Optimization control method of waste incineration power generation deacidification process

Abstract

The invention discloses an optimal control method and system for a waste incineration power generation deacidification process, and relates to the technical field of intelligent control, and the method comprises the steps: collecting the multi-source working condition information of a deacidification tower through arranging a plurality of sensors; preprocessing and fusing the multi-source working condition data to obtain a working condition feature vector and a working condition constraint condition; on the basis of the working condition feature vectors, utilizing an intelligent prediction model to predict the acid gas emission trend and the garbage heat value; optimal control parameters are calculated according to prediction results and working condition constraint conditions, and real-time flow and concentration adjustment is conducted on a spraying system in combination with online heat value feed-forward compensation and self-tuning PID feedback control. According to the method, the response speed and stability of the deacidification system to the furnace condition fluctuation are remarkably improved, the alkali liquor consumption and the emission fluctuation are reduced, and the environmental protection standard and the operation economy are guaranteed.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent control, and particularly relates to a method and system for optimizing control of a waste incineration power plant deacidification process. BACKGROUND

[0002] In recent years, with the acceleration of urbanization and the continuous growth of waste disposal capacity, waste incineration power generation technology has been widely used due to its large processing capacity, high power generation efficiency, and small land occupation. In particular, in the field of deacidification process, through the ways of spraying alkali solution and neutralization reaction, the acidic components in the incineration tail gas are removed, which has become a standard configuration of waste incineration power plants. The early deacidification system mainly relies on simple open-loop control and empirical parameter adjustment, which cannot respond to furnace condition fluctuations in real time. The subsequent feedback control based on PID enables the spraying amount and alkali concentration to track the emission concentration to some extent, but it is still difficult to make rapid compensation for sudden disturbances such as sharp changes in waste heat value and uneven temperature distribution in the furnace. In recent years, model predictive control (MPC) and online parameter identification have been introduced into the deacidification system by the academic and industrial communities, but most researches focus on single control loop or linear model, ignoring the fusion of multi-source working condition information, nonlinear characteristics and online heat value feedforward compensation, which is difficult to maintain optimal control effect in complex and variable operating environment.

[0003] Although the above-mentioned technologies improve the tracking accuracy and stability of the deacidification system to some extent, there are still some deficiencies: on the one hand, the sensor layout and data processing are mostly limited to single-point measurement and offline calibration, lacking the coordinated use of multi-source working condition information such as gas concentration, temperature distribution, slurry flow, pressure difference and vibration, resulting in insufficient working condition feature extraction; on the other hand, the existing control strategies mostly use fixed gain or PID control based on simplified process model, lacking feedforward compensation for changes in waste heat value and online self-tuning ability, when facing fluctuations in raw material composition and combustion conditions, the control performance decreases significantly, the alkali consumption increases and the emission of acid gases fluctuates. SUMMARY

[0004] In view of the above-mentioned existing problems, the present application is proposed.

[0005] Therefore, the present application provides a method for optimizing control of a waste incineration power plant deacidification process to solve the problems of insufficient utilization of working condition information, insufficient prediction and compensation ability, and lagging feedback regulation in the existing waste incineration power plant deacidification control technology.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] In a first aspect, the present application provides an optimization control method for waste incineration power generation deacidification process, which comprises collecting multi-source working condition information of the deacidification tower by arranging multiple sensors;

[0008] The multi-source working condition data is preprocessed and fused to obtain working condition characteristic vectors and working condition constraints;

[0009] Based on the working condition characteristic vectors, the intelligent prediction model is used to predict the emission trend of acid gases and the calorific value of waste;

[0010] According to the prediction results and the working condition constraints, the optimal control parameters are calculated, and the online calorific value feedforward compensation and self-tuning PID feedback control are combined to adjust the flow and concentration of the spraying system in real time.

[0011] As a preferred scheme of the optimization control method for the waste incineration power generation deacidification process, the multiple sensors arranged include electrochemical and optical laser absorption gas sensors arranged at the inlet, middle section and outlet of the deacidification tower; distributed optical fiber temperature sensing network is arranged in the inlet area, reaction area and cooling area, flow and pressure difference sensors are arranged before and after the slurry pipeline, and vibration and liquid level sensors are arranged at the atomizer oil tank and bearing for real-time acquisition of multi-source working condition information such as gas concentration, temperature distribution, slurry flow, pressure difference and vibration liquid level.

[0012] As a preferred scheme of the optimization control method for the waste incineration power generation deacidification process, the preprocessing and fusion include time alignment and synchronization of each sensor, detection and correction of abnormal values, and normalization and standardization after completion;

[0013] For the same type of redundant sensor data, the historical data and online signal-to-noise ratio are used to dynamically assign weights, and the weighted combination is performed to complete the preliminary fusion of multi-source working condition data;

[0014] The multi-source concentration and temperature readings of the sensors are respectively constructed into trustworthiness functions, the contradictory measurement values are removed through evidence synthesis rules to obtain a fusion confidence interval, and the fused scalar estimate is calculated to complete the secondary fusion of multi-source working condition data;

[0015] The fused scalar estimate is converted into a working condition characteristic vector, which specifically includes:

[0016] The instantaneous value is extracted, the trend index is calculated, the health level is encoded, the raw material characteristics and online calorific value estimation results are extracted.

[0017] As a preferred scheme of the optimization control method of the waste incineration power generation deacidification process, wherein: the intelligent prediction model comprises, collecting the working condition feature vector of the past L seconds from the current time, fitting the concentration sequence on the sliding time window by using the recursive least square method, and obtaining the current intercept β0 and the slope β1;

[0018] Recording the reference starting time t0 of the baseline, calculating the time offset, and calculating the linear trend value together with β0 and β1;

[0019] Subtracting the corresponding linear trend value from the observed concentration in the last L seconds to obtain a residual sequence with a length of L;

[0020] Performing TCN residual prediction, splicing the residual sequence with the slope β1 to form an L×2-dimensional input; sequentially passing through three layers of causal dilated convolution, the channel number is C ch , the convolution kernel width is K, the interlayer dilution rate is d1, d2, and d3 respectively, and each layer adopts residual connection and activation;

[0021] Adjusting the dilution rates of the middle and rear three layers online and dynamically, when the differential pressure index exceeds the threshold P thr , adding an increment Δd to each of the original d2 and d3;

[0022] The feature vector output by the last layer is mapped through full connection to generate a prediction sequence of future H seconds of residual; calculating a dynamic gating coefficient, and extracting gating features from the residual fluctuation amplitude of the last L seconds, the current filter screen differential pressure and the vibration index;

[0023] Inputting the gating features into a two-layer fully connected network to obtain the gating coefficients γ1…γ H for each prediction time;

[0024] Performing trend and residual fusion reconstruction, for each prediction time τ (1≤τ≤H), calculating the trend reference value T(τ) of the time by β0 and β1 and time offset;

[0025] Weighting and mixing the trend value and the residual prediction value by the gating coefficient γ τ to obtain the trend contribution ratio γ τ , and the residual contribution ratio 1-γ τ ;

[0026] Performing sliding smoothing and upper and lower limit clipping on the fusion result to output the final H-second prediction sequence.

[0027] As a preferred scheme of the optimization control method of the waste incineration power generation deacidification process, wherein: the prediction of the emission trend and the waste heat value comprises: using an autoregressive submodule to extract the linear trend parameters of the recent window, and performing detrending on the concentration component; inputting the detrended concentration residual and the current trend slope into a deep time sequence network to predict the future short-term concentration residual;

[0028] Through the gating network, the linear trend and the nonlinear residual are dynamically balanced according to the residual fluctuation amplitude, the pressure difference and the vibration and the like, and the two are superimposed to obtain an emission trend prediction curve;

[0029] The near-infrared spectrum feature is introduced, and is spliced with the temperature and flow characteristics to form an extended input, the extended feature is mapped, and the heat value estimation at the next moment is output; through a sliding window mode, the actual measured heat value each time is fed back to the regression network with a prediction error, and online incremental updating is performed, and the heat value prediction at the next moment and in the short term is output.

[0030] As a preferred scheme of the optimization control method of the waste incineration power generation deacidification process, wherein: the calculation of the optimal control parameters comprises: in each control period, collecting the working condition characteristic vector and the working condition constraint condition to obtain the current state and the measurement vector; the AR-TCN prediction model outputs the disturbance and the emission trend prediction and the residual statistics in the future H steps;

[0031] Based on the residual statistics, K disturbance paths are sampled from the predicted trajectory; a scenario tree is grown with a branch factor b in the first L steps to form multiple possible evolution paths; S representative scenarios are obtained by scenario reduction, and probability weights are assigned; the key boundary is tightened according to the robustness margin p;

[0032] The AR-TCN prediction model is used for rolling simulation, a process model that can be updated online is selected, and each scenario is rolled forward within H; the index values of the working condition characteristic vector are calculated at each node;

[0033] The objective function is obtained by weighting multiple terms, including: an emission deviation term, a reagent consumption term, an action smoothing term, and a risk term;

[0034] The working condition constraint condition comprises: a control constraint, a process constraint, an emission soft constraint, and an opportunity constraint.

[0035] As a preferred scheme of the optimization control method of the waste incineration power generation deacidification process, wherein: the online heat value feedforward compensation and the self-tuning PID feedback control select a QP solver according to the model and the constraint, and are started up with the solution of the previous control period;

[0036] Output the whole control sequence with length H within the given solving time limit, execute the first-step control quantity, and combine the feedforward quantity deviation generated according to the online heat value estimation with the first-step control quantity to synthesize the reference instruction;

[0037] The emission error residual is slightly corrected through the self-tuning PID, the final first-step control quantity is sent to the actuator through the field bus, and the steps of calculating the optimal control parameters are repeated every control period, if the solving fails, the degradation strategy is triggered, and the time domain is shortened to H ’ , adjust p until the feasibility is restored.

[0038] In the second aspect, the present application provides an optimization control system for a waste incineration power generation deacidification process, comprising a multi-source sensing and collecting module for completing high-frequency synchronous collection, timestamp marking, preliminary denoising and abnormal coarse screening; a working condition feature construction module for obtaining fused scalar estimation and working condition feature vectors, constraint side information; an intelligent prediction module for obtaining a future short-term emission trend prediction sequence and heat value prediction; and a feedforward-feedback integrated control module for parameter correction through self-tuning PID and output of an execution instruction.

[0039] In the third aspect, the present application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, any step of the optimization control method for the waste incineration power generation deacidification process according to the first aspect of the present application is implemented.

[0040] In the fourth aspect, the present application provides a computer readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, any step of the optimization control method for the waste incineration power generation deacidification process according to the first aspect of the present application is implemented.

[0041] The present application has the following beneficial effects: multi-channel sensors are used to collect real-time multi-source working condition information such as deacidification tower gas concentration, temperature, flow, pressure difference, etc., and layered weighting and evidence fusion are used to improve data reliability; a method combining recursive least squares and TCN residual prediction is used to realize high-precision second-level prediction of the emission trend of acid gas; near-infrared spectrum and online incremental regression are introduced to complete real-time estimation and feedforward compensation of the heat value of waste; based on the disturbance statistics and scene tree robust MPC output by the AR-TCN model, the QP solver warm start and self-tuning PID feedback are combined to form feedforward-feedback fusion control. The response speed and stability of the deacidification system to furnace condition fluctuations are significantly improved, the alkali consumption and emission fluctuations are reduced, and the environmental protection standard and operation economy are ensured. BRIEF DESCRIPTION OF DRAWINGS

[0042] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments description. Obviously, the drawings in the following description only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0043] Figure 1 The flow chart of the waste incineration power generation deacidification process optimization control method. DETAILED DESCRIPTION

[0044] In order to make the above-mentioned objects, features and advantages of the present application more apparent and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0045] In the following description, a large number of specific details are set forth in order to facilitate a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the connotation of the present application, therefore the present application is not limited by the specific embodiments disclosed below.

[0046] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an independent or alternative embodiment that excludes other embodiments.

[0047] Reference Figure 1 For one embodiment of the present application, the embodiment provides a waste incineration power generation deacidification process optimization control method, comprising the following steps:

[0048] S1: Collecting multi-source working condition information of the deacidification tower by arranging multiple sensors.

[0049] Further, the arrangement of multiple sensors includes arranging electrochemical and optical laser absorption gas sensors at the inlet, middle section and outlet of the deacidification tower; arranging a distributed optical fiber temperature sensing network in the inlet area, reaction area and cooling area, arranging flow and pressure difference sensors before and after the slurry pipeline, and arranging vibration and liquid level sensors at the atomizer oil tank and bearing, for real-time acquisition of gas concentration, temperature distribution, slurry flow, pressure difference and vibration liquid level multi-source working condition information.

[0050] Further, gas monitoring, electrochemical SO2 / HCl sensors (precision ±1mg / m 3 ) are arranged at the inlet of the deacidification tower, and one more optical laser absorption sensor is arranged in the tower body and at the outlet, for mutual redundancy and verification.

[0051] Slurry and filter screen monitoring, differential pressure sensor (0-100 kPa) and ultrasonic filter screen blockage monitor are arranged on the upstream and downstream of the slurry pipeline; at the same time, electromagnetic flowmeter (±0.5%) is retained for flow calibration.

[0052] Vibration and liquid level monitoring, three-axis acceleration sensor (sampling frequency 100 Hz) and ultrasonic liquid level meter are installed at the bottom of the atomizer oil tank and the bearing to realize real-time sensing of returned slurry and vibration.

[0053] Temperature control, FBGT optical fiber Bragg grating distributed temperature sensing network is arranged at the inlet of the deacidification tower, reaction and cooling sections to realize high-density temperature monitoring (sampling frequency 5 Hz).

[0054] It should be noted that based on historical deacidification operation data, a digital twin model of the deacidification tower is constructed, covering gas-solid coupling reaction, temperature field and flow field simulation. A lightweight prediction model (MPC combined with self-tuning PID) is pre-installed in the edge gateway, and an online heat value estimation model is loaded.

[0055] S2: Preprocessing and fusion of the multi-source working condition data to obtain working condition feature vectors and working condition constraints.

[0056] The preprocessing and fusion includes time series alignment and synchronization of each sensor, detection and correction of abnormal values, and normalization and standardization after completion.

[0057] Signal decomposition and denoising, including wavelet packet decomposition: decomposing SO2 / HCl concentration, flow, pressure difference and other signals into multi-scale subbands, removing high-frequency noise and retaining key fluctuation components.

[0058] Empirical mode decomposition, IMF component extraction is performed on temperature gradient and vibration index to eliminate the influence of turbulent flow and random vibration.

[0059] Dynamic time warping is used to correct the time difference between different sensors to ensure the consistency of the fused features in physical time. For 1Hz data with missing values, spline interpolation based on adjacent 5s is used to ensure continuity.

[0060] Normalization and adaptive standardization, adaptive batch standardization is applied to features of different dimensions, and the mean and variance are updated online, taking into account both stationary and sudden change scenarios.

[0061] For redundant sensor data of the same type, dynamic weight allocation is performed according to historical data and online signal-to-noise ratio to perform weighted combination and complete preliminary fusion of multi-source working condition data.

[0062] The trustworthiness function is constructed for multi-source concentration and temperature readings of the sensor respectively, and the mutually contradictory measurement values are removed by evidence combination rules to obtain a fused confidence interval, and a combined scalar estimate is calculated to complete secondary fusion of multi-source working condition data.

[0063] The fused scalar estimate is converted into a working condition feature vector, specifically including:

[0064] The instantaneous value is extracted, the trend index is calculated, the health level is encoded, the raw material features and online heat value estimation results are extracted. The fused "scalar estimate" is converted into "multi-dimensional features" required for decision-making.

[0065] The instantaneous value is extracted, including SO2HCl fused concentration, three-stage temperature average value, filter pressure difference, spray flow, vibration amplitude.

[0066] The trend index is calculated, including the average value of the last 5 minutes and the difference of the last 5 minutes, temperature gradient (reaction zone-inlet zone), pressure difference change amplitude (past 1 minute)

[0067] The health level is encoded, including filter clogging (light, medium, heavy), atomizer vibration (normal, warning, alarm). Raw material features and online heat value estimation results.

[0068] It should be noted that feature extraction is serially dependent on fusion output - the "fused" scalar must be used as input to calculate trends, gradients and levels.

[0069] S3: Based on the working condition feature vector, an intelligent prediction model is used to predict the trend of acid gas emission and the heat value of the garbage.

[0070] The intelligent prediction model includes collecting the working condition feature vectors of the past L seconds from the current time, fitting the concentration sequence on the sliding time window using the recursive least squares method, and obtaining the current intercept β0 and slope β1.

[0071] The reference starting time t0 is recorded, the time offset is calculated, and the linear trend value is calculated together with β0 and β1.

[0072] The observed concentration in the last L seconds is subtracted from the corresponding linear trend value shift to obtain a residual sequence of length L.

[0073] The TCN residual prediction is performed, the residual sequence and the slope β1 are spliced to form an L×2-dimensional input; sequentially through three layers of causal dilated convolution, the channel number is C ch , the convolution kernel width is K, the inter-layer dilution rate is d1, d2, d3 respectively, and residual connection and activation are used in each layer.

[0074] The dilution rates of the middle and rear three layers are adjusted dynamically online when the pressure difference index exceeds the threshold Pthr At this time, the increments Δd are added to the original d2 and d3 respectively.

[0075] The feature vector output by the last layer is mapped by full connection to generate a prediction sequence of the future H-second residual; a dynamic gating coefficient is calculated from the residual fluctuation amplitude of the last L seconds, the current filter pressure difference and the vibration index to extract the gating features.

[0076] The gating features are input into a two-layer fully connected network to obtain the gating coefficients γ1…γ H .

[0077] Trend and residual fusion reconstruction is performed, and for each prediction time τ (1≤τ≤H), the trend reference value T(τ) at this time is calculated using β0 and β1 and the time offset.

[0078] The gating coefficients γ τ are used to weight the trend value and the residual prediction value to obtain the trend contribution ratio γ τ and the residual contribution ratio 1-γ τ .

[0079] The fusion result is further smoothed and clipped to output the final H-second prediction sequence.

[0080] The specific implementation process is represented as:

[0081] Let the window length L=60 seconds, the prediction step H=30 seconds, the TCN network parameters be three layers of dilated convolution with channel number 16 and convolution kernel width 3, the initial dilated rate [d1,d2,d3]=[1,2,4], the AR model parameters be the parameter vector θ AR =[β0,β1], the RLS covariance matrix be initialized as a unit matrix multiplied by a constant, the decay factor ρ=0.99, the gating network parameters be two layers of full connection with 32 units of ReLU in the first layer and 30 units of Sigmoid in the second layer, the learning rate of TCN α TCN =1×10 -4 , and the learning rate of the gating network α gate =1×10 -3 .

[0082] The online prediction process is performed once per second, assuming that the current time is t. From t-L+1 to t seconds, the fused SO2 / HCl concentration C obs (k), temperature gradient, pressure difference, flow, vibration peak value, online heat value, etc. are collected to form a matrix

[0083] The Col1 (concentration) is slidingly standardized by normalizing it with the mean and standard deviation of the past one hour; the other columns are similarly standardized.

[0084] Trend decomposition (AR fitting) is performed, and the concentration data from k = t - 300 + 1 to t are taken, with the index set as {C obs (k)}. A linear model is fitted using the least squares method, and the formula is expressed as:

[0085] C lin (k) = β0 + β1(k - (t - 300 + 1))

[0086] The current β0, β1 are obtained. The reference time t0 = t - L + 1 is recorded.

[0087] For each time k ∈ [t0, t] in the window, the detrended residual is calculated, and the formula is expressed as:

[0088] r(k) = C obs (k) - [β0 + β1(k - (t - 300 + 1))]

[0089] The residual sequence is calculated, and the formula is expressed as:

[0090]

[0091] TCN residual prediction is performed, input splicing, and r t is expanded to L × 1, and is spliced with the corresponding 60 × 1 "current slope broadcast column" [β1, …, β1] to form an L × 2 input.

[0092] Three-layer causal dilated convolution: for the first layer, use the dilation rate d1; output dimension: L × 16. Residual connection: normalize and ReLU after adding the input and the output. Similarly, the second and third layers use dilation rates d2 and d3, respectively, and each layer does residual connection. Dynamic adjustment is performed, and if the current filter pressure difference > 50 kPa, then d2 and d3 are each increased by 1 before entering the second and third layers.

[0093] The last row of the output of the third layer (the 16-dimensional vector corresponding to time t) is taken as the hidden layer representation h t . Through a fully connected layer (16 → 30), h t is mapped to the residual prediction

[0094] Dynamic gating coefficient calculation, gating input features: standard deviation of residuals in the last 60 s, current filter pressure difference, current vibration peak value.

[0095] Forward calculation: first layer: 32 units, ReLU. Second layer: 30 units, Sigmoid. Output γ ∈ (0, 1) ^ {30}.

[0096] Fusion reconstruction and post-processing, for τ = 1 … H, calculate the linear trend value, and the formula is expressed as:

[0097] T(t) = b0 + b1 ((t + T) - (t - 300 + 1))

[0098] The fusion is performed, and the formula is expressed as:

[0099]

[0100] A three-point moving average is performed on (t + 1:t + H). All values are clipped to [0, 20] mg / m 3 The parameters of the model are updated online: the AR parameters are replaced by the newly fitted b0 and b1 at each time, and the covariance is updated by RLS without additional batch training.

[0101] TCN and gating network: every 3600 iterations (about 1 hour), the true residual and the predicted residual of the last 3600 time points are taken, and a small batch back propagation is performed on the TCN network and the gating network respectively. If the average error (MAE) increases after updating for 30s, it is rolled back to the last snapshot.

[0102] The prediction of the acid gas emission trend and the waste heat value includes extracting a recent window linear trend parameter by using an autoregressive submodule, and performing detrending on a concentration component; inputting the detrended concentration residual and a current trend slope into a deep time series network to predict a future short-term concentration residual.

[0103] The emission trend prediction curve is obtained by superimposing the linear trend and the nonlinear residual through the gating network according to the residual fluctuation amplitude, differential pressure and vibration and the like.

[0104] The near-infrared spectral feature is introduced, which is spliced with the temperature and flow characteristics to form an extended input. The extended feature is mapped to output the heat value estimation of the next time. The actual measured heat value and the prediction error are fed back to the regression network in a sliding window manner for online incremental updating to output the next time and short-term heat value prediction.

[0105] It should be noted that the linear trend extraction and detrending of the concentration sequence by the autoregressive submodule can effectively eliminate the overall drift caused by changes in the combustion load or the air inlet condition, so that the subsequent deep time series network only needs to focus on capturing the short-term nonlinear fluctuation component. The gating network dynamically allocates the weights of the linear trend and the network residual according to the residual fluctuation amplitude, filter differential pressure and atomizer vibration and the like, which not only retains the large-scale emission trend, but also takes into account the microscopic disturbance, so as to obtain a smooth and sensitive emission trend prediction curve. The online sliding window incremental updating mechanism of the regression network can not only correct the deviation of the heat value estimation model by using the latest measurement value, but also quickly adapt to the fluctuation of the waste composition or moisture, so as to ensure the long-term stability and reliability of the heat value prediction.

[0106] S4: Calculate the optimal control parameters according to the prediction results and the working condition constraints, combine online heat value feedforward compensation with self-tuning PID feedback control, and adjust the flow and concentration of the spraying system in real time.

[0107] Further, the calculation of the optimal control parameters includes, in each control period, collecting the working condition characteristic vector and the working condition constraints to obtain the current state and measurement vector; outputting the disturbance and emission trend prediction and residual statistics of the future H steps by the AR-TCN prediction model.

[0108] Based on the residual statistics, K disturbance paths are sampled from the predicted trajectory; a scenario tree is grown with a branch factor b for the first L steps, forming multiple possible evolution paths; S representative scenarios are obtained by scenario reduction and are assigned probability weights; and the key boundary is tightened according to the robustness margin p.

[0109] Rolling simulation is performed using the AR-TCN prediction model, a process model that can be updated online is selected, and each scenario is rolled forward within H; the index values of the working condition characteristic vector are calculated at each node.

[0110] The objective function is obtained by weighting multiple terms, including: emission deviation term, chemical consumption term, action smoothing term, and risk term.

[0111] The working condition constraints include: control constraints, process constraints, emission soft constraints, and opportunity constraints.

[0112] Initialization, set the control step size Δt (e.g., 60s) and the prediction time domain H (e.g., 30 steps).

[0113] Set the number of scenarios M (e.g., 20 trajectories). Scene tree generation: use TCN-AR to predict M emission and heat value trajectories for the future H steps, and construct an M-branch scenario tree according to the probability branching; each branch node is assigned a generation probability under the historical error companion.

[0114] Construct the optimization problem, and the decision variable formula is represented as:

[0115] {u flow (k,s),u conc (k,s)} k=0…H-1,s=1…M

[0116] Where s represents the scenario index. M represents the number of scenarios, representing different disturbance trajectories generated by the TCN-AR model. k represents the prediction step index in the scenario tree. u flow (k,s) represents the spraying flow control amount corresponding to scenario s at prediction step k. u conc (k,s) represents the slurry concentration control amount corresponding to scenario s at prediction step k.

[0117] Non-preemptive constraints: for all scenario indices s that have not diverged before step k, enforce:

[0118] u flow (k,s)=u flow (k,s′),u conc (k,s)=u conc (k,s′)

[0119] For all s,s' belonging to the same parent node.

[0120] Cost function, formulated as:

[0121]

[0122] where π s represents the probability of scenario s. ΔC(k,s) represents the emission deviation. V(k,s) represents the vibration energy consumption estimate. w1,w2,w3 represent the corresponding weights of each term.

[0123] Node constraints, impose upper limits on spray pressure, temperature, pressure difference, and vibration for all (k,s). Write the above objective and constraints in standard QP format, call the solver, and obtain the optimal values of u flow and u conc

[0124] The online heat value feedforward compensation and self-tuning PID feedback control selects a QP solver based on the model and constraints, and is started up with the solution of the previous control period.

[0125] Output the entire control sequence with length H within the given solving time limit, execute the first-step control quantity; feedforward-feedback synthesis and execution, generate the feedforward quantity deviation based on the online heat value estimate, and combine it with the first-step control quantity to synthesize the reference command.

[0126] Correct the emission error residual by self-tuning PID in small amplitude; issue the final first-step control quantity to the actuator through the field bus; repeat the steps of calculating the optimal control parameters every control period, if the solving fails, trigger the degradation strategy, and shorten the time domain to H ’ , adjust ρ, until the feasibility is restored.

[0127] ​It should be noted that in the scenario tree / SMPC solving process, first, representative disturbance paths are sampled from the residual statistics of the AR-TCN, and the size of the tree is controlled by the branching factor, and then a number of key scenarios are selected by scenario reduction techniques, the occurrence probability is assigned, and the robustness margin tightening is applied to important boundaries to balance the calculation efficiency and risk protection; After that, the process model is rolled in each scenario node, the working condition characteristic index is calculated, and these indexes are used to construct a multi-objective weighted cost function to realize the "emission compliance priority-minimum drug consumption-smooth action-risk controllable" step-by-step optimization. It should be noted that the combination of online heat value feedforward and self-tuning PID feedback not only utilizes feedforward compensation to offset the load disturbance caused by heat value mutation, but also corrects the residual emission error through PID, so as to meet the process and emission soft constraints while realizing the fastest and smoothest spraying response; If the quadratic programming (QP) cannot obtain a feasible solution within the given solving time limit, the system automatically triggers the degradation strategy-shorten the prediction time domain and limit the action amplitude, then adjust the model prediction step, until the feasibility is restored, to ensure the continuous and stable operation of the deacidification system.

[0128] The following is an embodiment of the present application, which provides an optimization control method for a waste incineration power plant deacidification process. In order to verify the beneficial effects of the present application, economic benefit calculation and simulation experiments are carried out for scientific demonstration.

[0129] A waste incineration plant reconstruction project.

[0130] Before reconstruction: SO2 hourly average set to 8 mg / m 3 , actual SO2 hourly average fluctuates greatly within 6-10 mg / m 3 , atomizer oil tank returns slurry 1 time / 3 months, filter screen clogs 1 time / day, and lime consumption is 9.2 kg / ton of waste.

[0131] After reconstruction: SO2 hourly average set to 8 mg / m 3 , actual SO2 hourly average is stable at 7.8 mg / m3; atomizer oil tank returns slurry 1 time / year; filter screen clogging frequency is reduced to 0.3 times / day; lime consumption is reduced to 6.9 kg / ton of waste.

[0132] Low load working condition test experiment.

[0133] Test conditions: waste treatment capacity 60% load, heat value fluctuation ±15%.

[0134] Test results: SO2 hourly average is stable at 7.8 mg / m 3 , atomizer vibration value <130 μm / s; smoke emission compliance rate is improved from 78% to 99.2%, and equipment failure-free operation time is >2000 hours.

[0135] The embodiment also provides an optimization control system of the waste incineration power generation deacidification process, comprising: a multi-source perception acquisition module, which is used for completing high-frequency synchronous acquisition, time stamp marking, preliminary denoising and abnormal rough screening.

[0136] A working condition feature construction module is used for obtaining fused scalar estimation and working condition feature vectors and constraint side information.

[0137] An intelligent prediction module is used for obtaining a future short-term emission trend prediction sequence and a calorific value prediction.

[0138] A feedforward-feedback integrated control module is used for parameter correction through self-tuning PID and output of an execution instruction.

[0139] The embodiment also provides a computer device suitable for the optimization control method of the waste incineration power generation deacidification process, comprising: a memory and a processor; the memory is used for storing computer executable instructions, and the processor is used for executing the computer executable instructions to realize the optimization control method of the waste incineration power generation deacidification process proposed in the above embodiment.

[0140] The computer device can be a terminal, and the computer device comprises a processor, a memory, a communication interface, a display screen and an input device which are connected through a system bus. The processor of the computer device is used for providing computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for running of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used for wired or wireless communication with external terminals. The wireless communication can be realized through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or can be a key, a trackball or a touchpad arranged on the shell of the computer device, or can be an external keyboard, a touchpad or a mouse and the like.

[0141] The embodiment also provides a storage medium on which a computer program is stored, the computer program being executed by a processor to implement the optimization control method for realizing the acid removal process of waste incineration power generation as proposed in the above embodiment; the storage medium can be realized by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk, or an optical disk.

[0142] To sum up, the present application is realized by: collecting multi-source working condition information of the acid removal tower through arranging multiple sensors; pre-processing and fusing the multi-source working condition data to obtain a working condition feature vector and a working condition constraint condition; based on the working condition feature vector, using an intelligent prediction model to predict the acid gas emission trend and the waste heat value; calculating optimal control parameters according to the prediction result and the working condition constraint condition, and combining online heat value feedforward compensation and self-tuning PID feedback control to adjust the flow and concentration of the spraying system in real time.

[0143] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, and all of them should be covered in the scope of the claims of the present application.

Claims

1. An optimization control method for waste incineration power generation deacidification process, characterized in that: The application relates to a method for real-time monitoring and control of a spray system of a deacidification tower. The method comprises the following steps: arranging multiple sensors to collect multi-source working condition information of the deacidification tower; preprocessing and fusing the multi-source working condition data to obtain working condition characteristic vectors and working condition constraint conditions; predicting the acid gas emission trend and the waste heat value by using an intelligent prediction model based on the working condition characteristic vectors; calculating optimal control parameters according to the prediction results and the working condition constraint conditions, and combining online heat value feedforward compensation with self-tuning PID feedback control to adjust the flow rate and concentration of the spray system in real time.

2. The method of claim 1, wherein the method is characterized by: The arrangement of multiple sensors comprises the following steps: arranging electrochemical and optical laser absorption gas sensors at the inlet, middle section and outlet of the deacidification tower; arranging a distributed optical fiber temperature sensing network in the inlet area, reaction area and cooling area; arranging flow rate and pressure difference sensors before and after the slurry pipeline; and arranging vibration and liquid level sensors at the atomizer oil tank and bearings, so as to obtain multi-source working condition information of gas concentration, temperature distribution, slurry flow rate, pressure difference and vibration liquid level in real time.

3. The method of claim 2, wherein the method is characterized by: The preprocessing and fusing comprise the following steps: performing time sequence alignment and synchronization on each sensor, detecting and correcting abnormal values, and then performing normalization and standardization; for the same type of redundant sensor data, dynamically assigning weights according to historical data and online signal-to-noise ratio, performing weighted combination, and completing preliminary fusion of the multi-source working condition data; respectively constructing a confidence function for the multi-source concentration and temperature readings of the sensors, removing mutually contradictory measurement values by using an evidence synthesis rule to obtain a fused confidence interval, calculating a fused scalar estimate, and completing secondary fusion of the multi-source working condition data; the fused scalar estimate is converted into a working condition characteristic vector, which specifically comprises the following steps: extracting instantaneous values, calculating trend indicators, performing health level coding, extracting raw material features and online heat value estimation results.

4. The method of claim 3, wherein the method is characterized by: The intelligent prediction model comprises the following steps: collecting working condition characteristic vectors of the past L seconds from the current time, fitting a concentration sequence on a sliding time window by using a recursive least squares method, and obtaining a current intercept beta0 and a slope beta1; recording a reference starting time t0, calculating a time offset, and calculating a linear trend value together with beta0 and beta1; subtracting the corresponding linear trend value from the observed concentration in the last L seconds to obtain a residual sequence with a length of L; The TCN residual prediction is performed, the residual sequence is spliced with the slope β1 broadcast to form an L×2-dimensional input; and sequentially passing through three layers of causal dilated convolution, the channel number is C ch , the convolution kernel width is K, the interlayer dilution rate is d1, d2 and d3 respectively, and each layer adopts residual connection and activation; The expansion rates of the middle, rear three layers are dynamically adjusted online, and when the pressure difference index exceeds the threshold value P thr , each of d2 and d3 is increased by an increment Δd; the feature vector output by the last layer is mapped through full connection to generate a prediction sequence of future H-second residuals; calculating a dynamic gating coefficient, and extracting gating features from the residual fluctuation amplitude of the last L seconds, the current filter screen pressure difference and vibration indicators; The gating feature is input into a two-layer fully connected network to obtain gating coefficients γ1…γ H ; performing trend and residual fusion reconstruction, and calculating a trend reference value T(τ) of each prediction time τ (1 <= tau <= H) by using beta0, beta1 and the time offset; With the gating coefficient γ τ The trend value and the residual prediction value are weighted and mixed to obtain a trend contribution ratio γ τ , and a residual contribution ratio 1-γ τ ; performing sliding smoothing and upper and lower limit clipping on the fusion result, and outputting a final H-second prediction sequence.

5. The method of claim 4, wherein the method is characterized by: The prediction of the acid gas emission trend and the waste heat value comprises the following steps: extracting the linear trend parameters of the recent window by using an autoregressive submodule, and performing detrending processing on the concentration component; inputting the detrended concentration residual and the current trend slope into a deep time sequence network to predict future short-term concentration residuals; dynamically balancing the linear trend and the nonlinear residual by using a gating network according to the residual fluctuation amplitude, the pressure difference and the vibration and the like, and superimposing the linear trend and the nonlinear residual to obtain an emission trend prediction curve. The near-infrared spectrum feature is introduced, and is spliced with temperature and flow characteristics to form an extended input. The extended feature is mapped, and a heat value estimation at the next moment is output. Through a sliding window mode, the actual measured heat value and prediction error are fed back to the regression network each time, and online incremental updating is performed to output heat value prediction at the next moment and in the short term.

6. The method of claim 5, wherein the method is characterized by: The computing optimal control parameters comprises, in each control period, collecting a working condition characteristic vector and a working condition constraint condition to obtain a current state and a measurement vector; a future H-step disturbance and emission trend prediction and residual error statistics are output by an AR-TCN prediction model; K disturbance paths are sampled from the predicted trajectory based on the residual error statistics; a scenario tree is grown with a branch factor b in the first L steps to form multiple possible evolution paths; S representative scenarios are obtained by scenario reduction, and probability weights are assigned; and a key boundary is tightened according to a robustness margin p; Rolling simulation is performed using the AR-TCN prediction model, a process model that can be updated online is selected, and each scenario is rolled forward within H; The index values of the working condition characteristic vector are calculated at each node; The objective function is obtained by weighting multiple items, including: an emission deviation item, a reagent consumption item, an action smoothing item, and a risk item; The working condition constraint conditions include: control constraints, process constraints, emission soft constraints, and opportunity constraints.

7. The method of claim 6, wherein the method further comprises: determining the amount of the acid to be injected into the flue gas based on the amount of the acid injected into the flue gas in the previous cycle and the amount of the acid injected into the flue gas in the current cycle. The online heat value feedforward compensation and self-tuning PID feedback control select a QP solver according to the model and the constraint, and are started with the solution of the previous control period; A control sequence with a length of H is output within a given solving time limit, and the first-step control amount is executed; feedforward-feedback synthesis and execution are performed, and a feedforward amount deviation is generated according to the online heat value estimation to synthesize a reference instruction with the first-step control amount; The small correction of the discharge error residual is performed through the self-tuning PID, the final first-step control quantity is issued to the actuator through the field bus, the step of calculating the optimal control parameter is repeated every control period, if the solution fails, the degradation strategy is triggered, and the time domain is shortened to H ’ , the feasibility is recovered.

8. An optimization control system of a waste incineration power plant deacidification process, based on the optimization control method of a waste incineration power plant deacidification process according to any one of claims 1 to 7, characterized in that: The multi-source perception acquisition module is used for high-frequency synchronous acquisition, timestamp marking, preliminary denoising, and abnormality rough screening; The working condition characteristic construction module is used for obtaining fused scalar estimation and working condition characteristic vectors, and constraint side information; the intelligent prediction module is used for obtaining a future short-term emission trend prediction sequence and heat value prediction; and the feedforward-feedback integrated control module corrects parameters through self-tuning PID and outputs an execution instruction. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is characterized in that: The processor executes the computer program to realize the steps of the optimization control method of the waste incineration power generation deacidification process according to any one of claims 1-7.

10. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to realize the steps of the optimization control method of the waste incineration power generation deacidification process according to any one of claims 1-7.

Images