Building energy-carbon double-control and equipment predictive maintenance linkage energy consumption optimization method and system

Abstract

The invention discloses a building energy-carbon double-control and equipment predictive maintenance linkage energy consumption optimization method, and aims to solve the problems that equipment efficiency degradation'no fault but high energy consumption 'is difficult to recognize in time and is difficult to link with energy-carbon control. A control state feature sequence is generated, an equipment health risk sequence is output, a health constraint multi-target optimization model of energy consumption, carbon emission and health risks is established, a control target parameter set containing reward and punishment parameters and constraint parameters is generated, and a safety reinforcement learning controller is adopted to generate, verify and correct control actions, issue and execute the control actions. Constructing an action condition expected energy consumption model to obtain expected energy consumption, forming a total residual error with actual energy consumption, combining execution feedback to obtain an execution residual error, decomposing the execution residual error to generate an efficiency residual error, screening a health window according to the health risk and the execution residual error, and self-updating a baseline model only by using health window data; and determining a degradation threshold based on the health window, calculating an efficiency degradation index, additional energy consumption and additional carbon emission, triggering a maintenance instruction, updating a control target parameter set, and achieving the technical effects of energy consumption and carbon emission optimization, efficiency degradation identification and maintenance linkage.

Description

Technical Field

[0001] This invention relates to the field of building energy management and equipment operation and maintenance, and in particular to a method and system for optimizing energy consumption by linking building energy and carbon dual control with predictive maintenance of equipment. Background Technology

[0002] Driven by national carbon peaking and carbon neutrality goals and building energy conservation policies, public buildings and industrial park buildings are generally equipped with building automation systems, energy management systems, and sub-metering platforms to collect and centrally manage data from energy-consuming equipment such as air conditioning cooling systems, terminal systems, water pumps, and fans. Existing technologies typically collect data on equipment energy consumption, operating conditions, control commands, execution feedback, indoor and outdoor environment, and load to conduct energy consumption monitoring, energy efficiency assessment, and energy-saving control. In terms of control optimization, existing solutions have evolved from traditional rule-based control to setpoint adjustment and start-stop scheduling based on model predictive control and data-driven optimization, and combine time-of-use pricing and electricity carbon emission factors for coordinated optimization of energy consumption and carbon emissions. Regarding equipment operation and maintenance, condition-based fault detection and diagnosis, as well as predictive maintenance, are increasingly being applied to identify sensor anomalies, actuator anomalies, and explicit faults, improving operation and maintenance efficiency and system reliability.

[0003] The existing technology still has the following shortcomings:

[0004] 1. Energy consumption and carbon emission optimization are mostly based on energy consumption indicators and comfort constraints. Equipment health risks are often separated from control strategies. There is a lack of a unified modeling mechanism that incorporates health risks into objective functions and operational constraints, which makes it difficult to reflect the risks of equipment efficiency degradation in control strategies in a timely manner.

[0005] 2. Fault detection and diagnosis is better at identifying fault-type anomalies. However, for situations where there is no fault but high energy consumption due to equipment efficiency degradation, it is easy to confuse it with energy consumption deviations caused by load fluctuations, changes in operating conditions, and control actions not being executed as expected. It is difficult to identify and quantify the degradation impact in a timely and accurate manner.

[0006] 3. The expected energy consumption baseline model relies heavily on long-term training with historical data. During operation, it is easily affected by policy changes and degraded data, which can cause drift and lead to unstable residual thresholds. This reduces the reliability of degradation judgment and maintenance triggering, making it difficult to form a closed-loop linkage between maintenance decisions and subsequent control optimization.

[0007] Therefore, a method and system for optimizing building energy and carbon dual control and predictive maintenance of equipment, which can solve the above-mentioned shortcomings of existing technologies, is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0008] One objective of this invention is to propose a method for energy consumption optimization that integrates building energy and carbon dual control with predictive maintenance of equipment. Addressing the problem in existing technologies where high energy consumption despite no apparent faults is difficult to identify in a timely manner due to equipment efficiency degradation, and where control optimization and maintenance decisions are difficult to coordinate, this invention proposes a method based on the construction of state characteristics using time-aligned and cleaned building operation data, and equipment health risk assessment. A multi-objective optimization model with health constraints, encompassing energy consumption, carbon emissions, and health risks, is established to generate reward / penalty parameters and constraint parameters. A safety reinforcement learning controller outputs control actions and executes them after constraint verification and correction, collecting feedback data. Based on the expected energy consumption model under the action conditions, a total residual is formed, and combined with the execution feedback, execution residuals and efficiency residuals are obtained. A health window is used to achieve baseline model self-updating, thereby determining the degradation judgment threshold and calculating the efficiency degradation index, additional energy consumption, and corresponding additional carbon emissions. Under trigger conditions, maintenance instructions are generated, and the control target parameter set is updated for subsequent cycles. This invention effectively distinguishes between efficiency degradation and control execution deviations, detects hidden high energy consumption in advance and quantifies carbon impact, and achieves closed-loop linkage between control and maintenance, thereby reducing energy consumption and carbon emissions.

[0009] This invention provides a method for optimizing energy consumption by linking building energy and carbon dual control with predictive maintenance of equipment, including:

[0010] S1. Acquire the building operation data of the target building and perform time alignment and data cleaning to obtain aligned and cleaned building operation data; S2. Generate a state feature sequence for control decision-making based on the aligned and cleaned building operation data, and output the equipment health risk sequence through a health assessment model; S3. Based on the state feature sequence, equipment health risk sequence, carbon emission factors, and energy carbon control indicators, establish and solve a health-constrained multi-objective optimization model of energy consumption, carbon emissions, and health risks to generate a control target parameter set, which includes reward / penalty parameters and constraint parameters; S4. Based on the control target parameter set and the state feature sequence, generate a control action sequence through a safety reinforcement learning controller, and verify and correct it according to the constraint parameters to obtain the execution control action sequence. Execute the execution control action sequence and collect actual energy consumption data. S5. Based on the state feature sequence and the execution control action sequence, output the expected energy consumption through the action condition expected energy consumption model, and generate the total residual from the actual energy consumption and the expected energy consumption; S6. Generate the execution residual from the execution control action sequence and the equipment execution feedback, and generate the efficiency residual based on the total residual and the execution residual; S7. Filter the health window set according to the equipment health risk sequence and the execution residual, and update the action condition expected energy consumption model only using the data corresponding to the health window set; S8. Determine the degradation judgment threshold based on the health window set, calculate the efficiency degradation index based on the efficiency residual and the degradation judgment threshold, and calculate the additional energy consumption and corresponding additional carbon emissions caused by efficiency degradation. When the efficiency degradation index meets the preset trigger condition, generate the maintenance instruction and update the control target parameter set for subsequent control cycles.

[0011] Optionally, S1 includes:

[0012] Within the preset sampling period, the target building's equipment energy consumption data, equipment operating condition data, equipment control action data, equipment execution feedback data, outdoor environmental data, indoor load data, electricity carbon emission factor data, energy consumption control indicators, and carbon emission control indicators are collected respectively, and corresponding timestamps are generated for each type of data record.

[0013] Based on the timestamp, time alignment processing is performed on various data records. The time alignment processing includes resampling data with different sampling frequencies according to the preset sampling period, and merging the resampled data records according to the same timestamp.

[0014] Data cleaning is performed on various data records after time alignment. The data cleaning process includes missing value handling, outlier handling, and numerical consistency processing. The numerical consistency processing includes unit conversion and dimension unification.

[0015] Output aligned and cleaned building operation data after the time alignment processing and data cleaning processing, wherein the aligned and cleaned building operation data includes equipment energy consumption data, equipment operating condition data, equipment control action data, equipment execution feedback data, outdoor environmental data, indoor load data, electricity carbon emission factor data, energy consumption control indicators, and carbon emission control indicators corresponding to the same timestamp.

[0016] Optionally, S2 includes:

[0017] The system calls up the aligned and cleaned building operation data, combines equipment energy consumption data, equipment operating condition data, equipment control action data, equipment execution feedback data, outdoor environmental data, and indoor load data according to timestamps, and performs numerical normalization and time series processing on the combined data to generate a state feature sequence for control decision-making.

[0018] Using the energy consumption data, operating condition data, and execution feedback data of the target equipment from the aligned and cleaned building operation data, health characteristics reflecting the energy efficiency deviation of the target equipment are calculated, and the health characteristics are input into the health assessment model to output an equipment health risk sequence that characterizes the risk of efficiency degradation of the target equipment.

[0019] Optionally, S3 includes:

[0020] Based on the state characteristic sequence and equipment health risk sequence, combined with power carbon emission factor data, energy consumption control indicators and carbon emission control indicators, a health-constrained multi-objective optimization model with equipment control actions as decision variables is established, wherein the equipment control actions include setpoint adjustment actions and start-stop scheduling actions.

[0021] In the health-constrained multi-objective optimization model, the energy consumption target is defined as the optimization target of the energy consumption assessment result of the equipment control action under the corresponding working conditions of the state characteristic sequence; the carbon emission target is defined as the optimization target of the carbon emission assessment result obtained by converting the energy consumption assessment result based on the electricity carbon emission factor data; and the health risk target is defined as the optimization target of the health risk assessment result corresponding to the equipment health risk sequence.

[0022] Meanwhile, in the health-constrained multi-objective optimization model, energy consumption constraints corresponding to the energy consumption control index, carbon emission constraints corresponding to the carbon emission control index, and equipment operation constraints are set. The equipment operation constraints include amplitude limits, rate of change limits, and start-stop interval limits for equipment control actions.

[0023] By solving the health-constrained multi-objective optimization model, a set of control objective parameters is generated. The set of control objective parameters includes reward and penalty parameters for the safety reinforcement learning controller to generate control action sequences, and constraint parameters for the feasibility verification and correction of the control action sequences.

[0024] Optionally, S4 includes:

[0025] Based on the set of control target parameters and combined with the state feature sequence, a sequence of control actions is generated by a safety reinforcement learning controller. In the interactive training process, the safety reinforcement learning controller uses the reward and punishment parameters corresponding to the energy consumption target, carbon emission target and health risk target as the reward function parameters, and uses the constraint parameters to limit the output range of the control actions.

[0026] The feasibility of the control action sequence is verified based on the constraint parameters. When the control action in the control action sequence does not meet the equipment operation constraints corresponding to the constraint parameters, the control action that does not meet the constraints is corrected to generate an execution control action sequence that meets the equipment operation constraints.

[0027] The execution control action sequence is sent to the target device and executed. During the execution, the actual energy consumption sequence and the device execution feedback sequence corresponding to the execution control action sequence are collected according to the preset sampling period.

[0028] Optionally, S5 includes:

[0029] An action condition feature sequence is constructed using a state feature sequence and an execution control action sequence. The action condition feature sequence is used to characterize the operating conditions of the target device under corresponding state features and corresponding execution control actions.

[0030] Input the action condition feature sequence into the action condition expected energy consumption model, and output the expected energy consumption sequence corresponding to the execution control action sequence;

[0031] The total residual sequence is generated by calculating the difference between the actual energy consumption sequence and the expected energy consumption sequence time by time.

[0032] Optionally, S6 includes:

[0033] For each control action in the sequence of execution control actions, an execution feedback value corresponding to the timestamp of the control action is determined in the device execution feedback sequence, and the deviation between the control action and the execution feedback value is calculated to obtain a control deviation sequence that characterizes the degree of deviation of the control action execution.

[0034] The control deviation is converted into an execution residual sequence based on the control deviation sequence. The execution residual sequence is used to characterize the energy consumption deviation caused by the control action not being executed as expected.

[0035] An efficiency residual sequence is calculated based on the total residual sequence and the execution residual sequence. The efficiency residual sequence is used to characterize the energy consumption deviation caused by equipment efficiency degradation when the control action is executed as expected.

[0036] Optionally, the S7 includes:

[0037] The equipment health risk sequence and execution residual sequence are divided into multiple time windows according to a preset window length;

[0038] Calculate the equipment health risk statistic and execution residual statistic in each time window, and determine the time window in which the equipment health risk statistic is lower than the first threshold and the execution residual statistic is lower than the second threshold as the health window, thereby generating a set of health windows;

[0039] The expected energy consumption model under action conditions is updated by using the state feature sequence, execution control action sequence, and actual energy consumption sequence corresponding to the health window set to generate an updated expected energy consumption model under action conditions. The parameter update is performed only based on the data corresponding to the health window set to avoid the expected energy consumption model under action conditions being shifted due to the data corresponding to equipment efficiency degradation participating in the update.

[0040] Optionally, S8 includes:

[0041] Based on the set of healthy windows, the efficiency residual benchmark statistic is calculated in the efficiency residual sequence corresponding to the set of healthy windows, and the degradation judgment threshold is determined based on the efficiency residual benchmark statistic.

[0042] The efficiency residual sequence is compared with the degradation judgment threshold, and an efficiency degradation index is generated according to a preset calculation rule, wherein the preset calculation rule includes accumulating the efficiency residuals that exceed the degradation judgment threshold and performing time smoothing.

[0043] The additional energy consumption caused by efficiency degradation is calculated based on the efficiency residual sequence, and the additional energy consumption is converted into additional carbon emissions based on the electricity carbon emission factor data.

[0044] When the efficiency degradation index meets the preset degradation triggering condition, a maintenance instruction is generated, and in the next control cycle, the weight of the health risk target in the health constraint multi-objective optimization model is increased or the equipment operation constraints are tightened to update the control target parameter set, so that the updated control target parameter set is used in step S4 of the next control cycle.

[0045] On the other hand, the present invention also provides a building energy and carbon dual control and equipment predictive maintenance linkage energy consumption optimization system, comprising:

[0046] The system comprises the following modules: a data preprocessing module for acquiring target building operation data and performing time alignment and data cleaning; a status and health assessment module for generating control status characteristics and outputting equipment health risks; a control target parameter generation module for establishing and solving a health-constrained multi-objective optimization model based on control status characteristics, equipment health risks, carbon emission factors, and energy and carbon control indicators, generating a control target parameter set containing reward and penalty parameters and constraint parameters; a safety reinforcement learning control module for generating control actions based on the control target parameter set, verifying and correcting them according to constraint parameters, and then issuing them for execution, while collecting actual energy consumption and equipment execution feedback; and a residual analysis and maintenance linkage module for obtaining total residual, execution residual, and efficiency residual based on control status characteristics, executed control actions, actual energy consumption, and equipment execution feedback. It also filters health windows based on equipment health risks and execution residuals, updates the expected energy consumption model for action conditions, determines degradation judgment thresholds based on health windows, generates an efficiency degradation index from efficiency residuals, calculates the additional energy consumption and corresponding additional carbon emissions caused by efficiency degradation, and generates maintenance instructions and updates the control target parameter set for subsequent control cycles when the efficiency degradation index meets the triggering conditions.

[0047] The beneficial effects of this invention are:

[0048] 1. By constructing an expected energy consumption model for action conditions and combining it with equipment execution feedback to form a decomposition mechanism for total residual, execution residual and efficiency residual, it is possible to effectively distinguish between energy consumption deviation caused by control actions not being executed as expected and hidden high energy consumption caused by equipment efficiency degradation, and realize timely identification and quantitative assessment of "high energy consumption without failure".

[0049] 2. By screening health windows based on equipment health risks and execution residuals, and using only health window data to self-update the expected energy consumption model, baseline drift caused by degradation data is avoided, making the degradation judgment threshold more stable and reducing false alarms and false negatives, thereby improving the reliability of efficiency degradation detection and maintenance triggering.

[0050] 3. By establishing a multi-objective optimization model of energy consumption, carbon emissions, and health risks, a set of control target parameters is generated and used for the reward and penalty setting of the safety reinforcement learning controller. After efficiency degradation is triggered, the weight of health risks is further increased or the operating constraints are tightened to achieve closed-loop linkage between energy and carbon optimization and maintenance decision-making, thereby reducing energy consumption and carbon emissions and controlling health risks. Attached Figure Description

[0051] 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:

[0052] Figure 1The flowchart shows a method and system for optimizing energy consumption by linking building energy and carbon dual control with predictive maintenance of equipment, as proposed in this invention. Detailed Implementation

[0053] 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.

[0054] refer to Figure 1 A method for optimizing building energy and carbon emissions through a linkage between predictive maintenance and equipment, comprising:

[0055] S1. Acquire the building operation data of the target building and perform time alignment and data cleaning to obtain aligned and cleaned building operation data; S2. Generate a state feature sequence for control decision-making based on the aligned and cleaned building operation data, and output the equipment health risk sequence through a health assessment model; S3. Based on the state feature sequence, equipment health risk sequence, carbon emission factors, and energy carbon control indicators, establish and solve a health-constrained multi-objective optimization model of energy consumption, carbon emissions, and health risks to generate a control target parameter set, which includes reward / penalty parameters and constraint parameters; S4. Based on the control target parameter set and the state feature sequence, generate a control action sequence through a safety reinforcement learning controller, and verify and correct it according to the constraint parameters to obtain the execution control action sequence. Execute the execution control action sequence and collect actual energy consumption data. S5. Based on the state feature sequence and the execution control action sequence, output the expected energy consumption through the action condition expected energy consumption model, and generate the total residual from the actual energy consumption and the expected energy consumption; S6. Generate the execution residual from the execution control action sequence and the equipment execution feedback, and generate the efficiency residual based on the total residual and the execution residual; S7. Filter the health window set according to the equipment health risk sequence and the execution residual, and update the action condition expected energy consumption model only using the data corresponding to the health window set; S8. Determine the degradation judgment threshold based on the health window set, calculate the efficiency degradation index based on the efficiency residual and the degradation judgment threshold, and calculate the additional energy consumption and corresponding additional carbon emissions caused by efficiency degradation. When the efficiency degradation index meets the preset trigger condition, generate the maintenance instruction and update the control target parameter set for subsequent control cycles.

[0056] In this specific embodiment, S1 includes:

[0057] The system completes the collection, time alignment, and data cleaning of building operation data for the target building according to a preset sampling period, and outputs the aligned and cleaned building operation data. The preset sampling period is set to [missing information]. and Using the unified clock of the building energy management system as the time reference and maintaining consistency of clocks across all data acquisition terminals via network time synchronization, the start time of the control cycle is recorded as... , will the Each sampling time is recorded as ,in The sampling sequence number is incremented starting from 0 and satisfies In the formula Indicates the aligned first A standard timestamp, Indicates the data start timestamp for this control cycle. Indicates the sampling sequence number. Indicates the preset sampling period;

[0058] At each sampling time Collect and record equipment energy consumption data, equipment operating condition data, equipment control action data, equipment execution feedback data, outdoor environmental data, indoor load data, electricity carbon emission factor data, energy consumption control indicators, and carbon emission control indicators. Write the original timestamp corresponding to the time of generation for each data record, and ensure that the original timestamps are uniformly in the same time zone and with the same time precision. Equipment energy consumption data is provided by individual metering meters. When the meter output is interval energy, it is directly recorded as the energy of the current sampling interval. When the meter output is cumulative energy, the interval energy is obtained by subtracting the cumulative energy of two adjacent sampling times and then aggregated to the sampling time. ;

[0059] The equipment operating condition data includes continuous or quasi-continuous quantities reported by sensors and equipment controllers, such as chilled water supply and return water temperature, cooling water supply and return water temperature, supply and return water pressure difference, water pump frequency, valve opening degree, fan frequency, and unit load rate.

[0060] The equipment control action data includes setpoint adjustment actions and start / stop scheduling actions. Setpoint adjustment actions record the setpoint in numerical form, while start / stop scheduling actions record the start / stop status in binary form.

[0061] The equipment execution feedback data includes execution feedback quantities that correspond one-to-one with the equipment control actions. For setpoint actions, the feedback is the actual setpoint or actual process quantity reached. For start-stop actions, the feedback is the equipment operating status and key interlock status.

[0062] Outdoor environmental data includes outdoor dry-bulb temperature, outdoor wet-bulb temperature, outdoor relative humidity, solar irradiance, etc., while indoor load data includes indoor dry-bulb temperature, indoor relative humidity, occupancy load, and fresh air volume.

[0063] Electricity carbon emission factor data is recorded as the carbon emissions per unit of electrical energy and uniformly converted to [amount missing]. ;

[0064] Energy consumption control targets and carbon emission control targets are written at the beginning of each control period and remain unchanged during that control period. Energy consumption control targets include the upper limit of energy consumption per unit time and the budget value of energy consumption per unit time, and carbon emission control targets include the upper limit of carbon emissions per unit time and the budget value of carbon emissions per unit time.

[0065] After completing the above data collection, time alignment processing is performed based on the original timestamps, and sampling frequencies higher than [the original timestamps] are adjusted accordingly. Data channels according to Resampling is performed over a time interval, with the resampling rule set to take the arithmetic mean of the continuous measurements over the interval and aggregate the average values. The cumulative amount is first divided into interval amounts and then aggregated. For event quantities and start / stop binary quantities, interval undefined values ​​are preserved and aggregated. Sampling frequency lower than The data channels maintain the most recent valid reported value forward until the next valid report is received and then aggregated to the corresponding channels. ;

[0066] After resampling, all data channels are... Merge them as primary keys to form each An aligned data table corresponding to each record, and each record simultaneously contains equipment energy consumption data, equipment operating condition data, equipment control action data, equipment execution feedback data, outdoor environmental data, indoor load data, electricity carbon emission factor data, energy consumption control indicators, and carbon emission control indicators.

[0067] In the data cleaning process, the merged aligned data table is processed for missing values, outliers, and numerical consistency through each channel. The missing value processing rule is set as follows: when the number of consecutive missing points does not exceed 2 sampling points, linear interpolation is used to fill in the missing points; when the number of consecutive missing points exceeds 2 sampling points, the channel is marked as missing within the corresponding time range, and the entire record within that time range is marked as unusable for subsequent modeling and control training.

[0068] The outlier handling rules are set by first determining the physical upper and lower limits for each channel based on the sensor range and the allowable range specified on the device nameplate. and ,in This represents the minimum allowable value for the channel, determined by the larger of the lower limit of the range and the lower limit of the process. This indicates the maximum allowable value for the channel, determined by the smaller of the upper limit of the measuring range and the upper limit of the process; exceeding this value will result in... Data points are identified as abnormal and marked as missing. Then, median filtering is performed on the remaining data points using a 7-point sliding window. Data points that deviate from the median in the window by more than three times the absolute median of the window are identified as abnormal and marked as missing. Subsequently, they are filled in or marked according to the missing value handling rules.

[0069] The numerical consistency processing rules are set to uniformly convert energy consumption to kWh, power to kW, and temperature to [missing value]. Traffic is uniformly converted to Pressure is uniformly converted to kPa, valve opening degree and frequency are uniformly converted to percentages, and electricity carbon emission factor is uniformly converted to... Furthermore, field mapping is performed on the naming of the same physical quantity in different systems to ensure that the same field corresponds to only one meaning;

[0070] After cleaning is completed, the aligned and cleaned building operation data is output, wherein the aligned and cleaned building operation data is in the form of... Create a time series table for the index.

[0071] In this specific embodiment, S2 includes:

[0072] Based on the aligned and cleaned building operation data, a state characteristic sequence for control decision-making is generated, and a health risk sequence of equipment is output through a health assessment model, with the target equipment set as the chiller unit of the target building.

[0073] First, retrieve the aligned and cleaned building operation data and sort it by timestamp. Data records from the same moment are combined into fields, and these combined fields are written into the same status record in a fixed field order to avoid feature misalignment. The fixed field order is as follows: unit input power and interval power in equipment energy consumption data; chilled water supply and return temperatures and chilled water flow rates, and cooling water supply and return temperatures and cooling water flow rates in equipment operating condition data; chilled water outlet temperature setpoint and unit start / stop commands in equipment control action data; actual chilled water outlet temperature and unit operating status in equipment execution feedback data; outdoor dry-bulb temperature and outdoor wet-bulb temperature in outdoor environmental data; and terminal return water temperature and terminal flow rate in indoor load data. Each combined status record is then recorded as a status feature vector and... Arrange the state features in ascending order to obtain a sequence of state features;

[0074] Subsequently, numerical normalization was performed on the state feature sequence. Normalization employed independent minimum-maximum normalization for each feature channel, and the normalized values ​​were then cropped to the specified values. The normalization parameters for the interval are calculated from the building operation data after 30 consecutive days of aligned cleaning following the completion of commissioning and acceptance of the target building. The minimum parameter for each characteristic channel is taken as the 1st percentile of the sample for that channel, and the maximum parameter is taken as the 99th percentile of the sample for that channel, to suppress the influence of extreme values ​​on the normalization scale. The normalization calculation satisfies:

[0075] ;

[0076] In the formula Represents the timestamp index and is One-to-one correspondence, This represents the feature channel index and corresponds one-to-one with a specific field in the fixed field order mentioned above. Indicates the timestamp First The original values ​​of each feature channel, This represents the normalized value. Indicates the first The normalized minimum parameter of each feature channel. Indicates the first The normalized maximum parameter of each feature channel, and in The channel The value is fixed at 0.5 to ensure numerical stability and avoid division by zero;

[0077] After normalization, the state feature sequence is subjected to time serialization. Time serialization uses a fixed-length sliding window to concatenate sequence samples, with the window length set to [value missing]. And corresponding to 60 minutes of historical information, each timestamp The corresponding sequence samples are from Continuous The normalized state feature vectors are stacked in chronological order to obtain the result, and the insufficient ones are... The initial segment samples are directly discarded to ensure that the input dimension is fixed, thereby obtaining the state feature sequence used for control decision-making;

[0078] When outputting the device health risk sequence, health characteristics are calculated by extracting device energy consumption data, device operating condition data, and device execution feedback data corresponding to the target device from the aligned and cleaned building operation data. These health characteristics consist of five categories of quantities, and each... A set of values ​​is calculated and arranged in a fixed order to form a health feature vector. The five types of quantities are, in order: deviation of the unit's overall energy efficiency, deviation of the condensing-side heat exchange approximation, deviation of the evaporating-side heat exchange approximation, deviation of the chilled water outlet temperature setpoint tracking, and deviation of start-up / shutdown and operation consistency. The deviation of the unit's overall energy efficiency is determined by the deviation between the instantaneous energy efficiency coefficient (EEC) calculated from the instantaneous cooling capacity and instantaneous input electrical power reported by the unit controller and the reference EEC of the rated performance curve under the same outdoor wet-bulb temperature and the same unit load rate. The rated performance curve is obtained by using discrete points from the unit's nameplate performance table. The regression coefficients obtained by quadratic polynomial regression fitting are fixed and saved during equipment commissioning and acceptance. The deviation of the condensing side heat exchange approximation is determined by the deviation of the difference between the cooling water outlet temperature and the condensing saturation temperature relative to its acceptance benchmark. The deviation of the evaporating side heat exchange approximation is determined by the deviation of the difference between the evaporation saturation temperature and the chilled water outlet temperature relative to its acceptance benchmark. The deviation of the chilled water outlet temperature setpoint is determined by the absolute deviation between the actual chilled water outlet temperature and the setpoint. The deviation of the start-up and shutdown consistency is determined by the duration of the logical inconsistency between the unit start-up and shutdown command and the unit operating status.

[0079] Each The health feature vector is input into the health assessment model to output a device health risk sequence. The health assessment model is constructed as a gradient boosting decision tree binary classification model and the output is a range of values. Risk score ,in Represents timestamp The corresponding probabilistic quantification result of the risk of efficiency degradation of the target equipment is used. The gradient boosting decision tree binary classification model is trained offline once using maintenance work orders and metering data after the target building is put into operation, and the parameters are kept unchanged during the operation period. The positive class samples of the training data are defined as the samples at each time point corresponding to the health feature vector sequence within 14 consecutive days before the closing time of the efficiency-related maintenance work order, and the label is 1. The negative class samples are defined as the samples at each time point corresponding to the health feature vector corresponding to the equipment execution feedback without abnormal alarms when the time window of any efficiency-related maintenance work order exceeds 30 days, and the label is 0. The hyperparameter of the model is fixed at the number of trees. Maximum depth Learning rate Row sampling ratio Column sampling ratio Minimum leaf sample weight Minimum loss from splitting decreases The model uses log loss as the objective function and a logistic function at the output to map the model score to a risk score. , will be according to All in ascending order This constitutes a sequence of equipment health risks.

[0080] In this specific embodiment, S3 includes:

[0081] At the beginning of each control cycle Based on state feature sequences and equipment health risk sequences, and combined with electricity carbon emission factor data, energy consumption control indicators, and carbon emission control indicators from aligned and cleaned building operation data, a health-constrained multi-objective optimization model with equipment control actions as decision variables is established and solved to generate a set of control target parameters. The target equipment is set as a chiller unit, and the preset sampling period is set to [missing information]. And set the control prediction time domain to Each sampling step corresponds to the next 60 minutes;

[0082] Define equipment control actions as a sequence of control actions. and The control actions at each moment It consists of setpoint adjustment actions and start / stop scheduling actions, and is denoted as... ,in This indicates the setpoint for the chilled water outlet temperature of the chiller unit, and the unit is... Indicates the start / stop command of the chiller unit and and Indicates power on, Indicates that the machine is stopped;

[0083] The energy consumption constraint is set as the upper limit of energy consumption in the prediction time domain. ,in We directly take the corresponding value of the unit time energy consumption limit in the energy consumption control index within the 60-minute window and unify the unit to kWh, and set the carbon emission constraint as the carbon emission limit within the prediction time domain. ,in Take the corresponding value of the upper limit of carbon emissions per unit time in the carbon emission control indicators within the 60-minute window and unify the units to . ;

[0084] The equipment operation constraints are set into three types: control action amplitude limit, rate of change limit, and start / stop interval limit, and written into the constraint parameters. The amplitude limit is fixed at [value missing]. And the endpoints of the interval are respectively used as constraint parameters and The rate of change limit is fixed at 1. and Based on the current device health risk Determine and write it as a constraint parameter, where Indicates the equipment health risk sequence in Risk score at any time and ,when Time setting ,when Time setting ,when Time setting The start-stop interval limit is fixed at the minimum number of steps between two consecutive start-stop state transitions. and Also by Determine and write it as a constraint parameter when Time setting ,when Time setting ,when Time setting And thereby constrain any time After a flip from 0 to 1 or from 1 to 0 occurs, in subsequent consecutive... Maintain within each sampling step No longer flipping;

[0085] In a health-constrained multi-objective optimization model, the energy consumption objective is defined as the energy consumption target in the state characteristic sequence corresponding to the operating condition and the candidate control action sequence. Predicted energy consumption assessment results ,in The expected energy consumption is obtained by reasoning about the action condition features at each moment in the prediction time domain using the action condition expected energy consumption model, and then accumulated to obtain the expected energy consumption when... At that moment, the expected energy consumption is set to 0. The rules for constructing the action condition features are consistent with those in step S5, and the outdoor environment and indoor load in the prediction time domain are used. The alignment and cleaning values ​​at each moment are forward-preserved to ensure that the evaluation input is determined;

[0086] The carbon emission target is defined as the predicted carbon emission assessment result obtained by converting energy consumption assessment results based on electricity carbon emission factor data. ,in It is obtained by multiplying and summing the stepwise expected energy consumption and stepwise carbon emission factors within the prediction time domain, and the carbon emission factors are adopted in the prediction time domain. The alignment and cleaning values ​​at each moment are forward maintained and the units are unified. ;

[0087] The health risk target is defined as the health risk assessment result corresponding to the equipment health risk sequence. ,in Based on the current risk score Together with the control motion stress term, the control motion stress term is determined. It is composed of the sum of the absolute changes in the setpoint within the prediction time domain and the number of start-stop / reverse cycles, and is then linearly weighted using fixed coefficients of 0.05 and 0.10. Add and cut off ;

[0088] The above three objectives are used to construct an objective function through weighted quantization, which is then used to generate reward and penalty parameters. The objective function is written as follows:

[0089] ;

[0090] In the formula Represents the sequence of control actions The scalarized comprehensive evaluation value, Indicates by The candidate control action sequence is composed of The energy consumption target weight is represented and written into the control target parameter set as a reward or penalty parameter. This represents the predicted energy consumption assessment result corresponding to the candidate control action sequence. This represents the upper limit of predicted time-domain energy consumption given by the energy consumption control index. The carbon emission target weights are represented and written into the control target parameter set as reward and penalty parameters. This indicates the predicted carbon emission assessment results corresponding to the candidate control action sequence. This represents the upper limit of predicted carbon emissions over the given time period, as indicated by the carbon emission control indicators. The weights of health risk targets are represented and written into the control target parameter set as reward and punishment parameters. This indicates the health risk assessment results corresponding to the candidate control action sequence;

[0091] The solution to the health-constrained multi-objective optimization model employs a fixed-iteration cross-entropy method with a fixed random number seed of 0 to ensure a deterministic solution process. In each outer iteration, the cross-entropy method generates a set of candidate control action sequences and calculates each sequence individually. and And thus obtain This will violate energy consumption constraints. Violation of carbon emission constraints Candidate sequences that violate equipment operation constraints are directly determined as infeasible solutions and eliminated; only those within the feasible solution set are considered as... Sort by size from smallest to largest before selecting As an elite set, the Gaussian sampling distribution mean and variance of the setpoint adjustment actions and the Bernoulli sampling distribution parameters of the start-stop scheduling actions are updated accordingly. The outer iteration count is fixed at 5, and the number of candidate sequences in each iteration is fixed at 256. After the last iteration, the feasible control action sequence with the smallest comprehensive evaluation value is output as the reference optimal sequence for that control cycle. Simultaneously, the sequence used to construct the objective function is... , Fixed setting and according to Segmented values ​​and corresponding to each segment and will Compared with the above energy consumption constraints Carbon emission constraints and equipment operating constraints They are jointly written into the control objective parameter set, so that the control objective parameter set can be used in step S4 as the reward function parameters and action feasible domain constraint parameters of the safety reinforcement learning controller.

[0092] In this specific embodiment, S4 includes:

[0093] Based on the set of control target parameters and the sequence of state features, a safety reinforcement learning controller is used at each sampling time. The control actions are generated and executed after being verified and corrected by the constraint parameters, while the actual energy consumption and equipment execution feedback are collected.

[0094] The security reinforcement learning controller consists of a policy network and a value network, and is trained and inferred online using a constrained policy optimization algorithm. The input to the policy network is the time step. Corresponding time-seriesd state input The From the recent The normalized state feature vectors of each sampling point are concatenated in chronological order and correspond to 60 minutes of historical information. The output of the policy network is an action consistent with the device control action. ,in The setpoint for chilled water outlet temperature and the unit is... For start / stop commands and Indicates power on, Indicates that the machine is stopped;

[0095] The policy network structure is fixed as a two-layer fully connected network with 128 neurons per layer. The activation function is fixed as ReLU, and a dual-output head is used to model mixed actions. The setpoint output head outputs a normalized continuous quantity. Candidate setpoints are obtained through linear mapping. Start-stop output head output power-on probability Candidate start / stop commands are then determined using a threshold of 0.5. ;

[0096] Value networks and strategy networks share the same dimension of input. Furthermore, the same two-layer fully connected structure is used to output state values ​​for training;

[0097] The security reinforcement learning controller is trained offline using an interactive playback environment constructed from historical operating data, and then updated online incrementally with a fixed step size after deployment. The training hyperparameters are fixed as a discount factor. Generalized dominance estimation parameters Strategy clipping factor Learning rate Number of sample steps for each parameter update Number of iteration rounds per batch Value loss coefficient Entropy regularity coefficient During the inference phase, random sampling is turned off, and the output is determined by the mean output of continuous actions and the start / stop probability threshold to ensure that the control action is determined.

[0098] The reward and penalty parameters in the control objective parameter set are directly used in the construction of the reward function. As a weight, and incorporating normalized values ​​of energy consumption and carbon emissions to unify the dimensions, the single-step return is denoted as... And it is defined as:

[0099] ;

[0100] In the formula Indicates the sampling time The single-step reward scalar corresponding to the control action. This represents the weight of the energy consumption target and is taken from the control target parameter set. Indicates the interval The actual electrical energy obtained from the equipment's energy consumption data, expressed in kWh. This represents the upper limit of energy consumption constraints, is taken from the control objective parameter set, and is consistent with the prediction time domain window. This represents the carbon emission target weights, taken from the control target parameter set. Indicates the interval Actual carbon emissions within, in units of And by The conversion rules are used to calculate the result. This indicates the data on the carbon emission factor of electricity at time [time]. The value of and the unit is , This represents the upper limit of carbon emission constraints, is taken from the control target parameter set, and is consistent with the prediction time domain window. This represents the weight of health risk objectives and is taken from the control objective parameter set. Indicates the equipment health risk sequence at time [time]. The risk score and its range is ;

[0101] In the action verification and correction process, the constraint parameters in the control target parameter set are treated as hard constraints and processed by the safety layer. The safety layer first performs amplitude verification and selects candidate setpoints. Crop to Within the range, a feasible set value for the amplitude can be obtained, where and All parameters are taken from the control objective parameter set and the unit is... ;

[0102] The security layer then performs a rate-of-change check and applies the previously issued actionable settings. Calculate the change in this case, and limit the change to a certain value. To obtain a feasible setpoint for the rate of change, where Taken from the control objective parameter set and in units of ;

[0103] The security layer then performs start-stop interval verification and maintains an index of the most recent occurrence time of the start-stop state reversal. When candidate start / stop command Compared to the previous start / stop command Inconsistent and When Forced correction to and maintain Unchanged, updated when flipping is allowed. ,in The minimum start-stop interval is taken from the control target parameter set;

[0104] After completing the above hard constraint processing, the resulting action is recorded as the execution control action. And write it into the execution control action sequence;

[0105] When control actions are issued and executed, the building automation system will... Write the chilled water outlet temperature setpoint to the chiller unit controller and The start / stop command point is written, employing a handshake mechanism of one-write-one-verification. After writing, the returned value and command point status are read to confirm successful transmission. Subsequently, within the interval... The actual energy consumption is collected according to the preset sampling period. The equipment execution feedback includes the actual chilled water outlet temperature, unit operating status, key interlock status, and execution confirmation quantities corresponding to control actions, along with timestamps. The binding forms a device execution feedback sequence.

[0106] In this specific embodiment, S5 includes:

[0107] Based on the state feature sequence and the execution control action sequence, an action condition feature sequence is constructed, and the expected energy consumption sequence is output through the action condition expected energy consumption model. Then, the total residual sequence is generated from the actual energy consumption sequence and the expected energy consumption sequence.

[0108] At each sampling time Read the normalized state feature vector aligned with this timestamp and denote it as... The Read the data record that corresponds to the same timestamp and has the same fixed field order used for control decisions in step S2, and execute the control action corresponding to that timestamp. And use it as the input for the action condition, where Indicates at time The issued chilled water outlet temperature setpoint and the unit is Indicates at time The issued start / stop command and and Indicates power on, Indicates that the machine is stopped;

[0109] Adjust the setpoint action according to the amplitude constraint parameter. and Perform linear normalization to obtain the normalized setpoint components and then... and The action condition feature vector is synthesized according to a fixed splicing order. The fixed splicing order is set to write first. All components are then written, followed by the normalized setpoint components, and finally the start / stop components. , and according to In ascending order will put all Composition of action condition feature sequence;

[0110] The expected energy consumption model under action conditions is denoted as: ,in For gradient boosting decision tree regression models, The model parameter set includes all tree structure parameters and leaf node value parameters. The model input is... And the model output is in the interval Corresponding expected energy consumption And the unit is kWh, and will be according to All in ascending order Composition of expected energy consumption sequence;

[0111] The When the system is put into operation, offline training is completed using 30 consecutive days of aligned and cleaned building operation data. In subsequent operations, self-updates are performed only according to step S7. The supervision label for offline training is the interval electrical energy in the equipment energy consumption data and recorded as the actual energy consumption. The training objective function is the mean squared error, and the training hyperparameter is fixed at the number of trees. Maximum depth Learning rate Row sampling ratio Column sampling ratio Minimum number of leaf samples L2 regularization coefficient And the random number seed is fixed to 0 to ensure that the training results are certain;

[0112] when The model will be output at that time. Fixed setting for expected power consumption during shutdown and standby ,in To filter in offline training data Furthermore, a constant obtained by taking the average actual energy consumption of samples with normal key interlock states is used to characterize the standby energy consumption of auxiliary equipment in the shutdown state.

[0113] After obtaining the desired energy consumption sequence, read the values ​​corresponding to each energy consumption sequence from the actual energy consumption sequence. Corresponding actual energy consumption and with expected energy consumption The total residual is generated by subtracting the values ​​at each time step. In the formula Represents timestamp The corresponding total residual and the unit is Representing an interval Actual energy consumption and unit Representing an interval Expected energy consumption and unit Representation and timestamp One-to-one corresponding sampling sequence number, and will be according to All in ascending order The total residual sequence is formed.

[0114] In this specific embodiment, S6 includes:

[0115] Generate execution residual sequence and efficiency residual sequence based on execution control action sequence, equipment execution feedback sequence and total residual sequence;

[0116] For each sampling time Execution control actions Extract the execution feedback value corresponding to the execution control action from the device execution feedback sequence and perform time attribution processing. The time attribution processing is set to use timestamps as... Device execution feedback values ​​are attributed to timestamps. The execution control action is used to characterize the action within the interval. The execution result within, and when When the timestamp is the last sample number of the current batch of data, it is set to... Device execution feedback values ​​are attributed to timestamps. The execution control action;

[0117] The execution feedback value includes the setpoint readback value. and running status readback value ,in This indicates the actual written value of the chilled water outlet temperature setpoint returned by the device controller, and the unit is... This indicates the unit operating status reported back by the equipment controller and and Indicates running, Indicates that the machine is stopped;

[0118] In obtaining A control deviation sequence is then generated, consisting of two parts: setpoint control deviation and start / stop control deviation. The setpoint control deviation is denoted as... And its value is and The difference, the start-stop control deviation is denoted as And its value is and The difference;

[0119] To avoid misjudgments caused by measurement jitter, a defined dead zone and persistence criterion are introduced when generating the control deviation sequence, and the setpoint dead zone is fixed at a certain value. ,when At that moment, the setpoint action is considered to have been executed as expected, and the feedback setpoint used for subsequent calculations is set to [value]. ,when The feedback setting value used for subsequent conversions will be set to... ;

[0120] The start / stop continuity criterion is fixed as continuous. Only when the sampling step size is inconsistent and mismatched is it confirmed that the start / stop operation has not been executed as expected. And in and Both times are satisfied The feedback start / stop status used for subsequent conversions will be set to [state]. Otherwise, the feedback start / stop status used for subsequent conversions will be set to 0. ;

[0121] The feedback action obtained according to the above rules is denoted as:

[0122] ;

[0123] in This represents the feedback setpoint used to perform residual conversion, and the unit is... This indicates the start / stop status of the feedback used to perform residual conversion and takes a value of 0 or 1;

[0124] Subsequently, the control deviations are converted into execution residual sequences based on the control deviation sequence. The conversion method is set to reuse the expected energy consumption model of the action conditions in step S5. And maintain model parameters Unchanged, for each To execute control actions respectively With feedback action Construct two action condition features and input them. The state part is taken from step S5 and... The corresponding normalized state feature vectors and the action parts are respectively taken and Thus, the expected energy consumption under the execution conditions is obtained. Expected energy consumption with feedback conditions ,in Represents the interval obtained by performing a control action. Expected energy consumption and unit This represents the interval obtained based on the feedback action. The expected energy consumption is given in kWh, and the residual will be calculated. Defined as Compared to The deviation is used to characterize the energy consumption deviation caused by the control action not being executed as expected, and when the device execution feedback value is marked as missing or abnormal in step S1, Set directly to Thus Fixed to 0 to avoid unreliable feedback driving misattribution;

[0125] After obtaining the execution residual sequence, an efficiency residual sequence is generated based on the total residual sequence and the execution residual sequence. The efficiency residuals are then distributed according to time intervals. The calculation is as follows:

[0126] ;

[0127] In the formula Represents timestamp The corresponding efficiency residual, expressed in kWh, is used to characterize the energy consumption deviation caused by equipment efficiency degradation when the control actions are executed as expected. Represents timestamp The corresponding total residual and the unit is Represents timestamp The corresponding execution residual and the unit is Representation and timestamp One-to-one corresponding sampling sequence number, and will be according to Increasing arrangement , and These are respectively composed of the control deviation sequence, the execution residual sequence, and the efficiency residual sequence output.

[0128] In this specific embodiment, S7 includes:

[0129] A set of health windows is selected based on the equipment health risk sequence and the execution residual sequence, and the expected energy consumption model of the action conditions is applied only using the data corresponding to the set of health windows. Perform parameter updates to obtain the updated expected energy consumption model for the action conditions;

[0130] First, the risk score in the equipment health risk sequence is recorded as follows: The execution residual sequence is denoted as and The unit is kWh, and both are set to the same timestamp. After alignment, the time window is divided into multiple time windows according to a preset window length, with the window length fixed at [value]. Each sampling point and the preset sampling period Corresponding to 60 minutes, the window is divided in a non-overlapping manner and the first... The starting sampling number of each window is denoted as . and The window number is incremented starting from 0;

[0131] Within each window, the equipment health risk statistic and the execution residual statistic are calculated separately. The equipment health risk statistic is the arithmetic mean of the risk scores within the window and is denoted as [missing value]. The execution residual statistic is the arithmetic mean of the absolute values ​​of the execution residuals within the window, and is denoted as . Its calculation satisfies:

[0132] ;

[0133] In the formula Indicates the first The device health risk statistics within a time window, with a value range of [value missing]. Indicates the first The execution residual statistics within each time window, with units of Indicates the window length and is fixed at 1. Indicates the window number. Indicates the first The starting sampling sequence number of each window. Indicates the offset index within the window, with a value ranging from 0 to... , Represents timestamp The corresponding risk score, Represents timestamp The corresponding execution residual;

[0134] The first threshold is fixed at [value]. The second threshold is fixed at 1. and will satisfy and The window is identified as a healthy window, thus forming a set of healthy windows. ;

[0135] When generating the sample set for updating, only the health window set is used. Samples are extracted from the covered timestamp set and validity screening is performed. The validity screening rule is set as follows: the timestamp was not marked as missing or abnormal in step S1, and the corresponding control action is executed. Both the device execution feedback sequence and the data have valid readbacks, thus obtaining a set of health samples. ,in It is an action condition feature vector, and its construction order and field meaning remain unchanged. This represents actual energy consumption in kWh.

[0136] To ensure the update process is deterministic and to prevent the baseline from being dragged down by degraded data, the model update employs a fixed-capacity sliding buffer mechanism and only writes to the healthy sample set. The samples in the buffer are denoted as And the capacity is fixed at Each sample is assigned a corresponding 30-day historical data set, and data is appended and written from oldest to newest according to timestamps. When the capacity is exceeded, the oldest sample is deleted in a first-in-first-out manner.

[0137] The model parameter update trigger rule is fixed at once every 72 sampling points, corresponding to a 6-hour interval, and is triggered from the buffer upon triggering. Read all samples for expected energy consumption model under action conditions Incremental training is performed, where the existing tree structure is kept unchanged and additional training is performed afterward. A new tree is added to correct the residuals, and the objective function for training is the mean squared error, with the training hyperparameter fixed at the maximum depth. Learning rate Row sampling ratio Column sampling ratio Minimum number of leaf samples L2 regularization coefficient The random number seed is set to 0, and the model parameters are changed from [previous settings] after the update is complete. Replace with the updated parameters And will As the action condition expected energy consumption model for the subsequent step S5 to calculate the expected energy consumption sequence, this ensures that the model self-update is driven only by the health window data and avoids baseline drift caused by the participation of data corresponding to equipment efficiency degradation in the update.

[0138] In this specific embodiment, S8 includes:

[0139] Based on health window set Determine the degradation threshold and base it on the efficiency residual sequence. The system calculates the efficiency degradation index and quantifies the additional energy consumption and carbon emissions. It also generates maintenance instructions and updates the control target parameter set when the preset degradation trigger conditions are met.

[0140] Among them, the health window collection It consists of several non-overlapping time windows with a fixed window length. Each sampling point corresponds to 60 minutes, and the efficiency residual... With timestamp One-to-one correspondence and the unit is kWh, first in the health window set Extract the efficiency residual sample set from all covered sampling points and denot it as and to Calculate the baseline efficiency residual statistic, which includes the baseline mean. Standard deviation from the benchmark ,in The arithmetic mean of the efficiency residuals within the health window, expressed in units of 1. The standard deviation of the efficiency residuals within the healthy window is expressed in kWh, and the degradation threshold is determined accordingly. and ,in This represents the threshold used to determine efficiency degradation, and the unit is kWh.

[0141] In obtaining Then, for each sampling time efficiency residual Calculate the amount exceeding the threshold And Defined as ,in Indicates time The energy consumption deviation exceeding the threshold due to efficiency degradation, expressed in kWh, will be... As a sample-point quantification result of the additional energy consumption caused by efficiency degradation, the electricity carbon emission factor is also read based on the aligned and cleaned building operation data. Additional carbon emissions per sampling point are defined as and The product of, where Represents timestamp The corresponding carbon emission factor for electricity, and the unit is and within the statistical period The additional energy consumption for the statistical period is obtained by accumulating point by point. The additional carbon emissions for the statistical period are obtained by summing up the emissions point by point.

[0142] In the calculation of the efficiency degradation index, a deterministic rule is adopted to accumulate and smooth the excess threshold over time, and a smoothing coefficient is set. and initialization And calculate the efficiency degradation index recursively according to the sampling sequence number:

[0143] ;

[0144] In the formula Represents timestamp The corresponding efficiency degradation index and the unit is Represents the time smoothing coefficient and is fixed at 1. Indicates the previous sampling time. The efficiency degradation index. Represents timestamp The efficiency residual and the unit is Indicates the degradation threshold and the unit is . Representation and timestamp One-to-one corresponding sampling sequence number;

[0145] The default degradation trigger condition is fixed to be set to occur continuously. All sampling points satisfy the following conditions. and ,in Indicates a continuous determination length that is fixed. The threshold value is expressed in kWh. When a preset degradation trigger condition is met, a maintenance instruction is immediately generated and written into the operation and maintenance system. The maintenance instruction includes the target device identifier and the trigger timestamp. Efficiency degradation index The statistical period includes additional energy consumption, additional carbon emissions during the statistical period, and a fixed set of maintenance tasks, wherein the set of maintenance tasks is fixedly set as condenser heat exchange surface cleaning, evaporator heat exchange surface cleaning, and refrigerant charge verification.

[0146] Simultaneously, the control target parameter set will be updated in the next control cycle to tighten the health-related control strategy and increase the weight of health risk targets. Fixed update to 2.0, setting the rate of change constraint parameter. The fixed update value is 0.1 and the unit is... and set start-stop interval constraint parameters The update is fixed at 12 sampling steps, so that the updated set of control target parameters is used in step S4 of subsequent control cycles to reduce further degradation accumulation and additional energy consumption growth.

[0147] 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.

[0148] This invention employs a linked architecture combining health-constrained multi-objective optimization and safety reinforcement learning control. It integrates energy consumption, carbon emission, and health risk objectives into the control objective parameter generation process. After constraint verification and correction, the safety reinforcement learning controller outputs executable control actions, thereby achieving energy consumption optimization while meeting energy and carbon emission indicators and equipment operation constraints. Simultaneously, this invention introduces a residual degradation detection mechanism based on policy consistency. It generates expected energy consumption through an action-condition expected energy consumption model and forms a total residual with the actual energy consumption. This residual is then combined with control actions and execution feedback to obtain the execution residual, further separating out the efficiency residual. This allows energy consumption deviations occurring even when control actions are executed as expected to be attributed to efficiency degradation, enabling timely identification of "high energy consumption without failure." Furthermore, it allows for the calculation of additional energy consumption and corresponding additional carbon emissions caused by efficiency degradation, forming a quantitative basis for maintenance decisions.

[0149] This invention improves the algorithm structure to enhance the stability of degradation identification: First, by modeling the residuals of action conditions, both operating conditions and control actions are used as inputs to the expected energy consumption, reducing misjudgments caused by load fluctuations and changes in operating conditions. Second, by decomposing the residuals, the impact of execution deviations is removed from the total residuals, resulting in efficiency residuals that better characterize the true efficiency degradation. Third, through health window screening and baseline self-updating, only window data with low health risk and small execution residuals are used to update the expected energy consumption model, avoiding baseline drift caused by degradation data participating in learning, thereby stabilizing the calculation of degradation thresholds and degradation indices. Fourth, through a residual feedback linkage mechanism, maintenance instructions are generated after degradation is triggered, and the weight of health risk targets is increased or operating constraints are tightened simultaneously, updating the control target parameter set for subsequent cycles. This enables the control strategy and maintenance strategy to form a closed-loop synergy, further improving energy saving and carbon reduction effects and suppressing the accumulation of equipment health risks.

Claims

1. A method for optimizing energy consumption by linking building energy and carbon dual control with predictive maintenance of equipment, comprising: S1. Obtain the building operation data of the target building and perform time alignment and data cleaning to obtain aligned and cleaned building operation data; S2. Generate a state feature sequence for control decisions based on the aligned and cleaned building operation data, and output an equipment health risk sequence through a health assessment model; S3. Based on the state feature sequence, equipment health risk sequence, carbon emission factors, and energy carbon control indicators, establish and solve a health-constrained multi-objective optimization model for energy consumption, carbon emissions, and health risks, generating a control target parameter set, which includes reward / penalty parameters and constraint parameters; S4. Based on the control target parameter set and the state feature sequence, generate a control action sequence through a safety reinforcement learning controller, and verify and correct it according to the constraint parameters to obtain an execution control action sequence. Execute the execution control action sequence and collect actual energy consumption and equipment execution feedback; S5. Based on the state feature sequence and the execution control action sequence, output the expected energy consumption through an action condition expected energy consumption model, and generate a total residual from the actual energy consumption and the expected energy consumption; S6. Generate an execution residual from the execution control action sequence and equipment execution feedback, and generate an efficiency residual based on the total residual and the execution residual; S7. Based on the equipment health risk sequence and execution residuals, screen the health window set, and update the expected energy consumption model of the action conditions using only the data corresponding to the health window set; S8. Determine the degradation judgment threshold based on the health window set, calculate the efficiency degradation index based on the efficiency residual and the degradation judgment threshold, and calculate the additional energy consumption and corresponding additional carbon emissions caused by efficiency degradation. When the efficiency degradation index meets the preset trigger conditions, generate maintenance instructions and update the control target parameter set for subsequent control cycles.

2. The method for optimizing building energy and carbon dual control and predictive maintenance of equipment according to claim 1, S1 includes: Within the preset sampling period, the target building's equipment energy consumption data, equipment operating condition data, equipment control action data, equipment execution feedback data, outdoor environmental data, indoor load data, electricity carbon emission factor data, energy consumption control indicators, and carbon emission control indicators are collected respectively, and corresponding timestamps are generated for each type of data record. Based on the timestamp, time alignment processing is performed on various data records. The time alignment processing includes resampling data with different sampling frequencies according to the preset sampling period, and merging the resampled data records according to the same timestamp. Data cleaning is performed on various data records after time alignment. The data cleaning process includes missing value handling, outlier handling, and numerical consistency processing. The numerical consistency processing includes unit conversion and dimension unification. Output aligned and cleaned building operation data after the time alignment processing and data cleaning processing, wherein the aligned and cleaned building operation data includes equipment energy consumption data, equipment operating condition data, equipment control action data, equipment execution feedback data, outdoor environmental data, indoor load data, electricity carbon emission factor data, energy consumption control indicators, and carbon emission control indicators corresponding to the same timestamp.

3. The method for optimizing building energy and carbon dual control and predictive maintenance of equipment according to claim 1, S2 includes: The system calls up the aligned and cleaned building operation data, combines equipment energy consumption data, equipment operating condition data, equipment control action data, equipment execution feedback data, outdoor environmental data, and indoor load data according to timestamps, and performs numerical normalization and time series processing on the combined data to generate a state feature sequence for control decision-making. Using the energy consumption data, operating condition data, and execution feedback data of the target equipment from the aligned and cleaned building operation data, health characteristics reflecting the energy efficiency deviation of the target equipment are calculated, and the health characteristics are input into the health assessment model to output an equipment health risk sequence that characterizes the risk of efficiency degradation of the target equipment.

4. The method for optimizing building energy and carbon dual control and predictive maintenance of equipment according to claim 1, S3 includes: Based on the state characteristic sequence and equipment health risk sequence, combined with power carbon emission factor data, energy consumption control indicators and carbon emission control indicators, a health-constrained multi-objective optimization model with equipment control actions as decision variables is established, wherein the equipment control actions include setpoint adjustment actions and start-stop scheduling actions. In the health-constrained multi-objective optimization model, the energy consumption target is defined as the optimization target of the energy consumption assessment result of the equipment control action under the corresponding working conditions of the state characteristic sequence; the carbon emission target is defined as the optimization target of the carbon emission assessment result obtained by converting the energy consumption assessment result based on the electricity carbon emission factor data; and the health risk target is defined as the optimization target of the health risk assessment result corresponding to the equipment health risk sequence. Meanwhile, in the health-constrained multi-objective optimization model, energy consumption constraints corresponding to the energy consumption control index, carbon emission constraints corresponding to the carbon emission control index, and equipment operation constraints are set. The equipment operation constraints include amplitude limits, rate of change limits, and start-stop interval limits for equipment control actions. By solving the health-constrained multi-objective optimization model, a set of control objective parameters is generated. The set of control objective parameters includes reward and penalty parameters for the safety reinforcement learning controller to generate control action sequences, and constraint parameters for the feasibility verification and correction of the control action sequences.

5. The method for optimizing building energy and carbon dual control and predictive maintenance of equipment according to claim 1, S4 includes: Based on the set of control target parameters and combined with the state feature sequence, a sequence of control actions is generated by a safety reinforcement learning controller. In the interactive training process, the safety reinforcement learning controller uses the reward and punishment parameters corresponding to the energy consumption target, carbon emission target and health risk target as the reward function parameters, and uses the constraint parameters to limit the output range of the control actions. The feasibility of the control action sequence is verified based on the constraint parameters. When the control action in the control action sequence does not meet the equipment operation constraints corresponding to the constraint parameters, the control action that does not meet the constraints is corrected to generate an execution control action sequence that meets the equipment operation constraints. The execution control action sequence is sent to the target device and executed. During the execution, the actual energy consumption sequence and the device execution feedback sequence corresponding to the execution control action sequence are collected according to the preset sampling period.

6. The method for optimizing building energy and carbon dual control and predictive maintenance of equipment according to claim 1, S5 includes: An action condition feature sequence is constructed using a state feature sequence and an execution control action sequence. The action condition feature sequence is used to characterize the operating conditions of the target device under corresponding state features and corresponding execution control actions. Input the action condition feature sequence into the action condition expected energy consumption model, and output the expected energy consumption sequence corresponding to the execution control action sequence; The total residual sequence is generated by calculating the difference between the actual energy consumption sequence and the expected energy consumption sequence time by time.

7. The method for optimizing building energy and carbon dual control and predictive maintenance of equipment according to claim 1, S6 includes: For each control action in the sequence of execution control actions, an execution feedback value corresponding to the timestamp of the control action is determined in the device execution feedback sequence, and the deviation between the control action and the execution feedback value is calculated to obtain a control deviation sequence that characterizes the degree of deviation of the control action execution. The control deviation is converted into an execution residual sequence based on the control deviation sequence. The execution residual sequence is used to characterize the energy consumption deviation caused by the control action not being executed as expected. An efficiency residual sequence is calculated based on the total residual sequence and the execution residual sequence. The efficiency residual sequence is used to characterize the energy consumption deviation caused by equipment efficiency degradation when the control action is executed as expected.

8. The method for optimizing building energy and carbon dual control and predictive maintenance of equipment according to claim 1, S7 includes: The equipment health risk sequence and execution residual sequence are divided into multiple time windows according to a preset window length; Calculate the equipment health risk statistic and execution residual statistic in each time window, and determine the time window in which the equipment health risk statistic is lower than the first threshold and the execution residual statistic is lower than the second threshold as the health window, thereby generating a set of health windows; The expected energy consumption model under action conditions is updated by using the state feature sequence, execution control action sequence, and actual energy consumption sequence corresponding to the health window set to generate an updated expected energy consumption model under action conditions. The parameter update is performed only based on the data corresponding to the health window set to avoid the expected energy consumption model under action conditions being shifted due to the data corresponding to equipment efficiency degradation participating in the update.

9. The method for optimizing building energy and carbon dual control and predictive maintenance of equipment according to claim 1, S8 includes: Based on the set of healthy windows, the efficiency residual benchmark statistic is calculated in the efficiency residual sequence corresponding to the set of healthy windows, and the degradation judgment threshold is determined based on the efficiency residual benchmark statistic. The efficiency residual sequence is compared with the degradation judgment threshold, and an efficiency degradation index is generated according to a preset calculation rule, wherein the preset calculation rule includes accumulating the efficiency residuals that exceed the degradation judgment threshold and performing time smoothing. The additional energy consumption caused by efficiency degradation is calculated based on the efficiency residual sequence, and the additional energy consumption is converted into additional carbon emissions based on the electricity carbon emission factor data. When the efficiency degradation index meets the preset degradation triggering condition, a maintenance instruction is generated, and in the next control cycle, the weight of the health risk target in the health constraint multi-objective optimization model is increased or the equipment operation constraints are tightened to update the control target parameter set, so that the updated control target parameter set is used in step S4 of the next control cycle.

10. A building energy and carbon dual control and equipment predictive maintenance linkage energy consumption optimization system, used to execute any one of the building energy and carbon dual control and equipment predictive maintenance linkage energy consumption optimization methods according to claims 1 to 9, comprising: The data preprocessing module is used to acquire the target building's operational data and perform time alignment and data cleaning. The module for status and health assessment generates control status characteristics and outputs equipment health risks. The module for generating control target parameters establishes and solves a health-constrained multi-objective optimization model based on control status characteristics, equipment health risks, carbon emission factors, and energy and carbon control indicators, generating a set of control target parameters that includes reward and penalty parameters and constraint parameters. The module for safety reinforcement learning control generates control actions based on the set of control target parameters, verifies and corrects them according to constraint parameters, and then issues them for execution, collecting actual energy consumption and equipment execution feedback. The module for residual analysis and maintenance linkage obtains total residual, execution residual, and efficiency residual based on control status characteristics, executed control actions, actual energy consumption, and equipment execution feedback. It filters health windows based on equipment health risks and execution residuals and updates the expected energy consumption model for action conditions. Based on the health windows, it determines the degradation judgment threshold and generates an efficiency degradation index from the efficiency residual. It calculates the additional energy consumption and corresponding additional carbon emissions caused by efficiency degradation. When the efficiency degradation index meets the triggering conditions, it generates maintenance instructions and updates the set of control target parameters for subsequent control cycles.

Images