Thermal power plant carbon dioxide control system and method based on double-neural-network adaptive feedback

Abstract

The invention discloses a double neural network adaptive feedback thermal power plant carbon dioxide control system and method. The system comprises a data input unit, a GRU unit, an LSTM unit, an error feedback unit and an output control unit. The data input unit inputs data; the GRU controls information flow of input data of the data input unit through a gating mechanism, and a reset gate affects a hidden state vector ht-1 at the previous moment to the extent that the candidate hidden state at the current moment is affected; the LSTM unit performs linear transformation on the input data to generate a final carbon dioxide predicted value of the thermal power plant; an error feedback unit adjusts weight parameters according to the error performance of the GRU unit and the LSTM unit after each iteration; and the output control unit outputs the final carbon dioxide prediction value of the thermal power plant, provides operation suggestions for a regulation and control system of the thermal power plant according to the prediction result, and is connected with the real-time regulation and control module to realize automatic regulation of operation parameters of the thermal power plant. According to the method, the accuracy and the self-adaptive capability of carbon dioxide emission prediction are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of thermal power generation, and particularly relates to a double neural network adaptive feedback carbon dioxide control system and method for a thermal power plant. BACKGROUND

[0002] With the increasing demand for carbon emission reduction and environmental protection worldwide, the monitoring and prediction technology of carbon dioxide emissions in thermal power plants has gradually become an important means to improve the operation efficiency of power plants and reduce carbon emissions. However, the existing carbon dioxide monitoring and prediction systems still face many challenges in practical application. Most of the current carbon dioxide monitoring systems in thermal power plants rely on a single data source and traditional algorithms to predict emissions, but such systems have weak response capability to complex working condition changes, limited prediction accuracy, and are difficult to adapt to the dynamic load conditions and variable combustion process of thermal power plants. The existing technology mostly uses simple linear regression models or traditional time series algorithms for carbon dioxide prediction. Although these methods can provide certain prediction effect in stable environments, they have greatly reduced prediction performance when facing complex, nonlinear, and multi-time-varying combustion environments, and are difficult to capture the influence of multiple factors such as fuel composition, combustion efficiency, and equipment state on carbon dioxide emissions. In addition, the model update frequency of traditional systems is low, and it is difficult to adapt to changes under different operating conditions, especially when the load fluctuates and the fuel type changes, it is difficult to provide high-precision real-time prediction, affecting the optimal adjustment of power plant operating parameters.

[0003] Although there are some improved carbon dioxide monitoring systems on the market that attempt to integrate more data sources and simple control algorithms, these systems still mainly rely on rule-driven or static models based on historical data, and fail to effectively utilize real-time data and working condition information for adaptive adjustment, resulting in large prediction errors and difficulty in maintaining high prediction performance under variable working conditions. Therefore, the existing carbon dioxide monitoring and prediction systems urgently need to introduce more intelligent solutions to better cope with the complexity and dynamics of the operating environment of thermal power plants.

[0004] The current carbon dioxide monitoring and prediction system of thermal power plants mostly relies on a single data source and traditional algorithms, lacks intelligent prediction and control means, and is difficult to achieve accurate response to dynamic working condition changes. In the existing technology, linear regression models or simple time series analysis are mainly used for carbon dioxide emission prediction. These methods cannot fully utilize real-time data and multi-dimensional feature information, resulting in low prediction accuracy and slow response when facing complex combustion environments. In addition, the model updating frequency of existing systems is low, and they lack self-adaptive ability, especially when the load of the thermal power plant fluctuates and the fuel composition changes, they cannot provide accurate prediction. This makes the monitoring and control of carbon dioxide emissions not flexible enough, directly affecting the environmental protection and energy efficiency performance of the power plant. At the same time, traditional technologies usually cannot effectively integrate multiple data sources such as equipment status, combustion efficiency and external environmental factors, resulting in a lack of comprehensive analysis capability and inability to achieve efficient multi-dimensional optimization prediction. These deficiencies limit the reliability and effectiveness of existing technologies in practical applications, especially when large-scale deployment is required, which shows the problem of insufficient scalability. SUMMARY

[0005] Based on the deficiencies of the existing technology, the present application provides a double neural network adaptive feedback carbon dioxide control system and method for thermal power plants. The present application fully combines the advantages of GRU in short-term memory and computational efficiency and the ability of LSTM in capturing long-term dependencies to improve the accuracy and self-adaptive ability of carbon dioxide emission prediction. By integrating information from multiple real-time data sources such as fuel composition, load changes and environmental conditions, the system realizes the fusion of multi-dimensional data, enhancing the perception and response ability to complex combustion environments. The adaptive feedback mechanism introduced by the system enables the model to dynamically adjust parameters based on real-time monitoring of carbon dioxide concentration and prediction error, thereby optimizing prediction results and plant operating parameters in real time, achieving accurate control and effective management of carbon dioxide emissions. This innovative solution not only improves the environmental performance of thermal power plants, but also provides a practical solution for achieving efficient energy management and sustainable development goals, helping the thermal power industry to transform towards low carbonization.

[0006] In order to achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows:

[0007] The double neural network adaptive feedback carbon dioxide control system for thermal power plants comprises a data input unit, a GRU unit, an LSTM unit, an error feedback unit and an output control unit.

[0008] The data input unit is used for inputting data, including a current time input vector x t , representing the previous time hidden state vector h t-1 , representing the hidden information stored at the previous time.

[0009] The GRU unit is used for controlling the flow of information through the gating mechanism of the input data of the data input unit, and the reset gate affects the hidden state vector h of the previous moment t-1 Influence the candidate hidden state of the current moment to what extent;

[0010] The LSTM unit is used for generating a final carbon dioxide prediction value of the thermal power plant through linear transformation of the input data;

[0011] The error feedback unit is used for adjusting the weight parameters according to the error performance of the GRU unit and the LSTM unit after each iteration, so that the smaller error occupies a larger proportion in future prediction, and the prediction accuracy is improved;

[0012] The output control unit is used for outputting the final carbon dioxide prediction value of the thermal power plant, and providing operation suggestions to the regulation system of the thermal power plant according to the prediction result, and is connected with the real-time regulation module to realize automatic adjustment of the operation parameters of the thermal power plant.

[0013] Further improvement of the application is that the GRU unit controls the flow of information through the gating mechanism of the input data of the data input unit, and the reset gate affects the hidden state vector h of the previous moment t-1 Influence the candidate hidden state of the current moment to what extent, the calculation formula is as follows:

[0014] r t =σ*(W r *[h t-1 ,x t ]+b r ) (1)

[0015] Wherein, W r and b r are the reset gate weight matrix and the bias term, sigma is the sigmoid activation function, r t is the reset gate output, and the value range is between [0, 1], when r t approaches 0, the reset gate ignores the influence of the previous hidden state; when r t approaches 1, the historical information is retained;

[0016] The update gate decides how much of the hidden state h t of the previous moment is retained h t-1 , that is, the information retention degree, and the calculation formula is as follows:

[0017] z t =σ*(W z *[h t-1 ,x t ]+b z ) (2)

[0018] Wherein, Wz and b z is to update the gate weight matrix and bias term, z t is the update gate output; after the reset gate and the update gate are calculated, the input x of the current moment is combined t and part of the hidden state h of the previous moment is controlled by the reset gate t-1 The candidate hidden state is calculated, and the calculation formula is as follows:

[0019]

[0020] wherein, is the candidate hidden state, which represents the current moment hidden information that the GRU may update; the final output hidden state h t is whether to retain the old hidden state or accept the new hidden state through the update gate on this basis; W h is the trainable weight matrix of the candidate hidden state; tanh is the hyperbolic tangent activation function, which compresses the linearly transformed result to the range of [-1, 1];

[0021] The final hidden state is combined according to the update gate z t output value, the hidden state h t-1 of the previous moment is linearly combined with the candidate hidden state of the current moment , and the formula is as follows:

[0022]

[0023] When z t is close to 1, it means that more information of the previous moment is retained at the current moment; when z t is close to 0, it means that more information of the current moment is used, and the final GRU unit output is the final hidden state.

[0024] The further improvement of the present application is that the update gate output is between [0, 1].

[0025] The further improvement of the present application is that the LSTM unit comprises an input gate, a forget gate, an output gate, a cell state update and an output layer.

[0026] The input gate determines how much part of the input information xt of the current moment can enter the cell state of the current moment, and the calculation formula is:

[0027] i t =σ*(W i *x t +U i *h t-1 +b i ) (5)

[0028] wherein, i tis the activation value of the forget gate, indicating the importance of the previous cell state in the current update; W i is the weight matrix of the forget gate, mapping the input x t ; U i is the weight matrix of the forget gate, mapping the hidden state h t-1 of the previous time; b t is the bias term of the forget gate;

[0029] The forget gate is used to determine the part of the current time cell state C t-1 that needs to be forgotten, that is, whether to retain some historical information at the current time, and its calculation formula is as follows

[0030] f t = σ*(W f *x t + U f *h t-1 + b f ) (6)

[0031] Where f t is the activation value of the forget gate, indicating the importance of the previous cell state in the current update; W f is the weight matrix of the forget gate, mapping the input x t ; U f is the weight matrix of the forget gate, mapping the hidden state h t-1 of the previous time; b f is the bias term of the forget gate;

[0032] The output gate is used to determine the part of the current time cell state C t as the hidden state h t output, and is used for next time calculation, and the calculation formula is:

[0033] O t = σ*(W O *x t + U O *h t-1 + b O ) (7)

[0034] Where O t is the activation value of the output gate, controlling the output amount from the cell state to the hidden state; W o is the weight matrix of the output gate, mapping the current input x t ; U o is the weight matrix of the output gate, mapping the hidden state h t-1 of the previous time; b o is the bias term of the output gate;

[0035] The cell state update and the cell state update calculation formula of the output layer are as follows:

[0036] C t = f t *C t-1 +i t *C t (8)

[0037] h t = O t *tanh(C t +b c ) (9)

[0038] Wherein, C t is the current moment cell state, fusion past and current information; f t *C t-1 The last moment cell state after the forgetting gate regulation information amount;I t *C t The current input information C t After the input gate regulation new information is added;B c It is used for the bias term of candidate cell state calculation.

[0039] The further improvement of the application is that the hidden state h t The final thermal power plant carbon dioxide prediction value is generated by linear transformation:

[0040] Y pred =W out *h t +b out (10)

[0041] Wherein, W out The weight of the output layer, b out Bias term, after calculation, get LSTM prediction value.

[0042] The further improvement of the application is that the activation value i t Normalized to 0 and 1 by Sigmoid activation function.

[0043] The double neural network adaptive feedback thermal power plant carbon dioxide control method comprises:

[0044] Data input unit input data, including current input vector x t , indicates the current moment external input data hidden state vector h t-1 , indicates the last moment stored implicit information;

[0045] The input data of the data input unit is controlled by the gating mechanism of the GRU unit to control the flow of information, and the reset gate affects the last moment hidden state vector h t-1To what extent does the candidate hidden state at the current time point affect;

[0046] The LSTM unit generates the final carbon dioxide prediction value of the thermal power plant through linear transformation of the input data;

[0047] The error feedback unit adjusts the weight parameters according to the error performance of the GRU unit and the LSTM unit after each iteration, so that the smaller error occupies a larger proportion in future prediction, thereby improving the prediction accuracy;

[0048] The output control unit outputs the final carbon dioxide prediction value of the thermal power plant, and provides operation suggestions to the regulation system of the thermal power plant according to the prediction result, and is connected with the real-time regulation module to realize automatic adjustment of the operation parameters of the thermal power plant.

[0049] Further improvement of the present application is that the GRU unit controls the flow of information through the gating mechanism of the input data of the data input unit, and the reset gate affects the hidden state vector h t-1 To what extent does the candidate hidden state at the current time point affect, the calculation formula is as follows:

[0050] r t =σ*(W r *[h t-1 ,x t ]+b r ) (1)

[0051] Wherein, W r and b r are the reset gate weight matrix and bias term, sigma is the sigmoid activation function, r t is the reset gate output, and the value range is between [0, 1], when r t is close to 0, the reset gate ignores the influence of the previous hidden state; when r t is close to 1, the historical information is retained;

[0052] The update gate determines how much of the previous hidden state h t is retained in the current hidden state h t-1 , that is, the information retention degree, and the calculation formula is:

[0053] z t =σ*(W z *[h t-1 ,x t ]+b z ) (2)

[0054] Wherein, W z and b z are the update gate weight matrix and bias term, z t is the update gate output; after the reset gate and the update gate are calculated, the current input xt and part of the reset gate controls the hidden state h of the last time t-1 The candidate hidden state is calculated, and the calculation formula is as follows:

[0055]

[0056] wherein, is the candidate hidden state, representing the current time hidden information that GRU can update; the final output hidden state h t is the update gate that controls whether to retain the old hidden state or accept the new hidden state on this basis; W h is the trainable weight matrix of the candidate hidden state; tanh is the hyperbolic tangent activation function, which compresses the linearly transformed result to the range of [-1, 1];

[0057] The final hidden state is updated according to the update gate z t The output value is linearly combined with the hidden state h t-1 of the last time and the candidate hidden state of the current time, and the formula is as follows:

[0058]

[0059] When z t is close to 1, it means that more information of the last time is retained at the current time; when z t is close to 0, it means that more information of the current time is used, and the final GRU unit output is the final hidden state.

[0060] The further improvement of the application is that the update gate output is between [0, 1].

[0061] The further improvement of the application is that the LSTM unit includes an input gate, a forget gate, an output gate, a cell state update and an output layer.

[0062] The input gate determines how much of the current input information xt can enter the cell state at the current time, and the calculation formula is:

[0063] i t =σ*(W i *x t +U i *h t-1 +b i ) (5)

[0064] wherein, i t is the activation value, indicating the importance of the current input in updating the cell state; W i is a weight matrix that maps the input x t to the hidden layer; U iThe weight matrix maps the hidden state h from the previous time step. t-1 b t It is a bias term;

[0065] The forget gate is used to determine the current cell state C. t-1 The part that needs to be forgotten, i.e. whether certain historical information should be retained at the current moment, is calculated using the following formula:

[0066] f t =σ*(W f *x t +U f *h t-1 +b f (6)

[0067] Among them, f t W is the activation value of the forget gate, representing the importance of the cell state from the previous time step in the current time step's update; f It is the weight matrix of the forget gate, mapping to the input x. t ;U f It is the weight matrix of the forget gate, which maps to the hidden state h in the previous time step. t-1 b f It is the bias term of the forget gate;

[0068] The output gate is used to determine the cell state C at the current moment. t h is the hidden state t The output is used for calculation in the next time step, and the calculation formula is:

[0069] O t =σ*(W O *x t +U O *h t-1 +b O (7)

[0070] Among them, O t It is the activation value of the output gate, controlling the output amount from the cell state to the hidden state; W o It is the weight matrix of the output gate, mapping to the current input x. t ;U o It is the weight matrix of the output gate, which maps to the hidden state h of the previous time step. t-1 b o It is the output gate bias term;

[0071] The formulas for calculating cell state updates and output layer cell state updates are as follows:

[0072] C t =f t *C t-1 +i t *Ct (8)

[0073] h t =O t *tanh(C t +b c )(9)

[0074] where C t is the current cell state, fusing information from past and current time; f t *C t-1 represents the information amount of the cell state at the last time step after being adjusted by the forget gate; i t *C t represents the new information added after the current input information C t is adjusted by the input gate; b c is a bias term for candidate cell state calculation.

[0075] Compared with the prior art, the present application has at least the following beneficial technical effects:

[0076] The present application proposes a double neural network adaptive feedback carbon dioxide control system and method for thermal power plants, which has significant technical advantages. By combining the characteristics of GRU and LSTM networks, the system can efficiently process complex carbon dioxide emission data of thermal power plants and accurately predict future emission trends. The efficient calculation of GRU units and the long-term dependence capturing ability of LSTM units enable the system to fully adapt to the dynamic changes of the combustion environment and effectively fuse and analyze multi-source data such as real-time load fluctuations and fuel composition changes, thereby improving prediction accuracy and system response capability. The adaptive feedback mechanism in the present application can dynamically adjust model parameters according to real-time monitoring of carbon dioxide concentration and prediction error, thereby optimizing thermal power plant operation strategies, reducing excessive emissions, and improving overall environmental performance. Through this intelligent monitoring and prediction mechanism, the system not only can accurately monitor the current emission situation, but also can make pre-adjustment according to future trends to ensure that the power plant emissions meet environmental standards and reduce the risk of exceeding standards.

[0077] The present application solves the problems of insufficient prediction accuracy and limited adaptive ability of existing technologies in handling complex environments and dynamic power demand, significantly improves the accuracy and efficiency of carbon dioxide emission monitoring and prediction, provides reliable support for more environmentally friendly and intelligent energy management, and has good market application prospect and promotion value. BRIEF DESCRIPTION OF DRAWINGS

[0078] In order to more clearly illustrate the technical solutions in the specific embodiments of the present application or the prior art, the accompanying drawings needed to be used in the specific embodiments or prior art description will be briefly introduced. Obviously, the accompanying drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative effort.

[0079] Figure 1 The GRU unit operation flow chart of the double neural network adaptive feedback carbon dioxide control system of the power plant of the present application.

[0080] Figure 2 The LSTM unit operation flow chart of the double neural network adaptive feedback carbon dioxide control system of the power plant of the present application.

[0081] Figure 3 The system flow chart of the double neural network adaptive feedback carbon dioxide control system of the power plant of the present application. DETAILED DESCRIPTION

[0082] In the following, only some exemplary embodiments are simply described. As those skilled in the art can recognize, the described embodiments can be modified in various different ways without departing from the spirit or scope of the present application. Therefore, the drawings and the description are considered to be exemplary in nature rather than limiting.

[0083] In the description of the present application, it should be understood that, when used in the specification and the appended claims, the terms "comprise" and "include" indicate the presence of described features, whole, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, whole, steps, operations, elements, components and / or sets thereof.

[0084] It should also be understood that the terms used in the present application specification are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the present application specification and the appended claims, the singular forms "a", "an" and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0085] It should be further understood that the term "and / or" used in the present application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes these combinations.

[0086] The various structural diagrams according to the disclosed embodiments of the present application are shown in the drawings. These diagrams are not drawn to scale, in which certain details are exaggerated for clarity and others omitted. The shapes and relative sizes of the various regions, layers, and their relative positions illustrated in the drawings are merely exemplary and may deviate in actuality due to manufacturing tolerances or technical limitations, and regions / layers with different shapes, sizes, and relative positions can be additionally designed according to actual needs by those skilled in the art.

[0087] The embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0088] Embodiment 1

[0089] The double neural network adaptive feedback carbon dioxide control system for thermal power plants provided by the present application comprises a data input unit, a GRU unit, an LSTM unit, an error feedback unit, and an output control unit.

[0090] The data input unit is configured to input data, including a current time input vector x t , representing the previous time hidden state vector h t-1 , representing the hidden information stored at the previous time.

[0091] The GRU unit is configured to control the flow of information through the gating mechanism of the input data of the data input unit, and the reset gate affects the previous time hidden state vector h t-1 to what extent it affects the candidate hidden state at the current time.

[0092] The LSTM unit is configured to generate a final carbon dioxide prediction value for the thermal power plant by linearly transforming the input data.

[0093] The error feedback unit is configured to adjust the weight parameters according to the error performance of the GRU unit and the LSTM unit after each iteration, so that the smaller error occupies a larger proportion in future prediction, thereby improving the prediction accuracy.

[0094] The output control unit is configured to output the final carbon dioxide prediction value for the thermal power plant, and provide operation suggestions to the regulation system of the thermal power plant according to the prediction results, and is connected with a real-time regulation module to realize automatic adjustment of the operating parameters of the thermal power plant.

[0095] In this embodiment, please refer to Figure 1 , in which the GRU (Gated Recurrent Unit) processes time series data through a gating mechanism, and the GRU unit input includes a current time input vector x t , representing the current time external input data (carbon dioxide concentration of the thermal power plant); a previous time hidden state vector h t-1, which combines the features of previous time steps. According to the input data, the GRU unit controls the flow of information through a gating mechanism, the reset gate affects the previous hidden state vector h t-1 to the extent to which the candidate hidden state of the current moment is affected, the calculation formula is as follows:

[0096] r t =σ*(W r *[h t-1 ,x t ]+b r ) (1)

[0097] where W r and b r are the reset gate weight matrix and bias term, σ is the sigmoid activation function, r t is the reset gate output, and the value range is between [0, 1], when r t is close to 0, the reset gate ignores the influence of the previous hidden state; when r t is close to 1, more historical information is retained.

[0098] The update gate determines how much of the previous hidden state h t-1 should be retained in the current hidden state h t , i.e. the degree of information retention. The calculation formula is as follows:

[0099] z t =σ*(W z *[h t-1 ,x t ]+b z ) (2)

[0100] where W z and b z are the update gate weight matrix and bias term, and z t is the update gate output. Similar to the reset gate, the update gate output is also between [0, 1]. After the reset gate and the update gate are calculated, the candidate hidden state can be calculated by combining the current input x t and part of the previous hidden state h t-1 controlled by the reset gate, and the calculation formula is as follows:

[0101]

[0102] where h h is the candidate hidden state, representing the current hidden information that GRU may update (including historical information and current input features). The final output hidden state h t is controlled by the update gate to determine whether to retain the old hidden state or accept the new hidden state; Wh is the trainable weight matrix of the candidate hidden state; tanh is the hyperbolic tangent activation function, which compresses the result after linear transformation to the range of [-1,1], enabling the model to better handle complex nonlinear relationships.

[0103] The final hidden state depends on the update gate z t The output value is the hidden state h from the previous time step. t-1 With the current candidate hidden state The formula for linear combination is as follows:

[0104]

[0105] When z t When z approaches 1, it indicates that the current time step retains more information from the previous time step; when z approaches 1... t When the value is close to 0, it indicates that more information at the current moment is used, and the final output of the GRU unit is the final hidden state.

[0106] In this embodiment, please refer to Figure 2 An LSTM unit includes an input gate, a forget gate, an output gate, a cell state update layer, and an output layer. The main function of the input gate is to determine how much of the input information xt at the current time step can enter the cell state at the current time step. The calculation formula is:

[0107] i t =σ*(W i *x t +U i *h t-1 +b i (5)

[0108] Among them, i t The activation value represents the importance of the current input in updating the cell state. Its value is normalized to between 0 and 1 using the Sigmoid activation function; W i Multimatrix mapping input x t To the hidden layer; U i The weight matrix maps the hidden state h from the previous time step. t-1 b t It is a bias term that helps adjust the linear transformation of the input.

[0109] The forgetting gate determines the current cell state C. t-1 The part that needs to be forgotten, i.e. whether certain historical information should be retained at the current moment, is calculated using the following formula:

[0110] f t =σ*(W f *x t +U f *h t-1+b f ) (6)

[0111] where f t is the activation value of the forget gate, indicating the importance of the cell state at the previous time in the update at the current time; W f is the weight matrix of the forget gate, mapping the input x t ; U f is the weight matrix of the forget gate, mapping the hidden state h t-1 at the previous time; b f is the bias term of the forget gate, adjusting the linear transformation of the input.

[0112] The output gate determines the hidden state h t in the cell state C t at the current time as the output part, and is used for the next time calculation, and the calculation formula is:

[0113] O t = σ*(W O *x t + U O *h t-1 + b O ) (7)

[0114] where O t is the activation value of the output gate, controlling the output amount from the cell state to the hidden state; W o is the weight matrix of the output gate, mapping the current input x t ; U o is the weight matrix of the output gate, mapping the hidden state h t-1 at the previous time; b o is the bias term of the output gate, adjusting the linear transformation of the input.

[0115] The key of the LSTM unit is the cell state C t , which determines whether the long-term memory is retained, and the cell state update is jointly determined by the input gate and the forget gate, which determines the retention and update of the cell state at the current time. The cell state update calculation formula is:

[0116] C t = f t *C t-1 + i t *C t (8)

[0117] h t = O t *tanh(C t + b c ) (9)

[0118] where C tis the current cell state, which integrates the information of past and current time; f t *C t-1 represents the information of the last cell state after forgetting gate regulation; i t *C t represents the current input information C t after input gate regulation; b c is the bias term for candidate cell state calculation, which adjusts the linear transformation.

[0119] The hidden state h t output by the LSTM unit generates the final carbon dioxide prediction value of the power plant through linear transformation:

[0120] Y pred = W out *h t +b out (10)

[0121] where W out is the weight of the output layer, and b out is the bias term. After calculation, the LSTM prediction value can be obtained.

[0122] In this embodiment, please refer to Figure 3 , the carbon dioxide monitoring and prediction system of the power plant runs the GRU unit and the LSTM unit in parallel. The GRU unit quickly responds to data with strong short-time dependence and can quickly capture the influence of short-term fluctuations of power plant operating parameters on carbon dioxide concentration. The LSTM unit is good at processing data sequences with long-time dependence and can learn the cumulative effect or slow change trend of carbon dioxide concentration in the power plant over a long period of time. The system compares the outputs of the GRU unit and the LSTM unit and combines them through weighted combination to form the final prediction result. The prediction result of the GRU unit is multiplied by w The prediction result of the LSTM unit is multiplied by w The mean square error formula is used to calculate the error:

[0123]

[0124]

[0125] In order to fully utilize the advantages of the network, the prediction results of the two networks are combined through weighted average method to form the final prediction:

[0126]

[0127] where w GRU and w LSTMare the weights of two neural networks, which can be dynamically adjusted based on the real-time error of each model. According to the error performance of the two models after each iteration, the weight parameters are adjusted, so that the model with smaller error occupies a larger proportion in future prediction, thereby improving the prediction accuracy. Through the cooperative work of the two network units, both short-term and long-term carbon dioxide concentration changes can be focused on, the overall prediction accuracy is improved, and flexible adjustment can be made under different operating conditions. The output control unit outputs the final power plant carbon dioxide prediction value, and provides operation suggestions to the regulation system of the power plant according to the prediction result, and is connected with the real-time regulation module to realize automatic adjustment of the operating parameters of the power plant.

[0128] Embodiment 2

[0129] The double neural network adaptive feedback power plant carbon dioxide control method provided by the application comprises:

[0130] The data input unit inputs data, including the current time input vector x t , which represents the hidden state vector h t-1 of the previous time external input data, which represents the hidden information stored at the last time;

[0131] The GRU unit controls the flow of information through the gating mechanism of the input data of the data input unit, and the reset gate affects the hidden state vector h t-1 of the previous time to a certain extent, which affects the candidate hidden state at the current time;

[0132] The LSTM unit generates the final power plant carbon dioxide prediction value through linear transformation of the input data;

[0133] The error feedback unit adjusts the weight parameters according to the error performance of the GRU unit and the LSTM unit after each iteration, so that the model with smaller error occupies a larger proportion in future prediction, thereby improving the prediction accuracy;

[0134] The output control unit outputs the final power plant carbon dioxide prediction value, and provides operation suggestions to the regulation system of the power plant according to the prediction result, and is connected with the real-time regulation module to realize automatic adjustment of the operating parameters of the power plant.

[0135] In this embodiment, the GRU unit controls the flow of information through the gating mechanism of the input data of the data input unit, and the reset gate affects the hidden state vector h t-1 of the previous time to a certain extent, which affects the candidate hidden state at the current time, and the calculation formula is as follows:

[0136] r t =σ*(W r *[h t-1 ,x t ]+b r) (1)

[0137] where W r and b r are the reset gate weight matrix and bias term, σ is the sigmoid activation function, r t is the reset gate output, which ranges from 0 to 1, when r t is close to 0, the reset gate ignores the influence of the previous hidden state; when r t is close to 1, the historical information is retained;

[0138] The update gate determines how much of the previous hidden state h t is retained in the current hidden state h t-1 , i.e., the information retention degree, and the calculation formula is:

[0139] z t = σ*(W z *[h t-1 ,x t ]+b z ) (2)

[0140] where W z and b z are the update gate weight matrix and bias term, and z t is the update gate output; after the calculation of the reset gate and the update gate, the current input x t and the previous hidden state h t-1 controlled by the reset gate are combined to calculate the candidate hidden state, and the calculation formula is as follows:

[0141]

[0142] where h t is the candidate hidden state, representing the current hidden information that the GRU may update; the final output hidden state h h is determined by whether to retain the old hidden state or accept the new hidden state through the update gate; W t is the trainable weight matrix of the candidate hidden state; tanh is the hyperbolic tangent activation function, which compresses the linearly transformed result to the range of [-1, 1];

[0143] The final hidden state is output according to the update gate z t-1 , and the previous hidden state h t and the current candidate hidden state h are linearly combined, and the formula is as follows:

[0144]

[0145] When z tClose to 1, it means that the current time more information to retain the last time; when z t Close to 0, it means more use of the current time information, the final GRU unit output as the final hidden state.

[0146] In this embodiment, the update gate output is between [0, 1].

[0147] In this embodiment, the LSTM unit includes an input gate, a forget gate, an output gate, a cell state update and an output layer;

[0148] The input gate determines how much of the current time input information xt can enter the current time cell state, and the calculation formula is:

[0149] i t =σ*(W i *x t +U i *h t-1 +b i ) (5)

[0150] Where i t is the activation value, indicating the importance of the current input in updating the cell state; W i is a weight matrix that maps the input x t to the hidden layer; U i is a weight matrix that maps the hidden state h t-1 of the previous time; b t is a bias term;

[0151] The forget gate is used to determine the part of the current time cell state C t-1 that needs to be forgotten, that is, whether to retain some historical information at the current time, and the calculation formula is as follows

[0152] f t =σ*(W f *x t +U f *h t-1 +b f ) (6)

[0153] Where f t is the activation value of the forget gate, indicating the importance of the previous time cell state in the current time update; W f is the weight matrix of the forget gate, which maps the input x t ; U f is the weight matrix of the forget gate, which maps the hidden state h t-1 of the previous time; b f is the bias term of the forget gate;

[0154] The output gate is used to determine the current time cell state Ct h as hidden state h t Output part, and used for next time calculation, the calculation formula is:

[0155] O t = σ * (W O * x t + U O * h t-1 + b O ) (7)

[0156] Wherein, O t is the activation value of the output gate, controlling the output quantity from the cell state to the hidden state; W o is the weight matrix of the output gate, mapping the current input x t ; U o is the weight matrix of the output gate, mapping the hidden state h t-1 of the previous moment; b o is the output gate bias term;

[0157] Cell state update and cell state update calculation formula of output layer:

[0158] C t = f t * C t-1 + i t * C t (8)

[0159] h t = O t * tanh (C t + b c ) (9)

[0160] Wherein, C t is the current time cell state, which integrates the information of the past and the current time; f t * C t-1 represents the information quantity of the cell state of the previous moment after being adjusted by the forgetting gate; i t * C t represents the new information added after the current input information C t is adjusted by the input gate; b c is the bias term used for candidate cell state calculation.

[0161] The foregoing merely illustrates the principles of the application and application of its leading features. This application is not limited to the exact details shown above and described herein, and obvious modifications will occur to those skilled in the art upon reading the foregoing description. Therefore, the scope of the application is not to be determined by the specific examples shown above, but only by the claims below. Any reference signs in the claims should not be construed as limiting the scope of the claims.

[0162] Furthermore, it should be understood that although the description above relates to embodiments, not every embodiment contains only one independent technical solution, and the description above is only for the sake of clarity, and those skilled in the art should understand the description as a whole, and the technical solutions in each embodiment can be appropriately combined to form other embodiments that those skilled in the art can understand. The above is only to illustrate the technical idea of the present application, and cannot limit the protection scope of the present application. Any modification made on the basis of the technical idea of the present application, which is within the scope of the technical solutions, falls within the protection scope of the claims of the present application.

Claims

1. A double neural network adaptive feedback carbon dioxide control system for a thermal power plant, characterized in that, The data input unit, the GRU unit, the LSTM unit, the error feedback unit and the output control unit are included. The data input unit is configured to input data, including a current time input vector x t , representing a previous time hidden state vector h t-1 , representing the stored implicit information at the previous time; The GRU unit is used for controlling the flow of information through the gating mechanism of the input data of the data input unit, and the reset gate affects the hidden state vector h of the previous moment t-1 The influence on the candidate hidden state of the current moment The LSTM unit is configured to generate a final carbon dioxide prediction value of the thermal power plant through linear transformation of the input data. The error feedback unit is configured to adjust the weight parameters according to the error performance of the GRU unit and the LSTM unit after each iteration, so that the smaller error occupies a larger proportion in future prediction, and the prediction accuracy is improved. The output control unit is configured to output the final carbon dioxide prediction value of the thermal power plant, and provide operation suggestions to a regulation system of the thermal power plant according to the prediction result, and is connected with a real-time regulation module to realize automatic adjustment of operation parameters of the thermal power plant.

2. The dual neural network adaptive feedback carbon dioxide control system for thermal power plants of claim 1, wherein, The GRU unit controls the flow of information through the gated mechanism of the input data of the data input unit, and the reset gate affects the previous time hidden state vector h t-1 The extent to which the current time candidate implicit state is affected is calculated as follows: r t = σ * (W r * [h t-1 , x t ] + b r ) (1) where W r and b r are the reset gate weight matrix and bias term, σ is the sigmoid activation function, r t is the reset gate output, which is in the range [0, 1], when r t is close to 0, the reset gate ignores the influence of the previous hidden state; when r t is close to 1, the historical information is retained; The update gate determines the hidden state h at the current time t How much of the previous hidden state h is retained t-1 That is, the degree of information retention, and the calculation formula is: z t =σ*(W z *[h t-1 ,x t ]+b z ) (2) where W z and b z are updated gate weight matrix and bias term, z t is the updated gate output; after the reset gate and the update gate are calculated, the input x t at the current moment and the part of the hidden state h t-1 controlled by the reset gate at the last moment are combined to calculate the candidate hidden state, and the calculation formula is as follows: wherein, is the candidate hidden state, representing the current time hidden information that the GRU can update; the final output hidden state h t is the update gate that controls whether to retain the old hidden state or accept the new hidden state on this basis; W h is the trainable weight matrix of the candidate hidden state; tanh is the hyperbolic tangent activation function, which compresses the linearly transformed result to the range of [-1, 1]; The final hidden state is updated according to the update gate z t The output value is obtained by performing linear combination on the hidden state h t-1 The current time candidate hidden state h The linear combination is performed, and the formula is as follows: When z t approaches 1, it means that more information of the previous time is reserved for the current time; when z t approaches 0, it means that more information of the current time is used, and the final GRU unit output is the final hidden state.

3. The dual neural network adaptive feedback carbon dioxide control system for a thermal power plant of claim 2, wherein, The update gate output is between 0 and 1.

4. The dual neural network adaptive feedback carbon dioxide control system for thermal power plants of claim 1, wherein, The LSTM unit includes an input gate, a forget gate, an output gate, a cell state update and an output layer. The input gate determines how much of the input information xt at the current time can enter the cell state at the current time, and the calculation formula is: i t = σ * (W i * x t + U i * h t-1 + b i ) (5) where i t is the activation value, indicating the importance of the current input in updating the cell state; W i is the weight matrix mapping the input x t to the hidden layer; U i is the weight matrix mapping the hidden state h t-1 at the previous time step to the hidden state h t is the bias term. The forget gate is used to decide the current cell state C t-1 The part that needs to be forgotten, i.e. whether to retain some historical information at the current moment, is calculated as follows f t = σ * (W f * x t + U f * h t-1 + b f ) (6) wherein f t is the activation value of the forget gate, indicating the importance of the cell state at the previous time in the update at the current time; W f is the weight matrix of the forget gate, mapping the input x t ; U f is the weight matrix of the forget gate, mapping the hidden state h t-1 at the previous time; b f is the bias term of the forget gate; The output gate is used to determine the cell state C at the current time t The hidden state h at the current time is calculated as follows: t The output part is used to calculate the next time, and the calculation formula is as follows: O t = σ * (W O * x t + U O * h t-1 + b O ) (7) wherein, O t is the activation value of the output gate, controlling the output quantity from the cell state to the hidden state; W o is the weight matrix of the output gate, mapping the current input x t ; U o is the weight matrix of the output gate, mapping the hidden state h t-1 of the previous moment; b o is the output gate bias term; The cell state update of the cell state update and the output layer is calculated by the following formula: C t = f t * C t-1 + i t * C t (8) h t =O t *tanh(C t +b c (9) Among them, C t It represents the current cell state, integrating information from the past and present; f t *C t-1 This represents the amount of information about the cell state at the previous moment after it has been regulated by the forgetting gate; i t *C t Indicates the current input information C t New information added after input gate adjustment; b c It is a bias term used for calculating candidate cell states.

5. The dual neural network adaptive feedback carbon dioxide control system for thermal power plants of claim 4, wherein, The hidden state h output by the LSTM unit t Generating final power plant CO2 prediction values through linear transformation: Y pred = W out h t + b out (10) where W out is the weight of the output layer, b out is the bias term, and the LSTM prediction is obtained after the calculation.

6. The dual neural network adaptive feedback carbon dioxide control system for thermal power plants of claim 4, wherein, activation value i t normalized between 0 and 1 by a Sigmoid activation function.

7. A method for carbon dioxide control in a thermal power plant using dual neural network adaptive feedback, characterized by, It includes: The data input unit inputs data, including a current time input vector x t , represents a previous time hidden state vector h t-1 , represents the stored implicit information at the previous time The GRU unit controls the flow of information through the input data of the data input unit by a gating mechanism, and the reset gate affects the hidden state vector h of the previous moment t-1 To what extent does it affect the candidate hidden state at the current moment; The LSTM unit generates a final carbon dioxide prediction value of the thermal power plant through linear transformation of the input data. The error feedback unit adjusts the weight parameters according to the error performance of the GRU unit and the LSTM unit after each iteration, so that the smaller error occupies a larger proportion in future prediction, and the prediction accuracy is improved. The output control unit outputs the final carbon dioxide prediction value of the thermal power plant, and provides operation suggestions to a regulation system of the thermal power plant according to the prediction result, and is connected with a real-time regulation module to realize automatic adjustment of operation parameters of the thermal power plant.

8. The dual neural network adaptive feedback method for carbon dioxide control in a thermal power plant according to claim 7, characterized in that, The GRU unit controls the flow of information through the gated mechanism of the input data of the data input unit, and the reset gate affects the previous time hidden state vector h t-1 The extent to which the current time candidate implicit state is affected is calculated as follows: r t = σ * (W r * [h t-1 , x t ] + b r ) (1) where W r and b r are the reset gate weight matrix and bias term, σ is the sigmoid activation function, r t is the reset gate output, which is in the range [0, 1], when r t is close to 0, the reset gate ignores the influence of the previous hidden state; when r t is close to 1, the historical information is retained; The update gate determines the hidden state h at the current time t How much of the previous hidden state h is retained t-1 That is, the degree of information retention, and the calculation formula is: z t = σ * (W z * [h t-1 , x t ] + b z ) (2) where W z and b z are updated gate weight matrix and bias term, z t is the updated gate output; after the reset gate and the update gate are calculated, the input x t at the current moment and the part of the hidden state h t-1 controlled by the reset gate at the last moment are combined to calculate the candidate hidden state, and the calculation formula is as follows: wherein, is the candidate hidden state, representing the current time hidden information that the GRU can update; the final output hidden state h t is the update gate that controls whether to keep the old hidden state or accept the new hidden state on this basis; W h is the trainable weight matrix of the candidate hidden state; tanh is the hyperbolic tangent activation function, which compresses the linearly transformed result to the range of [-1, 1]; The final hidden state is updated according to the update gate z t The output value is obtained by performing linear combination on the hidden state h t-1 The current time candidate hidden state h The linear combination is performed, and the formula is as follows: When z t is close to 1, it means that more information of the previous time step is reserved for the current time step; when z t is close to 0, it means that more information of the current time step is used, and the final GRU unit output is the final hidden state.

9. The dual neural network adaptive feedback method for carbon dioxide control in a thermal power plant according to claim 8, wherein, The update gate output is between 0 and 1.

10. The dual neural network adaptive feedback method for carbon dioxide control in a thermal power plant according to claim 7, wherein, The LSTM unit includes an input gate, a forget gate, an output gate, a cell state update and an output layer. The input gate determines how much of the input information xt at the current time can enter the cell state at the current time, and the calculation formula is: i t = σ * (W i * x t + U i * h t-1 + b i ) (5) where i t is the activation value, indicating the importance of the current input in updating the cell state; W i is the weight matrix mapping the input x t to the hidden layer; U i is the weight matrix mapping the hidden state h t-1 at the previous time step; b t is the bias term. The forget gate is used to decide the current cell state C t-1 The part that needs to be forgotten, i.e. whether to retain some historical information at the current time, is calculated as follows f t = σ * (W f * x t + U f * h t-1 + b f ) (6) wherein f t is the activation value of the forget gate, indicating the importance of the cell state at the previous time in the update at the current time; W f is the weight matrix of the forget gate, mapping the input x t ; U f is the weight matrix of the forget gate, mapping the hidden state h t-1 at the previous time; b f is the bias term of the forget gate; The output gate is used to determine the cell state C at the current time t The hidden state h at the current time is calculated as t The output part, and is used to calculate the next time, and the calculation formula is: O t = σ * (W O * x t + U O * h t-1 + b O ) (7) wherein, O t is the activation value of the output gate, controlling the output quantity from the cell state to the hidden state; W o is the weight matrix of the output gate, mapping the current input x t ; U o is the weight matrix of the output gate, mapping the hidden state h t-1 of the previous moment; b o is the output gate bias term; The cell state update of the cell state update and the output layer is calculated by the following formula: C t = f t * C t-1 + i t * C t (8) h t = O t * tanh(C t + b c )(9) where C t is the current cell state, fusing information from past and current time instants; f t *C t-1 denotes the amount of information from the previous cell state regulated by the forget gate; i t *C t denotes the new information added after regulating the current input information C t by the input gate; b c is the bias term used for the candidate cell state computation.

Images