CN-120540074-B - Intelligent building self-adaptive comprehensive management and control method and system based on deep learning

CN120540074BCN 120540074 BCN120540074 BCN 120540074BCN-120540074-B

Abstract

The invention discloses a self-adaptive comprehensive control method and system for intelligent building based on deep learning, which relates to the technical field of intelligent building, and comprises the steps of effectively reducing the time sequence dimension and removing redundant trend information through a neural clustering and DCT frequency domain compression mechanism, enhancing the prediction stability and generalization capability of a model under long time span, carrying out global optimization on LSTM model parameters through a self-adaptive genetic algorithm and an APSO algorithm, the method has the advantages that the optimal model search with extremely small jump-out part and faster convergence is realized, so that the generalization capability and prediction precision of the model to the user behavior change are enhanced, the performance and self-adaptive optimization capability of the user behavior prediction model are remarkably improved, the strategy migration is realized through a MAML quick parameter updating mechanism, the immediate adaptation of the control model under abnormal conditions is realized, the response speed is improved, and the continuous optimization capability is realized in a dynamic scene.

Inventors

YAN LONGLONG
ZENG YIXIN
Cui Shuoshuo
WANG CHONGYING
FAN SHIMING
MIAO RAN
WANG JIALI
LIU WEI

Assignees

朗高科技有限公司

Dates

Publication Date: 20260508
Application Date: 20250519

Claims (8)

1. An intelligent building self-adaptive comprehensive control method based on deep learning is characterized by comprising the following steps of, Collecting multi-source data for preprocessing to construct a building space diagram structure, and predicting the building environment state based on neural clustering and a transducer model combined with a frequency domain compression technology; The multi-source data comprises environmental sensor data and video monitoring data; Based on the prediction result, predicting a future path of the user by utilizing an LSTM model, and outputting control actions according to the environment state prediction and the future path of the user through deep reinforcement learning; The abnormal condition of the management and control implementation is monitored in real time, the management and control actions are adjusted by combining a meta-learning mechanism, and the adjustment management and control actions are uploaded and stored; the prediction of building environment states based on neural clustering and transform model combined frequency domain compression techniques includes, Through the building space diagram structure G, feature vectors of each node and surrounding nodes in the diagram structure are fused by using a graph neural network, and the enhanced feature vector representation is obtained Forming a post-enhancement time series matrix Dividing the neural cluster into a plurality of dynamic similar cluster blocks, initializing k cluster centers by adopting k-means++, and calculating the state after structure enhancement With cluster center Membership degree of (C) Distributing cluster centers with the largest membership degree to form a cluster sequence set Converting the cluster sequence set into a frequency domain, compressing the representation, and then restoring to obtain a reconstructed compressed time sequence ; Will compress the time series In the input transducer model encoder structure, modeling the dependency relationship among the states at each moment through a time sequence self-attention mechanism to generate a prediction result sequence of the building environment state Calculating the total loss function according to the prediction result Optimizing parameters of a transducer model to obtain a prediction result of the final building environment state ; The predicting of the future path of the user by using the LSTM model means that the building environment state is predicted to be a result sequence based on the predicted result The input attention weighting module obtains an attention score ; Obtaining a sequence of attention scores by calculating the attention score of the environmental state for each time step and converting into normalized weights Obtaining a new weighted state sequence Input into variant LSTM cell to obtain hidden state of each step ; Selecting the hidden state of the last time step as the internal state vector formed after the user perceives the whole prediction environment Outputting user behavior predicted value by inputting to full connection layer ; Calculating a loss function between a predicted value of the user behavior and the actual observation behavior, calculating the gradient of model parameters of the loss function through back propagation, and carrying out iterative update of the parameters with gradient descent by adopting an Adam optimizer; taking the model parameters in the neural network as individuals, optimizing the model parameters in the neural network through an adaptive genetic algorithm, and reserving the individuals with the highest fitness Global optimization is carried out through APSO algorithm to obtain final optimal parameter solution Predicting user behavior by combining environmental state prediction results 。
2. The intelligent building self-adaptive comprehensive control method based on deep learning according to claim 1, wherein the prediction of the environmental state and the output control action of the future path of the user by the deep reinforcement learning refer to the prediction result of the environmental state Prediction result of user behavior Construction of state vectors in conjunction Defining target control variables of required control equipment and constructing continuous value action space Calculating a joint rewards function With weighted ambient state vectors As input, control action As an output, the SAC algorithm is adopted to train the control strategy, and the strategy parameters are obtained by maximizing the expected return function Desired target value ; Continuously optimizing the desired target value during the training process Obtaining current policy parameters State vector is put into As a function of policy The input and output of the control operation probability distribution in the state Selecting an action with the highest probability density by a maximum probability mode After the control action is generated, converting the instruction into an MQTT protocol structure according to the control parameter format of each device, sending the MQTT protocol structure to target execution equipment through an equipment gateway, and automatically recording the whole process log of the current round control strategy after the control execution is completed.
3. The intelligent building self-adaptive comprehensive control method based on deep learning as claimed in claim 2, wherein the real-time monitoring of abnormal conditions of control implementation and the combination of element learning mechanism to adjust control actions means to dynamically detect abnormal values by using abnormal deviation measurement indexes When the abnormal deviation is identified to be out of limit, a MAML quick adjustment mechanism is adopted to automatically migrate from the current control task to the output of the new scene control, and update and obtain parameters ; By updated policy network parameters Inputting the probability distribution of the current building environment state calculation selection action And selecting the optimal action through sampling, executing the selected action, feeding back rewards in the environment, and carrying out the next decision through the new state.
4. The intelligent building self-adaptive comprehensive control method based on deep learning as claimed in claim 3, wherein the uploading and storing of the adjustment and control actions means that control parameters of each device are converted into unified control instruction structure format binding meta information, triggering background information of each action is packaged together with the action to form an action and context data packet as a control execution unit, and each stage in the control process is to push the synchronization and organization record unit to an event log module, and the synchronization and the context data packet is uploaded to a central database in batches after unified organization by a data management module.
5. The intelligent building self-adaptive comprehensive control method based on deep learning as set forth in claim 4, wherein the preprocessing of the collected multi-source data to construct a building space structure figure is to extract room units and connection relations through a BIM model of a building to construct a node set and an edge set Obtaining an adjacency matrix ; The multi-source data are collected through a temperature and humidity sensor, a noise sensor and a high-definition camera, data uploading and equipment control are carried out by combining an MQTT protocol, coding is carried out in a unified format through a central data platform, node state vectors are extracted, time synchronization and alignment are carried out by adopting a sliding window mechanism, and feature vectors are formed Repairing the missing data by using a latest value filling method and performing Z-Score standardization to construct a standardized feature matrix ; Constructing a time sequence matrix according to the normalized eigenvectors Combining adjacency matrices A space diagram structure G is established.
6. An intelligent building self-adaptive comprehensive control system based on deep learning is characterized by comprising the intelligent building self-adaptive comprehensive control method based on the deep learning according to any one of claims 1-5, The diagram construction and preprocessing module is used for analyzing the building structure and the multi-source data and generating a standardized diagram structure and a characteristic sequence; The feature enhancement compression modeling module is used for extracting main trend features of the dynamic environment by utilizing the graph neural network and clustering compression; the environment state prediction modeling module is used for predicting the environment state based on the transducer and modeling the time dependence and the sequence relationship; the behavior perception prediction module is used for fusing attention and LSTM modeling user paths and outputting behavior probability distribution; the intelligent control and execution module is used for adopting SAC algorithm decision-making equipment to control, combining the state and behavior prediction optimization strategy, and generating an optimal control scheme for real-time regulation and control.
7. A computer device comprises a memory and a processor, wherein the memory stores a computer program, and the computer program is characterized in that the processor realizes the steps of the intelligent building self-adaptive comprehensive management and control method based on deep learning according to any one of claims 1-5 when executing the computer program.
8. A computer readable storage medium, on which a computer program is stored, is characterized in that the computer program, when being executed by a processor, implements the steps of the intelligent building adaptive comprehensive management and control method based on deep learning as set forth in any one of claims 1 to 5.

Description

Intelligent building self-adaptive comprehensive management and control method and system based on deep learning Technical Field The invention relates to the technical field of intelligent buildings, in particular to a self-adaptive comprehensive control method and system for intelligent buildings based on deep learning. Background With the rapid development of information technology and intelligent perception technology, intelligent buildings serve as important components of novel urban construction, and become key means for improving building operation efficiency, energy utilization rate and personnel comfort. However, the traditional building automation system mainly relies on a rule preset or static scheduling mode to carry out environmental control, lacks of real-time sensing capability and dynamic response capability to building space states, and is difficult to adapt to the development trend of diversified use scenes and complicated personnel behaviors. In recent years, with the fusion application of Building Information Model (BIM), internet of things (IoT), edge computing and artificial intelligence and other technologies, a new idea is provided for comprehensive management and control of intelligent buildings. However, there are still many limitations to the existing intelligent building environment management and control scheme. Firstly, the space-time isomerism of the multisource perception data is strong, the quality is different, the modeling of the environment state is difficult, the traditional method usually ignores the topological structure characteristics and the personnel flow rules of the building space, the dynamic relation among nodes is difficult to capture, secondly, when the existing model is used for carrying out building environment prediction and behavior modeling, a time sequence is usually used as a core, the comprehensive utilization of the local space aggregation mode and the frequency domain information is lacked, and the prediction precision and the generalization capability are limited. In addition, a control method based on a predefined strategy or linear optimization is generally adopted in a system control layer, linkage consideration of complex environment evolution and user behavior response is lacked, and problems of lag in regulation and control, high energy consumption, poor user experience and the like can occur in actual operation. Particularly, under abnormal scenes such as abrupt change of user behaviors or model mismatch, an effective abnormality diagnosis and strategy self-adaption mechanism is lacked, and safety and robustness of the intelligent building system are restricted. Therefore, a new building comprehensive control method with sufficient data fusion, high prediction precision, strong behavior adaptability and sustainable self-optimization is needed. Disclosure of Invention The present invention has been made in view of the above-described problems occurring in the prior art. Therefore, the invention provides a self-adaptive comprehensive control method for intelligent buildings based on deep learning, which solves the problems of insufficient data fusion, poor adaptability of a prediction model, weak behavior adaptability and optimal static stiffness. In order to solve the technical problems, the invention provides the following technical scheme: in a first aspect, the invention provides a smart building adaptive comprehensive control method based on deep learning, which comprises the following steps of, Collecting multi-source data for preprocessing to construct a building space diagram structure, and predicting the building environment state based on neural clustering and a transducer model combined with a frequency domain compression technology; The multi-source data comprises environmental sensor data and video monitoring data; Based on the prediction result, predicting a future path of the user by utilizing an LSTM model, and outputting control actions according to the environment state prediction and the future path of the user through deep reinforcement learning; and (3) monitoring abnormal conditions of management and control implementation in real time, adjusting management and control actions by combining a meta-learning mechanism, and uploading and storing the adjustment management and control actions. As an optimal scheme of the intelligent building self-adaptive comprehensive management and control method based on deep learning, the method comprises the steps of carrying out building environment state prediction based on the neural clustering and a transducer model by combining a frequency domain compression technology, Through a building space diagram structure G, fusing feature vectors of each node and surrounding nodes in the diagram structure by utilizing a diagram neural network to obtain an enhanced feature vector representation x i (t) ', forming an enhanced time sequence matrix B i ', dividing the enhanced time sequence matrix