Public building-based efficiency-increasing and carbon-reducing operation optimization system and method

Abstract

The invention relates to the technical field of public building energy management, and discloses an efficiency-increasing and carbon-reducing operation optimization system and method based on a public building. The method comprises the following steps: establishing an operation characteristic spectrum containing multi-dimensional data of equipment, a system and an environment, and carrying out time sequence slicing on historical data; and extracting a steady-state operation mode from the time slice by using a self-learning dynamic energy reference model, and generating a reference feature flow. Online operation feature flow is collected in real time, the most similar reference segment is positioned in the benchmark feature flow through a dynamic window matching algorithm, and the multi-dimensional deviation degree is calculated to generate a comprehensive abnormal index. And when the index exceeds the limit, starting a feature tracing program, performing contribution degree decomposition on the deviation degree of each dimension, and identifying a dominant abnormal dimension. And according to the dominant dimension matching optimization strategy knowledge base, outputting a specific energy efficiency optimization action instruction. According to the invention, accurate identification and root positioning of the abnormal state of the public building energy consumption system are realized, and a targeted optimization scheme can be automatically generated.

Description

Technical Field

[0001] This invention relates to the field of energy management technology for public buildings, specifically to an efficiency-enhancing and carbon-reducing operation optimization system and method for public buildings. Background Technology

[0002] In the energy conservation and carbon reduction operation management of public buildings, the common method to detect anomalies is to compare real-time energy consumption data with preset fixed benchmarks. These fixed benchmarks are usually based on design values, historical average levels, or simple time-sharing and zone-based static thresholds. However, because the operating load, environmental parameters, and usage patterns of public buildings exhibit significant dynamic and periodic changes, fixed comparison benchmarks cannot accurately reflect the normal energy consumption level of the system under current external conditions, leading to frequent false alarms and missed alarms, and making it difficult to truly identify energy efficiency degradation problems.

[0003] Existing technologies, upon detecting abnormal energy consumption, typically only provide an overall alarm or a conclusion of excessively high energy efficiency, lacking in-depth diagnosis of the root cause. Anomalies may be caused by the suboptimal operation of one or more devices across multiple subsystems, such as chillers, lighting systems, and ventilation equipment. However, conventional methods cannot distinguish the contribution of each subsystem to the overall anomaly. This makes it difficult for operation and management personnel to quickly pinpoint the problem, and subsequent optimization adjustments often rely on trial and error based on personal experience, resulting in low efficiency and uncertain effects, making it difficult to achieve precise and automated energy efficiency improvements and carbon reductions. Summary of the Invention

[0004] The purpose of this invention is to provide an efficiency-enhancing and carbon-reducing operation optimization system and method for public buildings, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides a method for optimizing the operation of public buildings to improve efficiency and reduce carbon emissions, the method comprising: Establish a multi-dimensional operational characteristic map of the energy consumption system of public buildings. The multi-dimensional operational characteristic map includes equipment-level, system-level, and environmental-level characteristic data streams related to energy consumption and carbon emissions. The historical feature data stream in the multi-dimensional running feature map is time-series sliced ​​to form continuous time slice segments with periodic regularity. Using a dynamic energy benchmark model with self-learning capabilities, steady-state operation modes are extracted from the time slice segments to generate a benchmark feature stream that reflects the inherent operating laws of the system. Real-time acquisition of online operational feature streams of public buildings, and synchronous mapping with the baseline feature streams for the corresponding time periods; A dynamic window matching algorithm is used to locate the reference segment in the baseline feature stream that has the highest similarity to the online running feature stream; The deviation between the online running feature stream and the reference segment is calculated across multiple feature dimensions, and then integrated to generate a comprehensive anomaly index. When the comprehensive anomaly index exceeds the preset dynamic tolerance threshold, the feature tracing procedure is initiated. The feature tracing program decomposes the deviation of each feature dimension that constitutes the comprehensive anomaly index into a contribution factor and identifies the dominant dimension. Based on the identification results of the dominant dimension, the system matches the preset optimization strategy knowledge base and outputs an energy efficiency optimization scheme containing specific optimization action instructions.

[0006] Preferably, the establishment of a multi-dimensional operational characteristic map of the energy consumption system of public buildings includes: Collect data on the operating current of chiller units, chilled water pump frequency, cooling tower fan power, and the opening degree of air conditioning terminal valves in each area of ​​the building's HVAC system to form an equipment-level characteristic data stream; Collect zonal illuminance sensor data, zonal lighting circuit power data, and indoor and outdoor light sensing data of the building's lighting system to form a system-level characteristic data stream; Collect data on carbon dioxide concentration distribution within buildings, indoor and outdoor temperature and humidity distribution, and heat map data on personnel distribution to form an environmental-level characteristic data stream; The device-level feature data stream, system-level feature data stream, and environment-level feature data stream are aligned and fused according to a unified timestamp to generate a multi-dimensional operational feature map.

[0007] Preferably, the step of performing time-series slicing on the historical feature data stream in the multi-dimensional running feature map includes: The historical feature data stream is divided into daily segments according to the cycle of a complete natural day; Within each daily segment, identify the start and end boundaries of peak energy consumption periods, moderate energy consumption periods, and low energy consumption periods, in hourly increments. Based on the start and end boundaries, each daily segment is further divided into short time-segment slices corresponding to different operating intensities; The characteristic data belonging to the same hourly slice in all historical data are clustered and integrated to form a representative library of typical time slice segments.

[0008] Preferably, the step of extracting the steady-state operating mode of the time slice segment using a dynamic energy benchmark model with self-learning capability includes: Segments belonging to the same season and similar weather types are selected from the time slice fragment library as the model training set; The training set drives the dynamic energy benchmark model to learn the inherent coupling relationships and change boundaries between the feature dimensions under specific external conditions. After the model converges, for each time slice, the model can generate a set of feature value sequences that reflect the time slice under ideal and efficient conditions, and the feature value sequences constitute the baseline feature flow. The dynamic energy benchmark model can update its internal coupling parameters online as new historical data is added.

[0009] Preferably, the step of employing a dynamic window matching algorithm to locate the reference segment in the baseline feature stream that has the highest similarity to the online running feature stream includes: The length of the online feature stream acquired in real time is used as the width of the initial matching window; On the baseline feature stream, the initial matching window is moved in a sliding step manner to extract multiple candidate segments; Calculate the comprehensive distance between the online running feature stream and each candidate fragment across all feature dimensions; The candidate segment with the smallest overall distance is selected as the initial reference segment; Within the time range before and after the initial reference segment, local fine matching is performed with smaller sliding steps to ultimately determine the reference segment with the highest similarity and its precise time boundary.

[0010] Preferably, the step of calculating the deviation between the online running feature stream and the reference segment across multiple feature dimensions and integrating them to generate a comprehensive anomaly index includes: For each feature dimension, the Euclidean distance between the feature value sequence of the online running feature stream and the feature value sequence of the reference segment is calculated as the basic deviation of the feature dimension; For the basic deviation of each dimension, normalization is performed by combining the normal fluctuation range of the feature dimension in historical operation to obtain the standardized deviation. A weighting coefficient is assigned to each feature dimension, and the weighting coefficient is preset according to the sensitivity of the feature dimension to overall carbon emissions; The standardized deviation of all feature dimensions is weighted and summed with their corresponding weight coefficients, and the result is used as a comprehensive anomaly index.

[0011] Preferably, the feature tracing program decomposes the deviation of each feature dimension constituting the comprehensive anomaly index into a contribution factor, identifying the dominant dimension, including: Freeze the calculation paths of the dynamic energy benchmark model, the dynamic window matching algorithm, and the comprehensive anomaly index; From the perspective of the reverse computation graph, calculate the gradient of the comprehensive anomaly index with respect to the original value of each feature dimension in the input online running feature stream; Take the absolute value of all the calculated gradient values ​​and normalize them again to obtain the relative contribution of each feature dimension to the comprehensive anomaly index. The relative contributions of all feature dimensions are ranked, and the feature dimension with the highest ranking is determined as the dominant dimension.

[0012] Preferably, the energy efficiency optimization scheme that matches the preset optimization strategy knowledge base and outputs specific optimization action instructions includes: The optimization strategy knowledge base pre-stores multiple abnormal patterns, each of which is associated with a dominant dimension combination and a set of specific optimization action instructions. The identified dominant dimensions are matched with abnormal patterns in the knowledge base to find the most similar pre-stored abnormal patterns. Extract all optimized action instructions associated with the most similar pre-stored anomaly pattern; Based on the specific values ​​of the real-time running feature stream, the parameters in the extracted optimization action instructions are adjusted and instantiated to generate an energy efficiency optimization scheme that can be directly sent to the building equipment management system.

[0013] Preferably, it also includes a closed-loop verification step for the implementation effect of the energy efficiency optimization scheme: Within a preset time period after the energy efficiency optimization scheme is implemented, new online operating characteristic streams are continuously collected; Calculate the comprehensive anomaly index corresponding to the newly collected online running feature stream; Compare the decrease in comprehensive abnormal indicators before and after implementation; If the reduction reaches the preset optimization target, the energy efficiency optimization scheme and its corresponding abnormal mode characteristics will be stored in the success case library. If the optimization target is not achieved, a secondary optimization process is triggered, which generates a corrected energy efficiency optimization scheme based on the newly collected operating feature stream.

[0014] Preferably, when the processor executes the computer program, it implements the steps of the operation optimization method based on improving efficiency and reducing carbon emissions in public buildings as described in any of the above-mentioned methods.

[0015] Compared with the prior art, the beneficial effects of the present invention are: By constructing a dynamic energy benchmark model capable of self-learning from time-series slices of historical operational data, the traditional fixed benchmark is replaced. This model automatically extracts and generates benchmark feature streams that reflect the inherent operational patterns of the system under different periods and operating conditions. This allows the reference standard for anomaly detection to adaptively update with the actual operating status of the building, eliminating benchmark mismatch problems caused by seasonal changes and usage mode shifts. This improves the accuracy of anomaly identification and reduces misjudgments caused by benchmark rigidity.

[0016] A dynamic window matching algorithm is employed to locate the most similar reference segment of real-time data in a multi-dimensional baseline feature stream. Based on this, a feature tracing program decomposes the contribution of multi-dimensional deviations in the comprehensive anomaly index, thereby identifying the dominant feature dimension causing the anomaly. This achieves a leap from "anomaly detection" to "anomaly root cause dimension location." It can clearly indicate whether the anomaly originates from specific aspects such as cooling system efficiency, lighting power, or environmental control parameters, making subsequent optimization action commands more targeted, avoiding blind adjustments, and improving the accuracy and execution efficiency of operational optimization measures. Attached Figure Description

[0017] Figure 1 This is a schematic diagram illustrating the working principle of the public building efficiency improvement and carbon reduction operation optimization method described in this invention.

[0018] Figure 2 A flowchart for establishing a multi-dimensional operational feature map.

[0019] Figure 3 A flowchart for extracting steady-state operating modes from a dynamic energy benchmark model.

[0020] Figure 4 A standardized deviation bar chart for the multi-dimensional characteristics of public buildings.

[0021] Figure 5 A bar chart showing the energy consumption percentage of various equipment systems in a public building. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Please see Figure 1This invention provides a method for optimizing the operation of public buildings to improve efficiency and reduce carbon emissions. The method includes: establishing a multi-dimensional operational feature map of the public building's energy system, which contains equipment-level, system-level, and environmental-level feature data streams related to energy consumption and carbon emissions; performing time-series slicing on the historical feature data streams in the multi-dimensional operational feature map to form continuous and periodic time slices; using a dynamic energy benchmark model with self-learning capabilities to extract steady-state operation modes from the time slices, generating a benchmark feature stream reflecting the inherent operating laws of the system; collecting the online operational feature stream of the public building in real time and synchronously mapping it with the benchmark feature stream for the corresponding time period; using a dynamic window matching algorithm to locate the reference segment in the benchmark feature stream that has the highest similarity to the online operational feature stream; calculating the deviation between the online operational feature stream and the reference segment in multiple feature dimensions and integrating them to generate a comprehensive anomaly index; and initiating a feature tracing program when the comprehensive anomaly index exceeds a preset dynamic tolerance threshold. The feature tracing program decomposes the deviation of each feature dimension constituting the comprehensive anomaly index into a contribution factor, identifying the dominant dimension. Based on the identification results of the dominant dimension, the system matches the preset optimization strategy knowledge base and outputs an energy efficiency optimization scheme containing specific optimization action instructions.

[0024] In one embodiment of the present invention, see [reference] Figure 2 Establishing a multi-dimensional operational characteristic map of a public building's energy consumption system involves the following processes: First, collecting data on the operating current of chiller units, chilled water pump frequency, cooling tower fan power, and the opening degree of air conditioning terminal valves in each area of ​​the building's HVAC system to form an equipment-level characteristic data stream. Second, collecting data on zonal illuminance sensors, zonal lighting circuit power, and indoor and outdoor light sensing data for the building's lighting system to form a system-level characteristic data stream. Third, collecting data on carbon dioxide concentration distribution, indoor and outdoor temperature and humidity distribution, and occupant distribution heatmap data to form an environmental-level characteristic data stream. The equipment-level, system-level, and environmental-level characteristic data streams are then aligned and fused using a unified timestamp to generate a multi-dimensional operational characteristic map. The historical characteristic data stream in the multi-dimensional operational characteristic map is then time-series sliced, dividing the historical characteristic data stream into daily segments according to the cycle of a complete natural day. Within each daily segment, the start and end boundaries of peak energy consumption periods, periods of moderate energy consumption, and periods of low energy consumption are identified on an hourly basis. Based on start and end boundaries, each daily segment is further divided into smaller time-segment slices corresponding to different operational intensities. Feature data belonging to the same hourly segment from all historical data are clustered and integrated to form a representative library of typical time-segment segments.

[0025] In practical implementation, taking an office building with a central air conditioning system, intelligent lighting system, and environmental monitoring system as an example, this paper illustrates the process of constructing a multi-dimensional operational characteristic map. During the operation of the office building, data is collected in real time from the building equipment management system and the Internet of Things (IoT) sensor network. This includes data on the chiller operating current, chilled water pump frequency, cooling tower fan power, and the opening degree of air conditioning terminal valves in each area of ​​the building's HVAC system, forming an equipment-level characteristic data stream. This data is recorded once per minute, with chiller operating current in amperes, chilled water pump frequency in hertz, cooling tower fan power in kilowatts, and air conditioning terminal valve opening expressed as a percentage. Data from the building's lighting system, including zoned illuminance sensor data, zoned lighting circuit power data, and indoor and outdoor light sensing data, forms a system-level characteristic data stream. Zoned illuminance sensor data is in lux, zoned lighting circuit power data is in kilowatts, and indoor and outdoor light sensing data are dimensionless brightness values. Finally, data on carbon dioxide concentration distribution, indoor and outdoor temperature and humidity distribution, and personnel distribution heatmap data are collected to form an environmental-level characteristic data stream. Carbon dioxide concentration data is expressed in ppm, while temperature and humidity data are expressed in degrees Celsius and percentages, respectively. The population distribution heatmap data is generated by regional population statistics sensors, representing the population density in each area. In practice, device-level, system-level, and environmental-level feature data streams are aligned and fused according to a unified timestamp to generate a multi-dimensional operational feature map. The alignment and fusion process is performed in a time-series database, where each row of data corresponds to a timestamp accurate to the second, and each column corresponds to a specific feature dimension, thus forming a structured map containing multi-dimensional time-series data.

[0026] In some embodiments, the time-series slicing process is based on historical multi-dimensional operational feature maps. Time-series slicing of the historical feature data stream in the multi-dimensional operational feature map includes the following operations: dividing the historical feature data stream into daily segments according to the cycle of a complete natural day. The start point of each daily segment is typically set to 00:00:00 of the current day, and the end point is set to 00:00:00 of the next day, forming a data block containing complete 24-hour data. Within each daily segment, the start and end boundaries of peak energy consumption periods, moderate energy consumption periods, and low energy consumption periods are identified in hourly units. The identification process is based on historical time-series data of the building's total electricity consumption. A clustering algorithm identifies consecutive hourly periods with power consistently above a first set threshold as peak energy consumption periods, consecutive hourly periods with power consistently below a second set threshold as low energy consumption periods, and the remaining periods as moderate energy consumption periods. It can be understood that the first and second set thresholds are determined by analyzing the statistical distribution of historical power data. Based on the start and end boundaries of the identified peak energy consumption periods, moderate energy consumption periods, and low energy consumption periods, each daily segment is further divided into smaller time-segment slices corresponding to different operating intensities. For example, a daily segment is divided into multiple smaller time-segment slices with specific type labels, such as "Weekday - Peak Period", "Weekday - Moderate Period", and "Weekday - Low Period".

[0027] In some embodiments, forming a typical time slice library involves integrating a large amount of historical slice data. Feature data belonging to the same hourly segment from all historical data are clustered and integrated to form a representative typical time slice library. Clustering and integration are performed separately for each feature dimension. For short-time slices with the same type of label, the corresponding feature data sequences for all historical dates within that time period are extracted. Similar sequences are aggregated by calculating the similarity distance between these sequences. A method for calculating the clustering distance between sequences is described. The method is shown in the formula:

[0028] in: The cluster distance between two sequences is represented by T, which represents the total number of time points contained in the hourly slice. This represents the eigenvalue of the first sequence at time t. This represents the eigenvalue of the second sequence at time t. This represents the average of the feature values ​​of all sequences to be aggregated at time t. Optionally, the clustering process uses a hierarchical clustering algorithm to group sequences whose distance is less than a set clustering threshold into one class. From each class, the actual historical sequence with the smallest distance to the center sequence of that class is selected as the representative time slice segment of that class and stored in the typical time slice segment library.

[0029] In one embodiment of the present invention, see [reference] Figure 3This study utilizes a self-learning dynamic energy benchmark model to extract steady-state operating modes from time slice segments. The process includes selecting segments from a time slice library that fall within the same season and similar weather types as the model training set. The training set drives the dynamic energy benchmark model to learn the inherent coupling relationships and change boundaries between various feature dimensions under specific external conditions. After model convergence, for each time slice segment, the model generates a set of feature numerical sequences reflecting the ideal, efficient state of the time slice segment. These feature numerical sequences constitute the benchmark feature flow. The dynamic energy benchmark model can update its internal coupling parameters online with newly added historical data.

[0030] In its implementation, the dynamic energy benchmark model employs a neural network-based sequence generation model structure. The model's input consists of time slices labeled with environmental conditions, and its output is the corresponding benchmark feature stream. Slices belonging to the same season and similar weather types are selected from a time slice library as the model's training set. Each slice in the library is labeled with its respective season and weather type; seasons can be categorized as spring, summer, autumn, and winter, and weather types can be categorized as sunny, cloudy, rainy, and snowy based on meteorological data. For example, to train a dynamic energy benchmark model suitable for sunny summer days, the trainers extract all typical time slices labeled "summer" and "sunny" from the time slice library, forming the sample set for model training. The training set drives the dynamic energy benchmark model to learn the inherent coupling relationships and change boundaries between various feature dimensions under specific external conditions.

[0031] In some embodiments, the learning process of the dynamic energy baseline model is specified. Coupling relationships refer to the dynamic correlations between different feature variables; for example, the operating current of the chiller unit and the frequency of the chilled water pump should maintain a proportional range under efficient operating conditions, and there is a linkage between indoor carbon dioxide concentration and the opening degree of the fresh air valve. Change boundaries refer to the upper and lower limits of the reasonable values ​​for each feature dimension under specific external conditions. The model learns these relationships through a multi-layer network structure, and the loss function used in training... Aimed at measuring the deviation between model output and ideal state, the Dynamic Energy Benchmark Model (VEBM) learns the coupling relationships and change boundaries between various feature dimensions through a multi-layered network structure. Its implementation is based on a neural network architecture. The model input consists of time slices labeled with season and weather, drawn from a library of typical time slices to ensure training data is under similar external conditions. Hidden layers in the multi-layered network process the input features level by level, capturing the dynamic correlations between equipment-level, system-level, and environmental-level feature data streams through nonlinear transformations. Examples include the coordinated change patterns of chiller operating current and chilled water pump frequency, or the linkage between carbon dioxide concentration and fresh air valve opening. The network output layer generates a benchmark feature stream, reflecting the feature value sequence under ideal high-efficiency conditions. During training, the model aims to minimize the loss function, adjusting network weights through backpropagation to make the output sequence approximate historical high-efficiency operating data, thereby learning the inherent coupling relationships and reasonable value boundaries of each feature dimension. The formula is expressed as follows:

[0032] in: The loss value represents the baseline model, N represents the number of time slices in the training set, and D represents the total number of feature dimensions. This represents the true efficiency value of the i-th training segment in the d-th feature dimension. This represents the benchmark value generated by the dynamic energy benchmark model for the i-th training segment on the d-th feature dimension. This represents the importance weight of the d-th feature dimension in the loss calculation. It can be understood that the true efficient value... It is actual operating data selected from historical data during the period when the system's energy efficiency was at its best under the same external conditions.

[0033] In some embodiments, a baseline feature stream is generated after model convergence. After convergence, for each time slice, the model generates a set of feature value sequences that reflect the ideal efficient state of the time slice. These feature value sequences constitute the baseline feature stream. For any time slice in the library with seasonal and weather labels, it is input into the corresponding trained dynamic energy baseline model. The model outputs predicted values ​​for each feature dimension step by step. These predicted values ​​are concatenated to form the baseline feature stream corresponding to that slice. It can be understood that the baseline feature stream is not a single value, but a complete data sequence containing all feature dimensions with the same length as the original time slice. It represents the theoretically optimal or most stable operating trajectory of the system under specific external conditions during that specific time period. The dynamic energy baseline model can update its internal coupling parameters online with newly added historical data. Optionally, the online update mechanism adopts an incremental learning algorithm. When the system confirms that its operating state is efficient and stable within a certain period, the multi-dimensional operating feature map data corresponding to that period is automatically labeled as new positive samples and injected into the training process of the dynamic energy baseline model at a certain frequency to fine-tune the model parameters. Optionally, multiple dynamic energy benchmark models can be established to address different scenarios. Independent dynamic energy benchmark models can be trained for different seasons, different building functional areas, or different operating modes, forming a model group. In practical applications, the corresponding dynamic energy benchmark model is called based on the real-time identified scenario labels to generate the benchmark feature stream.

[0034] In one embodiment of the present invention, a dynamic window matching algorithm is used to locate the reference segment with the highest similarity to the online running feature stream in the benchmark feature stream. This involves the following steps: using the time length of the real-time acquired online running feature stream as the width of the initial matching window; moving the initial matching window in a sliding step manner on the benchmark feature stream to extract multiple candidate segments; calculating the comprehensive distance between the online running feature stream and each candidate segment across all feature dimensions; selecting the candidate segment with the smallest comprehensive distance as the initial reference segment; performing local fine-tuning matching within the time range before and after the initial reference segment with a smaller sliding step to ultimately determine the reference segment with the highest similarity and its precise time boundary; and calculating the deviation between the online running feature stream and the reference segment across multiple feature dimensions and integrating them to generate a comprehensive anomaly index. This involves the following steps: for each feature dimension, calculating the Euclidean distance between the feature value sequence of the online running feature stream and the feature value sequence of the reference segment as the basic deviation of the feature dimension; and normalizing the basic deviation of each dimension by combining it with the normal fluctuation range of the feature dimension in historical operation to obtain the standardized deviation. Each feature dimension is assigned a weight coefficient, which is pre-set based on the sensitivity of the feature dimension to overall carbon emissions. The standardized deviations of all feature dimensions are weighted and summed with their corresponding weight coefficients, and the result is used as a comprehensive anomaly index.

[0035] In its implementation, the dynamic window matching algorithm begins by acquiring the real-time online running feature stream and the corresponding time-segment baseline feature stream. The baseline feature stream is generated by the dynamic energy baseline model based on the current season, weather type, and time slice type. The length of the real-time online running feature stream is used as the width of the initial matching window. For example, if the online running feature stream is a data sequence from the past 30 minutes, the width of the initial matching window is set to 30 minutes. The initial matching window is moved along the baseline feature stream using a sliding step, extracting multiple candidate segments. The sliding step can be set to 1 minute. Starting from the beginning of the baseline feature stream, the window slides backward by 1 minute each time until the end of the window reaches the end of the baseline feature stream, thus obtaining a series of candidate baseline segments of the same length as the online running feature stream. The comprehensive distance between the online running feature stream and each candidate segment is calculated across all feature dimensions. The calculation formula is as follows:

[0036] in: Indicates the online running feature flow and the k-th The comprehensive distance between candidate segments, where D represents the total number of feature dimensions. This represents the matching weight coefficient pre-set for the d-th feature dimension, and T represents the total number of time points within the window. This represents the feature value of the online running feature stream at time t and dimension d. Let represent the feature value of the k-th candidate segment at time t and dimension d. The candidate segment with the smallest comprehensive distance is selected as the initial reference segment, which represents the historical steady-state pattern that is most similar to the current real-time running state in the baseline feature stream.

[0037] In some embodiments, the dynamic window matching algorithm includes a step of local fine-grained matching. Within the time range before and after the initial reference segment, local fine-grained matching is performed with smaller sliding steps to ultimately determine the reference segment with the highest similarity and its precise time boundaries. For example, a candidate interval is formed by extending 15 minutes before and after the center time point of the initial reference segment. Within this interval, the matching window is moved with a sliding step of 10 seconds, and the comprehensive distance at each position is recalculated. The minimum will be obtained The window position of the value is determined as the final reference segment. In some embodiments, the calculation of the comprehensive anomaly index is based on the difference between the online operating feature stream and the finally determined reference segment. Calculating the deviation between the online operating feature stream and the reference segment across multiple feature dimensions and integrating them to generate a comprehensive anomaly index includes the following operations: For each feature dimension, the Euclidean distance between the feature value sequence of the online operating feature stream and the feature value sequence of the reference segment is calculated as the basic deviation of the feature dimension. The basic deviation of each dimension is normalized by combining the normal fluctuation range of the feature dimension in historical operation to obtain the standardized deviation. The normal fluctuation range is obtained by statistically analyzing the standard deviation or quantile of the dimension's values ​​in historical steady-state data. Normalization maps the basic deviation to a unitless relative scale. A weight coefficient is assigned to each feature dimension. The weight coefficient is pre-set according to the sensitivity of the feature dimension to overall carbon emissions; for example, the weight coefficient for chiller unit operating current is higher than the weight coefficient for lighting circuit power.

[0038] Optionally, the standardized deviation can be calculated using different mathematical expressions. One approach is to divide the base deviation by the standard deviation of that feature dimension in historical steady-state data, so that the magnitude of the standardized deviation can intuitively reflect the multiple of the current deviation relative to historical normal fluctuations. It can be understood that the weighting coefficients used in the weighted summation are related to the calculation of the overall distance. Matching weight coefficients used at the time It can be the same set of parameters, or two independent parameter systems, used for matching similarity and evaluating the degree of anomaly, respectively.

[0039] See Figure 4This is a standardized deviation bar chart of multi-dimensional features of public buildings, primarily showing the degree of deviation of each feature dimension relative to historical benchmarks. The standardized deviation of equipment-level features (chiller current, chilled water pump frequency) is significantly higher than other features (all exceeding 12), classifying them as high-deviation dimensions; the deviations of system-level and environmental-level features are close to 0, falling within the normal fluctuation range. This chart is used in the deviation calculation phase of public buildings, intuitively locating abnormal feature dimensions and helping technicians quickly identify high-risk links in the energy system. It serves as a core reference tool for subsequent anomaly tracing and optimization strategy matching.

[0040] In one embodiment of the present invention, the feature tracing program decomposes the deviation of each feature dimension constituting the comprehensive anomaly index into a contribution factor, identifies the dominant dimension, and includes the following process: freezing the dynamic energy benchmark model, the dynamic window matching algorithm, and the calculation path of the comprehensive anomaly index. From the perspective of the reverse computation graph, the gradient of the comprehensive anomaly index with respect to the original value of each feature dimension in the input online running feature flow is calculated. The absolute value of all calculated gradient values ​​is taken and normalized again to obtain the relative contribution of each feature dimension to the comprehensive anomaly index. The relative contributions of all feature dimensions are sorted, and the feature dimension with the highest ranking is determined as the dominant dimension. Matching a preset optimization strategy knowledge base and outputting an energy efficiency optimization scheme containing specific optimization action instructions includes the following process: the optimization strategy knowledge base pre-stores multiple anomaly patterns, each anomaly pattern is associated with a combination of dominant dimensions and a set of specific optimization action instructions. The identified dominant dimension is matched with the anomaly patterns in the knowledge base to find the most similar pre-stored anomaly pattern. All optimization action instructions associated with the most similar pre-stored anomaly pattern are extracted. Based on the specific values ​​of the real-time running feature stream, the parameters in the extracted optimization action instructions are adjusted and instantiated to generate an energy efficiency optimization scheme that can be directly sent to the building equipment management system.

[0041] In practice, the feature tracing procedure is initiated when the comprehensive anomaly index exceeds a preset dynamic tolerance threshold. The procedure decomposes the deviation of each feature dimension constituting the comprehensive anomaly index into its contribution, identifying the dominant dimension. The specific operation of the feature tracing procedure includes the following steps: freezing the dynamic energy baseline model, the dynamic window matching algorithm, and the computation path of the comprehensive anomaly index. This means that during the tracing calculation, the internal parameters and intermediate computation states of the aforementioned models and algorithms are fixed, ensuring that gradient calculations are based on the current anomaly state. From the perspective of the reverse computation graph, the gradient of the comprehensive anomaly index with respect to the original values ​​of each feature dimension in the online feature stream is calculated. The gradient value reflects the instantaneous sensitivity of the comprehensive anomaly index relative to the input value of each feature dimension. The absolute values ​​of all calculated gradient values ​​are taken and normalized again to obtain the relative contribution of each feature dimension to the comprehensive anomaly index. The relative contributions of all feature dimensions are sorted, and the feature dimension with the highest ranking is determined as the dominant dimension. (Relative contribution) The calculation formula is expressed as follows:

[0042] in: This represents the relative contribution of the i-th feature dimension to the comprehensive anomaly index. This represents the gradient of the comprehensive anomaly index with respect to the original value of the i-th feature dimension, where N represents the total number of feature dimensions. The gradient can be understood as... Obtained by backpropagation over the entire frozen computation graph using automatic differentiation techniques.

[0043] In some embodiments, the identification results of the dominant dimension are used to match the optimization strategy knowledge base. Matching the preset optimization strategy knowledge base and outputting an energy efficiency optimization scheme containing specific optimization action instructions includes the following process: the optimization strategy knowledge base pre-stores multiple abnormal patterns, and each abnormal pattern is associated with a combination of dominant dimensions and a set of specific optimization action instructions. See Table 1.

[0044] Table 1: An exemplary storage structure table

[0045] It is understandable that the dominant dimension combination is the key index for identifying anomalous patterns, while the optimization action instructions are a set of pre-defined, executable control commands for that pattern. The identified dominant dimensions are matched against anomalous patterns in the optimization strategy knowledge base to find the most similar pre-stored anomalous pattern. The matching process can calculate the similarity between the current dominant dimension set and the dominant dimension set of each pre-stored pattern in the knowledge base, for example, using the Jaccard similarity coefficient, and selecting the pre-stored pattern with the highest similarity as the matching result. All optimization action instructions associated with the most similar pre-stored anomalous pattern are extracted; these instructions are textualized, parameterized operation descriptions.

[0046] In some embodiments, the instantiation of optimization action instructions generates the final energy efficiency optimization scheme. Based on the specific values ​​of the real-time operating feature stream, the parameters in the extracted optimization action instructions are adjusted and instantiated to generate an energy efficiency optimization scheme that can be directly sent to the building equipment management system. For example, if the abnormal pattern PATTERN_01 is matched, the extracted optimization action instructions include the parameters "baseline value X℃" and "reduce chilled water pump frequency ZHz". The feature tracing program reads the actual set temperature value of the chiller unit and the actual value of the chilled water supply and return temperature difference in the current online operating feature stream, and calculates the specific adjustment parameters by combining them with the corresponding base values ​​in the base feature stream generated by the dynamic energy base model. Optionally, the calculated specific adjustment parameters and the complete operation instructions are encapsulated into a standard format message recognizable by the building equipment management system, such as BACnet or Modbus instructions. The final generated energy efficiency optimization scheme is a complete instruction set containing specific operation targets, execution device addresses, set parameter values, and execution timing.

[0047] In one embodiment of the present invention, the closed-loop verification step of the energy efficiency optimization scheme execution effect includes the following process: Within a preset time period after the energy efficiency optimization scheme is executed, new online operating feature streams are continuously collected. The comprehensive anomaly index corresponding to the newly collected online operating feature streams is calculated. The decrease in the comprehensive anomaly index before and after execution is compared. If the decrease reaches a preset optimization target, the energy efficiency optimization scheme and its corresponding anomaly mode features are stored in a success case database. If the optimization target is not reached, a secondary optimization process is triggered, which generates a corrected energy efficiency optimization scheme based on the newly collected operating feature streams.

[0048] In practical implementation, new online operational characteristic streams are continuously collected within a preset time period after the energy efficiency optimization plan is implemented. The length of the preset time period is determined based on system inertia; for example, for HVAC systems, the preset time period is set to 60 to 180 minutes after the optimization command is executed. During this period, the data acquisition system continues to collect building operation data at the same frequency and dimensions as before, forming new online operational characteristic streams reflecting the optimized state. The comprehensive anomaly index corresponding to the newly collected online operational characteristic stream is calculated. This involves dynamically matching the new online operational characteristic stream with the newly generated baseline characteristic stream in the same time period, calculating the deviation, and performing weighted integration to obtain a new comprehensive anomaly index value. The decrease in the comprehensive anomaly index before and after implementation is compared. The formula for calculating the decrease R is as follows:

[0049] Where: R represents the magnitude of the decrease in the comprehensive abnormality index. This represents the original comprehensive anomaly index value calculated before implementing the energy efficiency optimization plan. This represents the new comprehensive abnormal index value calculated within a preset time period after implementing the energy efficiency optimization plan. It can be understood that the decrease R is a percentage value; the larger the value, the more significant the positive effect of the optimization plan.

[0050] In some embodiments, closed-loop verification performs triage based on the comparison between the reduction magnitude R and a preset optimization target. If the reduction magnitude reaches the preset optimization target, the energy efficiency optimization scheme and its corresponding abnormal mode characteristics are stored in the success case library. The preset optimization target is a predefined threshold, such as requiring the reduction magnitude R to be no less than 15%. The information stored in the success case library includes the original abnormal mode characteristics that triggered this optimization, the complete set of energy efficiency optimization scheme instructions executed, and the final reduction magnitude R value achieved.

[0051] Optionally, different strategies exist for implementing the secondary optimization process. One strategy is to restart the entire process from feature tracing to strategy matching, recalculating the comprehensive anomaly index based on the latest online operational feature flow, re-identifying the dominant dimension, and matching it to different anomaly patterns in the optimization strategy knowledge base, thereby generating a new set of optimization action instructions. Another strategy is to adjust parameters based on the initial optimization plan. For example, if the initial optimization plan is to adjust the chiller unit's set temperature by 2°C, but the effect is not as expected, the secondary optimization process attempts to modify the adjustment range to 3°C, or add another auxiliary adjustment instruction. The secondary optimization process generates a corrected energy efficiency optimization plan based on the newly collected operational feature flow. The corrected energy efficiency optimization plan will be issued and executed again, entering a new round of closed-loop verification. In some embodiments, a maximum number of optimization iterations can be set to avoid getting stuck in an infinite loop in uncorrectable fault scenarios.

[0052] See Figure 5 This is a bar chart showing the energy consumption percentage of various equipment systems in a public building, primarily illustrating the contribution of different equipment to the building's total energy consumption. Chillers account for the highest energy consumption (nearly 35%), making them the core component of building energy use; chilled water pumps (approximately 22%) and cooling towers (approximately 18%) follow, all three belonging to the HVAC system, accounting for over 75% combined; lighting and ventilation systems have lower energy consumption percentages. This chart is used in the characteristic analysis phase of public building energy systems, visually presenting the energy consumption weight of each piece of equipment, helping technicians clarify the priorities for energy efficiency optimization, and serving as a fundamental reference tool for formulating efficiency improvement and carbon reduction strategies.

[0053] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0054] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An operational optimization method for improving efficiency and reducing carbon emissions in public buildings, characterized in that: Includes the following steps: Establish a multi-dimensional operational characteristic map of the energy consumption system of public buildings. The multi-dimensional operational characteristic map includes equipment-level, system-level, and environmental-level characteristic data streams related to energy consumption and carbon emissions. The historical feature data stream in the multi-dimensional running feature map is time-series sliced ​​to form continuous time slice segments with periodic regularity. Using a dynamic energy benchmark model with self-learning capabilities, steady-state operation modes are extracted from the time slice segments to generate a benchmark feature stream that reflects the inherent operating laws of the system. Real-time acquisition of online operational feature streams of public buildings, and synchronous mapping with the baseline feature streams for the corresponding time periods; A dynamic window matching algorithm is used to locate the reference segment in the baseline feature stream that has the highest similarity to the online running feature stream; The deviation between the online running feature stream and the reference segment is calculated across multiple feature dimensions, and then integrated to generate a comprehensive anomaly index. When the comprehensive anomaly index exceeds the preset dynamic tolerance threshold, the feature tracing procedure is initiated. The feature tracing program decomposes the deviation of each feature dimension that constitutes the comprehensive anomaly index into a contribution factor and identifies the dominant dimension. Based on the identification results of the dominant dimension, the system matches the preset optimization strategy knowledge base and outputs an energy efficiency optimization scheme containing specific optimization action instructions.

2. The operation optimization method for improving efficiency and reducing carbon emissions in public buildings according to claim 1, characterized in that, The establishment of a multi-dimensional operational characteristic map of the energy consumption system of public buildings includes: Collect data on the operating current of chiller units, chilled water pump frequency, cooling tower fan power, and the opening degree of air conditioning terminal valves in each area of ​​the building's HVAC system to form an equipment-level characteristic data stream; Collect zonal illuminance sensor data, zonal lighting circuit power data, and indoor and outdoor light sensing data of the building's lighting system to form a system-level characteristic data stream; Collect data on carbon dioxide concentration distribution within buildings, indoor and outdoor temperature and humidity distribution, and heat map data on personnel distribution to form an environmental-level characteristic data stream; The device-level feature data stream, system-level feature data stream, and environment-level feature data stream are aligned and fused according to a unified timestamp to generate a multi-dimensional operational feature map.

3. The operation optimization method for improving efficiency and reducing carbon emissions in public buildings according to claim 1, characterized in that, The step of performing time-series slicing on the historical feature data stream in the multi-dimensional running feature map includes: The historical feature data stream is divided into daily segments according to the cycle of a complete natural day; Within each daily segment, identify the start and end boundaries of peak energy consumption periods, moderate energy consumption periods, and low energy consumption periods, in hourly increments. Based on the start and end boundaries, each daily segment is further divided into short time-segment slices corresponding to different operating intensities; The characteristic data belonging to the same hourly slice in all historical data are clustered and integrated to form a representative library of typical time slice segments.

4. The operation optimization method for improving efficiency and reducing carbon emissions in public buildings according to claim 1, characterized in that, The step of extracting the steady-state operating mode of the time slice segment using a dynamic energy benchmark model with self-learning capability includes: Segments belonging to the same season and similar weather types are selected from the time slice fragment library as the model training set; The training set drives the dynamic energy benchmark model to learn the inherent coupling relationships and change boundaries between the feature dimensions under specific external conditions. After the model converges, for each time slice, the model can generate a set of feature value sequences that reflect the time slice under ideal and efficient conditions, and the feature value sequences constitute the baseline feature flow. The dynamic energy benchmark model can update its internal coupling parameters online as new historical data is added.

5. The operation optimization method for improving efficiency and reducing carbon emissions in public buildings according to claim 1, characterized in that, The dynamic window matching algorithm is used to locate the reference segment in the baseline feature stream that has the highest similarity to the online running feature stream, including: The length of the online feature stream acquired in real time is used as the width of the initial matching window; On the baseline feature stream, the initial matching window is moved in a sliding step manner to extract multiple candidate segments; Calculate the comprehensive distance between the online running feature stream and each candidate fragment across all feature dimensions; The candidate segment with the smallest overall distance is selected as the initial reference segment; Within the time range before and after the initial reference segment, local fine matching is performed with smaller sliding steps to ultimately determine the reference segment with the highest similarity and its precise time boundary.

6. The operation optimization method for improving efficiency and reducing carbon emissions in public buildings according to claim 5, characterized in that, The calculation of the deviation between the online running feature stream and the reference segment across multiple feature dimensions, and the integration to generate a comprehensive anomaly index, includes: For each feature dimension, the Euclidean distance between the feature value sequence of the online running feature stream and the feature value sequence of the reference segment is calculated as the basic deviation of the feature dimension; For the basic deviation of each dimension, normalization is performed by combining the normal fluctuation range of the feature dimension in historical operation to obtain the standardized deviation. A weighting coefficient is assigned to each feature dimension, and the weighting coefficient is preset according to the sensitivity of the feature dimension to overall carbon emissions; The standardized deviation of all feature dimensions is weighted and summed with their corresponding weight coefficients, and the result is used as a comprehensive anomaly index.

7. The operation optimization method for improving efficiency and reducing carbon emissions in public buildings according to claim 1, characterized in that, The feature tracing program decomposes the deviation of each feature dimension constituting the comprehensive anomaly index into its contribution, identifying the dominant dimension, including: Freeze the calculation paths of the dynamic energy benchmark model, the dynamic window matching algorithm, and the comprehensive anomaly index; From the perspective of the reverse computation graph, calculate the gradient of the comprehensive anomaly index with respect to the original value of each feature dimension in the input online running feature stream; Take the absolute value of all the calculated gradient values ​​and normalize them again to obtain the relative contribution of each feature dimension to the comprehensive anomaly index. The relative contributions of all feature dimensions are ranked, and the feature dimension with the highest ranking is determined as the dominant dimension.

8. The operation optimization method for improving efficiency and reducing carbon emissions in public buildings according to claim 7, characterized in that, The matching preset optimization strategy knowledge base outputs energy efficiency optimization schemes containing specific optimization action instructions, including: The optimization strategy knowledge base pre-stores multiple abnormal patterns, each of which is associated with a dominant dimension combination and a set of specific optimization action instructions. The identified dominant dimensions are matched with abnormal patterns in the knowledge base to find the most similar pre-stored abnormal patterns. Extract all optimized action instructions associated with the most similar pre-stored anomaly pattern; Based on the specific values ​​of the real-time running feature stream, the parameters in the extracted optimization action instructions are adjusted and instantiated to generate an energy efficiency optimization scheme that can be directly sent to the building equipment management system.

9. The operation optimization method for improving efficiency and reducing carbon emissions in public buildings according to claim 1, characterized in that, It also includes a closed-loop verification step for the implementation effect of energy efficiency optimization solutions: Within a preset time period after the energy efficiency optimization scheme is implemented, new online operating characteristic streams are continuously collected; Calculate the comprehensive anomaly index corresponding to the newly collected online running feature stream; Compare the decrease in comprehensive abnormal indicators before and after implementation; If the reduction reaches the preset optimization target, the energy efficiency optimization scheme and its corresponding abnormal mode characteristics will be stored in the success case library. If the optimization target is not achieved, a secondary optimization process is triggered, which generates a corrected energy efficiency optimization scheme based on the newly collected operating feature stream.

10. An operational optimization system for improving efficiency and reducing carbon emissions in public buildings, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the operation optimization method based on improving efficiency and reducing carbon emissions in public buildings as described in any one of claims 1 to 9.

Images