Building energy efficiency digital management system and method based on big data

Abstract

The invention discloses a building energy efficiency digital management system and method based on big data. Acquiring and preprocessing historical and real-time operation data of the building; carrying out probability prediction on the environment quantity of each partition of the prediction time domain by utilizing a time sequence fusion Transform, and outputting a prediction quantile, a comfort default probability and a default duration probability; carrying out online calibration on the quantile and the probability based on the actual environment quantity in a calibration period; distributing a default probability upper limit of each moment of each partition based on the total comfort risk budget in combination with the partition risk level and the occupancy parameter, and generating a chance constraint; identifying a short-period start-stop event, calculating an equivalent damage index, and mapping the equivalent damage index into a differentiable segmented convex penalty function to obtain a life loss model; opportunity constraint multi-target model predictive control is constructed and solved, and a controlled quantity is issued and is updated in a rolling manner, so that the comfort risk is controllable, the energy consumption is reduced, and the start-stop damage is reduced.

Description

Technical Field

[0001] This invention relates to the field of building energy management, and in particular to a digital management system and method for building energy efficiency based on big data. Background Technology

[0002] Building operational energy consumption accounts for a significant portion of total social end-use energy consumption, with HVAC systems, cold and heat source systems, and electromechanical equipment such as fans and pumps being the main energy consumers. With the widespread application of building automation systems and building energy management systems, building sites are gradually gaining the ability to continuously collect, transmit, and store data on zoned temperature, humidity, energy consumption, equipment operating status, equipment control parameters, occupancy information, and outdoor weather conditions. Based on this, existing technologies have evolved from timing strategies, empirical rule control, and proportional-integral-derivative (PID) control to optimization control based on mechanistic models or data-driven models. Simultaneously, prediction methods for load and indoor environment have evolved from traditional statistical models to deep learning models, and probabilistic prediction methods considering prediction uncertainty have emerged, along with model predictive control, robust control, and chance-constrained control that combine prediction results to optimize energy consumption and indoor environmental quality under constraints.

[0003] Existing technologies still have the following shortcomings in multi-zone energy efficiency management of buildings: First, comfort is difficult to quantify and effectively incorporate into control decisions. Many solutions treat comfort as a post-event evaluation indicator or only use deterministic temperature and humidity threshold constraints, making it difficult to incorporate the risk of comfort exceeding limits into optimal control in a calculable and constrained form; moreover, probabilistic predictions are prone to deviations during operation, making it difficult to reliably realize opportunity constraints.

[0004] Second, there is a lack of a comprehensive comfort risk management mechanism that addresses multiple zones and time periods. Existing strategies often apply uniform or fixed default limits to different zones and different time periods, making it difficult to dynamically allocate limits based on the importance of the zone and the intensity of occupancy. This can easily lead to overly conservative control or uneven comfort guarantees between regions.

[0005] Third, the description of the impact on equipment lifespan is rather rough. Many solutions approximate equipment wear and tear by the number of start-stop cycles or the duration of operation, making it difficult to identify the accelerated degradation caused by frequent start-stop cycles in short periods and drastic changes in control variables. This may lead to frequent start-stop cycles and reduced equipment lifespan during energy-saving or comfort optimization processes.

[0006] Therefore, a digital management method and system for building energy efficiency that can overcome the shortcomings of the existing technology is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose a digital management system and method for building energy efficiency based on big data. Addressing the challenges of unifying and quantifying energy consumption, comfort, and equipment lifespan targets for coordinated optimization in existing technologies, the difficulty in realizing comfort exceedance risks due to fixed threshold settings for comfort constraints, and the inability to suppress short-cycle start-stop degradation by approximating lifespan loss solely through start-stop frequency, this invention proposes a method that uses a time-series fusion Transformer model for probabilistic prediction of zonal environmental quantities and adds a comfort default risk output to the output. It also involves online calibration of the predicted quantiles and comfort default probabilities using a calibration period. Based on the total comfort risk budget and combined with zonal risk levels and occupancy parameters, it dynamically allocates the upper limit of comfort default probabilities for each zone at each time point to construct opportunity constraints. Furthermore, it identifies short-cycle start-stop events, calculates the equivalent damage index, and maps it to a differentiable piecewise convex penalty function to form a lifespan loss model. Finally, it constructs and solves the output control quantity for the multi-objective model predictive control problem with opportunity constraints. This invention achieves the technical effects of making comfort exceedance risks quantifiable and controllable, reducing building energy consumption while satisfying equipment operation constraints and comfort opportunity constraints, and reducing short-cycle start-stops and equivalent damage to extend equipment lifespan.

[0008] This invention provides a digital management method for building energy efficiency based on big data, comprising: S1. Acquire historical and real-time operational data of the building and preprocess them; S2. Train a time-series fusion Transformer model based on historical operational data, and then predict the environmental quantities of each zone within the prediction time domain based on real-time operational data to obtain the predicted quantiles of the environmental quantities of each zone. Set the comfort default risk output at the model output end so that the model outputs the comfort default probability and comfort default duration probability of each zone; S3. Acquire actual environmental quantities during the calibration period, and calculate the prediction deviation distribution based on the actual environmental quantities, predicted quantiles, and comfort default probabilities. Perform online calibration of the predicted quantiles, comfort default probabilities, and comfort default duration probabilities based on the prediction deviation distribution to obtain the calibrated predicted quantiles, calibrated comfort default probabilities, and calibrated comfort default duration probabilities; S4. Acquire zone risk level parameters and... The following steps are taken: S5, identifying short-cycle start-stop events based on equipment operating status data in real-time operation data, and calculating the equivalent damage index based on short-cycle start-stop events, equipment control quantity data, and equipment rated parameters. The equivalent damage index is mapped to a differentiable piecewise convex penalty function to obtain the equipment life loss model. S6, constructing a multi-objective model predictive control problem based on building energy consumption target, comfort target, and equipment life target, and solving it to obtain the current control quantity. S7, issuing the current control quantity to building equipment for execution, collecting real-time operation data after control execution for preprocessing to update the real-time operation data.

[0009] Optionally, S1 includes: Acquire the historical operating data and the real-time operating data, and associate timestamps with the partition energy consumption data, partition environmental data, equipment operating status data, equipment control quantity data, occupancy data and outdoor meteorological data in the historical operating data and the real-time operating data respectively; The timestamps of various types of data are mapped to a unified timeline and resampled to a preset sampling interval to complete time alignment; Missing value processing is performed on various types of time-aligned data. The missing value processing includes interpolation to complete the missing value when the missing duration is not greater than a preset missing duration threshold, and using statistical values ​​based on historical running data to complete the missing value or retaining the missing value marker when the missing duration is greater than the preset missing duration threshold. Outlier processing is performed on various types of time-aligned data. The outlier processing includes removing or correcting out-of-range data based on a preset physical range threshold and a preset rate of change threshold. The processed data is then standardized in terms of unit and dimension, and the preprocessed historical running data and the preprocessed real-time running data are output.

[0010] Optionally, S2 includes: Training samples are constructed based on the preprocessed historical operating data. Each training sample includes an input sequence and a labeled sequence. The input sequence includes partition energy consumption data, partition environmental data, equipment operating status data, equipment control quantity data, occupancy data, and outdoor meteorological data within a preset historical window. The labeled sequence includes partition environmental data within the prediction time domain corresponding to the input sequence. A time series fusion Transformer model is trained using the training samples. The time series fusion Transformer model outputs predicted quantiles corresponding to a preset quantile set for the environmental quantities of each partition within the prediction time domain. The predicted quantiles are trained using a quantile loss function. A comfort default risk output layer is set at the output of the time series fusion Transformer model. The comfort default risk output layer outputs the comfort default probability and the comfort default duration probability based on the implicit representation of the time series fusion Transformer model. The comfort default probability is the probability that the comfort index of each partition exceeds a preset comfort threshold at any given time in the predicted time domain. The comfort default duration probability is the probability that the comfort index of the partition exceeds the preset comfort threshold and the duration is not less than a preset duration threshold. The preprocessed real-time operating data, including partition energy consumption data, partition environmental data, equipment operating status data, equipment control quantity data, occupancy data, and outdoor meteorological data within a preset real-time window, is input into the trained time series fusion Transformer model, which outputs the predicted quantile, the comfort default probability, and the comfort default duration probability.

[0011] Optionally, S3 includes: Within a preset calibration period, actual environmental quantities for each partition are extracted from the preprocessed real-time operating data, and the predicted quantiles and comfort default probabilities at the corresponding times are extracted simultaneously. Based on the actual environmental quantities and the predicted quantiles, the quantile prediction errors for each environmental quantity in each partition are calculated, and the prediction deviation distribution is statistically obtained based on the quantile prediction errors. The predicted quantiles are calibrated for quantile coverage according to the prediction deviation distribution, and the calibrated predicted quantiles are output. The comfort default probability and the comfort default duration probability are calibrated for probability consistency according to the prediction deviation distribution, and the calibrated comfort default probability and the calibrated comfort default duration probability are output. The probability consistency calibration includes updating the calibration mapping relationship based on a sliding time window to ensure that the statistical coverage of the calibrated comfort default probability meets a preset coverage condition.

[0012] Optionally, S4 includes: Obtain the partition risk level parameters and partition occupancy parameters of each partition at each time point within the prediction time domain. Determine the risk weight coefficient of each partition based on the partition risk level parameters and the occupancy weight coefficient of each partition at each time point based on the partition occupancy parameters. Calculate the allocation weight of each partition at each time point based on the risk weight coefficient and the occupancy weight coefficient, and allocate the total comfort risk budget to each partition at each time point within the prediction time domain according to the allocation weight, obtaining the upper limit of the comfort default probability for each partition at each time point. The upper limit of the comfort default probability satisfies that the sum of the upper limits of the comfort default probabilities for each partition at each time point is not greater than the total comfort risk budget. Based on the upper limit of the comfort default probability, the calibrated comfort default probability, and the calibrated comfort default duration probability, generate opportunity constraint parameters. The opportunity constraint parameters include the upper limit of the comfort default probability for each partition at each time point and the comfort constraint conditions corresponding to the upper limit of the comfort default probability, such that when solving the multi-objective model predictive control problem subsequently, the calibrated comfort default probability is not greater than the upper limit of the comfort default probability, and the calibrated comfort default duration probability is not greater than a preset duration probability threshold.

[0013] Optionally, S5 includes: Extract the equipment operation status data and equipment control quantity data of each building equipment from the preprocessed real-time operation data, and obtain the rated parameters of the equipment corresponding to the building equipment; identify the start and stop status change time based on the equipment operation status data, and calculate the start and stop interval between two adjacent start and stop status change times; When the start-stop interval is less than the preset short-cycle threshold, and the start-stop state change includes shutting down and restarting or starting and shutting down again, the corresponding start-stop state change is recorded as a short-cycle start-stop event; for each short-cycle start-stop event, the control quantity change amplitude and control quantity change rate are calculated based on the equipment control quantity data, and the control quantity change amplitude and control quantity change rate are normalized based on the equipment rated parameters. The equivalent damage index is calculated based on the number of starts and stops, the start and stop interval, the normalized change amplitude of the control quantity, and the normalized change rate of the control quantity for the short-cycle start and stop events. The equivalent damage index is mapped to a differentiable piecewise convex penalty function, wherein the piecewise convex penalty function includes a first penalty segment corresponding to the equivalent damage index being lower than a first damage threshold, a second penalty segment corresponding to the equivalent damage index being between the first damage threshold and a second damage threshold, and a third penalty segment corresponding to the equivalent damage index being higher than the second damage threshold, and the penalty intensity of each penalty segment increases as the equivalent damage index increases. An equipment life loss model is generated based on the piecewise convex penalty function, and the equipment life loss model is used to characterize the equipment life loss under a given change in equipment control quantity.

[0014] Optionally, S6 includes: The system state at the current moment is determined based on the equipment operating status data and equipment control quantity data in the preprocessed real-time operating data, and the control quantity of each building equipment in the predicted time domain is used as the decision variable of the control sequence; a building energy consumption objective function is constructed based on the partition energy consumption data in the preprocessed real-time operating data, a comfort objective function and comfort constraint conditions are constructed based on the chance constraint parameters, and an equipment life objective function is constructed based on the equipment life loss model. The building energy consumption objective function, the comfort objective function, and the equipment lifespan objective function are combined into a multi-objective model predictive control problem; equipment operation constraints are set in the multi-objective model predictive control problem, including equipment output range constraints, minimum start-up time constraints, minimum shutdown time constraints, and control variable change rate constraints; The comfort constraint is set as an opportunity constraint, such that the calibrated comfort default probability at each time in each partition within the prediction time domain is not greater than the corresponding upper limit of the comfort default probability. The multi-objective model predictive control problem is solved under the conditions of satisfying the equipment operation constraint and the opportunity constraint to obtain the control sequence, and the control quantity at the current time in the control sequence is determined as the current control quantity.

[0015] Optionally, the S7 includes: The current control quantity is converted into a control command consistent with the building equipment communication protocol, and the control command is sent to the building equipment corresponding to the current control quantity to execute the control. Real-time operating data after control execution is collected within a preset acquisition period. The real-time operating data after control execution includes zone energy consumption data, zone environmental data, equipment operating status data, and equipment control quantity data. Time alignment and missing value processing are performed on the real-time operating data after control execution to obtain updated preprocessed real-time operating data. The updated preprocessed real-time operating data replaces the preprocessed real-time operating data in step S1, and in the rolling time domain, the updated preprocessed real-time operating data is used as the input for subsequent steps S2 to S6 to update the predicted quantiles, comfort default probability, and comfort default duration probability; update the calibrated predicted quantiles, calibrated comfort default probability, and calibrated comfort default duration probability; update the opportunity constraint parameters; update the equipment life loss model; and update the current control quantity, thereby cyclically executing the digital management of building energy efficiency.

[0016] Optionally, the risk budget allocation rule further includes: setting a lower limit for the upper limit of the comfort default probability for at least one high-priority partition, and redistributing the remaining total comfort risk budget according to the allocation weight, provided that the lower limit is met.

[0017] Optionally, the preset short-cycle threshold is a threshold related to the device type and is adaptively updated based on the historical start-stop interval statistics of the device operating status data.

[0018] On the other hand, the present invention also provides a digital building energy efficiency management system based on big data, comprising: The data processing module acquires and preprocesses historical and real-time building operation data; the probability prediction module trains a time-series fusion Transformer model based on historical operation data, performs probability predictions of environmental quantities in each zone within the prediction time domain based on real-time operation data, and outputs prediction quantiles, comfort default probability, and comfort default duration probability; the online calibration module performs online calibration of the prediction quantiles, comfort default probability, and comfort default duration probability based on the actual environmental quantities during the calibration period; and the risk budget module is used to calculate the total comfort risk budget and combine it with zone risk level parameters and zone occupancy parameters. The algorithm assigns upper limits to the comfort default probability for each partition at each time point and generates opportunity constraint parameters. A lifespan loss module identifies short-cycle start-stop events and calculates the equivalent damage index, mapping the equivalent damage index to a differentiable piecewise convex penalty function to obtain the equipment lifespan loss model. A multi-objective control module constructs and solves an opportunity-constrained multi-objective model predictive control problem based on energy consumption, comfort, and equipment lifespan objectives to obtain the current control quantity. An update module sends the current control quantity to building equipment for execution and collects feedback data, updating the probability prediction, online calibration, opportunity constraint, and model predictive control processes in the rolling time domain.

[0019] The beneficial effects of this invention are: 1. By setting the comfort default risk output at the output end of the time series fusion Transformer model, and calibrating the predicted quantile, comfort default probability, and comfort default duration probability online, the comfort exceedance risk is transformed from a post-evaluation indicator into a calculable and constrainable risk quantity that enters the opportunity constraint, thereby improving the probability prediction coverage and the fulfillment of opportunity constraints, and reducing the probability of comfort exceedance and the risk of exceedance duration.

[0020] 2. By introducing a total comfort risk budget and combining it with zoning risk level parameters and zoning occupancy parameters for dynamic allocation, the upper limit of comfort default probability for each zoning at each time is obtained and opportunity constraint parameters are generated. This enables differentiated comfort protection and overall optimization across multiple zoning and time periods, reducing unnecessary conservative controls while meeting overall comfort goals, thereby reducing building energy consumption and improving the rationality of comfort management between regions.

[0021] 3. By identifying short-cycle start-stop events and calculating the equivalent damage index by combining the amplitude and rate of change of control variables, the index is mapped to a differentiable piecewise convex penalty function and incorporated into the objective function or constraint of multi-objective model predictive control. This makes equipment life loss quantifiable and optimizable, thereby reducing the equivalent damage caused by short-cycle start-stop events and drastic changes in control variables, slowing down equipment degradation, and extending equipment life. Attached Figure Description

[0022] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a digital management method for building energy efficiency based on big data, as proposed in this invention. Detailed Implementation

[0023] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0024] refer to Figure 1 A digital management method for building energy efficiency based on big data includes: S1. Acquire historical and real-time operational data of the building and preprocess them; S2. Train a time-series fusion Transformer model based on historical operational data, and then predict the environmental quantities of each zone within the prediction time domain based on real-time operational data to obtain the predicted quantiles of the environmental quantities of each zone. Set the comfort default risk output at the model output end so that the model outputs the comfort default probability and comfort default duration probability of each zone; S3. Acquire actual environmental quantities during the calibration period, and calculate the prediction deviation distribution based on the actual environmental quantities, predicted quantiles, and comfort default probabilities. Perform online calibration of the predicted quantiles, comfort default probabilities, and comfort default duration probabilities based on the prediction deviation distribution to obtain the calibrated predicted quantiles, calibrated comfort default probabilities, and calibrated comfort default duration probabilities; S4. Acquire zone risk level parameters and... The following steps are taken: S5, identifying short-cycle start-stop events based on equipment operating status data in real-time operation data, and calculating the equivalent damage index based on short-cycle start-stop events, equipment control quantity data, and equipment rated parameters. The equivalent damage index is mapped to a differentiable piecewise convex penalty function to obtain the equipment life loss model. S6, constructing a multi-objective model predictive control problem based on building energy consumption target, comfort target, and equipment life target, and solving it to obtain the current control quantity. S7, issuing the current control quantity to building equipment for execution, collecting real-time operation data after control execution for preprocessing to update the real-time operation data.

[0025] In this specific embodiment, S1 includes: The data processing module executes at the beginning of each rolling control cycle. It reads historical operating data from the historical database of the building energy management system and real-time operating data from the real-time acquisition gateway. It associates timestamps with the zonal energy consumption data, zonal environmental data, equipment operating status data, equipment control quantity data, occupancy data, and outdoor meteorological data in the historical and real-time operating data. The timestamps are uniformly converted to Beijing time and absolute time in seconds and written to the time field of the corresponding data record. In the case where there are multiple records at the same time in the original data, the latest record at the acquisition time is retained to eliminate duplicates. Then, the timestamps of various data types are mapped to a unified timeline and resampled to a preset sampling interval to complete time alignment. The unified timeline is defined by the formula: ; in Represents the first on the unified timeline The timestamp of each sampling moment This indicates the start timestamp of the data window corresponding to the merged historical and real-time operational data. This indicates the sampling point number, which starts from 0 and increments. This indicates a preset sampling interval of 300 seconds, and each original record is assigned to a sampling interval containing that timestamp based on its timestamp. To form a sequence indexed along a unified timeline; During resampling, the resampled value for each zone is generated by taking the arithmetic mean of all original values ​​within the sampling interval for the zone environmental data, outdoor meteorological data, and equipment control quantity data. For the zone energy consumption data, the interval energy consumption is generated by taking the difference of the cumulative meter readings, and the difference between the last cumulative reading in the interval and the last cumulative reading in the previous interval is taken as the zone energy consumption data for that interval. For the equipment operating status data, the interval end value retention method is used to generate resampled values, and the operating status most recently reported before the end of the sampling interval is taken as the equipment operating status data for that interval. For the occupancy data, the interval end value retention method is used to generate resampled values, and the number of occupants most recently reported before the end of the sampling interval is taken as the occupancy data for that interval. If there are no original records in the interval that can be used to generate resampled values, the corresponding field of the interval is marked as missing. Subsequently, missing value processing is performed on various types of time-aligned data. Missing value processing is calculated by converting the number of consecutive missing points on a unified time axis into missing duration, and the missing duration threshold is set to 1800 s. When the consecutive missing duration is not greater than the missing duration threshold, the consecutive missing segments are completed by linear interpolation, and the interpolation endpoints are the two nearest non-missing resampled values ​​before and after the missing segment. When the consecutive missing duration is greater than the missing duration threshold, the consecutive missing segments are completed by statistical values ​​based on historical running data, and the statistical values ​​are the median of the same partition, the same data field, the same day type, and the same sampling point number in the historical running data of the most recent 30 days. At the same time, a missing flag is written in the corresponding field to indicate that the data is obtained by statistical completion. If there are no historical samples that meet the above constraints in the most recent 30 days, the missing flag is retained and no completion value is written to avoid introducing unfounded data. After missing value handling is completed, outlier handling is performed on all types of time-aligned data. Outlier handling simultaneously uses preset physical range thresholds and preset rate of change thresholds to identify outliers and resets the missing data points to their missing value flags, thus incorporating them into the same missing value handling logic. The physical range threshold and rate of change threshold are fixedly configured according to the data field, with the partition temperature belonging to... The temperature is within 2 degrees Celsius and the rate of temperature change between adjacent sampling points does not exceed 2 degrees Celsius per 300 seconds. The relative humidity of the zone belongs to... And the humidity change rate between adjacent sampling points does not exceed Every 300 seconds, the outdoor temperature is... Temperature changes at adjacent sampling points do not exceed 3 degrees Celsius per 300 seconds. Zoned energy consumption data is not less than 0. An anomaly is defined as an interval energy consumption exceeding three times the median energy consumption of the same sampling point interval over the past seven days within the same zone. Equipment control data is then normalized to [a specific value]. Furthermore, the rate of change of the control quantity of adjacent sampling points shall not exceed 0.1 every 300 seconds. The equipment operation status data is only allowed to take 0 or 1, and when a non-zero or non-1 value appears, it shall be judged as abnormal. The occupancy data shall not be less than 0 and not greater than the number of people designed to occupy the partition, and when a non-integer value appears, it shall be rounded to an integer. Finally, the units and dimensions of the processed data were standardized. Specifically, zoned energy consumption data was standardized to kWh, temperature in zoned environmental data and outdoor meteorological data was standardized to degrees Celsius, and relative humidity was standardized to percentage. Equipment control data was uniformly mapped to... The dimensionless control quantity unifies the equipment operation status data into a binary state, with 1 representing operation and 0 representing stop. The occupancy data is unified into the number of people and stored as an integer. The preprocessed historical operation data and the preprocessed real-time operation data obtained from the above processing are encapsulated into a data structure with a missing marker field according to a unified time axis and written into the input cache of the subsequent probability prediction module.

[0026] In this specific embodiment, S2 includes: This is executed by the probability prediction module, which constructs training samples for training the time series fusion Transformer model based on preprocessed historical running data. The training samples are generated in a sliding window manner along a unified time axis, with the sampling time corresponding to the start of each window as the sample index time. The input sequence takes the continuous data before the index time. Data from each sampling point and Take 288 to cover the 24-hour history window and match step S1. The sampling intervals are consistent, and the labeled sequence is continuous after the index time. Data from each sampling point and 24 are selected to cover the 2-hour prediction time domain. The input sequence of each training sample contains partition energy consumption data, partition environmental data, equipment operation status data, equipment control quantity data, occupancy data and outdoor meteorological data within a preset historical window. The data are concatenated in the field order after time alignment in the feature dimension to form a multivariate time series. At the same time, the missing markers are converted into missing indicator features of the same dimension and input together with the original features so that the model can explicitly perceive the historical missing situation. The labeled sequence contains the partitioned environmental data in the prediction time domain corresponding to the input sequence, and in this embodiment, the partitioned temperature is taken as the prediction object of the partitioned environmental quantity, so as to form a univariate supervision signal for each partition at each time in the prediction time domain. The probability prediction module uses the training samples to train a time-series fusion Transformer model. This model consists of a feature embedding layer, a variable selection network, an encoder-decoder sequence modeling layer, an attention fusion layer, and an output layer. Continuous features are embedded into the model via linear mapping. Hidden space and The variable selection network calculates the gating weights for the multivariate inputs at each time step and uses a gated residual network to perform nonlinear feature filtering to integrate the contributions of partition energy consumption data, partition environmental data, equipment operating status data, equipment control quantity data, occupancy data, and outdoor meteorological data to the future partition temperature. The encoder-decoder sequence modeling layer uses a two-layer gated cyclic unit with a length of... Encode the historical sequence and the length is The future sequence is decoded with a hidden state dimension of 64. The attention fusion layer uses a multi-head self-attention mechanism to enhance the characterization of long-range dependencies and has 4 attention heads. The output layer targets a preset set of quantiles. Output predicted quantiles and in this embodiment And for each partition, the predicted quantile of the partition temperature at each time point in the prediction time domain is output respectively; A comfort default risk output layer is set at the output of the time series fusion Transformer model. The comfort default risk output layer reuses the decoded implicit representation and generates the comfort default probability and the comfort default duration probability through two parallel fully connected probability output branches. The comfort default probability is defined as "the probability that the comfort index of each zone exceeds the preset comfort threshold at each time point in the predicted time domain". In this embodiment, the zone comfort index is set as the zone temperature and the preset comfort threshold is set as the upper and lower limits of the temperature range. and and The first probability output branch outputs a comfort default probability between 0 and 1 for each moment in the prediction time domain. The comfort default duration probability is defined as "the probability that the comfort index of a region exceeds the preset comfort threshold and the duration is not less than the preset duration threshold in the prediction time domain", and the preset duration threshold is set to 1. Each sampling point corresponds to a 30-minute continuous over-limit, and the second probability output branch outputs a comfort default duration probability between 0 and 1 for each partition. The second probability output branch uses one-dimensional convolution and global max pooling in the prediction temporal dimension to perform temporal aggregation on the decoded latent representation to ensure that its length is not less than [a certain value]. The continuous out-of-limit mode is sensitive and the kernel length of the one-dimensional convolution is taken as... It uses 32 convolutional channels and the activation function is a rectified linear unit; During the model training phase, the predicted quantiles are trained using the quantile loss function, and the comfortable default probability and comfortable default duration probability are trained using the binary cross-entropy loss function. A weighted sum is then used to form the total loss for joint training. The total loss is calculated as follows: ; in Indicates the total loss. Indicates the number of building zones. Represents the time index within the prediction time domain, and takes values ​​from 1 to... Indicates a quantile and belongs to a predefined set of quantiles. Indicates the first Partition in the prediction time domain The actual zone temperature label value at any given time. The first output of the model represents the... Partition in the prediction time domain Time corresponding quantile Predicted quantiles of regional temperatures, Indicates quantile loss and by quantile point Positive and negative errors are weighted by a piecewise linear method. Represents the binary cross-entropy loss. Indicates the first Partition in the prediction time domain The comfort level at any given moment is marked as a breach of contract and the actual zone temperature is as follows. Falling Take 1 if it is not otherwise taken The first output of the model represents the... Partition in the prediction time domain The probability of default at any given moment. Indicates the first Does the partition have at least one continuous interval throughout the entire prediction time domain? The duration of comfort default at each sampling point is labeled, and a value of 1 is assigned if the default exists, otherwise a value is assigned. The first output of the model represents the... The probability of comfort breach duration in each zone. and In this embodiment, the loss weight is represented. and ; The model was trained using the Adam optimizer with a learning rate set to [value missing]. The batch size is 128 and the number of training rounds is 50. Before training, each continuous feature in the input sequence is standardized according to the statistical mean and standard deviation of the training set. During inference, the same statistical measure is reused to standardize the real-time running data in the same way to ensure consistent distribution. During the model inference stage, the preprocessed real-time running data within a preset real-time window will be used for continuous... The energy consumption data, environmental data, equipment operation status data, equipment control quantity data, occupancy data, and outdoor meteorological data of each sampling point are used to construct an input sequence in the same field order and standardization method as in the training phase. This sequence is then input into the trained time series fusion Transformer model. The output is the predicted quantile of the temperature of each zone in the prediction time domain, the comfort default probability of each zone at each time, and the comfort default duration probability of each zone.

[0027] In this specific embodiment, S3 includes: The online calibration module performs this operation within each rolling control cycle, setting the preset calibration period to the most recent one. The time interval corresponding to each sampling point is to cover the most recent 24 hours and the sampling interval is preset. The actual environmental quantities of each partition within the calibration period are extracted from the preprocessed real-time running data, and in this embodiment, the actual environmental quantities are determined as the actual partition temperature. At the same time, the predicted quantiles and comfort default probabilities output in step S2 under the same partition and the same timestamp are extracted and point-by-point alignment is completed. The actual partition temperature is given by the resampled value of the partition temperature sensor at the sampling time of the unified time axis. The predicted quantiles and comfort default probabilities are read from the prediction output cache of the Transformer model of time series fusion at the same batch of sampling times in the previous rolling cycle to ensure time consistency. The online calibration module calculates the quantile prediction error based on the actual zone temperature and predicted quantiles within the calibration period, and statistically analyzes the prediction deviation distribution accordingly. The prediction deviation distribution is represented by an empirical distribution, and the cumulative distribution is obtained by sorting all error samples to avoid introducing additional distribution assumptions. Based on the prediction deviation distribution, the module performs quantile coverage calibration, which is achieved through "additive correction at each quantile point." The calibrated predicted quantiles are output using the following formula: and ; in This represents the calibrated predicted quantiles. Indicates a partitioned index and and Indicates the number of partitions. This indicates the sampling time index within the calibration period and corresponds to a unified timeline sampling point. Indicates quantile and and This refers to the preset set of quantiles used in step S2. This indicates the predicted quantiles for temperature in different zones. This indicates the actual partition temperature extracted from real-time operational data. Indicates quantile The corresponding additive calibration offset, Quantile This indicates that the first error sample is taken from the set of all error samples within the calibration period. The statistical calculation of quantiles is performed by sorting the samples in ascending order of value and assigning them a sequence number "". The sample values ​​are realized and Indicates the total number of error samples; The online calibration module performs probability consistency calibration on the probability of comfort default and the probability of comfort default duration based on the same prediction bias distribution. The calibration of the comfort default probability uses default observation markers within the calibration period as supervision signals, and the default observation markers are defined as those when the actual zone temperature... Exceeding the preset comfort threshold range in step S2 The value is set to 1 if the condition is met and 0 otherwise. Paired samples of predicted comfort default probability and default observation labels are collected during the calibration period, and then a monotonically non-decreasing calibration mapping relationship is constructed through order-preserving regression. The uncalibrated comfort default probability is mapped to the calibrated comfort default probability. The order-preserving regression uses the pool adjacent violation algorithm to generate a piecewise constant mapping and outputs a set of breakpoints and corresponding mapping values ​​arranged in ascending order of probability values ​​so that the mapping can be completed by looking up a table during online inference. The calibration of the probability of comfort default duration uses duration event observations within the calibration period as supervision signals, and defines duration event observations as events with a length of... There are at least 10 consecutive sampling points in the prediction time domain. For each sampling point, a value of 1 is assigned if a comfort default occurs, and 0 is assigned otherwise. Paired samples of the predicted comfort default duration probability and duration event observations are collected during the calibration period and then compared with... The same ordination-preserving regression method is used to construct the calibration mapping relationship. Output the probability of the duration of comfort default after calibration; Probabilistic consistency calibration uses a sliding time window to update the calibration mapping relationship. The update method is to add the paired samples corresponding to the latest sampling point and remove the paired samples corresponding to the earliest sampling point in each rolling control cycle to keep the window length constant. After each update, the calibrated comfort default probability is divided into 20 probability bins with a width of 0.05. The absolute deviation between the empirical default frequency and the bin center value within each bin is calculated, and the statistical coverage condition is set to ensure that the absolute deviation of all bins is no greater than 0.03. If the statistical coverage condition is not met, an action is triggered. and The refit yields the calibrated comfort default probability and the calibrated comfort default duration probability that satisfy the statistical coverage condition, and these are compared with the calibrated prediction quantile.

[0028] In this specific embodiment, S4 includes: This is executed by the risk budgeting module. The risk budgeting module first reads the partition risk level parameters for each partition from the building configuration library and records them as follows: ,in For partitioned index and ,and This indicates the number of building zones and the risk level parameter for each zone. The set of values ​​is It is agreed that the larger the value, the higher the requirement for comfort assurance in that zone; The risk budgeting module then extracts the partition occupancy parameters for each partition at each time point within the prediction time domain from the preprocessed real-time runtime data and records them as follows: ,in To predict the index of time in the time domain and ,and To predict the time domain length and take 24, the partitioning parameter is used. The number of people must be a non-negative integer and aligned with a unified timeline; The risk budgeting module determines the risk weight coefficient for each partition based on the partition risk level parameter and records it as follows: Among them, risk weight coefficient Take as To ensure that higher-risk partitions correspond to lower risk weight coefficients, thereby achieving a stricter upper limit on default probability during budget allocation, the risk budget module determines the occupancy weight coefficient for each partition at each time point based on the partition occupancy parameter and records it as follows: The weighting coefficient is used for Take as This is to ensure that the more people occupying a space, the smaller the occupancy weight coefficient will be, thus obtaining a stricter upper limit on the probability of default when allocating the budget; The risk budgeting module multiplies the risk weight coefficient by the occupancy weight coefficient to obtain the allocation weight for each partition at each time point, and records it as follows: ,in The total comfort risk budget is defined as the upper limit of the sum of the upper limits of comfort default probabilities for all partitions and all times in the prediction time domain. This total comfort risk budget is set as the total budget corresponding to "the average default probability allowed for each partition at each time is 0.02", thus making the total comfort risk budget equal to... ; After obtaining the assigned weights, the risk budget module allocates the total comfort risk budget to each partition and time point within the prediction time domain according to the assigned weights to obtain the upper limit of the comfort default probability, and then follows the formula: ; in Indicates the prediction of the first time domain Partition number The upper limit of the probability of default at any given moment. Indicates the number of building zones. Indicates the length of the prediction time domain. Indicates the first Partition number Weighting of time intervals This represents the summation of the allocation weights for all partitions and all times within the prediction time domain to achieve normalized allocation, thereby ensuring... This ensures that the sum of the upper limits of the comfort default probability for each partition at each time point does not exceed the total comfort risk budget. After completing the allocation of the comfort default probability cap, the risk budget module will calibrate the comfort default probability. Probability of duration of comfort default after calibration Read in and combine Perform index alignment, where This indicates the prediction of the th time term within the prediction time domain under the current rolling period. Partition number The calibration probability estimate of when the comfort index exceeds the preset comfort threshold. This indicates that under the current rolling period, the first... The partition occurs within the prediction time domain and its duration is not less than a preset duration threshold. Calibrated probability estimates of comfort default events at each sampling point; The risk budgeting module generates opportunity constraint parameters based on this and outputs them in the form of a structured parameter table. These opportunity constraint parameters include a comfortable default probability upper limit for each partition and each time step. calibrated comfort default probability Preset comfort threshold range And the corresponding comfort constraints, which are defined as mandatory conditions to be met when solving the multi-objective model predictive control problem. At the same time, each partition includes a preset duration probability threshold and is recorded as follows. And take The duration chance constraint is defined as something that must be satisfied when solving the multi-objective model predictive control problem. This allows the instantaneous risk of comfort default to be controlled by the budgeted upper limit of each partition at each time, and the persistent risk of comfort default to be controlled by a uniform duration probability threshold.

[0029] In this specific embodiment, S5 includes: The life loss module performs the operation. It extracts the equipment operation status data and equipment control quantity data of each building equipment from the pre-processed real-time operation data, and reads the equipment rated parameters corresponding to the building equipment from the equipment ledger for normalization calculation. The equipment operation status data is expressed as a binary sequence on a unified time axis, with 0 indicating shutdown and 1 indicating operation. The equipment control quantity data is expressed as a continuous sequence on a unified time axis, and the values ​​fall within the equipment's rated control range. The life loss module is in length of Within the monitoring window of each sampling point, the equipment operation status data of each building equipment is scanned point by point to identify the time of start-stop status change. Specifically, when the status value changes between adjacent sampling points, the timestamp of the next sampling point is recorded as the time of start-stop status change and the direction of the change is recorded to obtain a sequence of start-stop status change times arranged in chronological order, and the start-stop interval between two adjacent start-stop status change times is calculated. The lifespan loss module sets the preset short-cycle threshold to a fixed threshold related to the equipment type and determines the threshold value using an equipment type table, where the chiller unit compressor corresponds to... Water pump corresponding corresponding to the wind turbine And during the determination, the start-stop interval will be compared with... Compare; When the start-stop interval between two adjacent start-stop state changes is less than When the two start-stop state changes constitute a round-trip switching relationship of "shutting down and then restarting" or "starting and then shutting down again," the two start-stop state changes are combined into a short-cycle start-stop event and assigned an event number, so that each short-cycle start-stop event contains two start-stop state changes and the number of start-stops is fixed. ; For each short-cycle start / stop event, the lifespan loss module extracts the control quantity sequence from the equipment control quantity data within the time interval covered by the event and calculates the control quantity change amplitude and control quantity change rate. The control quantity change amplitude is the difference between the maximum and minimum control quantity values ​​within the time interval, and the control quantity change rate is the absolute value of the difference between the control quantities of adjacent sampling points within the time interval divided by the sampling interval. The maximum value after normalization is obtained, and the amplitude and rate of change of the control quantity are normalized using the equipment rated parameters to eliminate the influence of dimensions. The equipment rated parameters include at least the equipment rated lower control limit. Equipment rated control upper limit and the rate of change of equipment rated control And based on this, the normalized control quantity change range is obtained. Rate of change of normalized control quantity ; The lifespan loss module calculates the equivalent damage index based on the number of start / stop events, start / stop intervals, normalized control quantity change amplitude, and normalized control quantity change rate of short-cycle start / stop events. This equivalent damage index is then accumulated within the monitoring window according to events to obtain the window equivalent damage index corresponding to the building equipment. Simultaneously, the window equivalent damage index is mapped to a differentiable piecewise convex penalty function to form the equipment lifespan loss model. The window equivalent damage index and its penalty function are determined by the following formula: ; ; in Indicates building equipment The window equivalent damage index within the monitoring window Indicates building equipment The number of short-cycle start-stop events identified within the monitoring window. This represents an index of short-cycle start / stop events. Indicates building equipment The The number of start-stop events in a short-cycle start-stop event is always equal to the number of start-stop events in this embodiment. Indicates building equipment The The start-stop interval between two start-stop state changes in a short-cycle start-stop event. Indicates building equipment Preset short-cycle thresholds for corresponding device types Indicates building equipment The The normalized change range of the control quantity after a short-cycle start-stop event is calculated by dividing the change range of the control quantity by ( )get, Indicates building equipment The The normalized rate of change of control quantity for short-cycle start-stop events, divided by the rate of change of control quantity. get, Indicates building equipment The equipment's rated control lower limit, Indicates building equipment The rated control limit of the equipment, Indicates building equipment The rated control change rate of the equipment, Indicates the weight of the number of start and stop operations and takes Indicate the start / stop interval weight and take Indicates the weight of the change in the control quantity and takes This represents the weight of the rate of change of the control variable, and is set to 0.25. This represents a piecewise convex penalty function that maps the window-equivalent damage index to lifetime loss. Represents the first damage threshold and takes This represents the second damage threshold, which is set to 3.0. This indicates that the coefficient of the first penalty segment is 1. This indicates that the coefficient for the second penalty segment is 5. This represents the coefficient of the third penalty segment, which is set to 20, and is calculated on the second penalty segment. and Make the third penalty segment in The function value and first derivative at the second penalty segment are continuous, thus satisfying differentiability and preserving convexity, and the penalty intensity varies with... Increase and increase; The life loss module will It is used as the output penalty term of the equipment life loss model and is combined with it in the calculation. of They are encapsulated together as a parameter set for the equipment life loss model, so that the equipment life loss can be quantified in a differentiable form and incorporated into the multi-objective model predictive control solution process under a given change in equipment control variable.

[0030] In this specific embodiment, S6 includes: At the start of the current rolling cycle, the multi-objective control module reads the current system state from the preprocessed real-time operating data and constructs initial optimization values. The system state includes the current operating state of each building device and the current control variables, which are denoted as follows: and ,in Represents device index and Indicates the quantity of building equipment. It is binary and Indicates running and Indicates shutdown. It is a normalized control quantity and its value range is And it is consistent with the unit unification and dimension consistency processing in step S1; The multi-objective control module uses the control quantities of each building device in the predicted time domain as the decision variables of the control sequence and denotes them as follows: Simultaneously, the operating status of each building's equipment within the predicted time domain is used as an auxiliary decision variable and denoted as... ,in Represents the index of the time in the prediction time domain and , To predict the time domain length, and in this embodiment, it is taken as 24; The multi-objective control module reads the opportunity constraint parameters from step S4 and obtains the upper limit of the comfort default probability for each partition at each time step. and preset duration probability threshold ,in Indicates a partitioned index and Indicates the number of partitions; The multi-objective control module calls the time series fusion Transformer model and selects candidate control sequences in each optimization iteration. Write the future control variable data field into the model input, and simultaneously write the occupancy data and outdoor meteorological data into the future known exogenous variable field of the model input according to the prediction time domain, and then add the most recent data. The partition energy consumption data, partition environmental data, equipment operating status data, and historical equipment control data of each sampling point are written into the historical window field of the model input to complete the input tensor construction. This yields the original comfort default probability and the original comfort default duration probability of each partition at each time point in the prediction time domain under the candidate control sequence. Furthermore, a calibration mapping relationship is applied to map the original probabilities to the calibrated comfort default probabilities. Probability of duration of comfort default after calibration For use in opportunity constraint determination; The multi-objective control module calls the equipment life loss model and, under the candidate control sequence, bases its decisions on... Start-stop changes and The amplitude and rate of change are used to calculate the equivalent damage index in the prediction time domain, and the equipment life loss penalty value is obtained through a piecewise convex penalty function. To characterize the lifetime loss caused by candidate control sequences; The multi-objective control module constructs a building energy consumption objective function based on the equipment's rated parameters and denotes the equipment's rated active power as... And with The sampling interval corresponds to the duration, and the energy consumption in the prediction time domain is estimated as a discrete integral of the power of each device at each moment to obtain the energy consumption target term; Based on this, the multi-objective control module combines the building energy consumption objective, comfort objective, and equipment lifespan objective into a multi-objective model predictive control problem and solves it using the following formula to obtain the current control variable: , st ; in This represents the comprehensive objective function of a multi-objective model predictive control problem. Indicates the weight of the energy consumption target and takes Indicates the weight of the comfort target and takes Indicates the target weight of equipment lifespan and takes Indicates equipment The rated active power of the equipment, Indicates equipment In the prediction time domain Normalized control quantity at time, This indicates the duration corresponding to the sampling interval. Indicates the first in the candidate control sequence Partition number The probability of comfort default after calibration at any given moment. This indicates the allocation obtained in step S4. Partition number The upper limit of the probability of default at any given moment. Indicates the first in the candidate control sequence The partition occurs within the prediction time domain and its duration is not less than a preset duration threshold. The calibrated duration probability of comfort default events at each sampling point Indicates the preset duration probability threshold and takes Indicates equipment The piecewise convex penalty value output by the equipment life loss model. Indicates equipment The equivalent damage index in the prediction time domain Indicates equipment In the prediction time domain The running status at any given moment; In the multi-objective model predictive control problem, the multi-objective control module further sets equipment operating constraints and incorporates them into the solution process. These equipment operating constraints include equipment output range constraints, and are achieved by... and The binding ensures that the output is zero when shutting down and that the output is not less than the device's minimum allowable output during operation. For frequency converters, the minimum output is set to 0.2, ensuring that when... hour And when hour The equipment operation constraints also include minimum power-on duration constraints and minimum power-off duration constraints, which are implemented by applying a hold rule to the start-stop switching. Specifically, when a start-stop switching occurs at a certain time... Forced subsequent continuation Maintaining operation at any given moment means that for any... satisfy When a shutdown / switching occurs at a certain time, Forced subsequent continuation Keep it off at all times, that is, for any satisfy and the chiller unit compressor and And the fan and water pump and In accordance with the minimum start-up and shutdown duration, the equipment operation constraints also include control quantity change rate constraints and limit the change of control quantity between adjacent time moments to no more than 0.1 in order to suppress drastic adjustments and thus maintain consistency with the outlier change rate threshold in step S1; The multi-objective control module employs a combined solution process of branch and bound and sequential quadratic programming for handling binary variables. With continuous variables This paper solves a multi-objective predictive control problem with opportunity constraints and sets the feasibility tolerance as... And set the optimality gap to Furthermore, the maximum solution time for a single rolling cycle is set to 20 seconds, and the first [number]th ... Control of time As the current control variable.

[0031] In this specific embodiment, S7 includes: This is executed by the update module and uses the same unified timeline and preset sampling interval as step S1. ; The update module receives the current control quantity and represents it as the target value of the control command for each building device in the current rolling cycle. ,in Represents device index and Indicates the quantity of building equipment. It is a normalized control quantity and its value range is ; The update module reads the communication protocol type, device address, register or object point, range parameters, and write permissions for each building device from the device communication configuration table, and converts the normalized control quantity into a physical control setpoint consistent with the building device's communication protocol and encapsulates it into a control command. The physical control setpoint is configured as follows: ; in Indicates that the data has been sent to the device. Physical control setpoints, Indicates equipment The normalized control quantity in the current rolling cycle, Indicates equipment The minimum physical setting value that can be written to the corresponding control point. Indicates equipment The maximum physical setting value that can be written to the corresponding control point; The update module is generated Then, boundary clamping and resolution quantization are performed, where boundary clamping will... Limited to To ensure that the limits are not exceeded, the resolution quantization is based on the control resolution recorded in the device communication configuration table. Perform rounding to match the minimum acceptable step size on the device side, and then quantize the result. Combined with device address and location information, a control command frame is generated; The update module groups control commands by device and sends them to the corresponding building devices through a unified communication gateway to execute the control. After sending, it enters the acknowledgment process. The acknowledgment process sets an acknowledgment timeout of 3 seconds and a resend count of 2 for each control command. When a successful acknowledgment is received from the device side within the acknowledgment timeout, the control command is marked as successfully executed and the execution timestamp is recorded. If a successful acknowledgment is not received after continuous resends, the control command is marked as failed and the effective setting value of the previous rolling cycle of the device remains unchanged. At the same time, the failure event is written to the alarm log for operation and maintenance to handle. After the control command is issued, the update module collects real-time operating data after the control is executed within a preset collection cycle. The real-time operating data includes partition energy consumption data, partition environment data, equipment operating status data, and equipment control quantity data. Each data record collected is accompanied by its collection timestamp and uniformly converted to a second-level Beijing time timestamp. The update module performs time alignment and missing value processing on the collected real-time running data to obtain updated preprocessed real-time running data. Time alignment involves mapping the collected timestamps to... A unified timeline is constructed and resampled according to the sampling interval. For missing value handling, the threshold for continuous missing duration is set to 1800 s. When the missing duration is not greater than the threshold, linear interpolation is used to complete the missing value. When the missing duration is greater than the threshold, the missing value is retained to avoid introducing unfounded data. The update module writes the updated preprocessed real-time running data into the real-time data cache and replaces the preprocessed real-time running data in step S1. Simultaneously, it maintains the real-time window cache to ensure it always contains the most recently updated data. The input sequence data of each sampling point is used to maintain a calibration window cache that always contains the most recent data. The calibration period data of each sampling point is used as the input for subsequent execution steps S2 to S6 in the rolling time domain, thereby updating the predicted quantile, comfort default probability, and comfort default duration probability in sequence. The calibrated predicted quantile, calibrated comfort default probability, and calibrated comfort default duration probability are updated, the opportunity constraint parameters are updated, the equipment life loss model is updated, and a new current control quantity is calculated, thereby cyclically executing the digital management of building energy efficiency.

[0032] In this specific embodiment, when executing the risk budget allocation rule in step S4, the risk budget module sets a lower limit value of the upper limit of the comfort default probability for at least one high priority partition and executes a two-stage allocation process. The risk budget module reads high-priority flag parameters from the building configuration library and records them as follows. ,in Indicates a partitioned index and and Indicates the number of partitions. and Indicates the first The partition is a high-priority partition and Indicates the first The partition is a non-high-priority partition, and the upper and lower limits of the comfort default probability corresponding to the high-priority partition are read from the same configuration library and recorded as follows. ,in It is a non-negative real number and in this embodiment satisfies partition take And to satisfy partition take This allows lower bound constraints to be applied only to high-priority partitions; The risk budget module records the total comfort risk budget as And in this embodiment, ,in This indicates the prediction time domain length, which is 24, and the assigned weights obtained in step S4. Based on this, a lower limit of the budget is first reserved for all high-priority partitions at each time point in the prediction time domain. Then, the remaining budget is redistributed according to the allocation weights to obtain the upper limit of the comfortable default probability for each partition at each time point. ,in Represents the index of the time in the prediction time domain and And complete the budget reservation and redistribution according to the following formula: ; in This represents the total comfort risk budget remaining after meeting the upper and lower limits of the comfort default probability for high-priority zones. This represents the total comfort risk budget. Indicates a high-priority marker parameter. This represents the upper and lower limits of the comfort default probability for high-priority partitions. Indicates the prediction of the first time domain Partition number The upper limit of the probability of default at any given moment. This indicates the result of step S4. Partition number Weighting of time intervals This represents the summation of the allocation weights for all partitions and all times within the prediction time domain to achieve normalized allocation; The risk budget module has its constraints fixed in the configuration library. To ensure that there is still an allocable budget after the lower limit is reserved, and to calculate the... Write the opportunity constraint parameter table so that step S6 can calculate the calibrated comfort default probability. and Establish opportunity constraints to redistribute the remaining total comfort risk budget while meeting the lower limit reservation requirements for high-priority zones.

[0033] In this specific embodiment, when performing step S5, the life loss module sets the preset short-cycle threshold to a threshold related to the device type and updates it adaptively. The life loss module reads the equipment type of each piece of building equipment from the equipment ledger and records it as follows: ,in Represents device index and Indicates the quantity of building equipment. The equipment type is identified and its value is either a chiller compressor, water pump, or fan. During system initialization, the initial short-cycle threshold corresponding to the equipment type is read from the equipment type threshold table and written into the equipment threshold cache, where it is recorded as... ,in Indicates equipment The short-cycle threshold is in seconds, and the threshold table is fixed according to equipment type as follows: 1800 s for chiller compressor, 600 s for water pump, and 600 s for fan; The life loss module extracts equipment data from preprocessed real-time operating data at the end of each rolling control cycle. recent The device operating status data corresponding to each sampling point covers the most recent 7 days and is consistent with the preset sampling interval. The system is consistent and identifies the start / stop status change moments within this window, calculating the start / stop interval between two adjacent start / stop status change moments to form a historical start / stop interval sample set. ,in Indicates equipment Within the statistics window A sample of start-stop intervals, in seconds. Indicates equipment The number of start-stop interval samples obtained within the statistical window; The lifespan depletion module sets the update cycle of the short-cycle threshold to 1 day and checks the conditions when the update cycle is reached. The device performs adaptive threshold updates and updates for devices that do not meet the requirements. The device remains current Keep it unchanged to avoid threshold fluctuations caused by insufficient samples; The threshold adaptive update adopts a quantile update rule based on historical start-stop interval statistics and introduces upper and lower bounds corresponding to the device type to ensure that the threshold is within a reasonable range. The updated short-cycle threshold is calculated as follows: Quantile ; in Indicates equipment Updated short-cycle threshold, Indicates device type The corresponding lower bound of the short-cycle threshold is in seconds and is fixed at 900 s for the chiller compressor, 300 s for the water pump, and 300 s for the fan. Indicates device type The corresponding short-cycle threshold upper bound, in seconds, is fixed at 3600 s for the chiller compressor, 1200 s for the water pump, and 1200 s for the fan. This represents the statistical operator that takes the 0.2 quantile of the start-stop interval sample set and calculates it by... Sort by numerical value in ascending order and take the sequence number. The sample values ​​are implemented; The life loss module is obtained Then write it back to the device threshold cache and replace the original one. The updated version is used in the short-cycle start-stop event recognition in subsequent step S5. The system determines whether the start-stop interval is less than the short-cycle threshold, thereby enabling the setting of a short-cycle threshold that is related to the equipment type and can be adaptively updated based on historical start-stop interval statistics.

[0034] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0035] This invention addresses the technical problem of the difficulty in quantifying the multi-objective trade-offs between energy consumption, comfort, and equipment lifespan, and the lack of a unified evaluation and decision-making mechanism. It establishes a closed-loop linkage between probabilistic prediction, opportunity-constrained multi-objective model predictive control, and lifespan loss modeling: First, a time-series fusion Transformer is used to probabilistically predict environmental variables in different zones, outputting predicted quantiles and directly providing the probability of comfort violation and the probability of comfort violation duration, transforming comfort from a post-hoc statistical indicator into a calculable risk quantity. Then, online calibration ensures that the predicted quantile coverage and violation probability consistently meet preset conditions during operation, thus making opportunity constraints feasible. Next, energy consumption targets, comfort opportunity constraints, and equipment operation constraints are simultaneously introduced into the model predictive control, and the penalty term corresponding to the lifespan loss model is incorporated into the optimization solution, outputting the current control quantity and updating it continuously. This allows for the reduction of comfort exceedance probability and persistent risk, lowering energy consumption, and suppressing losses caused by frequent equipment start-ups and shutdowns, while meeting the comfort risk ceiling and equipment constraints, achieving quantifiable synergistic optimization between energy efficiency, comfort, and lifespan.

[0036] This invention addresses key technical challenges in building scenarios by making improvements in three aspects: First, it adds a "comfort default risk output" and an online calibration module to the prediction end, enabling comfort risk to directly enter opportunity constraints and maintain constraint fulfillment under uncertainty drift conditions. Second, it proposes a risk budget allocation mechanism based on the total comfort risk budget, combining zonal risk level parameters and zonal occupancy parameters to dynamically allocate the upper limit of default probability by zonal and time period, avoiding excessive conservatism or uneven regional protection caused by fixed thresholds, thereby further reducing energy consumption and unnecessary adjustments under the same comfort objectives. Third, it upgrades the lifespan target from the number of start-stop cycles to an equivalent damage index, using short-cycle start-stop identification and weighting of control quantity changes in magnitude and rate of change, and mapping it to a differentiable piecewise convex penalty function. This makes lifespan loss optimizable and constrainable, more effectively reducing degradation caused by short-cycle start-stop cycles and drastic control changes, achieving quantitative protection of equipment lifespan and improving overall operational efficiency.

Claims

1. A digital management method for building energy efficiency based on big data, characterized in that, include: S1. Obtain historical and real-time operational data of the building and perform preprocessing; S2. Train a time series fusion Transformer model based on historical operational data, and then make probabilistic predictions of environmental quantities in each partition within the prediction time domain based on real-time operational data to obtain the predicted quantiles of environmental quantities in each partition. Set the comfort default risk output at the output end of the model so that the model outputs the comfort default probability and comfort default duration probability for each partition. S3. Obtain the actual environmental quantities during the calibration period, and calculate the prediction deviation distribution based on the actual environmental quantities, predicted quantiles, and comfort default probabilities. Based on the prediction deviation distribution, perform online calibration of the predicted quantiles, comfort default probabilities, and comfort default duration probabilities to obtain the calibrated predicted quantiles, calibrated comfort default probabilities, and calibrated comfort default duration probabilities. S4. Obtain the risk level parameters and occupancy parameters of each zone. Construct a risk budget allocation rule based on the total comfort risk budget and allocate it to obtain the upper limit of the comfort default probability at each moment in each zone within the prediction time domain. Generate opportunity constraint parameters based on the upper limit of the comfort default probability, the calibrated comfort default probability, and the probability of comfort default duration. S5. Identify short-cycle start-stop events based on the equipment operating status data in the real-time operation data. Calculate the equivalent damage index based on the short-cycle start-stop events, equipment control quantity data, and equipment rated parameters. Map the equivalent damage index to a differentiable piecewise convex penalty function to obtain the equipment life loss model. S6. Construct a multi-objective model predictive control problem using building energy consumption targets, comfort targets, and equipment life targets. Solve the problem to obtain the current control quantity. S7. Send the current control quantity to the building equipment for execution. Collect the real-time operation data after control execution and preprocess it to update the real-time operation data.

2. The digital management method for building energy efficiency based on big data according to claim 1, characterized in that, S2 include: Training samples are constructed based on the preprocessed historical operating data. Each training sample includes an input sequence and a labeled sequence. The input sequence includes partition energy consumption data, partition environmental data, equipment operating status data, equipment control quantity data, occupancy data, and outdoor meteorological data within a preset historical window. The labeled sequence includes partition environmental data within the prediction time domain corresponding to the input sequence. A time series fusion Transformer model is trained using the training samples. The time series fusion Transformer model outputs predicted quantiles corresponding to a preset quantile set for the environmental quantities of each partition within the prediction time domain. The predicted quantiles are trained using a quantile loss function. A comfort default risk output layer is set at the output of the time series fusion Transformer model. The comfort default risk output layer outputs the comfort default probability and the comfort default duration probability based on the implicit representation of the time series fusion Transformer model. The comfort default probability is the probability that the comfort index of each partition exceeds a preset comfort threshold at any given time in the predicted time domain. The comfort default duration probability is the probability that the comfort index of the partition exceeds the preset comfort threshold and the duration is not less than a preset duration threshold. The preprocessed real-time operating data, including partition energy consumption data, partition environmental data, equipment operating status data, equipment control quantity data, occupancy data, and outdoor meteorological data within a preset real-time window, is input into the trained time series fusion Transformer model, which outputs the predicted quantile, the comfort default probability, and the comfort default duration probability.

3. The digital management method for building energy efficiency based on big data according to claim 1, characterized in that, S3 includes: Within a preset calibration period, actual environmental quantities for each partition are extracted from the preprocessed real-time operating data, and the predicted quantiles and comfort default probabilities at the corresponding times are extracted simultaneously. Based on the actual environmental quantities and the predicted quantiles, the quantile prediction errors for each environmental quantity in each partition are calculated, and the prediction deviation distribution is statistically obtained based on the quantile prediction errors. The predicted quantiles are calibrated for quantile coverage according to the prediction deviation distribution, and the calibrated predicted quantiles are output. The comfort default probability and the comfort default duration probability are calibrated for probability consistency according to the prediction deviation distribution, and the calibrated comfort default probability and the calibrated comfort default duration probability are output. The probability consistency calibration includes updating the calibration mapping relationship based on a sliding time window to ensure that the statistical coverage of the calibrated comfort default probability meets a preset coverage condition.

4. The digital management method for building energy efficiency based on big data according to claim 1, characterized in that, S4 includes: Obtain the partition risk level parameters and partition occupancy parameters of each partition at each time point within the prediction time domain. Determine the risk weight coefficient of each partition based on the partition risk level parameters and the occupancy weight coefficient of each partition at each time point based on the partition occupancy parameters. Calculate the allocation weight of each partition at each time point based on the risk weight coefficient and the occupancy weight coefficient, and allocate the total comfort risk budget to each partition at each time point within the prediction time domain according to the allocation weight, obtaining the upper limit of the comfort default probability for each partition at each time point. The upper limit of the comfort default probability satisfies that the sum of the upper limits of the comfort default probabilities for each partition at each time point is not greater than the total comfort risk budget. Based on the upper limit of the comfort default probability, the calibrated comfort default probability, and the calibrated comfort default duration probability, generate opportunity constraint parameters. The opportunity constraint parameters include the upper limit of the comfort default probability for each partition at each time point and the comfort constraint conditions corresponding to the upper limit of the comfort default probability, such that when solving the multi-objective model predictive control problem subsequently, the calibrated comfort default probability is not greater than the upper limit of the comfort default probability, and the calibrated comfort default duration probability is not greater than a preset duration probability threshold.

5. The digital management method for building energy efficiency based on big data according to claim 1, characterized in that, S5 includes: Extract the equipment operation status data and equipment control quantity data of each building equipment from the preprocessed real-time operation data, and obtain the rated parameters of the equipment corresponding to the building equipment; identify the start and stop status change time based on the equipment operation status data, and calculate the start and stop interval between two adjacent start and stop status change times; When the start-stop interval is less than the preset short-cycle threshold, and the start-stop state change includes shutting down and restarting or starting and shutting down again, the corresponding start-stop state change is recorded as a short-cycle start-stop event; for each short-cycle start-stop event, the control quantity change amplitude and control quantity change rate are calculated based on the equipment control quantity data, and the control quantity change amplitude and control quantity change rate are normalized based on the equipment rated parameters. The equivalent damage index is calculated based on the number of starts and stops, the start and stop interval, the normalized change amplitude of the control quantity, and the normalized change rate of the control quantity for the short-cycle start and stop events. The equivalent damage index is mapped to a differentiable piecewise convex penalty function, wherein the piecewise convex penalty function includes a first penalty segment corresponding to the equivalent damage index being lower than a first damage threshold, a second penalty segment corresponding to the equivalent damage index being between the first damage threshold and a second damage threshold, and a third penalty segment corresponding to the equivalent damage index being higher than the second damage threshold, and the penalty intensity of each penalty segment increases as the equivalent damage index increases. An equipment life loss model is generated based on the piecewise convex penalty function, and the equipment life loss model is used to characterize the equipment life loss under a given change in equipment control quantity.

6. The digital management method for building energy efficiency based on big data according to claim 1, characterized in that, S6 include: The system state at the current moment is determined based on the equipment operating status data and equipment control quantity data in the preprocessed real-time operating data, and the control quantity of each building equipment in the predicted time domain is used as the decision variable of the control sequence. Based on the partitioned energy consumption data in the preprocessed real-time operating data, a building energy consumption objective function is constructed; based on the opportunity constraint parameters, a comfort objective function and comfort constraint conditions are constructed; and based on the equipment life loss model, an equipment life objective function is constructed. The building energy consumption objective function, the comfort objective function, and the equipment lifespan objective function are combined into a multi-objective model predictive control problem; equipment operation constraints are set in the multi-objective model predictive control problem, including equipment output range constraints, minimum start-up time constraints, minimum shutdown time constraints, and control variable change rate constraints; The comfort constraint is set as an opportunity constraint, such that the calibrated comfort default probability at each time in each partition within the prediction time domain is not greater than the corresponding upper limit of the comfort default probability. The multi-objective model predictive control problem is solved under the conditions of satisfying the equipment operation constraint and the opportunity constraint to obtain the control sequence, and the control quantity at the current time in the control sequence is determined as the current control quantity.

7. The digital management method for building energy efficiency based on big data according to claim 1, characterized in that, S7 includes: The current control quantity is converted into a control command consistent with the building equipment communication protocol, and the control command is sent to the building equipment corresponding to the current control quantity to execute the control; real-time operation data after control execution is collected within a preset collection period, and the real-time operation data after control execution includes zone energy consumption data, zone environmental data, equipment operation status data, and equipment control quantity data; The real-time running data after the control execution is time-aligned and missing values ​​are processed to obtain updated preprocessed real-time running data. The updated preprocessed real-time operating data is used to replace the preprocessed real-time operating data in step S1. In the rolling time domain, the updated preprocessed real-time operating data is used as the input for subsequent steps S2 to S6 to update the predicted quantiles, comfort default probability, and comfort default duration probability. The calibrated predicted quantiles, calibrated comfort default probability, and calibrated comfort default duration probability are updated. The opportunity constraint parameters are updated, the equipment life loss model is updated, and the current control variables are updated, thereby cyclically executing the digital management of building energy efficiency.

8. The digital management method for building energy efficiency based on big data according to claim 4, characterized in that, The risk budget allocation rule further includes: setting a lower limit for the upper limit of the comfort default probability for at least one high-priority partition, and redistributing the remaining total comfort risk budget according to the allocation weight, provided that the lower limit is met.

9. The digital management method for building energy efficiency based on big data according to claim 5, characterized in that, The preset short-cycle threshold is a threshold related to the device type and is adaptively updated based on the historical start-stop interval statistics of the device operating status data.

10. A building energy efficiency digital management system based on big data, used to execute the building energy efficiency digital management method based on big data as described in any one of claims 1 to 9, comprising: The data processing module is used to acquire and preprocess historical and real-time operational data of the building. The probability prediction module is used to train a time series fusion Transformer model based on historical operating data, and to make probability predictions on environmental quantities in each partition within the prediction time domain based on real-time operating data, outputting prediction quantiles, comfort default probability, and comfort default duration probability. The online calibration module is used to perform online calibration of the predicted quantiles, comfort default probability, and comfort default duration probability based on the actual environmental quantities during the calibration period. The risk budget module is used to allocate the upper limit of the comfort default probability for each partition at each time based on the total comfort risk budget and combined with the partition risk level parameter and partition occupancy parameter, and to generate opportunity constraint parameters. The lifespan loss module is used to identify short-cycle start-stop events and calculate the equivalent damage index, and map the equivalent damage index to a differentiable piecewise convex penalty function to obtain the equipment lifespan loss model; the multi-objective control module is used to construct and solve the chance-constrained multi-objective predictive control problem based on energy consumption target, comfort target and equipment lifespan target to obtain the current control quantity; The update module is used to send the current control quantity to the building equipment for execution and collect feedback data, and update the probability prediction, online calibration, chance constraint and model predictive control process in the rolling time domain.

Images