An application method and system of combining a DNN and an LSTM model in an intelligent building control system

Abstract

The application discloses a kind of application method and system of combining DNN and LSTM model in intelligent building control system, it is related to intelligent building control and artificial intelligence fusion technical field, including based on BIM model construction space digital twin system, extraction building physical structure and equipment layout relationship, using multidimensional input information, construct space static parameter coding for intelligent control;Build IoT middle station driven multi-source heterogeneous perception channel, fusion dynamic environment information and space static parameter coding, using fixed time interval and unified time stamp mechanism, construct multi-dimensional time series input;Build LSTM, Attention and DNN heterogeneous fusion multilayer neural inference network, fusion time series modeling and feature mapping capability, build building information and multi-device linkage data closed loop architecture;The method described in the application realizes the fusion modeling of space structure and time series data, improves the precision and self-adaptive ability of multi-device linkage control of intelligent building.

Description

Technical Field

[0001] This invention relates to the field of intelligent building control and artificial intelligence integration technology, specifically to a method and system for applying DNN and LSTM models in intelligent building control systems. Background Technology

[0002] With the deepening of the concept of smart buildings and the continuous improvement of high-quality living standards for "good houses," building environment control systems based on information technology, sensing systems, control systems, and artificial intelligence are gradually becoming one of the infrastructures of intelligent buildings. Especially in the context of the interconnection of multiple subsystems such as building air conditioning, fresh air, lighting, shading, and underfloor heating, how to achieve refined, interconnected, and predictive automatic control has become a key issue in the current development of smart buildings. However, current intelligent building control systems still face significant technical shortcomings in structural modeling, data fusion, feature extraction, and dynamic decision-making.

[0003] First, existing technologies generally lack structured representation of architectural spatial information. Although some projects introduce BIM models for auxiliary design and collaborative management during the construction phase, these BIM models often fail to effectively link with IoT control systems during the building operation phase, failing to establish a real-time response relationship between architectural spatial structure and environmental control. Particularly regarding static information such as building equipment layout, spatial attributes, and orientation, there is a lack of quantifiable coding mechanisms to support algorithm model input. This makes it difficult for AI control systems to understand the contextual structure between spatial layout and equipment interaction.

[0004] Secondly, while a large number of sensors have been deployed at the IoT data level, inconsistent data upload schedules, missing time stamps, and non-standard device coding make it difficult to effectively integrate multi-source heterogeneous data. Most building IoT platforms still use single-point collection and flat display methods for data integration, lacking system-level time-series construction logic, making it difficult to form a high-dimensional time-series input structure for intelligent algorithms. At the same time, most mainstream air conditioning, ventilation, and lighting control systems on the market rely on a "single variable + current state" control strategy, that is, adjusting only based on a certain indicator (such as temperature), failing to consider the changing trends and interactions of indoor temperature, humidity, light, and other indicators within a certain time range, and ignoring the cumulative effect of user comfort and delayed response during long-term user experience.

[0005] Secondly, at the level of AI algorithm integration, although some building control systems have introduced models such as Deep Neural Networks (DNNs) or Long Short-Term Memory Networks (LSTMs) for single-point prediction, two core problems remain: limited ability to extract features from multi-dimensional environmental data, especially lacking scalable input mechanisms for the fusion and encoding of different types of data; and time-series data is only processed using simple sliding windows, failing to deeply explore the delayed impact of historical states on current decisions. In this context, relying solely on a single model is insufficient to fully reflect the dynamic scenarios, complex relationships, and multi-dimensional dependencies required for building control, resulting in limited intelligence and control strategies often exhibiting problems such as delayed response, inappropriate linkage, and significant differences in user perception. For example, in traditional air conditioning systems, there is often a distributed inconsistency in user perception, with some people feeling hot while others feel cold; this is a typical manifestation of the lack of time-scale judgment and spatial awareness encoding.

[0006] Furthermore, judging from existing technical literature and commercial products, the application of composite neural networks integrating DNN and LSTM structures in intelligent building control scenarios is still in its early exploratory stage, lacking systematic supporting mechanisms such as data input modeling, spatial information structured design, time series consistency construction, and equipment command mapping. Therefore, how to form a closed-loop AI control process from spatial structure modeling and data perception fusion to intelligent control decision-making remains a key challenge in the current research and practical deployment of smart building technologies. Summary of the Invention

[0007] In view of the above-mentioned problems, the present invention is proposed.

[0008] Therefore, the technical problem solved by this invention is that existing intelligent building control methods suffer from problems such as difficulty in structured input of spatial structural information, asynchronous and scattered environmental perception data, weak temporal modeling capability of control models, and low coupling degree of multi-system linkage, resulting in low control accuracy, untimely response, and coarse granularity of comfort control. The invention also addresses the problem of how to achieve spatial semantic modeling, multi-source perception fusion, temporal feature learning, and closed-loop coordination of equipment control within a unified framework.

[0009] To address the aforementioned technical problems, this invention provides the following technical solution: a method for applying DNN and LSTM models in intelligent building control systems, comprising: constructing a spatial digital twin system based on a BIM model; extracting the physical structure of the building and the relationship between equipment layout; employing multi-dimensional input information to construct spatial static parameter encoding for intelligent control; constructing a multi-source heterogeneous sensing channel driven by an IoT platform; fusing dynamic environmental information and spatial static parameter encoding; employing a fixed time interval and a unified timestamp mechanism to construct a multi-dimensional time series input; constructing a multi-layer neural inference network that heterogeneously integrates LSTM, Attention, and DNN; fusing temporal modeling and feature mapping capabilities; and constructing a data closed-loop architecture for building information and multi-device linkage; employing a fixed time interval and a unified timestamp mechanism includes setting a standard data acquisition cycle during the uploading of IoT sensing data and generating aligned time labels for multiple types of sensors; the data closed-loop architecture includes integrating the spatial dimension vectors provided by the BIM model, the real-time sensing data collected by the IoT, and the inference results of the AI ​​module, and feeding them back to the spatial intelligent devices through an intelligent control interface.

[0010] As a preferred embodiment of the application method of combining DNN and LSTM models in intelligent building control systems described in this invention, the construction of a spatial digital twin system based on a BIM model includes: constructing a 1:1 replica of a three-dimensional as-built model through the Revit modeling platform, extracting spatial structural information, and embedding intelligent devices installed in the building into corresponding spatial units in the building model according to their positions.

[0011] As a preferred embodiment of the application method of combining DNN and LSTM models in intelligent building control systems described in this invention, the construction of spatial static parameter encoding for intelligent control includes normalizing and encoding the spatial structural elements extracted from the three-dimensional as-built model as spatial dimension vectors for the input layer of the neural network.

[0012] As a preferred embodiment of the application method of combining DNN and LSTM models in intelligent building control systems as described in this invention, the spatial dimension vector serving as the input layer of the neural network includes spatial static parameter encoding and sensing data establishing a correspondence through a spatial index binding mechanism, synchronously participating in the construction of sequence input.

[0013] As a preferred embodiment of the application method of combining DNN and LSTM models in intelligent building control systems described in this invention, the construction of a multi-source heterogeneous sensing channel driven by an IoT platform includes integrating multiple types of sensor devices deployed in the spatial area through a smart building control platform, setting a unified transmission cycle, and centrally aggregating and scheduling real-time sensing data.

[0014] As a preferred embodiment of the application method of combining DNN and LSTM models in intelligent building control systems described in this invention, the construction of multi-dimensional time series input includes attaching time labels to multiple types of sensor data and generating a standardized time series tensor structure by combining device number and sampling step size.

[0015] As a preferred embodiment of the method for applying the combination of DNN and LSTM models in intelligent building control systems as described in this invention, the construction of a multi-layer neural inference network that heterogeneously integrates LSTM, Attention and DNN includes: inputting a temporal tensor into an LSTM to extract hidden states, generating temporal features through multi-head attention weighting, and inputting the temporal features and spatial static parameters into a DNN to generate a control vector.

[0016] As a preferred embodiment of the application method of combining DNN and LSTM models in intelligent building control systems as described in this invention, the fusion of temporal modeling and feature mapping capabilities includes: standardizing the encoding format of temporal data and spatial static parameters, constructing a spliced ​​input structure, and feeding the data into LSTM and DNN through preset channels for joint modeling.

[0017] As a preferred embodiment of the application method of combining DNN and LSTM models in intelligent building control system described in this invention, the construction of a data closed-loop architecture for building information and multi-device linkage includes mapping the DNN output to device control commands, transmitting them to the corresponding devices based on the communication interface, and feeding back the status feedback to the perception layer.

[0018] Another objective of this invention is to provide an application system combining DNN and LSTM models in intelligent building control systems. This system can solve the problems of poor coupling and insufficient real-time performance in current intelligent building control systems, such as the lack of collaborative modeling of spatial structure and sensor data, weak temporal response capability, and isolated distribution of equipment control commands. This is achieved by constructing a control flow scheme that takes spatial static coding and IoT sensing data as input and integrates LSTM, Attention mechanism and DNN inference structure.

[0019] As a preferred embodiment of the application system combining DNN and LSTM models in intelligent building control as described in this invention, it includes: a spatial modeling and static parameter encoding module, a multi-source perception fusion and time series construction module, and a neural inference and equipment linkage control module; the spatial modeling and static parameter encoding module is used to construct a digital twin system through a BIM model, extract building structure and equipment layout information, standardize and encode spatial attributes, and generate static input fields adapted to the neural network; the multi-source perception fusion and time series construction module is used to collect data from multiple types of environmental sensors, complete multi-source alignment through a unified timestamp mechanism, and fuse static encoding to generate standardized multi-dimensional time series input; the neural inference and equipment linkage control module is used to fuse spatial static parameter encoding and time series data input into a neural network structure, complete feature modeling and control command generation, and link intelligent equipment to dynamically respond.

[0020] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement a method for combining DNN and LSTM models in an intelligent building control system.

[0021] A computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the steps of a method for applying a combination of DNN and LSTM models in an intelligent building control system.

[0022] The beneficial effects of this invention are as follows: The method for applying DNN and LSTM models in intelligent building control systems provided by this invention constructs a spatial twin system through BIM, realizing structured input of equipment and structural data and enhancing the spatial perception capability of the control model; it constructs a perception channel through a unified sampling mechanism and time labels, realizing temporal alignment and standardized input of heterogeneous data; and it integrates LSTM, Attention and DNN structures to achieve collaborative modeling of spatial and temporal features, improving the accuracy of multi-device linkage control. This invention achieves better results in semantic modeling of building spatial structure, standardized processing of multi-source heterogeneous perception data, and spatiotemporal feature fusion reasoning for linkage control. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1The first embodiment of the present invention provides an overall flowchart of a method for applying DNN and LSTM models in an intelligent building control system.

[0025] Figure 2 This is a schematic diagram of an application system combining DNN and LSTM models in an intelligent building control system, provided as a third embodiment of the present invention. Detailed Implementation

[0026] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0027] Example 1, referring to Figure 1 As an embodiment of the present invention, a method for applying a combination of DNN and LSTM models in an intelligent building control system is provided, comprising:

[0028] S1: Construct a spatial digital twin system based on the BIM model, extract the relationship between the building's physical structure and equipment layout, and construct spatial static parameter coding for intelligent control by using multi-dimensional input information.

[0029] Furthermore, building a spatial digital twin system based on BIM models includes constructing a 1:1 replica of a 3D as-built model using the Revit modeling platform, extracting spatial structural information, and embedding smart devices installed in the building into corresponding spatial units of the building model according to their locations.

[0030] Furthermore, the use of multi-dimensional input information includes integrating IoT information from various sensors and devices through the smart building management platform. For example, through wired and wireless signal transmission, the total number of operating devices, the number of deployed devices, and the number of faults for specific equipment types such as regional fire water tanks, air quality, lighting, and air conditioning are statistically analyzed. For instance, 514 lighting devices are in operation with zero faults, indicating that the platform has good monitoring coverage in spatial lighting control. This demonstrates the real-time monitoring capability of the smart building control system for the status of various equipment in the building and the actual manifestation of multi-dimensional data input. Sensor information is transmitted to the platform to realize the storage and recording of multi-dimensional information (temperature, humidity, wind speed, light intensity) and time series data, which are then correlated with the building space structure and equipment operation data to form a complete time series tensor.

[0031] Furthermore, room units and smart devices in the Revit model have associated attributes, and the smart device deployment points have corresponding identifiers in the BIM model. By mapping through spatial geometric centers, boundary conditions, and device coordinates, a correspondence between "spatial nodes and device entities" is established, forming a spatial semantic map for the control system.

[0032] It should be noted that constructing spatial static parameter encoding for intelligent control includes normalizing and encoding the spatial structural elements extracted from the 3D as-built model as fields, and using them as spatial dimension vectors for the input layer of the neural network.

[0033] It should be noted that field normalization and feature encoding include converting spatial area, orientation, and floor height attributes of different dimensions into a standard form ranging from 0 to 1, constructing a vector matrix with a unified structure, and setting each space as a spatial feature vector with a fixed dimension, which facilitates subsequent splicing with dynamic data and input into the neural network.

[0034] It should also be noted that the spatial dimension vector used as the input layer of the neural network includes spatial static parameter encoding and perceptual data, which establish a correspondence through a spatial index binding mechanism and participate synchronously in the construction of sequence input.

[0035] S2: Construct a multi-source heterogeneous sensing channel driven by an IoT middle platform, integrate dynamic environmental information and spatial static parameter encoding, and construct a multi-dimensional time series input by adopting a fixed time interval and a unified timestamp mechanism.

[0036] Furthermore, the construction of a multi-source heterogeneous sensing channel driven by an IoT platform includes integrating various types of sensor devices deployed in the spatial area through a smart building control platform, setting a unified transmission cycle, and centrally aggregating and scheduling real-time sensing data.

[0037] Furthermore, setting a unified transmission cycle includes setting the time period length to the data monitored within 1-2 hours, with a data interval of 10 minutes, that is, 6-12 time series data. For the temperature intelligent control system, the parameter range includes the real-time temperature, real-time humidity, and real-time light intensity monitored for the previous 1-2 hours, as well as the size and orientation of the room where the intelligent device is located.

[0038] It should be noted that constructing multidimensional time series inputs includes attaching time labels to multiple types of sensor data and generating a standardized time series tensor structure by combining device numbers and sampling step sizes.

[0039] It should also be noted that by constructing a multi-source heterogeneous sensing channel driven by an IoT middle platform, unified access and time synchronization processing of multiple types of environmental sensors can be achieved, generating a multi-dimensional tensor structure with temporal and spatial identification. Combined with spatial static parameter encoding, a unified input vector is constructed, providing a structured spatiotemporal fusion data foundation for subsequent neural networks, forming an input closed-loop mechanism that supports linkage control.

[0040] S3: Construct a multi-layer neural inference network that integrates LSTM, Attention and DNN heterogeneously, and integrates temporal modeling and feature mapping capabilities to build a data closed-loop architecture for building information and multi-device linkage.

[0041] Furthermore, the construction of a multi-layer neural inference network that heterogeneously integrates LSTM, Attention, and DNN includes inputting a temporal tensor into an LSTM to extract hidden states, generating temporal features through multi-head attention weighting, and then concatenating the temporal features with spatial static parameters before inputting them into a DNN to generate control vectors.

[0042] Furthermore, LSTM stands for Long Short-Term Memory Neural Network. Through its gating mechanism, it can solve the problem of long-term dependencies in long-term sequences and is an effective way to extract features from time series elements. Attention represents a multi-head attention mechanism, expressed as:

[0043] ;

[0044] ;

[0045] in, The query vector represents the intelligent device information retrieval request proposed at the current time step. The linear projection weight matrix of the query vector represents the smart device query demand at the current time step during attention retrieval. For time steps The hidden state is the first state in a smart building that the LSTM network represents. The output after modeling the perceived data sequence of each time slice represents the comprehensive semantic representation of that time node in the temporal features; The key vector represents the identity descriptor of this time step in the attention mechanism; Let be the linear projection weight matrix of the key vectors; This is a value vector, representing the information carrier of the intelligent device that is passed to the final representation at this time step; The linear projection weight matrix of the value vector; For each time step The attention weight represents the importance of that time step to the overall input sequence; The normalization function transforms the matching scores at all time steps into a probability distribution. This is the transpose matrix of the key vectors for all time steps; The scaling factor is used; the LSTM output is weighted at each time step using multi-head attention, which is represented as:

[0046] ;

[0047] in, The output of the multi-head attention mechanism represents the attention result after weighted summation over all heads. Indicates the number of attention heads. For the concatenation operation function, the weighted representation is calculated as follows:

[0048] ;

[0049] in, The temporal feature output for a single attention head. For the time range of smart devices that need weighting. For a certain time step, For each time step The attention weight represents the importance of that time step to the overall input sequence. Let be a value vector, representing the information carrier of the intelligent device passed to the final representation at this time step. DNN represents a deep neural network, represented as:

[0050] ;

[0051] in, To generate the final control command from the intermediate state obtained after fusing spatial static parameter encoding and time-weighted features through multiple mappings, For the first The weight matrix of a layer represents the weights of the output from the previous layer. Connection strength to the current layer The output of the previous DNN layer is fused with attention features. The input is concatenated with spatial static encoding. For the first The layer bias vector is a constant term that matches the number of neurons in the current layer. The activation function is used to introduce nonlinearity. Through nonlinear calculation, it can be expressed as:

[0052] ;

[0053] in, To predict the current control status of the smart building, The weight matrix of the output layer maps the feature vectors of the last layer of the neural network to the control instruction space. This is the output feature vector of the last layer of the neural network, which is the feature extraction result of the preceding modules LSTM→Attention→DNN. The output layer bias term is a constant offset corresponding to the control output dimension, providing a mapping from input features to output features.

[0054] It should be noted that the integration of temporal modeling and feature mapping capabilities includes standardizing the encoding format of temporal data and spatial static parameters, constructing a concatenated input structure, and feeding the data into LSTM and DNN via preset channels for joint modeling, as shown below:

[0055] ;

[0056] in, To predict the current control status of the smart building, The weight matrix of the output layer maps the feature vectors of the last layer of the neural network to the control instruction space. To introduce a multi-head attention mechanism onto the hidden state sequence, this mechanism dynamically and selectively represents temporal information by weighted focusing on key time points. Modeling time series data, extracting long-term dependencies and evolutionary trends, and generating a hidden state sequence for each time step. The output layer bias term is a constant offset that corresponds to the control output dimension.

[0057] It should be noted that the model input uses a tensor quantization interface, the LSTM structure processes the temporal feature dimension sequence, and the DNN structure processes the spatial feature vector. During training, both networks participate in backpropagation simultaneously, and the weights are optimized through a joint loss function, expressed as:

[0058] ;

[0059] in, This represents the average error of the current batch of samples in predicting the control quantity. Indicates the number of samples. For the first The predicted output for each sample. For the first The true target values ​​corresponding to each sample form a fusion learning structure.

[0060] It should also be noted that building a data closed-loop architecture for building information and multi-device linkage includes mapping the DNN output to device control commands, transmitting them to the corresponding devices via communication interfaces, and feeding back the status feedback to the perception layer.

[0061] Example 2, refer to Figure 2 As an embodiment of the present invention, an application system combining DNN and LSTM models in an intelligent building control system is provided, including a spatial modeling and static parameter encoding module 100, a multi-source perception fusion and time series construction module 200, and a neural reasoning and equipment linkage control module 300.

[0062] X1: Spatial modeling and static parameter encoding module 100 includes spatial digital twin modeling submodule 101 and spatial semantic feature encoding submodule 102.

[0063] It should be noted that the spatial digital twin modeling submodule 101 is used to construct a three-dimensional BIM model, extract room structure and equipment layout information, establish the correspondence between spatial units and equipment entities, and form a spatial semantic map that can be used for calculation; the spatial semantic feature encoding submodule 102 is used to normalize spatial attributes into fixed-length vectors, bind equipment through spatial indexes, realize the vectorization of spatial static parameters, and construct a neural network spatial input channel.

[0064] It should also be noted that the spatial digital twin modeling submodule 101 provides spatial objects, and the spatial semantic feature encoding submodule 102 completes semantic abstraction, forming a closed-loop process from spatial geometry to network-usable vectors, and transmits it to the multi-source perception fusion and time series construction module 200.

[0065] X2: The multi-source sensing fusion and time series construction module 200 includes the heterogeneous sensing access and time specification sub-module 201 and the spatiotemporal tensor construction and index binding sub-module 202.

[0066] It should be noted that the heterogeneous sensor access and time standardization submodule 201 is used to uniformly schedule multiple types of sensors, set the sampling period and timestamp, and generate standardized input segments; the spatiotemporal tensor construction and index binding submodule 202 is used to construct the sensing data tensor and splice it with the spatial static vector to complete the construction of the multidimensional input format and realize spatiotemporal integrated modeling.

[0067] It should also be noted that the heterogeneous sensing access and time standardization submodule 201 provides dynamic input with time alignment, and the spatiotemporal tensor construction and index binding submodule 202 realizes spatial injection and structural fusion, completing the input path from sensing data to network tensor, and providing a structurally complete input tensor for the neural reasoning and device linkage control module 300.

[0068] X3: The neural reasoning and device linkage control module 300 includes an LSTM temporal modeling submodule 301, an attention feature aggregation submodule 302, and a DNN decision mapping and control generation submodule 303.

[0069] It should be noted that the LSTM temporal modeling submodule 301 is used to model the input temporal tensor, extract the dependencies between time steps, and form a hidden state sequence; the attention feature aggregation submodule 302 is used to calculate the temporal attention weights using a multi-head attention mechanism, identify key time periods, and improve the model's response to sudden events or high-variability intervals; the DNN decision mapping and control generation submodule 303 is used to concatenate the temporal features with static encoding and input them into the DNN network to generate multi-dimensional control vectors and distribute them in real time, feeding back the device status to the perception layer to form a closed loop.

[0070] It should also be noted that the neural reasoning and device linkage control module 300 relies on the spatial structure semantics provided by the spatial modeling and static parameter encoding module 100 and the temporal dynamic features provided by the multi-source perception fusion and time series construction module 200. The three work together to complete the closed-loop task from spatial recognition, environmental perception to control decision-making.

Claims

1. A method for applying DNN and LSTM models in intelligent building control systems, characterized in that, include: Based on the BIM model, a spatial digital twin system is constructed, the relationship between the building's physical structure and equipment layout is extracted, and multi-dimensional input information is used to construct spatial static parameter coding for intelligent control. Construct a multi-source heterogeneous sensing channel driven by an IoT middleware, integrate dynamic environmental information and spatial static parameter encoding, and construct a multi-dimensional time series input by adopting a fixed time interval and a unified timestamp mechanism. Construct a multi-layer neural inference network that integrates LSTM, Attention and DNN heterogeneously, and integrate temporal modeling and feature mapping capabilities to build a data closed-loop architecture that links building information and multiple devices. The use of fixed time intervals and a unified timestamp mechanism includes setting a standard data acquisition cycle during the upload of IoT sensing data and generating aligned time tags for multiple types of sensors. The data closed-loop architecture integrates the spatial dimension vectors provided by the BIM model, real-time sensing data collected by the Internet of Things, and inference results from the AI ​​module, and feeds them back to the spatial intelligent devices through the intelligent control interface. The construction of a multi-layer neural inference network that heterogeneously integrates LSTM, Attention and DNN includes: inputting temporal tensors into LSTM to extract hidden states, generating temporal features through multi-head attention weighting, and concatenating the temporal features and spatial static parameters into DNN to generate control vectors; The integration of temporal modeling and feature mapping capabilities includes standardizing the encoding format of temporal data and spatial static parameters, constructing a concatenated input structure, and feeding the data into LSTM and DNN via preset channels for joint modeling, as shown below: ， in, To predict the current control status of the smart building, The weight matrix of the output layer maps the feature vectors of the last layer of the neural network to the control instruction space. To introduce a multi-head attention mechanism onto the hidden state sequence, this mechanism dynamically and selectively represents temporal information by weighted focusing on key time points. Modeling time series data, extracting long-term dependencies and evolutionary trends, and generating a hidden state sequence for each time step. The output layer bias term is a constant offset that corresponds to the control output dimension.

2. The application method of combining DNN and LSTM models in intelligent building control systems as described in claim 1, characterized in that: The spatial digital twin system built based on the BIM model includes, A 1:1 scale 3D as-built model is constructed using the Revit modeling platform, spatial structural information is extracted, and smart devices installed in the building are mapped and embedded into the corresponding spatial units of the building model according to their location.

3. The application method of combining DNN and LSTM models in intelligent building control systems as described in claim 1 or 2, characterized in that: The construction of spatial static parameter encoding for intelligent control includes... The spatial structural elements extracted from the 3D as-built model are normalized and encoded as features, and used as the spatial dimension vector of the neural network input layer.

4. The application method of combining DNN and LSTM models in intelligent building control systems as described in claim 3, characterized in that: The spatial dimension vector used as the input layer of the neural network includes, Spatial static parameter encoding and sensing data establish a correspondence through a spatial index binding mechanism, and synchronously participate in the construction of sequence input.

5. The application method of combining DNN and LSTM models in intelligent building control systems as described in any one of claims 1, 2, and 4, characterized in that: The construction of a multi-source heterogeneous sensing channel driven by an IoT middleware platform includes... By integrating various types of sensor devices deployed in the space area through a smart building control platform, setting a unified transmission cycle, and centrally collecting and scheduling real-time sensing data.

6. The application method of combining DNN and LSTM models in intelligent building control systems as described in claim 5, characterized in that: The input for constructing the multidimensional time series includes, Time labels are added to various types of sensor data, and standardized time series tensor structures are generated by combining device numbers and sampling step sizes.

7. The application method of combining DNN and LSTM models in intelligent building control systems as described in any one of claims 1, 2, 4, and 6, characterized in that: The fusion of temporal modeling and feature mapping capabilities includes, A unified format standard is established for encoding time-series data and spatial static parameters. A spliced ​​input structure is constructed, and the data flows into LSTM and DNN for joint modeling according to preset channels.

8. The application method of combining DNN and LSTM models in intelligent building control systems as described in claim 7, characterized in that: The data closed-loop architecture for building information and multi-device linkage includes, The DNN output is mapped to device control commands, which are then transmitted to the corresponding devices via a communication interface, and the status feedback is sent back to the perception layer.

9. An application system combining DNN and LSTM models in an intelligent building control system, employing the application method of combining DNN and LSTM models in an intelligent building control system as described in any one of claims 1 to 8, characterized in that: It includes a spatial modeling and static parameter encoding module (100), a multi-source perception fusion and time series construction module (200), and a neural reasoning and device linkage control module (300). The spatial modeling and static parameter encoding module (100) is used to construct a digital twin system through the BIM model, extract building structure and equipment layout information, standardize and encode spatial attributes, and generate static input fields adapted to neural networks. The multi-source sensing fusion and time series construction module (200) is used to collect data from multiple environmental sensors, complete multi-source alignment through a unified timestamp mechanism, and fuse static coding to generate standardized multi-dimensional time series input; The neural reasoning and device linkage control module (300) is used to integrate spatial static parameter encoding and temporal data input into a neural network structure to complete feature modeling and control command generation, and to link intelligent devices to respond dynamically.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method for applying the combination of DNN and LSTM models in an intelligent building control system as described in any one of claims 1 to 8.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for applying the combination of DNN and LSTM models in an intelligent building control system as described in any one of claims 1 to 8.

Images