Building system adaptive control method based on data truncation and modular transfer learning
By modularly modeling and trunculating the building system, and combining it with transfer learning of hybrid recurrent neural networks, the problem of model adaptability of building systems under sudden changes is solved, and rapid and accurate energy consumption prediction and control performance improvement are achieved.
Patent Information
- Application Number
- CN202511600404.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-04
- Publication Date
- 2026-01-23
AI Technical Summary
Existing machine learning-based predictive control models struggle to adapt quickly to sudden system changes in building systems, leading to model distortion and decreased control performance. In particular, they lack the ability to distinguish between outdated and new data domains, making it impossible to achieve rapid and reliable model relearning with limited new samples.
A method based on data truncation and modular transfer learning is adopted to modularly model each component of the building system. Only the parameters of the modules affected by the mutation are unfrozen and fine-tuned, and the data before the mutation is discarded. The model transfer learning is carried out using a hybrid recurrent neural network. The influence of the old domain is eliminated by data truncation, and only the prediction error output by the module affected by the mutation is used as the loss function during the transfer process.
It significantly improved the model's adaptability after sudden system changes, reducing the prediction error by 32.4% after 2 hours and 25.4% after 22 hours, thus enhancing the system's robustness and energy consumption prediction accuracy.
Smart Images
Figure CN121386410A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of intelligent building energy consumption optimization and control technology, and in particular to an adaptive control method for building systems based on data truncation and modular transfer learning. Background Technology
[0002] Model predictive control (MPC) based on machine learning (ML) has shown great potential in the field of smart buildings. Numerous studies have demonstrated the significant effectiveness of MPC in smart building control, and its online model adaptation technology is often integrated to improve long-term performance. However, existing adaptive algorithms are mostly based on the assumption of continuous stationarity and can only cope with slow drifts in system parameters (such as building aging and seasonal climate changes). For example, to achieve greater energy savings, MPC systems are applied to an office or an auditorium. Compared to traditional rule-based control or PID control, the machine learning (ML)-based MPC framework integrates recurrent neural networks (RNNs), exhibiting good performance in temperature regulation and energy saving. However, buildings are highly dynamic systems with inherent uncertainties. Fixed machine learning models tend to accumulate prediction errors over time, leading to model distortion and unreliable MPC control performance. Sudden system changes, such as resident interventions, are difficult to capture by traditional online ML model adaptation mechanisms.
[0003] Recent research has focused on adaptive machine learning methods, which continuously learn from real-time operational data and update predictive models to maintain long-term reliable accuracy. For example, the model predictive control framework based on adaptive machine learning proposed in the existing paper "Yang Si, Wan Ming, Chen Wei, Wu Ning, Zhai Da. Adaptive Robust Model Predictive Control for Indoor Climate Optimization and Uncertainty Handling in Buildings. Architecture and Environment 2019" provides a solution for achieving robust indoor climate regulation. Such adaptive strategies can effectively cope with slowly changing temporal characteristics (such as building aging or gradual climate change) by utilizing accumulated operational data for optimization. However, sudden systemic changes (such as human intervention or major structural modifications) may trigger instantaneous and significant domain shifts, resulting in a domain shift between the historical data domain and the new operating domain. Historical data reflects outdated domains and misleads the learning process, leading to severe model distortion, accumulated prediction errors, and ultimately weakening the control effect of MPC.
[0004] While existing online adaptive methods can update models using real-time data, they lack the ability to distinguish between "outdated data domains" and "new data domains," making it difficult to achieve fast and reliable model relearning with limited new samples and struggling to cope with sudden systemic changes. Therefore, a new method is needed that can both discard invalid data and efficiently reuse existing knowledge. Summary of the Invention
[0005] To overcome the shortcomings of the prior art, this application provides an adaptive control method for building systems based on data truncation and modular transfer learning. This method addresses the problem that existing ML-MPC models cannot quickly adapt to new thermodynamic states when building systems undergo sudden changes, leading to inaccurate energy consumption predictions and decreased control performance. This application aims to solve the problems of low domain transfer efficiency and interference of old domain data with new domain learning.
[0006] To achieve the above objectives, this application adopts the following technical solution, including: An adaptive control method for building systems based on data truncation and modular transfer learning is shown below: Modular modeling of each component of the building system is performed, and a hybrid recurrent neural network is used as the architecture of the modular model of the building system. The hybrid recurrent neural network includes multiple recurrent neural network units, each of which corresponds to a component module in the building system. After a mutation is detected in the building system, only the module parameters affected by the mutation are unfrozen for fine-tuning, while the module parameters unaffected by the mutation remain frozen. At the same time, the data before the mutation is discarded to eliminate the influence of the old domain, and only the data after the mutation is retained for fine-tuning the unfrozen module parameters.
[0007] Preferably, during model transfer learning, only the prediction error output by the module affected by the mutation is used as the loss function.
[0008] Preferably, the building heating system includes four components: a ground source heat pump, a phase change material water tank, a radiator, and a building structure, which are respectively constructed as a ground source heat pump module, a phase change material water tank module, a radiator module, and a building module; The input to the ground source heat pump module is the ground source heat pump heat flow rate. q gshp Water temperature inside the tank T wtr The output is the electricity consumption of the ground source heat pump. P gshp ; The input to the phase change material water tank module is the heat flow rate of the ground source heat pump. q gshp Auxiliary electric heater heat flow rate q eh Ambient temperature T amb Water temperature inside the tank T wtr The output is the liquid fraction of the phase change material. f liq Water temperature inside the tank T wtr ; The input to the heat sink module is the heat flow rate of the heat sink. qrdtr Hot zone temperature T z Water temperature inside the tank T wtr The output is the actual heat flux. q act ; The input to the building module is the ambient temperature. T amb Solar radiation R sloar Internal heat absorption and thermal gain v IG Hot zone temperature T z The output is the temperature of the hot zone. T z ; For the ground source heat pump module, phase change material water tank module and building module, the cooperative function of the hidden tensor is the same, and only the final hidden state is passed to form the final output of the hybrid recurrent neural network; The heat sink module employs a unique hidden tensor cooperative function, which generates a time series of the actual heat flux as the output by concatenating all hidden states. The actual heat flux sequence is synchronized with the input sequence of the hybrid recurrent neural network and concatenated with other physical quantity sequences as the input of other modules, thereby ensuring the accuracy of time coupling.
[0009] Preferably, the building resistance-capacitance model (BRCM) is used to simulate and analyze the thermal dynamics of a building. The model simplifies the building structure into a combination of resistance and capacitance to describe the heat transfer and storage process in the building.
[0010] Preferably, in the study of abrupt changes in building systems, changes in the state of the window are taken into consideration. The window is quantified by two parameters, namely the G value and the U value, which correspond to the solar radiation heat flux and the heat entering the heat zone through conduction / convection through the window, respectively.
[0011] This application also provides a readable storage medium having a computer program stored thereon, which, when executed, implements the aforementioned adaptive control method for building systems based on data truncation and modular transfer learning.
[0012] This application also provides an electronic device including a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned adaptive control method for building systems based on data truncation and modular transfer learning.
[0013] This application also provides a computer program product comprising a computer program / instruction that, when executed by a processor, implements the aforementioned adaptive control method for building systems based on data truncation and modular transfer learning.
[0014] The advantages of this application are: (1) This application proposes an innovative Adaptive Transfer Learning-Assisted Modular Learning (ATML) framework, combined with data truncation technology. Data truncation, i.e., removing data collected before the mutation of the building system, can effectively eliminate the negative impact of outdated domains. When the available data is limited due to data truncation, ATML can further improve the model's rapid adaptability.
[0015] (2) The main contributions of this application are as follows: (1) The impact of domain shift caused by systematic mutation was studied and the effectiveness of data truncation was verified. (2) To achieve fast and effective model adaptation using limited data after data truncation, the ATML framework was proposed and evaluated.
[0016] (3) The impact of changes in the window opening status of residential buildings is studied: Compared with the traditional online ML model adaptive method, ATML can reduce the prediction error by 32.4% after 2 hours and by 25.4% after 22 hours, which fully demonstrates the effectiveness of the framework in dealing with sudden system changes.
[0017] (4) This application can be applied to the following scenarios: Intelligent building energy management: used to improve the energy prediction and optimization control performance of residential buildings, office buildings, and commercial complexes; Heat pump and thermal storage system control: Improve the modeling accuracy and energy efficiency control of ground source heat pumps (GSHP) and phase change energy storage systems (PCM Tank); Smart Home and Building Automation: Combining sensor networks to achieve intelligent control strategies for multi-device collaboration; Industrial and regional energy system optimization: dynamic modeling and predictive control for multi-energy systems (heat, electricity, cooling).
[0018] (5) The method of this application can be embedded into a building management system (BMS) or energy management platform, and automatically adapt to system mutations (such as window opening, equipment replacement, climate anomalies, etc.) through data-driven models, which significantly improves system robustness and energy consumption prediction accuracy.
[0019] (6) Compared with traditional adaptive learning methods, the beneficial effects of the improved technical solution in this application are as follows: Rapid adaptability: After 2 hours of system change, the prediction error decreased by 32.4%; after 22 hours, it still decreased by 25.4%. Negative migration suppression: prevents performance degradation of unchanged modules and ensures the stability of the overall model; Physical consistency: Maintaining the thermal coupling between modules improves physical interpretability and generalization ability; Scalability: Applicable to multiple scenarios such as multi-energy systems, regional energy management, and smart home control. Attached Figure Description
[0020] Figure 1 This is a schematic diagram of a hybrid recurrent neural network structure.
[0021] Figure 2 This is a schematic diagram of modular transfer learning in a hybrid recurrent neural network.
[0022] Figure 3 This is a schematic diagram of a building's heating system.
[0023] Figure 4 This is a diagram illustrating how a form changes from closed to open.
[0024] Figure 5 A comparison diagram of the truncated and non-truncated adaptive processes of HRNN and RNN.
[0025] Figure 6 This is a comparison chart of ATML for HRNN and retraining after truncation.
[0026] Figure 7 A comparison chart of multi-output sensitivity analysis of target dataset availability for ATML and RNN. Detailed Implementation
[0027] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0028] Example 1
[0029] This application provides an adaptive control method for building systems based on data truncation and modular transfer learning, mainly applied to model predictive control (MPC) of building systems such as heating, ventilation and air conditioning (HVAC) systems, as detailed below: Modular modeling of each component of the building system is performed, and a hybrid recurrent neural network (HRNN) is used as the architecture of the modular model of the building system. The hybrid recurrent neural network includes multiple recurrent neural network units (RNN), and each recurrent neural network unit corresponds to each component module in the building system. Upon detecting a mutation in the building system, only the parameters of the modules affected by the mutation are unfrozen for fine-tuning, while the parameters of modules unaffected by the mutation remain frozen (i.e., a modular transfer learning strategy is employed). Simultaneously, pre-mutation data is discarded to eliminate the influence of the old domain, and only post-mutation data is retained for fine-tuning the unfrozen module parameters (i.e., a data truncation strategy is employed). Furthermore, to accelerate the convergence of corrections for modules affected by the mutation, during model transfer learning, only the prediction errors output by the affected modules are used as the loss function.
[0030] In this application, the Adaptive Transfer Learning Assisted Modular Learning (ATML) framework achieves local fine-tuning by constructing a modular model and employing modular transfer learning. The Hybrid Recurrent Neural Network (HRNN) developed in this application serves as the core architecture of the modular model. The core design concept of this model is to establish multiple recurrent neural network units (RNNs), each corresponding to a specific component in the system.
[0031] In this embodiment, the building heating system comprises four core components: a ground source heat pump (GSHP), a phase change material (PCM) water tank, radiators, and a building structure. Correspondingly, a ground source heat pump (GSHP) module, a PCM water tank module, a radiator module, and a building module are constructed. A schematic diagram of the corresponding hybrid recurrent neural network (HRNN) structure is shown below. Figure 1 As shown.
[0032] The input to the ground source heat pump module is the ground source heat pump heat flow rate. q gshp Water temperature inside the tank T wtr The output is the electricity consumption of the ground source heat pump. P gshp ; The input to the phase change material water tank module is the heat flow rate of the ground source heat pump. q gshp Auxiliary electric heater heat flow rate q eh Ambient temperature T amb Water temperature inside the tank T wtr The output is the liquid fraction of the phase change material. f liq Water temperature inside the tank T wtr ; The input to the heat sink module is the heat flow rate of the heat sink. q rdtr Hot zone temperature T z Water temperature inside the tank T wtr The output is the actual heat flux.q act ; The input to the building module is the ambient temperature. T amb Solar radiation R sloar Internal heat absorption and thermal gain v IG Hot zone temperature T z The output is the temperature of the hot zone. T z .
[0033] The meanings of the relevant characters in the building heating system of this embodiment are shown in Table 1 below: Table 1
[0034] The overall HRNN model's topology is derived from physical principles, but each module is implemented through a dedicated RNN sub-model, resulting in differences in the cooperative functions of their hidden tensors. For the GSHP, PCM water tank, and building modules, the cooperative functions of the hidden tensors are identical; only the final hidden state is passed to form the final output of the HRNN. The radiator module employs a unique hidden tensor cooperative function, generating a time series of the actual heat flux as its output by concatenating all hidden states. This actual heat flux sequence is synchronized with the HRNN's input sequence and concatenated with other physical quantity sequences before being used as input to other modules, thus ensuring accurate temporal coupling.
[0035] In modular transfer learning, the impact of task-specific loss function weights on model learning efficiency is taken into consideration. Only output bias caused by component changes is included in the loss function. To prioritize learning changed components and prevent negative transfer from unchanged components, the effective knowledge domain captured by the previous model is preserved during the transfer process, the sub-model parameters related to unchanged components are frozen, and only the parameters related to changed components are fine-tuned. Figure 2 This demonstrates the fine-tuning of a hybrid recurrent neural network (HRNN) based on modular transfer learning after a mutation occurs in a building structural component.
[0036] In this embodiment, the building heating system includes a standardized prototype single-family home heated by radiators. These radiators are connected via a hydraulic system to an outdoor phase change material (PCM) water tank, which is jointly filled by a ground source heat pump (GSHP) and an auxiliary electric heater (EH). Both the ground source heat pump and the auxiliary electric heater are powered directly from the power grid. The specific structure of the building heating system is as follows: Figure 3As shown. To establish the thermodynamic model of this residence, the Building Resistance-Capacitance Modeling (BRCM) toolkit was used. BRCM (Building Resistance-Capacitance Model) is used to simulate and analyze the thermal dynamics of buildings. The model describes the transfer and storage of heat in building materials by simplifying the building structure into a combination of resistance and capacitance.
[0037] In systemic catastrophe studies, changes in window state are taken into consideration. As the most common resident intervention in residential buildings, changes in window state (from closed to open) are... Figure 4 As shown, the BRCM (Building Resistance-Capacitance Model) quantifies windows using two parameters (G-value and U-value). The G-value corresponds to the solar radiation heat flux and the heat entering the heat zone through conduction / convection via the window, respectively. Specifically, the G-value is the ratio of solar radiation heat entering the room through the window to the total incident solar radiation; it includes two parts: directly transmitted solar energy and heat absorbed by the glass and then transferred to the room. The U-value is the heat transfer rate per unit area of the window per unit time, reflecting the window's (including glass, frame, and spacer) heat insulation capability; the lower the U-value, the better the insulation performance.
[0038] When the window is partially open, the open area no longer supports conductive heat transfer because conduction requires a solid or stationary medium as a carrier. At this time, heat exchange is mainly dominated by convection and atmospheric flow through the opening. In the absence of external wind, this heat exchange is primarily driven by temperature difference, i.e., the buoyancy-driven ventilation column effect. The corresponding heat transfer rate can be expressed as: ; ; in, For ventilation heat flux, air density, For air heat capacity, The ambient temperature, For the temperature of the hot zone, For emission coefficients, Let be the area of the opening. g It is the acceleration due to gravity. h The opening height, V For volumetric flow rate, This refers to the temperature difference.
[0039] Due to the lack of a solid medium, the conduction effect at the window opening is negligible, and the corresponding U value can be approximated as zero. Furthermore, since there is no refraction or reflection caused by glass in the opening region, the G value can be considered equal to 1. Given that the contributions of G and U values to heat flux are linearly related, the parametric expression for the entire window can be expressed as: ; ; in, The G value for the entire form. The U value for the entire form. The form opening ratio. The G value for the closed area of the form. The U value for the closed area of the form.
[0040] Considering the upper limit of the radiator's heating power, the window opening ratio is 5%.
[0041] To accurately reflect changes in electricity prices and weather conditions, this embodiment uses typical meteorological data from a certain city in a given year, as well as real-time wholesale electricity price data from 2019.
[0042] For the Hybrid Recurrent Neural Network (HRNN) model, parameter settings are detailed in Table 2. The sequence length was set to 5 time steps, each corresponding to 5 minutes. Two datasets were constructed: a resource dataset (December 1st to 13th) representing the system in an unchanged state (window closed), and a target dataset (December 14th to 16th) representing the system in a changed state (window open). During dataset generation, a randomized control strategy was employed to ensure comprehensive coverage of the feature space. These datasets were used for model training and testing, respectively. All training data was divided into two subsets: 80% for parameter optimization and 20% for setting a 16-step early stopping strategy. All simulation experiments were performed on a workstation equipped with an Intel Core i5-12500H processor (2.5GHz).
[0043] Table 2 Configuration of Hybrid Recurrent Neural Network (HRNN)
[0044] Two adaptive training strategies were compared: the first strategy retrained the model from scratch using the complete cumulative data (without data truncation); the second strategy used only data collected after the system mutation for adaptive training (with data truncation). The learning curves of the two methods were analyzed by gradually increasing the available target training dataset (December 14-15). The availability metric refers to the duration of target data available during model training. The evaluation metric was the mean absolute error (MAE) between the 1-hour recursive predictions and actual values on the test dataset, which included all available data from December 16.
[0045] like Figure 5As shown, both HRNN and RNN (both using a single hidden layer with 17 hidden units and the Tanh activation function) exhibit convergence and transcendence characteristics. This indicates that although the model using the truncation method performs poorly initially due to the smaller training set, the untruncation method, by retaining outdated domain information, actually hinders the model's adaptability.
[0046] While truncation is effective in addressing the domain shift problem, its performance remains poor in the initial hours after a system mutation. To evaluate the adaptability of the Adaptive Transfer Learning-Assisted Modular Learning (ATML) framework under conditions of scarce target data, we conducted an in-depth analysis of its learning curve. This method first trains a hybrid recurrent neural network (HRNN) from scratch using the resource dataset, and then performs incremental modular fine-tuning only on the target data. Figure 6 As shown, ATML protects the model's potential by truncating outdated domains while reusing effective subdomains from preceding models, thereby significantly improving the model's adaptive speed.
[0047] To investigate the impact of target dataset availability on the prediction accuracy of multi-output prediction tasks during transfer learning, this embodiment plots learning curves for all output variables, using the normalized mean absolute error (NMAE) between the 1-hour recursive predicted values and actual values in the target test dataset as the evaluation metric. This metric is obtained by performing min-max normalization on the mean absolute error (MAE).
[0048] To evaluate the adaptive performance of ATML, this embodiment also develops another model adaptive framework (single RNN) using a truncation method: first, the RNN is trained from scratch using the resource dataset (its hyperparameters are the same as in Section 3.1), and then fully fine-tuned using the same loss function of modular transfer learning. Figure 7 As shown, as the size of the target training dataset increases, the single RNN exhibits a decline in both transfer performance and adaptive capability, while the ATML model effectively mitigates this negative impact.
[0049] This application addresses the domain migration problem caused by sudden systemic changes by proposing the ATML framework, which incorporates data truncation techniques. This strategy removes outdated domains while accelerating the adaptive capabilities of machine learning models. Comparative analysis shows that, compared to traditional adaptive schemes, ATML enables control-oriented models to reduce prediction errors by 32.4% within just 2 hours after sudden systemic changes, and further reduces them by 25.4% after 22 hours.
[0050] Example 2
[0051] An electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method of Embodiment 1 described above.
[0052] The electronic device in this application embodiment may be the mobile device itself, or a standalone device independent of it. The standalone device may communicate with the mobile device to receive the collected input signals from it and send the selected target decision behavior to it.
[0053] An electronic device includes one or more processors and memory. The processor may be a central processing unit (CPU) or other processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device to perform desired functions. The memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor may execute the program instructions to implement the decision-making behavior and decision-making methods of the various embodiments of this application described above, and / or other desired functions.
[0054] Electronic devices may also include input devices and output devices.
[0055] Example 3
[0056] In addition to the methods and devices described above, embodiments of this application may also be computer program products, which include computer program instructions that, when executed by a processor, cause the processor to perform the steps in the decision-making behavior method according to various embodiments of this application as described in Embodiment 1 above.
[0057] The computer program product can be written in any combination of one or more programming languages to perform the operations of the embodiments of this application. The programming languages include object-oriented programming languages such as Java and C++, as well as conventional procedural programming languages such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.
[0058] Example 4
[0059] Embodiments of this application may also be computer-readable storage media storing computer program instructions thereon, which, when executed by a processor, cause the processor to perform the steps in the decision-making behavior decision-making method according to various embodiments of this application described in Embodiment 1 above.
[0060] The computer-readable storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may, for example, include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.
[0061] The above are merely preferred embodiments of this application and are not intended to limit the scope of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application shall be included within the protection scope of this application.
Claims
1. An adaptive control method for building systems based on data truncation and modular transfer learning, characterized in that, As shown below: Modular modeling of each component of the building system is performed, and a hybrid recurrent neural network is used as the architecture of the modular model of the building system. The hybrid recurrent neural network includes multiple recurrent neural network units, each of which corresponds to a component module in the building system. After a mutation is detected in the building system, only the module parameters affected by the mutation are unfrozen for fine-tuning, while the module parameters unaffected by the mutation remain frozen. At the same time, the data before the mutation is discarded to eliminate the influence of the old domain, and only the data after the mutation is retained for fine-tuning the unfrozen module parameters.
2. The adaptive control method for building systems based on data truncation and modular transfer learning according to claim 1, characterized in that, In the model transfer learning process, only the prediction error output by the module affected by the mutation is used as the loss function.
3. The adaptive control method for building systems based on data truncation and modular transfer learning according to claim 1 or 2, characterized in that, The building heating system consists of four components: ground source heat pump, phase change material water tank, radiator and building structure, which are respectively constructed as ground source heat pump module, phase change material water tank module, radiator module and building module; The input to the ground source heat pump module is the ground source heat pump heat flow rate. q gshp Water temperature inside the tank T wtr The output is the electricity consumption of the ground source heat pump. P gshp ; The input to the phase change material water tank module is the heat flow rate of the ground source heat pump. q gshp Auxiliary electric heater heat flow rate q eh Ambient temperature T amb Water temperature inside the tank T wtr The output is the liquid fraction of the phase change material. f liq Water temperature inside the tank T wtr ; The input to the heat sink module is the heat flow rate of the heat sink. q rdtr Hot zone temperature T z Water temperature inside the tank T wtr The output is the actual heat flux. q act ; The input to the building module is the ambient temperature. T amb Solar radiation R sloar Internal heat absorption and thermal gain v IG Hot zone temperature T z The output is the temperature of the hot zone. T z ; For the ground source heat pump module, phase change material water tank module and building module, the cooperative function of the hidden tensor is the same, and only the final hidden state is passed to form the final output of the hybrid recurrent neural network; The heat sink module employs a unique hidden tensor cooperative function, which generates a time series of the actual heat flux as the output by concatenating all hidden states. The actual heat flux sequence is synchronized with the input sequence of the hybrid recurrent neural network and concatenated with other physical quantity sequences as the input of other modules, thereby ensuring the accuracy of time coupling.
4. The adaptive control method for building systems based on data truncation and modular transfer learning according to claim 3, characterized in that, The Building Resistance-Capacitance Model (BRCM) is used to simulate and analyze the thermal dynamics of buildings. The model simplifies the building structure into a combination of resistance and capacitance to describe the heat transfer and storage process in the building.
5. The adaptive control method for building systems based on data truncation and modular transfer learning according to claim 3, characterized in that, In the study of aberrations in building systems, changes in the state of windows are taken into consideration. The windows are quantified using two parameters, namely the G value and the U value, which correspond to the solar radiation heat flux and the heat entering the heat zone through conduction / convection through the windows, respectively.
6. A readable storage medium, characterized in that, It stores a computer program, which, when executed, implements the adaptive control method for building systems based on data truncation and modular transfer learning as described in any one of claims 1 to 5.
7. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the adaptive control method for building systems based on data truncation and modular transfer learning as described in any one of claims 1 to 5.
8. A computer program product, characterized in that, It includes a computer program / instruction that, when executed by a processor, implements the adaptive control method for building systems based on data truncation and modular transfer learning as described in any one of claims 1 to 5.