A method for predicting the output power of a steam turbine based on vacuum degree guidance

Abstract

A method for predicting the output power of a steam turbine based on vacuum degree guidance, comprising: collecting the original data of the steam turbine and the condenser, and dividing it into a historical data sequence, the steam turbine data and the condenser data of the node to be predicted; adopting an encoder-decoder framework and additionally proposing a connection module; in the encoder, using an LSTM network to encode the condenser working condition information and predict the vacuum degree, and encoding the vacuum degree information into the hidden state representation; in the connection module, capturing the local and global vacuum information of the encoder through the attention mechanism and CNN respectively; in the decoder, using the hidden state and cell state containing the vacuum degree information of the encoder for initialization, and using the local and global vacuum information and the working condition information of the steam turbine as the input of the decoder to encode and predict the output power of the steam turbine until the MSE loss function converges to complete the network training. Predicting the output power of the steam turbine through the trained network.

Description

Technical Field

[0001] The present invention relates to the technical field of operation optimization of thermal power plant equipment, and particularly relates to a method for predicting the output power of a steam turbine based on vacuum degree guidance. Background Art

[0002] The development of the economy requires the support of electricity, and a stable power supply is an important guarantee for effective production. By establishing a model to achieve power prediction for thermal power plants, it can provide reference for power plants when formulating power generation plans, and has high research value.

[0003] A complete thermal power generation system includes equipment such as circulating water pumps, condensers, and steam turbines. Among them, the steam turbine is the main power generation equipment, and the condenser is the main auxiliary equipment. The vacuum degree of the condenser directly reflects the operating state of the condenser and plays a very important role in ensuring the safe operation of the unit. Summary of the Invention

[0004] In view of the deficiencies in the prior art, the present invention provides a method for predicting the output power of a steam turbine based on vacuum degree guidance; by introducing the coupling relationship with the condenser to explore the prediction of the output power of the steam turbine, the natural medium for the coupling of the condenser and the steam turbine is the vacuum degree of the condenser (i.e., an index reflecting the working state of the condenser and also an important index for measuring the operation of the generating unit), which is a key factor for predicting the output power of the steam turbine; the prediction method provided by the present invention adopts an encoder-decoder framework, models the condenser and the steam turbine at the encoder and decoder respectively, and additionally proposes a connection module for modeling the coupling relationship between the condenser and the steam turbine. In the encoder part, the working condition data of each moment of the condenser is read in through a multi-layer long short-term memory network LSTM, and the vacuum degree at the moment to be predicted is predicted. In addition, a connection module composed of an attention mechanism and a convolutional neural network is proposed to capture the local vacuum information and global vacuum information of the encoder. In the decoder part, the local vacuum information, global vacuum information and the historical working condition data of the steam turbine are spliced as the input of another multi-layer long short-term memory network LSTM, and finally the output power at the moment to be predicted is predicted.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for predicting the output power of a steam turbine based on vacuum degree guidance, comprising the following steps:

[0007] S1. Collect several groups of original historical data of a thermal power plant, and perform preprocessing and normalization on each group of original historical data to obtain multiple groups of training data, and a condenser working condition data sequence X t-b+1:t and a steam turbine historical working condition data sequence Y t-b+1:t, the condenser vacuum data P of the predicted node t , the turbine output power data Q of the predicted node t ;

[0008] S2. Based on each set of training data in step S1 and the encoder-decoder framework, establish an encoder-decoder network; among them, establish a condenser model and a turbine model at the encoder and decoder respectively, and design a connection module to establish the coupling relationship between the condenser model and the turbine model; the encoder is used to dynamically predict the vacuum degree v at the target time of the condenser t ; the decoder is used to predict the turbine output power p at the target time t ; the connection module consists of an attention mechanism and a convolutional neural network CNN, which is used to capture the local and global vacuum information of the encoder, and then assist the above decoder to predict the turbine output power p at the target time t ;

[0009] Based on each set of training data in step S1, respectively use the predicted vacuum degree v t , the turbine output power p t to calculate the MSE loss function until convergence, and complete the end-to-end training of the encoder-decoder network;

[0010] S3. Real-time collect the operating condition data related to the condenser and the turbine, and through the trained encoder-decoder network, predict the vacuum degree of the condenser and the output power of the turbine at the target time node.

[0011] To optimize the above technical solution, the specific measures taken also include:

[0012] Furthermore, the specific content of preprocessing and normalizing each set of original historical data in step S1 is: for the original historical data, filter and delete the time points with outliers, and then according to the actual scenario of thermal power generation, select the characteristic data related to the operation of the condenser and the turbine, and for each characteristic data, use maximum or minimum normalization to achieve data standardization.

[0013] Furthermore, the specific content of establishing an encoder-decoder network based on each set of training data in step S1 and the encoder-decoder framework to predict the vacuum degree v at the target time of the condenser t and the turbine output power p at the target time t is as follows:

[0014] S211. Design both the encoder and the decoder to use an L-layer long short-term memory network LSTM as the basic framework;

[0015] S212. Input each group of condenser condition data sequences \(X\) t-b+1:t =\(\{x\) t-b+1 ,…, \(x\) i ,…, \(x\) t \}\) into the first - layer long short - term memory network LSTM of the encoder, so as to obtain the hidden representations corresponding to all moments of the condenser condition data sequence wherein, \(x\) i \(\in X\) t-b+1:t represents the condenser condition data sequence at the \(i\) - th time node, and \(b\) is the number of time nodes of the data sequence;

[0016] S213. Take the single - layer hidden state representation obtained in step S212 as the input of the next - layer long short - term memory network LSTM of the encoder, so as to obtain the hidden representations corresponding to all moments of the condenser condition data sequence Similarly, take the output of the upper - layer long short - term memory network LSTM of the encoder as the input of the next - layer long short - term memory network LSTM of the encoder until the hidden state representation of the data sequence and the cell state of the LSTM are output by the last layer of the encoder long short - term memory network LSTM and the cell state of the LSTM

[0017] Step S214. Pass the hidden state at time \(t\) of the encoder through a linear layer and a RELU activation function, and then obtain the predicted vacuum degree value \(v\) at time \(t\) t

[0018]

[0019] In the formula, \(W1\) and \(W2\) are both parameters for training;

[0020] Step S215. Through the connection module, use the internal attention mechanism and convolutional neural network CNN to capture the local vacuum degree information \(L\) and global vacuum degree information \(G\) in the condenser respectively;

[0021] For the local vacuum degree information \(L = [l1,\ldots,l\) i ,\ldots,l\) t , where the local feature at time step \(i\) is \(l\) i ; Determine the semantic correlation \(A\) between the decoder hidden state at time step \(i - 1\) e and \(H\) l by calculating the dot product:

[0022]

[0023] Therefore, the calculation method of the local feature \(l\) i at time step \(i\) is as follows:

[0024] l i = A l H e

[0025] Similarly, according to the above formula, the local vacuum degree information L = [l1,…,l i ,…,l t ;

[0026] For the global vacuum degree information G, in the convolutional neural network CNN, first use one-dimensional convolution to perform feature mapping on H e :

[0027]

[0028]

[0029] In the formula, s is the size of the filter, n is the step size of the convolution, w i represents the parameter of the i-th filter, f(.) represents the activation function, represents the vector obtained by concatenating the hidden state levels from time n to n + s - 1 in H e , represents the representation obtained by performing a convolution operation on by the i-th filter, represents the representation obtained by successively performing convolution operations on H e by the i-th filter after concatenation, and then, max(.) represents max pooling to reduce the dimension of the convolution output to obtain the final output o e of the i-th filter on H i . Connect the outputs of all filters to obtain the global feature G:

[0030] G = [o1, o2,…, o m

[0031] In the formula, o1 represents the output of the 1st filter on H e , o2 represents the output of the 2nd filter on H e , o m represents the output of the m-th filter on H e , and m represents the number of filters;

[0032] Step S216, for each group of steam turbine condition data sequences Y t-b+1:t = {y t-b+1 ,…, y i ,…, y t ​} Concatenate with the local vacuum degree information L and the global vacuum degree information G to obtain the input I = [L:Y:G], and input it into the first layer of the long short-term memory network LSTM of the decoder to obtain the hidden representations corresponding to all moments of the steam turbine operating condition data sequence where, y i ∈Y t-b+1:t represents the operating condition data of the steam turbine at the i-th time node, and b is the number of time nodes in the data sequence;

[0033] Step S217: Use the single-layer hidden state representation obtained in step S216 as the input of the next layer of the long short-term memory network LSTM of the decoder, so as to obtain the hidden representations corresponding to all moments of the steam turbine operating condition data sequence Similarly, use the output of the previous layer of the long short-term memory network LSTM of the decoder as the input of the next layer of the long short-term memory network LSTM of the decoder until the output of the last layer of the long short-term memory network LSTM of the decoder is the hidden state representation of the data sequence

[0034] Step S218: Pass the hidden state at the last moment of the decoder through the RELU activation function, and then obtain the predicted steam turbine output power magnitude p t :

[0035]

[0036] where, both W3 and W4 are parameters for training.

[0037] Furthermore, after steps S211 - S218, an encoder-decoder network is established to predict the vacuum degree magnitude v t and the steam turbine output power p t ; before applying the established encoder-decoder network, initialization and training are required. Therefore, based on each group of training data in step S1, the predicted vacuum degree magnitude v t and the steam turbine output power p t are respectively used to calculate the MSE loss function until convergence to complete the end-to-end training of the encoder-decoder network. The specific content is as follows:

[0038] Step S221: Use zero vectors as the initial hidden state and cell state respectively to initialize the encoder LSTM, where L represents the number of layers of the LSTM network, |x| represents the number of data in x, and d represents the size of the hidden state; Use the historical condenser operating condition data sequence As input, len represents the data contained in the data sequence, which passes through the L-layer LSTM network in sequence to obtain the hidden state representation H of the data sequence of the last layer e and the cell state C e , and the hidden state corresponding to time t passes through a linear layer and an activation function to obtain the predicted vacuum degree v at time t of the encoder-decoder framework model t ;

[0039] Step S222: Use the hidden state representation H at time t of the last layer of the encoder e and the cell state C e to initialize the decoder. Calculate the local vacuum degree information L and the global vacuum degree information G through a connection module, splice the working condition data sequence of the steam turbine with the local vacuum degree information L and the global vacuum degree information G to obtain the input I = [L:Y:G], and input it into the L-layer long short-term memory network LSTM of the decoder in sequence to obtain the hidden state representation H of the data sequence of the last layer d , and the hidden state at time t of the decoder passes through a linear layer and an activation function to obtain the predicted output power p of the steam turbine at time t of the encoder-decoder framework model t ;

[0040] Step S223: Calculate the MSE loss function for the predicted vacuum degree result v at the current time node t to be predicted t and the true vacuum degree P at time t of the target output in step S1 t , and calculate the MSE loss function for the predicted output power result p at time t t and the true steam turbine output power Q at time t of the target output in step S1 t . Repeat steps S221 - S223 for each set of original historical data in sequence until the MSE loss function converges, and complete the training of the end-to-end encoder-decoder framework network

[0041] Furthermore, the calculation formula of the MSE loss function described in step S223 is as follows

[0042]

[0043] In the formula, n represents the number of groups of original historical data, j represents the index of n, Y j represents the jth target output v t or p t , represents the vacuum degree P t or the output power Q t of the predicted result at the jth time node t to be predicted

[0044] The beneficial effects of the present invention are as follows: The steam turbine output power prediction method based on vacuum degree guidance of the present invention adopts an encoder-decoder framework to jointly model the condenser and the steam turbine. In the encoder part, the condition data sequence of the condenser is modeled, and the vacuum degree at time t is predicted. In order to model the coupling relationship between the steam turbine and the condenser, an additional connection module is proposed, which captures the local and global vacuum information of the encoder part by using the attention mechanism and CNN respectively. The hidden state and cell state containing the vacuum degree information in the encoder are used to initialize the decoder. In the decoder, the local vacuum degree information, the global vacuum degree information and the condition data of the steam turbine are spliced as the input of the decoder, and the output power at time t is predicted. Compared with some existing output power prediction methods, by encoding the historical information of the past condenser and steam turbine, combining the encoder-decoder framework and the long short-term memory network LSTM, the periodic characteristics and time series characteristics of the vacuum degree change hidden in the historical condition sequence and data of the condenser and steam turbine can be captured, and a connection module is proposed to model the coupling relationship between the condenser and the steam turbine. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a schematic flow chart of the steam turbine output power prediction method based on vacuum degree guidance of the present invention.

[0046] Figure 2 is a schematic diagram of the model structure of the steam turbine output power prediction method based on vacuum degree guidance of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0047] The present invention will be further described in detail below with reference to the accompanying drawings.

[0048] Refer to Figure 1 - Figure 2 . A steam turbine output power prediction method based on vacuum degree guidance includes the following steps:

[0049] S1. Collect several groups of original historical data of a thermal power plant, and perform preprocessing and normalization processing on each group of original historical data to obtain multiple groups of training data. Each group of training data is divided into a condenser condition data sequence X t-b+1:t , a steam turbine historical condition data sequence Y t-b+1:t , the condenser vacuum degree data P of the predicted node t , and the steam turbine output power data Q of the predicted node t ;

[0050] S2. Based on each set of training data in step S1 and the encoder-decoder framework, an encoder-decoder network is established. Among them, a condenser model and a steam turbine model are respectively established at the encoder and the decoder, and a connection module is designed to establish the coupling relationship between the condenser model and the steam turbine model. The encoder is used to dynamically predict the vacuum degree v of the condenser at the target moment t ; the decoder is used to predict the output power p of the steam turbine at the target moment t ; the connection module consists of an attention mechanism and a convolutional neural network CNN, and is used to capture the local vacuum information and global vacuum information of the encoder, so as to assist the above decoder to predict the output power p of the steam turbine at the target moment t ;

[0051] Based on each set of training data in step S1, the predicted vacuum degree v t and the output power p of the steam turbine t are respectively used to calculate the MSE loss function until convergence, and the end-to-end training of the encoder-decoder network is completed;

[0052] S3. The working condition data related to the condenser and the steam turbine are collected in real time, and through the trained encoder-decoder network, the vacuum degree of the condenser and the output power of the steam turbine at the target time node are predicted.

[0053] Furthermore, the specific content of preprocessing and normalizing each set of original historical data in step S1 is as follows: for the original historical data, the time points with outliers are filtered and deleted, and then according to the actual scenario of thermal power generation, the characteristic data related to the operation of the condenser and the steam turbine are selected, and for each characteristic data, maximum or minimum normalization is adopted to realize the standardization of the data.

[0054] Furthermore, the specific content of establishing the encoder-decoder network based on each set of training data in step S1 and the encoder-decoder framework to predict the vacuum degree v of the condenser at the target moment t and the output power p of the steam turbine at the target moment t is as follows:

[0055] S211. Both the encoder and the decoder are designed to use the L-layer long short-term memory network LSTM as the basic framework;

[0056] S212. Each set of condenser working condition data sequences X t-b+1:t = {x t-b+1 , …, x i , …, x t} are input into the first layer of the long short-term memory network LSTM of the encoder, so as to obtain the hidden representations corresponding to all moments of the condenser working condition data sequence where x i ∈ X t-b+1:t represents the condenser condition data sequence at the i-th time node, and b is the number of time nodes in the data sequence;

[0057] S213. Use the single-layer hidden state representation obtained in step S212 as the input of the long short-term memory network (LSTM) in the next layer of the encoder, so as to obtain the hidden representation corresponding to all moments of the condenser condition data sequence Similarly, use the output of the LSTM in the previous layer of the encoder as the input of the LSTM in the next layer of the encoder until the hidden state representation of the data sequence is output by the last layer of the encoder LSTM and the cell state of the LSTM

[0058] Step S214. Pass the hidden state at time t of the encoder through a linear layer and a ReLU activation function to obtain the predicted vacuum degree v at output time t t

[0059]

[0060] where W1 and W2 are both parameters for training;

[0061] Step S215. Through a connection module, use the internal attention mechanism and convolutional neural network (CNN) to capture the local vacuum degree information L and global vacuum degree information G in the condenser respectively;

[0062] For the local vacuum degree information L = [l1,..., l i ,..., l t , where the local feature at time step i is l i ; determine the semantic correlation A between the decoder hidden state at time step i - 1 e and H l by calculating the dot product:

[0063]

[0064] Therefore, the calculation method of the local feature l i at time step i is as follows:

[0065] l i = A l H e

[0066] Similarly, according to the above formula, the local vacuum degree information L = [l1,..., l i ,..., lt ;

[0067] For the global vacuum degree information G, first use one-dimensional convolution in the convolutional neural network CNN for H e to perform feature mapping:

[0068]

[0069]

[0070] In the formula, s is the size of the filter, n is the step of convolution, w i represents the parameter of the i-th filter, f(.) represents the activation function, represents H e The vector obtained by concatenating the hidden states at these moments from subscript n to n + s - 1 in it, represents the i-th filter for The representation obtained after performing the convolution operation, represents the i-th filter for H e The representation obtained after sequentially performing the convolution operation The vector obtained after concatenation. Then, max(.) represents max pooling to reduce the dimension of the convolution output to obtain the final output o e of the i-th filter for H i , and connect the outputs of all filters to obtain the global feature G:

[0071] G = [o1, o2, …, o m

[0072] In the formula, o1 represents the output of the 1st filter for H e , o2 represents the output of the 2nd filter for H e , o m represents the output of the m-th filter for H e , and m represents the number of filters;

[0073] Step S216, Concatenate each group of steam turbine condition data sequences Y t-b+1:t = {y t-b+1 , …, y i , …, y t} with the local vacuum degree information L and the global vacuum degree information G to obtain the input I = [L:Y:G], and input it into the first layer of the long short-term memory network LSTM of the decoder to obtain the hidden representations corresponding to all moments of the steam turbine condition data sequence Among them, y i ∈Y t-b+1:t represents the condition data of the steam turbine at the i-th time node, and b is the number of time nodes of the data sequence;​

[0074] Step S217: Use the single-layer hidden state representation obtained in step S216 as the input of the next layer of the decoder's long short-term memory network (LSTM), so as to obtain the hidden representation corresponding to all moments of the steam turbine operating condition data sequence Similarly, use the output of the previous layer of the decoder's long short-term memory network (LSTM) as the input of the next layer of the decoder's long short-term memory network (LSTM), until the hidden state representation of the data sequence is output by the last layer of the decoder's long short-term memory network (LSTM)

[0075] Step S218: Use the hidden state at the last moment of the decoder through the RELU activation function to obtain the predicted steam turbine output power magnitude p at output time t t :

[0076]

[0077] where both W3 and W4 are parameters for training.

[0078] Furthermore, after steps S211 - S218, an encoder-decoder network is established to predict the vacuum degree magnitude v t and the steam turbine output power p t ; before applying the established encoder-decoder network, initialization and training are required. Therefore, based on each set of training data in step S1, the predicted vacuum degree magnitude v t and the steam turbine output power p t are used to calculate the MSE loss function until convergence, and the end-to-end training of the encoder-decoder network is completed. The specific content is as follows:

[0079] Step S221: Use zero vectors to initialize the encoder LSTM as the initial hidden state and cell state respectively, where L represents the number of layers of the LSTM network, |x| represents the number of data in x, and d represents the size of the hidden state; use the historical condenser operating condition data sequence as the input, len represents the data included in this data sequence, and pass through the L-layer LSTM network in sequence to obtain the hidden state representation H e and the cell state C e , and pass the hidden state corresponding to time t through a linear layer and an activation function to obtain the predicted vacuum degree v of the encoder-decoder framework model at time t t ;

[0080] Step S222: Use the hidden state representation H of the encoder at time t e and the cell state C e to initialize the decoder. Calculate the local condenser vacuum degree information L and the global vacuum degree information G through the connection module, and splice the operating condition data sequence of the steam turbine with the local vacuum degree information L and the global vacuum degree information G to obtain the input I = [L:Y:G]. Then, input it into the L-layer long short-term memory network (LSTM) of the decoder in sequence to obtain the hidden state representation H of the data sequence of the last layer d . Use the hidden state of the decoder at time t, through the linear layer and the activation function, to obtain the predicted output power p of the steam turbine at time t by the encoder-decoder framework model t ;

[0081] Step S223: Calculate the MSE loss function for the vacuum degree prediction result v at the current time node t to be predicted t and the true vacuum degree P at time t of the target output in step S1 t . Also, calculate the MSE loss function for the output power prediction result p t at time t and the true steam turbine output power Q at time t of the target output in step S1 t . Repeat steps S221 - S223 for each set of original historical data in sequence until the MSE loss function converges, completing the training of the end-to-end encoder-decoder framework network.

[0082] Furthermore, the calculation formula of the MSE loss function in step S223 is as follows:

[0083]

[0084] In the formula, n represents the number of groups of original historical data, j represents the index of n, and Y j represents the j-th target output v t or p t , represents the predicted result of the vacuum degree P t or the output power Q t at the j-th time node t to be predicted.

[0085] The steam turbine output power prediction method of the present invention performs the prediction of the steam turbine output power through an encoder-decoder framework. In the encoder part, the condenser is modeled. The long short-term memory network (LSTM) is used to read the operating condition data of the condenser at each moment and predict the vacuum degree at the moment to be predicted. In the decoder part, the steam turbine is modeled. The hidden state representations at each moment containing the vacuum degree information in the encoder are concatenated with the operating condition data of the steam turbine at each moment as the input of the decoder, and finally the output power at the moment to be predicted is predicted. Compared with the method of only considering the steam turbine part for power prediction, the method proposed by the present invention can take into account the vacuum degree information of the condenser and has higher prediction accuracy and stability.

[0086] In a specific embodiment of the present application, the original data of a power plant is collected and classified. The number of layers L of the long short-term memory network (LSTM) is set to 2, the encoding vector length len of the historical data sequence is set to 64, and the length of the historical data sequence is 10. The historical data is encoded with the above-set parameters, and the encoder-decoder network is trained end-to-end using the Adam optimizer with a learning rate of 1e-3. The output power of the steam turbine is predicted through the trained end-to-end encoder-decoder network. The mean absolute error (MAE) and root mean square error (RMSE) are used as evaluation indicators. The lower the indicator, the better the prediction effect. In this scheme, the MAE is 1.3635 and the RMSE is 1.7597, which are better than the method of using the LSTM network to separately model the steam turbine in the baseline model, where its MAE is 2.1656 and the RMSE is 2.5947. The steam turbine output power prediction method based on vacuum degree guidance of the present invention can effectively improve the prediction accuracy of the steam turbine output power and greatly improve the prediction stability.

[0087] It should be noted that the terms such as "up", "down", "left", "right", "front", "back", etc. cited in the invention are only for the convenience of clear narration and are not used to limit the scope of implementation of the present invention. The change or adjustment of their relative relationship, without substantial change in the technical content, should also be regarded as the scope of implementation of the present invention.

[0088] The above is only the preferred embodiment of the present invention. The protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the idea of the present invention belong to the protection scope of the present invention. It should be pointed out that for those of ordinary skill in the art in this technical field, several improvements and refinements made without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A method for predicting the output power of a steam turbine based on vacuum degree guidance, characterized in that It includes the following steps: S1. Collect several groups of original historical data of the thermal power plant, and perform preprocessing and normalization on each group of original historical data to obtain multiple groups of training data. And in each group of training data, there are sequences of condenser operating conditions , sequences of steam turbine operating conditions , predict the vacuum degree of the condenser at the target moment , the output power of the steam turbine at the target moment ; S2. Based on each set of training data in step S1 and the encoder-decoder framework, an encoder-decoder network is established; wherein, a condenser model and a turbine model are established at the encoder and decoder respectively, and a connection module is designed to establish a coupling relationship between the condenser model and the turbine model; the encoder is used to dynamically predict the vacuum degree of the condenser at the target time , with a zero vector 、 As the initial hidden state and cell state, the encoder LSTM is initialized, where Represents the number of layers of the LSTM network, express The number of data in Represents the size of the hidden state; the historical condenser operating data sequence As input, Indicates the data contained in the data sequence, which is passed in sequence Layer LSTM network, get the hidden state representation of the last layer of data sequence and cell status ,Will The hidden state corresponding to the moment Through the linear layer and activation function, the encoder-decoder framework model is obtained. The vacuum degree of the prediction at the moment ; The decoder is used to predict the output power of the steam turbine at the target time ; The connection module consists of an attention mechanism and a convolutional neural network CNN, and is used to capture the local vacuum information and global vacuum information of the encoder, so as to assist the above decoder to predict the output power of the steam turbine at the target time ; Based on each set of training data in step S1, respectively use the predicted vacuum degree magnitude , the output power of the steam turbine to calculate the MSE loss function until convergence, and complete the end-to-end training for the encoder-decoder network; S3. Real-time collect the operating condition data related to the condenser and the steam turbine, and through the trained encoder-decoder network, predict the vacuum degree of the condenser and the output power of the steam turbine at the target time node.

2. The steam turbine output power prediction method based on vacuum degree guidance according to claim 1, characterized in that The specific content of preprocessing and normalizing each group of original historical data described in step S1 is as follows: for the original historical data, filter and delete the time points with outliers, and then according to the actual scenario of thermal power generation, select the characteristic data related to the operation of the condenser and the steam turbine, and for each characteristic data, adopt maximum or minimum normalization to achieve data standardization.

3. A method for predicting the output power of a steam turbine based on vacuum degree guidance according to claim 1, characterized in that, Based on each set of training data in step S1 and the encoder-decoder framework, establish an encoder-decoder network in step S2 to predict the vacuum degree of the condenser at the target moment and the output power of the steam turbine at the target moment The specific content is as follows: S211. Design both the encoder and the decoder to adopt an LSTM (Long Short-Term Memory) network with S212. Input each group of condenser condition data sequences into the first layer of long short-term memory network (LSTM) of the encoder to obtain the hidden representations corresponding to all time moments of the condenser condition data sequences , where represents the condenser condition data sequence at the -th time node, and is the number of time nodes of the data sequence; S213. Use the single-layer hidden state representation obtained in step S212 as the input of the next layer of the encoder's long short-term memory network (LSTM), so as to obtain the hidden representations corresponding to all moments of the condenser operating condition data sequence . Similarly, use the output of the previous layer of the encoder's long short-term memory network (LSTM) as the input of the next layer of the encoder's long short-term memory network (LSTM), until the hidden state representation of the data sequence is output by the last layer of the encoder's long short-term memory network (LSTM) and the cell state of the LSTM ; Step S214: The hidden state at the encoder time is passed through a linear layer and a ReLU activation function to obtain the predicted vacuum degree at the output time ​ wherein, and are both parameters for training; Step S215: Through the connection module, use its internal attention mechanism and convolutional neural network (CNN) to capture the local vacuum information and global vacuum information in the condenser respectively and global vacuum information ; Regarding local vacuum degree information , where the local feature of time step is ; Determine the semantic correlation between the decoder hidden state of time step and by calculating the dot product : Therefore, the time step of the local feature is calculated as follows: Similarly, according to the above formula, the local vacuum degree information is obtained ; Regarding the global vacuum degree information , in the convolutional neural network CNN, first use the one-dimensional convolution method to perform feature mapping: In the formula, is the size of the filter, is the stride of the convolution, represents the parameters of the th filter, denotes the activation function, in the subscript from to the vector obtained by concatenating the hidden state at these moments, represents the th filter's operation on which is the representation obtained after successive convolution operations, then, after concatenation, the resulting vector, and then, represents max pooling to reduce the dimension of the convolution output to obtain the th filter's final output , and connecting the outputs of all filters to obtain the global feature : Wherein, represents the output of the first filter for , represents the output of the second filter for , represents the output of the th filter for , represents the number of filters; Step S216: Concatenate each group of steam turbine operating condition data sequences with the local vacuum degree information and the global vacuum degree information to obtain the input , and input it into the first layer of the long short-term memory network LSTM of the decoder to obtain the hidden representations corresponding to all moments of the steam turbine operating condition data sequences ; where represents the operating condition data of the steam turbine at the i th time node, b is the number of time nodes of the data sequence Step S217: Use the single-layer hidden state representation obtained in step S216 as the input to the next layer of the decoder's long short-term memory network (LSTM), so as to obtain the hidden representations corresponding to all moments of the steam turbine operating condition data sequence Similarly, use the output of the previous layer of the decoder's long short-term memory network (LSTM) as the input to the next layer of the decoder's long short-term memory network (LSTM), until the hidden state representation of the data sequence is output from the last layer of the decoder's long short-term memory network (LSTM) ; Step S218: Take the hidden state at the last moment of the decoder and pass it through the ReLU activation function to obtain the output moment for predicting the output power of the steam turbine : Among them, and are all parameters for training.

4. A method for predicting the output power of a steam turbine based on vacuum degree guidance according to claim 3, characterized in that, After steps S211 - S218 are implemented to establish an encoder - decoder network for predicting the vacuum degree magnitude and the steam turbine output power ; before applying the established encoder - decoder network, initialization and training are required. Therefore, based on each set of training data in step S1, the predicted vacuum degree magnitude , the steam turbine output power are used to calculate the MSE loss function until convergence, and the specific content of the end - to - end training for the encoder - decoder network is as follows: The hidden state representation at the last moment of the encoder is used to initialize the decoder. Calculate the local condenser vacuum degree information through the connection module and the cell state to calculate the local condenser vacuum degree information and the global vacuum degree information , and splice the operating condition data sequence of the steam turbine with the local vacuum degree information , the global vacuum degree information to obtain the input , and input it into the layer long short-term memory network LSTM of the decoder in sequence to obtain the hidden state representation of the data sequence at the last layer . Pass the hidden state of the decoder at the moment through a linear layer and an activation function to obtain the predicted output power of the steam turbine by the encoder-decoder framework model at the moment ; The current time node to be predicted Vacuum degree prediction result And the target output in step S1 True vacuum degree at the moment Calculate the MSE loss function and Output power prediction result at the moment And the target output in step S1 True steam turbine output power at the moment Calculate the MSE loss function, and repeat steps S221 - S223 for each set of original historical data in turn until the MSE loss function converges, and complete the training of the end-to-end encoder-decoder framework network.

5. The turbine output power prediction method based on vacuum degree guidance according to claim 4, characterized in that The calculation formula of the MSE loss function is as follows: In the formula, represents the number of groups of original historical data, represents the index of represents the th target output or , represents the th time node to be predicted vacuum degree or output power prediction result.

Images