Artificial intelligence-based building energy consumption optimization method

Abstract

The present application relates to the technical field of energy consumption simulation, and more particularly to a building energy consumption optimization method based on artificial intelligence. The method comprises the following steps: obtaining multi-modal sensor data and performing device physical space semantic mapping modeling to obtain a sensor spatio-temporal fusion feature set; extracting device energy transfer influence factors according to the sensor spatio-temporal fusion feature set to obtain a dynamic energy consumption influence factor matrix; constructing an energy consumption probability cloud map according to the dynamic energy consumption influence factor matrix; performing adversarial strategy optimization on the energy consumption probability cloud map to obtain a multi-objective energy-saving control strategy manifold; generating an energy-saving optimization instruction set based on the multi-objective energy-saving control strategy manifold; obtaining real-time building energy consumption data and implementing device control local comfort instruction optimization on the real-time building energy consumption data according to the energy-saving optimization instruction set to obtain a personalized control instruction set. The present application improves the intelligence, refinement and personalization of building energy consumption management.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of energy consumption simulation, and in particular to a building energy consumption optimization method based on artificial intelligence. BACKGROUND

[0002] The existing building energy saving system usually relies on preset rules or static control strategies (such as fixed temperature threshold, device start-stop control based on time period), although it reduces energy waste to some extent, but it is difficult to adapt to complex environmental dynamic changes. For example, weather conditions (temperature, humidity, light intensity) may fluctuate rapidly over time, the flow density inside the building, the use of equipment also presents nonlinear changes, a single fixed strategy often cannot adjust the control parameters in time, leading to energy management lag, and poor energy efficiency optimization effect. At present, most energy saving management systems still mainly optimize single equipment, for example, the air conditioning system adjusts the refrigerating capacity according to the room temperature, the lighting system controls the switch according to the preset time or the light sensor, and the elevator, fresh air system, hot water supply and other equipment run independently. This isolated optimization method ignores the coupling relationship between multiple devices inside the building. For example, the energy consumption of the air conditioning system and the fresh air system influences each other, the change of the air exchange rate of the fresh air system may cause the air conditioning load to fluctuate sharply, and then affect the overall energy consumption balance. Similarly, the elevator operation mode and the lighting load distribution also indirectly affect the heat load of the air conditioning system, and if there is no cross-system collaborative optimization, it will lead to low efficiency of overall energy scheduling. In addition, although the application of Internet of Things (IoT) technology improves the building energy consumption monitoring capability, it can collect multi-dimensional data such as temperature and humidity, light intensity, equipment running state in real time, but the utilization rate of these data is low, and the existing system is mainly used for state monitoring and early warning, and lacks deep data analysis and intelligent decision-making capability. In most scenarios, the energy saving strategy still relies on manual adjustment or simple rule setting, and it is difficult to realize self-learning and dynamic optimization. For example, when the personnel density changes or the external environment fluctuates, the traditional system usually cannot respond quickly, leading to unreasonable energy scheduling, affecting user comfort and energy use efficiency. SUMMARY

[0003] Therefore, it is necessary to provide a building energy consumption optimization method based on artificial intelligence to solve at least one of the above technical problems.

[0004] To achieve the above purpose, a building energy consumption optimization method based on artificial intelligence comprises the following steps:

[0005] Step S1: acquiring multi-modal sensor data and constructing a heterogeneous sensor dynamic spatio-temporal adjacency matrix; performing device physical space semantic mapping modeling according to the heterogeneous sensor dynamic spatio-temporal adjacency matrix to obtain a sensor spatio-temporal fusion feature set;

[0006] Step S2: cross-energy-consumption influence mapping topology modeling is performed on the sensing spatio-temporal fusion feature set, to obtain a cross-energy-consumption correlation topology graph; and a device energy transmission influence factor is extracted according to the cross-energy-consumption correlation topology graph, to obtain a dynamic energy consumption influence factor matrix;

[0007] Step S3: multi-scale energy consumption evolution deduction is performed according to the dynamic energy consumption influence factor matrix, to generate a building energy consumption evolution graph; and an energy consumption probability cloud map is constructed according to the building energy consumption evolution graph;

[0008] Step S4: an antagonistic strategy optimization is performed on the energy consumption probability cloud map, to obtain a multi-objective energy-saving control strategy manifold; an energy-saving optimization instruction set is generated based on the multi-objective energy-saving control strategy manifold, and is uploaded to a building energy consumption control platform, to control devices;

[0009] Step S5: real-time building energy consumption data is acquired, device predicted energy consumption deviation bloodline tracking is performed on the real-time building energy consumption data according to the energy-saving optimization instruction set, and device control local comfort degree instruction optimization is implemented, to obtain an individualized control instruction set, which is uploaded to the building energy consumption control platform, to control devices.

[0010] Optionally, step S1 specifically comprises:

[0011] Step S11: multi-modal sensor groups are used to acquire multi-modal sensing data, wherein the multi-modal sensing data comprises temperature and humidity sensing data, illumination sensing data, people flow density infrared sensing data, and device state sensing data;

[0012] Step S12: time step time sequence alignment is performed on the multi-modal sensing data, to obtain a time sequence aligned environment data set;

[0013] Step S13: non-Euclidean topology mapping is performed on the time sequence aligned environment data set, and an initial space semantic topology structure graph is constructed based on the topology mapping result;

[0014] Step S14: neighborhood adaptive topology optimization is performed based on the initial space semantic topology structure graph, to obtain a device space semantic graph, and a heterogeneous sensing dynamic spatio-temporal adjacency matrix is obtained based on fusion of device operation modes and environment data based on the device space semantic graph;

[0015] Step S15: a device physical space semantic mapping model is established according to the heterogeneous sensing dynamic spatio-temporal adjacency matrix, feature screening and correlation optimization are performed on the model establishment result, and a sensing spatio-temporal fusion feature set is obtained.

[0016] Optionally, step S13 specifically comprises:

[0017] Step S131: building area device operation data is acquired, and synchronous time sequence alignment is performed on the building area device operation data and the time sequence aligned environment data set, to obtain time sequence aligned device operation data;

[0018] Step S132: Non-Euclidean topological mapping is performed on the time alignment device operation data and the time alignment environment data set, a space-time relationship graph between the devices and the environment variables in the building area is constructed, and a building area space-time relationship graph is obtained;

[0019] Step S133: Physical space interaction modeling is performed based on the building area space-time relationship graph, and the physical space interaction modeling result is mapped to a preset building physical space coordinate system, a node in the space-time relationship graph is converted into an actual physical space coordinate node, and a device-environment physical space interaction data is constructed;

[0020] Step S134: Node space layout analysis is performed on the device-environment physical space interaction data, and the interaction relationship between the devices and the environment is mapped according to the node space layout analysis result, and an environment physical space interaction graph is obtained.

[0021] Step S135: Graph optimization processing is performed based on the environment physical space interaction graph, and an initial space semantic topology structure graph is constructed.

[0022] Optionally, step S14 is specifically:

[0023] Step S141: Topological node local density calculation is performed according to the initial space semantic topology structure graph, the calculation neighborhood radius is set to 2, the local density threshold is set to 0.1, and the nodes below the local density threshold are regarded as sparse areas, so as to obtain a topological structure local density distribution;

[0024] Step S142: The minimum sample number is set to 5, the density threshold is set to 0.3, the node neighborhood adaptive clustering is performed on the initial space semantic topology structure graph, the weight range is set to [0, 1], the neighborhood adaptability weight is calculated based on the clustering result, and the neighborhood adaptive topology optimization initial parameter is generated;

[0025] Step S143: The weight adjustment coefficient is set to 0.5, the neighborhood topology weight adjustment is performed on the initial space semantic topology structure graph according to the neighborhood adaptive topology optimization initial parameter, the optimization scale [1, 3] is set to perform multi-scale topology structure optimization, and an optimized device space semantic graph is obtained.

[0026] Step S144: Based on the optimized device space semantic graph, a device operation mode feature with a feature dimension of 5 is extracted, and the time alignment environment data set is fused, the device operation mode feature is normalized, and standardized device operation mode data is generated;

[0027] Step S145: The standardized device operation mode data is embedded in the topology structure of the optimized device space semantic graph, the interaction modeling of the device operation mode and the environment data is performed, and a preliminary heterogeneous sensor dynamic space-time adjacency matrix is obtained.

[0028] Step S146: Feature screening and topological correlation optimization are performed on the preliminary heterogeneous sensor dynamic spatio-temporal adjacency matrix, key interaction relationships are strengthened, and low correlation edge weights are removed, thereby generating a heterogeneous sensor dynamic spatio-temporal adjacency matrix.

[0029] Optionally, step S2 specifically comprises:

[0030] Step S21: Cross-device energy consumption influence factor screening is performed on the sensor spatio-temporal fusion feature set, a correlation threshold of [0.7, 0.8] is set to retain high correlation features, energy consumption influence weights of each device are calculated, and an initial energy consumption influence feature matrix is constructed based on the energy consumption influence weights;

[0031] Step S22: Non-Euclidean graph topological mapping is performed on the initial energy consumption influence feature matrix, a maximum connection threshold of 0.5 is set for the adjacency matrix, an energy consumption correlation topological structure across devices and spatial regions is constructed, and an initial cross-device energy consumption correlation topological graph is generated;

[0032] Step S23: Spectral clustering analysis is performed on the initial cross-device energy consumption correlation topological graph, the first 20 feature vectors are selected for Laplace feature decomposition, and a feature cutoff threshold of 0.1 is set to extract main energy consumption influence patterns, thereby generating a cross-device energy consumption correlation topological graph;

[0033] Step S24: Energy consumption transmission relationship modeling between devices is performed based on the cross-device energy consumption correlation topological graph, an energy consumption transmission rate of [0.05, 0.2] is set, a dynamic weight variation interval of [0, 1] is set to assign a dynamic weight, and an initial dynamic energy consumption influence factor matrix is generated;

[0034] Step S25: Normalization processing is performed on the initial dynamic energy consumption influence factor matrix, an influence factor confidence interval of [0.95, 1.05] is set to correct the robustness of the influence factor, and energy consumption influence parameters are optimized, thereby outputting a dynamic energy consumption influence factor matrix.

[0035] Optionally, the multi-scale energy consumption evolution deduction in step S3 specifically comprises:

[0036] Time series decomposition is performed on the dynamic energy consumption influence factor matrix, multi-scale time steps are set, trend decomposition and period decomposition are performed, and an energy consumption time series decomposition data set is generated;

[0037] Energy consumption propagation path causal inference is performed on the energy consumption time series decomposition data set, energy consumption correlation strengths between devices are calculated, and an energy consumption propagation network is constructed, thereby generating an initial energy consumption propagation structure;

[0038] Based on the initial energy consumption propagation structure, energy consumption propagation mode clustering is performed, energy consumption propagation paths are screened, and low-contribution energy consumption paths are removed, thereby generating an optimized energy consumption propagation network;

[0039] Time series feature extraction is performed on the optimized energy consumption propagation network, energy consumption time sequence segments are divided, and empirical mode tendency decomposition is performed on the energy consumption time sequence segments to obtain energy consumption time sequence feature data;

[0040] Autoregressive integral short-term energy consumption tendency modeling is performed based on the energy consumption time sequence feature data to generate short-term energy consumption prediction data;

[0041] Residual analysis is performed on the short-term energy consumption prediction data, and the residual analysis result is input into a preset long short-term memory network to perform long-term energy consumption evolution modeling to obtain long-term energy consumption prediction data;

[0042] The long-term energy consumption prediction data and the short-term energy consumption prediction data are fused to construct a comprehensive energy consumption prediction sequence; multi-resolution wavelet transform is performed based on the comprehensive energy consumption prediction sequence to extract energy consumption change patterns at different time scales to generate a multi-scale energy consumption evolution prediction matrix;

[0043] State probability modeling is performed based on the multi-scale energy consumption evolution prediction matrix to calculate an energy consumption state transition matrix, and the energy consumption state transition matrix is combined with a building energy consumption evolution graph to obtain the building energy consumption evolution graph.

[0044] Optionally, the long-term energy consumption evolution modeling specifically includes:

[0045] The short-term energy consumption prediction data and the residual analysis result are time step aligned to obtain a time sequence aligned residual data set;

[0046] Local spatiotemporal feature extraction is performed on the time sequence aligned residual data set to generate residual feature data;

[0047] A long short-term memory network model is constructed based on the residual feature data set;

[0048] The long short-term memory network model is back propagated, network parameters are optimized using the gradient descent method, the time sequence aligned residual data set is modeled for long-term energy consumption evolution, and a long-term energy consumption evolution model is obtained;

[0049] The long-term energy consumption evolution model is used to perform long-term energy consumption prediction on the residual feature data to obtain long-term energy consumption prediction data.

[0050] Optionally, the generating of the energy-saving optimization instruction set in step S4 specifically includes:

[0051] Device category strategy adaptability analysis is performed based on the multi-objective energy-saving control strategy manifold, and device control demand hierarchical clustering is performed to generate a device control classification index;

[0052] The multi-objective energy-saving control strategy manifold is mapped for energy-saving strategy constraints based on the device control classification index to obtain a constraint optimization strategy set;

[0053] Time series scheduling optimization is performed on the constraint optimization strategy set, the strategy execution time is adjusted, and a time-optimized strategy set is generated.

[0054] Strategy execution priorities are allocated based on the time-optimized strategy set, and strategy execution parameters are adaptively adjusted in combination with the time sequence alignment environment data set to generate an adaptive energy-saving control instruction set.

[0055] The adaptive energy-saving control instruction set is subjected to consistency verification to verify the execution feasibility of the adaptive energy-saving control instruction set, and an energy-saving optimization instruction set is obtained and uploaded to a building energy consumption control platform for execution of device control tasks.

[0056] Optionally, the device predicted energy consumption deviation in step S5 is specifically:

[0057] Building device historical energy consumption data is obtained, and a baseline energy consumption prediction model is established based on the building device historical energy consumption data and the energy-saving optimization instruction set, thereby obtaining device baseline energy consumption prediction data.

[0058] Real-time building energy consumption data is obtained, error comparison is performed on the device baseline energy consumption prediction data and the real-time building energy consumption data, a predicted energy consumption deviation metric is calculated, and preliminary screening of abnormal energy consumption is performed based on a preset deviation prediction, thereby obtaining a device energy consumption deviation metric matrix.

[0059] Device control logs are obtained, and a device energy consumption change causal relationship graph is constructed according to the device control logs and the dynamic energy consumption influence factor matrix, energy consumption influence factor causal verification is performed, and a device energy consumption bloodline tracking chain is generated.

[0060] The device energy consumption deviation metric matrix is subjected to energy consumption anomaly tracing in combination with the device bloodline tracking chain, and device abnormal energy consumption mode classification is performed based on the energy consumption anomaly tracing result, thereby obtaining an energy consumption deviation backtracking analysis report.

[0061] Optionally, the establishment of the baseline energy consumption prediction model is specifically

[0062] Building device historical energy consumption data is obtained, and the building device historical energy consumption data is subjected to time sequence alignment, thereby obtaining a standardized device historical energy consumption data set.

[0063] The device historical energy consumption data is subjected to energy-saving optimization influence quantification based on the energy-saving optimization instruction set and the dynamic energy consumption influence factor matrix, thereby generating an energy-saving optimization influence factor data set.

[0064] The standardized device historical energy consumption data set and the energy-saving optimization influence factor data set are used to perform baseline energy consumption prediction modeling through a preset regression algorithm, thereby obtaining a device baseline energy consumption prediction model.

[0065] Cross-validation is performed on the device benchmark energy consumption prediction model, and the model parameters are iteratively adjusted to obtain the benchmark energy consumption prediction model;

[0066] The energy-saving optimization instruction set is input into the benchmark energy consumption prediction model to obtain device benchmark energy consumption prediction data.

[0067] The present application can effectively improve the building energy efficiency and optimize the energy consumption control strategy of the equipment. First, through the environmental data and equipment state information obtained by the multi-modal sensor, combined with the non-Euclidean topological mapping method, the space-time relationship between the equipment and the environment is modeled, and a high-precision space-time fusion feature set is generated. This provides a precise physical and semantic mapping basis for subsequent energy consumption optimization, enabling the energy consumption optimization process to better reflect the real interaction and changes of the equipment and the environment. Then, through cross-energy consumption influence factor screening and related topological modeling, a dynamic energy consumption influence factor matrix is obtained, which can help accurately quantify the specific influence of various factors (such as temperature and humidity changes, equipment load, etc.) on building energy consumption, and thus provide a scientific basis for the construction of building energy consumption evolution atlas. Through multi-scale energy consumption evolution deduction, not only can the short-term fluctuations of building energy consumption be predicted, but also the long-term energy consumption trend can be scientifically predicted, thereby providing a basis for the formulation of energy-saving optimization strategies. Through the construction of the energy consumption propagation network and time series analysis, the building energy consumption can be carefully decomposed and analyzed, thereby improving the understanding of the building energy consumption evolution process. Based on these analysis results, energy-saving optimization instruction sets can be generated through multi-objective energy-saving control strategy manifold, and real-time uploaded to the building energy consumption control platform for execution, realizing personalized equipment energy consumption control. This process helps to dynamically adjust according to the needs of different equipment and environment, thereby improving building energy efficiency and ensuring that the operation of various equipment is more in line with energy-saving requirements. Through the technology of device predicted energy consumption deviation bloodline tracking, when the real-time energy consumption data deviates from the predicted value, the abnormality can be quickly identified and traced back to the source for analysis, thereby finding the root cause of the energy consumption anomaly and making timely adjustments. The combination of this energy consumption deviation measurement matrix and the energy consumption change causal relationship diagram can effectively enhance the system's control ability over equipment energy consumption fluctuations, further improving the accuracy and flexibility of energy consumption management. Finally, based on the generation of the energy-saving optimization influence factor data set, the influence of the energy-saving optimization instruction set can be accurately quantified, providing data support for the device benchmark energy consumption prediction model, ensuring the accuracy of the device energy consumption prediction, and improving the reliability of the system. In summary, the present application combines multi-modal data acquisition, dynamic energy consumption influence quantification, cross-energy consumption analysis, and multi-objective energy-saving optimization strategy, not only can accurately predict and optimize building energy consumption, but also can respond to changes in building energy efficiency in real time, ultimately achieving the maximization of building energy efficiency and the optimization of equipment energy consumption control. BRIEF DESCRIPTION OF DRAWINGS

[0068] Other features, objects, and advantages of the application will become more apparent from the following detailed description when read in connection with the following drawings:

[0069] Fig. 1 A schematic diagram of the step flow of the building energy consumption optimization method based on artificial intelligence of the present application;

[0070] Fig. 2 A schematic diagram of the detailed step flow of step S1 in the present application;

[0071] The implementation, functional features and advantages of the present application will be further described with reference to the embodiments in conjunction with the drawings. DETAILED DESCRIPTION

[0072] The technical method of the present application will be described clearly and completely below in conjunction with the drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0073] In addition, the drawings are only schematic illustrations of the present application and are not necessarily drawn to scale. The same reference numerals in the drawings represent the same or similar parts, and thus repeated descriptions thereof will be omitted. Some of the block diagrams shown in the drawings are functional entities, which do not necessarily have to correspond to physically or logically independent entities. The functional entities can be implemented in the form of software, or in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0074] It should be understood that although the terms "first", "second" and the like can be used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another. For example, without departing from the scope of the exemplary embodiments, a first element can be referred to as a second element, and similarly a second element can be referred to as a first element. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0075] To achieve the above-mentioned purpose, please refer to Figs. 1-2 The present application provides a building energy consumption optimization method based on artificial intelligence, which comprises the following steps:

[0076] Step S1: Obtain multi-modal sensor data and construct a heterogeneous sensor dynamic spatio-temporal adjacency matrix; perform device physical space semantic mapping modeling according to the heterogeneous sensor dynamic spatio-temporal adjacency matrix to obtain a sensor spatio-temporal fusion feature set;

[0077] In this embodiment, real-time data collection is performed within the building using multiple multimodal sensor devices (such as infrared sensors, ultrasonic sensors, light sensors, temperature and humidity sensors, etc.). Taking the collection time interval Δt = 0.1 seconds as an example, data on the operating status of building equipment (such as the status of fans, air conditioners, and lighting equipment) and environmental parameters (such as temperature, humidity, and carbon dioxide concentration) are obtained. This sensor data is first subjected to noise removal processing, using the Kalman filter algorithm to remove noise components from the sensor readings. At the same time, missing data is interpolated using nearest neighbor interpolation or spline interpolation algorithms to ensure data integrity. Then, time synchronization is performed, normalizing all sensor data to a unified time step to ensure data consistency. Next, based on the spatiotemporal characteristics of the data, a heterogeneous sensor dynamic spatiotemporal adjacency matrix is ​​constructed using a graph neural network (GNN). Each node in the matrix represents a device or environmental parameter, and the edges represent the degree of association between different devices or sensors. The edge weights are calculated based on the spatial and temporal associations between devices. Finally, based on this adjacency matrix, semantic mapping is performed in the three-dimensional building physical space, the operating status and environmental parameters of the equipment are mapped to specific locations in the physical space, and the sensing spatiotemporal fusion feature set is extracted as input for subsequent steps.

[0078] Step S2: Perform cross-domain energy consumption impact mapping topology modeling on the sensing spatiotemporal fusion feature set to obtain a cross-domain energy consumption correlation topology map; extract the device energy transfer impact factor based on the cross-domain energy consumption correlation topology map to obtain a dynamic energy consumption impact factor matrix;

[0079] In this embodiment, based on the constructed sensor spatiotemporal fusion feature set, cross-domain data collaborative modeling is performed for multiple devices in a building. At the local device end of each building area, the sensor data is first denoised, normalized, and feature extracted. Then, a weighted average model is used to learn and aggregate the data with the global model to ensure data privacy. Based on the learning results of these device ends, a graph neural network (GNN) is used to map energy consumption influencing factors. Each node in the graph represents a building device, and the edges between nodes represent the energy consumption transfer relationship between different devices. By calculating the spatiotemporal correlation between devices and setting the maximum connection threshold T = 0.6, nodes with low correlation are screened out to construct a cross-domain energy consumption correlation topology. Based on this topology, the energy transfer impact factor of each device is calculated, and the least squares method (LSE) is used to estimate the device's contribution to overall energy consumption. The dynamic energy consumption impact factor matrix is ​​formed by the energy transfer rate and transfer relationship between devices. The specific rate is set between 0.05 and 0.2 to reflect the changes in energy consumption between devices.

[0080] Step S3: Perform multi-scale energy consumption evolution deduction according to the dynamic energy consumption influence factor matrix to generate a building energy consumption evolution graph; and construct an energy consumption probability cloud map according to the building energy consumption evolution graph;

[0081] In this embodiment, the long short-term memory network (LSTM) is used to model and predict the historical building energy consumption data according to the dynamic energy consumption influence factor matrix generated in the previous step. In order to better capture the short-term energy consumption trend, a 24-hour time sliding window is used for data analysis, and a short-term trend model is established. The empirical mode decomposition (EMD) is used to perform multi-scale decomposition on the building energy consumption data to extract the energy consumption patterns at different time scales. This step generates a building energy consumption evolution graph, which shows the long-term and short-term energy consumption evolution process of each device. Based on the energy consumption evolution graph, the Gaussian kernel density estimation (KDE) is used to calculate the energy consumption probability distribution of the device, and then the energy consumption probability cloud map is generated to reflect the energy consumption fluctuation of each device in the building under different environmental and usage conditions.

[0082] Step S4: Perform adversarial strategy optimization on the energy consumption probability cloud map to obtain a multi-objective energy-saving control strategy manifold; generate an energy-saving optimization instruction set based on the multi-objective energy-saving control strategy manifold and upload it to the building energy consumption control platform to control the devices;

[0083] In this embodiment, the generated energy consumption probability cloud map is used to perform adversarial optimization of the energy-saving control strategy using the generative adversarial network (GAN). In order to improve the robustness of the energy-saving strategy, the training rounds are set to 1000 rounds, and the model is adapted to adjust the energy-saving control strategy of the building devices through adversarial training. During this optimization process, for HVAC devices, the energy-saving control objective function is set, including controlling the indoor temperature deviation to be less than ±0.5℃, the humidity control to be within ±2%, and reducing the device energy consumption by at least 10%. The optimized energy-saving strategy manifold is adjusted by a multi-objective optimization algorithm (such as NSGA-II), and is confirmed by consistency verification. The verified energy-saving optimization instruction set is uploaded to the building energy consumption control platform through the MQTT protocol, and the instruction set automatically performs device control tasks, including adjusting the operation state of air conditioners, lighting, and other devices.

[0084] Step S5: Obtain real-time building energy consumption data, perform device predicted energy consumption deviation bloodline tracking on the real-time building energy consumption data according to the energy-saving optimization instruction set, and implement device control local comfort instruction optimization to obtain a personalized control instruction set, and upload it to the building energy consumption control platform to control the devices.

[0085] In this embodiment, real-time building energy consumption data is obtained through the building energy consumption control platform, and the baseline energy consumption of the equipment is predicted in combination with the energy-saving optimization instruction set. The predicted equipment energy consumption is compared with the real-time building energy consumption data through deviation measurement, the deviation trend of the equipment energy consumption is analyzed using the Markov hidden state model, and the deviation degree threshold is set to 5%, to screen out possible abnormal equipment. Combined with the dynamic energy consumption influence factor matrix, the Granger Causality Analysis (GCA) is used to trace the blood relationship of energy consumption deviation, analyze which equipment operation state leads to abnormal fluctuation of energy consumption, and generate an energy consumption deviation backtracking analysis report. The report lists the abnormal energy consumption mode of the equipment and the corresponding influence factors in detail. According to the analysis report, the energy-saving control strategy of the equipment will be further optimized, and the control of specific equipment such as air conditioners will be adjusted according to the difference in personnel distribution, automatically adjusting the comfort range to 22-26°C, and adjusting the brightness of the lighting equipment to keep it between 300 lux and 500 lux. The optimized personalized control instruction set will be uploaded to the building energy consumption control platform to automatically execute the equipment control task, ensuring that the building energy consumption is effectively controlled within a reasonable range.

[0086] Optionally, step S1 is specifically:

[0087] Step S11: Obtain multi-modal sensing data through a multi-modal sensor group, wherein the multi-modal sensing data includes temperature and humidity sensing data, illumination sensing data, people flow density infrared sensing data, and equipment state sensing data.

[0088] In this embodiment, a plurality of multi-modal sensor devices (such as infrared sensors, ultrasonic sensors, illumination sensors, temperature and humidity sensors, etc.) are responsible for collecting different types of data. The temperature and humidity sensor monitors the indoor temperature and humidity changes in real time, the illumination sensor captures the change of ambient light intensity, the infrared sensor monitors the people flow density, and the equipment state sensor obtains the working state of the equipment (such as air conditioners, lights, fans, etc.) in the building. The collection frequency is set to once per second (sampling frequency is 1 Hz) to ensure the real-time and integrity of the data. The collected data is pre-processed to remove outliers and noise data, and standardized to convert all data to the standard interval of 0 to 1, ensuring the compatibility and comparability of different sensor data.

[0089] Step S12: Time step time sequence alignment is performed on the multi-modal sensing data to obtain a time sequence aligned environment data set.

[0090] In this embodiment, the multi-modal data obtained from each sensor is aligned by timestamp. Due to the differences in collection frequency and timing of each sensor, time step adjustment is needed. For this purpose, a method combining linear interpolation and nearest neighbor interpolation is used to ensure that the data time steps are uniform and no information is lost. For example, for temperature and humidity and light sensor data, a uniform time step of 0.1 seconds is set, that is, a data point is generated every 0.1 seconds, and the missing time step data is filled by interpolation method. The aligned data set will be the basis for subsequent analysis and modeling, ensuring the consistency and timing of the data.

[0091] Step S13: Non-Euclidean topological mapping is performed on the time-aligned environmental data set, and an initial spatial semantic topological structure graph is constructed based on the topological mapping results;

[0092] In this embodiment, based on the time-aligned environmental data set, Non-Euclidean Graph is used for topological mapping, and Graph Neural Network (GNN) is used to model the relationship between data. Each sensor is a node in the graph, and different sensors are connected by edges, and the weight of the edge represents the association strength between them. For the relationship between device status and environmental data, by calculating the Euclidean distance and correlation, the weight threshold is set to 0.5, the associated edges with larger weights are retained, and the redundant connections are removed. Through this Non-Euclidean topological mapping, an initial spatial semantic topological structure graph of devices and environmental factors in the building space is generated, which is used for subsequent neighborhood optimization and feature fusion.

[0093] Step S14: Neighborhood adaptive topological optimization is performed based on the initial spatial semantic topological structure graph to obtain a device spatial semantic graph, and device operation mode and environmental data are fused based on the device spatial semantic graph to obtain a heterogeneous sensor dynamic spatio-temporal adjacency matrix;

[0094] In this embodiment, based on the generated initial spatial semantic topological structure graph, a graph adaptive algorithm is used to optimize the topology. The algorithm adaptively adjusts the neighborhood relationship between devices according to the actual operation of building devices and sensors, optimizes the topological structure, and makes the spatial semantic graph better reflect the physical space relationship between devices. For example, if a device is frequently used in a certain period of time and has strong correlation with the energy consumption of adjacent devices, the connection strength between them will automatically increase. The optimized device spatial semantic graph further fuses the operation mode and environmental data of the device, and generates a more representative heterogeneous sensor dynamic spatio-temporal adjacency matrix through weighted average method. This matrix not only contains the physical location relationship between devices, but also reflects the dynamic energy interaction between devices and environment.

[0095] Step S15: Establish a device physical space semantic mapping model based on the heterogeneous sensor dynamic spatio-temporal adjacency matrix. Feature selection and correlation optimization are performed on the model establishment results to obtain a set of sensor spatio-temporal fusion features.

[0096] In this embodiment, a device physical space semantic mapping model is established based on the heterogeneous sensor dynamic spatio-temporal adjacency matrix. Through the deep learning capability of the graph neural network, the model can map the device and environment data to the physical space of the building, and construct a more accurate device location and energy consumption correlation graph. In order to improve the accuracy and robustness of the model, feature selection and correlation optimization techniques are used. First, principal component analysis (PCA) is used to reduce the dimensionality of the sensor data and reduce redundant features. Second, correlation analysis (such as Pearson correlation coefficient) is used to select device features that have a greater impact on energy consumption. Finally, a set of high-quality sensor spatio-temporal fusion features is obtained by combining the spatial semantic mapping model, which can better reflect the relationship between device energy consumption and environmental changes, providing important input data for subsequent energy-saving strategy formulation and optimization.

[0097] Optionally, step S13 is specifically:

[0098] Step S131: Obtain building area device operation data, and synchronize the time sequence of the building area device operation data and the time sequence aligned environment data set to obtain time sequence aligned device operation data;

[0099] In this embodiment, the running data of all devices in the building area is obtained through the device monitoring system, including the real-time running state of air conditioning, lighting, heating and ventilation system (HVAC), fan and water pump and other devices. The running data of all devices is sampled at a frequency of one second, and is collected through standard network protocols (such as Modbus or BACnet). After obtaining these device data, the synchronized time sequence alignment is performed with the aligned environment data set. The environment data set includes temperature, humidity, illumination, and personnel density, and the sampling time step is set to 5 seconds. For the collected device running data and environment data, the missing data is supplemented by interpolation method (such as linear interpolation or spline interpolation), to ensure that all data are strictly aligned in time, so as to obtain complete time sequence aligned device running data. The key of this step is to ensure that all devices and environment data are accurately synchronized in time, which lays a foundation for subsequent analysis.

[0100] Step S132: Perform non-Euclidean topology mapping on the time sequence aligned device operation data and the time sequence aligned environment data set to construct a spatio-temporal relationship graph between devices and environmental variables in the building area, and obtain a building area spatio-temporal relationship graph;

[0101] In this embodiment, the topological mapping method of non-Euclidean graph is used to analyze the time alignment device operation data and environmental data. Each device and environmental variable is regarded as a node in the graph, and the relationship between the device operation state and the environmental variable is represented by the edge of the graph. A graph neural network (GNN) model is used to capture the spatio-temporal dependency between the devices and environmental variables, and the weight of the edge is dynamically adjusted based on the correlation and influence of the devices and environmental variables. For example, when the running state of the air conditioning device has a strong correlation with the indoor temperature change, the weight between them is larger, and vice versa. In this way, a spatio-temporal relationship graph between devices and environmental variables in the building area is constructed. This graph can reveal the complex interaction between devices and environment, providing data support for subsequent optimization decisions.

[0102] Step S133: Based on the spatio-temporal relationship graph of the building area, the physical space interaction is modeled, and the physical space interaction modeling result is mapped to the preset building physical space coordinate system, the nodes in the spatio-temporal relationship graph are converted to actual physical space coordinate nodes, and the device-environment physical space interaction data is constructed;

[0103] In this embodiment, based on the spatio-temporal relationship graph of the building area, the positions of the devices (such as the installation positions of HVAC devices and lighting devices) and their influences on the surrounding environment are analyzed. For example, the air conditioning device is located in a room, and the temperature change in the room will directly affect the energy efficiency performance of the air conditioning device. The devices and environmental variables (such as temperature and humidity) are modeled through physical space relationship, and the model is mapped to the actual physical coordinate system of the building. Through the physical space coordinate system, each node in the spatio-temporal relationship graph can be converted to the actual position in the building plan, making the relationship between each device and environmental variable more intuitive. For example, the positions of the air conditioner and the temperature and humidity sensor will directly affect their energy efficiency performance, so their spatial coordinates are accurately mapped to the building plan.

[0104] Step S134: The node space layout analysis is performed on the device-environment physical space interaction data, and the interaction relationship between the devices and the environment is mapped according to the node space layout analysis result, to obtain an environmental physical space interaction graph;

[0105] In this embodiment, the device-environment physical space interaction data is optimized through node space layout analysis. First, the space layout of the devices in the building area is analyzed using Voronoi diagram or spatial clustering algorithm. The interaction between the devices and the environment is mapped between the nodes, and the energy efficiency impact of the devices is analyzed through the spatial relationship of the nodes (such as the relative distance between the device and the temperature and humidity sensor, and the distribution of the air conditioner and the light control device). The energy efficiency of the device is usually reflected by its impact on the environment. For example, the air conditioner directly affects the environment by adjusting the indoor temperature and humidity. If the air conditioner is close to the environmental sensor (such as the temperature and humidity sensor), its control over the environment is more accurate and efficient, and therefore, the energy efficiency of the air conditioner has a greater impact on the environment, especially in energy saving and load optimization. Assuming that the spatial distance between the air conditioner and the temperature and humidity sensor in a certain area is close, the interaction between the two is strong, and the impact on the environment is greater, so the energy efficiency impact is greater. Through spatial layout analysis, strong interaction between devices is identified, and basis is provided for subsequent energy efficiency optimization. In this step, the algorithm used also includes spatial weighted analysis to ensure that the interaction between the device and the environment is accurately reflected. Finally, an environment physical space interaction graph is generated, which clearly indicates which devices have a close relationship with the environment and which devices have a greater impact on the environment in the energy efficiency optimization process.

[0106] Step S135: Perform graph optimization processing based on the environment physical space interaction graph, and construct an initial spatial semantic topology structure graph.

[0107] In this embodiment, based on the environment physical space interaction graph, an optimization algorithm in graph theory is used to optimize the graph, with the purpose of improving the expression ability of the graph and reducing the computational complexity. The key to graph optimization is to filter out the most energy-efficient device and environmental interaction relationship through algorithms (such as minimum spanning tree, maximum flow algorithm), and delete redundant low-impact nodes. For example, if the interaction between some devices and the environment is determined to have a small impact (such as being far away, having a weak impact) in the optimization process, these nodes can be deleted in the optimized graph to simplify the calculation. The optimized graph can more accurately reveal the main interaction between the device and the environment, thereby providing a basis for subsequent energy efficiency optimization decisions. Finally, an initial spatial semantic topology structure graph is generated, which can clearly express the core interaction between the devices and the environment in the building area and provide support for subsequent energy saving optimization.

[0108] Optionally, step S14 is specifically:

[0109] Step S141: Perform topology node local density calculation according to the initial spatial semantic topology structure graph, set the calculation neighborhood radius to 2, set the local density threshold to 0.1, and regard the nodes below the local density threshold as sparse areas, thereby obtaining the topology structure local density distribution;

[0110] In this embodiment, when calculating the local density of the nodes based on the initial spatial semantic topology graph, the neighborhood radius is set to 2, which means that each node in the initial spatial semantic topology graph will consider the neighborhood nodes within a distance of 2 nodes. The local density of each node is calculated by this radius, and the local density threshold is set to 0.1. Nodes below this threshold are considered sparse areas. The purpose of this step is to identify sparse areas in the topology, which represent areas with less impact on energy efficiency or lower environmental impact in the system. By calculating the local density of the nodes, we can effectively filter out areas that need to be optimized and determine areas with low density in the system, which is important for energy efficiency improvement. In actual operation, the local density can be calculated using Python graph analysis libraries such as NetworkX, considering the neighborhood information of the nodes and generating the corresponding local density distribution graph. Assuming that the initial topology graph has 1000 nodes, the local_density() function will check the 2 nodes around each node, calculate their density, and mark the nodes with a local density below 0.1 as sparse areas. For example, in a building's ventilation system, some areas have only a small number of devices, and the calculated density value is below the set threshold of 0.1. These areas are low-energy sparse areas and are worth further optimization.

[0111] Step S142: Set the minimum sample size to 5 and the density threshold to 0.3, perform node neighborhood adaptive clustering on the initial spatial semantic topology graph, and set the weight range to [0, 1] to calculate the neighborhood adaptability weight based on the clustering results, generating the neighborhood adaptive topology optimization initial parameters;

[0112] In this embodiment, the minimum sample size is set to 5, which means that the minimum number of nodes in each cluster cannot be less than 5 when clustering. The density threshold is set to 0.3, indicating that only areas with a density higher than 0.3 in the neighborhood will be considered as a cluster during clustering. Based on these settings, the density peak clustering algorithm is used to adaptively divide the topology graph, find areas with similar characteristics, and divide them into the same category. The similarity between nodes (such as Euclidean distance or Manhattan distance) is used to determine the neighborhood adaptability weight, with a range of [0, 1]. Nodes with higher weights indicate that they are more closely related to other nodes in the neighborhood, and have higher priority for optimizing energy efficiency and environmental impact. Through this step, the neighborhood adaptive topology optimization initial parameters can be obtained, providing a basis for subsequent multi-scale optimization. The clustering results can be visualized on the physical space layout of the building.

[0113] Step S143: Set the weight adjustment coefficient to 0.5, and perform neighborhood topological weight adjustment on the initial spatial semantic topology graph according to the neighborhood adaptive topology optimization initial parameter, and set the optimization scale [1, 3] to perform multi-scale topology optimization, to obtain an optimized device spatial semantic graph;

[0114] In this embodiment, in the neighborhood topological weight adjustment stage, the weight adjustment coefficient is set to 0.5, which means that the neighborhood weight of each node will be adjusted according to the influence of adjacent nodes in the optimization process. The process of neighborhood topological weight adjustment is performed by setting the optimization scale to [1, 3] to perform multi-scale topology optimization in the scale range of 1 to 3, so that the optimized graph more accurately reflects the relationship between nodes and its impact on energy efficiency. This optimization process can be performed by a weighted shortest path algorithm in the algorithm, which adjusts the weight by using the distance and relevance between nodes to reduce redundant calculations and optimize nodes with greater energy efficiency impact. The weighted shortest path algorithm can calculate the "distance" between each node, which can be defined by energy consumption, workload or other related metrics between devices. Nodes with higher relevance (such as the relationship between air conditioners and temperature and humidity sensors) will be given higher weights to optimize the energy efficiency of these nodes. This optimization method can efficiently integrate the interaction between various devices and the environment in the building, ensuring that the optimized graph more accurately reflects the key areas for energy efficiency improvement.

[0115] Step S144: Extract device operation mode features with a feature dimension of 5 based on the optimized device spatial semantic graph, and fuse the time sequence aligned environment data set to normalize the device operation mode features, to generate standardized device operation mode data;

[0116] In this embodiment, the device operation mode data is standardized, and the feature dimension is set to 5, that is, 5 important device operation features are selected, such as power consumption of air conditioning equipment, indoor temperature, humidity sensor reading, lighting equipment state and personnel density, etc. Through these features, the interaction between the running state of the equipment and the environmental variables can be understood in detail. The device operation mode features are fused with the time-aligned environmental data. This fusion process is carried out by a weighted average method based on data correlation. The running mode features of each device (such as the power consumption of the air conditioner, the on-off state of the lighting, etc.) will be given different weights according to their performance under certain environmental conditions (such as temperature and humidity changes). For example, when the indoor temperature is high, the energy consumption feature of the air conditioner will be higher, so the correlation between the air conditioner and the temperature is stronger, and the corresponding weight is also larger. In this way, the fusion of device operation mode and environmental data can accurately reflect the energy efficiency of the equipment under different environmental conditions. In the standardization process, the Z-score standardization method is used, so that the data value of each feature is converted to a distribution with a mean of 0 and a standard deviation of 1, which can avoid the bias of different dimensional features on the model. Assuming that the power consumption of the air conditioning equipment is 0.5 after standardization, it means that the equipment is running normally, while the equipment with large temperature and humidity changes has a standardization value of -1.2, indicating that the energy efficiency of the equipment is unstable. Through these standardized feature data, more accurate and consistent data can be provided for the next modeling.

[0117] Step S145: Embedding the standardized device operation mode data into the topology of the optimized device space semantic graph, modeling the interaction between device operation mode and environmental data, and obtaining a preliminary heterogeneous sensor dynamic spatio-temporal adjacency matrix;

[0118] In this embodiment, the standardized device operation mode data is embedded into the optimized device space semantic graph to model the interaction between the device and the environment. To achieve this, the device nodes and environmental variable nodes are combined together, and a graph neural network (GNN) model is used to learn the interaction between the device and the environment. Assuming that the air conditioning equipment, temperature and humidity sensor and lighting equipment are selected as the input nodes of the model. During model training, the edge weight between the air conditioning equipment node and the temperature and humidity sensor node will be adjusted according to their interaction strength. For example, if the air conditioning equipment runs in a low temperature environment, its adjustment effect on temperature is small, and the edge weight between the air conditioning node and the temperature and humidity sensor node will be small. In this way, an interaction model between the device operation mode and the environmental data can be constructed, and a preliminary heterogeneous sensor dynamic spatio-temporal adjacency matrix is generated.

[0119] Step S146: Feature screening and topological correlation optimization are performed on the preliminary heterogeneous sensor dynamic spatio-temporal adjacency matrix to strengthen key interaction relationships and remove low correlation edge weights, thereby generating a heterogeneous sensor dynamic spatio-temporal adjacency matrix.

[0120] In this embodiment, in the feature screening and topological correlation optimization stage, a graph analysis method is used to optimize the preliminary heterogeneous sensor dynamic spatio-temporal adjacency matrix. By removing low correlation edge weights, the key interaction relationships between nodes are strengthened, and the final heterogeneous sensor dynamic spatio-temporal adjacency matrix is generated. This step uses a feature selection algorithm (such as the L1 regularization method or principal component analysis PCA) to screen the features of the matrix, remove nodes and edges that have less impact on energy efficiency optimization, and thereby reduce the computational complexity. The final heterogeneous sensor dynamic spatio-temporal adjacency matrix will more accurately reflect the key interaction relationships between devices and the environment within the building, providing more effective data support for subsequent energy efficiency optimization. This process helps to extract the most influential device and environment relationships, ensuring that optimization work is focused on the most critical interaction areas.

[0121] Optionally, step S2 is specifically:

[0122] Step S21: Cross-energy consumption influence factor screening is performed on the sensor spatio-temporal fusion feature set, a correlation threshold of [0.7, 0.8] is set to retain high correlation features, energy consumption influence weights of each device are calculated, and an initial energy consumption influence feature matrix is constructed based on the energy consumption influence weights;

[0123] In this embodiment, the energy consumption influence factor screening is performed on the sensor spatio-temporal fusion feature set, the purpose of which is to select the most influential part of the energy consumption from a large number of features. For this purpose, a correlation threshold range of [0.7, 0.8] is set, and correlation analysis is performed on the spatio-temporal feature data of all devices. Only when the correlation of the features of two devices exceeds 0.7 and is less than 0.8, these features are retained. For example, the power consumption of an air conditioning device may have a strong correlation with indoor temperature and humidity, and such high correlation features are selected for retention. Next, the energy consumption influence weights of each device are calculated based on the selected features, and the higher the weight, the greater the influence of the device on the overall energy consumption. Specifically, the weight calculation uses weighted regression analysis based on the energy efficiency performance of each device in historical operation and the correlation with environmental factors. After feature screening and weight calculation, an initial energy consumption influence feature matrix is constructed based on the weights. This matrix includes the energy efficiency influence degree of each device and their mutual relationships.

[0124] Step S22: Non-Euclidean graph topological mapping is performed on the initial energy consumption influence feature matrix, a maximum connection threshold of 0.5 is set for the adjacency matrix, an energy consumption correlation topological structure across devices and spatial regions is constructed, and an initial cross-energy consumption correlation topological graph is generated;

[0125] In this embodiment, the initial energy consumption influence feature matrix is mapped to a non-Euclidean graph topology. The "non-Euclidean graph topology" here means that instead of using the traditional Euclidean space, the devices are connected according to their energy consumption correlation and spatial layout. For example, if the energy efficiency correlation between the air conditioning device and the temperature and humidity sensor device is greater than 0.5, a connection between the two devices will be created in the topology graph. In order to describe the complex energy consumption relationship between devices, the maximum connection threshold of the adjacency matrix is set to 0.5. When the energy consumption influence degree of two devices is greater than 0.5, they are considered to have a strong correlation, and an edge is created in the graph. For example, assuming that the air conditioning device and the temperature and humidity sensor have a strong correlation in energy consumption (correlation coefficient is 0.6), while the correlation between the air conditioner and the light sensor is weak (correlation coefficient is 0.4), only the connection between the air conditioner and the temperature and humidity sensor will be formed, and the connection with the light sensor will be ignored. In this way, the energy consumption data of different devices and space areas is converted into a graph structure, with devices as nodes and energy consumption transmission relationships as edges, and a preliminary cross-device and cross-space area energy consumption correlation topology graph is constructed. This provides a basis for the next step of spectral clustering analysis.

[0126] Step S23: Perform spectral clustering analysis on the initial cross-energy consumption correlation topology graph, select the top 20 eigenvectors for Laplace eigenvalue decomposition, and set the feature truncation threshold to 0.1 to extract the main energy consumption influence mode, and generate a cross-energy consumption correlation topology graph;

[0127] In this embodiment, when performing spectral clustering analysis on the initial cross-energy consumption correlation topology graph, the top 20 eigenvectors are selected for Laplace eigenvalue decomposition. The purpose of this process is to extract the most representative energy consumption influence mode. The selection of eigenvectors is based on the significance of the energy consumption mode, and the top 20 eigenvectors represent the areas and devices with the greatest energy consumption influence. In this process, the feature truncation threshold is set to 0.1, which means that when the contribution of some features is less than 0.1, these features will be excluded. In this way, noise and irrelevant features can be removed, and only the energy consumption influence mode that is most helpful for energy efficiency optimization is retained. The generated cross-energy consumption correlation topology graph can more accurately reflect the energy efficiency interdependence relationship between devices in the building, and provide accurate data support for subsequent optimization steps.

[0128] Step S24: Model the energy consumption transmission relationship between devices based on the cross-energy consumption correlation topology graph, set the energy consumption transmission rate to [0.05, 0.2], and set the dynamic weight change interval to [0, 1] to assign a dynamic weight, and generate an initial dynamic energy consumption influence factor matrix;

[0129] In this embodiment, the energy consumption transmission relationship between devices is modeled using a cross-energy consumption correlation topology graph. By setting the energy consumption transmission rate to [0.05, 0.2], the transmission efficiency of energy consumption between different devices can be simulated. For example, the air conditioning device directly affects the energy consumption of the temperature and humidity sensor, while the temperature and humidity sensor affects the operating load of the air conditioner. By setting the range of energy consumption transmission rate to 0.05 to 0.2, different situations of this transmission process can be simulated. Then, the dynamic weight variation interval is set to [0, 1], which enables the energy consumption transmission relationship to be dynamically adjusted under different working conditions. For example, when the external temperature rises sharply, the energy consumption influence weight of the air conditioner will increase, while the influence of other devices such as lighting will decrease. Through this dynamic weight adjustment mechanism, a more accurate dynamic energy consumption influence factor matrix can be obtained, further reflecting the energy consumption transmission relationship between devices.

[0130] Step S25: Normalize the initial dynamic energy consumption influence factor matrix, set the influence factor confidence interval to [0.95, 1.05] to correct the influence factor robustness and optimize the energy consumption influence parameter, and output the dynamic energy consumption influence factor matrix.

[0131] In this embodiment, the initial dynamic energy consumption influence factor matrix is normalized to ensure that all influence factors are within a reasonable range. For example, the confidence interval is set to [0.95, 1.05] to ensure that the fluctuation of the energy consumption influence factor within this interval does not exceed the predetermined value. If the energy efficiency factor of a device exceeds this range, for example, there is an abnormal energy efficiency change (such as a significant increase), the energy efficiency of the device is considered to be abnormal and further adjustment is needed (including data smoothing, recalibration, and abnormal data rejection or correction). For example, there is a sudden change in environmental factors such as external temperature and humidity, resulting in a sharp change in device energy efficiency. In this case, the environmental parameters need to be compensated to ensure the stability and accuracy of the model. The sensor itself may malfunction or the data input may be biased, resulting in inaccurate energy consumption data. In this case, the energy consumption influence factor will lose credibility and needs to be adjusted through data cleaning or recalibration of the sensor. For example, if the energy consumption factor of a device exceeds 1.05 at a certain moment, it means that the energy consumption performance of the device is much higher than the normal range, which may be caused by device failure, external environmental change, or decrease in device operating efficiency. If it is lower than 0.95, it may indicate that the energy consumption performance of the device is abnormally low, which may be caused by improper operation of the device or failure of some sensors, and a warning needs to be issued for the faulty device. Through normalization and adjustment of the influence factor, the robustness of the energy consumption matrix can be enhanced, making it more stable and accurate, and providing reliable data support for optimization decisions. Ultimately, through this optimized matrix, the system can more efficiently allocate resources, improve overall energy efficiency, and ensure the stability and energy-saving effect of system operation under different environmental conditions and device loads.

[0132] Optionally, the multi-scale energy consumption evolution deduction in step S3 is specifically:

[0133] Time series decomposition is performed on the dynamic energy consumption influence factor matrix, multi-scale time steps are set, trend decomposition and periodic decomposition are performed, and an energy consumption time series decomposition dataset is generated;

[0134] In this embodiment, time series decomposition is performed on the dynamic energy consumption influence factor matrix, and the purpose is to extract energy consumption influence patterns at different time scales. For this purpose, multi-scale time steps (such as 5 minutes, 1 hour, 12 hours and 24 hours) are selected for decomposition. These time scales can adapt to different granularity energy consumption change requirements. For example, 5-minute steps are suitable for capturing minor fluctuations in device operation, and 24-hour steps are used to analyze daily energy consumption change patterns. The Trend Decomposition method is used to remove long-term trends and extract periodic energy consumption changes. Seasonal Decomposition helps identify daily or seasonal energy consumption patterns and extract periodic fluctuations. Finally, the obtained decomposition dataset is used for subsequent analysis, such as energy consumption propagation path causal inference.

[0135] Causal inference is performed on the energy consumption time series decomposition dataset, the energy consumption correlation strength between devices is calculated, and an energy consumption propagation network is constructed to generate an initial energy consumption propagation structure;

[0136] In this embodiment, causal inference analysis is performed on the energy consumption time series decomposition dataset. By using the Granger Causality Test, it can be detected whether there is a causal propagation relationship of energy between devices. For example, if the energy consumption fluctuation of device A leads to the energy consumption fluctuation of device B, it can be considered that device A has a causal impact on the energy consumption of device B. Based on this causal inference result, the energy consumption correlation strength between each pair of devices is calculated (such as by measuring the Pearson correlation coefficient). Then, these correlation strengths are converted into an energy consumption propagation network, and by setting a threshold (such as 0.7), edges with a strength lower than the threshold are removed, thereby constructing a preliminary energy consumption propagation network structure.

[0137] Based on the initial energy consumption propagation structure, energy consumption propagation pattern clustering is performed, energy consumption propagation paths are screened, low-contribution energy consumption paths are removed, and an optimized energy consumption propagation network is generated;

[0138] In this embodiment, the initial structure of energy consumption propagation is used for pattern clustering analysis. The spectral clustering algorithm is used to cluster the energy consumption propagation paths to identify groups of devices with similar propagation characteristics. By analyzing the clustering results, high-contribution propagation paths are selected. For example, if the energy consumption fluctuations of some paths can significantly affect the overall energy efficiency level, they will be retained; while for those with less contribution and weak propagation effect, they will be removed. After this optimization step, a simplified and efficient energy consumption propagation network is obtained, which can more accurately reflect the energy flow patterns between devices.

[0139] Time series feature extraction is performed on the optimized energy consumption propagation network, energy consumption time series segments are divided, and empirical mode trend decomposition is performed on the energy consumption time series segments to obtain energy consumption time series feature data;

[0140] In this embodiment, the optimized energy consumption propagation network enters the time series feature extraction stage. First, the energy consumption time series segments are divided according to time periods, such as daily, weekly, and monthly data. Then, the empirical mode decomposition (EMD) method is used to decompose the energy consumption data of each time period to extract different feature modes (such as trend, periodic component, noise, etc.) of the time series data. This method can adaptively extract key features in the signal without relying on traditional statistical models. Each time series segment is decomposed into multiple intrinsic mode functions (IMF), providing rich time series feature data for subsequent short-term and long-term energy consumption prediction.

[0141] Based on the energy consumption time series feature data, autoregressive integral short-term energy consumption trend modeling is performed to generate short-term energy consumption prediction data;

[0142] In this embodiment, based on the energy consumption time series feature data obtained by empirical mode decomposition, the autoregressive integral moving average (ARIMA) model is used for short-term energy consumption trend modeling. This model can model the trend, seasonality, and randomness of historical data to generate short-term energy consumption prediction. The prediction time range is set to 7 days, and the parameters (such as AR order, MA order, etc.) in the ARIMA model are further adjusted to make the prediction results more accurate. The generated short-term energy consumption prediction data will help decision-makers in daily energy management and scheduling.

[0143] Residual analysis is performed on the short-term energy consumption prediction data, and the residual analysis results are input into a pre-set long short-term memory network for long-term energy consumption evolution modeling to obtain long-term energy consumption prediction data;

[0144] In this embodiment, residual analysis is performed to compare the differences between actual energy consumption and predicted energy consumption. Residual analysis can help identify potential biases or error sources in the prediction model. If the residual is large, it indicates that the model needs further adjustment. Then, the residual data is input into a pre-set long short-term memory network (LSTM) for long-term energy consumption evolution modeling. The LSTM network has memory and can effectively capture long-term dependencies to generate long-term energy consumption prediction data. This data can be used for monthly, quarterly or annual energy budgeting and scheduling.

[0145] The long-term energy consumption prediction data and the short-term energy consumption prediction data are fused to construct a comprehensive energy consumption prediction sequence; and multi-resolution wavelet transform is performed on the comprehensive energy consumption prediction sequence to extract energy consumption change patterns at different time scales and generate a multi-scale energy consumption evolution prediction matrix.

[0146] In this embodiment, long-term energy consumption prediction data and short-term energy consumption prediction data are fused to construct a comprehensive energy consumption prediction sequence. Weighted average method or other fusion algorithms are used to assign different weights to short-term and long-term data according to the length of the prediction. For example, the weight of short-term prediction can be set to 0.7 and the weight of long-term prediction can be set to 0.3. The specific weights can be adjusted according to actual application requirements. Ultimately, the fused energy consumption prediction sequence can provide more accurate and comprehensive energy consumption prediction to provide a basis for energy efficiency management. Multi-resolution wavelet transform is performed on the comprehensive energy consumption prediction sequence to extract energy consumption change patterns at different time scales. Wavelet transform can analyze data at multiple levels and frequencies, helping to identify short-term fluctuations and long-term change patterns. After wavelet transform, a multi-scale energy consumption evolution prediction matrix is generated, which contains energy consumption patterns at different time scales. This multi-scale analysis method can provide more accurate basis for energy consumption scheduling, load balancing and resource optimization.

[0147] Based on the multi-scale energy consumption evolution prediction matrix, a state probability model is established, a state transition matrix is calculated, and the state transition matrix is combined with the building energy consumption evolution map to obtain the building energy consumption evolution map.

[0148] In this embodiment, based on the multi-scale energy consumption evolution prediction matrix, a state probability model is established. The transition probability of the energy consumption state is calculated by Markov chain and other methods to obtain the energy consumption state transition matrix. This matrix can describe the transition rules and probabilities between different energy consumption states. By combining the energy consumption state transition matrix with the building energy consumption evolution map, a map that comprehensively reflects the building energy efficiency evolution is obtained, which can play a key role in building energy efficiency management and optimization, and support energy use prediction and decision-making.

[0149] Optionally, the long-term energy consumption evolution modeling specifically comprises:

[0150] The short-term energy consumption prediction data is time step aligned with the residual analysis result to obtain a time alignment residual data set;

[0151] In this embodiment, the short-term energy consumption prediction data is time step aligned with the residual analysis result, which requires ensuring consistency of the two data in the time dimension. Usually, interpolation or resampling techniques can be used to complete the alignment. A suitable time step (e.g., 1 hour or 30 minutes) is selected for data alignment to ensure consistency of the time interval of the data and to ensure that the time step does not have a significant impact on the prediction result.

[0152] Local spatio-temporal feature extraction is performed on the time alignment residual data set to generate residual feature data;

[0153] In this embodiment, the time alignment residual data set is processed by a local spatio-temporal feature extraction method. A convolutional neural network (CNN) is used to extract local features from the data, including short-term energy consumption trend and long-term and short-term time dependence. In this process, a sliding window with a window size of 30 time steps (e.g., 30 hours) is used for local feature extraction to capture the pattern of energy consumption fluctuations.

[0154] Based on the residual feature data set, a long short-term memory network model is constructed;

[0155] In this embodiment, based on the extracted residual feature data set, a long short-term memory (LSTM) network model is constructed to learn the time sequence dependence of the data. The LSTM network has 3 layers, the number of neurons in the hidden layer is 256, the optimization algorithm uses the Adam optimizer, and the learning rate is set to 0.001. Through training of historical data, the LSTM model can capture the time sequence rule of energy consumption data, thereby providing accurate modeling for long-term energy consumption evolution.

[0156] The long short-term memory network model is backpropagated, the gradient descent method is used for network parameter optimization, the time alignment residual data set is modeled for long-term energy consumption evolution, and a long-term energy consumption evolution model is obtained;

[0157] In this embodiment, in the backpropagation phase, the model parameter optimization is performed by using the gradient descent method. The batch training method can be used, the batch size is set to 64, and the network parameters are updated by the backpropagation algorithm to make the model more accurately fit the training data and reduce errors. In this process, the model is trained for 50 iterations to ensure the accuracy of the long-term energy consumption evolution.

[0158] The long-term energy consumption evolution model is used to perform long-term energy consumption prediction on the residual feature data to obtain long-term energy consumption prediction data.

[0159] In this embodiment, the trained long-term energy consumption evolution model will be used to perform long-term energy consumption prediction on the residual feature data. By inputting the time-aligned residual feature data, the model can output energy consumption prediction results for a certain number of future time steps (e.g. 72 hours or 168 hours). The prediction process not only relies on the learning of the training data, but also dynamically adjusts the prediction window and the parameters of the model to provide a dynamic evolution trend of long-term energy consumption, thereby providing data support for energy-saving optimization of building equipment.

[0160] Optionally, the generating of the energy-saving optimization instruction set in step S4 is specifically:

[0161] Performing device category strategy adaptability analysis based on the multi-objective energy-saving control strategy manifold, and performing device control demand hierarchical clustering to generate a device control classification index;

[0162] In this embodiment, by analyzing the functional characteristics of various devices in the building (such as HVAC systems, lighting, elevators, smart sockets, etc.), important parameters such as device operating mode, energy consumption characteristics, and load response capability are extracted. Specifically, the characteristics of air conditioning equipment include its temperature adjustment range, power consumption, response speed, etc.; the characteristics of lighting equipment include brightness adjustment range, power consumption, and light intensity change response capability; and elevators have different operating modes, motorized window blinds and fans have control requirements for automatic adjustment based on time or sensor data. Through data collection and analysis, a feature vector is formed for each device. Taking air conditioning as an example, its feature vector may include "maximum temperature adjustment range: ±2℃", "maximum power consumption: 3kW", "adjustment response time: 10 minutes", etc. The strategy manifold can be regarded as a high-dimensional embedding space, where each strategy point represents an energy-saving control strategy. To evaluate the adaptability of each device to the energy-saving strategy, the feature vector of the device is compared with the strategy points in the manifold, and the similarity between the device features and the strategy points is calculated. Common similarity evaluation methods include cosine similarity, Euclidean distance, and KL divergence, etc. For example, the temperature control characteristics of the air conditioning equipment can be compared with the strategy points in the strategy manifold that represent temperature control optimization strategies, and a high similarity indicates that the device is suitable for the strategy. For lighting devices, the matching degree of their brightness adjustment capability with the light management strategies in the manifold is calculated. After evaluating the adaptability of devices to energy-saving strategies, the next step is to group all devices in the building and classify them based on their adaptability to energy-saving strategies. For this purpose, hierarchical clustering algorithms (such as K-means or DBSCAN) are used to group devices. Devices will be classified into three categories: high-priority, medium-priority, and low-priority devices, based on their energy efficiency impact, control requirements, and operating mode. For example, air conditioning, lighting, and electric heating devices usually belong to the high-priority device group because they have a greater impact on building energy efficiency and require precise control; while motorized window blinds and fans belong to the low-priority device group because they have a relatively small impact on energy efficiency. In device classification, not only the load response capability and control requirements of the device are considered, but also the energy efficiency boundary parameters of the device are introduced to ensure that the control strategy of the device matches its actual capability. As the building environment and device operating state change, the load capacity, control time window, and environmental sensitivity of the device may change, so the device classification needs to be dynamically adjusted. For example, the adjustment capability of an air conditioning device within the temperature control range may be affected by external temperature changes, so it may need to be adjusted to a high-priority device for real-time energy-saving optimization in certain situations. To ensure the accuracy of the analysis, energy efficiency boundary parameters are set for devices, such as the temperature adjustment range of an air conditioning device (±1℃) and the maximum brightness adjustment range of a lighting device (±20%), to ensure that the classification of each device is consistent with its actual control capability. Through the above steps, a device control classification index is obtained.The index contains key information such as the priority of each device, energy efficiency impact, and control requirements, and can be dynamically adjusted as needed. The device control classification index will serve as the basis for subsequent energy-saving strategy constraint mapping, ensuring that devices can be accurately adapted when implementing energy-saving strategies. When generating the device control classification index, in addition to considering the control characteristics and energy efficiency impact of the device, the actual operating state of the device during building operation is also considered to improve the execution accuracy and effectiveness of the energy-saving strategy.

[0163] According to the device control classification index, the multi-objective energy-saving control strategy manifold is mapped to obtain a set of constraint optimization strategies;

[0164] In this embodiment, the specific control requirements of the devices in the device control classification index are mapped to the energy-saving control strategy. First, through the device category index, the energy efficiency requirements of the devices are mapped to the appropriate energy-saving strategy manifold, ensuring that each device can reduce energy consumption to the maximum extent under a specific strategy. For example, air conditioning devices will adopt the "temperature rise limit strategy", which sets the indoor temperature within a relatively stable range to avoid frequent on-off caused by excessive temperature fluctuations, thereby reducing energy consumption; while for lighting devices, the "natural light sensing dimming strategy" is adopted, which adjusts the indoor lighting brightness in real time based on external light intensity. In this process, the strategy is further refined according to the environmental requirements of the building, load distribution, and comfort requirements, such as setting the temperature control range of air conditioning devices in summer to 23-26°C and in winter to 18-22°C, and adjusting the indoor brightness of lighting devices according to external light intensity to between 200 and 500 lux. In addition, other important constraints such as the maximum delay of control time not exceeding 5 minutes are also set to ensure the real-time execution of the strategy.

[0165] The set of constraint optimization strategies is optimized in time series to adjust the execution time of the strategy, generating a set of time-optimized strategies;

[0166] In this embodiment, time series scheduling optimization is performed according to the device type and its control constraints. For example, air conditioning devices usually have high load during peak hours in the daytime, and less demand during off-peak hours at night. By setting a time window, the running time of the device is optimized and adjusted. In the optimization process, based on historical data, environmental data (such as temperature changes, personnel activities) and prediction models (such as ARIMA model), the start time and running period of air conditioning and lighting devices are dynamically adjusted. For example, it is determined that the air conditioning device will be adjusted to low power (such as from 20℃ to 24℃) during the idle period at night, and will be adjusted to high power to maintain temperature stability during the daytime high temperature period. In addition, the lighting device will also be adjusted in real time according to the change of natural light, and will not waste energy when the light is sufficient. Through this time scheduling optimization, the peak and off-peak load of building energy consumption can be effectively allocated, and the start and stop time of the device can be dynamically adjusted according to the actual demand, so as to optimize the overall energy efficiency.

[0167] The time optimization strategy set is distributed to the strategy execution priority, and the strategy execution parameter is adaptively adjusted based on the time sequence alignment environment data set to generate an adaptive energy-saving control instruction set.

[0168] In this embodiment, based on the time optimization strategy set obtained by the foregoing optimization, the execution priority of each device in each period is further analyzed. First, considering the energy efficiency influence of the device, the device priority and operation mode already allocated in the device control classification index, for example, the air conditioning device is usually allocated as a high priority device due to its large energy efficiency influence and importance to indoor comfort, while the lighting device, heater and the like are less important, and the fan, smart socket and the like are usually allocated a lower priority. The operation mode of the device also affects its priority, for example, the air conditioning device will be temporarily adjusted to high priority during the high temperature period in summer due to the need to frequently adjust the temperature, and can be adjusted to medium priority during the low temperature period to save energy. During specific execution, the execution parameter of the device is adaptively adjusted according to the time sequence alignment environment data set (including indoor and outdoor temperature and humidity, personnel activity and external climate change). For example, if the system detects that the indoor temperature is high and there are more personnel activities, the running power of the air conditioner will be immediately increased. At the same time, the system will combine the natural light sensing function of the lighting device, and if the external light intensity is sufficient, the lighting brightness will be reduced to reduce unnecessary energy waste. In addition, based on the dynamic device control priority, the system will prioritize adjusting those devices with large energy efficiency influence, such as air conditioning and lighting devices, while ensuring comfort, to maximize energy efficiency.

[0169] The adaptive energy-saving control instruction set is verified for consistency, the execution feasibility of the adaptive energy-saving control instruction set is verified, an energy-saving optimization instruction set is obtained, and the energy-saving optimization instruction set is uploaded to the building energy consumption control platform to perform the device control task.

[0170] In this embodiment, the consistency of the adaptive energy-saving control instruction set is verified to ensure the feasibility and reliability of the execution of each instruction. First, all generated energy-saving control instructions are checked by preset rules to ensure that each instruction meets the operating range and environmental constraints of the device. For example, the temperature adjustment instruction of the air conditioning device will not exceed the set operating temperature range (such as between 18°C and 26°C), and the brightness adjustment of the lighting device will also be adjusted according to the real-time changes of external light intensity. The system will detect and check whether the control instructions can be effectively executed in real time based on historical operation data, current building energy efficiency, and environmental feedback, to avoid problems such as overwork or underwork of the device. For any abnormal instruction execution, the system will automatically adjust and correct the strategy parameters based on feedback to ensure that the executed energy-saving instructions meet the actual needs. Finally, the energy-saving optimization instruction set verified for consistency will be uploaded to the building energy consumption control platform to start executing specific device control tasks, ensuring that the building maintains comfort while saving energy.

[0171] Optionally, the device predicted energy consumption deviation bloodline tracking specifically comprises:

[0172] The historical energy consumption data of the building device is obtained, and a baseline energy consumption prediction model is established based on the historical energy consumption data of the building device and the energy-saving optimization instruction set, so as to obtain device baseline energy consumption prediction data;

[0173] In this embodiment, the building energy consumption control platform is used to collect data and obtain historical energy consumption data. These historical data usually include the power consumption, operation period, external environment (such as temperature and humidity, light intensity), and load state of each device in different time periods. The energy consumption data of the device is collected in real time by Internet of Things sensors and intelligent metering devices (such as electricity meters, temperature control sensors, etc.), and stored in the cloud database of the building energy consumption control platform. Then, combined with the historical energy consumption data of the device and the energy-saving optimization instruction set (such as temperature control range, lighting brightness control, etc.), a baseline energy consumption prediction model is constructed. This model can be trained using machine learning algorithms (such as support vector machine regression, decision tree regression, etc.) to predict the baseline energy consumption of each device in future time periods based on historical data. During the training process of the model, multi-dimensional features such as environmental temperature and humidity, device load, etc. are used for data feature engineering to optimize the prediction accuracy of the model. Finally, device baseline energy consumption prediction data based on device characteristics, historical energy consumption data, and energy-saving instructions are obtained.

[0174] Real-time building energy consumption data is obtained, error comparison is performed on the device baseline energy consumption prediction data and the real-time building energy consumption data, the predicted energy consumption deviation metric is calculated, and preliminary screening of abnormal energy consumption is performed based on the preset deviation prediction, to obtain a device energy consumption deviation metric matrix;

[0175] In this embodiment, real-time energy consumption data is obtained through the building energy consumption control platform. These data are fed back to the cloud platform or local control system in real time through intelligent metering systems (such as smart meters, temperature and humidity sensors, etc.). Real-time energy consumption data usually includes the current power consumption value of each device, load changes, and real-time data of external environment (such as air temperature, humidity, illumination, etc.). Then, the real-time energy consumption data is compared with the previously established device benchmark energy consumption prediction data, the error is calculated, and analysis is performed based on the deviation. Common methods for calculating error include mean square error (MSE), mean absolute error (MAE), etc. Through the deviation metric calculated, the difference between the actual energy consumption of the device and the predicted value can be judged, and then the devices with large energy consumption deviation are screened out to form a device energy consumption deviation metric matrix. This matrix contains the difference between the predicted energy consumption and the actual energy consumption of each device at each time, the error size and its relative deviation. For devices with deviation exceeding a set threshold (such as 10%), preliminary screening of abnormal energy consumption is performed, which serves as the basis for further diagnosis.

[0176] Obtain device control logs and construct a device energy consumption change causal relationship diagram based on the device control logs and the dynamic energy consumption influence factor matrix, perform energy consumption influence factor causal inspection, and generate a device energy consumption bloodline tracking chain;

[0177] In this embodiment, the control logs of the devices are obtained through the building energy consumption control platform, which record the running state of the devices, operation instructions (such as power on / off, temperature adjustment, brightness adjustment, etc.) and the influence of environmental variables (such as external air temperature, humidity, etc.) on the devices. By analyzing the device control logs, the control behavior of each device at different time periods can be understood, and then combined with the historical energy consumption data of the devices and the current environmental data, a device energy consumption change causal relationship diagram is constructed. This diagram models the causal relationship between the control operation of the device and its energy consumption change, environmental factors, etc. To ensure the accuracy of the model, a dynamic energy consumption influence factor matrix is used, which calculates the influence degree of different control factors on the energy consumption of the device by analyzing the energy consumption fluctuations of the device under various environmental conditions. Through inspection of these causal relationships and using causal inference methods (such as Granger causality test, Bayesian network, etc.), the causal relationship of each node in the relationship diagram is ensured to be accurate. Finally, a device energy consumption bloodline tracking chain is generated, which records the energy consumption change path and influencing factors of each device. For example, the energy consumption change of an air conditioning device may be closely related to the changes of external air temperature, room occupancy, temperature control settings, etc. Through the bloodline tracking chain, the root cause of the energy consumption change can be traced back.

[0178] Combine the device bloodline tracking chain with the device energy consumption deviation metric matrix to trace the energy consumption anomaly, and based on the results of the energy consumption anomaly tracing, classify the abnormal energy consumption patterns of the devices to obtain an energy consumption deviation backtracking analysis report.

[0179] In this embodiment, the energy consumption bloodline tracking chain of the device is combined with the device energy consumption deviation metric matrix to trace the energy consumption anomaly. Through the bloodline tracking chain, the root cause of the energy consumption deviation of each device can be determined, and the factors causing the deviation can be found, such as abnormal control instructions, changes in external environment, aging of devices, etc. By tracing each device with a large deviation in the deviation metric matrix one by one, the pattern of energy consumption change is analyzed to further determine whether there is an abnormal energy consumption behavior. According to the tracing result, the devices are divided into different energy consumption anomaly patterns. For example, the air conditioning device may deviate from the energy consumption due to the sudden change of external temperature, while the lighting device may increase unnecessary energy consumption due to the failure of the automatic control system. Finally, an energy consumption deviation backtracking analysis report is generated, which includes the energy consumption deviation of each device, abnormal reason analysis, abnormal energy consumption pattern classification and suggested improvement measures, providing support for subsequent energy saving optimization and device management.

[0180] Optionally, the establishing a baseline energy consumption prediction model specifically comprises

[0181] The historical energy consumption data of the building equipment is obtained, and the historical energy consumption data of the building equipment is time-aligned to obtain a standardized device historical energy consumption dataset;

[0182] In this embodiment, the historical energy consumption data of the building equipment is obtained through the building energy consumption control platform. These data include energy consumption records of the equipment in different time periods, such as hourly or minute energy consumption values. After obtaining, the data needs to be time-aligned, which is usually completed by resampling or interpolation of data from different time sources. Specifically, a uniform time step can be set, such as 1 hour, and linear interpolation method is applied to align the data with unequal time intervals to the uniform time point. After alignment, all data is standardized, and the zero-mean unit-variance standardization method is adopted to make the mean value of the data 0 and the standard deviation 1, ensuring the comparability and stability of the data in subsequent modeling, to obtain a standardized device historical energy consumption dataset.

[0183] Based on the energy saving optimization instruction set and the dynamic energy consumption influence factor matrix, the energy saving optimization influence of the device historical energy consumption data is quantified to generate an energy saving optimization influence factor dataset;

[0184] In this embodiment, the energy-saving optimization instruction set includes a series of energy-saving measures in the operation process of the device, which usually include adjusting the temperature setting value of the air conditioner, changing the brightness of the light, adjusting the working mode of the heater, etc. Each energy-saving operation can affect the energy consumption of the device by changing the running state of the device or the environmental conditions. Therefore, it is necessary to analyze and quantify the specific influence of these factors on the energy consumption of the device according to the influencing factors of each energy-saving optimization instruction (such as environmental temperature, humidity, device load, working time, etc.). Based on the dynamic energy consumption influence factor matrix, weighted average method or regression analysis method can be used for quantitative analysis of energy-saving optimization instructions. Specifically, through the regression model or weighted average model, the influence factors corresponding to each operation in the energy-saving optimization instruction set are combined with the factors in the dynamic energy consumption influence factor matrix to quantify the influence degree of each energy-saving operation on the energy consumption of the device. For example, if a certain energy-saving operation instruction adjusts the air conditioner temperature setting value, and according to the historical data, the influence coefficient of temperature change on energy consumption is 0.15, then the energy-saving effect of the instruction can be calculated by multiplying this coefficient by the adjusted temperature change amplitude, and finally the energy-saving optimization influence factor is obtained. Finally, through the above calculation process, the influence of each energy-saving optimization instruction on the energy consumption of the device is quantified as a numerical value, and an energy-saving optimization influence factor data set is generated. The data set records the quantitative relationship between each energy-saving optimization instruction and the energy consumption of the device.

[0185] Based on the standardized device historical energy consumption data set and the energy-saving optimization influence factor data set, a baseline energy consumption prediction model of the device is obtained by a preset regression algorithm.

[0186] In this embodiment, based on the standardized device historical energy consumption data set and the energy-saving optimization influence factor data set, a baseline energy consumption prediction model is obtained by a preset regression algorithm. Support vector regression (SVR) algorithm is used as the regression model. In specific operation, a suitable kernel function (such as RBF kernel) is selected, and the parameters C and ε are set. In the training process, the standardized device historical energy consumption data and the energy-saving optimization influence factor data are used as inputs to train the model to predict the baseline energy consumption of the device without energy-saving optimization intervention. The SVR model generates a device baseline energy consumption prediction model by minimizing the error function, which can accurately reflect the normal energy consumption trend of the device.

[0187] The device baseline energy consumption prediction model is cross-validated, and the model parameters are iteratively adjusted to obtain the baseline energy consumption prediction model.

[0188] In this embodiment, the k-fold cross-validation method is adopted, and k=5 is set. The energy consumption data of the device is divided into 5 subsets, and each subset is used as the validation set in turn, and the remaining subsets are used as the training set for model training and validation. After each validation, the prediction error is calculated, and the parameters of the SVR model, especially the C value and the epsilon value, are adjusted according to the error to optimize the generalization ability and accuracy of the model. Through multiple rounds of cross-validation, the optimal device baseline energy consumption prediction model is finally obtained.

[0189] The energy-saving optimization instruction set is input into the baseline energy consumption prediction model to obtain device baseline energy consumption prediction data.

[0190] In this embodiment, the energy-saving optimization instruction set is input into the baseline energy consumption prediction model to obtain device baseline energy consumption prediction data. At this time, the energy-saving optimization instruction set contains operation instructions for adjusting the operation mode of the device. By inputting these instructions into the obtained baseline energy consumption prediction model, the energy consumption prediction data of the device after executing these energy-saving optimization instructions is generated. For example, if the energy-saving instruction set contains instructions to raise or lower the temperature of the air conditioner, the model will calculate the energy consumption change of the device under this condition, and finally obtain the optimized device baseline energy consumption prediction data. These data will be used for subsequent energy efficiency evaluation and optimization decision-making.

[0191] Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, the scope of the present application being defined by the attached claims and not by the above description, therefore all variations falling within the meaning and scope of the equivalent elements of the application file are intended to be included in the present application.

[0192] The above description is only a specific embodiment of the present application, enabling those skilled in the art to understand or implement the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A building energy consumption optimization method based on artificial intelligence, characterized in that: The following steps are involved: Step S1: Acquire multimodal sensor data and construct a heterogeneous sensor dynamic spatiotemporal adjacency matrix; perform device physical space semantic mapping modeling based on the heterogeneous sensor dynamic spatiotemporal adjacency matrix to obtain a sensor spatiotemporal fusion feature set; Step S2: Perform cross-domain energy consumption impact mapping topology modeling on the sensing spatiotemporal fusion feature set to obtain a cross-domain energy consumption correlation topology map; extract the device energy transfer impact factor based on the cross-domain energy consumption correlation topology map to obtain a dynamic energy consumption impact factor matrix; Step S3: Perform multi-scale energy consumption evolution deduction based on the dynamic energy consumption influencing factor matrix to generate a building energy consumption evolution map; construct an energy consumption probability cloud map based on the building energy consumption evolution map; Step S4: performing adversarial strategy optimization on the energy consumption probability cloud map to obtain a multi-objective energy-saving control strategy manifold; generating an energy-saving optimization instruction set based on the multi-objective energy-saving control strategy manifold and uploading it to the building energy consumption control platform to control the equipment; Step S5: Acquire real-time building energy consumption data, perform equipment predicted energy consumption deviation lineage tracking on the real-time building energy consumption data according to the energy-saving optimization instruction set, implement equipment control local comfort instruction optimization, obtain personalized control instruction set, and upload it to the building energy consumption control platform to control the equipment.

2. The method for optimizing building energy consumption based on artificial intelligence according to claim 1, characterized in that: Step S1 is specifically as follows: Step S11: acquiring multimodal sensing data through a multimodal sensor group, wherein the multimodal sensing data includes temperature and humidity sensing data, light sensing data, crowd density infrared sensing data, and equipment status sensing data; Step S12: performing time step timing alignment on the multimodal sensing data to obtain a time-series aligned environment data set; Step S13: performing non-Euclidean topological mapping on the temporally aligned environment dataset, and constructing an initial spatial semantic topological structure graph based on the topological mapping result; Step S14: performing neighborhood adaptive topology optimization based on the initial spatial semantic topology graph to obtain a device spatial semantic graph, and fusing the device operation mode and environmental data based on the device spatial semantic graph to obtain a heterogeneous sensor dynamic spatiotemporal adjacency matrix; Step S15: Establish a device physical space semantic mapping model based on the heterogeneous sensor dynamic spatiotemporal adjacency matrix, perform feature screening and association optimization on the model establishment result, and obtain a sensor spatiotemporal fusion feature set.

3. The method for optimizing building energy consumption based on artificial intelligence according to claim 2, characterized in that: Step S13 is specifically as follows: Step S131: Acquire the building area equipment operation data, and synchronize the building area equipment operation data and the timing alignment environment data set to obtain the timing aligned equipment operation data; Step S132: Perform non-Euclidean topological mapping on the time-aligned device operation data and the time-aligned environment data set to construct a spatiotemporal relationship map between the devices and the environmental variables within the building area, thereby obtaining a spatiotemporal relationship map of the building area; Step S133: Perform physical space interaction modeling based on the spatiotemporal relationship graph of the building area, map the physical space interaction modeling results to a preset building physical space coordinate system, convert the nodes in the spatiotemporal relationship graph into actual physical space coordinate nodes, and construct device-environment physical space interaction data; Step S134: performing a node space layout analysis on the device-environment physical space interaction data, and mapping the interaction relationship between the device and the environment based on the node space layout analysis results to obtain an environment physical space interaction map; Step S135: Optimize the graph based on the environmental physical space interaction graph and construct an initial spatial semantic topology graph.

4. The method for optimizing building energy consumption based on artificial intelligence according to claim 2, characterized in that: Step S14 is specifically as follows: Step S141: Calculate the local density of topological nodes based on the initial spatial semantic topological structure graph, set the calculation neighborhood radius to 2, set the local density threshold to 0.1, and regard nodes below the local density threshold as sparse areas, thereby obtaining the local density distribution of the topological structure; Step S142: Set the minimum number of samples to 5 and the density threshold to 0.3, perform neighborhood adaptive clustering on the initial spatial semantic topology graph, set the weight range to [0, 1], calculate the neighborhood adaptive weight based on the clustering results, and generate the initial parameters for neighborhood adaptive topology optimization; Step S143: Setting the weight adjustment coefficient to 0.5, adjusting the neighborhood topology weights of the initial spatial semantic topology graph based on the initial parameters of the neighborhood adaptive topology optimization, and setting the optimization scale to [1, 3] to perform multi-scale topology optimization to obtain an optimized device spatial semantic graph; Step S144: extracting device operation mode features with a feature dimension of 5 based on the optimized device space semantic graph, fusing the time-series aligned environment dataset, normalizing the device operation mode features, and generating standardized device operation mode data; Step S145: embedding the standardized device operation mode data into the topological structure of the optimized device space semantic graph, performing interaction modeling between the device operation mode and the environmental data, and obtaining a preliminary heterogeneous sensor dynamic spatiotemporal adjacency matrix; Step S146: Perform feature screening and topological correlation optimization on the preliminary heterogeneous sensing dynamic spatiotemporal adjacency matrix, strengthen key interaction relationships, and remove low-correlation edge weights, thereby generating a heterogeneous sensing dynamic spatiotemporal adjacency matrix.

5. The method for optimizing building energy consumption based on artificial intelligence according to claim 1, characterized in that: Step S2 is specifically as follows: Step S21: Screen the cross-domain energy consumption influencing factors of the sensor spatiotemporal fusion feature set, set the correlation threshold to [0.7, 0.8] to retain high correlation features, calculate the energy consumption impact weight of each device, and construct the initial energy consumption impact feature matrix based on the energy consumption impact weight; Step S22: performing non-Euclidean graph topology mapping on the initial energy consumption impact feature matrix, setting the maximum connection threshold of the adjacency matrix to 0.5, constructing a cross-device and cross-spatial energy consumption correlation topology structure, and generating an initial cross-domain energy consumption correlation topology map; Step S23: Perform spectral clustering analysis on the initial cross-domain energy consumption correlation topology map, select the first 20 eigenvectors for Laplace eigendecomposition, and set the feature truncation threshold to 0.1 to extract the main energy consumption impact patterns, thereby generating a cross-domain energy consumption correlation topology map; Step S24: Modeling the energy consumption transfer relationship between devices based on the cross-domain energy consumption correlation topology graph, setting the energy consumption transfer rate to [0.05, 0.2], and setting the dynamic weight change interval to [0, 1] to assign dynamic weights, and generating an initial dynamic energy consumption impact factor matrix; Step S25: normalize the initial dynamic energy consumption impact factor matrix, set the impact factor confidence interval [0.95, 1.05] to correct the robustness of the impact factor, optimize the energy consumption impact parameters, and output the dynamic energy consumption impact factor matrix.

6. The method for optimizing building energy consumption based on artificial intelligence according to claim 1, characterized in that: The multi-scale energy consumption evolution deduction described in step S3 is specifically as follows: Perform time series decomposition on the dynamic energy consumption influencing factor matrix, set multi-scale time steps, and perform trend decomposition and cycle decomposition to generate an energy consumption time series decomposition data set; Perform causal inference on the energy consumption propagation path of the energy consumption time series decomposition dataset, calculate the energy consumption correlation strength between devices, and construct an energy consumption propagation network to generate the initial energy consumption propagation structure; Based on the initial structure of energy consumption propagation, energy consumption propagation patterns are clustered, energy consumption propagation paths are screened, low-contribution energy consumption paths are removed, and an optimized energy consumption propagation network is generated; Extract time series features from the optimized energy consumption propagation network, divide the energy consumption time series segments, and perform empirical mode trend decomposition on the energy consumption time series segments to obtain energy consumption time series feature data; Based on the energy consumption time series characteristic data, autoregressive integral short-term energy consumption trend modeling is performed to generate short-term energy consumption forecast data; Perform residual analysis on short-term energy consumption forecast data, input the residual analysis results into the preset long short-term memory network to perform long-term energy consumption evolution modeling, and obtain long-term energy consumption forecast data; The long-term energy consumption forecast data is integrated with the short-term energy consumption forecast data to construct a comprehensive energy consumption forecast sequence. Based on the comprehensive energy consumption forecast sequence, a multi-resolution wavelet transform is performed to extract the energy consumption change patterns at different time scales and generate a multi-scale energy consumption evolution prediction matrix. Based on the multi-scale energy consumption evolution prediction matrix, state probability modeling is performed, the energy consumption state transfer matrix is ​​calculated, and the energy consumption state transfer matrix is ​​combined with the building energy consumption evolution map to obtain the building energy consumption evolution map.

7. The method for optimizing building energy consumption based on artificial intelligence according to claim 6, characterized in that: The long-term energy consumption evolution modeling is specifically as follows: The short-term energy consumption forecast data and the residual analysis results are time-step aligned to obtain a time-series aligned residual dataset; Extract local spatiotemporal features from the time series alignment residual dataset to generate residual feature data; Based on the residual feature dataset, a long short-term memory network model is constructed; Backpropagation is performed on the long short-term memory network model, and the network parameters are optimized using the gradient descent method. The long-term energy consumption evolution model is modeled on the time-series alignment residual data set to obtain the long-term energy consumption evolution model. The long-term energy consumption evolution model is used to perform long-term energy consumption prediction on the residual feature data to obtain long-term energy consumption prediction data.

8. The method for optimizing building energy consumption based on artificial intelligence according to claim 1, characterized in that: The generation of the energy-saving optimization instruction set described in step S4 is specifically as follows: Based on the multi-objective energy-saving control strategy manifold, the adaptability analysis of device category strategy is carried out, and the device control demand hierarchical clustering is performed to generate the device control classification index; According to the equipment control classification index, energy-saving strategy constraints are mapped to the multi-objective energy-saving control strategy manifold to obtain the constrained optimization strategy set; Perform time series scheduling optimization on the constraint optimization strategy set, adjust the strategy execution time, and generate a time optimization strategy set; Assign policy execution priorities based on the time-optimized policy set, and adaptively adjust policy execution parameters based on the timing alignment environment data set to generate an adaptive energy-saving control instruction set; The adaptive energy-saving control instruction set is verified for consistency, the feasibility of executing the adaptive energy-saving control instruction set is verified, the energy-saving optimization instruction set is obtained, and uploaded to the building energy consumption control platform to execute the equipment control task.

9. The method for optimizing building energy consumption based on artificial intelligence according to claim 1, characterized in that: The device predicted energy consumption deviation lineage tracking described in step S5 is specifically as follows: Obtain historical energy consumption data of building equipment, and establish a baseline energy consumption prediction model based on the historical energy consumption data of building equipment and the energy-saving optimization instruction set, thereby obtaining equipment baseline energy consumption prediction data; Obtain real-time building energy consumption data, perform error comparison between equipment baseline energy consumption forecast data and real-time building energy consumption data, calculate the predicted energy consumption deviation metric, and perform preliminary screening of abnormal energy consumption based on the preset deviation forecast to obtain the equipment energy consumption deviation metric matrix; Obtain device control logs, and build a causal relationship diagram of device energy consumption changes based on the device control logs and the dynamic energy consumption influencing factor matrix, perform causal testing of energy consumption influencing factors, and generate a device energy consumption lineage tracking chain; The energy consumption deviation measurement matrix of the equipment is combined with the equipment lineage tracing chain to trace the energy consumption anomaly. Based on the energy consumption anomaly tracing results, the abnormal energy consumption patterns of the equipment are classified to obtain an energy consumption deviation backtracking analysis report.

10. The method for optimizing building energy consumption based on artificial intelligence according to claim 9, characterized in that: The establishment of the benchmark energy consumption prediction model is specifically as follows: Obtain historical energy consumption data of building equipment and perform time series alignment on the historical energy consumption data of building equipment to obtain a standardized historical energy consumption data set of equipment; Based on the energy-saving optimization instruction set and the dynamic energy consumption impact factor matrix, the energy-saving optimization impact of the equipment's historical energy consumption data is quantified to generate an energy-saving optimization impact factor data set; Based on the standardized equipment historical energy consumption data set and the energy-saving optimization influencing factor data set, a baseline energy consumption prediction model is built using a preset regression algorithm to obtain the equipment baseline energy consumption prediction model; Cross-validate the equipment benchmark energy consumption prediction model, iteratively adjust the model parameters, and obtain the benchmark energy consumption prediction model; The energy-saving optimization instruction set is input into the baseline energy consumption prediction model to obtain the equipment baseline energy consumption prediction data.

Images