Green building energy consumption intelligent monitoring and diagnosis system based on multi-source data fusion

Abstract

This invention discloses a green building energy consumption intelligent monitoring and diagnosis system based on multi-source data fusion, belonging to the field of building energy management technology. It employs an improved Transformer-LSTM model and dynamic window alignment technology to solve the problems of time synchronization and missing values ​​in heterogeneous data, thus improving the input quality of the diagnostic algorithm. Through device-level energy consumption decomposition and multimodal root cause localization, it improves the accuracy of anomaly diagnosis and can accurately locate faulty equipment. Policy generation based on GNN behavior perception and PPO reinforcement learning reduces overall energy consumption while ensuring comfort. The system has data closed-loop management capabilities, achieving continuous self-learning through reverse annotation of diagnostic results and dynamic threshold optimization. Three-dimensional visualization and an interactive interface significantly improve operation and maintenance efficiency. This system significantly enhances the accuracy, automation level, and economic benefits of building energy consumption management.

Description

Technical Field

[0001] This invention belongs to the field of building energy management technology, specifically relating to a green building energy consumption intelligent monitoring and diagnosis system based on multi-source data fusion. Background Technology

[0002] With the widespread adoption of green building concepts, intelligent building energy consumption monitoring systems have become a key technological means to achieve energy conservation and emission reduction goals. However, existing systems face a core bottleneck in practical applications: the fusion of multi-source heterogeneous data. Building energy consumption monitoring involves multiple subsystems such as electricity, HVAC, and lighting. The energy consumption data generated by these systems is typically collected at a frequency of minutes or even seconds, while related meteorological parameters, environmental data, and equipment operating status are often recorded at time intervals of hours or longer. This mismatch in time scales leads to difficulties in data synchronization, and traditional fixed-window alignment methods struggle to accurately capture the lag correlation between temperature changes and air conditioning energy consumption.

[0003] Regarding data quality, issues such as sensor malfunctions and communication interruptions lead to data loss and frequent outliers. Existing systems often employ static thresholding or simple linear interpolation for data processing; for example, an anomaly is identified when energy consumption data exceeds twice the historical average. While this method is simple and easy to implement, it cannot adapt to the dynamic changes in building energy consumption caused by factors such as seasons, weather, and human activity. More seriously, the time-lag correlation between temperature and air conditioning energy consumption is often ignored in existing models, resulting in an accuracy rate for energy consumption anomaly diagnosis generally below 60%.

[0004] In the diagnostic phase, most existing technologies remain at the level of overall energy consumption analysis, lacking the ability to decompose energy consumption at the device level. When an anomaly is detected, the system can only indicate "abnormal energy consumption" without being able to pinpoint the specific device or its root cause (such as filter blockage, refrigerant leakage, or sensor drift). This coarse-grained diagnostic result means that maintenance personnel still need to manually inspect and troubleshoot the problem, significantly reducing maintenance efficiency.

[0005] Furthermore, existing systems generally lack the ability to perceive and learn user behavior patterns. The distribution and activity patterns of people within a building significantly impact energy consumption, but traditional systems treat human factors as uncontrollable variables, failing to establish a dynamic relationship between "people, equipment, and environment." In terms of strategy generation, most systems still employ preset, fixed control strategies, unable to adaptively adjust according to actual usage scenarios, resulting in a difficulty in balancing energy efficiency and user comfort.

[0006] Finally, most existing monitoring systems are open-loop architectures, meaning that diagnostic results and optimization strategies cannot be fed back to the data preprocessing and model training stages, resulting in a lack of continuous system evolution capabilities. As buildings age and equipment deteriorates, fixed anomaly detection thresholds and diagnostic rules gradually become ineffective, leading to a decline in system performance.

[0007] Therefore, the current field of green building energy consumption monitoring urgently needs an intelligent system that can achieve dynamic fusion of multi-source data, possess device-level precise diagnostic capabilities, support user behavior perception, and form a closed-loop optimization, in order to break through existing technological bottlenecks and truly realize refined and intelligent management of building energy consumption. Summary of the Invention

[0008] To address the shortcomings of the existing technologies, this application provides a smart monitoring and diagnostic system for green building energy consumption based on multi-source data fusion.

[0009] In the first aspect, this application proposes a green building energy consumption intelligent monitoring and diagnosis system based on multi-source data fusion, including: a multi-source data dynamic standardization and anomaly detection module, an equipment-level energy consumption decomposition and root cause localization module, a user behavior perception and strategy generation module, a system-level data closed-loop management module, and a visual interactive terminal. The multi-source data dynamic standardization and anomaly detection module is used to align the time axis of heterogeneous data, process missing values, and detect abnormal data. The device-level energy consumption decomposition and root cause localization module is used to realize device-level energy consumption decomposition and root cause diagnosis. The user behavior perception and strategy generation module is used to construct a dynamic association graph, identify behavior patterns, and generate optimization strategies. The system-level data closed-loop management module is used to build a standardized data warehouse and realize reverse annotation of diagnostic results and dynamic optimization of thresholds; The visual interactive terminal is used to display energy consumption status, generate diagnostic reports, and support user feedback.

[0010] In some embodiments, the multi-source data dynamic standardization and anomaly detection module includes: a time alignment processing unit, a missing data completion unit, a data standardization unit, and an anomaly detection unit; The time alignment processing unit is used to calculate the moving average of energy consumption data based on the time of meteorological data using dynamic window alignment technology. The missing data completion unit is used to dynamically calculate the completion value based on the seasonal coefficient and weekday pattern using a weighted interpolation method based on historical data from the same period. The data standardization unit is used to employ an improved Transformer-LSTM hybrid model, including a multi-head self-attention mechanism to capture lag correlations, a bidirectional LSTM to handle temporal fluctuations, and batch normalization to output a standardized matrix. The anomaly detection unit is used to construct a decision tree using the isolated forest algorithm. When the energy consumption data deviates from the historical normal operation range by more than the preset tolerance range, and the deviation cannot be explained by external environmental factors, the system determines that the data is abnormal and triggers a recalculation and verification mechanism.

[0011] In some embodiments, the device-level energy consumption decomposition and root cause localization module includes: a multi-scale feature extraction unit, a device energy consumption decomposition unit, and a multi-modal root cause diagnosis unit; The multi-scale feature extraction unit is used to employ an improved Transformer-LSTM hybrid architecture, wherein the Transformer layer is designed with a lag attention weight matrix and the LSTM layer introduces a time decay factor. The device energy consumption decomposition unit is used to output the energy consumption ratio of each device through a softmax classifier. The multimodal root cause diagnosis unit is used to output a root cause probability ranking by combining decision tree coarse screening with Bayesian network fine localization.

[0012] In some embodiments, the lag attention weight matrix is ​​used to capture the temporal correlation between historical meteorological features and the current energy consumption status. By measuring the correlation strength between features at different times and combining a time decay mechanism to reduce the influence weight of long-term historical data, the expression of recent and periodic correlation features is strengthened.

[0013] In some embodiments, the user behavior perception and strategy generation module includes: a dynamic association graph construction unit, a behavior pattern recognition unit, and a strategy generation and optimization unit; The dynamic association graph construction unit is used to model the relationship between people, equipment, and environment using a graph neural network and to introduce a time decay factor. The behavior pattern recognition unit is used to employ an LSTM+DBSCAN hybrid algorithm, wherein DBSCAN uses dynamic time warping as a distance metric. The policy generation and optimization unit is used to design a reward function that incorporates energy saving, comfort, and user feedback using the PPO reinforcement learning framework.

[0014] In some embodiments, the reward function of the PPO reinforcement learning forms a guiding signal for strategy optimization by comprehensively evaluating the performance of three dimensions: energy saving effect, environmental comfort, and user satisfaction, in a weighted aggregation manner. Here, the energy saving reward reflects the degree of energy consumption reduction after the strategy is implemented, the comfort reward represents the degree of matching between indoor environmental parameters and the human comfort range, and the user feedback reward reflects the occupants' subjective evaluation of the system adjustment effect.

[0015] In some embodiments, the system-level data closed-loop management module includes: a standardized data warehouse, a diagnostic result reverse annotation unit, and a threshold dynamic optimization unit; The standardized data warehouse is used to employ a layered storage structure, including a raw data layer, a feature engineering layer, and a baseline curve layer; The diagnostic result reverse annotation unit is used to achieve knowledge transfer using a label propagation algorithm; The threshold dynamic optimization unit is used to adaptively adjust the anomaly detection threshold based on KL divergence.

[0016] In some embodiments, the system includes: the threshold dynamic optimization unit adaptively adjusts the anomaly judgment threshold according to the degree of difference between the current data distribution and the historical benchmark distribution, so that the detection sensitivity changes dynamically with the data evolution trend. When the monitoring data distribution shifts, the system automatically learns new data feature patterns and correspondingly relaxes or tightens the anomaly judgment boundary to adapt to the long-term evolution of the building's operating status.

[0017] In some embodiments, the visualization interactive terminal includes: a three-dimensional energy consumption heat map generation subsystem, a root cause diagnosis report automatic generator, and a strategy adjustment interactive interface; The three-dimensional energy consumption heat map generation subsystem is built on WebGL and uses Gaussian kernel density estimation to generate spatial heat distribution. The root cause diagnosis report generator is used to output structured reports using a template engine and natural language generation technology. The strategy adjustment interface is used for parameter adjustment, real-time preview, and user feedback collection.

[0018] In some embodiments, in the strategy adjustment interface, user feedback rewards are dynamically adjusted based on real-time collected comfort scores. Through an iterative learning mechanism, the system gradually approaches the user's personalized preferences. The system progressively adjusts the weight contribution of feedback rewards based on the degree of deviation between each user feedback and historical expectations, thereby achieving continuous learning of user preferences and strategy optimization.

[0019] Secondly, this application proposes a method for intelligent monitoring and diagnosis of energy consumption in green buildings based on multi-source data fusion, including the following steps: The system performs time synchronization alignment, missing value completion, and standardization on building energy consumption data and heterogeneous meteorological and environmental data. It also detects abnormal data in real time based on an improved Transformer-LSTM hybrid model and isolated forest algorithm, and outputs high-quality standardized data. Based on the standardized data, device-level energy consumption decomposition is achieved through multi-scale feature extraction, and a multi-modal diagnostic algorithm that integrates decision trees and Bayesian networks is used to output root cause probability ranking. Construct a dynamic relationship graph of personnel, equipment, and environment; identify user behavior patterns based on LSTM and DBSCAN; and use the PPO reinforcement learning algorithm to generate equipment optimization strategies. Establish a standardized data warehouse to achieve reverse annotation of diagnostic results and dynamic optimization of anomaly detection thresholds, forming a closed-loop management of data flow; The system displays the device-level energy consumption breakdown results through a 3D heatmap, automatically generates root cause diagnosis reports, and provides a strategy adjustment interface to collect user feedback.

[0020] Thirdly, this application proposes an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described above.

[0021] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described above.

[0022] The beneficial effects of this invention are: By dynamically fusing multi-source data and implementing an intelligent diagnostic closed loop, a systematic improvement in green building energy consumption monitoring is achieved. An improved Transformer-LSTM model and dynamic window alignment technology are employed to solve the problems of time synchronization and missing values ​​in heterogeneous data, thus improving the input quality of the diagnostic algorithm. Equipment-level energy consumption decomposition and multimodal root cause localization enhance the accuracy of anomaly diagnosis and enable precise location of faulty equipment. Policy generation based on GNN behavior perception and PPO reinforcement learning reduces overall energy consumption while ensuring comfort. The system possesses data closed-loop management capabilities, achieving continuous self-learning through reverse annotation of diagnostic results and dynamic threshold optimization. Three-dimensional visualization and an interactive interface significantly improve operational efficiency. Ultimately, a complete technical closed loop is formed, from data perception to intelligent diagnosis, policy generation, and feedback optimization, significantly improving the accuracy, automation level, and economic benefits of building energy consumption management. Attached Figure Description

[0023] Figure 1 This is a system principle block diagram of the present invention.

[0024] Figure 2 This is the overall flowchart of the present invention. Detailed Implementation

[0025] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein; rather, these embodiments are provided so that a more thorough understanding of the invention can be achieved and that the full scope of the invention can be conveyed to those skilled in the art.

[0026] Firstly, this application proposes a green building energy consumption intelligent monitoring and diagnosis system based on multi-source data fusion, such as... Figure 1As shown, it includes: a multi-source data dynamic standardization and anomaly detection module, a device-level energy consumption decomposition and root cause localization module, a user behavior perception and strategy generation module, a system-level data closed-loop management module, and a visual interactive terminal; The multi-source data dynamic standardization and anomaly detection module is used to align the time axis of heterogeneous data, process missing values, and detect abnormal data. The multi-source data dynamic standardization and anomaly detection module employs a multi-stage data processing workflow to achieve accurate fusion of heterogeneous data. During the time synchronization phase, a dynamic window alignment technique is implemented to address the time axis differences between minute-level energy consumption data (such as meter readings) and hourly meteorological data (such as temperature and humidity): First, using the timestamp of the meteorological data as a reference point, the energy consumption data is averaged by hourly windows (60-minute window width, 1-minute step), generating an energy consumption feature sequence that matches the time resolution of the meteorological data. For missing periods (such as nighttime without meteorological records), a weighted interpolation method based on historical data from the same period is used to fill in the gaps. The weights are dynamically adjusted based on seasonal coefficients and weekday patterns, for example: 0.6 for summer / 0.4 for winter, and 0.7 for weekdays / 0.3 for weekends. During the data standardization phase, an improved Transformer-LSTM model with a three-layer structure was constructed: the first layer, a Transformer encoder, captures the lag correlation between temperature and air conditioning energy consumption through a multi-head self-attention mechanism (with a maximum lag step size of 12 hours); the second layer, a bidirectional LSTM, processes the minute-level fluctuation characteristics of energy consumption data; and the third layer, a batch normalization layer, outputs a Z-score standardized matrix. In the anomaly detection phase, an isolated forest algorithm was used to construct 100 decision trees. When an energy consumption value exceeds the historical baseline by 150% and no abnormal fluctuations in meteorological characteristics are detected (temperature change <3℃ / h), a data recalculation process is triggered: the baseline value is automatically recalculated using data from the same period of the last 7 days, and the data validity is verified using the 3σ principle.

[0027] Taking the energy consumption monitoring of the air conditioning system of an office building on [Date] as an example, the raw data includes: air conditioning meter readings at 1-minute intervals (fields: timestamp, power_kW) and hourly temperature data provided by the weather station (fields: hour, temp_C). The implementation steps are as follows: The time alignment processing unit: takes the meteorological data from 9:00 to 10:00 (temp_C=28.5) as a benchmark, calculates the moving average of 60 energy consumption data points (power_kW from 9:00:00 to 9:59:00) within this period, and obtains the aligned energy consumption characteristic value of 29.8kW; The missing data completion unit: Since there are no meteorological records at 3:00 AM on the same day, the temperature data at 3:00 AM of the five most recent Mondays (26.2, 25.8, 26.5, 26.0, 25.9℃) are taken, and the weighted average of 26.1℃ is calculated according to the seasonal weight as the completion value; The data standardization unit: The Transformer layer identifies a significant temporal correlation between the peak energy consumption of the air conditioner on the current day and the temperature change on the previous night. It captures the lag effect of temperature change on energy consumption through an attention mechanism and assigns a high importance weight to this correlation feature. The anomaly detection unit detected an instantaneous power of 78.6kW (249% of the baseline value of 31.5kW) at 14:15, but the temperature change was only 1.8℃ / h. After triggering the recalculation process, it was confirmed that the sensor was faulty and replaced with an average value of 32.1kW over the previous 5 minutes.

[0028] In some embodiments, the multi-source data dynamic standardization and anomaly detection module includes: a time alignment processing unit, a missing data completion unit, a data standardization unit, and an anomaly detection unit; The time alignment processing unit is used to calculate the moving average of energy consumption data based on the time of meteorological data using dynamic window alignment technology. Example: Air conditioning energy consumption in office buildings is collected at minute intervals, while meteorological data (such as temperature) is provided at hourly intervals, resulting in inconsistent timelines.

[0029] The time alignment processing unit in this embodiment employs dynamic window alignment technology. Using hourly meteorological data as a reference, it calculates a moving average of 60 energy consumption data points within that hour. The sliding window width is 60 minutes, and the step size is 1 minute, generating an energy consumption characteristic sequence that matches the time resolution of the meteorological data.

[0030] Output: The temperature at 9:00 AM is 28.5℃. The average value of 60 air conditioner power values ​​from 9:00:00 to 9:59:00 during this period is calculated, resulting in an aligned energy consumption characteristic value of 29.8kW.

[0031] The missing data completion unit is used to dynamically calculate the completion value based on the seasonal coefficient and weekday pattern using a weighted interpolation method based on historical data from the same period. Example: Data is missing during nighttime or periods when no weather records are available.

[0032] The missing data completion unit in this embodiment uses weighted interpolation of historical data from the same period. It selects the data from the same period of the last 5 days of the week and calculates a weighted average as the completion value based on the seasonal coefficient (0.6 for summer, 0.4 for winter) and the weekday pattern (0.7 for weekdays, 0.3 for weekends).

[0033] Output: No weather records were available at 3:00 AM. Temperature data at 3:00 AM from the last five Mondays (26.2℃, 25.8℃, 26.5℃, 26.0℃, 25.9℃) were used. The weighted average of the data was calculated based on seasonal weights and was 26.1℃.

[0034] The data standardization unit is used to employ an improved Transformer-LSTM hybrid model, including a multi-head self-attention mechanism to capture lag correlations, a bidirectional LSTM to handle temporal fluctuations, and batch normalization to output a standardized matrix. The data normalization unit in this embodiment uses an improved Transformer-LSTM hybrid model.

[0035] The Transformer layer uses eight attention heads to capture the lagged correlation between temperature and air conditioning energy consumption, with a maximum lag step of 12 hours. The attention mechanism identifies the impact of historical temperature changes on current energy consumption.

[0036] LSTM layer: Uses bidirectional LSTM to process energy consumption fluctuations per minute, with 128 hidden units. Output gating introduces a time decay factor to reduce the impact of long-standing historical data.

[0037] Output results: After batch normalization, the output is standardized data with a mean of 0 and a variance of 1. Specifically, the Transformer layer identifies a strong correlation between the current peak energy consumption of the air conditioner (35.2kW) and the temperature drop (4.2℃ drop) at 22:00 the previous day, with an attention weight of 0.73.

[0038] The anomaly detection unit is used to construct a decision tree using the isolated forest algorithm. When the energy consumption data deviates from the historical normal operation range by more than the preset tolerance range, and the deviation cannot be explained by external environmental factors, the system determines that the data is abnormal and triggers a recalculation and verification mechanism.

[0039] The anomaly detection unit in this embodiment uses the isolated forest algorithm to construct 100 decision trees. If the current energy consumption value exceeds 150% of the historical baseline and the temperature change is less than 3°C per hour, a data recalculation process is triggered.

[0040] Recalculation process: The baseline value is automatically recalculated using data from the same period of the most recent 7 days, and the validity of the data is verified by the 3σ principle (mean ± 3 times standard deviation).

[0041] Output results: At 14:15 in the afternoon, the air conditioner power was detected to be 78.6kW (249% of the baseline of 31.5kW), but the temperature change was only 1.8℃ / hour. It was determined to be a sensor failure and replaced with the average value of 32.1kW for the previous 5 minutes.

[0042] The device-level energy consumption decomposition and root cause localization module is used to realize device-level energy consumption decomposition and root cause diagnosis. Specifically, the Transformer-LSTM hybrid model is used to extract multi-scale energy consumption features to achieve device-level energy consumption decomposition; and a multi-modal diagnostic algorithm that integrates decision trees and Bayesian networks is used to output root cause probability ranking.

[0043] In some embodiments, the device-level energy consumption decomposition and root cause localization module includes: a multi-scale feature extraction unit, a device energy consumption decomposition unit, and a multi-modal root cause diagnosis unit; The multi-scale feature extraction unit is used to employ an improved Transformer-LSTM hybrid architecture, wherein the Transformer layer is designed with a lag attention weight matrix and the LSTM layer introduces a time decay factor. Among them, the multi-scale feature extraction adopts an improved Transformer-LSTM hybrid architecture, in which the Transformer layer is designed with a lag attention weight matrix and the LSTM layer introduces a time decay factor.

[0044] Among them, multi-scale feature extraction uses an improved Transformer-LSTM hybrid architecture to process energy consumption data: Transformer encoding layer: A lag attention mechanism is designed to capture meteorological cycle features. This mechanism measures the correlation strength between features at different times and reduces the influence weight of long-term historical data by combining a time decay mechanism, thereby strengthening the expression of recent and periodic correlation features. It is specifically designed to capture the lag correlation between temperature changes and energy consumption response.

[0045] LSTM Temporal Layer: The bidirectional LSTM network reduces the impact of long-standing historical data by introducing a time decay mechanism, making the model focus more on recent temporal fluctuations.

[0046] The device energy consumption decomposition unit is used to output the energy consumption ratio of each device through a softmax classifier.

[0047] Specifically, a device feature vector is constructed, and the input features are linearly transformed using trainable parameters. The transformed feature vector is then converted into a probability distribution output using a normalized exponential function, so that the sum of the energy consumption proportions of each device is 1, thereby decomposing the total energy consumption among the devices.

[0048] The multimodal root cause diagnosis unit is used to output a root cause probability ranking by combining decision tree coarse screening with Bayesian network fine localization.

[0049] Multimodal root cause diagnosis is implemented in two stages: Decision tree coarse screening: Construct a multi-level decision tree and set splitting conditions based on key operating parameters, including the temperature difference threshold of air conditioning inlet and outlet water, the fluctuation rate of lighting power, and the start-stop frequency of the fresh air system, to quickly screen out the abnormal candidate set.

[0050] Bayesian Precise Localization: A dynamic Bayesian network is constructed, comprising observation nodes (device status parameters) and hidden nodes (root cause types). A Bayesian inference mechanism is employed to infer the posterior probability distribution of potential root cause types based on the observed device status, achieving precise localization from apparent anomalies to deep-seated faults. The prior probabilities are obtained through statistical analysis of historical maintenance records.

[0051] Specifically, let's continue with the example of diagnosing air conditioning systems in office buildings: Input data: Total energy consumption over multiple days and minute-by-minute data, and status of each device; The Transformer layer identified a strong correlation between peak energy consumption at specific times of the day and outdoor temperature at the same time the previous day, capturing a significant lag correlation. The LSTM layer detected a power fluctuation of a second in an air conditioning unit and determined it to be an abnormal compressor start-stop operation. Output the energy consumption percentage of each device: percentage for main unit, water pump, and terminal unit respectively; The decision tree identifies that the temperature difference between the inlet and outlet water of the host exceeds the standard, triggering Bayesian network calculation. Input observation: Abnormal compressor current fluctuation rate; Calculate the posterior probability: rank the probabilities of filter blockage, insufficient refrigerant, and sensor failure; The diagnostic report outputs the following recommendation: Clean the filter first.

[0052] The user behavior perception and strategy generation module is used to construct a dynamic association graph, identify behavior patterns, and generate optimization strategies. In some embodiments, the user behavior perception and strategy generation module includes: a dynamic association graph construction unit, a behavior pattern recognition unit, and a strategy generation and optimization unit; The dynamic association graph construction unit is used to model the relationship between people, equipment and environment using graph neural networks and introduces a time decay factor.

[0053] Dynamic Relationship Graph Construction: A graph neural network is used to model the "personnel-equipment-environment" ternary relationship. The graph structure is defined to include a set of nodes (personnel, equipment, and environment nodes) and an edge feature matrix.

[0054] An improved graph attention network is designed to compute edge weights and learn the importance relationships between nodes through trainable parameters. A time decay mechanism is added to give higher weight to recent interactions, reflecting the timeliness of human behavior.

[0055] The behavior pattern recognition unit is used to employ a hybrid LSTM+DBSCAN algorithm, wherein DBSCAN uses dynamic time warping as a distance metric.

[0056] Sensor data is processed using a hybrid LSTM+DBSCAN algorithm. The LSTM network takes time series data as input and outputs latent feature representations, mapping high-dimensional time series data to a low-dimensional feature space.

[0057] Improved DBSCAN clustering: The distance metric employs dynamic time warping to measure the similarity of time series of different lengths. By aligning local features of time series data, it identifies patterns that are similar in shape but have time axis scaling. Adaptive parameter adjustment: Automatically optimizes cluster radius based on silhouette coefficient; The policy generation and optimization unit is used to design a reward function that incorporates energy saving, comfort, and user feedback using the PPO reinforcement learning framework.

[0058] Design a reinforcement learning framework based on PPO: State space: contains device state, environmental parameters, and behavioral pattern encoding; Action space: Equipment control commands (such as temperature setpoint adjustment); The reward function comprehensively evaluates the performance across three dimensions: energy saving effect, environmental comfort, and user satisfaction, and uses a weighted aggregation method to form a guiding signal for strategy optimization. Specifically, the energy saving reward reflects the degree of energy consumption reduction after the strategy is implemented, the comfort reward characterizes the degree of matching between indoor environmental parameters and the human comfort range, and the user feedback reward reflects the subjective evaluation of the system's adjustment effect by users.

[0059] The policy network is updated using generalized advantage estimation.

[0060] Specifically, let's take the optimization of air conditioning in office buildings as an example: 1. Data Acquisition Phase: Personnel Node: Infrared sensors in the conference room detected a gathering of people; Equipment nodes: Increased air conditioning terminal opening and increased lighting power; Environmental factors: CO2 concentration increases, temperature increases; 2. Construction of the relationship graph: The graph neural network calculates the association weights between people and air conditioning, and between people and lighting; The time decay mechanism gives the current meeting mode a higher weight than the previous cleaning mode; 3. Behavior recognition: LSTM processes sensor time series and outputs feature vectors; DBSCAN clustering identifies it as "meeting mode," which is different from regular office mode; 4. Strategy Generation: PPO network input current status: air conditioning setpoint, actual temperature, behavior pattern code, CO2 concentration; Output action: Adjust the air conditioning setting to increase the fresh air volume; The procedure is executed after user confirmation, and a comfort rating is obtained.

[0061] The system-level data closed-loop management module is used to build a standardized data warehouse and realize reverse annotation of diagnostic results and dynamic optimization of thresholds; In some embodiments, the system-level data closed-loop management module includes: a standardized data warehouse, a diagnostic result reverse annotation unit, and a threshold dynamic optimization unit; The standardized data warehouse is used to employ a layered storage structure, including a raw data layer, a feature engineering layer, and a baseline curve layer; Raw data layer: Stores unprocessed raw sensor data, organized by timestamp, device identifier, measurement value, and quality mark; Feature engineering layer: Stores a data matrix that has undergone dynamic normalization and handles missing data through a periodic completion mechanism; Baseline curve layer: Stores the energy consumption baseline of each device under typical operating conditions, and establishes a reference standard for normal operation through Gaussian process regression calculation.

[0062] The diagnostic result reverse annotation unit is used to achieve knowledge transfer using a label propagation algorithm; A label propagation algorithm is used to achieve knowledge transfer.

[0063] Knowledge transfer is achieved using an improved label propagation algorithm: Construct a labeled graph, where nodes are data samples and edge weights are calculated based on feature similarity and time decay; Iterative updates to the label distribution enable the knowledge of labeled samples to spread to similar unlabeled samples, thereby achieving automatic expansion of diagnostic experience.

[0064] The threshold dynamic optimization unit is used to adaptively adjust the anomaly detection threshold based on KL divergence.

[0065] Threshold dynamic optimization: An adaptive threshold adjustment algorithm based on distribution differences is proposed: Calculate the degree of difference between the current data distribution and the baseline distribution to measure the data evolution shift; The detection threshold is dynamically adjusted so that the detection sensitivity changes dynamically with the data evolution trend. When the distribution of monitored data shifts, the system automatically learns new data feature patterns and adjusts the anomaly judgment boundary accordingly to adapt to the long-term evolution of the building's operating status.

[0066] Specifically: Taking the abnormal detection of the chiller unit in the target building as an example: 1. Data preparation stage: The power data of chiller units and ambient temperature from the past few days were loaded from the data warehouse to form a baseline matrix; Calculate the hourly baseline mean and variance; 2. Real-time detection phase: The power and ambient temperature are measured at the current moment; Calculate the degree of distribution difference, and determine it as abnormal if it exceeds the threshold; Dynamically adjust the detection threshold; 3. Reverse annotation stage: The anomaly was confirmed to be caused by scaling in the cooling tower; a new node was created in the annotation diagram. Distribute tags to historical data of similar operating conditions; Update the variance parameter for the corresponding time period in the baseline curve; 4. Strategy optimization phase: Maintenance recommendations are generated based on the new threshold.

[0067] The visual interactive terminal is used to display energy consumption status, generate diagnostic reports, and support user feedback.

[0068] In some embodiments, the visualization interactive terminal includes: a three-dimensional energy consumption heat map generation subsystem, a root cause diagnosis report automatic generator, and a strategy adjustment interactive interface; The three-dimensional energy consumption heat map generation subsystem is built on WebGL and uses Gaussian kernel density estimation to generate spatial thermal distribution.

[0069] A dynamic rendering pipeline based on WebGL is built to map device-level power consumption data to spatial thermal distribution: Data preprocessing: Normalize the equipment coordinates and energy consumption values; Thermal field calculation: A continuous distribution is generated using the kernel density estimation method. The energy consumption of discrete devices is extended into a continuous field distribution through spatial smoothing. The bandwidth parameter controls the degree of smoothing. Dynamic coloring: Color space conversion is applied based on energy consumption intensity; low-energy areas are presented with cool tones. High-energy-consuming areas are presented in warm colors, forming an intuitive visualization of thermal gradients.

[0070] The root cause diagnosis report generator is used to output structured reports using a template engine and natural language generation technology.

[0071] Employing template engines and natural language generation techniques: Structured data extraction: Obtain fields such as abnormal device identifier, occurrence time, and root cause probability from the diagnostic module; Report content assembly: Generate a structured diagnostic report according to a preset template, including sections such as equipment information, anomaly time, root cause analysis (sorted by confidence level), and recommended actions; Visual chart embedding: Automatically generate historical comparison curves and probability distribution histograms: In some embodiments, the strategy adjustment interface is used for parameter adjustment, real-time preview, and user feedback collection.

[0072] The strategy adjustment interface includes implementing a two-way feedback mechanism: Parameter adjustment panel: Provides slider controls for air conditioning temperature setting, lighting brightness, etc., supporting fine-tuning; Real-time preview function: Adjusting parameters instantly displays the predicted energy consumption changes, and the energy-saving effect after strategy adjustment is estimated through the energy consumption prediction model; User feedback collection: A multi-level comfort rating is set up, and the user feedback reward is dynamically adjusted based on the comfort rating collected in real time. Through an iterative learning mechanism, the system gradually approaches the user's personalized preferences.

[0073] Specifically: Taking the optimization of the air conditioning system of the target building as an example: 1. Heatmap generation process: Input data: Coordinates and current power of multiple air conditioners; Calculate the normalized value; Generate thermal fields: Render a gradient effect from low power consumption to high power consumption in the 3D model; 2. Diagnostic report generation: Input data: An air conditioner malfunctioned, the root cause being "clogged filter" and "insufficient refrigerant"; Automatically generate PDF reports, including comparisons of historical power curves; Output recommendations: "Clean the filter immediately", "Check refrigerant pressure within 48 hours"; 3. Strategy adjustment interaction: Users set the air conditioner temperature higher, and the interface displays the expected energy-saving effect in real time. Submit a comfort rating, and the system will update the reward function of the reinforcement learning model; The following day, under the same operating conditions, the system will prioritize recommending the optimized temperature control strategy.

[0074] Secondly, this application proposes a method for intelligent monitoring and diagnosis of energy consumption in green buildings based on multi-source data fusion, such as... Figure 2 As shown, it includes the following steps: S100: Performs time synchronization alignment, missing value completion, and standardization processing on building energy consumption data and heterogeneous meteorological and environmental data. Based on the improved Transformer-LSTM hybrid model and isolated forest algorithm, it detects abnormal data in real time and outputs high-quality standardized data. S200: Based on the standardized data, device-level energy consumption decomposition is achieved through multi-scale feature extraction, and a multi-modal diagnostic algorithm that integrates decision trees and Bayesian networks is used to output the root cause probability ranking. S300: Construct a dynamic relationship graph of personnel, equipment, and environment; identify user behavior patterns based on LSTM and DBSCAN; and use the PPO reinforcement learning algorithm to generate equipment optimization strategies. S400: Establish a standardized data warehouse to achieve reverse annotation of diagnostic results and dynamic optimization of anomaly detection thresholds, forming a closed-loop management of data flow; S500: Displays device-level energy consumption breakdown results through a 3D heat map, automatically generates root cause diagnosis reports, and provides a strategy adjustment interface to collect user feedback.

[0075] Thirdly, this application proposes an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described above.

[0076] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described above.

[0077] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0078] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0079] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this disclosure.

[0080] In the embodiments provided in this disclosure, it should be understood that the disclosed apparatus / computer devices and methods can be implemented in other ways. For example, the apparatus / computer device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. Multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0081] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0082] Furthermore, the functional units in the various embodiments of this disclosure can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0083] If an integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program may include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. A computer-readable medium may include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in a computer-readable medium may be appropriately added to or subtracted according to the requirements of legislation and patent practice in a jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0084] The above are merely preferred embodiments of the present invention. It should be noted that any modifications and improvements made by those skilled in the art without departing from the present technical solution should also be considered to fall within the scope of protection claimed by the present solution.

Claims

1. A green building energy consumption intelligent monitoring and diagnosis system based on multi-source data fusion, characterized in that, include: Multi-source data dynamic standardization and anomaly detection module, device-level energy consumption decomposition and root cause localization module, user behavior perception and strategy generation module, system-level data closed-loop management module, and visual interactive terminal; The multi-source data dynamic standardization and anomaly detection module is used to align heterogeneous data timelines, handle missing values, and detect abnormal data. The device-level energy consumption decomposition and root cause localization module is used to realize device-level energy consumption decomposition and root cause diagnosis. The user behavior perception and strategy generation module is used to build dynamic relationship graphs, identify behavior patterns and generate optimization strategies. The system-level data closed-loop management module is used to build a standardized data warehouse and realize reverse annotation of diagnostic results and dynamic optimization of thresholds; A visual interactive terminal is used to display energy consumption status, generate diagnostic reports, and support user feedback.

2. The system according to claim 1, characterized in that: The multi-source data dynamic standardization and anomaly detection module includes: a time alignment processing unit, a missing data completion unit, a data standardization unit, and an anomaly detection unit; The time alignment processing unit is used to calculate the moving average of energy consumption data based on the time of meteorological data using dynamic window alignment technology. The missing data completion unit is used to dynamically calculate the completion value based on the seasonal coefficient and weekday pattern using a weighted interpolation method based on historical data from the same period. The data normalization unit is used to employ an improved Transformer-LSTM hybrid model, including a multi-head self-attention mechanism to capture lag correlations, a bidirectional LSTM to handle temporal fluctuations, and batch normalization to output a normalized matrix. The anomaly detection unit is used to construct a decision tree using the isolated forest algorithm. When the energy consumption data deviates from the historical normal operation range by more than the preset tolerance range, and the deviation cannot be explained by external environmental factors, the system determines that the data is abnormal and triggers a recalculation and verification mechanism.

3. The system according to claim 2, characterized in that: The device-level energy consumption decomposition and root cause localization module includes: a multi-scale feature extraction unit, a device energy consumption decomposition unit, and a multi-modal root cause diagnosis unit; A multi-scale feature extraction unit is used to employ an improved Transformer-LSTM hybrid architecture, in which the Transformer layer is designed with a lag attention weight matrix and the LSTM layer introduces a time decay factor. The equipment energy consumption decomposition unit is used to output the energy consumption ratio of each device through a softmax classifier. The multimodal root cause diagnosis unit is used to output a root cause probability ranking by combining decision tree coarse screening with Bayesian network fine localization.

4. The system according to claim 3, characterized in that: The lag attention weight matrix is ​​used to capture the temporal correlation between historical meteorological features and current energy consumption status. By measuring the correlation strength between features at different times and combining a time decay mechanism to reduce the influence weight of long-term historical data, the expression of recent and periodic correlation features is strengthened.

5. The system according to claim 4, characterized in that: The user behavior perception and strategy generation module includes: a dynamic association graph construction unit, a behavior pattern recognition unit, and a strategy generation and optimization unit; The dynamic association graph construction unit is used to model the relationship between people, equipment, and environment using graph neural networks and introduces a time decay factor; The behavior pattern recognition unit is used to employ the LSTM+DBSCAN hybrid algorithm, where DBSCAN uses dynamic time warping as a distance metric. The policy generation and optimization unit is used to design reward functions that incorporate energy efficiency, comfort, and user feedback using the PPO reinforcement learning framework.

6. The system according to claim 5, characterized in that: The reward function of PPO reinforcement learning comprehensively evaluates the performance of three dimensions: energy saving effect, environmental comfort and user satisfaction. It forms a guiding signal for strategy optimization in a weighted aggregation manner. Among them, the energy saving reward reflects the degree of energy consumption reduction after the strategy is implemented, the comfort reward represents the degree of matching between indoor environmental parameters and human comfort range, and the user feedback reward reflects the occupants' subjective evaluation of the system adjustment effect.

7. The system according to claim 6, characterized in that: The system-level data closed-loop management module includes: a standardized data warehouse, a diagnostic result reverse annotation unit, and a threshold dynamic optimization unit; A standardized data warehouse is used to employ a layered storage structure, including a raw data layer, a feature engineering layer, and a baseline curve layer; The diagnostic result reverse annotation unit is used to achieve knowledge transfer using the label propagation algorithm; A threshold dynamic optimization unit is used to adaptively adjust the anomaly detection threshold based on KL divergence. In the threshold dynamic optimization unit, the threshold adjustment formula is: the threshold of the current time step is equal to the threshold of the previous time step plus the learning rate multiplied by the KL divergence between the current data distribution and the benchmark distribution.

8. The system according to claim 7, characterized in that, The visual interactive terminal includes: a 3D energy consumption heat map generation subsystem, a root cause diagnosis report automatic generator, and a strategy adjustment interactive interface; A three-dimensional energy consumption heat map generation subsystem is built based on WebGL and uses Gaussian kernel density estimation to generate spatial heat distribution. A root cause diagnosis report generator that uses a template engine and natural language generation technology to output structured reports; The strategy adjustment interface is used for parameter adjustment, real-time preview, and user feedback collection.

9. The system as claimed in claim 1, characterized in that, User feedback updates the reinforcement learning reward function using the following formula: the user feedback reward at the current time step equals the user feedback reward at the previous time step plus the learning rate coefficient multiplied by the difference between the user comfort score and the reward at the previous time step.

10. A method for intelligent monitoring and diagnosis of energy consumption in green buildings based on multi-source data fusion, characterized in that, Includes the following steps: The system performs time synchronization alignment, missing value completion, and standardization on building energy consumption data and heterogeneous meteorological and environmental data. It also detects abnormal data in real time based on an improved Transformer-LSTM hybrid model and isolated forest algorithm, and outputs high-quality standardized data. Based on the standardized data, device-level energy consumption decomposition is achieved through multi-scale feature extraction, and a multi-modal diagnostic algorithm that integrates decision trees and Bayesian networks is used to output root cause probability ranking. Construct a dynamic relationship graph of personnel, equipment, and environment; identify user behavior patterns based on LSTM and DBSCAN; and use the PPO reinforcement learning algorithm to generate equipment optimization strategies. Establish a standardized data warehouse to achieve reverse annotation of diagnostic results and dynamic optimization of anomaly detection thresholds, forming a closed-loop management of data flow; The system displays the device-level energy consumption breakdown results through a 3D heatmap, automatically generates root cause diagnosis reports, and provides a strategy adjustment interface to collect user feedback.

Images