Energy supply energy efficiency optimization method and system for gas-steam boiler

Abstract

The invention discloses a gas-steam boiler energy supply efficiency optimization method and system, and relates to the technical field of boiler control, and the method comprises the steps: collecting gas component data in real time, synchronously obtaining boiler operation parameters, and carrying out the data preprocessing of the gas component data and the boiler operation parameters to generate a standardized input matrix; real-time operation data flow of the boiler is obtained and input into the trained neural differential equation prediction model, and a state prediction sequence in a future time domain is obtained through calculation and transmitted to the dynamic optimization controller; in the dynamic optimization controller, a rolling horizon optimization problem is constructed based on a state prediction sequence, a control instruction set is converted into a physical operation quantity, a regulation gas flow, a combustion-supporting air quantity and a water supply flow through an execution mechanism, and after the boiler operation state is updated, actual operation data is fed back to a constraint updating process. Through dynamic coupling physical constraint and data driving, the real-time performance, adaptability and physical consistency of boiler energy efficiency optimization are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of boiler control, and in particular to a method and system for optimizing the energy supply efficiency of a gas-fired steam boiler. Background Art

[0002] In recent years, energy efficiency optimization methods for gas-fired steam boilers have gradually evolved from traditional PID control to a fusion of data-driven and physical models. In existing technologies, predictive control based on neural networks (such as LSTM and Transformer) can handle nonlinear dynamic characteristics, while physical models (such as heat balance equations and combustion dynamics equations) provide interpretable constraints. In addition, the combination of genetic programming and neural differential equations has shown potential in complex system modeling, such as mining implicit physical laws through symbolic regression. However, existing methods mostly use static constraints or pure data-driven methods, lack dynamic coupling mechanisms, and are difficult to adapt to real-time operating conditions such as gas composition fluctuations and load changes.

[0003] A core flaw in existing technologies lies in the disconnect between physical constraints and data-driven models. In traditional approaches, physical constraints are often hard-coded into the control algorithm as fixed equations, preventing dynamic adjustment to changing fuel characteristics. While purely data-driven models can capture dynamic characteristics, they are prone to generating spurious correlations that violate the laws of thermodynamics. Furthermore, the real-time updating of constraints in rolling horizon optimization is inefficient, leading to control delays or overshoot, especially when gas composition fluctuates at high frequencies. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a method for optimizing the energy efficiency of a gas-fired steam boiler to solve the problems of poor dynamic adaptability of physical constraints and real-time control delay in energy efficiency optimization of a gas-fired steam boiler.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In the first aspect, the present invention provides a method for optimizing the energy efficiency of a gas-fired steam boiler, which includes: real-time collection of gas composition data, synchronous acquisition of boiler operating parameters, and data preprocessing of the gas composition data and boiler operating parameters to generate a standardized input matrix; based on the standardized input matrix, physical constraints are mined through a genetic programming algorithm to generate a set of constraint equations in differential algebraic form, the constraint equations are encoded as Lagrange multiplier terms and injected into a neural differential equation to construct a neural differential equation prediction model with physical constraints; the real-time operation data stream of the boiler is obtained and input into a trained neural differential equation prediction model, a state prediction sequence in the future time domain is calculated, and the sequence is transmitted to a dynamic optimization controller; in the dynamic optimization controller, a rolling time domain optimization problem is constructed based on the state prediction sequence, the control instruction set is converted into a physical operation quantity through an actuator, the gas flow rate, combustion-supporting air volume and feed water flow rate are adjusted, and after the boiler operation state update is completed, the actual operation data is fed back to the constraint update process.

[0008] As a preferred solution of the gas steam boiler energy supply efficiency optimization method described in the present invention, the gas composition data includes the volume percentage concentrations of methane, ethane, propane, carbon monoxide and hydrogen; the boiler operating parameters include steam pressure, exhaust temperature, oxygen content, feed water flow and load rate.

[0009] As a preferred solution of the method for optimizing the energy efficiency of a gas-fired steam boiler according to the present invention, the set of constraint equations includes a heat balance equation, an oxygen content constraint, and a tube wall temperature gradient constraint;

[0010] The constraint equations are encoded as Lagrange multiplier terms and injected into the neural differential equation. The specific steps are as follows:

[0011] The heat balance equation, oxygen content limit and pipe wall temperature gradient constraint are input into the differential algebraic encoder, and the partial derivatives of the constraint equation with respect to the hidden state are calculated through the gradient field generation layer, and the constraint gradient tensor is output;

[0012] Dynamically couple the constrained gradient tensor with the forward propagation result of the neural differential equation, and generate a constraint-enhanced gradient flow through the Lagrange multiplier weighted layer;

[0013] The coupled gradient flow is input into the adjoint optimizer to perform joint backpropagation calculations and synchronously update the neural network parameters and Lagrange multipliers.

[0014] The optimized gradient flow is input into the prediction corrector, and the heat balance equation, oxygen content limit and pipe wall temperature gradient constraint are forced to be satisfied simultaneously through the interior point solver.

[0015] As a preferred solution of the gas steam boiler energy supply efficiency optimization method described in the present invention, the input layer of the neural differential equation prediction model receives real-time sensor data stream, the hidden layer describes the combustion dynamics process through neural ordinary differential equations, and the output layer generates state prediction values in the future time period. The state prediction values include thermal efficiency, nitrogen oxide concentration and pipe wall stress distribution. During the training process, the parameters of the neural differential equation prediction model are optimized based on the adjoint method, and the loss function integrates the prediction error and constraint violation penalty.

[0016] As an optimal solution of the gas steam boiler energy supply efficiency optimization method described in the present invention, the control variables of the rolling time domain optimization problem include the gas valve opening, the fan speed and the water supply pump frequency, and the control instruction set is generated in real time through the optimization algorithm, and the control instruction set is sent to the execution mechanism.

[0017] As a preferred solution of the method for optimizing energy efficiency of gas steam boiler energy supply of the present invention, the actual operation data refers to the real-time operating condition verification data after the gas steam boiler executes the control instruction;

[0018] The real-time operating condition verification data includes control response data, operating condition verification parameters and energy efficiency indicators.

[0019] As a preferred solution of the gas steam boiler energy supply efficiency optimization method described in the present invention, the constraint update process refers to a closed-loop optimization process of dynamically correcting physical constraints based on the real-time operation data flow of the boiler.

[0020] In a second aspect, the present invention provides a gas steam boiler energy supply efficiency optimization system, comprising a data acquisition module, a constraint modeling module, a real-time prediction module and an optimization control module; the data acquisition module is used to collect gas composition data in real time, synchronously obtain boiler operating parameters, and generate a standardized input matrix by preprocessing the gas composition data and boiler operating parameters; the constraint modeling module is used to mine physical constraints based on historical operating data and the standardized input matrix through a genetic programming algorithm, generate a set of constraint equations in differential algebraic form, encode the constraint equations as Lagrange multiplier terms and inject them into neural differential equations to construct a neural differential equation prediction model with physical constraints; the real-time prediction module is used to obtain the real-time operation data stream of the boiler, input it into the trained neural differential equation prediction model, calculate the state prediction sequence in the future time domain, and transmit it to the dynamic optimization controller; the optimization control module is used to construct a rolling time domain optimization problem based on the state prediction sequence in the dynamic optimization controller, convert the control instruction set into physical operation quantities through the actuator, adjust the gas flow rate, combustion-supporting air volume and feed water flow rate, and after completing the boiler operation state update, feed back the actual operation data to the constraint update process.

[0021] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the gas steam boiler energy supply efficiency optimization method as described in the first aspect of the present invention is implemented.

[0022] In a fourth aspect, the present invention 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 method for optimizing energy supply efficiency of a gas steam boiler as described in the first aspect of the present invention is implemented.

[0023] The beneficial effects of the present invention are: dynamic mining of differential algebraic constraint equations based on genetic programming to achieve deep coupling of physical laws and data-driven models; dynamic injection of constraint gradient tensors through Lagrange multiplier weighted layers, combined with edge computing acceleration and buffering mechanisms, significantly reducing constraint violation detection delays; improved hierarchical solution strategy realizes rapid rolling optimization with dynamic constraints and improves control instruction response speed; reverse symbolic deduction and verification of constraint equations ensures the physical rationality of prediction results; and adaptively updates constraint equations when gas composition changes suddenly to maintain prediction accuracy and control stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0025] Figure 1 Flowchart of the energy efficiency optimization method for powering a gas-fired steam boiler.

[0026] Figure 2 Schematic diagram of the coupling between physical constraint modeling and neural differential equations in Example 1.

[0027] Figure 3 This is a flow chart of the dynamic rolling time domain optimization control in Example 1.

[0028] Figure 4 This is a diagram of the closed-loop feedback and constraint dynamic update mechanism in Example 1. DETAILED DESCRIPTION

[0029] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0030] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0031] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0032] Example 1, with reference to Figures 1 to 4 This embodiment provides a method for optimizing the energy efficiency of a gas-fired steam boiler, comprising the following steps:

[0033] S1. Collect gas composition data in real time, obtain boiler operating parameters synchronously, and generate a standardized input matrix by preprocessing the gas composition data and boiler operating parameters.

[0034] Specifically, the following steps are included:

[0035] The fuel gas composition (volume percentage concentrations of methane, ethane, propane, carbon monoxide, and hydrogen) is collected through an online gas chromatograph (1 Hz), and the boiler parameters (steam pressure, exhaust temperature, oxygen content, feed water flow rate, and load rate) are obtained through a distributed control system (DCS system), and a unified data format is defined.

[0036] The clocks of each device are calibrated based on the NTP protocol, and all data are timestamped. Low-frequency data of gas composition (such as 5s / time for chromatograph) is linearly interpolated to generate a 1Hz continuous sequence.

[0037] A circular buffer is set up at the edge gateway to detect and eliminate outliers outside 3σ in real time. A sliding window (60 seconds) is used to establish a parameter baseline. Linear interpolation is used to fill in ≤3 consecutive missing points. An alarm is triggered when there are >3 points and the value is filled with the previous valid value.

[0038] Specifically, when deploying a circular buffer at the edge gateway, a fixed-size thread-safe circular queue structure is used to allocate independent buffers for gas composition data and boiler parameters (gas data cache capacity is 60 seconds, boiler parameter cache capacity is 10 seconds), and multi-threaded safe access is ensured through pre-allocated memory and mutex lock mechanisms. The data acquisition thread writes the sensor data to the buffer header in timestamp order. When the buffer is full, it automatically overwrites the oldest data and records an alarm; the pre-processing thread reads data from the end of the buffer at a frequency of 1Hz, and automatically triggers linear interpolation compensation when the timestamp is discontinuous. The buffer has a built-in integrity check mechanism to check the timestamp monotonicity and numerical rationality of each data packet. It also supports the function of resuming transmission after network disconnection. It can maintain a data cache of at least 10 seconds in the event of a network interruption, and resend the data in timestamp sorting after recovery. To optimize performance, memory mapping technology is used to reduce data copy overhead, and batch reading (10-second window data each time) is used to reduce thread switching frequency. Ultimately, reliable data buffering is achieved with end-to-end processing delay ≤ 20ms and data loss rate ≤ 0.1% within 10 seconds of network interruption. All abnormal events and buffer status are recorded in real time to the local log database for monitoring and analysis.

[0039] The Z-Score parameters (mean μ, standard deviation σ) are further calculated through a sliding window (60 seconds) to normalize the boiler parameters in real time. The gas concentration is first logarithmically transformed to compress the dynamic range before the Z-Score is executed.

[0040] The normalized gas composition data (5-dimensional) and boiler operating parameters (5-dimensional) are concatenated into a matrix by timestamp, where each row represents a moment (1 Hz) and each column corresponds to a feature, generating the standardized input matrix U:

[0041] U=[x1,x2,…,x 10 ] T ,U∈R N×10 ;

[0042] Where N is the time step of the real-time sliding window (typical value is 30).

[0043] S2. Based on the standardized input matrix, the physical constraints are mined through the genetic programming algorithm to generate a set of constraint equations in differential algebraic form. The constraint equations are encoded as Lagrange multipliers and injected into the neural differential equation to construct a neural differential equation prediction model with physical constraints.

[0044] Specifically, the following steps are included:

[0045] A symbolic space consisting of a variable set, an operator set, and a constant pool is defined based on the normalized matrix X.

[0046] Specifically, the variable set: {x1,x2,…,x 10}; (corresponding to 10-dimensional features, such as x1 = methane concentration, x2 = steam pressure, etc.); operator set: including differential Integration (∫), arithmetic operations (+, -, ×, ÷), and nonlinear functions Constant pool: fixed constants (0, 1, π, e) and random numbers (range [-5, 5]);

[0047] 200 initial equations are randomly generated. The initial equation structure is tree-like (depth ≤ 6) for population initialization, providing a basis for candidate constraint equations for subsequent evolution. For example, the candidate constraint equation F(x) is expressed as:

[0048]

[0049] Where, represents the rate of change of methane (CH4) concentration over time (instantaneous change rate), d is the differential operator, which means the differential operation is performed on the infinitesimal change of the variable, dt represents the small change in time (i.e. time differential), represents the nonlinear effect of steam pressure (heat balance equation prototype), x3 is the exhaust gas temperature, x4 is the oxygen content, It represents the exhaust heat loss under unit oxygen content, implying the balance between combustion efficiency and oxygen utilization rate (oxygen content limited prototype). 0 is the nature of the constraint, that is, it follows the mass-energy conservation law.

[0050] For each candidate constraint equation F(x), traverse the historical data to calculate its constraint violation degree Fitness, which is expressed as:

[0051]

[0052] Where T represents the length of the historical data time window for calculating fitness (total number of sampling points), F(x(t)) represents the value of the candidate constraint equation at time t, Represents the normalization coefficient, which is used to eliminate the influence of window length and make the fitness of different T comparable;

[0053] It should be noted that the closer the fitness is to 0, the stronger the constraint of the candidate constraint equation on the data.

[0054] Furthermore, the iterative operation is evolved by defining selection, crossover, mutation and termination conditions.

[0055] Specifically, selection: tournament selection (selecting the best from 5 individuals each time); crossover: subtree exchange (probability 60%); mutation: node replacement / insertion / deletion (probability 30%); termination condition: 50 generations of evolution or the optimal fitness change rate <1e-4.

[0056] Furthermore, through iterative optimization through evolutionary operations (selection, crossover, and mutation), the equations in the population gradually converge to candidate solutions with high fitness. At this point, the optimal individuals in the final generation need to be mathematically simplified to eliminate redundant structures and improve the interpretability of the equations.

[0057] Specifically, through mathematical equivalence simplification (such as combining like terms), the final output is a simplified set of differential-algebraic equations in the form of:

[0058]

[0059] Where, Represents the kth set of constraint equations (such as heat balance equation, oxygen content limit, pipe wall temperature gradient constraint), k represents the index of the constraint equation, represents the rate of change of the state vector over time, which is obtained by discrete difference or neural network output, and K represents the total number of valid constraint equations mined by genetic programming;

[0060] It should be noted that the candidate constraint equations are a set of initial equations randomly generated by genetic programming. These initial equations may contain redundant structures or parts that do not conform to physical laws and need to be screened and optimized through fitness evaluation and evolutionary operations. The constraint equation set is the final valid equations retained from the candidate equations after strict screening, mathematical simplification and physical verification. They not only meet the requirements of data consistency (high fitness) and mathematical simplicity (no redundant terms), but also must comply with engineering laws such as mass / energy conservation. Finally, they are expressed in the standard form C k =0 is injected into the neural differential equation to directly guide the training and prediction of the neural differential equation prediction model.

[0061] The core difference between the two is that the candidate equations are intermediate products in the evolutionary process, which are huge in number and require further verification; the set of constraint equations is the reliable result after refinement, which is concise in number and has clear physical meaning, and together constitute a complete process from data-driven to physical law integration.

[0062] Input the physical constraint equations (heat balance equation, oxygen content limit, pipe wall temperature gradient constraint) into the differential algebraic encoder, and calculate the partial derivatives of the constraint equations with respect to the hidden state through the gradient field generation layer And output the constrained gradient tensor (the dimension is consistent with the hidden state);

[0063] The constrained gradient tensor is dynamically coupled with the forward propagation result of the neural differential equation, and the constraint-enhanced gradient flow is generated through the Lagrange multiplier weighted layer, which is expressed as:

[0064]

[0065] Where, f θ(X,t) represents the original state derivative predicted by the neural network (with parameter θ), Indicates the partial derivative of the constraint equation with respect to X, represents the kth constraint equation The total differential of is the differential of the state vector X, representing the infinitesimal change in state, μ k is the Lagrange multiplier;

[0066] Perform joint backpropagation to synchronously update the neural network parameters θ and the Lagrange multiplier μ k ;

[0067] Enforced by the interior point solver to satisfy (For example, heat balance equation, oxygen content limit, and pipe wall temperature gradient constraint are all valid at the same time)

[0068] The constraint-enhanced gradient flow is generated through the Lagrange multiplier weighted layer to ensure that the state derivatives meet the physical constraints; the neural differential square network architecture is constructed, and a 4-layer fully connected network (hidden layer dimension 64, Swish activation function) is set up, and the input is The output is The constrained gradient terms are computed dynamically during the forward pass.

[0069] Randomly sample 1000 time windows (60 seconds of data per window) from the normalized matrix U;

[0070] The loss function is defined as:

[0071]

[0072] Where, X pred (t) represents the state vector (10 dimensions) predicted at time t, X true (t) The true state vector at time t (from historical data), |||| 2 represents the Euclidean norm (L2 norm);

[0073] The AdamW optimizer (learning rate 3e-4, weight decay 1e-5) is used to dynamically update parameters based on the adjoint method;

[0074] Dynamically update the Lagrange multiplier:

[0075]

[0076] Where, ρ (n) Represents the global constraint penalty coefficient, which increases with the training round n. is the truncation function, which means Restricted to the interval [-1,1];

[0077] Specifically, the truncation function is expressed as:

[0078]

[0079] Preferably, by defining a truncation function, it is possible to prevent a single constraint from being violated drastically (e.g., a sensor failure causing Mutation) causes μ k to force the constraints to be satisfied.

[0080] If a constraint If the violation decreases by <5% in 10 consecutive iterations, its weight is decayed by 50%;

[0081] Calculate the average violation of each constraint equation on the test set and compare the violation reduction ratio between the unconstrained model (Baseline) and the constrained model;

[0082] Calculate the mean square error (MSE) and mean absolute percentage error (MAPE) of predictions for key parameters (such as steam pressure and exhaust temperature);

[0083] It should be noted that the purpose of calculating the average violation of the constraint equation on the test set and comparing the reduction ratio of the violation between the unconstrained model and the constrained model is to verify the correction effect of the physical constraint on the prediction results (such as the thermal balance deviation is reduced from 15% to 3%) and to ensure that the model follows the engineering law; calculating the MSE and MAPE of the key parameters verifies the numerical accuracy (such as the exhaust temperature error is within ±2°C). These two steps form a double check: the former guarantees physical rationality, and the latter guarantees data accuracy, which together provide a credible basis for the subsequent "symbolic inverse verification" - only the constraint equations that meet both low violation and low error can be confirmed to be consistent with the combustion reaction principle through symbolic inverse (such as the relationship between oxygen content and exhaust temperature in the constraint equation conforms to the combustion heat loss law), and finally prove that the constraint equations generated by the algorithm truly reflect the physical mechanism of the boiler, rather than false correlations caused by data noise, thereby ensuring the engineering credibility of the prediction model in control optimization.

[0084] Constraint equations Perform symbolic inversion and verify its physical meaning (such as the relationship between methane consumption rate and oxygen supply) in combination with the boiler combustion chemical reaction formula to ensure compliance with engineering laws such as heat balance equations and oxygen content limits. For example:

[0085]

[0086] It should be noted that this verification step ensures that the constraint equations obtained by data driving are consistent with the physical and chemical principles of the combustion process, avoiding the algorithm generating false relationships that are mathematically reasonable but physically invalid, thereby improving the engineering credibility of the neural differential equation prediction model and confirming that the constraint equations do reflect the real physical mechanism in boiler operation, rather than just the result of data fitting.

[0087] S3. Obtain the real-time operation data stream of the boiler and input it into the trained neural differential equation prediction model to calculate the state prediction sequence in the future time domain and transmit it to the dynamic optimization controller.

[0088] Specifically, the following steps are included:

[0089] A lightweight data synchronization service is deployed on the edge gateway to subscribe to boiler operating parameters (steam pressure, exhaust temperature, etc.) from the DCS system in real time through the OPC UA protocol, and to receive gas composition data from the online gas chromatograph through the Modbus TCP protocol.

[0090] Based on the timestamp, a sliding time window (window length 30 seconds, step length 1 second) is used to dynamically align the sampling time deviation of the gas composition data and the boiler operating parameters, eliminating the timing misalignment problem caused by sensor response delay (such as the oxygen content sensor lag of 2 seconds).

[0091] The trained neural differential equation prediction model is deployed on edge computing nodes (such as NVIDIA JetsonAGX Xavier), and the model is converted into an inference-optimized version with FP16 precision through the TensorRT engine, achieving a single prediction time of ≤50ms.

[0092] Specifically, after training the neural differential equation prediction model, the neural differential equation prediction model is exported to the ONNX format, specifying the dimensions and dynamic axis parameters of the input and output tensors. The model is loaded using TensorRT's ONNX parser, and a TRTBuilder instance is created on the NVIDIA Jetson AGX Xavier edge computing node. An optimization profile is set to define the size range of the input and output tensors. FP16 precision mode is enabled and layer fusion optimization is added. Targeted optimization is applied to LSTM units to ensure numerical stability. The Builder.buildEngine method is called to generate the optimized inference engine, setting the maximum workspace to 1GB and the maximum batch size to 8 to balance memory usage and throughput. The computation time of each layer is verified using TRTProfiler. After the generated serialized engine file is deployed to the edge node, the measured single inference latency is reduced from 120ms of the original model to 43ms, meeting real-time requirements.

[0093] According to the characteristics of input feature dimension and hidden state dimension, a double buffer queue is pre-allocated in memory.

[0094] It should be noted that the feature dimensions correspond to the real-time collected variables in the standardized input matrix generated in step S1, including: 5-dimensional gas composition data (methane, ethane, propane, carbon monoxide, and hydrogen volume percentage concentration); 5-dimensional boiler operating parameters (steam pressure, exhaust temperature, oxygen content, feed water flow, and load rate); the role in the double buffer queue: the input data at each moment (1Hz) is composed of these 10-dimensional features, and the pre-allocated current moment input buffer (30-second capacity) stores a continuous 30-second 10-dimensional feature sequence for temporal dependency calculation of model inference.

[0095] The hidden state dimension, derived from the hidden layer construction of the neural differential equation prediction model in step S2, includes: predicted output targets: direct control variables such as thermal efficiency, NOx concentration, and tube wall stress distribution; intermediate state variables: combustion dynamics characteristics derived from solving the differential equation (such as instantaneous combustion rate and smoke retention coefficient). Its role in the dual buffer queue: The pre-allocated prediction result buffer (60-second capacity) stores the 15-dimensional hidden state sequence output by the model, providing the dynamic optimization controller with a predicted state trajectory in the future time domain, ensuring the temporal consistency of control command generation.

[0096] Among them, the current moment input buffer (storing the latest 30 seconds of standardized data) and the prediction result buffer (storing the next 60 seconds prediction sequence).

[0097] Furthermore, every time new data is received for 1 second, the neural differential equation prediction model inference process is triggered.

[0098] Specifically, the latest 30 seconds of standardized data is input into the neural differential equation prediction model, and forward propagation is performed to calculate the state prediction sequence (thermal efficiency, NOx concentration, etc.) for the next 60 seconds. The output sequence is physically verified for rationality: whether the thermal efficiency is in the range of [20%, 95%], whether the NOx concentration exceeds the environmental protection limit (such as ≤50mg / m 3 ), if violated, the neural differential equation prediction model will be triggered to re-infer (up to 3 times).

[0099] The prediction sequence is encapsulated into a binary data packet through a time series compression algorithm (such as Delta encoding + ZigZag compression) and transmitted to the dynamic optimization controller via the MQTT protocol. The end-to-end transmission delay is ≤100ms.

[0100] A prediction confidence evaluation unit is set on the dynamic optimization controller side, which automatically reduces the prediction time domain length when the following two situations are detected: specifically, when the gas composition fluctuation rate exceeds ±10% for 5 consecutive seconds (indicating fuel instability); when the deviation between the predicted value and the actual value of steam pressure is continuously greater than 5% for 3 seconds.

[0101] S4. In the dynamic optimization controller, a rolling time domain optimization problem is constructed based on the state prediction sequence. The control instruction set is converted into physical operation quantities through the actuator to adjust the gas flow, combustion air volume and feed water flow. After completing the boiler operation status update, the actual operation data is fed back to the constraint update process.

[0102] The rolling horizon optimization objective function is defined based on economic indicators, environmental indicators and safety indicators.

[0103] Among them, the economic index is: minimizing gas consumption (quadratically related to valve opening); the environmental index is: limiting the rate of increase of NOx concentration (linearly related to air volume adjustment); and the safety index is: the absolute value of the tube wall temperature gradient change rate is ≤3°C / s.

[0104] An improved sequential quadratic programming (SQP) algorithm is used to construct a hierarchical solution strategy for the control variable dimensions (gas valve opening, fan speed, and feedwater pump frequency):

[0105] Specifically, the first layer calculates the gas flow set value based on the prediction sequence, with the constraints being the heat balance equation and the pipe wall temperature gradient. The second layer fixes the gas flow and optimizes the ratio of the combustion-supporting air volume and the feed water flow, with the constraints being the oxygen content limit and the steam pressure.

[0106] It should be noted that the upper limit of the solution time is set to 200ms, and the backup plan is activated when the solution fails to converge after the timeout.

[0107] Specifically, when the optimization solution fails to converge within 200ms, the dynamic optimization controller immediately executes the emergency control strategy: when the load rate is greater than 80%, the steam pressure control target is locked, the preset air-coal ratio parameters are used, and the pressure closed-loop PID adjustment is activated; when the load rate is less than or equal to 80%, the controller switches to the efficiency priority mode, generates control instructions based on the interpolation of the historical data of the optimal operating conditions in the last 24 hours, and superimposes the feedforward compensation of the water flow. During the transition process, the steam pressure fluctuation is limited to ±0.2MPa, the oxygen content deviation is controlled within ±0.5%, the pipe wall temperature gradient does not exceed 3℃ / s, and the NOx concentration is maintained at 50mg / m 3 The following ensures that the control mode switch is completed within 5 seconds. The preset air-to-fuel ratio parameter refers to the fuel-to-air ratio baseline value pre-stored in the boiler control and verified by historical operating data.

[0108] Convert the optimized solution into actuator instructions.

[0109] Specifically, for the gas regulating valve, feedforward-feedback composite control is adopted, with the feedforward quantity coming from the optimized solution and the feedback quantity based on the pressure deviation PID correction; for the variable frequency fan, the frequency set value is generated by interpolation based on the air volume-speed characteristic curve, and surge prevention logic is superimposed (avoiding the 35-45Hz resonance zone); for the feed water pump, the lag effect caused by the boiler water volume is offset through flow differential compensation.

[0110] An instruction smoothing unit is deployed in the PLC to apply first-order inertia filtering to instruction jumps between adjacent control cycles (such as sudden changes in valve opening > 5%) to prevent mechanical shock.

[0111] Specifically, the instruction smoothing unit deployed in the PLC control program uses dynamic inertia filtering technology to process control instructions. The instruction smoothing unit continuously monitors the change amplitude of key control instructions such as the gas regulating valve opening, fan speed and water pump frequency. When it detects an instruction jump in adjacent control cycles, it automatically activates the smoothing algorithm. The filter intensity is automatically adjusted according to the current operating conditions, focusing on maintaining the control response speed during the rapid load change stage and focusing on improving the instruction smoothness during the stable operation stage. This adaptive adjustment mechanism not only ensures dynamic response capability, but also controls the action rate of the actuator within a safe range, effectively avoiding mechanical shock and wear of valves, fans and other equipment caused by sudden changes in instructions. All smoothing processes are completed within a single control cycle without affecting the overall control timing.

[0112] It should be noted that when referring to command jumps within adjacent control cycles, the "commands" here refer to the direct operational commands generated in real time by the dynamic optimization controller and issued to the actuators (gas control valve, combustion-supporting fan, and feedwater pump). Examples include the gas control valve opening setpoint (percentage command, such as 50% opening), the combustion-supporting fan speed setpoint (speed command, such as 1200 rpm), and the feedwater pump frequency setpoint (Hz command, such as 45 Hz). These commands are converted by the PLC into physical actuators (such as valve opening signals and inverter control signals) to directly drive the boiler's gas flow, combustion-supporting air flow, and feedwater flow. These jumps (such as a sudden change in the opening command from 45% to 52% within adjacent control cycles) can cause a step change in the actuator's operating amplitude, triggering mechanical stress. Therefore, a first-order inertial filter is required to smooth the command sequence. While maintaining control accuracy, the actuator's operating rate is limited to a safe range (e.g., valve opening change ≤ 3% per second) to prevent mechanical shock or equipment life loss caused by sudden transient load changes.

[0113] Collect actual operation data after execution (actual value of steam pressure, NOx concentration measurement value, etc.) and calculate the prediction-actual deviation index.

[0114] Specifically, short-term deviations (<10 seconds) are used for online parameter self-tuning of the dynamic optimization controller; long-term deviations (≥30 minutes) trigger the constraint update process.

[0115] When any of the following conditions is detected, the dynamic update of the constraint equations of step S2 is initiated:

[0116] Specifically, the violation amount predicted by the same constraint equation for 10 consecutive times is greater than the thermal balance deviation threshold (such as thermal balance deviation > 8kW); the gas composition undergoes a qualitative change (such as a sudden increase in the proportion of hydrogen by > 5% and lasting for 1 minute).

[0117] Feedback data is collected and transmitted back to the edge computing node through an encrypted channel, triggering lightweight genetic programming retraining (only updating the mutation rate and selection pressure parameters), and it takes ≤2 minutes to complete the constraint iteration.

[0118] This embodiment also provides a gas steam boiler energy supply efficiency optimization system, including: a data acquisition module, a constraint modeling module, a real-time prediction module and an optimization control module; the data acquisition module is used to collect gas composition data in real time, synchronously obtain boiler operating parameters, and generate a standardized input matrix by preprocessing the gas composition data and the boiler operating parameters; the constraint modeling module is used to mine physical constraints based on historical operating data and the standardized input matrix through a genetic programming algorithm, generate a set of constraint equations in differential algebraic form, encode the constraint equations as Lagrange multiplier terms and inject them into the neural differential equation to construct a neural differential equation prediction model with physical constraints; the real-time prediction module is used to obtain the real-time boiler operation data stream, input it into the trained neural differential equation prediction model, calculate the state prediction sequence in the future time domain, and transmit it to the dynamic optimization controller; the optimization control module is used to construct a rolling time domain optimization problem based on the state prediction sequence in the dynamic optimization controller, convert the control instruction set into physical operation quantities through the actuator, adjust the gas flow rate, combustion-supporting air volume and feed water flow rate, and after completing the boiler operation state update, feed the actual operation data back to the constraint update process.

[0119] This embodiment also provides a computer device, which is suitable for the gas steam boiler energy supply energy efficiency optimization method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the gas steam boiler energy supply energy efficiency optimization method proposed in the above embodiment.

[0120] The computer device may be a terminal, comprising a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide 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 the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner may be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a button, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse.

[0121] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for optimizing the energy efficiency of a gas-fired steam boiler as proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, disk or optical disk.

[0122] In summary, the present invention achieves deep coupling of physical laws and data-driven models by: dynamically mining differential algebraic constraint equations based on genetic programming; dynamically injecting constraint gradient tensors through the Lagrange multiplier weighted layer, combining edge computing acceleration and buffering mechanisms to significantly reduce constraint violation detection delays; an improved hierarchical solution strategy realizes rapid rolling optimization with dynamic constraints, improving the response speed of control instructions; reverse symbolic deduction and verification of constraint equations ensures the physical rationality of prediction results; and adaptively updates constraint equations when gas composition changes suddenly to maintain prediction accuracy and control stability.

[0123] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention 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 invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for optimizing the energy efficiency of a gas-fired steam boiler, characterized by: include, Collect gas composition data in real time, obtain boiler operating parameters simultaneously, and generate a standardized input matrix by preprocessing the gas composition data and boiler operating parameters; Based on the standardized input matrix, the genetic programming algorithm is used to mine physical constraints, generate a set of constraint equations in differential algebraic form, encode the constraint equations as Lagrange multipliers and inject them into the neural differential equation to construct a neural differential equation prediction model with physical constraints. The real-time operation data stream of the boiler is obtained and input into the trained neural differential equation prediction model to calculate the state prediction sequence in the future time domain and transmit it to the dynamic optimization controller; In the dynamic optimization controller, a rolling time domain optimization problem is constructed based on the state prediction sequence. The control instruction set is converted into physical operation quantities through the actuator to adjust the gas flow, combustion air volume and feed water flow. After completing the boiler operation status update, the actual operation data is fed back to the constraint update process.

2. The method for optimizing energy efficiency of a gas-fired steam boiler according to claim 1, wherein: The gas composition data includes the volume percentage concentrations of methane, ethane, propane, carbon monoxide and hydrogen; the boiler operating parameters include steam pressure, exhaust gas temperature, oxygen content, feed water flow and load rate.

3. The method for optimizing energy efficiency of a gas-fired steam boiler according to claim 1, wherein: The constraint equation set includes a heat balance equation, an oxygen content limit, and a pipe wall temperature gradient constraint; The constraint equations are encoded as Lagrange multiplier terms and injected into the neural differential equation. The specific steps are as follows: The heat balance equation, oxygen content limit and pipe wall temperature gradient constraint are input into the differential algebraic encoder, and the partial derivatives of the constraint equation with respect to the hidden state are calculated through the gradient field generation layer, and the constraint gradient tensor is output; Dynamically couple the constrained gradient tensor with the forward propagation result of the neural differential equation, and generate a constraint-enhanced gradient flow through the Lagrange multiplier weighted layer; The coupled gradient flow is input into the adjoint optimizer to perform joint backpropagation calculations and synchronously update the neural network parameters and Lagrange multipliers. The optimized gradient flow is input into the prediction corrector, and the heat balance equation, oxygen content limit and pipe wall temperature gradient constraint are forced to be satisfied simultaneously through the interior point solver.

4. The method for optimizing energy efficiency of a gas-fired steam boiler according to claim 3, wherein: The input layer of the neural differential equation prediction model receives real-time sensor data streams, the hidden layer describes the combustion dynamics process through neural ordinary differential equations, and the output layer generates state prediction values for future time periods. The state prediction values include thermal efficiency, nitrogen oxide concentration, and pipe wall stress distribution. During the training process, the parameters of the neural differential equation prediction model are optimized based on the adjoint method, and the loss function integrates the prediction error and constraint violation penalty.

5. The method for optimizing energy efficiency of a gas-fired steam boiler according to claim 4, wherein: The control variables of the rolling horizon optimization problem include the gas valve opening, the fan speed, and the water pump frequency. A control instruction set is generated in real time through the optimization algorithm and sent to the actuator.

6. The method for optimizing energy efficiency of a gas-fired steam boiler according to claim 5, wherein: The actual operation data refers to the real-time operating condition verification data after the gas steam boiler executes the control instruction; The real-time operating condition verification data includes control response data, operating condition verification parameters and energy efficiency indicators.

7. The method for optimizing energy efficiency of a gas-fired steam boiler according to claim 6, wherein: The constraint update process refers to a closed-loop optimization process that dynamically corrects physical constraints based on the real-time operation data stream of the boiler.

8. A gas-fired steam boiler energy supply efficiency optimization system, based on the gas-fired steam boiler energy supply efficiency optimization method according to any one of claims 1 to 7, characterized in that: Including data acquisition module, constraint modeling module, real-time prediction module and optimization control module; The data acquisition module is used to collect gas composition data in real time, synchronously obtain boiler operating parameters, and generate a standardized input matrix by preprocessing the gas composition data and boiler operating parameters; The constraint modeling module is used to mine physical constraints based on historical operating data and a standardized input matrix through a genetic programming algorithm, generate a set of constraint equations in differential algebraic form, encode the constraint equations as Lagrange multiplier terms and inject them into the neural differential equation to construct a neural differential equation prediction model with physical constraints; The real-time prediction module is used to obtain the real-time operation data stream of the boiler, input it into the trained neural differential equation prediction model, calculate the state prediction sequence in the future time domain, and transmit it to the dynamic optimization controller; The optimization control module is used to construct a rolling time domain optimization problem based on the state prediction sequence in the dynamic optimization controller, convert the control instruction set into physical operation quantities through the actuator, adjust the gas flow, combustion-supporting air volume and feed water flow, and after completing the boiler operation status update, feed back the actual operation data to the constraint update process.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method for optimizing the energy supply efficiency of a gas-fired steam boiler according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for optimizing the energy efficiency of a gas-fired steam boiler according to any one of claims 1 to 7 are implemented.

Images