An incremental collaborative integrated learning power plant dry slagging machine intelligent air control method

Abstract

The present application belongs to the field of power plant dry slagging machine air control technology, especially relates to a power plant dry slagging machine intelligent air control method based on incremental collaborative integrated learning, which comprises the following steps: collecting multi-time sequence real-time features of different nodes of the slagging machine by using a sensor; inputting the multi-time sequence real-time features into a basic learning model to obtain predicted characteristic values, wherein the basic learning model comprises a collaborative graph neural network, a space-time long short-term memory neural network and an incremental learning network; the collaborative graph neural network is used for performing an aggregation operation on the multi-time sequence real-time features to obtain first predicted characteristic values; the space-time long short-term memory neural network is used for forward propagation learning of space-time sequence information of the first predicted characteristic values to obtain second predicted characteristic values; the incremental learning network is used for dynamically obtaining third predicted characteristic values from the multi-time sequence real-time features; and the predicted characteristic values are input into a meta-learning model to obtain air volume of the slagging machine at the next moment.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power plant dry slagging machine air control, and particularly relates to an incremental collaborative integrated learning intelligent air control method for a power plant dry slagging machine. BACKGROUND

[0002] The dry slagging machine is the core equipment of the power plant boiler slagging system, and the cooling air volume is controlled to realize the cooling, conveying and heat recovery of high-temperature slag. The air control technology directly determines the cooling efficiency, heat recovery, environmental protection performance, equipment life and other indicators of the system. The existing air control technology is mainly divided into traditional control methods and intelligent control methods. The traditional control methods mainly include PID control methods and static model prediction methods. The PID control technology mainly adjusts the air volume through fixed parameters. This method cannot adapt to dynamic conditions such as boiler load fluctuations and coal quality changes, and the response lag is obvious. The static model prediction method establishes a linear regression or an empirical formula through historical data, but ignores the influence of environmental variables (such as seasonal temperature and humidity), and has a high error rate. The intelligent control methods mainly include fuzzy control, single-model machine learning and deep reinforcement learning methods. However, due to load disturbance and coal quality changes, the model is inaccurate, the recovery efficiency is low, and it is easily affected by spatial characteristics and time characteristics. SUMMARY

[0003] To solve the above technical problems, the application provides an incremental collaborative integrated learning intelligent air control method for a power plant dry slagging machine. The time and space characteristics of the slagging machine system are considered. Integrated learning is an important data mining method, mainly using the integration of multiple learners to solve problems, which can significantly improve the generalization ability of the learning system, improve the prediction accuracy, adapt to dynamic conditions, and adaptively adjust the cooling air volume to improve the power generation efficiency of the power plant.

[0004] To achieve the above purpose, the application provides an incremental collaborative integrated learning intelligent air control method for a power plant dry slagging machine, which comprises the following steps:

[0005] Collecting multiple time sequence real-time features of different nodes of the slagging machine by using a sensor;

[0006] Inputting the multiple time sequence real-time features into a basic learning model to obtain predicted feature values, wherein the basic learning model comprises a collaborative graph neural network, a space-time long short-term memory neural network and an incremental learning network;

[0007] The collaborative graph neural network is used for aggregation operation on the multiple time sequence real-time features to obtain first predicted feature values;

[0008] The space-time long short-term memory neural network is used for forward propagation learning of space-time sequence information of the first predicted feature values to obtain second predicted feature values;

[0009] The incremental learning network is configured to dynamically obtain a third predicted feature value from the multi-time real-time features.

[0010] The predicted feature value is input into the meta-learning model to obtain a next-time blast furnace air volume.

[0011] Optionally, the multi-time real-time features include wind speed, wind direction, ambient temperature, air density, and air pressure.

[0012] Optionally, before training the base learner using the multi-time real-time features, the method further includes normalizing the multi-time real-time features.

[0013] ,

[0014] wherein x r represents historical data of the multi-time real-time features, x min represents a minimum value in the sensor node data, x max represents a maximum value in the sensor node data, and x represents the multi-time real-time features.

[0015] Optionally, the base learning model is configured to output the predicted feature value.

[0016] The predicted feature value is obtained based on the first predicted feature value, the second predicted feature value, and the third predicted feature value.

[0017] Optionally, the collaborative graph neural network includes a mapping module, a time attention module, a spatial attention module, a spatio-temporal feature aggregation module, and an output module.

[0018] The mapping module is configured to map the multi-time real-time features to a high-dimensional hidden space and extract mapped spatio-temporal features.

[0019] The time attention module is configured to capture key events in the historical time sequence features and obtain time weights.

[0020] The spatial attention module is configured to extract spatial correlations of different sensor nodes and obtain spatial weights.

[0021] The spatio-temporal feature aggregation module is configured to aggregate the mapped spatio-temporal features using the time weights and the spatial weights.

[0022] The output module is configured to output the aggregated first predicted feature value.

[0023] Optionally, the method for obtaining the first predicted feature value includes:

[0024] ,

[0025] wherein denotes the feature information after feature aggregation, denotes the activation factor, denotes the attention coefficient of the collaborative graph neural network, W denotes the weight parameter matrix, and v i denotes the intermediate feature of the real-time feature, v j denotes v i adjacent features of v j is the real-time feature of node j.

[0026] Optionally, the space-time long short-term memory neural network comprises a space-time feature extraction module, a space-time memory gate module, a space-time fusion module, a feature updating module and a feature output module.

[0027] The space-time feature extraction module is configured to extract spatial features and temporal features of the first predicted feature value respectively.

[0028] The space-time memory gate module is configured to input the spatial features and the temporal features into a spatial memory gate and a temporal memory gate respectively, and train in a forward propagation manner.

[0029] The space-time fusion module is configured to fuse the trained spatial features and temporal features to obtain a fused space-time sequence data set.

[0030] The feature updating module is configured to filter the fused space-time sequence data set by using an updating gate.

[0031] The feature output module is configured to output the filtered space-time sequence data set to obtain a second predicted feature value.

[0032] Optionally, the method for dynamically obtaining a third predicted feature value from the multi-time sequence real-time features comprises:

[0033] ,

[0034] wherein λ is a sampling coefficient from a random distribution, x i is the real-time feature of node i, x j is the real-time feature of node j, y i is the increased real-time feature of node i, y j is the increased multi-time sequence real-time feature of node j, is the differential of the original real-time feature, is the differential of the increased real-time feature.

[0035] Optionally, the model parameters of the meta-learning model are:

[0036] ,

[0037] wherein, φ is a model parameter, L loss is a loss function of the model, D test is a test set, is the optimal model parameter of the training set, is the training set.

[0038] Compared with the prior art, the present application has the following advantages and technical effects:

[0039] The present application is based on the concept of ensemble learning, which integrates three basic learners and a meta-learner, and can adapt to dynamic working conditions and adjust cooling air volume. The collaborative graph neural network of the basic learner 1 is good at capturing long-term dependencies in time series. The spatio-temporal long short-term memory neural network of the basic learner 2 can learn sequence information forwardly and provide more comprehensive feature capturing ability. The incremental learning of the basic learner 3 can dynamically obtain information from new data and gradually adjust the model to adapt to changing environments and data. Meta-learning can automatically learn the feature representation in the base learner, thereby generalizing on the meta-learner. BRIEF DESCRIPTION OF DRAWINGS

[0040] The accompanying drawings, which form a part of this application, are intended to provide further understanding of the application and are incorporated herein in their entirety, and the illustrative embodiments thereof and their description serve the purpose of explanations and are not intended to limit the present application. In the drawings:

[0041] Figure 1 is a flow chart of an incremental collaborative ensemble learning method for intelligent air control of a dry slag discharger of a power plant in an embodiment of the present application;

[0042] Figure 2 is a collaborative graph neural network structure diagram in an embodiment of the present application;

[0043] Figure 3 is a spatio-temporal long short-term memory neural network diagram in an embodiment of the present application;

[0044] Figure 4 is an ensemble learning framework diagram in an embodiment of the present application. DETAILED DESCRIPTION

[0045] It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0046] It should be noted that the steps shown in the flow chart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flow chart, in some cases, the steps shown or described herein can be executed in an order different from that shown herein.

[0047] The embodiment provides an intelligent air control method for a power plant dry slag discharger based on incremental collaborative integrated learning, as shown in the figure, and specifically comprises the following steps: Figure 1

[0048] Collecting multi-time sequence real-time features of different nodes of the slag discharger by using sensors;

[0049] Inputting the multi-time sequence real-time features into a basic learning model to obtain predicted feature values, wherein the basic learning model comprises a collaborative graph neural network, a space-time long short-term memory neural network and an incremental learning network;

[0050] The collaborative graph neural network is used for performing an aggregation operation on the multi-time sequence real-time features to obtain first predicted feature values;

[0051] The space-time long short-term memory neural network is used for forward propagation learning of space-time sequence information of the first predicted feature values to obtain second predicted feature values;

[0052] The incremental learning network is used for dynamically obtaining third predicted feature values from the multi-time sequence real-time features;

[0053] Inputting the predicted feature values into a meta-learning model to obtain air volume of the slag discharger at a next time.

[0054] Specifically, first, the spatial correlation of different sensor nodes (such as a slag bin, a slag crusher, an air duct, a slag well entrance, a middle part of a conveying belt, a slag crusher exit and the like) is collected, and weights are dynamically allocated. The time attention module captures feature data of key events (such as slag amount peak values and temperature abnormities) in historical time sequences and performs feature aggregation; then, the aggregated feature data is input into different basic learning devices for training and prediction; finally, the prediction results of the different basic learning devices are input into a meta-learning device based on meta-learning for training and prediction, so that the air volume of the slag discharger at a next time is obtained.

[0055] Further, the multi-time sequence real-time features comprise wind speed, wind direction, environmental temperature, air density and air pressure.

[0056] Specifically, the geometric structure of the slag discharger can be recorded as G= (V, A), wherein V represents a sensor node set, |V|=N, A={A k |k=1,2,..., N} represents an adjacency matrix between different nodes, and k is the number of sensors. The multi-element data at a certain time step can be represented as represents information collected by the i node at the time step t, such as wind speed, wind direction, environmental temperature, air density, air pressure, air pollution data and the like.

[0057] Further, before training the basic learning device by using the multi-time sequence real-time features, the multi-time sequence real-time features are subjected to normalization processing: ​

[0058] ,

[0059] wherein x r represents the historical data of the multi-time sequence real-time feature, x min represents the minimum value in the sensor node data, x max represents the maximum value of the sensor node data, and x represents the multi-time sequence real-time feature.

[0060] Further, the basic learning model is used to output the predicted feature value, which includes:

[0061] Based on the first predicted feature value, the second predicted feature value and the third predicted feature value, the predicted feature value is obtained.

[0062] Further, the collaborative graph neural network includes a mapping module, a time attention module, a spatial attention module, a space-time feature aggregation module and an output module.

[0063] The mapping module is used to map the multi-time sequence real-time feature to a high-dimensional hidden space and extract the mapped space-time feature.

[0064] The time attention module captures key events in the historical time sequence feature and obtains a time weight.

[0065] The spatial attention module is used to extract the spatial correlation of different sensor nodes and obtain a spatial weight.

[0066] The space-time feature aggregation module is used to aggregate the mapped space-time feature using the time weight and the spatial weight.

[0067] The output module is used to output the aggregated first predicted feature value.

[0068] Specifically, as shown in Figure 2 , the collaborative graph neural network maps the space-time sequence data collected by the sensor to a high-dimensional hidden space and extracts its space-time feature; the time attention module captures key events (such as slag amount peak value, temperature anomaly) in the historical time sequence, and enhances the modeling of long-term dependence; the spatial attention module mainly extracts the spatial correlation of different sensor nodes (such as slag bin, slag crusher, air duct, slag well entrance, middle part of conveying belt, slag crusher outlet, etc.), and dynamically allocates weights. The time and space attention weights are respectively applied to the input features, and the weighted sum is obtained, the aggregated adjacent node information and the previous layer information are spliced, and the weights are adaptively allocated through the activation function; finally, the feature information after the aggregation operation is output.

[0069] More specifically, the spatial attention module activation function adopts Leaky ReLu, thereby dynamically and adaptively allocating the weights of each feature information. The specific representation is as follows:

[0070]

[0071] ,

[0072] in, This represents the feature information after feature aggregation in a collaborative graph neural network. Indicates activator. V represents the attention coefficients of the collaborative graph neural network, W represents the weight parameter matrix, and v i Indicates intermediate features, v j v i Let N represent the number of features in the slag discharger system, and X represent adjacent features. j X represents the real-time features of node j. i Let i be the real-time feature of node i.

[0073] Furthermore, the spatiotemporal long short-term memory neural network includes: a spatiotemporal feature extraction module, a spatiotemporal memory gate module, a spatiotemporal fusion module, a feature update module, and a feature output module;

[0074] The spatiotemporal feature extraction module is used to extract the spatial and temporal features of the first predicted feature value, respectively.

[0075] The spatiotemporal memory gate module is used to input spatial features and temporal features into the spatial memory gate and the temporal memory gate respectively, and to train them using a forward propagation method.

[0076] The spatiotemporal fusion module is used to fuse the spatial and temporal features after training to obtain a fused spatiotemporal sequence data set.

[0077] The feature update module is used to filter the fused spatiotemporal sequence data group using an update gate;

[0078] The feature output module is used to output the filtered spatiotemporal sequence data set and obtain the second predicted feature value.

[0079] Specifically, such as Figure 3 As shown, the input to the spatiotemporal long short-term memory neural network is the spatiotemporal sequence data group after the spatiotemporal attention network aggregation operation. Spatial features and temporal features are placed into spatial memory gates and temporal memory gates respectively, and training is performed using a forward propagation method. The spatiotemporal long short-term memory neural network filters the input temporal data through update gates, i.e., memory and forgetting. The update gate is essentially an activation function. When the input feature is a vector, the output is a real number vector between [0,1]. When the update gate is 0, the update gate is closed, and no vector can pass through; when the update gate is 1, any vector can pass through. Therefore, the feature information is filtered by controlling the update gate:

[0080] ,

[0081] Update stage update gate z t forgetting and memory principle;

[0082] ,

[0083] The output of the spatiotemporal long short-term memory neural network at time t;

[0084] ,

[0085] In the formula, , g t , respectively represent the feature information after the feature aggregation of the collaborative graph neural network, the updated feature information, and the output feature information; W i , W f , W g , W o are weights; b i , b f , b g , b o are bias vectors; is a sigmoid activation function, x t is the node feature at time t, h t-1 is the hidden state at the previous time, z t is the intermediate calculation result of the gate unit at time t.

[0086] Further, incremental learning can solve the problem of low prediction accuracy on old samples during training, and the process of updating and improving the model itself using old samples is a gradual learning process that dynamically obtains information from new data and gradually adjusts the model, which can adapt to changing environments and data. In the test set D test , two different class samples (x i , y i ), (x j , y j ) are randomly selected, y i ≠ y j , to increase the sample type:

[0087] ,

[0088] where λ is a sampling coefficient from a random distribution, x i is the real-time feature of node i, x j is the real-time feature of node j, y i is the increased real-time feature of node i, y j is the real-time feature of node j, is the differential of the original real-time feature, To increase the differential of real-time features.

[0089] Further, the meta-learner fuses the outputs of the base learners based on the algorithms of the base learners to generate a final decision result. The application adopts a meta-learning algorithm as the meta-learner, and the meta-learning can automatically learn the feature representation in the base learners, thereby generalizing on the meta-learner. The purpose of the meta-learning is to train the test set D test Get the model parameters φ, so that the loss function L loss is minimum, and the optimal model parameters can be expressed as follows:

[0090] ,

[0091] Where φ is the model parameter, L loss is the loss function of the model, D test is the test set, is the optimal model parameter of the training set, is the training set.

[0092] As Figure 4 shown, the ensemble learning fuses the advantages of multiple algorithms to adapt to various working condition scenes. The ensemble learning mainly consists of base learners and meta-learners. First, train the learners with all the data, each base learner predicts the training data set, and the prediction result is taken as the feature value of the test set. Then, the new training sample data trains the meta-learner to obtain the air volume of the slag extractor at the next moment.

[0093] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited to this. Any changes or replacements that can be easily thought of by those skilled in the art within the technical range disclosed in the present application should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A power plant dry slagging machine intelligent air control method for incremental synergistic ensemble learning, characterized in that, The method comprises the following steps: Collecting multi-time real-time features of different nodes of the slag extractor by using sensors; Inputting the multi-time real-time features into a basic learning model to obtain predicted feature values, wherein the basic learning model comprises a collaborative graph neural network, a spatio-temporal long short-term memory neural network and an incremental learning network; The collaborative graph neural network is used for performing aggregation operation on the multi-time real-time features to obtain first predicted feature values; The spatio-temporal long short-term memory neural network is used for forward propagation learning of spatio-temporal sequence information of the first predicted feature values to obtain second predicted feature values; The incremental learning network is used for dynamically obtaining third predicted feature values from the multi-time real-time features; The basic learning model is used for outputting the predicted feature values, comprising: Obtaining the predicted feature values based on the first predicted feature values, the second predicted feature values and the third predicted feature values; Inputting the predicted feature values into a meta-learning model to obtain air volume of the slag extractor at the next time.

2. The method of claim 1, wherein the method is characterized by, The multi-time real-time features comprise wind speed, wind direction, environmental temperature, air density and air pressure.

3. The method of claim 1, wherein the method further comprises: Before training the basic learning model by using the multi-time real-time features, the multi-time real-time features are further subjected to normalization processing. ， where x r represents historical data of multi-chronological real-time features, x min represents a minimum value in sensor node data, x max represents a maximum value of sensor node data, x represents multi-chronological real-time features.

4. The method of claim 1, wherein the method further comprises: The collaborative graph neural network comprises a mapping module, a time attention module, a space attention module, a spatio-temporal feature aggregation module and an output module; The mapping module is used for mapping the multi-time real-time features to a high-dimensional hidden space and extracting mapped spatio-temporal features; The time attention module is used for capturing key events in historical time sequence features to obtain time weights; The space attention module is used for extracting spatial correlation of different sensor nodes to obtain space weights; The spatio-temporal feature aggregation module is used for aggregating the mapped spatio-temporal features by using the time weights and the space weights; The output module is used for outputting the aggregated first predicted feature values.

5. The method of claim 4, wherein the method further comprises: The method for obtaining the first predicted feature values comprises: ， wherein, denotes the feature information after feature aggregation, denotes an activation factor, denotes an attention coefficient of the collaborative graph neural network, W denotes a weight parameter matrix, v i denotes an intermediate feature of the real-time feature, v j denotes v i adjacent features of v j is the real-time feature of node j.

6. The method of claim 1, wherein the method is a method of incremental synergistic integrated learning for intelligent control of a power plant dry slag tap hole, characterized by, The spatio-temporal long short-term memory neural network comprises a spatio-temporal feature extraction module, a spatio-temporal memory gate module, a spatio-temporal fusion module, a feature update module and a feature output module; The spatio-temporal feature extraction module is used for respectively extracting spatial features and time features of the first predicted feature values; The spatio-temporal memory gate module is used for respectively inputting the spatial features and the time features into a spatial memory gate and a time memory gate and training in a forward propagation manner; The spatio-temporal fusion module is used for fusing the trained spatial features and time features to obtain fused spatio-temporal sequence data groups; The feature update module is used for screening the fused spatio-temporal sequence data groups by using an update gate; The feature output module is used for outputting the screened spatio-temporal sequence data groups to obtain second predicted feature values.

7. The method of claim 1, wherein the method further comprises: The method for dynamically obtaining third predicted feature values from the multi-time real-time features comprises: ， where λ is a sampling coefficient from a random distribution, x i is the real-time feature for node i, x j is the real-time feature for node j, y i is the increased real-time feature for node i, y j is the increased real-time feature for node j, is the differential of the original real-time feature, is the differential of the increased real-time feature.

8. The method of claim 1, wherein the method is a method of incremental synergistic integrated learning for intelligent control of a power plant dry slag tap hole, characterized by, The model parameters of the meta-learning model are: ， where φ is the model parameter, L loss is the loss function of the model, D test is the test set, is the optimal model parameter of the training set, is the training set.

Images