0098]

[0099] In the formula: x norm is the normalized eigenvalue; x is the original observed value of the feature; x min x is the minimum value of this feature in the historical running dataset; max This is the maximum value of this feature in the historical running dataset.

[0100] In a preferred embodiment, after data preprocessing, the data is further subjected to dimensionality reduction. Because there are numerous collected operating parameters, and significant multicollinearity exists among them (e.g., total coal feed is highly correlated with total air volume), directly using all parameters would result in information redundancy. Therefore, a linear transformation is used to convert the original high-dimensional set of correlated variables into a set of lower-dimensional and linearly independent variables, i.e., principal components.

[0101] This method sorts the original data according to the variance explained by each principal component, and then selects the top K principal components whose cumulative variance contribution rate reaches a preset threshold (e.g., 95%) as new features. These selected principal components will be used as elements in each feature vector in the feature vector sequence in the subsequent step S3, thereby achieving data dimensionality reduction and eliminating collinearity while retaining most of the original data information.

[0102] Specifically, before performing principal component analysis, the covariance matrix of the original high-dimensional feature variables is first calculated. Then, eigenvalue decomposition is performed on this covariance matrix to obtain a series of eigenvalues ​​and their corresponding eigenvectors. The eigenvalues ​​are arranged in descending order, and the cumulative variance contribution rate is calculated. The formula for calculating the cumulative variance contribution rate is as follows:

[0103]

[0104] In the formula: R cumulative (k) represents the cumulative variance contribution rate of the first k principal components; λ i Let λ be the i-th eigenvalue; m Let m be the m-th feature value; M is the total dimension of the original features; This indicates that the summation operation is performed on index i from 1 to k; This indicates that the summation operation is performed on the index m from 1 to M.

[0105] Select R cumulative (k) The minimum k value that first reaches the preset threshold (e.g., 95%) is then projected onto the subspace formed by these k feature vectors to obtain the new k-dimensional features after dimensionality reduction.

[0106] In step S3, the historical operation data after data preprocessing is constructed into a series of time series samples for the supervised learning of the long short-term memory network model. In an embodiment of the present invention, this construction process uniquely extracts discrete samples with input-output correspondence relationships from continuous time series data.

[0107] Specifically, first, two key parameters are defined: the preset time window length L and the prediction step P. The preset time window length L defines the number of consecutive time points included in an input sequence. For example, if the data acquisition frequency is 1 minute and L = 60 is set, it means using the data of the past 60 minutes as an input sequence. The prediction step P defines the time distance of the target vector to be predicted relative to the end of the input sequence. For example, setting P = 15 means predicting the unit's load-carrying capacity 15 minutes after the current moment.

[0108] For a preprocessed data set containing N time points, this data set can be represented as an N×K feature matrix, where K is the number of features after data preprocessing and dimensionality reduction. Moving point by point along the time axis, at each time point t (where L - 1 ≤ t < N - P), the following operations are performed to generate a sample:

[0109] Extract the input sequence: Intercept the continuous data from time point t - L + 1 to time point t to form a matrix of size L×K. This matrix is the input sequence of a time series sample, which contains the feature vectors of L consecutive time points.

[0110] Determine the target vector: Obtain the load-carrying capacity data of the coal-fired power unit at the future time point t + P. This data constitutes a target vector. In an embodiment, if it is necessary to predict the maximum stable load-carrying capacity y max_load (t + P) and the minimum stable load-carrying capacity y min_load (t + P) of the unit simultaneously, then the target vector is a two-dimensional vector in the form of [y max_load (t + P), [y min_load (t + P)].

[0111] By pairing the above input sequence with the corresponding target vector, a complete time series sample is formed. Subsequently, the entire time window is slid backward by one time step, that is, the above extraction and pairing processes are repeated at time point t + 1 to generate the next sample. This process continues until the end of the time window reaches the end of the data set (until a complete input sequence and the corresponding future target vector can no longer be taken out), and finally a training data set composed of multiple time series samples is generated.

[0112] In step S4, a Long Short-Term Memory (LSTM) network model is constructed and trained. The operating data of coal-fired power units is typical time-series data, with its current state correlated with its states at multiple past moments, exhibiting complex dynamic characteristics and long-range dependencies. The LTM network model, through its unique gating structure, can learn long-term dependency patterns in the time dimension, thus making it suitable for predicting the load-carrying capacity of the units.

[0113] In the model building phase, the core structure of the network is first determined. This Long Short-Term Memory (LSTM) network model consists of an input layer, one or more LSTM layers, and an output layer. The input layer receives the input sequence from the time-series samples constructed in step S3.

[0114] The input sequence is then fed into the LSTM layer. The core of the LSTM layer is its internal computation unit. This unit controls the flow of information through three gating structures: a forget gate, an input gate, and an output gate. At one time step t, the workflow of this unit is as follows:

[0115] Forgot Gate: This gate is based on the hidden state h from the previous time step. t-1 and the input feature vector x at the current time. t An activation vector f is calculated using a sigmoid activation function σ(·). t Each element in this vector has a value between 0 and 1, determining the state information storage unit C from the previous moment. t-1 How much information is forgotten or retained?

[0116] Input gate: This gate determines which new information will be stored in the state information storage unit. It consists of two parts: First, a sigmoid layer computes the activation vector i. t First, it determines which values ​​need to be updated; second, a tanh activation function layer creates a vector of candidate state information storage units. It contains new candidate information.

[0117] Update the state information storage unit: Update the state information storage unit C from the previous moment. t-1 With the activation vector f of the forget gate t Perform the Hadamard product operation to discard some old information; simultaneously, change the activation vector i of the input gate. t With candidate state information storage unit vector Perform the Hadamard product to determine the part of the new information that needs to be added. Add the two results to obtain the updated state information storage unit C at the current time. t .

[0118] Output gate: This gate determines the output at the current time step. It first calculates the activation vector o through a sigmoid layer. t Then, the updated state information is stored in unit C. t The process is performed using a tanh activation function, and the result is compared with the activation vector o of the output gate. t Perform the Hadamard product operation to obtain the hidden state h at the current time step. t .

[0119] The mathematical expression of the above process is as follows:

[0120] f t =σ(W f ·[h t-1 ,x t ]+b f );

[0121] i t =σ(W i ·[h t-1 ,x t ]+b i );

[0122]

[0123] o t =σ(W o ·[h t-1 ,x t ]+b o );

[0124] h t =o t ⊙tanh(C t );

[0125] In the formula: f t i is the activation vector of the forget gate; t The activation vector of the input gate; o t C is the activation vector of the output gate; t C is the vector of storage units for the current state information; t-1 This is the state information storage unit vector from the previous time step; h is the vector of candidate state information storage units at the current moment; t h is the hidden state vector at the current time step. t-1 x is the hidden state vector from the previous time step; t W is the input feature vector at time t; f W i W C W oThese are the weight matrices corresponding to the forget gate, input gate, state information storage unit, and output gate, respectively; b f b i b C b o σ(·) represents the bias vectors corresponding to the forget gate, input gate, state information storage unit, and output gate, respectively; σ(·) is the Sigmoid activation function; tanh(·) is the hyperbolic tangent activation function; [·,·] represents the vector concatenation operation; ⊙ represents the Hadamard product operation.

[0126] After the entire input sequence (of length L) is processed by the LSTM layer, the hidden state h output at the last time step L is taken. l This vector contains condensed information about the entire input sequence. This hidden state h... l The data is fed into a fully connected layer (DenseLayer). This fully connected layer linearly transforms the high-dimensional hidden state vector into an output vector with the same dimension as the target vector, for example, a two-dimensional vector, whose two elements correspond to the predicted values ​​of the unit's maximum stable load capacity and minimum stable load capacity at a preset future time.

[0127] In one specific embodiment, the Long Short-Term Memory (LSTM) network model can be configured with two stacked LSTM layers, each with 128 hidden units. To prevent overfitting during training, a Dropout layer with a dropout rate of 0.2 can be added after each LSTM layer. This configuration improves the model's generalization ability while ensuring it has sufficient capacity to learn complex temporal features.

[0128] During the model training phase, the time series sample set generated in step S3 is first divided into a training set, a validation set, and a test set, for example, in a ratio of 7:1.5:1.5. The training set is used to learn the model weight parameters; the validation set is used to monitor model performance and perform hyperparameter tuning during training to prevent overfitting; and the test set is used to objectively evaluate the generalization ability of the final model after training is completed.

[0129] The training process employs mini-batch gradient descent. In each iteration, a mini-batch of samples is drawn from the training set and fed into the model for forward propagation to obtain the predicted output. Then, the mean squared error (MSE) is used as the loss function to calculate the error between the model's predicted value and the true target vector. The mean squared error is calculated as follows:

[0130]

[0131] In the formula: L MSE N represents the mean squared error loss value. sampley is the total number of training samples; j Let j be the true target vector of the j-th sample; Let be the target vector predicted by the model for the j-th sample; This indicates that index j ranges from 1 to N. sample Perform a summation operation.

[0132] Subsequently, the gradient of the loss function with respect to all trainable parameters of the model (i.e., the weight matrices and bias vectors of each layer) is calculated using the backpropagation (BPTT) algorithm. Finally, an optimizer, such as the Adam optimizer, is used to update the model parameters based on the calculated gradients. This process is repeated for multiple epochs on the entire training set until the loss function value on the validation set converges to a stable value, completing the training and yielding a Long Short-Term Memory (LSTM) network model capable of predicting the load-carrying capacity of coal-fired power units in real time.

[0133] After obtaining a long short-term memory network model capable of predicting future load capacity through step S4, the present invention can further implement the following technical solutions to form a complete application from prediction to decision support and optimization control.

[0134] In a preferred embodiment, the method further includes constructing a real-time assessment model of the unit's load capacity and combining it with the prediction results of a Long Short-Term Memory (LSTM) network model. The purpose of this step is to determine the instantaneous safe operating boundary of the unit at the current moment, determined by equipment status and operating parameters. This assessment model can be constructed using the Support Vector Regression (SVR) method. The SVR model is also trained using historical operating data of the coal-fired power unit. Unlike the LSTM network model, the training sample input of the SVR model is a feature vector of the unit's operating state at a certain moment, and its output target is the actual maximum and minimum stable load that the unit can achieve at that moment. SVR establishes a nonlinear mapping relationship between the input features and the output target by finding a high-dimensional "pipeline" with the largest interval that includes all sample points. When constructing the model, a Radial Basis Function (RBF) can be used to map the original features to a high-dimensional space.

[0135] The Radial Basis Function (RBF) was chosen because it effectively handles the nonlinear relationship between input features and the output target. The key hyperparameters of the RBF kernel are the penalty coefficient C and the kernel parameter gamma. These two hyperparameters are optimized using methods such as grid search and cross-validation to determine the optimal model configuration. The penalty coefficient C balances the model's complexity and the fitting error to the training samples, while gamma defines the magnitude of the influence range of a single training sample.

[0136] After training, the real-time load capacity assessment model for the unit was deployed online. It receives real-time unit operating parameters as input and instantly calculates and outputs the unit's current safe operating boundary, [Current_min_safe_load, Current_max_safe_load]. Subsequently, the system integrates the future load capacity [Predicted_min_load(t+P), Predicted_max_load(t+P)] predicted by the Long Short-Term Memory network model with the current safe operating boundary calculated by the real-time assessment model. The integrated information simultaneously presents the unit's future adjustment potential and current operational limitations, providing operators with a complete data foundation for making load adjustment decisions, ensuring that any adjustment operation is smoothly initiated within the current safe boundary.

[0137] In another preferred embodiment, the method further includes constructing and solving an online multi-objective optimization model for boiler combustion to generate an optimal operating strategy. The goal of this model is to translate the results of the aforementioned predictions and evaluations into a specific, actionable set of equipment control parameters.

[0138] The construction of this multi-objective optimization model for boiler combustion includes the following three elements:

[0139] The objective function is to optimize the economic and environmental performance of the unit, specifically minimizing the unit's coal consumption rate and minimizing the emission concentration of major pollutants (such as NOx). This constitutes a multi-objective optimization problem.

[0140] minF(u)=[f coal (u),f NOx (u)];

[0141] In the formula: min(·) is the minimization operator; F(u) is the multi-objective optimization function; u is the decision variable vector; f coal (u) is the coal consumption rate calculation function; f NOx (u) is the NOx emission concentration calculation function. These two functions can be obtained based on mechanistic models or through data-driven methods (such as constructing additional surrogate models).

[0142] Decision variables: Select directly adjustable operating parameters in the boiler combustion system as the decision variable vector u. Specifically, this may include: the opening degree of the secondary air damper at each level, the opening degree of the burnout air damper, the coal feed rate or speed of each coal mill, etc.

[0143] Constraints: To ensure the safety and feasibility of the optimization results, the following constraints are set:

[0144] Load command constraint: The optimized operating parameters must ensure that the unit load reaches the target load value P issued by the dispatching system.target .

[0145] Safe operating boundary constraints: No parameter adjustment during the optimization process can cause the unit to exceed the current safe operating boundary determined by the real-time load capacity assessment model of the unit.

[0146] Equipment physical constraints: The values ​​of all decision variables must be within their physical limits, such as the damper opening between 0% and 100%.

[0147] Process parameter constraints: fluctuations in key process parameters (such as main steam temperature and reheat steam temperature) must be within the allowable range.

[0148] The boiler combustion multi-objective optimization model was deployed for online solution. Upon receiving a new target load command, the system initiates optimization calculations. Intelligent optimization algorithms, such as multi-objective genetic algorithms (e.g., NSGA-II), can be used to solve the model. The solution is not a single optimal solution, but rather a Pareto optimal set. Each solution in this set represents a different operating strategy, i.e., a specific value for a set of decision variables, which achieve different trade-offs between the two objectives of coal consumption rate and pollutant emissions.

[0149] The system can provide this Pareto optimal solution set to operators for selection, or automatically select an optimal operating strategy based on preset preference rules (e.g., prioritizing the strategy with the lowest coal consumption while meeting emission standards), and use it as a recommended operating instruction or directly issue it to the distributed control system (DCS) for automatic adjustment.

[0150] When providing the Pareto optimal solution set to operators, the system can visualize this set in a two-dimensional coordinate system. The horizontal axis of this coordinate system represents the coal consumption rate, and the vertical axis represents the NOx emission concentration. The Pareto optimal solution set forms a boundary curve from the upper left to the lower right in this coordinate system. Operators can intuitively select an equilibrium point on this curve that meets the current operating requirements through a human-machine interface, and the system will then display the specific operating parameters (such as damper opening and coal feed rate) corresponding to that point.

[0151] Embodiments of the present invention also provide a construction system for implementing the aforementioned method. This construction system can be deployed on one or more server computers. The server computer includes a processor, memory, a network interface, and input / output devices. The memory stores computer-executable instructions, which the processor executes to implement the functions of the following modules.

[0152] The data acquisition module 10 is used to establish a communication connection with the distributed control system (DCS) or monitoring information system (SIS) of the coal-fired power unit. This connection can be achieved through the OPC (Open Platform Communication) protocol or a dedicated database interface. Based on a preset list of measuring points and acquisition frequency, this module periodically reads real-time values ​​of unit performance parameters, boiler combustion parameters, and key auxiliary equipment parameters from the DCS or SIS, and writes this timestamped data into a local time-series database as the original historical operating dataset.

[0153] The data preprocessing module 20 processes the raw data collected by the data acquisition module 10. This module first performs data cleaning, including filling missing values ​​in the data sequence using linear interpolation and identifying and replacing outliers using the 3-sigma criterion. Subsequently, the module performs data normalization, loading the minimum and maximum values ​​of each feature from historical datasets and applying the max-min normalization formula to transform the new data. If dimensionality reduction analysis is configured, the module will also load a pre-calculated dimensionality reduction transformation matrix and project the normalized data onto the principal component space, outputting the dimensionality-reduced feature vectors.

[0154] The sample construction module 30 performs different functions during the model training phase (offline state) and the online prediction phase. During offline training, this module reads preprocessed historical data and, based on the set time window length L and prediction step size P, uses a sliding window mechanism to generate time series samples in batches from the dataset. Each sample contains an input sequence and a corresponding target vector. During online prediction, this module extracts the feature vectors of the most recent L time points from the latest real-time data stream to form an input sequence for real-time prediction.

[0155] Model building module 40 provides a configuration interface for defining the specific architecture of the Long Short-Term Memory (LSTM) network model, including the number of LSTM layers, the number of neurons per layer, and the type of activation function. Similarly, this module is used to define the parameters of the Support Vector Regression (SVR) evaluation model (such as the kernel function type and penalty coefficient C) and the mathematical form of the boiler combustion multi-objective optimization model (including the objective function, decision variables and their ranges, and all constraints). This configuration information is saved as a model structure file.

[0156] The model training module 50 runs offline. It loads the model structure defined by the model building module 40 and the training sample set generated by the sample building module 30. This module calls the backend of a deep learning framework (such as TensorFlow or PyTorch) to execute mini-batch gradient descent and backpropagation algorithms, using mean squared error as the loss function to iteratively update the weights and biases of the LSTM model. The training process monitors performance on the validation set and stops training when convergence conditions are met. The final trained model parameters are then saved as a model file and stored in the system for online access. Similarly, this module 50 is also responsible for training the SVR evaluation model using historical data.

[0157] When the system is put into online operation, the processor executes instructions to enable the modules to work together:

[0158] Data acquisition module 10 continuously acquires real-time data;

[0159] The data preprocessing module 20 processes the data in real time;

[0160] Sample construction module 30 constructs the current input sequence;

[0161] A prediction execution unit loads the LSTM model file generated by the model training module 50, performs forward computation on the input sequence, and outputs the predicted value of the future load capacity.

[0162] Meanwhile, another evaluation execution unit loads the SVR model file, calculates the latest feature vectors, and outputs the current safe operating boundary.

[0163] These results are displayed via input / output devices or further passed to the optimization unit for online solving of the boiler combustion multi-objective optimization model, thereby generating the optimal operating strategy.

[0164] This invention provides a complete technical solution from data acquisition and processing, model building and training, to decision support through real-time evaluation, and finally to the generation of optimized control strategies.

[0165] The scheme first preprocesses historical operating data of coal-fired power units and constructs time-series samples to prepare standardized inputs for model training. Secondly, it constructs and trains a long short-term memory network model, utilizing its ability to process time-series data to predict the future load capacity of the units.

[0166] This approach further combines the predictive model with a support vector regression model used to assess the current safety operating boundary, integrating forward-looking predictive information with real-time safety constraints to output a comprehensive decision-making basis.

[0167] Finally, this scheme uses the determined safe operating boundary as a hard constraint condition for the boiler combustion multi-objective optimization model. By solving this model online, the prediction and evaluation results are transformed into a set of specific, executable operating parameters oriented towards achieving economic and environmental goals, thus forming a closed-loop process from data analysis to optimized control. This invention also provides a system containing corresponding functional modules, in which a processor in a computing device executes instructions from memory to implement all the steps of the above method.

Claims

1. A method for constructing a real-time prediction model of the load-carrying capacity of coal-fired power units based on machine learning, characterized in that, include: S1. Obtain historical operating data of the coal-fired power unit, which includes unit performance parameters, boiler combustion parameters, and key auxiliary equipment parameters; S2. Perform data preprocessing on the historical operating data of the coal-fired power unit; S3. The historical operating data after the data preprocessing is constructed into time series samples. Each time series sample includes an input sequence consisting of a feature vector sequence within a preset time window length, and a target vector consisting of the load-carrying capacity of the coal-fired power unit at a future preset time corresponding to the input sequence. S4. Construct a long short-term memory network model, and train the long short-term memory network model using the time series samples to obtain a long short-term memory network model that can predict the load-carrying capacity of the coal-fired power unit in real time.

2. The method for constructing a real-time prediction model for the load-carrying capacity of coal-fired power units based on machine learning according to claim 1, characterized in that, Step S2 specifically includes: The historical operating data of the coal-fired power unit is subjected to maximum-minimum normalization processing, and the historical operating data is linearly mapped to a preset numerical range to obtain the preprocessed historical operating data.

3. The method for constructing a real-time prediction model for the load-carrying capacity of coal-fired power units based on machine learning according to claim 1, characterized in that, Between step S2 and step S3, the following is also included: The historical running data after data preprocessing is dimensionality reduced, and the principal components obtained from the dimensionality reduction are used as elements constituting each feature vector in the feature vector sequence.

4. The method for constructing a real-time prediction model for the load-carrying capacity of coal-fired power units based on machine learning according to claim 1, characterized in that, The load-carrying capacity of the coal-fired power unit includes the unit's maximum stable load-carrying capacity and the unit's minimum stable load-carrying capacity.

5. The method for constructing a real-time prediction model for the load-carrying capacity of coal-fired power units based on machine learning according to claim 1, characterized in that, The Long Short-Term Memory (LSTM) network model includes the following core units: A time-series state information storage unit used to selectively transmit and maintain information; A forget gate is used to determine the information to be discarded from the state information storage unit of the previous moment; An input gate is used to determine whether to update the state information storage unit with new information; An output gate is used to generate the hidden state at the current moment based on the updated state information storage unit.

6. The method for constructing a real-time prediction model for the load-carrying capacity of coal-fired power units based on machine learning according to claim 5, characterized in that, The Long Short-Term Memory (LSTM) network model transforms the hidden state of the last time step in the input sequence through a fully connected layer to output the load capacity of the unit at the future preset time corresponding to the target vector format, wherein the total number of time steps in the input sequence is determined by the preset time window length.

7. The method for constructing a real-time prediction model for the load-carrying capacity of coal-fired power units based on machine learning according to claim 1, characterized in that, Step S4, the step of training the long short-term memory network model includes: The mean squared error is used as the loss function, and the weight parameters of the long short-term memory network model are adjusted through the backpropagation algorithm until the long short-term memory network model converges to the value of the loss function, thus completing the training.

8. The method for constructing a real-time prediction model for the load-carrying capacity of coal-fired power units based on machine learning according to claim 1, characterized in that, Also includes: Using the support vector regression method and historical operating data of the coal-fired power unit, a real-time load capacity assessment model is constructed. This model is used to determine the current safe operating boundary of the coal-fired power unit based on its real-time operating parameters. The load capacity of the coal-fired power unit at a predetermined future time is predicted by the long short-term memory network model that can predict the load capacity of the coal-fired power unit in real time, and is combined with the safe operation boundary to provide decision support for the operation of the coal-fired power unit.

9. The method for constructing a real-time prediction model for the load-carrying capacity of coal-fired power units based on machine learning according to claim 8, characterized in that, Also includes: By defining functions with the coal consumption rate and pollutant emission concentration of the coal-fired power unit as optimization objectives, the adjustable operating parameters of the coal-fired power unit as decision variables, and the safe operating boundary as constraints, a multi-objective optimization model for boiler combustion is constructed. The multi-objective optimization model for boiler combustion is solved online to obtain the optimal operating strategy, which can be used to guide or automatically adjust the operating parameters of the coal-fired power unit.

10. A system for constructing a real-time prediction model of the load-carrying capacity of coal-fired power units based on machine learning, characterized in that, The method for constructing a real-time prediction model of the load-carrying capacity of a coal-fired power unit based on machine learning, as described in any one of claims 1-9, wherein the construction system comprises: The data acquisition module is used to acquire historical operating data of coal-fired power units, including unit performance parameters, boiler combustion parameters, and key auxiliary equipment parameters. The data preprocessing module is used to preprocess the historical operating data of the coal-fired power unit; The sample construction module is used to construct time series samples from the historical operating data after the data preprocessing. Each time series sample includes an input sequence consisting of a feature vector sequence within a preset time window length and a target vector consisting of the unit's load capacity at a future preset time corresponding to the input sequence. The model building module is used to build a long short-term memory network model; The model training module is used to train the long short-term memory network model using the time series samples to obtain a long short-term memory network model that can predict the load-carrying capacity of the coal-fired power unit in real time.

Images