Intelligent building heating intelligent optimization operation method and system based on deep learning

Abstract

The invention relates to the technical field of intelligent buildings and energy management, and particularly discloses an intelligent building heating intelligent optimization operation method and system based on deep learning, and the method comprises the steps: collecting building environment parameters and equipment operation data in real time through a distributed optical fiber sensing network and an infrared thermal imaging system; extracting minute-level fluid transmission and distribution parameters and hour-level building thermal inertia characteristics by adopting wavelet packet transformation and a graph neural network; then, constructing a neural differential equation prediction model fused with physical constraints, and outputting high-precision thermal load demand prediction through a differentiable heat conduction operator coupling multi-time scale feature; then, a mixed integer optimization model considering equipment life loss is established, and boiler start-stop combination and pipe network flow distribution are synchronously optimized by adopting a hierarchical decision-making mechanism; and finally, closed-loop control is realized through a multi-mode actuator network, and model parameters are dynamically adjusted in combination with an online learning mechanism.

Description

Technical Field

[0001] This invention relates to the field of intelligent building and energy management technology, specifically to a method and system for intelligent building heating optimization based on deep learning. Background Technology

[0002] With the acceleration of urbanization and the continuous growth of building energy consumption, the intelligent and energy-saving optimization of building heating systems has become an important issue in building energy management. Traditional heating systems mostly adopt rule-based control strategies or simple PID regulation, which are difficult to cope with the influence of multi-dimensional factors such as building thermal inertia, meteorological changes, and fluctuations in heat demand, resulting in energy waste and uneven heating quality. In recent years, the development of Internet of Things (IoT) technology and artificial intelligence has provided new ideas for heating system optimization. Data-driven predictive control methods are gradually replacing experience-based regulation. However, existing technologies still have problems such as insufficient model accuracy, single optimization objectives, and neglect of equipment losses, which restrict the improvement of overall system performance.

[0003] Existing optimization control technologies primarily focus on minimizing short-term energy consumption, neglecting the impact of equipment start-up and shutdown on lifespan. Frequent adjustments accelerate the aging of critical components. Traditional methods separate continuous regulation from discrete decision-making, making it difficult to achieve globally optimal hydraulic-thermal matching. Closed-loop control lacks dynamic evaluation of performance and parameter adaptation mechanisms, leading to gradual performance degradation over long-term operation. These issues make it difficult for existing systems to simultaneously address multiple objectives such as energy efficiency, comfort, and equipment reliability, necessitating a novel optimization method that integrates physical laws and data intelligence. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for intelligent optimization of building heating based on deep learning, so as to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A deep learning-based intelligent building heating system optimization operation method includes the following steps:

[0007] S1. Real-time acquisition of building environmental parameters and equipment operation data, extraction of dynamic features at multiple time scales, wherein the dynamic features at multiple time scales include minute-level fluid transport parameters and hour-level building thermal inertia parameters;

[0008] S2. Construct a deep time-series prediction model that integrates physical constraints, and couple the minute-level fluid transport parameters with the hour-level building thermal inertia parameters through neural differential equations to output the predicted heat load demand for future periods;

[0009] S3. Based on the predicted heat load demand, establish a mixed integer optimization model, take equipment operating cost and life loss as the joint objective function, simultaneously optimize continuous adjustment variables and discrete start-stop decision variables, and generate heating control commands;

[0010] S4. Dynamically adjust the water supply temperature, flow rate, and equipment start / stop status according to the heating control command to achieve closed-loop optimized operation of the building heating system.

[0011] As a further aspect of the present invention: the extraction process of the multi-timescale dynamic features specifically includes:

[0012] The temperature data of pipes in each zone of the building are collected in real time through a distributed optical fiber temperature measurement network. The original temperature signal is decomposed into feature components at different time scales using a wavelet packet transform algorithm. The first feature component corresponds to the minute-level fluid dynamic fluctuations, and the second feature component corresponds to the hour-level building thermal inertia changes.

[0013] Based on the pipeline pressure data obtained by the hydraulic sensor array, a fluid distribution map neural network model is constructed. Minute-level hydraulic coupling features are extracted by the node embedding method. The hydraulic coupling features include branch flow distribution coefficient and pipeline pressure drop gradient.

[0014] Infrared thermal imagers are used to monitor the surface temperature field of building envelopes. Combined with thermal parameters in building information models, hourly building thermal inertia parameters are obtained through three-dimensional finite element heat conduction inversion calculations. The thermal inertia parameters include the equivalent heat capacity of the wall and the window-to-wall ratio correction factor.

[0015] After spatiotemporal alignment of the first feature component, the second feature component, the hydraulic coupling feature, and the thermal inertia parameter, the input is fed into the feature fusion module, which outputs minute-level fluid transport parameters and hour-level building thermal inertia parameters with clear physical meaning.

[0016] As a further aspect of the present invention: the method for constructing a deep temporal prediction model that incorporates physical constraints includes:

[0017] A dual-channel neural differential equation framework is constructed. The first channel processes minute-level fluid transport parameters and uses a convolutional long short-term memory network to extract local hydraulic dynamic features. The second channel processes hour-level building thermal inertia parameters and uses an attention-based gated recurrent unit network to capture long-term heat conduction patterns.

[0018] Building thermodynamic constraints are embedded in the hidden layer of the neural differential equation. The features of the two channels are coupled at the physical level through a differentiable heat conduction operator to ensure that the prediction results comply with the law of conservation of energy.

[0019] A multi-task learning mechanism is adopted to simultaneously optimize the prediction accuracy and the physical consistency loss function, wherein the physical consistency loss function is calculated by comparing the deviation between the predicted heat flux density and the actual sensor measurement value;

[0020] The coupled neural differential equations are solved by an adaptive time-step numerical solver, and the predicted heat load demand sequence for multiple future time scales is output.

[0021] As a further aspect of the present invention: the method for embedding physical constraints in the neural differential equation framework specifically includes:

[0022] An equivalent thermal network model of the building envelope is established, and the wall thermal resistance and window heat transfer coefficient are incorporated as learnable physical parameters into the neural network weight initialization process.

[0023] In each time step of the neural differential equation calculation, thermal equilibrium constraints are enforced, and a differentiable heat transfer operator is used to ensure that the predicted rate of change of heat load does not exceed the maximum thermal conductivity of the building envelope.

[0024] The physical information regularization method is adopted to add a residual term of the partial differential equation based on Fourier's heat conduction law to the loss function to constrain the learning direction of the neural network.

[0025] By generating virtual samples that conform to physical laws through adversarial training mechanisms, the generalization ability of the model under extreme conditions is enhanced.

[0026] As a further aspect of the present invention, the method for generating the multi-timescale heat load demand prediction sequence specifically includes:

[0027] A hierarchical time attention mechanism is designed to focus on recent hydraulic fluctuation characteristics in the minute-level prediction subnetwork and on long-term meteorological change trends in the hour-level prediction subnetwork.

[0028] Temporal convolutional networks are used to extract temporal features at different scales, and dilated convolutions are used to expand the receptive field to capture periodic heat usage patterns.

[0029] A post-processing correction module for the prediction results is established to dynamically adjust the output value based on the statistical analysis of historical prediction errors, thereby eliminating systematic biases.

[0030] An uncertainty quantification method is introduced to provide confidence intervals for the prediction results, thus providing risk warning information for subsequent optimization control.

[0031] As a further aspect of the present invention: the specific method for establishing the mixed integer optimization model includes:

[0032] A multi-objective optimization framework is constructed, which quantifies equipment operating costs as the product of energy consumption costs and time-of-use electricity prices, and quantifies equipment lifespan loss as a weighted function of start-up and shutdown times and operating load.

[0033] The design employs a hierarchical decision-making mechanism. The upper-level optimization uses a branch pricing algorithm to solve the boiler group start-up and shutdown combination, while the lower-level optimization uses the interior point method to solve the flow distribution scheme under the hydraulic balance condition of the pipeline network.

[0034] The equipment health status constraints are embedded in the optimization model, and the minimum operating time limit of the boiler is dynamically adjusted by real-time collected vibration spectrum data.

[0035] Establish a rolling optimization window mechanism to adaptively adjust the optimization time domain length based on the prediction confidence interval, ensuring that a more conservative control strategy is adopted when the prediction uncertainty increases.

[0036] As a further aspect of the present invention: the method for quantifying equipment lifespan loss specifically includes:

[0037] A damage accumulation model for key boiler components is established based on the theory of material fatigue, with the rate of temperature change and the amplitude of pressure fluctuation as damage factors.

[0038] The expansion of internal defects in the equipment is monitored in real time using acoustic emission sensors, and the life loss weighting coefficient is dynamically adjusted.

[0039] Establish a sub-model for predicting the remaining service life of equipment, and input historical operating conditions and current state parameters into a convolutional neural network to predict the remaining service life;

[0040] By introducing the equipment residual value term into the optimization objective function, joint optimization of operating costs and asset depreciation can be achieved.

[0041] As a further aspect of the present invention, the implementation method of the hierarchical decision-making mechanism specifically includes:

[0042] Construct a device combination state tree and use a depth-first search strategy to traverse feasible start-stop combination schemes;

[0043] Design heuristic pruning rules to eliminate non-optimal branches in advance based on equipment efficiency curves and pipeline topology;

[0044] Establish a preprocessing module for hydraulic-thermal coupling to quickly assess the system feasibility under various start-stop combinations;

[0045] A parallel computing architecture is adopted to accelerate the optimization process and ensure that all computing tasks are completed within the control cycle.

[0046] As a further aspect of the present invention: the specific method for closed-loop optimized operation includes:

[0047] A multimodal actuator collaborative control network is constructed, and optimization instructions are received through a distributed programmable logic controller cluster. The water supply temperature regulation adopts a feedforward-feedback composite control algorithm, and the flow regulation adopts a fuzzy adaptive PID control algorithm.

[0048] Design an equipment start-stop sequence optimization module to dynamically calculate the optimal start-stop time interval based on equipment thermal status monitoring data, so as to avoid equipment damage caused by thermal shock;

[0049] Establish a real-time evaluation mechanism for control effectiveness, collect the actual room temperature distribution of each zone through a high-precision temperature sensor array, and calculate the deviation of control command execution.

[0050] A dynamic parameter adjustment strategy is implemented, which automatically triggers the online learning mechanism for model parameters when the deviation exceeds the threshold, updating the internal parameters of the prediction model and the optimization model.

[0051] A deep learning-based intelligent building heating system for optimized operation includes:

[0052] A multi-scale data acquisition and feature extraction module is used to acquire building environment parameters and equipment operation data in real time and extract dynamic features at multiple time scales.

[0053] The physical constraint deep prediction module is used to construct a deep time series prediction model that integrates physical constraints. It couples the minute-level fluid transport parameters with the hour-level building thermal inertia parameters through neural differential equations to output the predicted heat load demand value for future periods.

[0054] The mixed integer optimization decision module establishes a mixed integer optimization model based on the predicted heat load demand, takes equipment operating cost and life loss as the joint objective function, and simultaneously optimizes continuous adjustment variables and discrete start-stop decision variables to generate heating control commands.

[0055] The closed-loop execution and adaptive adjustment module dynamically adjusts the water supply temperature, flow rate, and equipment start / stop status according to the heating control command, thereby realizing the closed-loop optimized operation of the building heating system.

[0056] The beneficial effects of this invention are:

[0057] (1) This invention utilizes multi-source sensing technologies such as a distributed fiber optic temperature measurement network, a hydraulic sensor array, and an infrared thermal imaging system to collect building environmental parameters and equipment operation data in real time. It then employs a six-layer wavelet packet transform algorithm and a graph neural network model to extract dynamic features across multiple time scales, including minute-level fluid transport parameters and hourly-level building thermal inertia parameters. Based on these features, a deep time-series prediction model incorporating physical constraints is constructed. This model, through an innovative dual-channel neural differential equation framework, combines an improved long short-term memory network and a gated recurrent unit network with a bidirectional attention mechanism to achieve accurate prediction of future heat load demand. Furthermore, this prediction result is input into a mixed-integer optimization model, using equipment operating costs and lifespan loss as the joint objective function. Continuous adjustment variables and discrete start-stop decision variables are simultaneously optimized, ensuring that energy costs are reduced while extending equipment lifespan and minimizing the risk of mechanical damage caused by frequent start-stop operations or rapid temperature changes. These measures not only significantly improve the energy efficiency and operational safety of the heating system, but also ensure the comfort and stability of indoor temperature through intelligent closed-loop control strategies. This is particularly important in special areas such as hospital wards, ensuring that the high standards of the medical environment are met. This approach, which comprehensively utilizes advanced sensing technology, deep learning models, and optimization algorithms, represents a highly efficient, intelligent, and sustainable building heating solution.

[0058] (2) This invention innovatively introduces a multimodal collaborative control architecture at the control strategy level, employing a feedforward-feedback composite control algorithm for fine-tuning the water supply temperature. The feedforward control adjusts the boiler outlet temperature setpoint in advance based on a heat load prediction model, while the feedback control uses a high-precision temperature transmitter to detect pipeline temperature deviations in real time, achieving dynamic closed-loop correction and significantly improving the system's response speed and control accuracy. Regarding flow regulation, the system applies an improved fuzzy adaptive PID control algorithm. This algorithm can automatically identify changes in hydraulic characteristics based on the pipeline network's operating status and dynamically adjust proportional, integral, and derivative parameters, effectively overcoming the performance degradation problem of traditional PID control under varying operating conditions and ensuring the stability and adaptability of flow distribution. Simultaneously, by using an infrared thermal imager installed on the boiler body to monitor the metal wall temperature distribution in real time, the system constructs a thermal state evolution model of the equipment and optimizes the equipment start-up and shutdown sequence accordingly: automatically extending the preheating time during cold starts to avoid thermal stress concentration, and controlling the cooling rate within a safe range during hot shutdowns, thereby effectively preventing mechanical damage to the equipment caused by unsteady-state operation. The system also deploys a distributed temperature monitoring system based on a distributed fiber optic temperature measurement network and high-precision digital temperature sensors. Combined with an incremental online learning mechanism, it can evaluate control effectiveness in real time, identify sources of control deviations, and dynamically correct model parameters using a sliding window least squares method, ensuring the system's long-term stability and adaptability. Furthermore, the system supports dedicated drive algorithms for various types of actuators, such as position feedback fuzzy control for electric regulating valves, vibration monitoring feedback control for variable frequency water pumps, and self-optimizing air-fuel ratio control for burners, comprehensively improving the control quality and equipment safety at the actuator level. This series of intelligent and adaptive closed-loop control strategies not only significantly improves the operating efficiency and stability of the heating system but also achieves high precision and reliability in temperature control in special scenarios with extremely high environmental temperature control requirements, such as hospital wards. It fully meets the high standards of comfort, stability, and safety required in medical environments, marking a significant step towards intelligent and refined control of building heating systems. Attached Figure Description

[0059] The invention will now be further described with reference to the accompanying drawings.

[0060] Figure 1 This is a flowchart of the intelligent building heating intelligent optimization operation method based on deep learning according to the present invention;

[0061] Figure 2 This is a flowchart of the intelligent building heating intelligent optimization operation system based on deep learning in this invention. Detailed Implementation

[0062] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] Please see Figure 1 As shown, this invention is a method for intelligent optimization of building heating operation based on deep learning, comprising the following steps:

[0064] S1. Real-time acquisition of building environmental parameters and equipment operation data, extraction of dynamic features at multiple time scales, wherein the dynamic features at multiple time scales include minute-level fluid transport parameters and hour-level building thermal inertia parameters;

[0065] S2. Construct a deep time-series prediction model that integrates physical constraints, and couple the minute-level fluid transport parameters with the hour-level building thermal inertia parameters through neural differential equations to output the predicted heat load demand for future periods;

[0066] S3. Based on the predicted heat load demand, establish a mixed integer optimization model, take equipment operating cost and life loss as the joint objective function, simultaneously optimize continuous adjustment variables and discrete start-stop decision variables, and generate heating control commands;

[0067] S4. Dynamically adjust the water supply temperature, flow rate, and equipment start / stop status according to the heating control command to achieve closed-loop optimized operation of the building heating system.

[0068] In S1, building environmental parameters and equipment operation data are collected in real time, and multi-timescale dynamic features are extracted. These multi-timescale dynamic features include minute-level fluid transport parameters and hour-level building thermal inertia parameters, specifically including:

[0069] First, a distributed fiber optic temperature measurement network is used to monitor the temperature of heating pipes in each zone of the building in real time. This network employs wavelength scanning distributed fiber optic sensing technology, deploying sensing fibers along the heating pipes to continuously acquire surface temperature distribution data with a spatial resolution of 1 meter and a temporal resolution of 10 seconds. The acquired raw temperature signals are transmitted to a central processing unit, where a six-layer wavelet packet transform algorithm is used for multi-scale decomposition. The third-layer detail coefficients correspond to fluid dynamic fluctuation characteristics at a time scale of 2–5 minutes, while the sixth-layer approximation coefficients correspond to building thermal inertia changes at a time scale of 4–6 hours. This decomposition method effectively separates the physical process characteristics at different time scales contained in the pipe temperature signals.

[0070] Secondly, pressure data is acquired using a hydraulic sensor array deployed at key nodes of the pipeline network. This array contains 32 high-precision pressure sensors that record real-time pressure values ​​at each node of the pipeline network at a sampling frequency of 1Hz. This pressure data is then input into a pre-trained graph neural network model. This network employs a message-passing mechanism and extracts pipeline network topology features through three layers of graph convolution operations. In the node embedding layer, the flow distribution coefficient of each branch pipe is represented as a nonlinear function of the pressure difference between the nodes at both ends of that branch. Simultaneously, the pressure drop gradient between adjacent nodes is calculated as an important indicator of hydraulic coupling characteristics. This processing method can accurately reflect the fluid transport and distribution dynamics on a minute-scale timescale in the pipeline network.

[0071] Then, an infrared thermal imaging system was used to scan and monitor the building envelope. This system, equipped with a high-resolution infrared camera and a 3D laser scanner, performs a panoramic scan of the building's exterior walls, roof, and windows once per hour. The acquired surface temperature field data, along with thermal parameters from the building information model (including wall thickness and material thermal conductivity), were input into finite element analysis software. The equivalent thermal parameters of the building envelope were solved using a reverse heat conduction calculation method. Specifically, the equivalent heat capacity of the walls was calculated by comparing the rate of surface temperature change with the indoor-outdoor temperature difference, while the window-to-wall ratio correction factor was dynamically adjusted based on the temperature difference between the window and wall areas. These parameters accurately characterize the building's thermal inertia on an hourly timescale.

[0072] Finally, the features extracted in the above steps are spatiotemporally aligned and fused. Temporal alignment uses a sliding window matching algorithm to unify data from different sampling frequencies onto a common time reference; spatial alignment is achieved using a building 3D coordinate system. The feature fusion module employs an attention mechanism to dynamically assign weights to features from different sources: minute-level fluid dynamic fluctuation features have a weight of 0.4, hour-level thermal inertia change features have a weight of 0.3, hydraulic coupling features have a weight of 0.2, and thermal inertia parameters have a weight of 0.1. The fused output retains the physical meaning of features at each time scale while eliminating interference from sensor noise, providing high-quality input data for subsequent prediction models.

[0073] During the above implementation process, the arrangement density and sampling frequency of each sensor were rigorously tested and determined. The fiber optic temperature measurement network was laid with a fiber spacing of no more than 2 meters to ensure the capture of local temperature changes in the pipeline; the placement of the hydraulic sensor array was optimized through pipeline hydraulic calculations to cover all key branch nodes; the infrared thermal imaging system achieved a scanning resolution of 0.1℃, enabling the detection of minute temperature differences in the building envelope. This optimized combination of technical parameters ensured the accuracy and reliability of dynamic feature extraction across multiple time scales.

[0074] In S2, a deep time-series prediction model incorporating physical constraints is constructed. This model couples minute-level fluid transport parameters with hourly-level building thermal inertia parameters via neural differential equations, outputting predicted heat load demands for future periods. Specifically, this includes:

[0075] First, a dual-channel neural differential equation framework is constructed, consisting of two parallel feature processing channels. The first channel specifically handles minute-level fluid transport parameters, employing an improved long short-term memory (LSTM) network architecture with 3D convolutional kernels. This network uses 5×5 convolutional kernels to capture the spatial correlation of the pipe temperature field in the spatial dimension, and learns the short-term fluctuation patterns of fluid dynamics through a memory gating mechanism in the temporal dimension. The second channel handles hour-level building thermal inertia parameters, employing a gated recurrent unit network with a bidirectional attention mechanism. This network effectively captures the long-term evolution trend of building thermal inertia through temporal scans in both forward and backward directions, combined with an attention weight allocation mechanism. The intermediate layer outputs of the two channels are dimensionally aligned through a specific data interface, preparing for subsequent coupled computations.

[0076] In the design of the hidden layers of the neural differential equations, building thermodynamic constraints are innovatively embedded. Specifically, a differentiable heat conduction operator is added to the output of each hidden layer. This operator calculates the theoretical heat transfer rate in real time based on the thermal performance parameters of the building envelope. During forward propagation, the features extracted by the neural network are weighted and fused with the theoretical values ​​calculated by the heat conduction operator. During backward propagation, gradient calculation of the operator ensures that the update direction of the network parameters conforms to the laws of thermodynamics. This design guarantees that the prediction results possess both the flexibility of a data-driven model and strictly adhere to fundamental physical laws such as energy conservation.

[0077] The model training process employs a multi-task learning mechanism, simultaneously optimizing two objective functions. The first is a conventional prediction accuracy loss function, using smoothed mean absolute error as the evaluation metric. The second is a physical consistency loss function, calculated by comparing the deviation between the model's predicted heat flux density value and the actual sensor measurement. These two loss functions are dynamically balanced using adaptive weighting coefficients, prioritizing prediction accuracy in the early stages of training and gradually increasing the weight of physical constraints as training progresses. This training strategy ensures both the model's prediction accuracy and that the prediction results conform to physical laws.

[0078] The model is solved using an adaptive time-step numerical computation method. For minute-level prediction tasks, a fixed-step fourth-order Runge-Kutta method is used; for hourly-level prediction tasks, a variable-step implicit integration algorithm is employed. Before each solution, the system automatically evaluates the dynamic characteristics of the input data and selects the most suitable numerical computation scheme. This adaptive solution strategy ensures both computational efficiency and numerical stability. The final output prediction results include minute-level heat load fluctuation curves for the next 6 hours and hourly heat load trend predictions for the next 24 hours.

[0079] Regarding the specific embedding method of physical constraints, an equivalent thermal network model of the building envelope is first established. This model divides the walls into several thermal resistance-heat capacity units, and simplifies windows to equivalent heat transfer coefficients. These physical parameters are encoded into the network structure during the initialization of the neural network weights, serving as non-trainable prior knowledge. During model operation, thermal balance constraints are enforced at each time step, ensuring that the predicted rate of change of heat load does not exceed the maximum thermal conductivity of the building envelope through a differentiable heat transfer operator. This constraint is not a simple post-processing correction, but is deeply integrated into the forward computation process of the neural network.

[0080] To improve the physical consistency of the model, a physical information regularization method was employed during training. A residual term based on Fourier's law of heat conduction was added to the loss function, calculated using automatic differentiation techniques. Specifically, several spatial-temporal points were sampled from the prediction results, the partial derivatives of the temperature field with respect to time and space at these points were calculated, and then substituted into the heat conduction equation to calculate the residuals. This regularization method does not rely on additional labeled data; the physical equations themselves effectively constrain the learning direction of the neural network.

[0081] To enhance the model's generalization ability under extreme conditions, an adversarial training mechanism was employed to generate virtual samples. First, a generative adversarial network (GAN) was trained, whose generator learned to produce virtual operating condition data that conformed to physical laws, while the discriminator determined whether the data satisfied thermodynamic constraints. These virtual data were then added to the training set, enabling the prediction model to learn extreme operating condition features that are difficult to observe in normal data. This data augmentation method significantly improved the model's predictive robustness under abnormal weather conditions such as cold waves.

[0082] A hierarchical temporal attention mechanism was designed for the generation of multi-timescale prediction sequences. The minute-level prediction sub-network uses sliding window attention, focusing on the hydraulic fluctuation characteristics within the most recent 30 minutes; the hour-level prediction sub-network uses global attention, focusing on the meteorological change trends over the past 24 hours. The attention weights of the two sub-networks are dynamically adjusted through a gating mechanism to ensure that the most relevant features are captured at different time scales.

[0083] For feature extraction, an improved temporal convolutional network architecture is employed. This network uses dilated convolutional kernels with different dilation rates to process the input sequence in parallel. Kernels with small dilation rates capture short-term fluctuations, while kernels with large dilation rates identify long-term periodic patterns. The outputs of each convolutional branch are fused through attention-weighted fusion to form a unified representation with multi-scale temporal features. This design is particularly suitable for handling complex temporal patterns such as daily and weekly cycles present in heating systems.

[0084] The post-processing of the prediction results employs a dynamic correction algorithm. The system maintains a prediction error database, recording the deviations between historical predictions and actual measurements. Before outputting a new prediction result, the average error under similar operating conditions is queried from this database, and then the current prediction value is compensated accordingly. This post-processing method effectively eliminates systematic prediction biases caused by data distribution shifts.

[0085] In S3, a mixed-integer optimization model is established based on the predicted heat load demand. Using equipment operating costs and lifespan losses as the joint objective function, it simultaneously optimizes continuous adjustment variables and discrete start-stop decision variables to generate heating control commands, specifically including:

[0086] First, a multi-objective optimization framework is constructed, which quantifies both equipment operating costs and lifespan losses. Operating costs are calculated using a time-of-use electricity pricing model, multiplying energy price coefficients for different time periods by real-time data from gas flow meters and electricity meters to obtain energy costs accurate to the minute. Equipment lifespan losses are calculated using an established fatigue damage model, which quantifies the mechanical losses of each boiler start-up and shutdown into equivalent operating hours, while also considering the accelerated aging effect of operating load rates on key components. The optimal weighting ratio for the two objective functions is determined using Pareto optimal frontier analysis, and the weighting coefficients are dynamically adjusted based on the equipment health status before each optimization.

[0087] A hierarchical decision-making mechanism is adopted in the optimization algorithm design. The upper-level optimization focuses on the discrete decision-making problem of boiler group start-up and shutdown combinations, and is solved using an improved branch-and-price algorithm. This algorithm first constructs a solution space containing all possible start-up and shutdown combinations, and then gradually adds potential new solutions through column generation techniques. In the lower-level optimization, for each feasible start-up and shutdown combination, the pre-processed interior-point method is used to solve the flow allocation problem under the constraints of the network hydraulic balance equations. This hierarchical structure effectively reduces the computational complexity of mixed-integer programming problems, making real-time optimization possible.

[0088] The key innovation of this invention lies in the embedding of equipment health status constraints. Vibration acceleration sensors are installed at critical locations in the boiler room to collect mechanical vibration signals from the equipment in real time. Characteristic frequency components are extracted using Fast Fourier Transform (FFT) to establish a mapping relationship between the vibration spectrum and the equipment's fatigue level. When abnormal vibration characteristics are detected, the minimum running time constraint in the optimization model is automatically tightened to prevent frequent start-ups and shutdowns of the equipment in a sub-healthy state. Simultaneously, the system records the temperature change curves during each start-up and shutdown, imposing penalties for rapid temperature changes exceeding safe thresholds.

[0089] The rolling optimization window mechanism fully considers the impact of forecast uncertainty. The system monitors the confidence interval width of the predicted heat load value in real time. When the interval width exceeds a preset threshold, it automatically shortens the optimization time domain length and increases the safety margin. Specifically, a 24-hour optimization time domain is used under normal operating conditions, shortened to 8 hours when uncertainty is high, and further reduced to 4 hours under extreme uncertainty. This adaptive adjustment strategy ensures both optimization effectiveness and system operational safety.

[0090] In terms of quantifying equipment lifespan loss, a multi-parameter damage model was established based on materials fatigue theory. This model decomposes the damage accumulation process of the main pressure-bearing components of the boiler into three parts: basic operating loss, start-up and shutdown impact loss, and accelerated loss under abnormal operating conditions. The rate of temperature change is measured by a thermocouple array installed on the boiler body, while the pressure fluctuation amplitude is obtained through a high-frequency sampling pressure sensor. After normalization, these two parameters are input into a formula based on Miner's linear cumulative damage theory to calculate the real-time damage coefficient.

[0091] The application of acoustic emission monitoring systems further improves the accuracy of lifespan assessment. Broadband acoustic emission sensors are deployed in key components such as boiler drums and headers to capture elastic wave signals emitted by the propagation of micro-cracks within the materials in real time. Signal features are extracted using wavelet packet analysis to establish a correlation model between crack propagation rate and operating parameters. When abnormal acoustic emission signals are detected, the lifespan loss weighting coefficient is automatically increased to reflect the accelerated aging of the equipment.

[0092] Equipment remaining life prediction is achieved using deep learning. A neural network model with a three-dimensional convolutional structure is constructed, with inputs including historical operating parameters, real-time monitoring data, and equipment archive information. The model output is the failure probability distribution of key components at different future time points, and this data is converted into residual value coefficients and incorporated into the optimization objective function. This approach achieves a precise correspondence between the economic value and physical state of the equipment, providing a scientific basis for optimization decisions.

[0093] The hierarchical decision-making mechanism is implemented using a state tree search algorithm. First, a state tree of equipment combinations is constructed based on boiler performance parameters and pipeline topology, with each node representing a possible start-up / shutdown combination. The search process employs a depth-first search strategy, while also applying multiple heuristic pruning rules: efficiency threshold pruning eliminates combinations with excessively low thermal efficiency, hydraulic feasibility pruning removes schemes that may cause hydraulic imbalance in the pipeline network, and lifespan balancing pruning avoids overuse of certain equipment. These pruning rules significantly improve search efficiency.

[0094] The hydraulic-thermal coupling preprocessing module employs a simplified physical model to rapidly assess the feasibility of various solutions. Based on pre-calculated hydraulic characteristic curves of the pipe network, this module approximates the flow distribution under different start-stop combinations using interpolation methods. Simultaneously, it estimates the temperature response speed using the building's thermal inertia time constant, ensuring that control commands meet thermal demands without causing system oscillations. This approximate calculation method is two orders of magnitude faster than full physical simulation, enabling real-time optimization.

[0095] The parallel computing architecture fully utilizes modern multi-core processor resources. The optimization problem is decomposed into multiple subtasks, including start-stop combination generation, hydraulic balance calculation, and economic evaluation. These subtasks are distributed across different computing cores for parallel execution. Key data is exchanged at high speed via a shared memory mechanism, and the computation progress is uniformly coordinated by a central scheduler. This architecture design enables complex mixed-integer optimization problems to be solved within 5 minutes, fully meeting the time requirements of real-time control.

[0096] The optimization results are output using a standardized control command format. Each command includes the target device number, operation type, set parameter value, execution time window, and confidence index. The system automatically checks the logical consistency between commands to avoid contradictory operations. It also retains a certain manual intervention interface; when there is a significant discrepancy between the automatic optimization results and the operator's experience-based judgment, a manual review process can be initiated. This human-machine collaborative mechanism leverages the advantages of the optimization algorithm while ensuring the safe and reliable operation of the system.

[0097] In S4, the water supply temperature, flow rate, and equipment start / stop status are dynamically adjusted according to the heating control command to achieve closed-loop optimized operation of the building heating system, specifically including:

[0098] First, a multimodal actuator collaborative control network is constructed, consisting of intelligent controllers distributed at key nodes of the heating system. Each controller is equipped with dual communication interfaces, enabling it to receive central optimization commands via industrial Ethernet and interact directly with local actuators via fieldbus. Water supply temperature control employs a feedforward-feedback composite algorithm. The feedforward component adjusts the boiler outlet water temperature setpoint in advance based on predicted heat load values, while the feedback component corrects control deviations in real time using high-precision temperature transmitters installed on the pipeline. Flow regulation utilizes an improved fuzzy adaptive PID algorithm. This algorithm automatically identifies changes in pipeline characteristics and dynamically adjusts the proportional, integral, and derivative coefficients, effectively addressing the performance degradation problem of traditional PID control under varying operating conditions.

[0099] The equipment start-up and shutdown sequence optimization module is one of the key innovations of this invention. This module uses an infrared thermal imager installed on the boiler body to monitor the metal wall temperature distribution in real time and establishes a thermal state evolution model of the equipment based on historical operating data. When a start-up or shutdown command is received, the module dynamically calculates the optimal start-up and shutdown time interval: for cold start-up conditions, it automatically extends the preheating time to ensure uniform expansion of the heated surface; for hot shutdown conditions, it controls the cooling rate within a safe range. Simultaneously, the system records key parameter curves during each start-up and shutdown process, forming a dedicated start-up and shutdown characteristic database for the equipment, providing a reference for subsequent optimization.

[0100] The control effectiveness evaluation mechanism employs a distributed temperature monitoring network. High-precision digital temperature sensors with self-calibration capabilities are deployed in various functional areas of the building to collect actual room temperature data at 1-minute intervals. The evaluation algorithm not only calculates the overall temperature deviation but also analyzes the temperature non-uniformity coefficient and trend indicators of each area. For critical areas such as hospital wards, an additional temperature stability evaluation function is implemented to ensure that the specific requirements for medical environment control are met. All evaluation results generate visual reports to help operators fully understand the system's control quality.

[0101] The dynamic parameter adjustment strategy employs an incremental online learning approach. When a control deviation is detected to continuously exceed a set threshold, the system automatically initiates a model parameter update process. This process first analyzes the cause of the deviation: for deviations in the prediction model, the output layer weights of the neural network are adjusted using the sliding window least squares method; for deviations in the optimization model, the weight coefficients in the objective function are corrected. The learning process uses a conservative update strategy, with each adjustment not exceeding 5% of the original value to ensure a smooth transition of system parameters. Multiple safety verification mechanisms are also implemented to prevent parameter drift caused by abnormal data.

[0102] At the actuator control level, dedicated drive algorithms were designed for different types of regulating mechanisms. The electric regulating valve adopts fuzzy control with position feedback to solve the nonlinearity and hysteresis problems of traditional valve position control; the variable frequency water pump drive introduces mechanical vibration monitoring feedback to ensure the mechanical safety of the equipment while pursuing flow regulation accuracy; the boiler burner control integrates an air-fuel ratio self-optimization function to adjust the gas and air ratio in real time, maximizing combustion efficiency while meeting load requirements.

[0103] The system status monitoring interface integrates multi-dimensional operational indicator dashboards. In addition to standard parameters such as temperature, pressure, and flow rate, it also displays real-time comprehensive evaluation parameters such as equipment health status scores, energy efficiency performance indicators, and environmental comfort indices. All monitoring data is timestamped and quality-marked, supporting historical data retrospective analysis by any time range and region. When abnormal operating conditions are detected, the system automatically pops up a diagnostic suggestion window, providing possible cause analysis and handling countermeasures.

[0104] The anomaly handling mechanism employs a tiered response strategy. For general parameter exceedances, the system automatically initiates preset corrective measures; for severe faults, a protective shutdown procedure is triggered, simultaneously notifying relevant personnel via audible and visual alarms and mobile push notifications. All anomalies generate detailed log records, including the occurrence time, duration, handling measures, and final result, providing data support for subsequent fault analysis and preventative maintenance.

[0105] The system maintenance functions are designed with comprehensive self-diagnostic and remote maintenance interfaces. It periodically and automatically performs sensor calibration checks, communication link tests, and control loop response tests to ensure each subsystem is in optimal working condition. It also supports a remote expert diagnostic mode, allowing equipment manufacturer technicians to access the system via a secure channel when necessary for in-depth troubleshooting and performance optimization, significantly improving system maintainability and service response speed.

[0106] Finally, the system provides a flexible human-machine interface. Operators can access the system through workstations, mobile terminals, and other means, choosing between fully automatic operation mode and manual adjustment of key parameters as needed. All operation commands undergo multiple confirmations and rationality checks to prevent system malfunctions caused by human error. This human-machine collaborative control architecture leverages the advantages of automatic control while retaining necessary human intervention permissions, achieving an optimal balance between security and flexibility.

[0107] Please see Figure 2 As shown, the intelligent building heating system based on deep learning includes:

[0108] A multi-scale data acquisition and feature extraction module is used to acquire building environment parameters and equipment operation data in real time and extract dynamic features at multiple time scales.

[0109] The physical constraint deep prediction module is used to construct a deep time series prediction model that integrates physical constraints. It couples the minute-level fluid transport parameters with the hour-level building thermal inertia parameters through neural differential equations to output the predicted heat load demand value for future periods.

[0110] The mixed integer optimization decision module establishes a mixed integer optimization model based on the predicted heat load demand, takes equipment operating cost and life loss as the joint objective function, and simultaneously optimizes continuous adjustment variables and discrete start-stop decision variables to generate heating control commands.

[0111] The closed-loop execution and adaptive adjustment module dynamically adjusts the water supply temperature, flow rate, and equipment start / stop status according to the heating control command, thereby realizing the closed-loop optimized operation of the building heating system.

[0112] The working principle of this invention: This invention achieves efficient and precise control of the heating system through multi-source data fusion, physical constraint modeling, and mixed-integer optimization. The technical solution first uses a distributed fiber optic temperature measurement network, a hydraulic sensor array, and an infrared thermal imaging system to collect real-time data on the building environment and equipment operation. Minute-level fluid distribution parameters and hourly-level building thermal inertia characteristics are extracted using wavelet packet transform and graph neural networks. Then, a dual-channel neural differential equation prediction model is constructed, embedding a differentiable heat conduction operator into the neural network. Multi-task learning is used to synchronously optimize prediction accuracy and physical consistency, outputting multi-timescale heat load predictions. Next, a mixed-integer optimization model considering equipment lifespan loss is established, employing a hierarchical decision-making mechanism to synchronously optimize boiler start-up and shutdown combinations and pipeline flow distribution. Finally, closed-loop control is achieved through a multi-modal actuator network, combining a feedforward-feedback composite algorithm and fuzzy adaptive PID to adjust water supply parameters, and dynamically adjusting model parameters based on real-time monitoring data. This method innovatively combines deep learning with physical modeling, significantly improving energy efficiency and extending equipment lifespan while ensuring heating quality through equipment health state constraints and an adaptive optimization window mechanism, achieving fully automatic intelligent optimized operation of the building heating system.

[0113] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for intelligent optimization of building heating operation based on deep learning, characterized in that, Includes the following steps: S1. Real-time acquisition of building environmental parameters and equipment operation data, extraction of dynamic features at multiple time scales, wherein the dynamic features at multiple time scales include minute-level fluid transport parameters and hour-level building thermal inertia parameters; S2. Construct a deep time-series prediction model that integrates physical constraints, and couple the minute-level fluid transport parameters with the hour-level building thermal inertia parameters through neural differential equations to output the predicted heat load demand for future periods; S3. Based on the predicted heat load demand, establish a mixed integer optimization model, take equipment operating cost and life loss as the joint objective function, simultaneously optimize continuous adjustment variables and discrete start-stop decision variables, and generate heating control commands; S4. Dynamically adjust the water supply temperature, flow rate, and equipment start / stop status according to the heating control command to achieve closed-loop optimized operation of the building heating system.

2. The intelligent building heating intelligent optimization operation method based on deep learning according to claim 1, characterized in that, The extraction process of the multi-timescale dynamic features specifically includes: The temperature data of pipes in each zone of the building are collected in real time through a distributed optical fiber temperature measurement network. The original temperature signal is decomposed into feature components at different time scales using a wavelet packet transform algorithm. The first feature component corresponds to the minute-level fluid dynamic fluctuations, and the second feature component corresponds to the hour-level building thermal inertia changes. Based on the pipeline pressure data obtained by the hydraulic sensor array, a fluid distribution map neural network model is constructed. Minute-level hydraulic coupling features are extracted by the node embedding method. The hydraulic coupling features include branch flow distribution coefficient and pipeline pressure drop gradient. Infrared thermal imagers are used to monitor the surface temperature field of building envelopes. Combined with thermal parameters in building information models, hourly building thermal inertia parameters are obtained through three-dimensional finite element heat conduction inversion calculations. The thermal inertia parameters include the equivalent heat capacity of the wall and the window-to-wall ratio correction factor. After spatiotemporal alignment of the first feature component, the second feature component, the hydraulic coupling feature, and the thermal inertia parameter, the input is fed into the feature fusion module, which outputs minute-level fluid transport parameters and hour-level building thermal inertia parameters with clear physical meaning.

3. The intelligent building heating intelligent optimization operation method based on deep learning according to claim 1, characterized in that, The specific methods for constructing the deep temporal prediction model that incorporates physical constraints include: A dual-channel neural differential equation framework is constructed. The first channel processes minute-level fluid transport parameters and uses a convolutional long short-term memory network to extract local hydraulic dynamic features. The second channel processes hour-level building thermal inertia parameters and uses an attention-based gated recurrent unit network to capture long-term heat conduction patterns. Building thermodynamic constraints are embedded in the hidden layer of the neural differential equation. The features of the two channels are coupled at the physical level through a differentiable heat conduction operator to ensure that the prediction results comply with the law of conservation of energy. A multi-task learning mechanism is adopted to simultaneously optimize the prediction accuracy and the physical consistency loss function, wherein the physical consistency loss function is calculated by comparing the deviation between the predicted heat flux density and the actual sensor measurement value; The coupled neural differential equations are solved by an adaptive time-step numerical solver, and the predicted heat load demand sequence for multiple future time scales is output.

4. The intelligent building heating intelligent optimization operation method based on deep learning according to claim 3, characterized in that, The method for embedding physical constraints in the neural differential equation framework specifically includes: An equivalent thermal network model of the building envelope is established, and the wall thermal resistance and window heat transfer coefficient are incorporated as learnable physical parameters into the neural network weight initialization process. In each time step of the neural differential equation calculation, thermal equilibrium constraints are enforced, and a differentiable heat transfer operator is used to ensure that the predicted rate of change of heat load does not exceed the maximum thermal conductivity of the building envelope. The physical information regularization method is adopted to add a residual term of the partial differential equation based on Fourier's heat conduction law to the loss function to constrain the learning direction of the neural network. By generating virtual samples that conform to physical laws through adversarial training mechanisms, the generalization ability of the model under extreme conditions is enhanced.

5. The intelligent building heating intelligent optimization operation method based on deep learning according to claim 3, characterized in that, The method for generating the multi-timescale heat load demand forecast sequence specifically includes: A hierarchical time attention mechanism is designed to focus on recent hydraulic fluctuation characteristics in the minute-level prediction subnetwork and on long-term meteorological change trends in the hour-level prediction subnetwork. Temporal convolutional networks are used to extract temporal features at different scales, and dilated convolutions are used to expand the receptive field to capture periodic heat usage patterns. A post-processing correction module for the prediction results is established to dynamically adjust the output value based on the statistical analysis of historical prediction errors, thereby eliminating systematic biases. An uncertainty quantification method is introduced to provide confidence intervals for the prediction results, thus providing risk warning information for subsequent optimization control.

6. The intelligent building heating intelligent optimization operation method based on deep learning according to claim 1, characterized in that, The specific methods for establishing the mixed integer optimization model include: A multi-objective optimization framework is constructed, which quantifies equipment operating costs as the product of energy consumption costs and time-of-use electricity prices, and quantifies equipment lifespan loss as a weighted function of start-up and shutdown times and operating load. The design employs a hierarchical decision-making mechanism. The upper-level optimization uses a branch pricing algorithm to solve the boiler group start-up and shutdown combination, while the lower-level optimization uses the interior point method to solve the flow distribution scheme under the hydraulic balance condition of the pipeline network. The equipment health status constraints are embedded in the optimization model, and the minimum operating time limit of the boiler is dynamically adjusted by real-time collected vibration spectrum data. Establish a rolling optimization window mechanism to adaptively adjust the optimization time domain length based on the prediction confidence interval, ensuring that a more conservative control strategy is adopted when the prediction uncertainty increases.

7. The intelligent building heating intelligent optimization operation method based on deep learning according to claim 6, characterized in that, The method for quantifying equipment lifespan loss specifically includes: A damage accumulation model for key boiler components is established based on the theory of material fatigue, with the rate of temperature change and the amplitude of pressure fluctuation as damage factors. The expansion of internal defects in the equipment is monitored in real time using acoustic emission sensors, and the life loss weighting coefficient is dynamically adjusted. Establish a sub-model for predicting the remaining service life of equipment, and input historical operating conditions and current state parameters into a convolutional neural network to predict the remaining service life; By introducing the equipment residual value term into the optimization objective function, joint optimization of operating costs and asset depreciation can be achieved.

8. The intelligent building heating intelligent optimization operation method based on deep learning according to claim 6, characterized in that, The specific implementation methods of the hierarchical decision-making mechanism include: Construct a device combination state tree and use a depth-first search strategy to traverse feasible start-stop combination schemes; Design heuristic pruning rules to eliminate non-optimal branches in advance based on equipment efficiency curves and pipeline topology; Establish a preprocessing module for hydraulic-thermal coupling to quickly assess the system feasibility under various start-stop combinations; A parallel computing architecture is adopted to accelerate the optimization process and ensure that all computing tasks are completed within the control cycle.

9. The intelligent building heating intelligent optimization operation method based on deep learning according to claim 1, characterized in that, The specific methods for closed-loop optimized operation include: A multimodal actuator collaborative control network is constructed, and optimization instructions are received through a distributed programmable logic controller cluster. The water supply temperature regulation adopts a feedforward-feedback composite control algorithm, and the flow regulation adopts a fuzzy adaptive PID control algorithm. Design an equipment start-stop sequence optimization module to dynamically calculate the optimal start-stop time interval based on equipment thermal status monitoring data, so as to avoid equipment damage caused by thermal shock; Establish a real-time evaluation mechanism for control effectiveness, collect the actual room temperature distribution of each zone through a high-precision temperature sensor array, and calculate the deviation of control command execution. A dynamic parameter adjustment strategy is implemented, which automatically triggers the online learning mechanism for model parameters when the deviation exceeds the threshold, updating the internal parameters of the prediction model and the optimization model.

10. A deep learning-based intelligent building heating intelligent optimization operation system, characterized in that, The method for intelligent optimization operation of building heating based on deep learning as described in any one of claims 1-9 includes: A multi-scale data acquisition and feature extraction module is used to acquire building environment parameters and equipment operation data in real time and extract dynamic features at multiple time scales. The physical constraint deep prediction module is used to construct a deep time series prediction model that integrates physical constraints. It couples the minute-level fluid transport parameters with the hour-level building thermal inertia parameters through neural differential equations to output the predicted heat load demand value for future periods. The mixed integer optimization decision module establishes a mixed integer optimization model based on the predicted heat load demand, takes equipment operating cost and life loss as the joint objective function, and simultaneously optimizes continuous adjustment variables and discrete start-stop decision variables to generate heating control commands. The closed-loop execution and adaptive adjustment module dynamically adjusts the water supply temperature, flow rate, and equipment start / stop status according to the heating control command, thereby realizing the closed-loop optimized operation of the building heating system.

Images