Dynamic load response adjusting method and system for hot water boiler

Abstract

The invention discloses a dynamic load response adjustment method and system for a hot water boiler, and relates to the field of boiler load adjustment, and the method comprises the steps: carrying out the deep learning of boiler and weather data through LSTM, and precisely predicting the thermal load change trend in a period of time in the future; on the basis, the current system state of the boiler is estimated in real time in combination with a Kalman filter, and the predicted load, the real-time state and a dynamic system model of the boiler are incorporated into a model predictive control (MPC) framework. And the MPC calculates an optimal gas valve opening sequence considering the physical time lag of the boiler through rolling optimization in a prediction time domain. The prospective prediction is combined with optimization control based on model dynamic characteristics, so that the boiler can adjust fuel supply in advance, response delay caused by system thermal inertia is effectively overcome, rapid and accurate response to dynamic loads is achieved, and the defect that traditional control is always slow by half beat is overcome.

Description

Technical Field

[0001] This application relates to the field of boiler load regulation, and more specifically, to a method and system for dynamic load response regulation of a hot water boiler. Background Technology

[0002] In energy-intensive industries such as oil extraction and processing, hot water boilers play a crucial role in heat supply. Particularly in heavy oil thermal recovery operations, large quantities of high-temperature hot water are injected into underground oil reservoirs to reduce crude oil viscosity and thus improve recovery rates. Similarly, in oil refining, precise temperature control is a core element in ensuring product quality and process safety. These industrial scenarios share the characteristic of phased and fluctuating production processes, leading to significant dynamic changes in the demand for hot water. Therefore, enabling hot water boilers, as a core piece of equipment, to efficiently and stably adapt to and respond to such dynamic loads directly impacts oilfield extraction efficiency, energy consumption costs, and overall production stability. Developing an advanced dynamic load response and regulation scheme is of great practical significance in this field.

[0003] However, current regulation schemes commonly used in hot water boilers generally suffer from inherent hysteresis when dealing with such dynamic loads. Existing control strategies, such as traditional PID (proportional-integral-derivative) feedback control, are essentially "post-event regulation" mechanisms. This mechanism only initiates regulation after detecting deviations in key parameters such as the supply water temperature from preset target values. This passive regulation method means that the control system's response always lags behind the actual load changes. When external load demand changes rapidly, although the control system can detect the deviation and issue commands, due to the inherent physical thermal inertia of the hot water boiler system, the manifestation of the regulation effect requires a process. By this time, the external load may have changed again, resulting in the system always being in a "half-beat" passive catch-up state, unable to achieve precise, real-time load matching. The reason for this fundamental problem is that a hot water boiler is a large and complex "water-fire-metal" coupled thermodynamic system. From the release of heat by fuel combustion, to the heat conduction through metal components such as the furnace and pipe walls, and finally to the heating of circulating water to change the supply water temperature, the entire energy transfer chain is long and slow, creating a significant time delay. This inherent time delay characteristic makes it difficult for traditional passive control methods to overcome effectively, thus severely restricting the boiler's operating efficiency and heating quality under dynamic conditions. Summary of the Invention

[0004] Considering the limitations mentioned above in application, according to one aspect of this application, a dynamic load response regulation method for a hot water boiler is provided, comprising: acquiring real-time raw boiler data and real-time raw weather data; performing heat load prediction based on an LSTM model on the real-time raw boiler data and the real-time raw weather data to obtain a predicted load curve from time t to time t+h; performing real-time system state estimation on the real-time raw boiler data based on a Kalman filter and combined with a boiler dynamic system model to obtain a system state vector at time t; constructing an MPC optimization problem based on the predicted load curve from time t to time t+h, the system state vector at time t, the boiler dynamic system model, and a constraint set, and using a numerical optimization algorithm to optimize and solve the MPC optimization problem to obtain an optimal control sequence from time t to time t+h-1; extracting the optimal control at time t from the optimal control sequence from time t to time t+h-1, and generating a gas valve opening control command based on the optimal control at time t.

[0005] According to another aspect of this application, a dynamic load response regulation system for a hot water boiler is provided, comprising: a boiler weather data acquisition module for acquiring real-time raw boiler data and real-time raw weather data; a load prediction module for performing heat load prediction based on an LSTM model on the real-time raw boiler data and the real-time raw weather data to obtain a predicted load curve from time t to time t+h; a system state estimation module for performing real-time system state estimation on the real-time raw boiler data based on a Kalman filter and combined with a boiler dynamic system model to obtain a system state vector at time t; an optimal control module for constructing an MPC optimization problem based on the predicted load curve from time t to time t+h, the system state vector at time t, the boiler dynamic system model, and a constraint set, and using a numerical optimization algorithm to optimize and solve the MPC optimization problem to obtain an optimal control sequence from time t to time t+h-1; and a control command generation module for extracting the optimal control at time t from the optimal control sequence from time t to time t+h-1, and generating a gas valve opening control command based on the optimal control at time t.

[0006] Compared with existing technologies, this application provides a dynamic load response regulation method and system for hot water boilers. Firstly, it utilizes LSTM (Laser-Based Memory) to perform deep learning on boiler and weather data, accurately predicting the trend of heat load changes over a future period, thereby achieving early perception of load demand and transforming passive response into proactive prediction. Based on this, it combines a Kalman filter to estimate the current system state of the boiler in real time, and incorporates the predicted load, real-time state, and boiler dynamic system model into a model predictive control (MPC) framework. MPC calculates the optimal gas valve opening sequence considering the boiler's physical time delay through rolling optimization within the prediction time domain. This combination of forward-looking prediction and optimization control based on model dynamic characteristics enables the boiler to adjust fuel supply in advance, effectively overcoming the response lag caused by system thermal inertia, achieving rapid and accurate response to dynamic loads, and avoiding the drawback of traditional control methods that are always half a step behind. Attached Figure Description

[0007] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0008] Figure 1 This is a flowchart of a dynamic load response adjustment method for a hot water boiler according to an embodiment of this application.

[0009] Figure 2 This is a schematic diagram of data flow in the dynamic load response adjustment method for a hot water boiler according to an embodiment of this application.

[0010] Figure 3 This is a flowchart of step S2 in the dynamic load response adjustment method for a hot water boiler according to an embodiment of this application.

[0011] Figure 4 This is a block diagram of a dynamic load response adjustment system for a hot water boiler according to an embodiment of this application. Detailed Implementation

[0012] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0013] In view of the limitations of the existing background technology, this application proposes a dynamic load response regulation method for hot water boilers. Figure 1 This is a flowchart of a dynamic load response adjustment method for a hot water boiler according to an embodiment of this application. Figure 2 This is a schematic diagram of data flow in the dynamic load response adjustment method for a hot water boiler according to an embodiment of this application. Figure 1 and Figure 2 As shown, the dynamic load response adjustment method for a hot water boiler according to an embodiment of this application includes: S1, acquiring real-time raw boiler data and real-time raw weather data; S2, performing heat load prediction based on an LSTM model on the real-time raw boiler data and the real-time raw weather data to obtain a predicted load curve from time t to time t+h; S3, performing real-time system state estimation on the real-time raw boiler data based on a Kalman filter and combined with a boiler dynamic system model to obtain a system state vector at time t; S4, constructing an MPC optimization problem based on the predicted load curve from time t to time t+h, the system state vector at time t, the boiler dynamic system model, and the constraint set, and using a numerical optimization algorithm to optimize and solve the MPC optimization problem to obtain the optimal control sequence from time t to time t+h-1; S5, extracting the optimal control at time t from the optimal control sequence from time t to time t+h-1, and generating a gas valve opening control command based on the optimal control at time t.

[0014] In step S1, real-time raw data of the boiler and real-time raw weather data are acquired. As described in the background art, hot water boilers exhibit significant physical thermal inertia, causing traditional control methods to exhibit response lag when dealing with dynamic load changes. To overcome this inherent defect and achieve proactive load prediction and precise adjustment, it is necessary to establish a comprehensive and real-time perception of the current boiler operating status and external environmental conditions. Real-time boiler data reflects the internal thermodynamic state and energy input and output of the boiler, serving as the basis for assessing its current operating efficiency and predicting future conditions; while real-time weather data directly affects the building's heat load demand and is a key input for predicting external load change trends. Only by acquiring this comprehensive real-time data can accurate and reliable basis be provided for subsequent processing, thereby supporting the optimization decisions of model predictive control (MPC), achieving proactive boiler control, and effectively solving the response lag problem.

[0015] Specifically, in one embodiment of this application, the boiler data at each time point in the real-time raw boiler data includes supply water temperature, return water temperature, fuel flow rate, and flue gas temperature; the weather data at each time point in the real-time raw weather data includes outdoor temperature value, humidity value, and wind speed value.

[0016] Step S1 is performed as follows: Acquiring real-time raw boiler data involves continuous monitoring of key boiler operating parameters. This includes, but is not limited to, supply water temperature, return water temperature, fuel flow rate, and flue gas temperature. Specifically, the supply water temperature is measured in real time using a high-precision temperature sensor installed on the boiler outlet pipe; the return water temperature is acquired by a temperature sensor installed on the boiler return water pipe. Fuel flow rate, such as the natural gas flow rate in a gas-fired boiler, is measured in real time using a high-precision flow meter, such as a turbine flow meter or mass flow meter, installed on the fuel supply line. Flue gas temperature is monitored using a high-temperature thermocouple or resistance temperature detector installed in the boiler flue to monitor combustion efficiency and exhaust losses. The analog signals collected by all these sensors are converted into digital signals by a data acquisition module, such as a PLC or the I / O module of a distributed control system (DCS), and timestamped and stored at a preset sampling frequency, for example, once every 10 seconds, forming continuous real-time raw boiler data.

[0017] Meanwhile, acquiring real-time raw weather data is crucial for capturing external environmental factors affecting heat load. This primarily includes outdoor temperature, humidity, and wind speed. Outdoor temperature and humidity can be measured in real-time using integrated temperature and humidity sensors installed on the exterior of buildings. These devices are weatherproof and can transmit data to the data acquisition unit via wired or wireless means. Wind speed is measured using wind speed sensors installed in open areas, such as cup anemometers or ultrasonic anemometers. Additionally, data can be interfaced with professional meteorological service providers to obtain the latest meteorological data periodically, for example, hourly or every 30 minutes. This weather data is also timestamped and synchronized or nearly synchronized with boiler data to ensure the accuracy of subsequent data fusion and model predictions.

[0018] In step S2, the real-time raw data of the boiler and the real-time raw weather data are used to perform heat load prediction based on an LSTM model to obtain the predicted load curve from time t to time t+h. Correspondingly, the physical thermal inertia of hot water boilers leads to a significant hysteresis in their response to load changes. To overcome this inherent defect and achieve a leap from passive response to active regulation, it is necessary to be able to predict future heat load demands in advance. Therefore, this application utilizes deep learning of historical boiler operation data and weather data—that is, an LSTM model—to capture complex time-series patterns and nonlinear relationships, thereby accurately predicting the heat load over a future period. This forward-looking predictive capability allows subsequent model predictive control (MPC) to plan the gas valve opening in advance, effectively offsetting the boiler's inertial effect, ensuring the stability of the water supply temperature and energy utilization efficiency, and fundamentally solving the response hysteresis problem.

[0019] Specifically, in one embodiment of this application, Figure 3This is a flowchart of step S2 in the dynamic load response adjustment method for a hot water boiler according to an embodiment of this application. Figure 3 As shown, step S2, which involves performing heat load prediction based on an LSTM model on the real-time raw data of the boiler and the real-time raw data of the weather to obtain the predicted load curve from time t to time t+h, includes: S21, extracting boiler state time-series features based on an LSTM model from the real-time raw data of the boiler to obtain a boiler state time-series pattern feature encoding vector; S22, extracting weather state time-series features based on an LSTM model from the real-time raw data of the weather to obtain a weather state time-series pattern feature encoding vector; S23, fusing the boiler state time-series pattern feature encoding vector and the weather state time-series pattern feature encoding vector to obtain a boiler state-weather state time-series pattern joint encoding vector; and S24, decoding and predicting the boiler state-weather state time-series pattern joint encoding vector to obtain the predicted load curve from time t to time t+h.

[0020] Step S2 is performed as follows: First, step S21. The real-time raw boiler data includes multiple dimensions such as supply water temperature, return water temperature, fuel flow rate, and flue gas temperature, which exist in continuous time series form. Before being input into the LSTM model, this raw data needs to be preprocessed, such as data normalization. This is done by scaling data of different dimensions to similar numerical ranges, such as between 0 and 1. For example, supply water temperature, such as 60-80℃, and fuel flow rate, such as 0-100%, can be min-max normalized separately. Subsequently, these normalized data are organized into a fixed-length time series as input to the LSTM model. For example, each input sequence can be set to contain boiler data from the past 10 minutes, i.e., 60 time steps (if the sampling frequency is once every 10 seconds). Next, an LSTM model for feature extraction is constructed. This model uses a multi-layer stacked LSTM network structure to capture the complex nonlinear temporal dependencies in the boiler operation data. Specifically, the network contains two LSTM layers, each containing 128 hidden units. The first LSTM layer receives the preprocessed boiler data sequence as input and passes its output to the second LSTM layer. Internally, the LSTM layer uses its unique gating mechanism—input gate, forget gate, and output gate—to selectively remember or forget historical information, effectively handling long-term dependencies in long-sequence data. At the very end of the model, a fully connected layer maps the hidden state or output of the last LSTM layer to a fixed-dimensional feature vector. The specific encoding process is as follows: when the real-time raw boiler data is input into the trained LSTM model, the data propagates forward through the LSTM layers time-step by time. At each time step, the LSTM unit updates its memory based on the current input and its internal memory states, including hidden states and unit states. After the entire input sequence has been processed, the final hidden state of the last LSTM layer or the output of the fully connected layer connected to it is extracted. This fixed-dimensional numerical vector is the boiler state temporal pattern feature encoding vector. It is worth noting that all weights, biases, and other internal parameters in this LSTM model are obtained through offline training on a large amount of historical boiler operation data.

[0021] Next is step S22. This process is similar in principle and implementation to the feature extraction from boiler data in step S21. Specifically, the real-time raw weather data obtained in the previous steps is first used as input. Before being input into the LSTM model, this raw weather data also needs necessary preprocessing, such as data normalization. Next, an LSTM model for weather feature extraction is constructed. This model also uses a multi-layered stacked LSTM network structure, designed to capture complex nonlinear temporal dependencies in weather data and their potential impact on heat load. For example, a network containing two LSTM layers can be designed, each containing 64 hidden units. The first LSTM layer receives the preprocessed weather data sequence as input and passes its output to the second LSTM layer. Internally, the LSTM layer, through its unique gating mechanism, can effectively process and remember long-term dependencies in long-sequence weather data, such as the impact of seasonal temperature changes or consecutive days of high humidity on building heat load. At the very end of the model, a fully connected layer is connected to map the hidden state or output of the last LSTM layer to a fixed-dimensional feature vector. The specific encoding process is as follows: When real-time weather data is input into the trained LSTM model, the data is propagated forward through the LSTM layers step by step. At each time step, the LSTM unit updates its memory based on the current input and its internal memory state. After the entire input sequence has been processed, the final hidden state of the last LSTM layer is extracted. This fixed-dimensional numerical vector is the weather state temporal pattern feature encoding vector. All weights, biases, and other internal parameters in this LSTM model are obtained through offline training on a large amount of historical weather data and corresponding actual heat load data.

[0022] The next step is S23. Specifically, in one embodiment of this application, step S23, fusing the boiler state time-series pattern feature encoding vector and the weather state time-series pattern feature encoding vector to obtain a joint encoding vector of boiler state-weather state time-series patterns, includes: inputting the boiler state time-series pattern feature encoding vector and the weather state time-series pattern feature encoding vector into a fully connected layer to obtain the joint encoding vector of boiler state-weather state time-series patterns. First, the boiler state time-series pattern feature encoding vector and the weather state time-series pattern feature encoding vector are concatenated. Concatenation means connecting two independent vectors in a dimension to form a longer single vector. For example, connecting two vectors in a dimension to form a longer joint feature vector. Then, this joint feature vector is input into a fully connected layer. This fully connected layer further processes the concatenated feature vector through linear transformation and nonlinear activation functions such as ReLU to learn a deeper level of interaction between the internal state of the boiler and external weather conditions, and finally outputs a fixed-dimensional joint encoding vector of boiler state-weather state time-series patterns. Specifically, the weights and biases of this fully connected layer are optimized through backpropagation during the entire LSTM model training process.

[0023] Specifically, considering that the subsequent decoding and prediction of the joint encoding vector of the boiler state-weather state time series patterns differs from the usual single-value regression, the predicted load curve from time t to time t+h essentially corresponds to the predicted load under a series of continuous time series states. This requires that the dynamic mapping of the feature trajectory during the decoding process possess sequence-level joint smoothness. If confusion and overlap occur between different local time series patterns, it will lead to discontinuities or unreasonable jumps in the decoded load curve, thereby hindering the continuity and accuracy of the prediction sequence. Therefore, local pattern aliasing compensation is required before fusion and decoding. This is to integrate the originally dispersed potential state transition paths on the basis of smooth and continuous joint posterior states, making them develop towards a stable and clear convergent evolution trajectory, ultimately avoiding confusion and overlap of different local time series patterns in the feature vector, significantly improving the prediction effect of the load curve in the future, and ensuring the stability and efficiency of boiler control.

[0024] Based on this, in a preferred embodiment of this application, step S23, inputting the boiler state time-series pattern feature encoding vector and the weather state time-series pattern feature encoding vector into a fully connected layer to obtain the boiler state-weather state time-series pattern joint encoding vector, includes: first, constructing a first hidden state transition matrix of the boiler state time-series pattern feature encoding vector relative to the weather state time-series pattern feature encoding vector; and constructing a second hidden state transition matrix of the weather state time-series pattern feature encoding vector relative to the boiler state time-series pattern feature encoding vector, i.e.: ;in, It is the boiler state time-series pattern feature encoding vector. It is a feature encoding vector of weather state time series pattern. It is matrix multiplication. It is the first hidden state transition matrix. This is the second hidden state transition matrix. It should be understood that this is used to represent the hidden state transition relationships between different local time-series patterns. In practice, some form of matrix operation or transformation is needed to capture how the boiler state time-series pattern feature encoding vector is related to the weather state time-series pattern feature encoding vector through the first hidden state transition matrix, and how the weather state time-series pattern feature encoding vector is related to the boiler state time-series pattern feature encoding vector through the second hidden state transition matrix. These hidden state transition matrices represent the potential laws governing the mutual influence and transformation between the boiler's internal state patterns and external weather patterns. By constructing these matrices, the interdependence between the two feature patterns can be preliminarily quantified and represented, laying the foundation for subsequent joint smoothing.

[0025] Next, the first hidden state transition matrix and the second hidden state transition matrix are subjected to joint smoothing of the hidden state space dynamic global transition to obtain the boiler-weather state joint posterior transition state representation matrix, that is: ;in, It is subtracted based on position. To calculate the F-norm value, To take the absolute value, It is added based on the position point. It is a regularization weight parameter, which encourages... and To a certain extent, maintain consistency, making The smoothest and most reliable path representing the co-evolution of boiler and weather conditions in the latent space is determined through cross-validation or empirical trial and error during model training or offline debugging. For example, it can be set to 0.1. It is the joint posterior transition state representation matrix of boiler-weather state. Correspondingly, the separately constructed... and Local inconsistencies or noise may exist, failing to fully reflect the global smooth evolution of the boiler-weather joint state. Therefore, this application employs joint smoothing to impose global transition confidence constraints on local differences in transition states, and accordingly uses local dynamic prior confidence as a regularization term. This results in a more robust and globally consistent joint posterior transition state representation, leading to a matrix that reflects the smooth and continuous evolution of the boiler-weather joint state over time, providing a unified benchmark for subsequent feature compensation.

[0026] Then, the boiler state temporal pattern feature encoding vector and the weather state temporal pattern feature encoding vector are respectively compensated for temporal consistency constraints using the boiler-weather state joint posterior transition state representation matrix to obtain the boiler state temporal pattern compensated feature encoding vector and the weather state temporal pattern compensated feature encoding vector, that is: ;in, It is the boiler state time-series pattern compensation feature encoding vector. This is the weather state temporal pattern compensation feature encoding vector. That is, to encourage a smooth and continuous evolution trajectory of features under the joint posterior transition state, thereby directly eliminating local pattern aliasing, this application performs joint dynamic transition smoothing compensation on the original vector based on the boiler-weather state joint posterior transition state representation matrix. In other words, this represents... It was used to correct the original boiler state time series pattern feature encoding vector and weather state time series pattern feature encoding vector, making their temporal evolution more consistent with... This represents a smooth, consistent joint transition pattern. Through this compensation, if the original feature vectors exhibit local jumps or anomalies at a certain point in time that deviate from the overall smooth trend, they will be... The applied temporal consistency constraints pull the system back onto a more reasonable trajectory. Corrected boiler state temporal pattern compensation feature encoding vectors and weather state temporal pattern compensation feature encoding vectors are generated. These not only preserve the original feature information, but more importantly, their internal temporal patterns are smoothed and unified, avoiding confusion and overlap between different local patterns.

[0027] Finally, the boiler state time-series pattern compensation feature encoding vector and the weather state time-series pattern compensation feature encoding vector are input into the fully connected layer to obtain the joint encoding vector of the boiler state-weather state time-series patterns. In other words, after local pattern aliasing compensation, these two vectors have eliminated internal noise and inconsistencies, each representing a pure and smooth time-series pattern of the boiler and weather conditions. The fully connected layer can learn and capture the complex nonlinear relationships between these independent but interconnected feature vectors, organically integrating the boiler's internal thermodynamic state with the influence of external environmental factors on the heat load. This deep fusion not only extracts more discriminative joint features but also effectively reveals the deep coupling mechanism between boiler operation and weather changes. As input to the subsequent heat load prediction decoder, this ensures that the decoded predicted load curve has high continuity and physical rationality throughout the entire prediction time domain, thereby significantly improving prediction accuracy and providing a more reliable basis for the dynamic load response adjustment of hot water boilers. Specifically, the process of this step is the same as the implementation process described above.

[0028] Finally, step S24. The joint encoding vector of the boiler state-weather state time series pattern is used as the initial input to the decoder. This joint encoding vector highly condenses the information on the impact of the current boiler operating state and external weather conditions on the future heat load. Specifically, the decoder uses one or more Long Short-Term Memory (LSTM) layers as its core structure to handle the sequence generation task. For example, a decoder containing two LSTM layers, each containing 128 hidden units, can be designed. The initial hidden state and unit state of the decoder can be initialized by the joint encoding vector through a linear transformation, such as a fully connected layer, or the joint encoding vector can be directly used as the input to the first time step of the decoder. The decoding prediction process is a time-step sequence generation. At time t, the decoder receives the joint encoding vector as input and generates the predicted heat load value at time t. Subsequently, this predicted value can be used as the input to the decoder at the next time step t+1, and combined with the decoder's own internal state, the predicted heat load value at time t+1 is generated. This process continues for h time steps until the complete predicted load curve from time t to time t+h is generated. After the LSTM layer output at each time step, a fully connected layer is connected, mapping the hidden states of the LSTM to a single predicted heat load value. Specifically, the prediction time domain h is set based on the physical thermal inertia of the boiler system, the typical cycle of load changes, and the control system's need for future information, determined through experience, simulation, or optimization analysis to balance prediction accuracy and control performance. For example, if the prediction time domain h is set to 24 time steps, each representing 15 minutes, the decoder will generate a sequence of predicted heat load values ​​every 15 minutes for the next 6 hours. In particular, all weights, biases, and other internal parameters in this decoder model are obtained during the end-to-end training process of the entire LSTM-based heat load prediction framework.

[0029] In step S3, the system state is estimated in real time based on the boiler's real-time raw data using a Kalman filter and combined with the boiler dynamic system model to obtain the system state vector at time t. It should be understood that although the boiler's real-time raw data provides rich measurement information, these measurements are often affected by sensor noise, transmission interference, and other factors, and are not entirely accurate. More importantly, some key internal state variables of the boiler, such as furnace tube temperature, may not be directly measurable in real time by sensors, or the measurement cost may be too high. However, these unmeasured state variables are crucial for accurately understanding the boiler's current thermodynamic state and predicting its future behavior, especially in model predictive control (MPC), where accurate initial state information is needed to construct the optimization problem. The Kalman filter can effectively fuse noisy measurement data and noisy mathematical model predictions, thereby providing a real-time, optimal estimate of the boiler's dynamic state, overcoming the limitations of single measurement or pure model prediction, and providing a high-confidence system state vector for subsequent MPC optimization.

[0030] Specifically, in one embodiment of this application, the boiler dynamic system model is represented as follows: ;in, For time, This indicates taking the derivative with respect to time. For water supply temperature, For furnace tube temperature, The return water temperature, For ambient temperature, Fuel supply rate, For hot water flow rate, and The process noise represents the impact of random disturbances or measurement errors not considered in the model on changes in water supply temperature and boiler tube temperature. The metal-water heat transfer coefficient reflects the efficiency of heat exchange between the metal and water. The heat removal coefficient of supply and return water is related to the specific heat capacity and flow rate of water, and represents the amount of heat removed per unit flow rate and temperature difference. The fuel-metal heat conversion coefficient reflects the efficiency of fuel combustion and the efficiency of heat transfer to the furnace tubes. The metal-environment heat dissipation coefficient reflects the thermal insulation performance and heat dissipation area of ​​the boiler shell. The water-metal endothermic coefficient is related to the specific heat capacity and mass of the metal, reflecting the rate at which the metal's temperature decreases due to heat transfer to water. Specifically, , , , and A systematic identification method is employed, which involves analyzing a large amount of historical data collected during the actual operation of the boiler to determine its characteristics. For example, It is 0.05℃ / (min·℃). 0.002℃ / (min·(m) 3 / h)·℃), It is 0.1℃ / (min·%). It is 0.001℃ / (min·℃). The value is 0.08℃ / (min·℃). These numerical examples are for illustrative purposes only, and the actual values ​​will vary depending on the specific boiler model, design, insulation conditions, and operating conditions.

[0031] Specifically, the boiler dynamic system model is a simplified set of second-order ordinary differential equations used to describe the dynamic characteristics of the supply water temperature and furnace tube temperature changing over time during the operation of a hot water boiler. This model treats the boiler as having two main heat capacities: the furnace tubes and the circulating water, and is established based on the principle of energy balance. The first equation describes the rate of change of the supply water temperature. The term indicates that the heat transferred from the furnace tube metal to the circulating water causes an increase in the supply water temperature. The difference between the furnace tube temperature and the supply water temperature is the driving force for heat transfer. The term indicates that the circulating water carries away heat, causing a decrease in the supply water temperature. The second equation describes the rate of change of the boiler tube temperature. The term indicates that the heat released by fuel combustion is transferred to the furnace tube metal, causing the furnace tube temperature to rise. The term indicates that the heat lost from the furnace tube metal to the environment leads to a decrease in the furnace tube temperature. The term indicates that the furnace tube metal transfers heat to the circulating water, causing a decrease in the furnace tube temperature. This is consistent with the first equation. The physical meaning is the same: heat is transferred from metal to water. However, here it is reflected from the perspective of the metal's energy balance.

[0032] Step S3 is executed as follows: It is worth noting that the Kalman filter is a recursive linear quadratic Gaussian estimator capable of providing real-time and accurate estimation of the state of a dynamic process, achieving optimal estimation even in the presence of noise. It works iteratively through two main steps: prediction and update. The prediction step uses a known system dynamic model to predict the current state; the update step uses real-time measurement data to correct the predicted state, resulting in a more accurate estimate. Internally, the Kalman filter includes a state transition model to describe how the state evolves over time, a measurement model to describe how measurements relate to the state, and a statistical description of process noise and measurement noise, represented by a covariance matrix.

[0033] In practical implementation, the first step is to define the boiler's system state vector and measurement vector. Based on the aforementioned boiler dynamic system model, the system state vector can be defined as including the supply water temperature. and furnace tube temperature A set, such as The measurement vector consists of directly obtainable real-time raw data from the boiler, such as... That is, the actual measured water supply temperature and return water temperature To apply a continuous-time boiler dynamics system model to a discrete-time Kalman filter, it needs to be discretized, transforming it into a mathematical relationship describing how the boiler state at the current moment evolves based on the previous moment's state, applied control inputs (such as fuel supply rate, hot water flow rate, ambient temperature, etc.), and stochastic process noise. This yields the state transition matrix F and the control input matrix B. For example, the differential equation can be transformed into... In the form of, To control inputs, such as fuel supply rate, hot water flow rate, and ambient temperature, The process noise is a random vector that is randomly sampled from a probability distribution at each time step. This represents the system state vector at discrete time step t. Simultaneously, a measurement model needs to be established. ,in This is a measurement matrix; the elements of the matrix are determined based on the actual sensor measurement principles and physical relationships. To measure noise , This represents the measurement vector at discrete time step t.

[0034] The Kalman filter estimation process is iterative. At each time step t, a prediction step is first performed: using the optimal state estimate x_{t-1|t-1} from the previous time step and its covariance P_{t-1|t-1}, combined with the discretized boiler dynamic model, the current state x_{t|t-1} and its covariance P_{t|t-1} are predicted. This step is essentially based on the model's projection of the future state, reflecting prior knowledge of the boiler's dynamic behavior. The prediction equation is: ,as well as ,in It is the process noise covariance matrix, which reflects the impact of model uncertainty and unmodeled dynamics.

[0035] Subsequently, an update step is performed: when the real-time raw measurement data y_t of the boiler at time t is available, the Kalman filter calculates the Kalman gain K_t. The Kalman gain determines the weight of the predicted and measured values ​​in the final estimate, and it is dynamically adjusted based on the prediction covariance P_{t|t-1} and the measurement noise covariance R. The measurement noise covariance R reflects the reliability of the sensor measurement. The settings of the two covariance matrices, Q and R, are crucial; they are determined through statistical analysis of historical data, empirical settings, or offline optimization, such as maximum likelihood estimation. For example, the value of R can be determined by statistical analysis of sensor measurement errors, while Q can be adjusted by analyzing model prediction errors or through trial and error. Finally, using the Kalman gain K_t and the measurement residual (the difference between the actual and predicted measurements), the predicted state x_{t|t-1} is corrected to obtain the optimal state estimate x_{t|t} and its covariance P_{t|t} at the current time. The update equation is: ,as well as ,and The final system state vector at time t is obtained. This means that it is the optimal estimate of key state variables such as boiler water supply temperature and furnace tube temperature.

[0036] In step S4, an MPC optimization problem is constructed based on the predicted load curve from time t to time t+h, the system state vector at time t, the boiler dynamic system model, and the constraint set. A numerical optimization algorithm is then used to solve the MPC optimization problem to obtain the optimal control sequence from time t to time t+h-1. It should be understood that although step S2 provides an accurate prediction of future heat load, and step S3 provides an accurate estimate of the boiler's current state, traditional feedback control methods are still insufficient to convert this information into actual control commands and effectively overcome the inherent physical thermal inertia of the boiler. In this application, model predictive control, as an advanced control strategy, can proactively utilize future load prediction information, combined with the boiler's dynamic model, and, under the premise of satisfying various operating constraints, continuously optimize control inputs (such as gas valve opening) online, thereby achieving a rapid and stable response to dynamic loads and fundamentally solving the response hysteresis problem mentioned in the background art.

[0037] Step S4 is performed as follows: The model predictive control optimization problem is a constrained optimization problem, the core of which lies in defining an objective function and a series of constraints. The objective function aims to minimize the deviation between the boiler output (e.g., water supply temperature) and the predicted load (target water supply temperature) in the predicted time domain, while considering the stability of the control input and energy consumption. Specifically, the objective function is constructed as: minimizing the weighted sum of the squared errors between the model-predicted water supply temperature and the target water supply temperature over the next H time steps, plus the weighted sum of the squared changes in the fuel supply rate. The model-predicted water supply temperature is derived from the predicted load curve from time t to time t+h, estimated by a Kalman filter under different fuel supply rates, based on the boiler dynamic system model and the system state vector at the current moment. The target water supply temperature is derived from the predicted load curve from time t to time t+h. The weighting coefficients of each term, such as the weight of the water supply temperature deviation and the weight of the fuel supply rate change, need to be set according to actual operating requirements to balance tracking performance, energy efficiency, and control stability. For example, the weight of water supply temperature deviation can be set to 1 to prioritize ensuring heating quality; the weight of fuel supply rate change can be set to 0.1 to avoid excessively frequent or drastic control actions. These weights are obtained through empirical setting or repeated debugging and optimization in offline simulation.

[0038] In addition to the objective function, the optimization problem must also incorporate a series of constraints to ensure the safe and efficient operation of the boiler. Specifically, in one embodiment of this application, the constraint set includes a target water supply temperature greater than or equal to 55°C and less than or equal to 75°C, a current target water supply temperature of 75°C, a fuel supply rate greater than or equal to 0% and less than or equal to 100%, an absolute value of the difference between the fuel supply rate at time k and the fuel supply rate at time k-1 less than or equal to 5%, and a predicted water supply temperature at time k less than or equal to 85°C. Specifically, the setting of the target water supply temperature range (55°C to 75°C) and the current target water supply temperature (75°C) aims to ensure that the heating system can meet the user's comfort requirements while taking energy efficiency into account and avoiding unnecessary excessively high temperature output. The upper and lower limits of the fuel supply rate (0% to 100%) are inherent physical limitations of actuators such as gas valves, ensuring that control commands are within the equipment's capabilities. The limit on the rate of change of fuel supply rate at time k (absolute difference less than or equal to 5%) is to prevent overly drastic control actions. This helps protect the burner and boiler body, reduces mechanical wear and thermal stress, improves operational stability and safety, and avoids combustion instability caused by excessive fuel fluctuations. Finally, the upper limit of the predicted water supply temperature at time k (85°C) is a key safety constraint designed to prevent boiler overheating, avoid dangerous situations such as steaming or equipment damage, and ensure that the boiler operates within a safety margin. These values ​​are determined and adjusted comprehensively through boiler manufacturer's recommended parameters, industry standards, historical operating data analysis, and the experience of field engineers. These constraints collectively define the feasible solution space of the optimization problem in the prediction time domain.

[0039] After the model is constructed, numerical optimization algorithms are used to solve the predictive control optimization problem of the model. Since the boiler dynamic model may contain nonlinear terms and multiple constraints, this is a typical nonlinear programming problem. Commonly used numerical optimization algorithms include sequential quadratic programming, interior-point methods, or gradient-based optimization methods. These algorithms iteratively calculate, starting from an initial guess, and gradually adjust the control variables (i.e., the fuel supply rate for the next h time steps) to find the optimal solution that minimizes the objective function, while satisfying all constraints. For example, in each iteration, the sequential quadratic programming algorithm approximates the complex nonlinear problem into a more easily solvable quadratic programming problem, then solves this approximation to determine the next search direction and step size, and updates along this direction until the convergence criterion is met, finding the optimal control sequence that satisfies all constraints. Finally, the output of this step is the optimal control sequence from time t to time t+h-1, a sequence containing the optimal fuel supply rate for the next h time steps, providing a precise basis for subsequent control command generation.

[0040] In step S5, the optimal control at time t is extracted from the optimal control sequence from time t to time t+h-1, and a gas valve opening control command is generated based on the optimal control at time t. Correspondingly, although the optimization algorithm calculates the optimal control sequence for the next h time steps, the predictions and estimates for future times become more accurate over time due to dynamic changes in the boiler operating environment, uncertainties between the model and the actual process, and the continuous influx of new measurement data. Therefore, MPC does not apply the entire optimal control sequence to the boiler, but only extracts and executes the first control action in the sequence, i.e., the optimal control at the current time t. In the next control cycle, the entire optimization process restarts, using the latest real-time data for load forecasting and state estimation, and solving a new optimization problem again to generate a new optimal control sequence. This rolling decision-making mechanism enables the control strategy to continuously adapt to environmental changes, improving the robustness and accuracy of control.

[0041] Specifically, step S5 is performed as follows: First, the output from step S4 is received, namely the optimal control sequence from time t to time t+h-1. This sequence is a vector containing h elements, each element representing the optimal fuel supply rate at the corresponding time. The optimal control at time t is extracted from the optimal control sequence from time t to time t+h-1, i.e., the optimal fuel supply rate at the current time t, which is a percentage value, such as 50% or 75%. This extracted value represents the ideal level of gas supply that should be achieved at the current time in order for the boiler to optimally respond to the predicted load and meet all operating constraints in the future.

[0042] Secondly, based on the extracted optimal fuel supply rate at time t, a gas valve opening control command is generated. The fuel supply rate is an abstract percentage value, while the gas valve opening is the actual control quantity of the physical actuator. It may be an opening percentage of 0-100%, or a corresponding analog signal, such as a 4-20mA current signal, or a digital signal. Therefore, a mapping relationship is needed to convert the fuel supply rate into the actual gas valve opening command. This mapping relationship is obtained through calibration during the boiler design or commissioning phase. It may be a linear relationship or a nonlinear lookup table, depending on the characteristics of the gas valve and the design of the boiler burner. For example, if there is a linear relationship between the fuel supply rate and the gas valve opening, then the optimal control at time t, i.e., the optimal fuel supply rate at the current time t, may correspond to a gas valve opening of 50%. If it is a nonlinear relationship, it needs to be converted using a lookup table or a nonlinear function. For example, when the optimal fuel supply rate at the current time t is 70%, after conversion, a command corresponding to a gas valve opening of 65% may be generated, or a specific current value (such as 16mA) may be sent to the actuator.

[0043] Ultimately, the output of this step is a specific control command that can be directly sent to the boiler's gas valve actuator. This command can be an analog signal, such as voltage or current, or a digital signal, such as a pulse-width modulation signal, or a digital quantity transmitted via industrial communication protocols such as Modbus or Profibus. Once the command is generated, it is sent to the boiler's gas valve actuator, thereby adjusting the opening of the gas valve in real time, which in turn changes the fuel supply, enabling the boiler's heat output to dynamically track the predicted heat load demand. This process is repeated in each control cycle, for example, every 10 seconds or every minute, ensuring that the boiler is always in an optimal operating state, achieving a rapid and accurate response to dynamic loads.

[0044] In summary, the dynamic load response regulation method for hot water boilers based on the embodiments of this application is explained. First, it utilizes LSTM to perform deep learning on boiler and weather data to accurately predict the trend of heat load changes over a future period, thereby achieving early perception of load demand and transforming passive response into proactive prediction. Based on this, a Kalman filter is used to estimate the current system state of the boiler in real time, and the predicted load, real-time state, and boiler dynamic system model are incorporated into a model predictive control (MPC) framework. MPC calculates the optimal gas valve opening sequence considering the physical time delay of the boiler through rolling optimization within the prediction time domain. This combination of forward-looking prediction and optimization control based on model dynamic characteristics enables the boiler to adjust fuel supply in advance, effectively overcoming the response lag caused by system thermal inertia, achieving rapid and accurate response to dynamic loads, and avoiding the drawback of traditional control methods that are always half a step behind.

[0045] Figure 4 This is a block diagram of a dynamic load response adjustment system for a hot water boiler according to an embodiment of this application. Figure 4As shown, the hot water boiler dynamic load response regulation system 100 according to an embodiment of this application includes: a boiler weather data acquisition module 110, used to acquire real-time raw boiler data and real-time raw weather data; a load prediction module 120, used to perform heat load prediction based on an LSTM model on the real-time raw boiler data and the real-time raw weather data to obtain a predicted load curve from time t to time t+h; a system state estimation module 130, used to perform real-time system state estimation on the real-time raw boiler data based on a Kalman filter and combined with a boiler dynamic system model to obtain a system state vector at time t; an optimal control module 140, used to construct an MPC optimization problem based on the predicted load curve from time t to time t+h, the system state vector at time t, the boiler dynamic system model, and a constraint set, and use a numerical optimization algorithm to optimize and solve the MPC optimization problem to obtain an optimal control sequence from time t to time t+h-1; and a control command generation module 150, used to extract the optimal control at time t from the optimal control sequence from time t to time t+h-1, and generate a gas valve opening control command based on the optimal control at time t.

[0046] As described above, the hot water boiler dynamic load response regulation system 100 according to the embodiments of this application can be implemented in various wireless terminals, such as servers with hot water boiler dynamic load response regulation algorithms. In one possible implementation, the hot water boiler dynamic load response regulation system 100 according to the embodiments of this application can be integrated into the wireless terminal as a software module and / or hardware module. For example, the hot water boiler dynamic load response regulation system 100 can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the hot water boiler dynamic load response regulation system 100 can also be one of many hardware modules of the wireless terminal.

[0047] Alternatively, in another example, the hot water boiler dynamic load response regulation system 100 and the wireless terminal can also be separate devices, and the hot water boiler dynamic load response regulation system 100 can be connected to the wireless terminal via wired and / or wireless networks, and transmit interactive information in accordance with an agreed data format.

[0048] Here, those skilled in the art will understand that the specific operations of each step in the above-mentioned dynamic load response regulation system for hot water boilers have been referenced above. Figures 1 to 3 The dynamic load response regulation method for hot water boilers has been described in detail, and therefore, its repeated description will be omitted.

Claims

1. A method for dynamic load response regulation of a hot water boiler, characterized in that, include: Obtain real-time raw data of the boiler and real-time raw weather data; The real-time raw data of the boiler and the real-time raw data of the weather are used to perform heat load prediction based on the LSTM model to obtain the predicted load curve from time t to time t+h. Based on a Kalman filter and combined with a boiler dynamic system model, the real-time system state of the boiler is estimated in real time to obtain the system state vector at time t. Based on the predicted load curve from time t to time t+h, the system state vector at time t, the boiler dynamic system model, and the constraint set, an MPC optimization problem is constructed, and a numerical optimization algorithm is used to optimize and solve the MPC optimization problem to obtain the optimal control sequence from time t to time t+h-1. The optimal control at time t is extracted from the optimal control sequence from time t to time t+h-1, and a gas valve opening control command is generated based on the optimal control at time t.

2. The dynamic load response regulation method for hot water boilers according to claim 1, characterized in that, The boiler data at each time point in the real-time raw boiler data includes supply water temperature, return water temperature, fuel flow rate, and flue gas temperature; the weather data at each time point in the real-time raw weather data includes outdoor temperature, humidity, and wind speed.

3. The dynamic load response adjustment method for hot water boilers according to claim 2, characterized in that, The method involves performing LSTM-based heat load prediction on the real-time raw boiler data and the real-time raw weather data to obtain the predicted load curve from time t to time t+h. This includes: extracting boiler state temporal features from the real-time raw boiler data using an LSTM model to obtain a boiler state temporal pattern feature encoding vector; extracting weather state temporal features from the real-time raw weather data using an LSTM model to obtain a weather state temporal pattern feature encoding vector; fusing the boiler state temporal pattern feature encoding vector and the weather state temporal pattern feature encoding vector to obtain a boiler state-weather state temporal pattern joint encoding vector; and decoding and predicting the boiler state-weather state temporal pattern joint encoding vector to obtain the predicted load curve from time t to time t+h.

4. The dynamic load response adjustment method for hot water boilers according to claim 3, characterized in that, The process of fusing the boiler state temporal pattern feature encoding vector and the weather state temporal pattern feature encoding vector to obtain a joint encoding vector of boiler state-weather state temporal patterns includes: inputting the boiler state temporal pattern feature encoding vector and the weather state temporal pattern feature encoding vector into a fully connected layer to obtain the joint encoding vector of boiler state-weather state temporal patterns.

5. The dynamic load response regulation method for hot water boilers according to claim 4, characterized in that, The process of inputting the boiler state temporal pattern feature encoding vector and the weather state temporal pattern feature encoding vector into a fully connected layer to obtain the boiler state-weather state temporal pattern joint encoding vector includes: constructing a first hidden state transition matrix of the boiler state temporal pattern feature encoding vector relative to the weather state temporal pattern feature encoding vector; constructing a second hidden state transition matrix of the weather state temporal pattern feature encoding vector relative to the boiler state temporal pattern feature encoding vector; performing hidden state space dynamic overall transition joint smoothing on the first hidden state transition matrix and the second hidden state transition matrix to obtain a boiler-weather state joint posterior transition state representation matrix; applying temporal consistency constraint compensation to the boiler state temporal pattern feature encoding vector and the weather state temporal pattern feature encoding vector respectively using the boiler-weather state joint posterior transition state representation matrix to obtain a boiler state temporal pattern compensation feature encoding vector and a weather state temporal pattern compensation feature encoding vector; and inputting the boiler state temporal pattern compensation feature encoding vector and the weather state temporal pattern compensation feature encoding vector into a fully connected layer to obtain the boiler state-weather state temporal pattern joint encoding vector.

6. The dynamic load response regulation method for hot water boilers according to claim 1, characterized in that, The boiler dynamic system model is represented as follows: ;in, For time, This indicates taking the derivative with respect to time. For water supply temperature, For furnace tube temperature, The return water temperature, For ambient temperature, Fuel supply rate, For hot water flow rate, and For process noise, The metal-water heat transfer coefficient is... The coefficient for heat removal from supply and return water. The fuel-metal heat conversion coefficient, The metal-environment heat dissipation coefficient, The water-metal heat absorption coefficient is given.

7. The dynamic load response adjustment method for a hot water boiler according to claim 6, characterized in that, The constraint set includes a target water supply temperature greater than or equal to 55°C and less than or equal to 75°C, a current target water supply temperature of 75°C, a fuel supply rate greater than or equal to 0% and less than or equal to 100%, an absolute value of the difference between the fuel supply rate at time k and the fuel supply rate at time k-1 less than or equal to 5%, and a predicted water supply temperature at time k less than or equal to 85°C.

8. A dynamic load response regulation system for a hot water boiler, characterized in that, include: The boiler weather data acquisition module is used to acquire real-time raw boiler data and real-time raw weather data. The load forecasting module is used to perform heat load forecasting based on the LSTM model on the real-time raw data of the boiler and the real-time raw data of the weather to obtain the predicted load curve from time t to time t+h. The system state estimation module is used to perform real-time system state estimation on the boiler's real-time raw data based on a Kalman filter and a boiler dynamic system model to obtain the system state vector at time t. The optimal control module is used to construct an MPC optimization problem based on the predicted load curve from time t to time t+h, the system state vector at time t, the boiler dynamic system model, and the constraint set, and to use a numerical optimization algorithm to optimize and solve the MPC optimization problem to obtain the optimal control sequence from time t to time t+h-1. The control command generation module is used to extract the optimal control at time t from the optimal control sequence from time t to time t+h-1, and to generate a gas valve opening control command based on the optimal control at time t.

Images