Emission prediction and control methods for thermal power plants

Abstract

The invention discloses a method for predicting and controlling emissions from a thermal power plant, which relates to the field of electric power technology, and comprises the following steps: obtaining and preprocessing the combustion process data of a thermal power plant boiler; inputting the preprocessed combustion process data of the thermal power plant boiler into a pre-trained NO x The emission concentration prediction model predicts NO x Emission concentration data; the predicted NO x The emission concentration data and the pre-processed thermal power plant boiler combustion process data are input into the pre-trained control strategy optimization model based on reinforcement learning to output the optimal control parameters; the optimal control parameters are sent to the boiler's combustion system and SCR denitrification system through the PLC system to implement automatic adjustment control. This invention integrates advanced data processing and machine learning technologies to optimize NO in thermal power plants. x The prediction and control of emissions provides an efficient and accurate solution that helps achieve the dual goals of environmental protection and economic benefits of thermal power plants.

Description

Technical Field

[0001] The present invention relates to the field of electric power technology, and in particular to a method for predicting and controlling emissions from a thermal power plant. Background Art

[0002] Thermal power plants produce a large amount of nitrogen oxides (NO x ), this type of pollutant is not only one of the main causes of acid rain, but also causes serious harm to human health and the ecological environment. Therefore, controlling NO x At present, SCR (selective catalytic reduction) denitrification technology is widely used in thermal power plants. It injects a reducing agent (such as ammonia or urea) into the flue gas to remove NO. x However, in actual operation, there is a complex nonlinear coupling relationship between the boiler combustion system and the SCR denitrification system. x Control methods often adopt control strategies based on fixed rules, which are difficult to adapt to complex working conditions in real time, and may significantly increase operating costs (such as excessive consumption of reductants or reduced fuel utilization) while pursuing low emissions. Summary of the invention

[0003] The present invention provides a method for predicting and controlling emissions from a thermal power plant, which can solve the above-mentioned problem.

[0004] In order to solve the above problems, the technical solution adopted by the present invention is as follows:

[0005] The present invention provides a method for predicting and controlling emissions from a thermal power plant, comprising the following steps:

[0006] S1. Obtain the combustion process data of the thermal power plant boiler and perform preprocessing;

[0007] S2, input the pre-processed thermal power plant boiler combustion process data into the pre-trained NO x The emission concentration prediction model predicts NO x Emission concentration data;

[0008] S3, the predicted NO x The emission concentration data and the pre-processed thermal power plant boiler combustion process data are input into the pre-trained control strategy optimization model based on reinforcement learning to output the optimal control parameters;

[0009] S4. The optimal control parameters are sent to the boiler's combustion system and SCR denitrification system through the PLC system to implement automatic adjustment and control.

[0010] As a further description of the above technical solution, the preprocessing of boiler combustion process data includes data cleaning, data standardization and time series segmentation.

[0011] As a further description of the above technical solution, the NO x The emission concentration prediction model includes an input layer, a multi-scale time series feature extraction module, a Transformer module, a feature fusion module and an output layer; the multi-scale time series feature extraction module is used to extract the multi-scale features of boiler operating parameters; the Transformer module captures the long-range dependency features in the boiler combustion process through the self-attention mechanism; the feature fusion module fuses the time series features with the global features through a weighted mechanism to predict NO x Emission concentration data.

[0012] As a further description of the above technical solution, it also includes: transferring the NO x The emission concentration prediction model is incrementally trained to adapt it to seasonal changes and boiler load fluctuations. The incremental training method includes: cleaning and feature extraction of new data; calculating the error between the model prediction value and the actual monitoring value; fine-tuning the model parameters according to the error size and updating the weights.

[0013] As a further description of the above technical solution, the control strategy optimization model includes a state space, an action space, a reward function and a dual network structure based on a deep deterministic policy gradient algorithm; wherein the state space includes boiler combustion parameters, NO x Emission concentration data and SCR denitrification system parameters; the action space includes boiler combustion operation control parameters and the ammonia injection amount of the SCR system; the reward function comprehensively considers NO x Concentration control target and boiler combustion operation cost target.

[0014] As a further description of the above technical solution, the reward function of the control strategy optimization model is defined as:

[0015]

[0016] Among them, R is the reward value, NO x Emission concentration, C limit is the emission standard limit, C cost is the combustion operation cost, α and β are weight coefficients, which control the balance between the two objectives.

[0017] As a further description of the above technical solution, it also includes: summarizing historical data to generate NO x Emission trend chart and control measures effect analysis chart, then generate NO x Emissions reporting, including NOx Comparison of emission prediction and actual values, evaluation of the effect of control measures, and the effect of various operating parameters on NO x Sensitivity analysis of emissions.

[0018] As a further description of the above technical solution, the PLC system can adjust the boiler combustion parameters in real time according to the optimal control parameters, dynamically adjust the ammonia injection amount and reaction temperature of the SCR denitrification system, record all operating parameters during the control process and generate a log report.

[0019] Compared with the prior art, the present invention has the following beneficial effects:

[0020] 1) NO can be predicted more accurately x Emission concentration, output optimal control parameters, realize precise control of boiler combustion system and SCR denitrification system, realize real-time adjustment of boiler combustion process, improve system adaptability and flexibility; reduce unnecessary ammonia use, reduce denitrification cost, and reduce environmental pollution; reduce the system's sensitivity to external disturbances and improve system stability and reliability.

[0021] 2) The proposed NO x The emission concentration prediction model combines LSTM and 1D-CNN modules to realize time series feature extraction, and models global dependencies through the Transformer module, overcoming the shortcomings of traditional time series models in terms of long-term dependence and complex coupling characteristics.

[0022] 3) The control strategy optimization model based on deep reinforcement learning adopts a dual network structure (Actor-Critic) to reduce NO x At the same time as the emission concentration, it can dynamically balance the combustion cost and the amount of denitrification reducing agent to achieve multi-objective control optimization; the reward function design comprehensively considers NO x Emission limits and combustion economy ensure that the system can maintain efficient operation under different operating conditions.

[0023] 4) The introduction of transfer learning and incremental training technology can quickly adjust the prediction model according to the new data, so that it can adapt to boiler load fluctuations and seasonal changes, and improve the generalization performance of the model; the reinforcement learning model significantly enhances its adaptability to complex working conditions through real-time updates and multi-task optimization, avoiding the failure of traditional control methods under dynamic conditions.

[0024] 5) The PLC system realizes the automatic control of the boiler combustion system and SCR denitrification system, which reduces manual intervention and improves the operation efficiency and control accuracy; the control strategy optimization model can effectively reduce the consumption of reducing agent and fuel waste, and reduce NO x While reducing emission concentration, it can significantly reduce operating costs and improve economic benefits.

[0025] 6) Summarize and analyze historical data to generate NO x Emission trend graphs, sensitivity analysis reports and control measures effectiveness evaluation provide a scientific basis for operational optimization and contribute to long-term improvement in decision-making.

[0026] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the embodiments of the present invention are specifically cited below and described in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0028] Figure 1 It is a flow chart of the method for predicting and controlling emissions from a thermal power plant described in the embodiment. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0030] Please refer to Figure 1 The embodiment of the present invention discloses a method for predicting and controlling emissions from a thermal power plant, which is as follows:

[0031] 1. Obtain the combustion process data of the thermal power plant boiler and perform preprocessing.

[0032] We set up multiple types of sensors to collect key parameters of the combustion process of thermal power plant boilers, including: fuel composition, combustion temperature, flue gas flow rate, SCR reactor temperature, etc.

[0033] Methods for preprocessing the collected data include:

[0034] (1) Data cleaning: Eliminate outliers and missing values ​​and use interpolation to fill in the gaps. Among them, outlier elimination uses the 3 times standard deviation method to eliminate abnormal data.

[0035] (2) Data standardization: Normalize the input parameters, that is, normalize all features to the range [0,1].

[0036] (3) Time series segmentation: Generate training samples based on the sliding window method.

[0037] 2. Input the pre-processed thermal power plant boiler combustion process data into the pre-trained NO x The emission concentration prediction model predicts NO x Emission concentration data.

[0038] In this embodiment, NO x The emission concentration prediction model introduces a multi-scale time series feature extraction mechanism and a self-attention mechanism, which improves the model's ability to fit and predict complex combustion process data.

[0039] NO x The emission concentration prediction model includes an input layer, a multi-scale time series feature extraction module, a Transformer module, a feature fusion module and an output layer.

[0040] The input layer is used to input multi-dimensional boiler combustion parameters. The input data of the combustion process is Where N is the length of the time series and d is the feature dimension of each time point.

[0041] The multi-scale temporal feature extraction module combines LSTM and 1D convolutional network (1D-CNN).

[0042] LSTM is used to extract long-term dependent temporal features. The hidden state update formula of LSTM is:

[0043] h t =f(X t ,h t-1 )=σ(W f [h t-1 , X t ]+b f )·h t-1 +tanh(W c [h t-1 , X t ]+b c )

[0044] Among them, h t is the hidden state at time t, X t is the current time point of the input sequence, W f and W c is the weight matrix of LSTM; σ is the activation function.

[0045] 1D-CNN is used to capture short-term fluctuation characteristics. The 1D convolution operation formula is:

[0046]

[0047] Among them, W conv is the convolution kernel weight, and * is the convolution operation.

[0048] The Transformer module is used to capture the global dependencies between key time points based on the self-attention mechanism. The self-attention calculation formula is:

[0049]

[0050] Among them, Q, K, and V are the input query, key, and value matrices respectively, and d k is the dimension of the key matrix, used for scaling and normalization.

[0051] The feature fusion module is used to fuse the features output by LSTM, 1D-CNN and Transformer using a weighted fusion method, namely:

[0052]

[0053] Among them, α, β and γ are fusion weights, satisfying α+β+γ=1;

[0054] The final output prediction value is:

[0055] Y predicted =W out h fused +b out Among them, W out and b out are the output layer weights and biases respectively. The output layer is used to output NO using the fully connected layer (Fully Connected Layer) x Emission concentration predicted value Y predicted .

[0056] NO x The loss function used in the training of the emission concentration prediction model is defined as:

[0057] L total =L mse +λ1L time +λ2L smooth

[0058] Where: L mse Represents the mean square error loss, which measures the deviation between the predicted value and the true value; L time represents the temporal consistency loss, which is used to ensure the continuity of short-term predictions; L smooth It represents the smoothing loss of the predicted value, which is used to suppress excessive fluctuations in the predicted results; λ1 and λ2 are weight parameters, which are used to adjust the contribution of each part of the loss.

[0059] The above NO xThe structural design of the emission concentration prediction model, combining LSTM and 1D-CNN, improves the modeling ability of long- and short-term dependencies; the Transformer module captures the correlation of key moments in the boiler operation status and improves the prediction accuracy; the feature fusion mechanism effectively combines the advantages of different modules to make the model more adaptable to changes in boiler operating conditions; by introducing multi-task loss functions (temporal consistency loss and result smoothing loss), the prediction model can generate smoother and more stable prediction results, with the error reduced by more than 20% compared with traditional models.

[0060] 3. The predicted NO x The emission concentration data and the pre-processed thermal power plant boiler combustion process data are input into the pre-trained reinforcement learning-based control strategy optimization model to output the optimal control parameters.

[0061] Among them, the control strategy optimization model is based on reinforcement learning, using the deep deterministic policy gradient algorithm (DDPG) and dual network structure to improve optimization efficiency and adaptability. The control strategy optimization model architecture includes state space, action space, dual network structure and reward function.

[0062] The objective function of reinforcement learning for the control strategy optimization model is:

[0063]

[0064] Among them, J(θ) is the objective function, R t is the reward value at time t; γ is the discount factor, and its value range is 0<γ<10.

[0065] The state space includes boiler combustion parameters and predicted NO x concentration data. The action space includes combustion parameters (such as excess air coefficient) and SCR denitrification parameters (such as ammonia injection amount). In the dual network structure, one network is used as a strategy network to generate control parameters, and the other network is used as an evaluation network to calculate the benefits of the strategy.

[0066] The reward function will be NO x The concentration control goal is combined with the goal of minimizing boiler combustion costs.

[0067]

[0068] Among them, R is the reward value, NO x Emission concentration, C limit is the emission standard limit, C cost is the combustion operation cost, α and β are weight coefficients, which control the balance between the two objectives.

[0069] The parameter update formula of the control strategy optimization model is:

[0070]

[0071] Where Q(s,a) is the action value function that evaluates the output of the network; π θ (s) is the optimal strategy given by the policy network.

[0072] The structural design of the control strategy optimization model introduces a dual network architecture, separates the strategy generation and value evaluation processes, and improves the efficiency of strategy optimization; the dynamic balance mechanism achieves NO through weight adjustment of the reward function. x Dynamic balance between emission control and cost optimization; the introduction of the DDPG algorithm enables the model to adapt to high-dimensional continuous action space and improves its adaptability to complex combustion conditions.

[0073] 4. The optimal control parameters are sent to the boiler's combustion system and SCR denitrification system through the PLC system to achieve real-time NO x Emission control.

[0074] 5. At regular intervals, use the newly added data to x The emission concentration prediction model is incrementally trained and transfer learning technology is used to adapt the model to the operating conditions of different seasons and boiler load changes. During the incremental training process, the NO x Emission concentration data and the NO x NO emission concentration prediction model x Compare the emission concentration data, calculate the error, and adjust the NO according to the error. x Parameters of emission concentration prediction model and update model weights.

[0075] The error calculation formula is:

[0076]

[0077] Among them, E represents the mean square error, Y i actual For actual NO x Concentration value, Y i predicted To predict NO x concentration value, and N is the number of samples.

[0078] 6. Summarize historical data and generate NO x Emission trend chart and control measures effect analysis chart, then generate NO x Emissions reporting, including NO x Comparison of emission prediction and actual values, evaluation of the effect of control measures, and the effect of various operating parameters on NO x Sensitivity analysis of emissions, where the sensitivity analysis uses regression analysis method, specifically:

[0079]

[0080] Among them, S j represents the sensitivity coefficient X of the jth parameter j is the jth input parameter value, Y is NO x Emission concentration value, Indicates that the parameter is NO x Partial derivative of concentration.

[0081] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting and controlling emissions from a thermal power plant, characterized in that: The following steps are involved: S1. Obtain the combustion process data of the thermal power plant boiler and perform preprocessing; S2, inputting the pre-processed thermal power plant boiler combustion process data into the pre-trained NOx emission concentration prediction model to predict the NOx emission concentration data; S3, input the predicted NOx emission concentration data and the pre-processed thermal power plant boiler combustion process data into the pre-trained control strategy optimization model based on reinforcement learning, and output the optimal control parameters; S4. The optimal control parameters are sent to the boiler's combustion system and SCR denitrification system through the PLC system to implement automatic adjustment and control.

2. The method according to claim 1, characterized in that The preprocessing of boiler combustion process data includes data cleaning, data standardization and time series segmentation.

3. The method according to claim 1, characterized in that The NOx emission concentration prediction model includes an input layer, a multi-scale time series feature extraction module, a Transformer module, a feature fusion module and an output layer; wherein the multi-scale time series feature extraction module is used to extract the multi-scale features of boiler operating parameters; the Transformer module captures the long-range dependency features in the boiler combustion process through a self-attention mechanism; the feature fusion module fuses the time series features with the global features through a weighted mechanism to predict the NOx emission concentration data.

4. The method according to claim 1, characterized in that Also includes: The NOx emission concentration prediction model is incrementally trained through transfer learning to adapt it to seasonal changes and boiler load fluctuations. The incremental training method includes: cleaning and feature extraction of new data; calculating the error between the model prediction value and the actual monitoring value; fine-tuning the model parameters according to the error size and updating the weights.

5. The method according to claim 1, characterized in that The control strategy optimization model includes a state space, an action space, a reward function and a dual network structure based on a deep deterministic policy gradient algorithm; wherein the state space includes boiler combustion parameters, NOx emission concentration data and SCR denitrification system parameters; the action space includes boiler combustion operation control parameters and the ammonia injection amount of the SCR system; the reward function comprehensively considers the NOx concentration control target and the boiler combustion operation cost target.

6. The method according to claim 1, characterized in that The reward function of the control strategy optimization model is defined as: Among them, R is the reward value, is the NOx emission concentration, C limit is the emission standard limit, C cost is the combustion operation cost, α and β are weight coefficients, which control the balance between the two objectives.

7. The method according to claim 1, characterized in that Also includes: Summarize historical data to generate NOx emission trend charts and control measures effect analysis charts, and then generate NOx emission reports, including NOx emission forecast and actual value comparison results, control measures implementation effect evaluation, and sensitivity analysis of various operating parameters to NOx emissions.

8. The method according to claim 1, characterized in that The PLC system can adjust the boiler combustion parameters in real time according to the optimal control parameters, dynamically adjust the ammonia injection amount and reaction temperature of the SCR denitrification system, record all operating parameters during the control process and generate a log report.

Images