Operation control system and method for building equipment

Abstract

The invention belongs to the technical field of building equipment control, and provides an operation control system and method of building equipment, and the operation control system of the building equipment comprises the building equipment and control equipment; the building equipment is provided with an equipment detection terminal, and the equipment detection terminal is used for collecting multi-modal data of the building equipment; the control equipment is connected with the building equipment and the equipment detection terminal, and is configured to execute the following steps: receiving multi-modal data of the building equipment; determining time-space correlation characteristics of the building equipment based on the multi-modal data; deducing equipment state parameters of the building equipment according to the time-space correlation characteristics by adopting a target prediction model obtained by pre-training; and determining a corresponding target operation control strategy according to the equipment state parameter of the building equipment so as to control the building equipment to optimize the operation mode based on the target operation control strategy.

Description

Technical Field

[0001] The present application belongs to the field of building equipment control technology, and more specifically, relates to an operation control system and method for building equipment. Background Art

[0002] At present, in the field of building automation technology, traditional building energy management systems mainly rely on manual experience and fixed rules to control building equipment. Faced with the complex and changing building environment, they are unable to accurately match actual needs, resulting in limited energy consumption optimization effects of building equipment. Summary of the Invention

[0003] The purpose of the embodiments of the present application is to provide an operation control system and method for building equipment, aiming to solve the technical problem in related technologies that traditional building energy management systems mainly rely on manual experience and fixed rules to control building equipment.

[0004] To achieve the above-mentioned object, according to a first aspect of the present application, there is provided an operation control system for building equipment, comprising: building equipment and control equipment;

[0005] Building equipment is equipped with equipment detection terminals, which are used to collect multimodal data of building equipment;

[0006] Control devices, connected to building equipment and equipment detection terminals, are configured to perform:

[0007] Receive multimodal data of building equipment; determine the spatiotemporal correlation characteristics of the building equipment based on the multimodal data; use a pre-trained target prediction model to derive equipment status parameters of the building equipment based on the spatiotemporal correlation characteristics; and determine a corresponding target operation control strategy based on the equipment status parameters of the building equipment to control the building equipment to optimize the operation mode based on the target operation control strategy.

[0008] In one possible implementation, the system further includes:

[0009] The multi-protocol gateway connects the device detection terminal and the control device, is used to receive the multimodal data transmitted by the device detection terminal, and forward the multimodal data to the control device.

[0010] In one possible implementation, the control device is further configured to execute:

[0011] Based on the equipment status parameters of the building equipment, multiple operation control strategies are automatically generated; through the physical field inversion technology, transient simulation processing is performed on the multiple operation control strategies to obtain simulation results corresponding to the multiple operation control strategies; based on the simulation results corresponding to the multiple operation control strategies, a target operation control strategy is selected from the multiple operation control strategies; and the target operation control strategy is sent to the building equipment.

[0012] In one possible implementation, the control device deploys a graph attention network based on the spatial attention mechanism, a layer of convolution based on causality, and a spatiotemporal encoder;

[0013] A graph attention network based on the spatial attention mechanism is used to determine the spatial characteristics of building equipment based on multimodal data. The spatial characteristics are used to characterize the spatial topological relationship between the various types of equipment contained in the building equipment. A layer of convolution based on causality is used to determine the temporal dynamic characteristics of building equipment based on multimodal data. The temporal dynamic characteristics are used to characterize the energy consumption fluctuations and dynamic changes of equipment status over time of building equipment. A spatiotemporal encoder is used to determine the spatiotemporal correlation characteristics of building equipment based on spatial characteristics and temporal dynamic characteristics.

[0014] In one possible implementation, the multimodal data is collected in real time or at a predetermined frequency;

[0015] The spatiotemporal encoder is also used to compress the spatiotemporal correlation characteristics of building equipment after determining the spatiotemporal correlation characteristics; wherein the compression processing is used to remove redundant information in the spatiotemporal correlation characteristics, and retain at least one of the cross-device collaborative patterns, periodic laws, sudden changes and abnormal event correlation characteristics in the spatiotemporal correlation characteristics.

[0016] In one possible implementation, the control device is further configured to:

[0017] Pre-acquire multiple sample multimodal data of sample building equipment, and sample equipment status parameters corresponding to each of the multiple sample multimodal data, wherein each sample multimodal data includes at least one of temperature, humidity, light, crowd density, current, voltage, and power; generate a training set based on the multiple sample multimodal data and the sample equipment status parameters corresponding to each of the multiple sample multimodal data; add physical knowledge and physical constraints to a general physical model, wherein the physical knowledge includes at least multiple physical laws; under the guidance of the physical constraints, train the model parameters of the general physical model based on the training set, and obtain the target prediction model after multiple iterative optimizations.

[0018] In one possible implementation, the control device is further configured to:

[0019] Confidence assessment is performed on the multiple sample multimodal data by calculating the physical rationality scores of the sample device state parameters corresponding to each of the multiple sample multimodal data;

[0020] According to the confidence evaluation results of each sample multimodal data, the momentum accumulation mechanism is used to update the weight of each sample multimodal data;

[0021] Selecting a preset number of target sample multimodal data from the plurality of sample multimodal data according to the weight of each sample multimodal data;

[0022] A preset number of target sample multimodal data and sample device state parameters corresponding to the target sample multimodal data are used as a training set.

[0023] In one possible implementation, the control device is further configured to:

[0024] After the target prediction model is obtained through multiple iterative optimizations, only the physical inversion module deployed in the control device is optimized in the first fine-tuning stage, and the model parameters of the target prediction model and the physical inversion module are optimized in the second fine-tuning stage.

[0025] In one possible implementation, if there are multiple target operation control strategies, the control device is further configured to:

[0026] After controlling the building equipment to run the control strategy based on the multiple objectives to optimize the operation mode, obtaining multiple operation results obtained by running the building equipment in the multiple optimized operation modes;

[0027] sorting multiple target operation control strategies based on multiple operation results;

[0028] According to the sorting result, a predetermined proportion of target operation control strategies among the multiple target operation control strategies are operated to construct a dynamic memory library;

[0029] After each predetermined time interval, knowledge distillation is performed on the dynamic memory library to obtain the latest dynamic memory library, wherein the target prediction model adjusts the model parameters of the target prediction model by learning the data in the latest dynamic memory library.

[0030] According to a second aspect of the present application, a method for controlling operation of building equipment is provided, the method comprising:

[0031] Receive multimodal data from building equipment;

[0032] Determine the spatiotemporal correlation characteristics of building equipment based on multimodal data;

[0033] Using the pre-trained target prediction model, the equipment status parameters of building equipment are derived based on the spatiotemporal correlation characteristics;

[0034] The corresponding target operation control strategy is determined according to the equipment status parameters of the building equipment to control the building equipment to optimize the operation mode based on the target operation control strategy.

[0035] The second aspect and any implementation of the second aspect correspond to the first aspect and any implementation of the first aspect, respectively. The technical effects corresponding to the second aspect and any implementation of the second aspect can be referred to the technical effects corresponding to the first aspect and any implementation of the first aspect, and will not be repeated here.

[0036] Through the embodiments of the present application, the operation control system of the building equipment includes building equipment and control equipment. The building equipment is deployed with an equipment detection terminal, which is used to collect multimodal data of the building equipment. The control equipment determines the spatiotemporal correlation characteristics of the building equipment based on the multimodal data. The pre-trained target prediction model is used to derive the equipment status parameters of the building equipment according to the spatiotemporal correlation characteristics. The equipment status parameters of the building equipment can be determined without precise modeling and a large amount of historical data participating in the calculation, and are no longer restricted by manual experience and fixed rules. Furthermore, the control equipment can also determine the corresponding target operation control strategy based on the derived equipment status parameters of the building equipment to control the building equipment to dynamically optimize the operation mode based on the target operation control strategy. Without relying on manual experience and fixed rules, the energy-saving operation effect of the building equipment can be adaptively optimized, thereby reducing the energy consumption waste of the building equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments or descriptions of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0038] Figure 1 This is a schematic diagram of the structure of an operation control system for building equipment provided in an embodiment of the present application;

[0039] Figure 2 This is a schematic diagram of the structure of an optional operation control system for building equipment provided in an embodiment of the present application;

[0040] Figure 3 This is a flow chart of a method for controlling the operation of building equipment provided in an embodiment of the present application;

[0041] Figure 4 It is a structural diagram of a control device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0042] In the following description, specific details such as specific system structures and techniques are provided for purposes of illustration rather than limitation to facilitate a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid obscuring the description of the present application with unnecessary detail.

[0043] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or collections thereof.

[0044] It should also be understood that, in the description of this application, unless otherwise specified, "a plurality of" refers to two or more than two. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items.

[0045] In addition, to facilitate a clear description of the technical solutions of the embodiments of the present application, in the embodiments of the present application, words such as "first" and "second" are used to distinguish between identical or similar items with substantially the same functions and effects. Those skilled in the art will understand that words such as "first" and "second" do not limit the quantity or execution order, but are only used to distinguish the description. In addition, words such as "first" and "second" do not necessarily define differences, nor should they be understood to indicate or imply relative importance.

[0046] As used in this specification and the appended claims, the term "if" can be interpreted as "when" or "upon" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "upon determination" or "in response to determining" or "upon detection of [described condition or event]" or "in response to detecting [described condition or event]," depending on the context.

[0047] References to "one embodiment" or "some embodiments" in this specification mean that a particular feature, structure, or characteristic described in conjunction with that embodiment is included in one or more embodiments of the present application. Thus, phrases such as "in one embodiment," "in some embodiments," "in other embodiments," and "in other embodiments" appearing in various places in this specification do not necessarily refer to the same embodiment, but rather mean "one or more but not all embodiments," unless otherwise specifically emphasized. The terms "including," "comprising," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0048] This application example provides an example of an operation control system for building equipment. Please refer to Figure 1 As shown, Figure 1 The schematic diagram shows the structure of an operation control system of a building equipment provided in the present application. The system includes: building equipment 100 and control equipment 200.

[0049] The building equipment 100 is deployed with an equipment detection terminal 101, which is used to collect multimodal data of the building equipment.

[0050] The control device 200 is connected to the building equipment 100 and the equipment detection terminal 101 and is configured to perform:

[0051] Receive multimodal data of building equipment; determine the spatiotemporal correlation characteristics of the building equipment based on the multimodal data; use a pre-trained target prediction model to derive equipment status parameters of the building equipment based on the spatiotemporal correlation characteristics; and determine a corresponding target operation control strategy based on the equipment status parameters of the building equipment to control the building equipment to optimize the operation mode based on the target operation control strategy.

[0052] In some embodiments, the control device may be a mobile phone, a smart screen, a tablet computer, a wearable control device, an in-vehicle control device, an augmented reality (AR) device, a virtual reality (VR) device, a laptop computer, an ultra-mobile personal computer (UMPC), a netbook, a personal digital assistant (PDA), a projector, or a communication device such as a server, a storage device, a base station, or a smart car. The embodiments of the present application do not impose any restrictions on the specific type of the control device.

[0053] In some embodiments, building equipment refers to electrical equipment deployed in a building, such as at least one of an air conditioner (e.g., HVAC, refrigeration, etc.), an elevator, a lamp, and a sensor. It should be noted that there may be one or more building equipment, and there may be one or more types of building equipment. The examples and drawings herein use a single building equipment for ease of explanation and are not intended to limit the examples.

[0054] In some embodiments, multimodal data is collected by a device detection terminal deployed on the building equipment side, and includes different types of data of the building equipment, such as temperature, humidity, light, crowd density and other parameter variables. Multimodal data can be understood as a collection of different types of data including multiple physical quantities related to the operation of building equipment, environmental parameters, and equipment status information.

[0055] In some embodiments, the device detection terminal can be a non-invasive detection device, including a current ripple detection sensor and multi-input sensors (such as Hall effect sensors, temperature, humidity, and air quality sensors with built-in networking), capable of collecting multimodal data such as current, voltage, power, temperature, humidity, light intensity, and crowd density. For example, a Hall effect sensor collects device current data, while a temperature and humidity sensor senses changes in ambient temperature and humidity in real time.

[0056] In some embodiments, various types of device detection terminals can be deployed on the building equipment side. These terminals utilize non-intrusive detection technology to collect multimodal data from the operation of building equipment. Hard non-intrusive detection terminals do not require access to the internal circuitry of building equipment. For example, current ripple detection sensors can monitor current fluctuations in building equipment in real time without damaging the circuit structure. Soft non-intrusive detection terminals do not rely on inter-device communication protocols. For example, temperature and humidity sensors and air quality sensors with built-in networking capabilities can directly read operating data and environmental parameters of building equipment.

[0057] In some embodiments, the control device processes multimodal data based on a graph attention network (GAT) with a spatial attention mechanism to determine the spatial characteristics of building equipment, such as sensor location, connection dependencies, etc.; a layer of convolution based on causality (CausalConv1 D) analyzes multimodal data to obtain the temporal dynamic characteristics of building equipment, such as energy consumption fluctuations, equipment status changes, etc. The control device then uses a spatiotemporal encoder to integrate the spatial characteristics and temporal dynamic characteristics of the building equipment to obtain the spatiotemporal correlation characteristics of the building equipment.

[0058] In some embodiments, after the equipment detection terminal transmits the collected multimodal data to the control device, the control device can pre-train the general physical model based on physical knowledge and physical constraints to obtain a target prediction model. Then, after determining the spatiotemporal correlation characteristics of the building equipment, the spatiotemporal correlation characteristics are input into the target prediction model to use the target prediction model to automatically derive equipment state parameters that satisfy laws such as conservation of energy and conservation of mass, such as the actual working efficiency of the air-conditioning compressor and the lighting intensity of the lighting system.

[0059] In some embodiments, the control device can generate multiple operation control strategies based on the derived device state parameters, for example, in combination with equipment operation goals such as energy saving and comfort, through reinforcement learning and physical field inversion technology. Specifically, a variable air volume (VAV) curve is generated for air-conditioning equipment, a time-sharing scheduling strategy cluster is generated for the elevator system based on the passenger density field, and an illumination gradient matrix is ​​generated for the lighting system.

[0060] For another example, physical field inversion technology can be used to perform transient simulations on multiple operation control strategies, eliminate operation control strategies that violate the laws of thermodynamics, and finally select the target operation control strategy. The target operation control strategy is converted into a control signal that can be recognized by building equipment through an adaptive protocol converter, and the target operation control strategy is sent to the building equipment in the form of the control signal, so that the building equipment can adjust the equipment operation mode in real time based on the target operation control strategy, such as adjusting the air conditioning temperature, adjusting the elevator operation frequency, changing the lighting brightness, etc.

[0061] Through the embodiments of the present application, the operation control system of the building equipment includes building equipment and control equipment. The building equipment is deployed with an equipment detection terminal, which is used to collect multimodal data of the building equipment. The control equipment determines the spatiotemporal correlation characteristics of the building equipment based on the multimodal data. The target prediction model obtained in advance is used to deduce the equipment status parameters of the building equipment according to the spatiotemporal correlation characteristics. The equipment status parameters of the building equipment can be determined without precise modeling and a large amount of historical data participating in the calculation. It is no longer restricted by manual experience and fixed rules. Furthermore, the control equipment can also determine the corresponding target operation control strategy based on the derived equipment status parameters of the building equipment to control the building equipment to dynamically optimize the operation mode based on the target operation control strategy. It can achieve adaptive optimization of the operation effect of the building equipment without relying on manual experience and fixed rules, and reduce the energy waste of the building equipment.

[0062] In one possible implementation, Figure 2 As shown, the system also includes:

[0063] The multi-protocol gateway 300 is connected to the device detection terminal 101 and the control device 200, and is used to receive multimodal data transmitted by the device detection terminal and forward the multimodal data to the control device.

[0064] In some embodiments, the multi-protocol gateway can be installed in a building's weak current room. It can be equipped with multiple communication interfaces, such as Ethernet, Wi-Fi, and RS232, to meet the connection requirements of building equipment, device detection terminals, and control devices, breaking down protocol barriers between devices and enabling heterogeneous device interconnection. The multi-protocol gateway includes a rich set of built-in protocol stacks, including BACnet (Building Automation and Control Network), MODBUS (Serial Communication), and ZigBee, adapting to device detection terminals using different protocols.

[0065] In some embodiments, when the receiving device detects multimodal data sent by the terminal, protocol conversion and format unification are performed based on the dynamic point table configured by the multi-protocol gateway (which records the mapping relationship between different protocol data and the unified data model). For example, the ZigBee protocol data is converted into the JSON format of the MQTT protocol (a lightweight message transmission protocol), and the processed multimodal data is then forwarded to the control device via Ethernet. During the data transmission process, caching and flow control technologies can also be used to ensure stable transmission.

[0066] In one possible implementation, the control device is further configured to execute: automatically generate multiple operation control strategies based on the equipment status parameters of the building equipment; perform transient simulation processing on the multiple operation control strategies through physical field inversion technology to obtain simulation results corresponding to the multiple operation control strategies; select a target operation control strategy from the multiple operation control strategies based on the simulation results corresponding to the multiple operation control strategies; and send the target operation control strategy to the building equipment.

[0067] In some embodiments, after obtaining the equipment status parameters of building equipment, the control device comprehensively considers various factors, such as energy-saving targets, equipment lifespan, and user comfort. Reinforcement learning algorithms, combined with physical field inversion techniques, are used to generate multiple operational control strategies based on the equipment status parameters of building equipment. For example, physical field inversion techniques are used to construct a three-dimensional thermodynamic digital twin, mapping real-time data such as temperature and pedestrian flow to an LSTM-Transformer hybrid encoder to simulate the actual operating environment of building equipment.

[0068] In some embodiments, corresponding operational control strategies are generated for different types of building equipment. For example, for air conditioning equipment, a variable air volume (VAV) curve is output, and an action space containing multiple air supply modes (such as strong cooling mode, comfort mode, and energy-saving mode) is set. For elevator systems, a time-sharing scheduling strategy cluster is generated using a double-delayed deep deterministic policy gradient algorithm (TD3 algorithm) based on the inverted crowd density field. For lighting systems, an illumination gradient matrix is ​​generated.

[0069] A physics-based solver calculates inverse data corresponding to the state parameters of building equipment in real time. Indicators such as the predicted mean vote count (PMV) for comfort are integrated with this inverse data and used as a multi-objective reward function for the reinforcement learning algorithm, guiding it to generate optimal operational control strategies. The physics-based system then verifies feasible strategies and trains a pool of strategies, ultimately retaining several relatively optimal operational control strategies.

[0070] In some embodiments, the computational fluid dynamics (CFD) constraint solver in the physical field inversion technology can also be used to perform transient simulation processing on the multiple generated operation control strategies, such as a 0.5-second transient simulation. Specifically, each operation control strategy can be applied to the previously constructed three-dimensional thermodynamic digital twin to simulate the operation of the building equipment under the operation control strategy, thereby obtaining the simulation results corresponding to each operation control strategy. For example, the simulation results include key information such as temperature changes, energy consumption, and equipment load.

[0071] In some embodiments, the control device evaluates multiple operational control strategies based on simulation results corresponding to each strategy. Evaluation metrics may include energy-saving effects, compliance with physical laws (such as the laws of thermodynamics), and impact on user comfort. For example, operational control strategies that exhibit temperature changes that do not conform to physical laws (such as sudden and significant temperature drops or increases) during the simulation are eliminated. Then, from the remaining multiple operational control strategies, the one that performs best in comprehensive metrics such as energy-saving and comfort is selected as the target operational control strategy.

[0072] In some embodiments, the control device converts the target operation control strategy into an instruction format that can be recognized by the building equipment through an adaptive protocol converter. For example, if the building equipment adopts the building automation and control network BACnet protocol, the target operation control strategy is converted into control instructions that comply with the BACnet protocol, and then the converted control instructions are sent to the corresponding building equipment through a network connection to control the building equipment to adjust the operation mode according to the target operation control strategy.

[0073] By using physical field inversion technology to perform transient simulations on multiple operation control strategies, the operation effect can be pre-evaluated before the operation control strategy is implemented, that is, the simulation results corresponding to each operation control strategy can be obtained, avoiding the use of unreasonable operation control strategies and improving the scientificity and effectiveness of determining the target operation control strategy finally adopted.

[0074] In one possible implementation, the control device is deployed with a graph attention network based on a spatial attention mechanism, a layer of convolution based on causality, and a spatiotemporal encoder; the graph attention network based on the spatial attention mechanism is used to determine the spatial characteristics of building equipment based on multimodal data, and the spatial characteristics are used to characterize the spatial topological relationship between various types of equipment contained in the building equipment; the layer of convolution based on causality is used to determine the temporal dynamic characteristics of building equipment based on multimodal data, and the temporal dynamic characteristics are used to characterize the energy consumption fluctuations and dynamic changes of equipment status changes over time of building equipment; the spatiotemporal encoder is used to determine the spatiotemporal correlation characteristics of building equipment based on spatial characteristics and temporal dynamic characteristics.

[0075] In some embodiments, the control device inputs the received multimodal data into a graph attention network based on a spatial attention mechanism. First, based on the physical location, connection relationships, and other information of each device (or type of device) in the building equipment, each device is abstracted as a node in a relationship graph, and the relationships between devices are abstracted as edges, thereby constructing a graph structure. For example, air conditioners, lighting, elevators, and other equipment on the same floor are used as nodes, and the electrical connections and spatial proximity between these devices are used as edges to construct a graph representing the spatial relationships of building equipment.

[0076] In some embodiments, in a graph attention network, each node calculates an attention coefficient with respect to other nodes. This attention coefficient reflects the degree to which each node pays attention to the features of other nodes, thereby highlighting important spatial correlation information. For example, for an air conditioning device node, the attention coefficient for these nodes is calculated based on the connection relationship and data features with surrounding lighting devices and sensor nodes. In the specific calculation process, a learnable weight matrix is ​​used to perform a linear transformation on the node features, and then the attention coefficient is calculated using an activation function.

[0077] In some embodiments, the features of adjacent nodes are weighted and summed based on the calculated attention coefficient to update the feature representation of each node. After processing through multiple layers of graph attention, spatial features are ultimately obtained that characterize the spatial topological relationships between the various devices within a building. For example, this approach can capture spatial information such as the collaborative relationships between devices on different floors and the mutual influence of devices within the same area.

[0078] In some embodiments, the control device inputs time series data from multimodal data into a causal convolution layer (Causal Conv1 D). To ensure causality, that is, the output depends only on the current and previous inputs, the convolution operation uses a causal convolution approach. For example, for time series such as building equipment energy consumption data and equipment status data, a sliding convolution operation is performed along the time dimension with a fixed convolution kernel size (e.g., 3).

[0079] As the convolution kernel slides, a weighted summation is performed on the input time series data to extract local features from the time series data. This approach captures the dynamic patterns of energy consumption fluctuations and device status changes over time in building equipment. After the convolution operation is completed, it is processed through an activation function (such as ReLU) to obtain dynamic time series features.

[0080] In some embodiments, the spatiotemporal encoder receives the spatial features output by a graph attention network based on a spatial attention mechanism and the temporal dynamic features output by a layer of convolution based on causality, and splices or fuses the spatial features and the temporal dynamic features in other ways to obtain fused features. For example, the spatial features and the temporal dynamic features are spliced ​​in the feature dimension, and then processed by the encoding layer of the spatiotemporal encoder. The encoding layer can use a fully connected layer or a lightweight Transformer structure to specifically perform dimensionality reduction, nonlinear transformation and other operations on the fused features to remove redundant information and extract more representative spatiotemporal correlation features. For example, through encoding processing, it is possible to discover the changing patterns of the spatial relationships of devices at different time points, as well as information such as the long-term impact of spatial layout on device energy consumption and state changes.

[0081] By respectively utilizing a graph attention network based on the spatial attention mechanism and a layer of convolution based on causality to extract features from multimodal data from both spatial and temporal dimensions, the spatiotemporal correlation characteristics of building equipment can be captured more comprehensively and accurately. Moreover, since the spatiotemporal correlation characteristics take into account the spatial topological relationship and temporal dynamic changes of building equipment, they can more accurately reflect the actual operation of building equipment, thereby improving the accuracy of the target prediction model in predicting the state parameters of building equipment using spatiotemporal correlation features.

[0082] In one possible implementation, the spatiotemporal encoder is further used to compress the spatiotemporal correlation characteristics after determining the spatiotemporal correlation characteristics of building equipment; wherein the compression processing is used to remove redundant information in the spatiotemporal correlation characteristics, and retain at least one of the cross-device collaborative patterns, periodic laws, mutation laws and abnormal event correlation characteristics in the spatiotemporal correlation characteristics.

[0083] In some embodiments, multimodal data of building equipment can be collected by equipment detection terminals deployed on building equipment. Specifically, real-time collection and collection at a predetermined frequency can be adopted.

[0084] In some embodiments, real-time data collection is used for data that requires high real-time performance, such as electrical parameters such as current and voltage of building equipment. The equipment detection terminal continuously monitors these electrical parameters and transmits the collected electrical parameters to the control device in real time, allowing the control device to promptly detect abnormal operation of the building equipment.

[0085] In some embodiments, data that changes relatively slowly, such as environmental parameters such as temperature and humidity, can be collected at a predetermined frequency, for example, every 5 minutes, so as to reduce the pressure of data transmission and processing while meeting the collection requirements.

[0086] In an optional embodiment, after determining the spatiotemporal correlation features, the spatiotemporal encoder may implement compression processing on the spatiotemporal correlation features in the following ways:

[0087] Dimensionality reduction algorithms such as principal component analysis (PCA) are used to map high-dimensional spatiotemporal correlation features into a low-dimensional space, removing redundant information from the spatiotemporal correlation features. For example, spatiotemporal correlation features containing a large number of correlated parameters can be subjected to PCA dimensionality reduction to extract a small number of principal components that retain most of the original feature information.

[0088] During the compression of spatiotemporal correlation features, the attention mechanism is used to highlight important information within these features, such as cross-device collaboration patterns, periodic patterns, mutation patterns, and features associated with abnormal events. The attention mechanism assigns different weights to different feature components, giving higher weights to important features, thereby preserving this key information during the compression process.

[0089] By compressing the spatiotemporal correlation features and removing redundant information in the spatiotemporal correlation features, the dimension and complexity of the data are reduced, and the computing power and storage requirements of the control equipment in the subsequent processing process are reduced. At the same time, key information such as the cross-device collaborative mode, periodic laws, mutation laws and abnormal event correlation characteristics in the spatiotemporal correlation features are retained. For example, by retaining the periodic laws, the operating mode of the building equipment can be adjusted in advance according to the periodic changes in the energy consumption of the building equipment to achieve the purpose of energy saving.

[0090] By retaining the correlation characteristics of abnormal events, potential building equipment failures can be promptly identified, allowing appropriate maintenance measures to be taken to ensure the normal operation of building equipment. In addition, the retained key information can also include other information that is important for accurately analyzing the operating status of building equipment, predicting equipment failures, and formulating reasonable operation control strategies.

[0091] In one possible implementation, the control device is further configured to: before determining the spatiotemporal correlation characteristics of the building equipment based on the multimodal data, perform data cleaning processing on the multimodal data to obtain processed multimodal data.

[0092] In some embodiments, multimodal data includes physical quantities related to the operation of building equipment (such as current, voltage, power), environmental parameters (such as temperature, humidity, light intensity), and equipment status information (such as equipment on status, equipment off status, fault alarm information), etc.

[0093] In some embodiments, the cleaning process includes at least one of: a denoising process, a normalization process, and a feature extraction process.

[0094] In some embodiments, the control device may employ a filtering algorithm to remove noise interference that may exist in multimodal data. For example, for physical quantity data such as current and voltage, a moving average filtering algorithm can be used to smooth the data by calculating the average value of the data within a certain time window, thereby reducing the impact of random noise.

[0095] In some embodiments, when the data distribution range is known, linearly mapping each feature value in the multimodal data to the interval [0, 1] can ensure that different features have the same scale, facilitating feature extraction and model training. For cases where the data distribution is unknown or outliers are present, normalization can be performed based on the mean and standard deviation of the multimodal data.

[0096] In some embodiments, the main characteristic components can also be extracted from the multimodal data, for example, by projecting the original data into a new low-dimensional space through linear transformation so as to maximize the variance of the projected data. In this low-dimensional space, each principal component is a linear combination of the original features, and the principal components are mutually orthogonal. In addition, the correlation between different features can be analyzed to remove redundant features with high correlation. For example, for ambient temperature and humidity, if the correlation coefficient between these two features is high (such as greater than 0.8), you can choose to retain one of the features to reduce the complexity of the data.

[0097] In some embodiments, after completing the data cleaning process, the control device uses the processed multimodal data to determine the spatial characteristics of the building equipment through a graph attention network based on the spatial attention mechanism, and determines the temporal dynamic characteristics of the building equipment based on a layer of convolution based on causality, and then the spatiotemporal encoder combines these two characteristics to determine the spatiotemporal correlation characteristics of the building equipment.

[0098] In the embodiment of the present application, denoising is performed on multimodal data to remove noise interference in the multimodal data, making the multimodal data smoother and more accurate, and reducing the impact of noise on subsequent analysis and modeling. Normalizing the multimodal data allows different features to have the same scale, avoiding the problem of unstable model training caused by excessive differences in feature scales. Feature extraction processing on multimodal data reduces the dimension of the multimodal data, removes redundant information, and improves the effectiveness of the multimodal data. Based on the processed multimodal data, the accuracy and stability of spatiotemporal correlation features and target prediction models can be improved.

[0099] In one possible implementation, the target prediction model is obtained by training the model parameters of a general physical model based on a training set, physical knowledge, and physical constraints. The training set includes: multiple sample multimodal data, and sample device state parameters corresponding to each sample multimodal data.

[0100] In some embodiments, in an actual building environment, a device detection terminal deployed on the building equipment continuously collects multiple sample multimodal data, such as: physical quantities related to the operation of building equipment (such as current, voltage, power), environmental parameters (such as temperature, humidity, light intensity), and equipment status information (such as equipment on status, equipment off status, fault alarm information), etc.

[0101] In some embodiments, for each sample multimodal data, sample device state parameters corresponding to the sample multimodal data are determined. For example, for air-conditioning equipment, sample device state parameters including compressor operating efficiency, refrigerant flow, etc. are obtained by direct measurement; for elevator equipment, sample device state parameters including motor load conditions, operating wear degree, etc. are obtained by direct measurement.

[0102] Afterwards, the collected multimodal data samples and their corresponding device status parameters are combined into a training set. To ensure the quality and representativeness of the training set, the multimodal data samples can also be cleaned, such as removing outliers and performing normalization.

[0103] In some embodiments, a general physical model can be selected as the base model. This general physical model can be constructed based on physical knowledge that includes the basic physical laws and principles that describe the operation of building equipment, such as the laws of thermodynamics and the principles of fluid mechanics. For example, for air conditioning equipment, a physical model based on thermodynamic cycle principles can be selected; for power systems, a physical model based on circuit theory can be selected.

[0104] In some embodiments, relevant physical knowledge (such as the laws of thermodynamics, fluid mechanics, and other basic physical laws and principles) can also be integrated into the general physical model and reflected in the structural design, parameter settings, or loss function of the general physical model. The following example uses the air conditioning equipment in an office building as an example to illustrate the integration of sample structure, loss design, and other content into a real-world scenario:

[0105] Consider analyzing the operating status and model training of air conditioning equipment in an office building. For example, there is an initial equipment status sample (i.e., the original sample), which records the operating parameters of the air conditioning equipment at a certain moment, such as the current temperature setting of 24°C, the cooling power of 5 kilowatts, and the wind speed of medium. During sample construction, for the initial equipment status sample, the control device generates one positive sample through physical field perturbations and randomly samples four negative samples.

[0106] Specifically, when generating positive samples through physical field perturbations, some physical environment parameters are adjusted appropriately while keeping the device state parameters unchanged. For example, the initial humidity around the air conditioner is adjusted from 50% to 55%, while the core device state parameters such as the current temperature, cooling power, and wind speed of the air conditioner remain unchanged. This results in a positive sample with the same device state parameters as the original sample.

[0107] Next, negative samples are obtained through random sampling. For example, the first negative sample might be to raise the temperature setting to 28°C, reduce the cooling power to 3 kW, and set the wind speed to low. The second negative sample might be to set the temperature to 20°C, increase the cooling power to 7 kW, and set the wind speed to high. The third negative sample might change the air conditioner's operating mode from cooling to heating, while adjusting other parameters. The fourth negative sample might simulate an air conditioner failure, such as cooling power fluctuating unstably between 2 and 4 kW and the current temperature failing to maintain the set value. These four negative samples have different operating parameters from the original sample, representing different operating states.

[0108] In the loss design phase, the improved NT-Xent loss function is used to train the model. This loss function is used to determine how big the gap is between the model's predicted results and the actual operation results. Assuming that the model predicts the device state of a certain sample and obtains a predicted result, the control device can use the loss function to calculate the difference between the predicted result and the actual result.

[0109] It's important to note that there's a key factor in the loss function: the temperature coefficient τ, which determines how the model focuses on sample features. The value of τ isn't fixed, but rather dynamically adjusts based on the actual physical state of the office building.

[0110] For example, during hot summer weather, thermodynamic parameters within an office building (such as the indoor-outdoor temperature difference and air conditioning load) will change. These changes are reflected in the physical field parameter Px. If the contribution weight α of the physical field parameter Px to the temperature coefficient τ is large, it means that the physical field state has a significant impact on model training. For another example, when a large number of people suddenly enter the office building, causing the indoor temperature to rise rapidly, the physical field parameter Px will change significantly. α, as the contribution weight of the physical field parameter Px to the temperature coefficient τ, when the physical field quantitative indicators |Px| representing the real-time physical state of building equipment (such as thermodynamic and fluid dynamic parameters) change significantly, α will propagate the change in |Px| to τ. In this case, if the baseline value β (i.e., the temperature bias term) of the temperature coefficient τ is set to an appropriate baseline, it can ensure that τ does not become zero due to extreme conditions, maintaining computational stability.

[0111] For example, when the τ value is large, the model will focus more on the global characteristics of the sample. For example, the model will comprehensively consider the air conditioner's operating status at different times of the day, as well as multiple factors such as the number of people indoors and the outdoor temperature, to determine the overall operation of the air conditioner. When the τ value is small, the model will focus on difficult samples that are difficult to predict. For example, when the air conditioner experiences intermittent minor faults, resulting in unstable operating parameters, the model will focus more on analyzing the sample characteristics of these special cases, attempting to identify the fault patterns and improve the accuracy of predictions for such complex situations.

[0112] In some embodiments, physical constraints can be set based on actual physical limitations and operational requirements. For example, these physical constraints can include the operating range, performance indicators, safety requirements, etc. of the equipment. For example, for air conditioning equipment, constraints such as upper and lower limits on cooling capacity and the maximum operating pressure of the compressor can be set; for elevator equipment, constraints such as operating speed limits and upper load limits can be set.

[0113] Afterwards, the model parameters of the general physical model can be initialized using random initialization or initialization methods based on prior knowledge. The model can then be trained using the training set. During the training process, the model parameters are continuously adjusted using optimization algorithms (such as stochastic gradient descent and the Adam optimizer) to minimize the error between the model's predictions and the state parameters of the sample devices. Furthermore, physical constraints can be introduced into the optimization process using methods such as the Lagrange multiplier method and the projected gradient method to ensure that the model parameters meet the physical constraints.

[0114] In some embodiments, during the training process, a validation set can also be used to evaluate the performance of the general physical model. Based on the evaluation results, the structure, parameters or training strategy of the general physical model can be adjusted to improve the accuracy and generalization ability of the general physical model.

[0115] The model parameters of the general physical model are trained based on the training set, enabling the general physical model to learn the mapping relationship between sample multimodal data and sample equipment state parameters. At the same time, physical knowledge and physical constraints are introduced to ensure that the prediction results of the target prediction model conform to physical laws, thereby improving the accuracy of the target prediction model's predictions of building equipment state parameters. Furthermore, the introduction of physical knowledge and physical constraints allows the use of prior knowledge during model training, reducing reliance on large amounts of training data. Even with limited training data, reasonable predictions based on physical laws can be made, improving the generalization and adaptability of the target prediction model.

[0116] In one possible implementation, the control device is further configured to:

[0117] Pre-acquire multiple sample multimodal data of sample building equipment, and sample equipment status parameters corresponding to each of the multiple sample multimodal data, wherein each sample multimodal data includes at least one of temperature, humidity, light, crowd density, current, voltage, and power; generate a training set based on the multiple sample multimodal data and the sample equipment status parameters corresponding to each of the multiple sample multimodal data; add physical knowledge and physical constraints to a general physical model, wherein the physical knowledge includes at least multiple physical laws; under the guidance of the physical constraints, train the model parameters of the general physical model based on the training set, and obtain the target prediction model after multiple iterative optimizations.

[0118] In some embodiments, the control device collects data at a certain frequency through various sensors deployed in the sample building equipment, thereby obtaining multiple sample multimodal data. Each sample multimodal data includes at least one of temperature, humidity, light, crowd density, current, voltage, and power. For example, temperature and humidity sensors can regularly collect environmental temperature and humidity; light sensors record indoor and outdoor light intensity; crowd density sensors use infrared sensors or cameras to count the number of people in a specific area; current and voltage sensors and power meters monitor the current, voltage, and power of the equipment in real time. At the same time, using the monitoring system or professional detection tools provided by the sample building equipment, the sample equipment status parameters corresponding to each sample multimodal data are recorded, such as the operating efficiency and degree of wear of the sample building equipment.

[0119] In some embodiments, the control device can also pair and pre-process the acquired sample multimodal data and the corresponding sample device state parameters. For example, the sample multimodal data is first cleaned by removing outliers and noise data. For example, when a current value has a spike that is obviously beyond the normal range, the current value is determined to be an outlier and removed. The sample multimodal data is then normalized to unify the sample multimodal data of different ranges and magnitudes to the same scale. For example, all data are mapped to the [0,1] interval using the minimum-maximum normalization method. Finally, the processed data is divided into a training set and a validation set according to a certain ratio. For example, the training set accounts for 70%-80% and the validation set accounts for 20%-30%, but this is not limited to this. The examples in this application do not specifically limit this.

[0120] In some embodiments, multiple physical laws are incorporated into the general physical model as physical knowledge. These physical laws cover multiple disciplines. For example, when building an air conditioning equipment model, the Carnot cycle principle and heat transfer equations from thermodynamics are introduced to describe the energy conversion and heat transfer in the air conditioning cooling process. For the power equipment model, Ohm's law and Kirchhoff's laws from electromagnetism are applied. By incorporating these physical laws into the structure and calculation process of the general physical model, the general physical model can better simulate actual physical phenomena.

[0121] In some embodiments, physical constraints can be set based on actual physical limitations and equipment operating requirements as constraints during model training to ensure that the model's output conforms to physical reality. For example, these constraints can be categorized as equality constraints and inequality constraints. For example, for the energy conversion process of building equipment, equality constraints based on energy conservation can be used; for operating parameters of building equipment, such as temperature, pressure, and power, inequality constraints based on the range of values ​​can be used.

[0122] In some embodiments, the model parameters of the general physical model are first initialized. For example, random initialization can be used, or reasonable initial values ​​can be assigned based on prior knowledge. Under the guidance of physical constraints, the general physical model is trained using a training set. During the training process, optimization algorithms such as stochastic gradient descent and the Adam optimizer are used to continuously adjust the model parameters to minimize the error between the model's prediction results and the state parameters of the sample devices. At the same time, constrained optimization techniques (such as the Lagrange multiplier method and the projected gradient method) are used to ensure that the model parameters meet the physical constraints.

[0123] After each round of training, the model performance is evaluated using the validation set, and the prediction error of the model on the validation set is calculated, such as the mean square error, mean absolute error, etc. Based on the validation results, the model's hyperparameters (such as the learning rate, number of iterations, etc.) are adjusted, or the model structure is fine-tuned to improve the model's generalization ability and accuracy. The above training, validation, and adjustment process is repeated, and multiple iterative optimizations are performed until the model performance reaches a satisfactory level, ultimately obtaining the target prediction model.

[0124] In the embodiments of the present application, by using a large amount of sample multimodal data for training, the target prediction model is able to learn the complex relationship between multimodal data and equipment status parameters, thereby improving the accuracy of predictions of building equipment status parameters. At the same time, the introduction of physical knowledge and physical constraints makes the prediction results of the target prediction model more consistent with actual physical laws, reducing unreasonable prediction results. In addition, physical constraints limit the model's search space, preventing the model from overfitting the training data, thereby improving the model's ability to generalize to new data. Even when faced with new building environments and equipment operating conditions, reasonable predictions can be made based on physical laws.

[0125] In one possible implementation, the general physical model includes multiple physical differential equations, each of which is added as a regular term to the loss function of the general physical model, or the deep learning network structure of the general physical model is defined to meet the restrictions of each physical differential equation.

[0126] In some embodiments, the general physical model consists of multiple physical differential equations. These equations, based on knowledge from different physical fields, are used to describe the physical phenomena during the operation of building equipment and can determine the temporal and spatial relationships between various physical quantities in building equipment operation. For example, when describing the air conditioning and cooling process, the heat conduction differential equations in thermodynamics and the flow equations in fluid mechanics are involved; for power equipment, simplified forms of Maxwell's equations in electromagnetism are involved.

[0127] In some embodiments, each physical differential equation can also be added as a regularization term to the loss function of the general physical model (such as the mean square error). This loss function is usually used to measure the error between the model prediction results and the actual sample device state parameters. For example, for a heat conduction differential equation that describes temperature changes, after it is discretized, the residual of the equation is calculated at each time step and spatial point, and then the sum of the squares of these residuals is added to the loss function as a regularization term. In this way, during the model training process, the optimization algorithm not only minimizes the error between the model prediction results and the sample data, but also prompts the model to satisfy the physical laws described by the physical differential equation.

[0128] In some embodiments, appropriate weights are assigned to each regularization term to balance the impact of traditional loss terms and physical regularization terms. Weights can be selected and adjusted based on the specific application scenario and the importance of the physical equation. If a physical equation is crucial for describing the critical operating state of building equipment, a higher weight can be assigned to the regularization term corresponding to that equation; otherwise, a lower weight can be assigned to the regularization term corresponding to that equation.

[0129] In some embodiments, a deep learning network structure is defined for a general physical model to satisfy the constraints of each physical differential equation. This can be achieved through specialized design of the network's layer structure and neuron connectivity. For example, for a model that needs to satisfy the heat conduction equation, the convolution kernel parameters in the network layer can be designed so that they simulate the diffusion process of heat conduction during calculations. During the network's forward propagation, the computational results of each neuron are ensured to conform to the physical relationships specified by the physical differential equation.

[0130] During network training, constrained optimization methods (such as the Lagrange multiplier method) are used to ensure that the network's output meets the constraints of physical differential equations. When the network's output violates a physical differential equation, a corresponding penalty term is generated. The optimization algorithm adjusts the network parameters based on this penalty term, so that the network gradually meets the physical constraints.

[0131] After constructing the loss function and defining the deep learning network structure, the general physics model is trained using the previously generated training set. During training, optimization algorithms (such as stochastic gradient descent and the Adam optimizer) continuously adjust the model parameters to minimize the loss function. Furthermore, because the loss function includes the regularization term of the physical differential equation and the network structure satisfies the constraints of the physical differential equation, the model gradually learns parameters that conform to the laws of physics, ultimately achieving the target prediction model.

[0132] By integrating physical differential equations into the loss function and network structure, the model's predictions strictly adhere to physical laws, preventing the model from producing predictions that are inconsistent with physical reality and improving the model's consistency and reliability at the physical level. For example, when predicting the cooling effect of an air conditioner, the model will not produce predictions of temperature changes that violate the laws of thermodynamics.

[0133] Because the model learns physical laws during training, it can make reasonable extrapolations and predictions based on physical knowledge, even with limited training data, thereby improving its generalization capabilities. Furthermore, the introduction of physical differential equations gives the model's predictions clear physical meaning, making them easier for operators to understand and interpret.

[0134] In one possible implementation, the control device is also used to: train the model parameters of the general physical model based on the training set under the guidance of physical constraints, calculate the loss function through the back propagation algorithm, and optimize and adjust the model parameters of the general physical model based on the loss function and the adaptive fine-tuning strategy.

[0135] In some embodiments, when training a general physical model based on a training set, it is first necessary to define a suitable loss function, which can be the mean square error (MSE), mean absolute error (MAE), etc., to measure the difference between the prediction results of the general physical model and the actual sample device state parameters.

[0136] In some embodiments, physical constraints are incorporated into the loss function to ensure that the prediction results of the general physical model conform to physical laws. For example, the physical constraints can be added to the loss function in the form of regularization terms.

[0137] In some embodiments, the backpropagation algorithm is an efficient algorithm for calculating the gradient of the loss function with respect to the model parameters. In each training iteration, the samples in the training set are input into the universal physical model to obtain the prediction results of the universal physical model. Then, the loss value is calculated according to the loss function defined above. Next, the backpropagation algorithm is used to start from the loss function and calculate the partial derivatives (i.e., gradients) of the loss function with respect to each parameter in the model layer by layer. These gradients represent the rate of change of the loss function under the current parameter value and indicate the direction of parameter adjustment.

[0138] In the embodiment of the present application, according to the calculated gradient and adaptive fine-tuning strategy, the model parameters of the general physical model are adjusted. Wherein, the adaptive fine-tuning strategy can be based on the training situation of the model, using a variant algorithm of stochastic gradient descent (SGD), such as adaptive gradient algorithm Adagrad, adaptive learning rate adjustment algorithm Adadelta, adaptive moment estimation optimization algorithm Adam and other adaptive optimization algorithms, to dynamically adjust the hyperparameters such as learning rate, to improve training efficiency and stability. For example, taking the Adam algorithm as an example, the algorithm combines the idea of ​​momentum and adaptive learning rate, and dynamically adjusts the learning rate of each parameter according to the first-order moment estimation (mean) and second-order moment estimation (variance) of the gradient.

[0139] Repeating the above steps, the sample data from the training set is continuously fed into the universal physical model. The loss function is calculated, the gradient is calculated through backpropagation, and the parameters are adjusted using an adaptive fine-tuning strategy, performing multiple iterative optimizations. As the number of iterations increases, the parameters of the universal physical model are gradually adjusted to the optimal values ​​that minimize the loss function, resulting in a target prediction model that can accurately predict the state parameters of the sample device.

[0140] Through this application example, by using a backpropagation algorithm to calculate the loss function and adjusting parameters based on an adaptive fine-tuning strategy, the model can continuously learn the characteristics and patterns of samples in the training set, gradually reducing the error between the model's predictions and the true values. Furthermore, the incorporation of physical constraints ensures that the model's predictions conform to physical laws, avoiding unreasonable predictions and further improving the model's prediction accuracy.

[0141] Furthermore, the adaptive fine-tuning strategy dynamically adjusts hyperparameters such as the learning rate based on the model's training progress, avoiding the training instability that can occur with fixed learning rates in traditional optimization algorithms. For example, in the early stages of training, a larger learning rate allows the model to converge quickly; while in the later stages of training, a smaller learning rate allows the model to adjust parameters more finely, avoiding overshooting the optimal solution and thus improving the stability of model training.

[0142] In one possible implementation, each sample of multimodal data includes: generating a first number of positive sample data using a physical field perturbation algorithm, and randomly sampling a second number of negative sample data, wherein the first number and the second number conform to a predetermined positive-negative sample ratio.

[0143] In some embodiments, for each sample of multimodal data, in order to enrich the diversity of training data and improve the generalization ability of the model, two different methods can be used to generate positive sample data and negative sample data, and their number is made to conform to a predetermined positive-negative sample ratio. This allows the model to better learn the differences and features between the data during the training process, thereby improving the performance of the model.

[0144] It should be understood that the physical field perturbation algorithm is based on physical knowledge and principles, and performs reasonable perturbation operations on the physical field-related parameters in the original sample multimodal data. Therefore, in the example of this application, the physical field perturbation algorithm is used to generate a first number of positive sample data. Through the physical field changes that may occur in actual operation, the generated positive sample data has similar physical properties and device states as the original sample data and belongs to the same type of sample (i.e., positive sample). By repeating this perturbation operation multiple times, the first number of positive sample data can be obtained.

[0145] In this embodiment, a second amount of negative sample data is obtained through random sampling, where data is randomly extracted from an existing sample multimodal data set (or a larger related data set) as negative samples. It should be understood that during the extraction process, the specific relationship between the data and the original sample data is not considered, as long as the data format and type requirements are met.

[0146] For example, for a sample of multimodal data about building lighting equipment, the randomly extracted negative sample data may come from operating data of other types of equipment, or data of the same equipment under different operating conditions that differ significantly from the original sample data. In this way, the extracted negative sample data has obvious differences from the original sample data in physical characteristics and equipment status, and belongs to a different type of sample (i.e., negative sample). Therefore, through the random sampling operation, a second number of negative sample data is obtained.

[0147] It should be understood that the positive sample data generated by the physical field perturbation algorithm simulates the changes in the physical field in actual operation, and the randomly sampled negative sample data covers a variety of different situations. The diverse sample data generated in this way can allow the model to learn a wider range of features and laws, enhance the model's adaptability to different scenarios and changes, and improve the model's generalization ability in practical applications.

[0148] In addition, the examples in this application determine the specific values ​​of the first and second quantities based on a predetermined positive-negative sample ratio. For example, the predetermined positive-negative sample ratio is 1:4, meaning the number of negative samples is four times the number of positive samples. In actual operation, an appropriate positive-negative sample ratio can be set based on specific model training requirements and data characteristics to achieve optimal training results.

[0149] It should be understood that a reasonable ratio of positive and negative samples can enable the model to learn data information more comprehensively during the training process, avoiding the problem of model training being biased towards a certain type of sample due to an imbalance in the number of positive and negative samples, thereby improving training efficiency and reducing training time and resource consumption.

[0150] In some embodiments, the generated first number of positive sample data and the second number of negative sample data can be integrated with the original sample multimodal data to form a new sample set. This new sample set includes different types of sample data, which can provide richer information for model training and help the model better learn the relationship between the sample multimodal data and the sample device state parameters.

[0151] By using the examples in this application, by generating positive and negative sample data and controlling the positive-negative sample ratio, the model can more clearly distinguish different types of samples during training, learn similar features between positive samples and differences between positive and negative samples, thereby improving the ability to classify and predict samples. For example, when predicting the operating status of building equipment, it can more accurately determine whether it is in a normal operating state (positive sample) or an abnormal operating state (negative sample).

[0152] In one possible implementation, the control device is further used to: perform confidence assessment on multiple sample multimodal data by calculating the physical rationality scores of the sample device state parameters corresponding to each of the multiple sample multimodal data; update the weight of each sample multimodal data using a momentum accumulation mechanism based on the confidence assessment results of each sample multimodal data; select a preset number of target sample multimodal data from the multiple sample multimodal data based on the weight of each sample multimodal data; and use the preset number of target sample multimodal data and the sample device state parameters corresponding to the target sample multimodal data as a training set.

[0153] In some embodiments, the control device may first calculate the physical rationality scores of the sample device state parameters corresponding to each of the multiple sample multimodal data based on pre-set physical rules and constraints, and obtain a confidence assessment result for each sample multimodal data.

[0154] For example, for a sample of multimodal data involving an air conditioner, the corresponding sample device state parameters include information such as cooling capacity and energy consumption. Therefore, the control device can determine whether the sample device state parameters of the air conditioner are reasonable based on the laws of thermodynamics and the performance specifications of the air conditioner. If the cooling capacity exceeds the theoretical maximum cooling capacity of the air conditioner, or if the relationship between energy consumption and cooling capacity does not conform to the law of conservation of energy, the physical rationality score of the sample device state parameters will be relatively low.

[0155] In some embodiments, the control device may also use a momentum accumulation mechanism to update the weight of each sample multimodal data. It should be understood that the momentum accumulation mechanism is similar to the momentum optimization algorithm in deep learning, which takes into account the performance of the sample data in multiple evaluations. Specifically, the update of the current weight of each sample multimodal data depends not only on the current confidence assessment result, but also on the previous weight information, wherein the momentum coefficient can be used to control the degree of influence of the previous weight on the current weight, and the value range is usually between 0 and 1. In addition, the weights of sample data with higher confidence in multiple evaluations gradually increase, while the weights of sample data with lower confidence gradually decrease.

[0156] It should be understood that since the momentum accumulation mechanism takes into account the performance of sample data in multiple evaluations, it can avoid excessive weight fluctuations caused by single evaluation errors, allowing the model to learn data features more stably during training, reducing the risk of model overfitting or underfitting, and improving the stability and generalization ability of the model.

[0157] In some embodiments, the control device may further select a preset number of target sample multimodal data from the plurality of sample multimodal data based on the weight of each sample multimodal data. For example, the sample data may be sorted in descending order of weight, and then a preset number of samples may be selected as target samples. This ensures that the sample data in the training set is of high confidence and physical plausibility, thereby improving the quality of the training set.

[0158] In some embodiments, the control device can also combine a preset number of selected target sample multimodal data and their corresponding sample device state parameters to obtain a training set for subsequent training of the general physical model. Since the sample data has been screened and weighted, it can more effectively help the model learn accurate physical laws and data characteristics during the model training process.

[0159] It should be noted that selecting multimodal data of target samples with higher confidence as the training set can reduce the model's processing of invalid data during training, improve training efficiency, and enable the model to converge to the optimal solution faster, saving training time and computing resources.

[0160] In the embodiment of the present application, confidence assessment is performed through physical rationality scores, and the momentum accumulation mechanism is used to update sample weights, so that high-quality (more in line with physical laws) sample data can be screened out as a training set, which can reduce the interference of noise data and unreasonable data on model training, thereby improving the training effect and prediction accuracy of the model.

[0161] In one possible implementation, the control device is also deployed with a physical field inversion module and a target prediction model; the physical field inversion module is used to solve the inverse problem by substituting the time-space correlation characteristics into the physical differential equations in the target prediction model to derive the equipment state parameters of the building equipment that satisfy the conservation law.

[0162] In the example of this application, since the general physical model contains multiple physical differential equations, for example, equations describing physical processes such as heat conduction, energy conservation, and fluid mechanics, during the training process, the physical differential equations are added to the loss function as regular terms, or the deep learning network structure is defined to meet the restrictions of the physical differential equations, and the target prediction model is obtained through multiple iterative optimizations.

[0163] In some embodiments, the physical field inversion module substitutes the obtained spatiotemporal correlation characteristics into the physical differential equations in the target prediction model. Since the physical differential equations describe the changing relationship between physical quantities, the inverse problem is to use mathematical calculations and numerical solution methods, such as finite difference method, finite element method, etc., under certain known conditions (i.e., spatiotemporal correlation characteristics), to solve the unknown physical quantities (i.e., equipment state parameters of building equipment) that meet these conditions. In the solution process, it is necessary to ensure that the solved equipment state parameters satisfy conservation laws, such as conservation of energy, conservation of mass, etc. Specifically, the physical differential equations can be discretized and converted into a set of algebraic equations, and then these equations are solved iteratively to gradually approximate the equipment state parameter solutions that satisfy the conservation laws.

[0164] It should be understood that the physics-based inversion process is based on physical knowledge and laws, giving the prediction results clear physical meaning. Compared to purely data-driven models, this physics-based inversion approach can better explain the reasons behind the predicted device state parameters. Furthermore, for the complex physical processes involved in building equipment operation, such as heat transfer and fluid flow, this allows for the effective derivation of device state parameters even in complex scenarios.

[0165] Using the example of this application, since physical differential equations and conservation laws reflect the essential laws of physical processes, they can constrain the scope of solutions and avoid unreasonable prediction results. By solving the inverse problem through the physical field inversion module and combining the physical differential equations and conservation laws in the target prediction model, the equipment state parameters of the building equipment can be more accurately derived. Furthermore, based on the accurate equipment state parameters, the building equipment can be controlled to optimize its operating mode, achieving goals such as energy conservation and increasing equipment lifespan, thereby improving the overall operating efficiency and management level of the building system.

[0166] In one possible implementation, the control device is further configured to:

[0167] After the target prediction model is obtained through multiple iterative optimizations, only the physical inversion module deployed in the control device is optimized in the first fine-tuning stage, and the model parameters of the target prediction model and the physical inversion module are optimized in the second fine-tuning stage.

[0168] After multiple iterative optimizations, the target prediction model is obtained. The model has learned the parameters of the general physical model based on the training set, physical knowledge, and physical constraints, and has a certain predictive ability. At this time, the physical differential equations and related parameters in the target prediction model can describe the operating laws of building equipment to a certain extent.

[0169] In some embodiments, the physical inversion module is used to solve the inverse problem by substituting spatiotemporal correlation features into the physical differential equations in the target prediction model to derive the device state parameters of the building equipment that satisfy conservation laws. In the first fine-tuning phase, only the physical inversion module is optimized. For example, the first N layers of the Transformer structure are frozen and only the PDE-Ne of the physical inversion module is optimized. This allows the physical inversion module to better adapt to the physical laws and data characteristics of the target prediction model, thereby improving the accuracy and efficiency of solving the inverse problem.

[0170] Using a subset of validation data, the corresponding spatiotemporal correlation features are fed into the physical inversion module. The error loss is calculated by comparing the device state parameters output by the physical inversion module with the actual device state parameters. An optimization algorithm (such as stochastic gradient descent or the Adam optimizer) is then used to adjust the parameters in the physical inversion module, gradually reducing the error loss. During this process, the parameters of the target prediction model remain unchanged, focusing solely on optimizing the physical inversion module.

[0171] After optimizing the physical inversion module in the first fine-tuning stage, in order to further improve the performance of the entire system, the model parameters of the target prediction model and the physical inversion module are optimized simultaneously in the second fine-tuning stage. However, the gradient normalization algorithm (GradNorm) is used to constrain the gradient magnitude ratio of physical loss to contrast loss to 1:0.3, which can improve the synergy between the target prediction model and the physical inversion module and improve the system's prediction accuracy of building equipment status parameters.

[0172] For example, in the second fine-tuning phase, validation data can also be used to input spatiotemporal correlation features into the physical inversion module. This module, combined with the target prediction model, generates predictions for the device state parameters. These predictions are then compared with the actual device state parameters to calculate the error loss. An optimization algorithm is then used to simultaneously adjust the model parameters of the target prediction model and the parameters of the physical inversion module to minimize the error loss.

[0173] In the embodiments of the present application, through the first fine-tuning phase, the physical inversion module is optimized separately, which enables the module to better understand and apply the physical laws in the target prediction model, improve the accuracy and speed of solving the inverse problem, and help more accurately derive the equipment state parameters of the building equipment. The second fine-tuning phase simultaneously optimizes the target prediction model and the physical inversion module. The target prediction model can adjust its own parameters based on the feedback from the physical inversion module, and the physical inversion module can also better adapt to changes in the target prediction model. After two stages of fine-tuning, the prediction accuracy of the building equipment state parameters is significantly improved.

[0174] In one possible implementation, if there are multiple target operation control strategies, the control device is further configured to:

[0175] After controlling the building equipment to optimize the operating mode based on multiple target operation control strategies, multiple operating results obtained by the building equipment running the multiple optimized operating modes are obtained; based on the multiple operating results, the multiple target operation control strategies are sorted; based on the sorting results, a dynamic memory library is constructed for a predetermined proportion of the target operation control strategies among the multiple target operation control strategies; after each predetermined time interval, knowledge distillation processing is performed on the dynamic memory library to obtain the latest dynamic memory library.

[0176] The target prediction model adjusts the model parameters of the target prediction model by learning the latest data in the dynamic memory library.

[0177] In some embodiments, after controlling building equipment to optimize operating modes based on multiple objective operation control strategies, the control device collects multiple operating results generated by the building equipment under the multiple optimized operating modes. These operating results may include various indicators, such as building equipment energy consumption data, operating efficiency, equipment wear, and indoor environmental parameters (e.g., temperature, humidity, and air quality). The control device may collect these operating result data using various sensors deployed in the building equipment and the environment.

[0178] Based on the multiple operational results obtained, the control device ranks multiple target operational control strategies. The ranking criteria can also be determined based on specific application requirements and goals. For example, if the application goal of a building device is to achieve energy conservation, the operational control strategies are ranked from smallest to largest based on energy consumption, with strategies with lower energy consumption ranking higher. If the application goal of a building device is to improve operational efficiency, the strategies are ranked from highest to lowest based on operational efficiency.

[0179] In some embodiments, the control device selects a predetermined proportion of target operation control strategies from the plurality of operation control strategies based on the ranking result to construct a dynamic memory library. For example, if the predetermined proportion is 20%, the top 20% of the target operation control strategies are selected.

[0180] In some embodiments, the control device performs knowledge distillation on the dynamic memory library at predetermined intervals (e.g., one week or one month). It should be understood that knowledge distillation primarily involves screening, integrating, and refining the data in the dynamic memory library, removing redundant, erroneous, or outdated data, and merging and summarizing similar operational control strategies and operational results. Thus, the target prediction model learns from the latest data in the dynamic memory library, undergoes further training and optimization, and learns the relationship between the optimal target operational control strategy and the corresponding operational results, thereby adjusting its own model parameters.

[0181] By using the embodiments of this application, by evaluating and ranking the operational results of multiple target operation control strategies, optimal operation control strategies can be identified and stored in a dynamic memory library, facilitating continuous improvement of the operational modes of building equipment, enhancing their operational efficiency and performance. Furthermore, knowledge distillation effectively manages and optimizes the data in the dynamic memory library, removing useless information while retaining valuable data, enabling the target prediction model to more efficiently learn key knowledge.

[0182] As an optional embodiment, the control device integrates a graph attention network (GAT) based on a spatial attention mechanism, a single convolution layer based on causality (CausalConv1 D), and a spatiotemporal encoder. Specifically, the control device uses GAT to process multimodal data, constructing a topological relationship diagram between devices by analyzing the spatial position relationship, connection dependencies, and other information between building devices, thereby determining the spatial characteristics of the building devices. The control device also uses CausalConv1 D to analyze the time series in the multimodal data and extract dynamic characteristics over time, such as energy consumption fluctuations and device status changes. The control device also uses a spatiotemporal encoder to fuse the spatial and temporal dynamic features output by GAT and CausalConv1 D, and removes redundant information through an encoding layer (such as a fully connected layer or a lightweight Transformer structure), retaining key information such as cross-device collaboration patterns, periodic patterns (such as energy consumption changes during daily peak and trough periods), sudden changes (such as parameter mutations during sudden equipment failures), and abnormal event correlation characteristics, ultimately determining the spatiotemporal correlation characteristics of the building devices.

[0183] Furthermore, control devices use a pre-trained target prediction model to derive equipment status parameters for building equipment. This target prediction model is based on a general physics model consisting of multiple differential equations, encompassing physical laws such as thermodynamics and fluid mechanics, and incorporating domain constraints. It utilizes a Transformer architecture to support multimodal data fusion and incorporates a physical field inversion module as a post-processing layer.

[0184] During the derivation process, the physics inversion module substitutes the spatiotemporal correlation characteristics into the differential equations of the target prediction model. By solving the inverse problem, it automatically derives the device state parameters that satisfy conservation laws. For example, for air conditioning equipment, parameters such as the actual compressor efficiency and refrigerant flow rate can be derived; for elevator equipment, motor load conditions and operating wear can be calculated without relying on extensive historical data or precise modeling.

[0185] Based on the derived building equipment state parameters, the control device automatically generates multiple operational control strategies by combining reinforcement learning algorithms with physical field inversion technology, combining multiple objectives such as energy saving, comfort, and equipment life. The specific process may include, but is not limited to: environmental modeling (using physical field inversion to construct a three-dimensional thermodynamic digital twin to simulate the equipment operating environment), strategy space construction (generating corresponding control strategies for different equipment types, such as variable air volume curves for air conditioning, time-sharing scheduling strategy clusters for elevators, and illumination gradient matrices for lighting), multi-objective rewards (real-time calculations through a physical field solver, and fusing indicators such as comfort PMV with inversion data), and strategy distillation (using physical fields to verify feasible strategies, train a strategy pool, and retain the optimal strategy).

[0186] The control device then uses the CFD constraint solver in the physical field inversion technology to perform transient simulation processing (for example, 0.5-second transient simulation processing) on ​​the multiple generated operation control strategies, eliminating operation control strategies that violate the laws of thermodynamics, such as intercepting abnormal strategies where temperature changes do not conform to physical laws. Ultimately, the target operation control strategy is selected from the multiple operation control strategies, and the optimization instructions corresponding to the operation control strategy are sent to the building equipment through the adaptive protocol converter. The control building equipment automatically adjusts its operating mode based on the target operation control strategy to achieve energy saving, consumption reduction and efficient operation.

[0187] It should be understood that in the embodiments of the present application, the control device uses physical field inversion technology and target prediction models to adjust the target operation control strategy of building equipment in real time based on physical laws and the actual state of the equipment, thus breaking away from the limitations of manual experience. For example, the control device adjusts the cooling capacity of air-conditioning equipment and the lighting brightness of lighting systems according to real-time data such as crowd density and ambient temperature. Compared with traditional methods, this can significantly reduce the energy consumption of building equipment such as air-conditioning equipment and lighting systems (for example, a reduction of 30%-40%), achieving the energy-saving operation effect of adaptive optimization of building equipment.

[0188] In the embodiments of this application, the control device is based on physical field inversion technology, using real-time multimodal data and physical equations from building equipment to solve the inverse problem. This allows the inference of the equipment status parameters of the building equipment without requiring a large amount of historical data, thus resolving data dependency issues and achieving a cold start. Simultaneously, the device detection terminal is deployed as an edge computing node on the building equipment side, enabling rapid local processing of some data on the building equipment, reducing data transmission latency, and collaborating with the cloud-based control device to achieve low-latency (response latency ≤ 120ms) and high-precision joint optimization, enabling the building automation system to be rapidly deployed and accurately control the operating mode of the building equipment.

[0189] This application example provides an example of a method for controlling the operation of a building equipment. Please refer to Figure 3 As shown, Figure 3 The following is a schematic flow chart of a method for controlling the operation of building equipment provided by the present application. As an example and not a limitation, the method can be applied to or run in an operation control system of building equipment. The method includes:

[0190] S301: Receive multimodal data from building equipment.

[0191] S302: Determine spatiotemporal correlation characteristics of building equipment based on multimodal data.

[0192] S303: Using the pre-trained target prediction model, derive the equipment status parameters of the building equipment according to the spatiotemporal correlation features.

[0193] S304: Determine a corresponding target operation control strategy according to the device status parameters of the building equipment, so as to control the building equipment to optimize the operation mode based on the target operation control strategy.

[0194] In some embodiments, the execution subject of the building equipment operation control method is a control device in the building equipment operation control system.

[0195] Through the embodiments of the present application, a control device is used to receive multimodal data of building equipment. The multimodal data can be collected by the equipment detection terminal deployed by the building equipment. Then, the building equipment determines the spatiotemporal correlation characteristics of the building equipment based on the multimodal data; and uses a pre-trained target prediction model to deduce the equipment status parameters of the building equipment according to the spatiotemporal correlation characteristics. The equipment status parameters of the building equipment can be determined without precise modeling and a large amount of historical data participating in the calculation, and are no longer restricted by manual experience and fixed rules. Furthermore, the control device can also determine the corresponding target operation control strategy based on the derived equipment status parameters of the building equipment to control the building equipment to dynamically optimize the operation mode based on the target operation control strategy. Without relying on manual experience and fixed rules, the operation effect of the building equipment can be adaptively optimized and the energy waste of the building equipment can be reduced.

[0196] It should be noted that the above-mentioned embodiments of the building equipment operation control method embodiment and any implementation thereof respectively correspond to the above-mentioned building equipment operation control system embodiment and any implementation thereof. The technical effects corresponding to the above-mentioned building equipment operation control method embodiment and any implementation thereof can be found in the above-mentioned building equipment operation control system embodiment and any implementation thereof, and will not be further elaborated here.

[0197] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0198] An embodiment of the present application further provides a control device, the control device comprising one or more processors and a memory;

[0199] The memory is coupled to one or more processors, and is used to store computer program code. The computer program code includes computer instructions. The one or more processors call the computer instructions to enable the control device to execute the operation control method of the building equipment shown above.

[0200] Figure 4 This is a schematic diagram of the structure of a control device provided in an embodiment of the present application. The control device 400 can be a mobile phone, a smart screen, a tablet computer, a wearable control device, an in-vehicle control device, an augmented reality (AR) device, a virtual reality (VR) device, a laptop computer, an ultra-mobile personal computer (UMPC), a netbook, a personal digital assistant (PDA), a projector, or a communication device such as a server, storage device, or a base station, or a smart car. The embodiments of the present application do not impose any restrictions on the specific type of control device.

[0201] The memory 401 can be used to store computer software programs 402 and modules. The processor 403 executes various functional applications and data processing of the control device by running the software programs and modules stored in the memory 401. The memory 401 can mainly include a program storage area and a data storage area. The program storage area can store an operating system, at least one application required for a function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area can store data created based on the use of the control device (such as audio data, a phone book, etc.). In addition, the memory 401 can include a high-speed random access memory and a non-volatile memory, such as at least one disk storage device, a flash memory device, or other volatile solid-state storage device.

[0202] Among them, the processor 403 may include one or more processors such as a central processing unit, an application processor (AP), and a baseband processor. The processor can be the nerve center and command center of the wireless router. The processor 403 can generate operation control signals based on the instruction operation code and timing signals to complete the control of instruction fetching and execution. The memory 401 can be used to store computer executable program code, and the executable program code includes instructions. The processor 403 executes various functional applications and data processing of the network device by running the instructions stored in the memory. The memory 401 may include a program storage area and a data storage area, such as storing data of a sound signal to be played. For example, the memory can be a double data rate synchronous dynamic random access memory DDR or a flash memory.

[0203] An embodiment of the present application further provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer instructions; when the computer-readable storage medium is run on a control device, the control device executes the operation control method of the building equipment shown above.

[0204] An embodiment of the present application further provides a computer program product comprising computer instructions. When the computer program product is run on a control device, the control device can execute the aforementioned operation control method for building equipment.

[0205] The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a magnetic tape), an optical medium, or a semiconductor medium (e.g., a solid-state disk (SSD)).

[0206] The computer storage medium and computer program product provided in the above-mentioned embodiments of the present application are used to execute the method provided above. Therefore, the beneficial effects that can be achieved can refer to the corresponding beneficial effects of the method provided above, and will not be repeated here.

[0207] In the above embodiments, it can also be implemented in whole or in part by software, hardware, firmware or any combination thereof. When software is used for implementation, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from a website, computer, server or data center to another website, computer, server or data center via a wired (such as a coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) method. The computer-readable storage medium can be any available medium that a computer can access, or a data storage device such as a server or data center that includes one or more available media integrations. The storage medium may be a magnetic disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), a flash memory (Flash Memory), a hard disk (HDD) or a solid-state drive (SSD), etc. The storage medium may also include a combination of the above types of memory.

[0208] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0209] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments applied for herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0210] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present application, and should all be included in the scope of protection of the present application.

Claims

1. An operation control system for building equipment, characterized in that: include: building equipment and control equipment; The building equipment is deployed with an equipment detection terminal, and the equipment detection terminal is used to collect multimodal data of the building equipment; The control device is connected to the building equipment and the equipment detection terminal and is configured to perform: receiving multimodal data of the building equipment; determining spatiotemporal correlation characteristics of the building equipment based on the multimodal data; Using a pre-trained target prediction model, deriving device status parameters of the building equipment according to the spatiotemporal correlation features; as well as A corresponding target operation control strategy is determined according to the device status parameters of the building equipment to control the building equipment to optimize the operation mode based on the target operation control strategy.

2. The system according to claim 1, wherein: The system further comprises: A multi-protocol gateway is connected to the device detection terminal and the control device, and is used to receive the multimodal data transmitted by the device detection terminal and forward the multimodal data to the control device.

3. The system according to claim 1, wherein: The control device is further configured to perform: Automatically generate multiple operation control strategies based on the equipment status parameters of the building equipment; Performing transient simulation processing on the plurality of operation control strategies through physical field inversion technology to obtain simulation results corresponding to the plurality of operation control strategies; selecting a target operation control strategy from the plurality of operation control strategies according to simulation results corresponding to the plurality of operation control strategies; as well as The target operation control strategy is sent to the building equipment.

4. The system according to claim 1, wherein: The control device is deployed with a graph attention network based on a spatial attention mechanism, a layer of convolution based on causality, and a spatiotemporal encoder; wherein: The graph attention network based on the spatial attention mechanism is used to determine the spatial features of the building equipment according to the multimodal data, wherein the spatial features are used to characterize the spatial topological relationship between various types of equipment included in the building equipment; The causal relationship-based convolution layer is used to determine the time series dynamic characteristics of the building equipment based on the multimodal data, wherein the time series dynamic characteristics are used to characterize the dynamic changes of energy consumption fluctuations and equipment status changes of the building equipment over time; The spatiotemporal encoder is used to determine the spatiotemporal correlation characteristics of the building equipment based on the spatial characteristics and the temporal dynamic characteristics.

5. The system according to claim 4, characterized in that The multimodal data is collected in real time or at a predetermined frequency; The spatiotemporal encoder is further configured to compress the spatiotemporal correlation features after determining the spatiotemporal correlation features of the building equipment; The compression processing is used to remove redundant information in the spatiotemporal correlation features, and retain at least one of the cross-device collaboration pattern, periodicity, mutation and abnormal event correlation features in the spatiotemporal correlation features.

6. The system according to claim 1, wherein: The control device is also used to: Pre-acquiring a plurality of sample multimodal data of sample building equipment and sample equipment status parameters corresponding to each of the plurality of sample multimodal data, wherein each of the sample multimodal data includes at least one of temperature, humidity, light, crowd density, current, voltage, and power; generating a training set based on the plurality of sample multimodal data and sample device state parameters corresponding to each of the plurality of sample multimodal data; Adding physical knowledge and physical constraints to a general physical model, wherein the physical knowledge includes at least a plurality of physical laws; Under the guidance of the physical constraints, the model parameters of the general physical model are trained based on the training set, and the target prediction model is obtained after multiple iterative optimizations.

7. The system according to claim 6, characterized in that The control device is also used to: Performing confidence assessment on the plurality of sample multimodal data by calculating physical rationality scores of sample device state parameters corresponding to each of the plurality of sample multimodal data; According to the confidence evaluation result of each sample multimodal data, a momentum accumulation mechanism is adopted to update the weight of each sample multimodal data; Selecting a preset number of target sample multimodal data from the plurality of sample multimodal data according to the weight of each sample multimodal data; The preset number of target sample multimodal data and the sample device state parameters corresponding to the target sample multimodal data are used as the training set.

8. The system according to claim 6, wherein: The control device is further equipped with a physical field inversion module and the target prediction model; The physical field inversion module is used to solve the inverse problem by substituting the spatiotemporal correlation characteristics into the physical differential equations in the target prediction model, so as to derive the equipment state parameters of the building equipment that satisfy the conservation law.

9. The system according to any one of claims 1 to 8, characterized in that If there are multiple target operation control strategies, the control device is further configured to: After controlling the building equipment to optimize the operation mode based on the multiple target operation control strategies, obtaining multiple operation results obtained by the building equipment running the multiple optimized operation modes; sorting the target operation control strategies according to the plurality of operation results; Building a dynamic memory library for a predetermined proportion of the target operation control strategies among the plurality of target operation control strategies according to the sorting result; After each predetermined time interval, knowledge distillation is performed on the dynamic memory library to obtain the latest dynamic memory library, wherein the target prediction model adjusts the model parameters of the target prediction model by learning the data in the latest dynamic memory library.

10. A method for controlling the operation of building equipment, characterized in that: include: Receive multimodal data from building equipment; determining spatiotemporal correlation characteristics of the building equipment based on the multimodal data; Using a pre-trained target prediction model, deriving device status parameters of the building equipment according to the spatiotemporal correlation features; A corresponding target operation control strategy is determined according to the device status parameters of the building equipment to control the building equipment to optimize the operation mode based on the target operation control strategy.

Images