Deep peak regulation system suitable for thermal power generating unit

Abstract

The invention discloses a deep peak regulation system suitable for a thermal power generating unit, and relates to the technical field of operation control of the thermal power generating unit, the system comprises a data acquisition module, a multi-dimensional load prediction module, an improved coordination control module, a self-adaptive denitration optimization module and a monitoring early warning module, and the modules realize data interaction through data interfaces. The improved coordination control module is used for receiving a load prediction result and a real-time load deviation, and outputting control instructions of boiler fuel quantity, feed water flow and steam turbine control valve opening by adopting a dynamic weight adjustment algorithm based on model prediction control (MPC); by improving the LSTM-attention model and fusing multi-dimensional features, the load prediction precision is remarkably improved, and an accurate load reference is provided for deep peak regulation; the system realizes integration of load prediction, coordination control, denitration optimization and monitoring early warning, realizes real-time interaction of data of each module, and improves the overall peak regulation efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of thermal power unit operation control, and more particularly to a deep peak shaving system suitable for a thermal power unit. BACKGROUND

[0002] With the increasing proportion of new energy power generation (wind power, photovoltaic) in the power system, the intermittency and volatility of its output have a significant impact on the stability of the grid frequency and voltage quality. As the main regulating power source of the power grid, thermal power units need to have deep peak shaving capability to smooth out new energy output fluctuations and ensure the safe and stable operation of the power grid. Deep peak shaving usually requires thermal power units to reduce the operating load to below 50% of the rated load, and some units even need to reduce to 30% of the rated load.

[0003] However, existing thermal power units face many technical bottlenecks during deep peak shaving:

[0004] Low load prediction accuracy: Traditional load prediction models mostly use single time series algorithms (such as ARIMA, traditional LSTM), without fully considering multi-dimensional influencing factors such as environmental temperature, fuel characteristics, and grid dispatching instructions, resulting in large prediction errors (usually more than 5%), which cannot provide accurate load reference for unit peak shaving;

[0005] Response lag of coordinated control: Thermal power units are typical multivariable, strongly coupled, and large-lag systems. Traditional coordinated control uses PID control algorithm, which is difficult to cope with rapid changes in load during deep peak shaving, and is prone to load overshoot or response lag, resulting in load tracking error exceeding ±5% of rated load;

[0006] Unstable operation of denitration system: Under deep peak shaving conditions, the boiler outlet flue gas temperature decreases and the flow rate decreases, resulting in a decrease in the activity of the denitration catalyst. Traditional denitration control uses fixed PID parameters and cannot dynamically adjust the ammonia injection amount, which is prone to NOx emission exceeding the standard or high ammonia escape rate (more than 5ppm), causing catalyst poisoning and environmental pollution;

[0007] Low system integration: Existing peak shaving systems mostly optimize single links (such as load control or denitration control), without realizing integrated integration of load prediction, coordinated control, and denitration optimization, resulting in poor data interaction between systems and low overall peak shaving efficiency.

[0008] Therefore, a deep peak shaving system suitable for a thermal power unit is proposed to solve the problems in the above background. SUMMARY

[0009] In order to overcome the above-mentioned defects of the prior art, the present application provides a deep peak shaving system suitable for a thermal power unit to solve the problems raised in the above background.

[0010] To achieve the above object, the application provides the following technical scheme: a deep peak regulation system suitable for thermal power generating units, comprising a data acquisition module, a multi-dimensional load prediction module, an improved coordinated control module, a self-adaptive denitration optimization module and a monitoring and early warning module, data interaction is realized between the modules through a data interface, the data acquisition module is used for real-time acquisition of operation parameters of the thermal power generating unit, and an output end of the data acquisition module is electrically connected with an input end of the multi-dimensional load prediction module;

[0011] The multi-dimensional load prediction module realizes 1-24 hours of unit load prediction in the future based on the acquired operation parameters and power grid dispatching instructions through an improved long short-term memory network (LSTM) and attention mechanism fusion model.

[0012] The improved coordinated control module is used for receiving load prediction results and real-time load deviation, adopting a dynamic weight adjustment algorithm based on model predictive control (MPC), and outputting control instructions of a boiler fuel quantity, a feedwater flow and a turbine governing valve opening degree.

[0013] The self-adaptive denitration optimization module dynamically adjusts denitration ammonia injection quantity through an optimization model combining a fuzzy PID and a genetic algorithm according to boiler outlet flue gas parameters and denitration system operation states.

[0014] The monitoring and early warning module real-time monitors operation states of each module, issues an early warning signal when parameters exceed preset threshold values, and generates an emergency regulation and control scheme.

[0015] Further, the operation parameters of the thermal power generating unit acquired in the data acquisition module include but are not limited to boiler side data, turbine side data, denitration system side data and environment and dispatching side data; wherein the boiler side data includes boiler outlet steam pressure, steam temperature, steam flow, fuel quantity, feedwater flow and furnace pressure; the turbine side data includes turbine speed, power generation power, governing valve opening degree and exhaust temperature; the denitration system side data includes denitration system inlet flue gas temperature, flue gas flow, NO x Concentration, ammonia escape rate and ammonia injection quantity; the environment and dispatching side data includes environment temperature, power grid dispatching instructions and fuel heat value.

[0016] Further, the multi-dimensional load prediction module comprises a data preprocessing unit and an LSTM and attention mechanism fusion model construction unit, wherein the data preprocessing unit is used for preprocessing the acquired original data, and the preprocessing strategy of the data preprocessing unit is as follows:

[0017] Missing value filling: missing data is filled by using a linear interpolation method, and for data with continuous missing of more than 5 sampling points, the average value of the same period data of the previous 7 days is used for filling;

[0018] Outlier rejection: 3σ criterion is used to reject outliers, that is, when the data is out of the range of [μ-3σ, μ+3σ], it is determined as an outlier, and the average value of the adjacent time data is used to replace it;

[0019] Data normalization: Min-Max normalization is used to map all feature parameters to the interval [0, 1], and the normalization formula is: Where x is the original data, x min , x max and x norm are the minimum and maximum values of the data, respectively, and x e is the normalized data.

[0020] Further, the LSTM and attention mechanism fusion model construction unit is used to construct an LSTM and attention mechanism fusion model, and the LSTM and attention mechanism fusion model includes an input layer, an embedding layer, an improved LSTM layer, an attention layer, a fully connected layer, and an output layer. Specifically:

[0021] Input layer: receiving 4 types of feature parameters after preprocessing, and the 4 types of feature parameters are historical load, dispatching instruction, ambient temperature, and fuel heat value, the input data dimension is (N, T, 4), where N is set to 10000 historical samples, T is set to 24, that is, based on the previous 24 hours of data to predict the future 1 hour load, if the future 24 hours of load is predicted, then T is expanded to 48;

[0022] Embedding layer: a fully connected neural network is used to map the input features to a high-dimensional feature space, and the mapping formula is: E = W e ·X + b e , where E is the embedding feature matrix, W e is the embedding layer weight matrix, X is the input data matrix, and b e is the embedding layer bias term;

[0023] Improved LSTM layer: a temperature compensation factor α is introduced into the forgetting gate, and the forgetting gate calculation formula is modified as: f t = σ'(W f ·[h t-1 , x t ]+b f + α·ΔT), where f t is the output of the t-time forgetting gate, σ' is the sigmoid activation function, W f is the forgetting gate weight matrix, h t-1 is the t-1-time hidden layer state, x t is the t-time input feature, b f is the forgetting gate bias, ΔT is the deviation value of the ambient temperature and the rated temperature, and α is the temperature compensation coefficient and the value range is 0.01-0.05;

[0024] Attention layer: the weight of each time step feature is calculated by using additive attention mechanism, and the weight calculation formula is: wherein a t is the attention weight at time t, W a is the attention layer weight matrix, b a is the attention layer bias;

[0025] Fully connected layer: there are two layers in total, and the first layer adopts 64→32 dimension compression, and the activation function is ReLU; the second layer adopts 32→1 dimension mapping, and the activation function is Linear, and finally the load prediction value y pred is output, and the dimension is (N, 1).

[0026] Further, the dynamic weight MPC algorithm in the improved coordinated control module comprises a prediction model, a rolling optimization and a feedback correction, wherein the prediction model adopts a method of mechanism and data fusion to construct a unit load prediction model, and the expression is: y(k+1|k)=f(y(k),u(k),d(k)), wherein y(k+1|k) is a load value at k+1 time based on the state prediction at k time, y(k) is an actual load at k time, u(k) is a control input vector at k time, the input vector includes fuel quantity, feed water flow, and valve opening, d(k) is a disturbance vector, the disturbance vector includes fuel heat value fluctuation and power grid frequency deviation, f(·) is a nonlinear prediction function, and is obtained by offline training of historical operation data.

[0027] Further, a dynamic weight objective function is introduced in the rolling optimization stage of the dynamic weight MPC algorithm, and the formula is: wherein J is an optimization objective function, P' is a prediction time domain, M is a control time domain, y ref (k+i) is a load reference value at k+i time, Δu(k+i) is a control input increment, ω1(k), ω2(k) are dynamic weight coefficients, and ω1(k)+ω2(k)=1;

[0028] The dynamic weight coefficient is adaptively adjusted according to the load deviation Δy=|y ref (k)-y(k)|, and the adjustment rule is:

[0029] When Δy>5%P e , ω1(k)=0.8, ω2(k)=0.2;

[0030] When 2%P e ≤Δy≤5%P e , ω1(k)=0.6, ω2(k)=0.4;

[0031] When Δy<2%P eω1(k) = 0.4, ω2(k) = 0.6;

[0032] where P e is the rated load of the unit.

[0033] Further, the feedback correction stage in the dynamic weight MPC algorithm corrects the prediction error by a Kalman filter, and the correction formula is: y corr (k+i|k) = y(k+i|k) + K(k) e(k), where y corr (k+i|k) is the corrected predicted load, K(k) is the Kalman gain, and e(k) = y(k) - y(k|k-1) is the prediction error at time k.

[0034] Further, the fuzzy PID-genetic algorithm model in the adaptive denitration optimization module includes a fuzzy reasoning unit and a genetic algorithm optimization unit, and the fuzzy reasoning unit takes the NO x concentration deviation e = C NOx,ref -C NOx,act and the rate of change of deviation as input, and the proportional coefficient K p , the integral coefficient K i , and the differential coefficient K d of the PID controller as output.

[0035] The fuzzy subsets of e and ec are both {NB, NM, NS, ZO, PS, PM, PB}, corresponding to the domain [-6, 6]; K p , K i , and K d correction are all {NB, NM, NS, ZO, PS, PM, PB}, corresponding to the domains [-2, 2], [-0.5, 0.5], and [-0.1, 0.1] respectively.

[0036] The genetic algorithm optimization unit is used to optimize the initial parameters of the fuzzy PID controller, and the optimization goal is to minimize the ammonia escape rate and meet the NO x emission standards, and the fitness function is: F is the fitness value, λ1 and λ2 are weight coefficients, γ1 and γ2 are penalty coefficients, is the ammonia escape rate, is the NO x emission standard value.

[0037] The operation steps of the genetic algorithm include:

[0038] A1. Initialize the population: randomly generate 50 PID parameter combinations (K p0 , K i0 , Kd0 ), parameter range K p0 ∈ [5, 15], K i0 ∈ [0.1, 1.0], K d0 ∈ [0.01, 0.1];

[0039] A2. Selection operation: 30 individuals are selected into the next generation by roulette selection method;

[0040] A3. Cross operation: single-point crossover is adopted, and the crossover probability is 0.8;

[0041] A4. Mutation operation: Gaussian mutation is adopted, and the mutation probability is 0.05;

[0042] A5. Iteration termination: when the iteration number reaches 100 generations or the fitness value converges, that is, the change rate is less than 0.01 for 10 consecutive generations, the optimal PID initial parameters are output.

[0043] Further, the monitoring and early warning module adopts a distributed data acquisition architecture, including an edge computing gateway and a cloud platform; the edge computing gateway is deployed in the on-site control cabinet of the unit, and a 5G industrial gateway is used to realize real-time data transmission, and the sampling frequency is 1 Hz; the cloud platform uses a Hadoop distributed file system (HDFS) to store historical data, and the storage period is not less than 1 year, and supports data backtracking and offline analysis.

[0044] Further, the monitoring and early warning module is provided with a three-level early warning mechanism, specifically:

[0045] First-level warning: for slight abnormality, when the parameter exceeds the normal range but is within the safety threshold, an audible and visual alarm is issued to prompt the operator to pay attention;

[0046] Second-level warning: for moderate abnormality, when the parameter exceeds the safety threshold but does not trigger the protection action, the control parameter is automatically adjusted, and the adjustment suggestion is pushed to the operator;

[0047] Third-level warning: for serious abnormality, when the parameter is close to the protection action value, the emergency shutdown plan is triggered, and the denitration ammonia injection system is cut off.

[0048] Technical effects and advantages of the present application:

[0049] Compared with the prior art, the deep peak regulation system suitable for thermal power units significantly improves the load prediction accuracy by improving the LSTM-attention model to fuse multi-dimensional features, and provides accurate load reference for deep peak regulation;

[0050] Fast control response: the dynamic weight MPC algorithm can adaptively adjust the control priority according to the load deviation, realize fast and undamped tracking of the load, and solve the response lag problem of traditional PID control;

[0051] More stable denitrification: The optimization model combining fuzzy PID and genetic algorithm can dynamically adapt to fluctuations in flue gas parameters, ensuring NOx compliance while significantly reducing ammonia slip rate and protecting catalyst performance;

[0052] High system integration: It achieves integrated load forecasting, coordinated control, denitrification optimization and monitoring and early warning, and the data of each module interacts in real time, improving the overall peak shaving efficiency. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the system flow of this application.

[0054] The attached diagram is labeled as follows: 1. Data acquisition module; 2. Multi-dimensional load prediction module; 201. Data preprocessing unit; 202. LSTM and attention mechanism fusion model construction unit; 3. Improved coordinated control module; 4. Adaptive denitrification optimization module; 401. Fuzzy inference unit; 402. Genetic algorithm optimization unit; 5. Monitoring and early warning module. Detailed Implementation

[0055] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0056] Example

[0057] As attached Figure 1 The system shown is a deep peak shaving system suitable for thermal power units, comprising a data acquisition module 1, a multi-dimensional load prediction module 2, an improved coordination control module 3, an adaptive denitrification optimization module 4, and a monitoring and early warning module 5. The modules interact with each other through a data interface. The system is characterized in that the data acquisition module 1 is used to collect the operating parameters of the thermal power unit in real time, and the output end of the data acquisition module 1 is electrically connected to the input end of the multi-dimensional load prediction module 2.

[0058] The operating parameters of the thermal power unit collected in data acquisition module 1 include, but are not limited to, boiler-side data, turbine-side data, denitrification system-side data, and environmental and dispatching-side data. Boiler-side data includes boiler outlet steam pressure, steam temperature, steam flow rate, fuel quantity, feedwater flow rate, and furnace pressure. Turbine-side data includes turbine speed, power generation, valve opening, and exhaust temperature. Denitrification system-side data includes denitrification system inlet flue gas temperature, flue gas flow rate, and NO₂ levels. x Concentration, ammonia slip rate, and ammonia injection rate; environmental and dispatch data include ambient temperature, power grid dispatch instructions, and fuel calorific value.

[0059] The data acquisition module 1 adopts a distributed architecture of "edge computing + cloud platform" to realize real-time acquisition, transmission and storage of unit operation data. The edge computing gateway is deployed in the on-site control cabinet of the boiler, steam turbine and denitration system, and communicates with the field sensors and actuators through the industrial Ethernet (Profinet protocol); the edge computing gateway adopts a 5G industrial gateway, the data transmission rate is not less than 100 Mbps, the sampling frequency is 1 Hz, and the real-time performance of data is ensured; the cloud platform uses HDFS to store historical data, the storage capacity of a single node is not less than 10 TB, supports data compression and distributed computing, and the storage period is not less than 1 year, thereby providing data support for subsequent model training and offline analysis.

[0060] The multi-dimensional load prediction module 2 realizes the prediction of unit load in the future 1-24 hours based on the collected operation parameters and power grid dispatching instructions through an improved long short-term memory network (LSTM) and attention mechanism fusion model.

[0061] The multi-dimensional load prediction module 2 includes a data preprocessing unit 201 and an LSTM and attention mechanism fusion model construction unit 202, wherein the data preprocessing unit 201 is used for preprocessing the collected original data, and the preprocessing strategy of the data preprocessing unit 201 is as follows:

[0062] Missing value filling: linear interpolation method is used to fill the missing data, and for data with continuous missing of more than 5 sampling points, the average value of the same period data in the previous 7 days is used for filling;

[0063] Abnormal value elimination: 3σ criterion is used to eliminate abnormal values, that is, when the data exceeds the range of [μ-3σ, μ+3σ], it is determined as an abnormal value, and the average value of the adjacent time data is used to replace it;

[0064] Data normalization: Min-Max normalization is used to map all feature parameters to the interval [0, 1], and the normalization formula is: Wherein, x is the original data, x min , x max are the minimum and maximum values of the data respectively, and x norm is the normalized data.

[0065] The LSTM and attention mechanism fusion model construction unit 202 is used to construct an LSTM and attention mechanism fusion model, and the LSTM and attention mechanism fusion model includes an input layer, an embedding layer, an improved LSTM layer, an attention layer, a full connection layer and an output layer, and specifically:

[0066] Input layer: receives 4 types of pre-processed feature parameters, and the 4 types of feature parameters are historical load, dispatching instruction, environmental temperature and fuel heat value, the input data dimension is (N, T, 4), wherein N is set to 10,000 historical samples, T is set to 24, that is, based on the previous 24 hours of data to predict the future 1 hour load, if the future 24 hours of load is predicted, T is expanded to 48;

[0067] Embedding layer: a fully connected neural network is used to map the input features to a high-dimensional feature space, and the mapping formula is: E = W e · X + b e , wherein E is an embedding feature matrix, W e is an embedding layer weight matrix, X is an input data matrix, and b e is an embedding layer bias term; specifically, a 32-dimensional fully connected neural network is used to realize feature mapping, the weight matrix W e has a dimension of (4, 32), the bias vector b e has a dimension of (1, 32), the ReLU activation function is used to enhance the feature nonlinear expression ability, and the mapping obtains an embedding feature matrix E with a dimension of (N, T, 32);

[0068] Improved LSTM layer: a temperature compensation factor a is introduced into the forgetting gate, and the forgetting gate calculation formula is modified as: f t = σ'(W f · [h t-1 , x t ] + b f + a· ΔT), wherein f t is the output of the forgetting gate at t, σ' is a sigmoid activation function, W f is a forgetting gate weight matrix, h t-1 is the hidden layer state at t-1, x t is the input feature at t, b f is the forgetting gate bias, ΔT is the deviation value of the environmental temperature from the rated temperature, and a is the temperature compensation coefficient and takes a value range of 0.01-0.05; specifically: 2 layers of LSTM structure are set, the number of hidden units of each layer is 64, and the temperature compensation factor a is introduced into the forgetting gate; when the environmental temperature is lower than the rated temperature (such as 25℃), ΔT is negative, the forgetting gate output is modified through a· ΔT, and the memory ability of the model to the load change under low temperature environment is enhanced; when the environmental temperature is higher than the rated temperature, ΔT is positive, the memory weight of the historical load data is appropriately weakened, and the response speed of the model to the real-time load fluctuation is improved. The improved LSTM layer finally outputs a hidden layer state matrix H with a dimension of (N, T, 64);

[0069] Attention layer: an additive attention mechanism is used to calculate the weight of each time step feature, and the weight calculation formula is: wherein at is the attention weight at time t, W a is the attention layer weight matrix, b a is the attention layer bias; specifically, the hidden layer state matrix H is weighted, and the weight a of each time step is calculated through the additive attention mechanism t and the dimension is (N, T, 1), and after weighting, the global feature vector with a dimension of (N, 64) is obtained where h t represents a certain variable at time node t;

[0070] The fully connected layer includes two layers, and the first layer adopts a dimension compression of 64→32, and the activation function is ReLU; the second layer adopts a dimension mapping of 32→1, and the activation function is Linear, and finally outputs the load prediction value y pred and the dimension is (N, 1).

[0071] The LSTM and attention mechanism fusion model is trained, specifically: the data set is divided into a training set, a validation set and a test set in a ratio of 7:2:1 according to the preprocessed data set, the training set is used for model parameter iterative update, the validation set is used for adjusting model hyperparameters (such as learning rate, number of hidden units), and the test set is used for evaluating model prediction accuracy;

[0072] The training parameters are set as follows: the Adam optimizer is used, the initial learning rate is 0.001, the learning rate decay strategy is to decay to 0.8 times of the original value every 50 epochs; the loss function adopts the root mean square error (RMSE) formula: where y true,i is the true value of the i-th sample, y pred,i is the value predicted by the model for the i-th sample; the training epoch is set to 200, and when the validation set loss does not decrease for 10 consecutive epochs, the training is stopped to prevent model overfitting;

[0073] The accuracy verification shows that the average absolute error (MAE) of the model for predicting the future 1-hour load is 1.2%, and the prediction accuracy (MAPE) is 1.8%; the average MAPE for predicting the future 24-hour load is 4.5%, which meets the design requirement of "prediction accuracy not less than 95%";

[0074] The improved coordinated control module 3 is used to receive the load prediction result and the real-time load deviation, adopt a dynamic weight adjustment algorithm based on model predictive control (MPC), and output control instructions of the boiler fuel quantity, the feedwater flow, and the turbine governor opening degree;

[0075] The prediction model adopts a method of mechanism and data fusion to construct a unit load prediction model, and the expression is: y(k+1|k) = f(y(k), u(k), d(k)), wherein y(k+1|k) is a load value at k+1 time based on the state prediction at k time, y(k) is an actual load at k time, u(k) is a control input vector at k time, the input vector includes fuel quantity, feed water flow, and valve opening, d(k) is a disturbance vector, the disturbance vector includes fuel heat value fluctuation and power grid frequency deviation, f(·) is a nonlinear prediction function, and the nonlinear prediction function is obtained by offline training of historical operation data;

[0076] The formula derivation process is as follows: the prediction model adopts a method of mechanism and data fusion to construct a unit load prediction model, combines a boiler heat balance equation and a steam turbine power equation mechanism formula, compensates errors of the mechanism model through a gradient boosting tree, that is, an XGBoost algorithm, and finally forms a nonlinear prediction function f(·), and the specific strategy is as follows:

[0077] The boiler heat balance equation is: Q in =D·(h sg -h fu )+Q loss , wherein Q in is fuel input heat, D is steam flow, h sg is steam enthalpy, h fu is feed water enthalpy, and Q loss is boiler heat loss;

[0078] The steam turbine power equation is: P = η t ·D·(h sg -h ex ), wherein η t is a relative internal efficiency of the steam turbine, and h ex is exhaust steam enthalpy;

[0079] The data compensation model is: the load prediction value y mech output by the mechanism model, the control input vector u(k), and the disturbance vector d(k) are input features, the difference between the actual load y act and y mech is an output label, the XGBoost compensation model y comp =g(y mech , u(k), d(k)) is trained, wherein g is a mapping function relationship learned by the XGBoost compensation model;

[0080] The fusion prediction model is: y(k+1|k) = y mech (k+1|k) + y comp (k+1|k), that is, y(k+1|k) = f(y(k), u(k), d(k)).

[0081] The rolling optimization phase in the dynamic weight MPC algorithm introduces a dynamic weight objective function, which is given by:

[0082] where J is the optimization objective function, P' is the prediction horizon, M is the control horizon, y ref (k+i) is the load reference value at time k+i, Δu(k+i) is the control input increment, ω1(k), ω2(k) are dynamic weight coefficients, and ω1(k)+ω2(k)=1;

[0083] The dynamic weight coefficients are adjusted adaptively according to the load deviation Δy=|y ref (k)-y(k)|, and the adjustment rule is:

[0084] When Δy>5%P e , ω1(k)=0.8, ω2(k)=0.2;

[0085] When 2%P e ≤Δy≤5%P e , ω1(k)=0.6, ω2(k)=0.4;

[0086] When Δy<2%P e , ω1(k)=0.4, ω2(k)=0.6;

[0087] where P e is the rated load of the unit.

[0088] The optimization strategy of the dynamic weight MPC algorithm is:

[0089] Prediction and control horizon setting: according to the load response characteristics of thermal power units (usually 30-60 seconds), set the prediction horizon P'=10 (corresponding to 10 seconds) and the control horizon (M=5) (i.e. update the control command every 5 seconds);

[0090] Constraint condition setting: considering the safe operation boundary of the unit equipment, set the control input and output constraints:

[0091] Fuel quantity constraint: u fuel,min ≤u fuel (k+i)≤u fuel,max , where u fuel,min is the minimum stable combustion fuel quantity, usually 30% of the rated fuel quantity, and u fuel,max is the rated fuel quantity;

[0092] Feedwater flow constraint: u fw,min ≤u fw (k+i)≤u fw,max , where u fw,minu is the minimum opening degree of the damper, and is matched with the minimum fuel amount fw,max u is the rated opening degree of the damper;

[0093] u is the minimum opening degree of the damper, and is matched with the minimum fuel amount valve,min u is the minimum opening degree of the damper, and is matched with the minimum fuel amount valve u is the minimum opening degree of the damper, and is matched with the minimum fuel amount valve,max u is the minimum opening degree of the damper, and is matched with the minimum fuel amount valve,min u is the minimum opening degree of the damper, and is matched with the minimum fuel amount valve,min u is the minimum opening degree of the damper, and is matched with the minimum fuel amount valve,max u is the minimum opening degree of the damper, and is matched with the minimum fuel amount valve,max u is the minimum opening degree of the damper, and is matched with the minimum fuel amount

[0094] u is the minimum opening degree of the damper, and is matched with the minimum fuel amount e u is the minimum opening degree of the damper, and is matched with the minimum fuel amount

[0095] u is the minimum opening degree of the damper, and is matched with the minimum fuel amount

[0096] u is the minimum opening degree of the damper, and is matched with the minimum fuel amount corr u is the minimum opening degree of the damper, and is matched with the minimum fuel amount corr u is the minimum opening degree of the damper, and is matched with the minimum fuel amount

[0097] u is the minimum opening degree of the damper, and is matched with the minimum fuel amount T u is the minimum opening degree of the damper, and is matched with the minimum fuel amount T u is the minimum opening degree of the damper, and is matched with the minimum fuel amount -1 u is the minimum opening degree of the damper, and is matched with the minimum fuel amount

[0098] u is the minimum opening degree of the damper, and is matched with the minimum fuel amount

[0099] u is the minimum opening degree of the damper, and is matched with the minimum fuel amount x u is the minimum opening degree of the damper, and is matched with the minimum fuel amount NOx,ref u is the minimum opening degree of the damper, and is matched with the minimum fuel amount NOx,act u is the minimum opening degree of the damper, and is matched with the minimum fuel amount u is the minimum opening degree of the damper, and is matched with the minimum fuel amount p, integral coefficient K i , differential coefficient K d The correction amount is output;

[0100] Fuzzy subset division: the fuzzy subsets of e and ec are both {NB, NM, NS, ZO, PS, PM, PB}, corresponding to the domain [-6, 6]; K p , K i , K d The fuzzy subsets of the correction amount are both

[0101] {NB, NM, NS, ZO, PS, PM, PB}, corresponding to the domains [-2, 2], [-0.5, 0.5], [-0.1, 0.1] respectively;

[0102] ;

[0103] The genetic algorithm optimization unit is used to optimize the initial parameters of the fuzzy PID controller, and the optimization target is to minimize the ammonia escape rate and to meet the NOx emission standard, and the fitness function is: F is the fitness value, λ1 and λ2 are weight coefficients, γ1 and γ2 are penalty coefficients, e is the ammonia escape rate, and C is the NOx emission value; ; x ;

[0104] The initial parameters K p0 , K i0 , K d0 of the fuzzy PID controller are optimized by the genetic algorithm, and the optimized initial parameters are input into the fuzzy PID controller to realize dynamic correction of the parameters;

[0105] The constraint condition is: considering the hardware limitation of the ammonia injection system, the ammonia injection amount constraint u NH3,min ≤u NH3 (k)≤u NH3,max , where u NH3,min = 0 kg / h, and u NH3,min is the minimum ammonia injection amount, u NH3,max = 500 kg / h is the maximum ammonia injection amount, and the NOx emission constraint is C x ≤50 mg / Nm NOx ;

[0106] The operation steps of the genetic algorithm include:

[0107] A1. Initialize the population: randomly generate 50 PID parameter combinations (K p0 , K i0 , K d0 ), and the parameter range is K p0 ∈[5, 15], K i0 ∈[0.5, 2], and K d0 ∈[0.1, 0.5].​i0 ∈[0.1,1.0], K d0 ∈[0.01,0.1] ;

[0108] A2. Selection operation: 30 individuals are selected into the next generation by roulette selection method;

[0109] A3. Cross operation: single-point crossover is adopted, and the crossover probability is 0.8;

[0110] A4. Mutation operation: Gaussian mutation is adopted, and the mutation probability is 0.05;

[0111] A5. Iteration termination: when the iteration number reaches 100 generations or the fitness value converges, that is, the change rate is less than 0.01 for 10 consecutive generations, the optimal PID initial parameters are output.

[0112] The monitoring and early warning module 5 monitors the running state of each module in real time, sends an early warning signal when the parameter exceeds the preset threshold, and generates an emergency control scheme.

[0113] The monitoring and early warning module 5 adopts a distributed data acquisition architecture, including an edge computing gateway and a cloud platform; the edge computing gateway is deployed in the unit local control cabinet, and realizes real-time data transmission by using a 5G industrial gateway, and the sampling frequency is 1Hz; the cloud platform uses a Hadoop distributed file system (HDFS) to store historical data, and the storage period is not less than 1 year, and supports data backtracking and offline analysis.

[0114] The monitoring and early warning module 5 is provided with a three-level early warning mechanism, specifically:

[0115] First-level warning: for slight abnormality, when the parameter exceeds the normal range but is within the safety threshold, an audible and visual alarm is sent out to prompt the operator to pay attention;

[0116] Second-level warning: for moderate abnormality, when the parameter exceeds the safety threshold but does not trigger the protection action, the control parameter is automatically adjusted, and the control suggestion is pushed to the operator;

[0117] Third-level warning: for serious abnormality, when the parameter is close to the protection action value, the emergency shutdown plan is triggered, and the denitration ammonia injection system is cut off.

[0118] Safety indicators: boiler water level (normal range ± 50mm), steam temperature (normal range 540±5℃), furnace pressure (normal range -100100Pa), ammonia escape rate (normal range ≤2ppm);

[0119] Stability indicators: load tracking error (normal range ≤±2%P e ), NOx concentration (normal range ≤50mg / Nm 3 ), control instruction fluctuation amplitude (fuel quantity fluctuation ≤±5% / min);

[0120] Economic indicators: power supply coal consumption (normal range ≤ 300 g / kWh), ammonia injection unit consumption (normal range ≤ 0.2 kg / MWh.

[0121] Early warning threshold setting: based on the historical operation data of the unit and the safety regulations, the three-level early warning thresholds of each indicator are set, for example, the first-level early warning threshold of the steam temperature is 535-545℃, the second-level early warning threshold is 530-550℃, and the third-level early warning threshold is 525-555℃.

[0122] Finally: the above is only a preferred embodiment of the present application and is not used to limit the present application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A deep peak shaving system suitable for a thermal power unit, comprising a data acquisition module (1), a multi-dimensional load prediction module (2), an improved coordinated control module (3), an adaptive denitration optimization module (4), and a monitoring and early warning module (5), the modules are connected through data interfaces to realize data interaction, characterized in that: The data acquisition module (1) is used for collecting the operation parameters of the thermal power unit in real time, and the output end of the data acquisition module (1) is electrically connected with the input end of the multi-dimensional load prediction module (2); The multi-dimensional load prediction module (2) realizes the load prediction of the unit in the future 1-24 hours based on the collected operation parameters and the power grid dispatching instruction through the improved long short-term memory network (LSTM) and attention mechanism fusion model; The improved coordinated control module (3) is used for receiving the load prediction result and the real-time load deviation, adopting a dynamic weight adjustment algorithm based on model predictive control (MPC), and outputting the control instructions of the boiler fuel quantity, the feedwater flow and the turbine governing valve opening degree; The self-adaptive denitration optimization module (4) dynamically adjusts the ammonia injection quantity of the denitration system through the optimization model combining the fuzzy PID and the genetic algorithm according to the boiler outlet flue gas parameters and the denitration system operation state; The monitoring and early warning module (5) monitors the operation state of each module in real time, issues an early warning signal when the parameters exceed the preset threshold, and generates an emergency control scheme.

2. The system according to claim 1, characterized in that: The operation parameters of the thermal power unit collected in the data acquisition module (1) include but are not limited to the boiler side data, the turbine side data, the denitration system side data and the environment and dispatching side data; wherein the boiler side data includes boiler outlet steam pressure, steam temperature, steam flow, fuel amount, feed water flow and furnace pressure; the steam turbine side data includes steam turbine rotating speed, power generation, governing valve opening and exhaust temperature; the denitration system side data includes denitration system inlet flue gas temperature, flue gas flow, NO x concentration, ammonia escape rate and ammonia injection amount; the environment and dispatching side data includes environment temperature, power grid dispatching instruction and fuel heat value.

3. The system according to claim 2, characterized in that: The multi-dimensional load prediction module (2) includes a data preprocessing unit (201) and an LSTM and attention mechanism fusion model construction unit (202), wherein the data preprocessing unit (201) is used for preprocessing the collected original data, and the preprocessing strategy of the data preprocessing unit (201) is as follows: Missing value filling: the missing data is filled by using the linear interpolation method, and for the data with continuous missing of more than 5 sampling points, the average value of the same period data in the previous 7 days is used for filling; Abnormal value elimination: the 3σ criterion is used to eliminate abnormal values, that is, when the data exceeds the range of [μ-3σ, μ+3σ], it is determined as an abnormal value, and the average value of the adjacent time data is used to replace it; Data normalization: all feature parameters are mapped to the interval [0, 1] by using Min-Max normalization, and the normalization formula is: Where x is the original data, x min , x max are the minimum and maximum values of the data, respectively, and x norm is the normalized data.

4. The system according to claim 3, characterized in that: The LSTM and attention mechanism fusion model construction unit (202) is used for constructing the LSTM and attention mechanism fusion model, and the LSTM and attention mechanism fusion model includes an input layer, an embedding layer, an improved LSTM layer, an attention layer, a full connection layer and an output layer, and specifically: The input layer receives the preprocessed four types of feature parameters, and the four types of feature parameters are historical load, dispatching instruction, environmental temperature and fuel heat value, the input data dimension is (N, T, 4), wherein N is set to 10000 historical samples, T is set to 24, that is, based on the previous 24 hours of data to predict the load in the future 1 hour, if the load in the future 24 hours is predicted, T is expanded to 48; Embedding layer: adopt full connection neural network to map input feature to high dimensional feature space, mapping formula is: E=W e ·X+b e , wherein, E is embedding feature matrix, W e is embedding layer weight matrix, X is input data matrix, b e is embedding layer bias item quantity; Improved LSTM layer: introduce temperature compensation factor a in the forget gate, the forget gate calculation formula is modified as: f t = σ'(W f · [h t-1 , x t ] + b f + a· ΔT), where f t is the output of the forget gate at time t, σ' is the sigmoid activation function, W f is the weight matrix of the forget gate, h t-1 is the hidden layer state at time t-1, x t is the input feature at time t, b f is the bias of the forget gate, ΔT is the deviation value of the ambient temperature and the rated temperature, and a is the temperature compensation coefficient and the value range is 0.01-0.05; Attention layer: the weight of each time step feature is calculated by using additive attention mechanism, and the weight calculation formula is: wherein, a t is the attention weight at time t, W a is the attention layer weight matrix, b a is the attention layer bias; Fully connected layer: there are two layers in total, and the first layer adopts 64→32 dimension compression, and the activation function is ReLU; the second layer adopts 32→1 dimension mapping, and the activation function is Linear, and finally outputs the load prediction value y pred and the dimension is (N, 1).

5. The system according to claim 4, characterized in that: The dynamic weight MPC algorithm in the improved coordinated control module (3) includes a prediction model, a rolling optimization and a feedback correction, wherein the prediction model adopts a mechanism and data fusion method to construct a unit load prediction model, and an expression is: y(k+1|k) = f(y(k), u(k), d(k)), wherein y(k+1|k) is a load value at k+1 time based on a state prediction at k time, y(k) is an actual load at k time, u(k) is a control input vector at k time, the input vector includes a fuel quantity, a feed water flow and a valve opening, d(k) is a disturbance vector, the disturbance vector includes fuel heat value fluctuation and power grid frequency deviation, f(·) is a nonlinear prediction function, and the function is obtained by offline training based on historical operation data.

6. The system according to claim 5, characterized in that: The rolling optimization phase in the dynamic weight MPC algorithm introduces a dynamic weight objective function, which is given by where J is the optimization objective function, P' is the prediction horizon, M is the control horizon, y ref (k+i) is the load reference value at time k+i, Δu(k+i) is the control input increment, ω1(k), ω2(k) are dynamic weight coefficients, and ω1(k)+ω2(k)=1. The dynamic weight coefficient is adjusted adaptively according to the load deviation Ay = |y ref (k)-y(k)| adaptive adjustment, and the adjustment rule is: When Δy > 5% P e ω1(k) = 0.8, ω2(k) = 0.2; when 2% P e ≤ Ay ≤ 5% P e when 2% P ω1(k) = 0.6, ω2(k) = 0.4; When Δy < 2% P e ω1(k) = 0.4, ω2(k) = 0.6; where P e is the rated load of the machine set.

7. The system according to claim 6, characterized in that: The feedback correction stage in the dynamic weight MPC algorithm corrects the prediction error by using Kalman filter, and the correction formula is: y corr (k+i|k) = y(k+i|k) + K(k) e(k), where y corr (k+i|k) is the corrected predicted load, K(k) is the Kalman gain, and e(k) = y(k) - y(k|k-1) is the prediction error at time k.

8. The system according to claim 7, characterized in that: The fuzzy PID-genetic algorithm model in the adaptive denitration optimization module (4) comprises a fuzzy reasoning unit (401) and a genetic algorithm optimization unit (402), the fuzzy reasoning unit takes the denitration inlet NO x Concentration deviation e = C NOx,ref -C NOx,act And deviation change rate As input, the proportional coefficient K p , integral coefficient K i , and differential coefficient K d of the PID controller are taken as input, and the correction amount is taken as output; Fuzzy subset partition: the fuzzy subsets of e and ec are both {NB, NM, NS, ZO, PS, PM, PB}, corresponding to the domain [-6, 6]; K p , K i , K d The fuzzy subsets of the correction amount are {NB, NM, NS, ZO, PS, PM, PB}, corresponding to the domains [-2, 2], [-0.5, 0.5], [-0.1, 0.1], respectively. The genetic algorithm optimization unit is used for optimizing initial parameters of the fuzzy PID controller, and an optimization target is to minimize ammonia escape rate and NO x The discharge standard is met, and the fitness function is: F is a fitness value, λ1 and λ2 are weight coefficients, γ1 and γ2 are penalty coefficients, is the ammonia escape rate, is the NO x emission standard value; The operation steps of the genetic algorithm include: A1. Initialize population: 50 PID parameter combinations (K p0 , K i0 , K d0 ) are randomly generated with parameter ranges K p0 ∈ [5, 15], K i0 ∈ [0.1, 1.0], K d0 ∈ [0.01, 0.1]; A2. Selection operation: 30 individuals are selected into the next generation by using a roulette selection method; A3. Cross operation: single-point cross is used, and the cross probability is 0.8; A4. Mutation operation: Gaussian mutation is used, and the mutation probability is 0.05; A5. Iteration termination: when the iteration number reaches 100 generations or the fitness value converges, that is, the change rate is less than 0.01 for 10 generations in succession, the optimal PID initial parameters are output.

9. The system according to claim 8, characterized in that: The monitoring and early warning module (5) adopts a distributed data acquisition architecture, includes an edge computing gateway and a cloud platform; the edge computing gateway is deployed in a unit on-site control cabinet, realizes real-time data transmission by using a 5G industrial gateway, and the sampling frequency is 1 Hz; the cloud platform stores historical data by using a Hadoop distributed file system (HDFS), the storage period is not less than 1 year, and data backtracking and offline analysis are supported.

10. The system according to claim 9, characterized in that: The monitoring and early warning module (5) is provided with a three-level early warning mechanism, specifically: Primary warning: for slight abnormality, when a parameter exceeds a normal range but is within a safety threshold, a sound and light alarm is given to prompt an operator to pay attention; Secondary warning: for moderate abnormality, when a parameter exceeds the safety threshold but does not trigger a protection action, a control parameter is automatically adjusted, and a regulation and control suggestion is pushed to the operator; Tertiary warning: for serious abnormality, when a parameter is close to a protection action value, an emergency shutdown plan is triggered, and a denitration ammonia injection system is cut off.

Images