Thermal power generating unit space-time coupling control method based on double-gating LSTM network

Abstract

The invention relates to a thermal power generating unit space-time coupling control method based on a double-gating LSTM network, and the method comprises the steps: collecting the real-time data of a thermal power generating unit, including a target control variable and a plurality of coupling variables; real-time data of the thermal power generating unit are input into the double-gating LSTM network, and temperature curves and control parameters of a plurality of steps in the future are obtained through prediction; the double-gating LSTM network splits a forgetting gate of an LSTM into a short-term forgetting gate and a long-term forgetting gate, a short-term gating channel is started when the double-gating LSTM network is in a conventional working condition, and a long-term gating channel is started when the time delay exceeds a set time delay threshold value; generating a thermodynamic sensitivity weight matrix based on the thermodynamic sensitivity weight model, and driving the LSTM network to perform back propagation rolling optimization of parameters; and based on the output of the double-gate-control LSTM network, a secondary air door opening instruction of the fuel valve is generated through a model prediction controller. Compared with the prior art, the problem of control instability under complex working conditions such as deep peak regulation and coal type switching can be effectively solved.

Description

Technical Field

[0001] The present invention relates to the field of process control of coal-fired thermal power units and intelligent thermal power plants, and in particular to a spatiotemporal coupling control method for thermal power units based on a dual-gated LSTM network. Background Art

[0002] The thermal system of a thermal power plant is a typical high-dimensional nonlinear dynamic system characterized by strong internal coupling, large equipment response inertia, and frequent parameter changes. With the combined effects of various uncertainties during operation, such as coal quality fluctuations and equipment aging, the system's dynamic characteristics exhibit significant nonlinearity and time-varying characteristics. Consequently, thermal control systems in thermal power plants are commonly subject to severe signal fluctuations, complex interference noise, and frequent sudden disturbances. This makes it difficult for traditional control algorithms to meet the real-time and robustness requirements in highly dynamic scenarios. Furthermore, control algorithms based on precise models, such as predictive control and adaptive control, are difficult to effectively apply in thermal control systems.

[0003] Coal-fired power plant boiler and steam turbine temperature control systems exhibit significant time delays. During thermal power plant operation, the response to fuel quantity adjustment and steam temperature adjustment involves three delays: fuel transmission delay (5-8 seconds), heat conduction delay (15-25 seconds), and measurement lag delay (3-5 seconds), totaling over 48 seconds. This significant time delay causes traditional PID controllers to exceed temperature overshoot limits when load fluctuations exceed ±15%. Standard long-short time series prediction models have prediction errors as high as 18.7% for time series with delays exceeding 60 seconds.

[0004] Current thermal power plant boiler temperature control systems are primarily based on a distributed control system (DCS) architecture, utilizing modular hardware to achieve multivariable coordinated control. The core control strategy is a cascaded PID structure: the primary loop targets the boiler outlet temperature and outputs the setpoint for the secondary loop; the secondary loop uses furnace temperature or fuel flow as input to drive the actuator. Drum water level control utilizes a three-impulse PID scheme, integrating water level, steam flow, and feedwater flow signals. A feedforward-feedback mechanism suppresses false water level disturbances, while dynamic compensation of feedwater flow maintains water level stability.

[0005] Current process control technology for thermal power plants mainly relies on distributed control systems and classic PID controllers, which have significant defects in time-varying operating conditions, multivariable coupling, long-delay response, and system integration. The parameter tuning of traditional control schemes relies on engineering experience and experimental methods, and can maintain temperature errors within a controllable range under steady-state conditions. However, when the load suddenly changes or the coal quality fluctuates, the fixed parameters cannot match the time-varying characteristics. In long-delay scenarios, the estimator will also cause excessive overshoot due to model mismatch, resulting in an average annual unplanned shutdown loss of more than 2 million yuan for the unit, seriously restricting the improvement of thermal power flexibility. The industry urgently needs an adaptive control method that integrates thermodynamic mechanisms and data-driven models to achieve high-precision temperature control.

[0006] 1. Insufficient adaptability to time-varying operating conditions: The PID controller with fixed parameters cannot dynamically match the changes in the delay characteristics caused by unit load fluctuations. For example, the boiler temperature response delay is about 28 seconds at 60% load, while it increases to 46 seconds at 100% load. The fixed PID parameters produce an overshoot of more than 5.2°C in this time-varying process, far exceeding the national standard limit of ±2.5°C; especially during deep peak regulation (such as 30% load), the adjustment time is as long as 120 seconds, which can easily trigger superheater protection action. In addition, when the load change rate is higher than 3% / min, the coordinated control system causes pressure oscillations due to mode switching lag, further exacerbating control instability.

[0007] 2. Lack of multivariable coupling mechanism: Existing technologies treat strongly coupled variables such as boiler temperature, pressure, and fuel quantity as independent single-loop controls, ignoring thermodynamic interaction effects. When coal types are switched or the fuel calorific value fluctuates, unmodeled coupling effects cause temperature fluctuations of up to ±6.2°C. Furthermore, decoupling relies on static feedforward coefficients and is unable to adapt to dynamic disturbances. For example, main steam pressure control is achieved solely by adjusting the amount of pulverized coal in the furnace, but the energy balance formula does not account for energy losses during actual operation, resulting in a deviation of more than 15% between the theoretical model and actual operating conditions.

[0008] 3. Low control accuracy with long delays: Boiler steam temperature control has a three-stage delay (total delay exceeding 48 seconds). Traditional Smith predictors require a precise plant model. However, boiler heat transfer dynamics drift with scaling and aging, and model mismatch causes overshoot exceeding 8%. Although standard LSTM neural networks are used for prediction, they suffer from a prediction error of 18.7% for long-delay sequences exceeding 60 seconds due to vanishing gradients. Furthermore, they fail to integrate real-time operating data for rolling parameter updates. For example, the attemperated water control delay reaches nearly 5 minutes, and the lag in temperature feedback forces the actuator to frequently fully open and close, exacerbating valve wear. Summary of the Invention

[0009] The purpose of the present invention is to address the defects of current thermal power plant process control technology in time-varying operating conditions, multivariable coupling, long-delay response and system integration, and to provide a spatiotemporal coupling control method for thermal power units based on a dual-gated LSTM network.

[0010] The purpose of the present invention can be achieved by the following technical solutions:

[0011] A spatiotemporal coupling control method for a thermal power unit based on a dual-gated LSTM network, comprising the following steps:

[0012] Collect real-time data of thermal power units, including target control variables and multiple coupling variables;

[0013] The real-time data of the thermal power unit is input into a dual-gated LSTM network to predict the temperature curve and control parameters for several steps in the future. The dual-gated LSTM network splits the forget gate of the LSTM into a short-term forget gate and a long-term forget gate. When in normal operating conditions, the short-term gating channel is enabled, and the long-term gating channel is activated when the delay exceeds the set delay threshold.

[0014] Generate a thermodynamic sensitivity weight matrix based on the thermodynamic sensitivity weight model, and drive the dual-gated LSTM network to perform back-propagation rolling optimization parameters;

[0015] The temperature prediction curve is used as the rolling optimization input, and the control increment matrix of the objective function of the model predictive controller is dynamically adjusted using the control parameters. The fuel valve opening and secondary air valve opening instructions are generated by the model predictive controller.

[0016] As a preferred technical solution, the target control variable is the boiler temperature, and the coupled variables include the steam pressure and the feed water flow rate.

[0017] As a preferred technical solution, the method compares the energy deviation of the target control variable with the set value in real time. When the energy value deviation exceeds the dynamic threshold, the high-speed processing channel is activated and the sampling period is compressed until the energy value returns to the normal range. The dynamic threshold is based on the thermodynamic energy conservation equation and the real-time load rate fitting:

[0018]

[0019] Wherein, L is the load rate; is the change in heat transfer rate; β0β1β2 are empirical coefficients.

[0020] As an optimal technical solution, the delay threshold adopts a dual-threshold mechanism with load rate adaptation: when the load rate is less than the set value, the delay threshold is set to the second delay threshold; when the load rate is greater than the set value, the delay threshold is increased to the second delay threshold.

[0021] As an optimal technical solution, the dual-gated LSTM network adopts a three-layer architecture: the input layer receives real-time data; the gating layer includes: a short-term forget gate The input is the most recent 10-second window data, and the output is the gating coefficient; the long-term forgetting gate The input is 180 seconds of historical data, and the output is the gating coefficient; the output layer predicts the temperature curve and control parameters for several steps in the future.

[0022] As a preferred technical solution, in the dual-gated LSTM network, the cell state update equation of the long-term gated channel is:

[0023] c t =(λ·γ short +(1-λ)·γ long )⊙c t-1 +i t ⊙g t

[0024] The cell state update equation for short-term gated channels is:

[0025] c t =γ short ⊙c t-1 +i t ⊙g t

[0026] Where, γ short , γ long Represent the output gating coefficients of the short-term forget gate and the long-term forget gate respectively; c t-1 Indicates the cell state at the previous moment; i t represents the input layer input of the current time step; g t Represents the newly generated candidate state at the current time step; λ is the adaptive weight, which is updated in the LSTM network with different event triggers.

[0027] As a preferred technical solution, the rolling update of the adaptive weight is specifically as follows:

[0028]

[0029] Where σ is the Sigmoid activation function, W λ and b λ is the network weight matrix and bias parameter, which are iteratively calculated according to the following formula:

[0030]

[0031] Where L is the prediction loss of the control variable and η is the learning rate.

[0032] As a preferred technical solution, the gating parameters of the dual-gated LSTM network are dynamically adjusted by multivariate weights, and the sensitivity weight α i Through online back-propagation update, the deep integration of thermodynamic mechanism and data-driven is achieved, as shown in the following formula:

[0033]

[0034] Where η is the learning rate, J is the loss function, represents the partial derivative of the target control variable with respect to the coupling variable; α i is the thermodynamic sensitivity weight, calculated as follows:

[0035]

[0036] Among them, T is the target control variable; x i represents the coupled variable, It represents the absolute value of the partial derivative of the target control variable with respect to the coupling variable, and M is the total number of variables.

[0037] As a preferred technical solution, the objective function of the model predictive controller is:

[0038]

[0039] Among them, T pred (k) is the feedback value of the target control quantity at time k; T set is the set value of the target control quantity at time k; W T =diag(α T ,α P ,…) is the weight matrix; W u To control the incremental weight matrix, the formula Δu=[Δu fuel ,Δu air ] T To control the increment, Δu fuel ,Δu air They are the increments of the fuel valve opening and the secondary air valve opening respectively.

[0040] As an optimal technical solution, the secondary air door opening u air Calculated by combustion efficiency constraint:

[0041]

[0042] Where K represents the combustion efficiency compensation coefficient, u fuel is the fuel valve opening input, α F is the fuel quantity sensitivity weight, α T Temperature controls the dominant weight.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] This paper proposes a spatiotemporal coupled intelligent control algorithm based on a dual-gated LSTM network. This algorithm utilizes a staged gating structure to accurately model the dynamic process of boiler temperature and multivariable coupling. A short-term forget gate focuses on millisecond-level disturbance responses, promptly capturing transient operating conditions such as fuel calorific value fluctuations. A long-term forget gate enhances the memory of historical steady-state characteristics, effectively addressing the long-delay control lag caused by load switching. This algorithm significantly suppresses overshoot oscillations, eliminates reliance on manual parameter adjustments, and significantly improves system stability and flexibility in complex scenarios such as deep peak shaving and coal blending. It also reduces equipment wear and maintenance costs, providing a core technology supporting the intelligent transformation of thermal power generation.

[0045] 2) The present invention establishes a load rate adaptive threshold mechanism, calculates a dynamic threshold based on the thermodynamic energy conservation equation and the real-time load rate fitting, and compares the energy deviation amplitude between the controlled variable and the set value in real time, thereby reducing the amount of redundant data while ensuring that the data of the disturbance event is fully captured.

[0046] 3) This invention addresses the strong multivariable coupling characteristics of thermal power plants by constructing a thermodynamic sensitivity weight model. This model quantifies the influence between variables through thermodynamic partial derivatives, replacing the pure signal-dimensional spectral analysis in the original image. This model then generates a weight matrix to drive the LSTM network for backpropagation and rolling optimization of parameters. By explicitly quantifying the thermodynamic sensitivity weights of variables such as steam pressure and feedwater flow, the multi-physics coupling effects are transformed into computable coordinated control laws. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 The present invention is a flow chart of a spatiotemporal coupling control method for a thermal power unit based on a dual-gated LSTM network.

[0048] Figure 2 This is a schematic diagram of the structure of the dual-gated LSTM network proposed in this invention. DETAILED DESCRIPTION

[0049] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0050] Example 1

[0051] The present invention proposes a spatiotemporal coupling intelligent control method based on a dual-gated LSTM network, which improves the process control performance of thermal power units. This solution achieves accurate modeling of the dynamic process of boiler temperature and multi-variable coupling through a unique staged gating structure: focusing on millisecond-level disturbance response through a short-term forget gate, timely capturing transient operating conditions such as fuel calorific value fluctuations; strengthening historical steady-state feature memory through a long-term forget gate, effectively solving the problem of large-delay control lag caused by load switching. The spatiotemporal attention mechanism explicitly quantifies the thermodynamic sensitivity weights of variables such as steam pressure and feed water flow, and converts the multi-physical field coupling effect into a computable collaborative control law. In practical applications, the edge computing layer optimizes the model prediction instructions in real time, drives the actuator to move at high speed, and forms a closed-loop control chain of "dynamic perception-adjustment decision-rolling execution". The specific steps are as follows: Figure 1 As shown:

[0052] S1. Collect real-time data from the thermal power unit, including the target control variables and multiple coupled variables. This method's real-time data acquisition unit continuously acquires process variables such as temperature, pressure, flow rate, and speed using high-precision sensors placed at key locations such as the boiler superheater, turbine bearings, and fuel pipelines. The sampling frequency is set at 100 Hz to capture millisecond-level disturbances.

[0053] S2. Perform short-term energy estimation. Input the raw data into the short-term fluctuation estimation module, which uses a sliding time window (window length 10 seconds, step length 0.5 seconds) to calculate the energy eigenvalue. Specifically, the square accumulation algorithm is used to quantify the instantaneous fluctuation intensity of the process variable:

[0054]

[0055] Among them, E i represents the energy value of the i-th sampling window, x j is the jth data point in the window, and N is the window capacity (2000 data points). This energy statistics method can effectively characterize the transient disturbance characteristics of the combustion system. For example, a coal feeder tripping event can cause the energy value to surge to 2.3 times the normal level within 3 seconds.

[0056] S3. Calculate the dynamic energy threshold. The energy characteristic value is transferred to the dynamic energy threshold calculation unit, which integrates the unit operation status and thermodynamic model.

[0057] The threshold is used as the core judgment benchmark input to the decision node, and the energy deviation between the controlled quantity and the set value is compared in real time. The dynamic threshold calculation is based on the thermodynamic energy conservation equation and the real-time load rate fitting:

[0058]

[0059] Wherein, L is the load rate (30%-100%); is the change in heat transfer rate; β0β1β2 are coefficients set manually based on experience. The coefficients can be initially taken as empirical values ​​and adjusted through actual operating experimental data of different units.

[0060] S4 determines whether the current window energy exceeds the dynamic threshold. If the boiler temperature error energy value exceeds the dynamic threshold, the system immediately activates the high-speed processing channel, compressing the sampling period from the baseline 1 second to 200 milliseconds, increasing the sampling density fivefold, until the energy value returns to the normal range. If the energy value is within the normal fluctuation range, the original sampling period is maintained. This dynamic sampling strategy reduced redundant data by 67% in actual measurements of a 660MW unit, while ensuring a data capture completeness rate exceeding 99% for disturbance events.

[0061] S5. Execute sampling and determine delay, establishing a load-rate-adaptive dual-threshold mechanism: When the load rate is less than 70%, the delay threshold is set to 25 seconds; when the load rate is greater than 70%, the delay threshold is increased to 45 seconds. This design breaks through the limitations of traditional fixed thresholds and accurately matches delay characteristics under varying operating conditions.

[0062] S6. Based on the delay judgment result, the dual-gated LSTM network starts the short-term gating channel or the long-term gating channel to process data. Figure 2 As shown in the figure, the dual-gated LSTM network splits the single forget gate of the traditional LSTM into a short-term forget gate and a long-term forget gate. The network adopts a three-layer architecture. The input layer receives time series data of temperature, pressure, flow, etc., with a sampling interval of 200ms-1s. The gating layer includes a short-term forget gate The input is the most recent 10-second window data, and the output is the gating coefficient γ short ∈[0,1]; long-term forget gate The input is 180 seconds of historical data, and the output is the gating coefficient γ long ∈[0,1]. The output layer predicts the temperature curve T for several steps in the future pred And control parameters [K p ,K i ,K d ].

[0063] The phased gating design of network parameter propagation is determined by the previous threshold judgment results: when in normal working conditions, the short-term gating channel is enabled, and the cell state update equation is:

[0064] c t =γ short ⊙c t-1 +i t ⊙g t

[0065] When the delay exceeds the threshold, the long-term gating channel is activated to strengthen the memory of the historical stable working conditions. The cell state update equation is:

[0066] c t =(λ·γ short +(1-λ)·γ long )⊙c t-1 +i t ⊙g t

[0067] Where, γ short , γ long Represent the output gating coefficients of the short-term forget gate and the long-term forget gate respectively; c t-1 Indicates the cell state at the previous moment; i t represents the input layer input of the current time step; g t Represents the newly generated candidate state at the current time step; λ is the adaptive weight, which is updated in the LSTM network with different event triggers, as shown in the following formula:

[0068]

[0069] Where σ is the Sigmoid activation function, W λ and b λ is the network weight matrix and bias parameter, which are iteratively calculated according to the following formula:

[0070]

[0071] Where L = || T pred -T actual || 2 is the prediction loss function of the controlled variable, namely boiler temperature, and η is the learning rate.

[0072] To ensure long-term memory, set the minimum value λ min =0.05; to prevent overfitting, set the maximum value λ max =0.95; at the same time, to increase the robustness of the system, the rate of change of λ is limited to less than ±0.2 / s.

[0073] S7. Dynamically adjust network parameters. To improve the response speed to coal quality fluctuations, the present invention dynamically adjusts the gate parameters through multivariable weights. Sensitivity weight α i Through online back-propagation update, the deep integration of thermodynamic mechanism and data-driven is achieved, as shown in the following formula:

[0074]

[0075] Where J is the loss function and η = 0.01 is the learning rate. For example, when the coal type is switched, the system automatically increases the fuel weight α FFrom 0.15 to 0.32.

[0076] According to the multivariable strong coupling characteristics of thermal power units, a thermodynamic sensitivity weight model is constructed:

[0077]

[0078] Where, T is the target control variable (such as boiler temperature), x i represents the coupled variables (including steam pressure, feed water flow, etc.), represents the absolute value of the partial derivative of the target control variable with respect to the coupled variable, where M is the total number of variables. This formula quantifies the influence intensity between variables by using thermodynamic partial derivatives, replacing the spectrum analysis of the pure signal dimension in the original figure. Taking a 1000MW unit as an example, the typical weight distribution is: temperature sensitivity weight α T =0.78, pressure sensitivity weight α P =0.15, traffic-sensitive weight α F =0.07, and satisfies ∑α i = 1.0. Generate a new weight matrix to drive the LSTM network to perform backpropagation rolling optimization parameters.

[0079] Under the conditions of dynamically blending Indonesian coal (4500kcal / kg) and Mongolian coal (5500kcal / kg), this design compresses the temperature fluctuation from ±6.2°C of the traditional long-short time LSTM network to ±1.4°C.

[0080] S8, perform rolling optimization and output control signal. The final control signal is generated by the model prediction MPC controller. For the predicted temperature curve T output by the dual-gated LSTM network pred and PID control parameters [K p ,K i ,K d The MPC controller uses the temperature prediction curve as the rolling optimization input of the MPC to establish the prediction time domain for the next 60 steps. The control increment matrix W of the MPC objective function is dynamically adjusted using the PID control parameters. u , when K p When it increases, the control increment weight is reduced, allowing a larger adjustment range (such as coal type switching conditions). The objective function of the MPC controller is:

[0081]

[0082] Among them, T pred (k) is the feedback value at time k, T set is the set value at time k, W T =diag(α T ,α P ,…) is the weight matrix, Δu=[Δufnel ,Δu air ] T To control the increment, W u To control the incremental weight matrix, the formula Secondary air door opening u air By calculating the combustion efficiency constraint,

[0083]

[0084] Where K represents the combustion efficiency compensation coefficient, u fuel is the fuel valve opening input, α F is the fuel quantity sensitivity weight, α T Temperature control takes precedence, employing a 60-step rolling optimization strategy. The controller outputs a fuel valve opening command every five seconds, simultaneously adjusting the secondary air damper opening. The actuator utilizes a high-speed electro-hydraulic servo system, completing actuation responses within 80 milliseconds—an eight-fold increase compared to traditional pneumatic actuators. After the control loop is closed, data alignment is performed: using a cubic spline interpolation algorithm, data with varying sampling rates is aligned to the original time base, eliminating the effects of time domain misalignment. The processed data is then output to the DCS system via the OPC protocol.

[0085] This invention optimizes control effects through an improved LSTM network and establishes an adaptive control architecture based on dynamic energy threshold decision-making, effectively addressing control instability issues under complex operating conditions such as deep peak shaving and coal type switching. Compared to traditional solutions, the present invention significantly suppresses overshoot oscillations, eliminates reliance on manual parameter adjustment, and significantly improves system stability and flexibility in complex scenarios such as deep peak shaving and coal blending. It also reduces equipment wear and maintenance costs, providing a core supporting technology for the intelligent transformation of thermal power generation. It also has broad engineering application value and promotion prospects, providing key data assurance capabilities for building highly dynamic and robust industrial intelligent control systems.

[0086] Example 2

[0087] As a specific implementation example of the present invention, this embodiment adopts the method described in Example 1 to perform spatiotemporal coupled intelligent control on a 660MW supercritical unit. The implementation process is as follows:

[0088] 1. System architecture and dynamic disturbance detection

[0089] Input variables include superheater outlet temperature (range 400-600°C), main steam pressure (0-30 MPa), coal feeder speed (0-1000 rpm), and flue gas oxygen content (0-15%). The dynamic delay estimation module uses a three-layer CNN-BiLSTM model (5×5 convolution kernel to extract local features and 64-unit bidirectional LSTM to capture temporal dependencies). The training data is 30 days of historical operating data of the unit (sampling interval 1 second), covering 100%-30% load fluctuation conditions.

[0090] When the simulated coal calorific value suddenly dropped by 15%, the outlet temperature of the A / B mills plummeted by 28°C in 3 seconds. The short-term energy estimation module calculated the energy surge value based on the temperature change rate (dT / dt = 9.3°C / s) and pressure fluctuation (ΔP = 1.7 MPa), reaching 2.3 times the normal operating condition. The dynamic threshold unit activated the 45-second high-sensitivity threshold due to the current load factor of 82%, and the decision module immediately switched to a 200ms high-speed sampling mode to capture disturbances early.

[0091] 2. Multivariable coordinated control and delay compensation

[0092] The system assigns temperature weight α through the weight matrix T =0.78, pressure α p =0.78, fuel quantity α F =0.78, quantifying the multivariable coupling relationship and prioritizing superheater safety.

[0093] In this embodiment, the short-term forget gate time constant τ short = 2.3s, long-term forget gate τ long = 41.7s, achieving decoupled control of disturbance response and steady-state memory. This structure achieves a prediction error of only 4.7% in an 80-second delay scenario, significantly improving performance compared to the 16.3% error of a standard LSTM network. After long-term memory gating is activated, it traces back the previous 180 seconds of stable operating data (load rate 80%-85%), reconstructs model parameters, and predicts the temperature decay curve for the next 45 seconds, correcting the Smith predictor model deviation (traditional methods can overshoot by more than 8% due to model mismatch). The controller uses this to generate the objective function.

[0094] Under the constraints of main steam pressure fluctuations less than 2.5 MPa and oxygen content greater than 3.5%, the optimized output fuel valve is opened 15% (from 42% to 57%), and the secondary air damper is simultaneously raised by 8% to balance combustion efficiency. The actuator completes its actuation within 80 milliseconds, three times faster than traditional pneumatic valves, effectively offsetting the 12.7 MW power shortfall.

[0095] 3. Comparison of control performance

[0096] From the data in Table 1, we can see that the method proposed in this invention has the following three advantages over the traditional PID system:

[0097] Temperature stability: Under disturbance, the lowest temperature only drops to 567℃, which is 20℃ higher than the trip protection value (547℃), and the overshoot is ±3.5℃.

[0098] Response speed: The disturbance suppression time is 38 seconds, which is 63.8% shorter than that of the unmodified units in the same plant.

[0099] Actuator life: The action frequency is reduced to 3 times / minute, and valve wear is reduced by 75%.

[0100] The edge computing latency is stable at 22±3ms (data preprocessing, model inference, and communication interaction), meeting the 50ms industrial real-time upper limit and eliminating computing bottlenecks.

[0101] Table 1 Comparison of data between the method of the present invention and the traditional system

[0102] index Traditional PID system Method of the present invention Improvement Lowest temperature 551℃ 567℃ +16℃ Disturbance suppression time 105 seconds 38 seconds 63.8％ Overshoot ±8.2 degrees ±3.5℃ 57.3％ Actuator action 12 times / minute 3 times / minute 75% drop

[0103] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0104] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. A spatiotemporal coupling control method for a thermal power plant based on a dual-gated LSTM network, characterized in that the steps include: Collect real-time data of thermal power units, including target control variables and multiple coupling variables; The real-time data of the thermal power unit is input into a dual-gated LSTM network to predict the temperature curve and control parameters for several steps in the future. The dual-gated LSTM network splits the forget gate of the LSTM into a short-term forget gate and a long-term forget gate. When in normal operating conditions, the short-term gating channel is enabled, and the long-term gating channel is activated when the delay exceeds the set delay threshold. Generate a thermodynamic sensitivity weight matrix based on the thermodynamic sensitivity weight model, and drive the dual-gated LSTM network to perform back-propagation rolling optimization parameters; The temperature prediction curve is used as the rolling optimization input, and the control increment matrix of the objective function of the model predictive controller is dynamically adjusted using the control parameters. The fuel valve opening and secondary air valve opening instructions are generated by the model predictive controller.

2. A spatiotemporal coupling control method for a thermal power unit based on a dual-gated LSTM network according to claim 1, characterized in that: The target controlled variable is the boiler temperature, and the coupled variables include the steam pressure and the feed water flow rate.

3. A spatiotemporal coupling control method for a thermal power unit based on a dual-gated LSTM network according to claim 1, characterized in that: The method compares the energy deviation of the target control variable with the set value in real time. When the energy value deviation exceeds the dynamic threshold, the high-speed processing channel is activated and the sampling period is compressed until the energy value returns to the normal range. The dynamic threshold is based on the thermodynamic energy conservation equation and the real-time load rate fitting: Wherein, L is the load rate; is the change in heat transfer rate; β0β1β2 are empirical coefficients.

4. A spatiotemporal coupling control method for a thermal power plant based on a dual-gated LSTM network according to claim 1, characterized in that: The delay threshold adopts a load rate adaptive dual threshold mechanism: when the load rate is less than a set value, the delay threshold is set to the second delay threshold; when the load rate is greater than the set value, the delay threshold is increased to the second delay threshold.

5. A spatiotemporal coupling control method for a thermal power plant based on a dual-gated LSTM network according to claim 1, characterized in that: The dual-gated LSTM network adopts a three-layer architecture: the input layer receives real-time data; the gating layer includes: a short-term forget gate The input is the most recent 10-second window data, and the output is the gating coefficient; the long-term forgetting gate f t long The input is 180 seconds of historical data, and the output is the gating coefficient; the output layer predicts the temperature curve and control parameters for several steps in the future.

6. A spatiotemporal coupling control method for a thermal power plant based on a dual-gated LSTM network according to claim 5, characterized in that: In the dual-gated LSTM network, the cell state update equation of the long-term gating channel is: c t =(λ·γ short +(1-λ)·γ long )⊙c t-1 +i t ⊙g t The cell state update equation for short-term gated channels is: c t =γ short ⊙c t-1 +i t ⊙g t Where, γ short , γ long Represent the output gating coefficients of the short-term forget gate and the long-term forget gate respectively; c t-1 Indicates the cell state at the previous moment; i t represents the input layer input of the current time step; g t Represents the newly generated candidate state at the current time step; λ is the adaptive weight, which is updated in the LSTM network with different event triggers.

7. A spatiotemporal coupling control method for a thermal power plant based on a dual-gated LSTM network according to claim 6, characterized in that: The rolling update of the adaptive weight is as follows: Where σ is the Sigmoid activation function, W λ and b λ is the network weight matrix and bias parameter, which are iteratively calculated according to the following formula: Where L is the prediction loss of the control variable and η is the learning rate.

8. A method for spatiotemporal coupling control of a thermal power plant based on a dual-gated LSTM network according to claim 5, characterized in that: The gating parameters of the dual-gated LSTM network are dynamically adjusted by the multivariate weights, and the sensitivity weight α i Through online back-propagation update, the deep integration of thermodynamic mechanism and data-driven is achieved, as shown in the following formula: Where η is the learning rate, J is the loss function, represents the partial derivative of the target control variable with respect to the coupling variable; α i is the thermodynamic sensitivity weight, calculated as follows: Among them, T is the target control variable; x i represents the coupled variable, It represents the absolute value of the partial derivative of the target control variable with respect to the coupling variable, and M is the total number of variables.

9. A spatiotemporal coupling control method for a thermal power plant based on a dual-gated LSTM network according to claim 1, characterized in that: The objective function of the model predictive controller is: Among them, T pred (k) is the feedback value of the target control quantity at time k; T set is the set value of the target control quantity at time k; W T =diag(α T ,α P ,…) is the weight matrix; W u To control the incremental weight matrix, the formula Δu=[Δu fuel ,Δu air ] T To control the increment, Δu fuel ,Δu air They are the increments of the fuel valve opening and the secondary air valve opening respectively.

10. A method for spatiotemporal coupling control of a thermal power unit based on a dual-gated LSTM network according to claim 9, characterized in that: The secondary air door opening u air Calculated by combustion efficiency constraints: Where K represents the combustion efficiency compensation coefficient, u fuel is the fuel valve opening input, α F is the fuel quantity sensitivity weight, α T Temperature controls the dominant weight.

Images