A central air conditioning system energy-saving operation method based on multi-device cooperative control

Abstract

The application discloses a kind of central air conditioning system energy-saving operation methods based on multi-device cooperative control, establishes device topology information library and stores as multimodal operation measurement data;According to thermodynamic mechanism equation and energy conservation equation, construct digital twin baseline model;Generate random survival forest input feature set;Output cooperative hazard function;Produce mirror calibration trigger signal;Update digital twin baseline model, obtain mirror calibration after digital twin model;The optimal cooperative energy-saving strategy is decomposed into executable control instruction set, and is issued to the programmable logic controller or distributed control system of physical central air conditioning system execution.The application reduces the time error of device cooperative attenuation prediction, significantly improves the foresight and stability of the model in identifying high-risk operation interval in advance.

Description

Technical Field

[0001] This invention relates to the field of central air conditioning technology, and in particular to an energy-saving operation method for a central air conditioning system based on multi-device collaborative control. Background Technology

[0002] Central air conditioning systems, as core energy-consuming equipment in large public buildings and industrial parks, typically account for over 40% of the building's total energy consumption. With the increasing sophistication of energy management, data-driven and model-driven energy-saving control technologies are gradually being applied to the operation and management of central air conditioning systems. Among these, digital twin technology, capable of constructing high-fidelity mapping models of physical systems in virtual space, is widely used for operational simulation and energy-saving strategy optimization of equipment such as chillers, pumps, and cooling towers.

[0003] Most existing digital twin models for central air conditioning systems are based on factory performance parameters or calibration results from the initial commissioning phase, assuming that the equipment's performance parameters remain constant throughout the operating cycle. In reality, factors such as scale buildup in chiller heat exchangers, compressor wear, pump impeller corrosion, and aging of cooling tower packing can all cause continuous drift in heat transfer coefficients, volumetric efficiency, and system hydraulic characteristics. Current technologies typically assess equipment performance through periodic manual testing or simple threshold alarms, lacking a continuous quantification mechanism for the energy efficiency degradation process and failing to effectively handle censored data that, while not yet showing significant degradation, is already in a potentially risky state. Therefore, as operating time increases, a systematic deviation gradually develops between the energy consumption predicted by the digital twin model and the actual energy consumption, leading to distortion of the control strategy generated based on the model, resulting in decreased energy-saving effects or even operational instability.

[0004] In the collaborative control of multiple devices in central air conditioning systems, existing energy-saving control methods mostly employ static optimization algorithms or greedy strategies based on the current operating conditions, aiming to minimize the total power consumption of the system at the current moment. Because central air conditioning systems are complex systems with strong coupling, nonlinearity, and significant hysteresis, there are clear linkages between chiller load distribution, pump frequency regulation, and cooling tower fan speed settings. Traditional control methods only focus on instantaneous power indicators, lacking forward-looking constraints on the long-term energy efficiency degradation trend of equipment. This can easily lead to a critical device operating at high load for extended periods, accelerating its energy efficiency degradation process. This short-term optimal but long-term unbalanced control approach causes a significant decrease in the overall system energy efficiency in subsequent operating cycles. Summary of the Invention

[0005] One objective of this invention is to propose an energy-saving operation method for a central air conditioning system based on multi-device collaborative control. This invention reduces the time error in predicting equipment collaborative attenuation and significantly improves the model's foresight and stability in identifying high-risk operating intervals in advance.

[0006] An energy-saving operation method for a central air conditioning system based on multi-device collaborative control according to an embodiment of the present invention includes: Establish a device topology information database and store it as multimodal operation measurement data; Based on multimodal operation measurement data and equipment topology information, a digital twin baseline model is constructed according to the thermodynamic mechanism equation and the energy conservation equation; The multimodal operation measurement data is processed by time series structuring, and multi-scale embedding features are extracted through a spatiotemporal representation learning network to generate a random survival forest input feature set; Based on the operating condition where energy efficiency drops to a preset performance threshold, the energy efficiency degradation event of the equipment is marked, and the improved random survival forest algorithm is used to train the equipment energy efficiency survival probability prediction model, and the collaborative hazard function is output. The equipment energy efficiency survival probability prediction model is used to infer the latest multimodal operation measurement data, continuously generate equipment cooperative hazard function sequences and construct dynamic cooperative hazard function band index sequences. The dynamic cooperative hazard function band index sequences are compared with preset dynamic cooperative hazard function band thresholds. When any equipment index exceeds the threshold or shows a monotonically increasing trend within the sliding time window, a mirror calibration trigger signal is generated. After receiving the mirror calibration trigger signal, the physical-virtual residual vector between the energy consumption predicted by the digital twin baseline model and the actual energy consumption of the physical central air conditioning system is calculated, the digital twin baseline model is updated, and the digital twin model after mirror calibration is obtained. The mirror-calibrated digital twin model is imported into the risk perception model predictive control optimization module. Global collaborative optimization of multiple central air conditioning devices is performed in the prediction time domain to obtain the optimal collaborative energy-saving strategy. The optimal collaborative energy-saving strategy is decomposed into an executable control instruction set and sent to the programmable logic controller or distributed control system of the physical central air conditioning system for execution.

[0007] Optionally, establishing a device topology information database and storing it as multimodal operation measurement data includes: Construct a multi-modal operation measurement data acquisition system for central air conditioning systems to acquire in real time temperature, pressure, flow, power, valve opening parameters, external meteorological parameters, and building load parameters of chillers, chilled water pumps, cooling water pumps, and cooling towers.

[0008] Optionally, the construction of the digital twin baseline model according to the thermodynamic mechanism equation and the energy conservation equation includes: Based on the equipment topology information database, the connection relationships and loop directions between the chiller unit, chilled water pump, cooling water pump and cooling tower are determined, the set of equipment nodes and the set of connection edges are constructed, and the system topology diagram is formed. An energy conservation sub-model on the evaporator side of the chiller unit is constructed based on the system topology diagram as the cooling output channel of the digital twin baseline model; An energy conservation sub-model on the cooling water side is constructed based on the system topology diagram as the heat emission channel of the digital twin baseline model; A power sub-model of pump equipment is constructed based on the system topology diagram as the hydraulic execution channel of the digital twin baseline model; A cooling tower heat exchange sub-model is constructed based on the system topology diagram as an environmental coupling channel for the digital twin baseline model; The cooling output channel, heat discharge channel, hydraulic actuation channel, and environmental coupling channel are coupled according to the system topology diagram to form a set of state equations for the digital twin baseline model. The initial parameters of the set of state equations are calibrated based on the set of input variables to obtain the digital twin baseline model.

[0009] Optionally, the step of performing time-series structuring processing on the multimodal operation measurement data and extracting multi-scale embedding features through a spatiotemporal representation learning network includes: A unified time step is determined based on multimodal operation measurement data, and the multimodal operation measurement data is time-aligned and resampled according to the unified time step to obtain the set of input variables corresponding to the time index; Organize the set of input variables into column vectors of dimension d, and form a matrix of input variable sets; The input variable set matrix is ​​subjected to time series structuring processing. The length of the structuring window is set and a window sequence tensor with the time index as the endpoint is constructed. Perform normalization mapping on the window sequence tensor to generate a normalized window sequence tensor; Based on the standardized window sequence tensor, a standardized differential window sequence tensor is constructed to characterize the trend of changes in the operating state. The system topology graph is represented as a topological adjacency matrix; The standardized window sequence tensor, the standardized difference window sequence tensor, and the topological adjacency matrix are input into the spatiotemporal representation learning network. The spatiotemporal representation learning network performs temporal modeling and topological constraint aggregation on the input and extracts the embedded feature vector. The embedded feature vectors of each time scale branch are concatenated to form a multi-scale embedded feature vector; Constructing a random survival forest input feature set based on multi-scale embedded feature vectors; The input feature set of the random survival forest consists of multi-scale embedded feature vectors corresponding to each time index and their one-to-one correspondence with the time index of the time series of the input variable set. The input feature set of the random survival forest is then output to the device energy efficiency survival probability prediction model.

[0010] Optionally, the step of marking equipment energy efficiency degradation events based on the operating condition where energy efficiency drops to a preset performance threshold, and training an improved random survival forest algorithm to obtain an equipment energy efficiency survival probability prediction model, includes: Based on the digital twin baseline model, the power sensitivity of the power of any device a to the control setpoint of any device b is calculated. The power sensitivity is proportionally converted to the control setpoint of device b and the power of device a to obtain the power coupling coefficient between device a and device b at the same time step. The power coupling coefficients between all pairs of devices constitute the device cooperative coupling strength matrix. Based on the equipment cooperative coupling strength matrix, the absolute values ​​of the power coupling coefficients between equipment j and all other equipment are summed and a constant 1 is added to obtain the cooperative risk amplification coefficient of equipment j. Based on the digital twin baseline model, the power of device j is divided by the instantaneous cooling capacity of the chiller evaporator to obtain the energy consumption index per unit cooling capacity. Based on the reference unit cooling energy consumption index obtained by device j during the digital twin baseline model calibration stage, the relative deviation between the unit cooling energy consumption index and the reference unit cooling energy consumption index is calculated, and the relative deviation is multiplied by the collaborative risk amplification coefficient of device j to obtain the collaborative weighted energy efficiency deviation. When the deviation of the collaborative weighted energy efficiency is greater than or equal to the threshold of the collaborative weighted energy efficiency decay event, sample i is marked as the collaborative energy efficiency decay event of device j, and the collaborative decay event indicator variable of the corresponding sample is assigned a value of one; otherwise, the collaborative decay event indicator variable is assigned a value of zero. When constructing training samples for a random survival forest, the sampling probability of the samples is calculated based on the co-decay event indicator variable. Based on the sampling probability, the input feature set of the random survival forest is stratified and sampled to obtain a training sample subset. The training sample subset is then used to train the collaborative weighted random survival forest model, thus forming the collaborative weighted random survival forest model. After training, the cooperative weighted random survival forest model outputs a cooperative survival function and a cooperative hazard function after inputting multi-scale embedded feature vectors.

[0011] Optionally, training the collaboratively weighted random survival forest model using a subset of training samples includes: During the node splitting process of each tree model, the samples belonging to the current node are weighted according to the collaborative risk amplification coefficient of device j, and the difference between the number of samples that actually experience collaborative decay events and the expected number of events under the assumption of no decay is weighted and accumulated to obtain the collaborative risk weighted log-rank statistic. The feature that maximizes the absolute value of the collaborative risk weighted log-rank statistic is selected as the optimal splitting feature to form a collaborative weighted random survival forest model.

[0012] Optionally, the step of calculating the physical-virtual residual vector between the energy consumption predicted by the digital twin baseline model and the actual energy consumption of the physical central air conditioning system, and updating the digital twin baseline model, includes: After receiving the mirror calibration trigger signal, the actual energy consumption observation vector of the physical central air conditioning system is read from the multimodal operation measurement data under a unified time step. At the same time, the digital twin baseline model is called to calculate the predicted energy consumption vector under the same time index and the same set of input variables. By subtracting the actual energy consumption observation vector from the predicted energy consumption vector, the physical-virtual residual vector is obtained, and the physical-virtual residual vectors corresponding to all time indices within the calibration interval are stacked in chronological order to form a residual matrix; The parameters used to characterize equipment performance degradation and heat transfer deviation in the digital twin baseline model are arranged in a fixed order to form a thermodynamic parameter vector; The sum of squares of the physical-virtual residual vectors corresponding to each time index in the residual matrix is ​​obtained by summing the squares of the residuals for the corresponding time index. The sum of squares of the residuals for all time indices within the calibration interval is accumulated to obtain the vectorized parameter inversion objective function. In the digital twin baseline model, a small perturbation is applied to each parameter in the thermodynamic parameter vector, the change in each component of the predicted energy consumption vector is recorded, and the ratio of the change to the corresponding small perturbation is used as the partial derivative approximation of the predicted energy consumption vector with respect to the thermodynamic parameter vector. All partial derivative approximations are arranged in the order of components to form the Jacobian matrix. The thermodynamic parameter increment vector is calculated based on the residual matrix and the Jacobian matrix, and the thermodynamic parameter vector is updated to obtain the calibration parameter vector; The calibration parameter vector is written back to the cooling output channel, heat discharge channel, hydraulic execution channel, and environmental coupling channel of the digital twin baseline model. This updates the evaporator-side equivalent heat transfer coefficient, condenser-side equivalent heat transfer coefficient, chilled water pump efficiency coefficient, cooling water pump efficiency coefficient, and cooling tower heat transfer correction coefficient used in the calculations within each channel to the corresponding components in the calibration parameter vector, thus obtaining the mirror-calibrated digital twin model.

[0013] Optionally, the step of importing the mirror-calibrated digital twin model into the risk perception model predictive control optimization module to perform global collaborative optimization of multiple central air conditioning devices in the prediction time domain includes: In the prediction time domain, the total power consumption of the system is calculated based on the mirror-calibrated digital twin model; Based on the collaborative hazard function output by the collaborative weighted random survival forest model, the energy efficiency life loss cost of equipment is calculated in the prediction time domain. By weighted summing and accumulating the total system power consumption and equipment energy efficiency lifespan depreciation cost at each prediction time point in the prediction time domain, the objective function for predictive control optimization of the risk perception model is obtained. During the optimization process, operational constraints are set, including cooling balance constraints and equipment operating boundary constraints. Under the conditions of satisfying the cooling balance constraint and the equipment operation boundary constraint, the global collaborative optimization of the control variable sequence in the prediction time domain is carried out. By simulating different control variable sequences in the digital twin model after mirror calibration, the corresponding risk perception model predictive control optimization objective function value is calculated, and the control variable sequence that minimizes the risk perception model predictive control optimization objective function value is selected as the optimal collaborative energy saving strategy. The optimal collaborative energy-saving strategy is decomposed into a set of executable control instructions, which are then sent to the programmable logic controller or distributed control system of the physical central air conditioning system for execution via a communication interface.

[0014] The beneficial effects of this invention are: This invention extends the original survival analysis algorithm for single devices to a collaboratively weighted random survival forest model for multi-device strongly coupled systems by introducing a device collaborative coupling strength matrix and a collaborative risk amplification coefficient into the training process of the improved random survival forest. This allows the device energy efficiency degradation prediction to not only reflect the performance degradation of a single device but also quantify its linkage amplification effect on the total power consumption of the system. The collaborative risk weighted log-rank statistic is introduced into the node splitting criterion, and a collaborative risk stratified sampling mechanism is adopted in the sample sampling stage, so that high collaborative risk samples occupy a higher weight in forest training, which enhances the model's ability to identify key degradation patterns at the system level. Compared with the traditional random survival forest model, this invention reduces the time error of device collaborative degradation prediction and significantly improves the model's foresight and stability in early identification of high-risk operating intervals.

[0015] In the digital twin model calibration stage, this invention employs a vectorized parameter inversion algorithm based on the residual matrix and the Jacobian matrix to jointly update the thermodynamic parameter vectors. This enables synchronous correction of the heat transfer coefficients on the evaporator side, the heat transfer coefficients on the condenser side, the pump efficiency coefficients, and the cooling tower heat transfer correction coefficients. The physical-virtual residual vectors are uniformly mapped to the power dimension space. By constructing a multi-time residual sum of squares objective function and performing vectorized updates based on the Jacobian matrix, each sub-channel satisfies the power conservation and heat balance constraints at the same time step. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an energy-saving operation method for a central air conditioning system based on multi-device collaborative control, as proposed in this invention. Detailed Implementation

[0017] Example 1: Reference Figure 1 An energy-saving operation method for a central air conditioning system based on multi-device collaborative control includes: Establish a device topology information database and store it as multimodal operation measurement data; In this embodiment, a device topology information database is established and stored as multimodal operation measurement data, including: Construct a multi-modal operation measurement data acquisition system for central air conditioning systems to acquire in real time temperature, pressure, flow, power, valve opening parameters, external meteorological parameters, and building load parameters of chillers, chilled water pumps, cooling water pumps, and cooling towers.

[0018] Based on multimodal operation measurement data and equipment topology information, a digital twin baseline model is constructed according to the thermodynamic mechanism equation and the energy conservation equation; In this embodiment, a digital twin baseline model is constructed according to the thermodynamic mechanism equation and the energy conservation equation, including: Based on the equipment topology information database, the connection relationships and loop directions between the chiller unit, chilled water pump, cooling water pump and cooling tower are determined, the set of equipment nodes and the set of connection edges are constructed, and the system topology diagram is formed. The system topology diagram is used to constrain the transmission paths of chilled water loop variables, cooling water loop variables, and cooling tower heat exchange variables in the input variable set, so that the chilled water loop variables, cooling water loop variables, and cooling tower heat exchange variables are transmitted in a consistent manner according to the flow direction specified by the connection edge set.

[0019] An energy conservation sub-model on the evaporator side of the chiller unit is constructed based on the system topology diagram as the cooling output channel of the digital twin baseline model; In Example 1, the cooling capacity output channel reads the chilled water mass flow rate, the chiller evaporator inlet temperature, and the chiller evaporator outlet temperature from the input variable set. Using the specific heat capacity of water at constant pressure as a conversion factor, the chilled water mass flow rate is multiplied by the temperature difference between the chiller evaporator inlet temperature and the chiller evaporator outlet temperature to obtain the instantaneous cooling capacity of the chiller evaporator. The instantaneous cooling capacity of the chiller evaporator is used to represent the amount of heat removed from the chilled water by the chiller evaporator per unit time.

[0020] An energy conservation sub-model on the cooling water side is constructed based on the system topology diagram as the heat emission channel of the digital twin baseline model; In Example 1, the heat discharge channel reads the cooling water mass flow rate, cooling water inlet temperature, and cooling water outlet temperature from the input variable set, and uses the specific heat capacity of water at constant pressure as a conversion factor to multiply the cooling water mass flow rate with the temperature difference between the cooling water outlet temperature and the cooling water inlet temperature to obtain the instantaneous heat discharge on the cooling water side. The instantaneous heat discharge on the cooling water side is used to represent the amount of heat absorbed by the cooling water from the chiller unit and carried to the cooling tower per unit time.

[0021] A power sub-model of pump equipment is constructed based on the system topology diagram as the hydraulic execution channel of the digital twin baseline model; In Example 1, the hydraulic actuation channel obtains the hydraulic power of the chilled water pump by dividing the product of the inlet and outlet pressure difference of the chilled water pump and the volumetric flow rate of the chilled water by the overall efficiency of the chilled water pump, and obtains the hydraulic power of the cooling water pump by dividing the product of the inlet and outlet pressure difference of the cooling water pump and the volumetric flow rate of the cooling water by the overall efficiency of the cooling water pump. The hydraulic power of the chilled water pump and the hydraulic power of the cooling water pump correspond to the electrical power of the chilled water pump and the electrical power of the cooling water pump under the same power dimension to form a power consistency constraint for pump equipment.

[0022] A cooling tower heat exchange sub-model is constructed based on the system topology diagram as an environmental coupling channel for the digital twin baseline model; In Example 1, the environmental coupling channel calculates the heat exchange of the cooling tower to the cooling water by multiplying the cooling water mass flow rate, the water's specific heat capacity at constant pressure, and the temperature difference between the cooling water outlet temperature and the cooling water inlet temperature. It then establishes an input mapping relationship between the heat exchange of the cooling tower to the cooling water and the outdoor dry-bulb temperature and relative humidity, so that the digital twin baseline model outputs the corresponding cooling water temperature response when the outdoor dry-bulb temperature and relative humidity change.

[0023] The cooling output channel, heat discharge channel, hydraulic actuation channel, and environmental coupling channel are coupled according to the system topology diagram to form a set of state equations for the digital twin baseline model. The initial parameters of the set of state equations are calibrated based on the set of input variables to obtain the digital twin baseline model.

[0024] In Example 1, the cooling output channel, heat discharge channel, hydraulic execution channel and environmental coupling channel are mapped according to the equipment connection relationship and flow direction relationship defined in the system topology diagram. This enables the instantaneous cooling capacity on the evaporator side of the chiller unit, the instantaneous heat dissipation on the cooling water side, the hydraulic power of pump equipment and the heat exchange of cooling tower to cooling water to establish an algebraic coupling relationship and a power conservation constraint relationship under a unified time step.

[0025] Based on the algebraic coupling relationship and the power conservation constraint, a system of simultaneous equations containing temperature, flow and power state variables is constructed. The set of input variables is substituted into the system of simultaneous equations, and the state variable solution that satisfies the heat conservation constraint and the power consistency constraint is calculated through iterative solution.

[0026] The state variable solution is compared with the multimodal operation measurement data for error. The parameters of heat transfer efficiency coefficient, pump equipment efficiency coefficient and cooling tower heat transfer correction coefficient are updated by minimizing the sum of squared errors between the predicted value and the measured value. When the sum of squared errors is lower than the preset convergence threshold, the current parameter set and the corresponding simultaneous equations are fixed as the digital twin baseline model.

[0027] The multimodal operation measurement data is processed by time series structuring, and multi-scale embedding features are extracted through a spatiotemporal representation learning network to generate a random survival forest input feature set; In this embodiment, the multimodal operation measurement data undergoes time-series structuring processing, and multi-scale embedding features are extracted through a spatiotemporal representation learning network, including: A unified time step is determined based on multimodal operation measurement data, and the multimodal operation measurement data is time-aligned and resampled according to the unified time step to obtain the set of input variables corresponding to the time index; The time index represents the time index after being discretized at a uniform time step, and the set of input variables represents the set of input variables consisting of multimodal operation measurement data at the time index.

[0028] Organize the set of input variables into column vectors of dimension d, and form a matrix of input variable sets; The input variable set matrix is ​​subjected to time series structuring processing. The length of the structuring window is set and a window sequence tensor with the time index as the endpoint is constructed. Perform normalization mapping on the window sequence tensor to generate a normalized window sequence tensor; Based on the standardized window sequence tensor, a standardized differential window sequence tensor is constructed to characterize the trend of changes in the operating state. The standardized difference window sequence tensor representation characterizes the trend of system state changes over time by calculating the standardized difference between adjacent time steps. The difference calculation is the standardized value of the current time step minus the standardized value of the previous time step.

[0029] The system topology graph is represented as a topological adjacency matrix; The topological adjacency matrix represents the connection relationships between device nodes in the set of device nodes. The elements of the topological adjacency matrix indicate whether there is a connection between nodes. The topological adjacency matrix serves as the structural constraint input for the spatiotemporal representation learning network.

[0030] The standardized window sequence tensor, the standardized difference window sequence tensor, and the topological adjacency matrix are input into the spatiotemporal representation learning network. The spatiotemporal representation learning network performs temporal modeling and topological constraint aggregation on the input and extracts the embedded feature vector. The embedded feature vectors of each time scale branch are concatenated to form a multi-scale embedded feature vector; Constructing a random survival forest input feature set based on multi-scale embedded feature vectors; The input feature set of the random survival forest consists of multi-scale embedded feature vectors corresponding to each time index and their one-to-one correspondence with the time index of the time series of the input variable set. The input feature set of the random survival forest is then output to the device energy efficiency survival probability prediction model.

[0031] Based on the operating condition where energy efficiency drops to a preset performance threshold, the energy efficiency degradation event of the equipment is marked, and the improved random survival forest algorithm is used to train the equipment energy efficiency survival probability prediction model, and the collaborative hazard function is output. In this embodiment, the energy efficiency degradation event of the equipment is marked according to the operating condition where the energy efficiency drops to a preset performance threshold, and an improved random survival forest algorithm is used to train a prediction model for the equipment energy efficiency survival probability, including: Based on the digital twin baseline model, the power sensitivity of the power of any device a to the control setpoint of any device b is calculated. The power sensitivity is proportionally converted to the control setpoint of device b and the power of device a to obtain the power coupling coefficient between device a and device b at the same time step. The power coupling coefficients between all pairs of devices constitute the device cooperative coupling strength matrix. In Example 1, the power sensitivity is obtained by applying a small perturbation to the control setpoint of device b in the digital twin baseline model and calculating the ratio of the change in electrical power of device a to the small perturbation. The device cooperative coupling strength matrix is ​​used to quantify the seesaw effect between multiple devices in the central air conditioning system.

[0032] Based on the equipment cooperative coupling strength matrix, the absolute values ​​of the power coupling coefficients between equipment j and all other equipment are summed and a constant 1 is added to obtain the cooperative risk amplification coefficient of equipment j. In Example 1, the collaborative risk amplification coefficient of device j is used to represent the degree of amplification of the energy consumption of other devices caused by the energy efficiency degradation of device j.

[0033] Based on the digital twin baseline model, the power of device j is divided by the instantaneous cooling capacity of the chiller evaporator to obtain the energy consumption index per unit cooling capacity. Based on the reference unit cooling energy consumption index obtained by device j during the digital twin baseline model calibration stage, the relative deviation between the unit cooling energy consumption index and the reference unit cooling energy consumption index is calculated, and the relative deviation is multiplied by the collaborative risk amplification coefficient of device j to obtain the collaborative weighted energy efficiency deviation. When the deviation of the collaborative weighted energy efficiency is greater than or equal to the threshold of the collaborative weighted energy efficiency decay event, sample i is marked as the collaborative energy efficiency decay event of device j, and the collaborative decay event indicator variable of the corresponding sample is assigned a value of one; otherwise, the collaborative decay event indicator variable is assigned a value of zero. When constructing training samples for a random survival forest, the sampling probability of the samples is calculated based on the co-decay event indicator variable. In Example 1, the sampling probability of a sample is obtained by weighting and summing the indicator variables of the collaborative decay events on all devices for each sample with the collaborative risk amplification coefficient of the corresponding device, and then dividing the weighted sum by the corresponding weighted sum of all samples. By increasing the sampling probability of high collaborative risk samples, the collaborative weighted random survival forest model pays more attention to decay patterns that have a significant impact on the collaborative energy consumption of the system.

[0034] Based on the sampling probability, the input feature set of the random survival forest is stratified and sampled to obtain a training sample subset. The training sample subset is then used to train the collaborative weighted random survival forest model, thus forming the collaborative weighted random survival forest model. In this embodiment, a collaboratively weighted random survival forest model is trained using a subset of training samples, including: During the node splitting process of each tree model, the samples belonging to the current node are weighted according to the collaborative risk amplification coefficient of device j, and the difference between the number of samples that actually experience collaborative decay events and the expected number of events under the assumption of no decay is weighted and accumulated to obtain the collaborative risk weighted log-rank statistic. The feature that maximizes the absolute value of the collaborative risk weighted log-rank statistic is selected as the optimal splitting feature to form a collaborative weighted random survival forest model.

[0035] After training, the cooperative weighted random survival forest model outputs a cooperative survival function and a cooperative hazard function after inputting multi-scale embedded feature vectors. In Example 1, after training, for any input multi-scale embedding feature vector, the multi-scale embedding feature vector is input into each tree model in the cooperative weighted random survival forest model, so that the multi-scale embedding feature vector sinks to the corresponding leaf node along the splitting path of each tree model. In each leaf node, the cooperative decay event indicator variable and the corresponding survival time label of the samples contained in the leaf node are counted, and the local cooperative survival function of the leaf node is constructed based on the survival time distribution of the samples in the leaf node. The local cooperative survival functions of the corresponding leaf nodes of all tree models are averaged to obtain the cooperative survival function. The cooperative survival function represents the predicted probability that the device will maintain a high energy efficiency state after considering the cooperative coupling effect under the current operating characteristics.

[0036] The cooperative hazard function is calculated based on the discrete difference rate of change of the cooperative survival function in adjacent time intervals. The cooperative hazard function represents the instantaneous risk intensity of the equipment entering the cooperative decay state per unit time.

[0037] The remaining energy efficiency lifetime is calculated based on the integral result of the cooperative survival function on the time axis. The remaining energy efficiency lifetime represents the expected remaining time for device j to maintain high energy efficiency operation under the condition of considering the amplification effect of multi-device cooperative coupling.

[0038] The equipment energy efficiency survival probability prediction model is used to infer the latest multimodal operation measurement data, continuously generate equipment cooperative hazard function sequences and construct dynamic cooperative hazard function band index sequences. The dynamic cooperative hazard function band index sequences are compared with preset dynamic cooperative hazard function band thresholds. When any equipment index exceeds the threshold or shows a monotonically increasing trend within the sliding time window, a mirror calibration trigger signal is generated. After receiving the mirror calibration trigger signal, the physical-virtual residual vector between the energy consumption predicted by the digital twin baseline model and the actual energy consumption of the physical central air conditioning system is calculated, the digital twin baseline model is updated, and the digital twin model after mirror calibration is obtained. In this embodiment, the physical-virtual residual vector between the energy consumption predicted by the digital twin baseline model and the actual energy consumption of the physical central air conditioning system is calculated, and the digital twin baseline model is updated, including: After receiving the mirror calibration trigger signal, the actual energy consumption observation vector of the physical central air conditioning system is read from the multimodal operation measurement data under a unified time step. At the same time, the digital twin baseline model is called to calculate the predicted energy consumption vector under the same time index and the same set of input variables. In Example 1, when constructing the actual energy consumption observation vector, the power of the chiller unit, the power of the chilled water pump, the power of the cooling water pump, the power of the cooling tower fan, the instantaneous cooling capacity of the chiller unit evaporator, and the instantaneous heat dissipation of the cooling water in the physical central air conditioning system are arranged in a fixed order to form a vector. The instantaneous cooling capacity of the chiller unit evaporator is obtained by reading the chilled water mass flow rate, the chiller unit evaporator inlet temperature, and the chiller unit evaporator outlet temperature and converting them by multiplying the mass flow rate with the temperature difference. The instantaneous heat dissipation of the cooling water is obtained by reading the cooling water mass flow rate, the cooling water inlet temperature, and the cooling water outlet temperature and converting them by multiplying the mass flow rate with the temperature difference. This makes the actual energy consumption observation vector structurally completely consistent with the output structure of the digital twin baseline model.

[0039] When constructing the predicted energy consumption vector of the digital twin baseline model, the power output of the chiller unit, chilled water pump, cooling water pump, cooling tower fan, instantaneous cooling capacity of the chiller unit evaporator side, and instantaneous heat dissipation of the cooling water side, which are output by the digital twin baseline model under the same set of input variables, are arranged in the same order as the actual energy consumption observation vector to form the predicted energy consumption vector. This ensures that the predicted energy consumption vector and the actual energy consumption observation vector are completely consistent in terms of component meaning, arrangement order, and physical dimensions.

[0040] By subtracting the actual energy consumption observation vector from the predicted energy consumption vector, the physical-virtual residual vector is obtained, and the physical-virtual residual vectors corresponding to all time indices within the calibration interval are stacked in chronological order to form a residual matrix; In Example 1, each component of the physical-virtual residual vector represents the difference between the predicted value and the actual value of the corresponding physical device at the same time index.

[0041] The parameters used to characterize equipment performance degradation and heat transfer deviation in the digital twin baseline model are arranged in a fixed order to form a thermodynamic parameter vector; The thermodynamic parameter vector includes the equivalent heat transfer coefficient on the evaporator side, the equivalent heat transfer coefficient on the condenser side, the efficiency coefficient of the chilled water pump, the efficiency coefficient of the cooling water pump, and the heat transfer correction coefficient of the cooling tower. Each parameter in the thermodynamic parameter vector is used to correct the cooling output channel, the heat discharge channel, the hydraulic actuation channel, and the environmental coupling channel, so that the digital twin baseline model can reflect the performance degradation of the physical central air conditioning system through parameter changes.

[0042] The sum of squares of the physical-virtual residual vectors corresponding to each time index in the residual matrix is ​​obtained by summing the squares of the residuals for the corresponding time index. The sum of squares of the residuals for all time indices within the calibration interval is accumulated to obtain the vectorized parameter inversion objective function. In the digital twin baseline model, a small perturbation is applied to each parameter in the thermodynamic parameter vector, the change in each component of the predicted energy consumption vector is recorded, and the ratio of the change to the corresponding small perturbation is used as the partial derivative approximation of the predicted energy consumption vector with respect to the thermodynamic parameter vector. All partial derivative approximations are arranged in the order of components to form the Jacobian matrix. The thermodynamic parameter increment vector is calculated based on the residual matrix and the Jacobian matrix, and the thermodynamic parameter vector is updated to obtain the calibration parameter vector; The calibration parameter vector is written back to the cooling output channel, heat discharge channel, hydraulic execution channel, and environmental coupling channel of the digital twin baseline model. This updates the evaporator-side equivalent heat transfer coefficient, condenser-side equivalent heat transfer coefficient, chilled water pump efficiency coefficient, cooling water pump efficiency coefficient, and cooling tower heat transfer correction coefficient used in the calculations within each channel to the corresponding components in the calibration parameter vector, thus obtaining the mirror-calibrated digital twin model.

[0043] The mirror-calibrated digital twin model is imported into the risk perception model predictive control optimization module. Global collaborative optimization of multiple central air conditioning devices is performed in the prediction time domain to obtain the optimal collaborative energy-saving strategy. The optimal collaborative energy-saving strategy is decomposed into an executable control instruction set and sent to the programmable logic controller or distributed control system of the physical central air conditioning system for execution.

[0044] In this embodiment, the mirror-calibrated digital twin model is imported into the risk perception model predictive control optimization module to perform global collaborative optimization of multiple central air conditioning devices in the prediction time domain, including: In the prediction time domain, the total power consumption of the system is calculated based on the mirror-calibrated digital twin model; The total power consumption of the system is obtained by summing the power of the chiller unit, the chilled water pump, the cooling water pump, and the cooling tower fan output from the mirror-calibrated digital twin model under the same power dimension.

[0045] Based on the collaborative hazard function output by the collaborative weighted random survival forest model, the energy efficiency life loss cost of equipment is calculated in the prediction time domain. In Example 1, the equipment energy efficiency life loss cost is obtained by multiplying the collaborative hazard function value of each device at the current predicted time point with the life loss weight coefficient of that device, and then summing the product results of all devices.

[0046] By weighted summing and accumulating the total system power consumption and equipment energy efficiency lifespan depreciation cost at each prediction time point in the prediction time domain, the objective function for predictive control optimization of the risk perception model is obtained. During the optimization process, operational constraints are set, including cooling balance constraints and equipment operating boundary constraints. In Example 1, the cooling balance constraint is obtained by comparing the instantaneous cooling capacity of the chiller evaporator side output by the mirror-calibrated digital twin model with the building terminal cooling load demand. When the instantaneous cooling capacity of the chiller evaporator side is less than the building terminal cooling load demand, it is determined to be a violation of the cooling balance constraint.

[0047] Equipment operating boundary constraints are obtained by determining whether the chiller unit's supply water temperature setpoint is between the minimum and maximum allowable values, whether the chilled water pump frequency is between the minimum and maximum allowable values, whether the cooling water pump frequency is between the minimum and maximum allowable values, and whether the cooling tower fan speed setpoint is between the minimum and maximum allowable values.

[0048] Under the conditions of satisfying the cooling balance constraint and the equipment operation boundary constraint, the global collaborative optimization of the control variable sequence in the prediction time domain is carried out. By simulating different control variable sequences in the digital twin model after mirror calibration, the corresponding risk perception model predictive control optimization objective function value is calculated, and the control variable sequence that minimizes the risk perception model predictive control optimization objective function value is selected as the optimal collaborative energy saving strategy. The optimal collaborative energy-saving strategy is decomposed into a set of executable control instructions, which are then sent to the programmable logic controller or distributed control system of the physical central air conditioning system for execution via a communication interface.

[0049] In Example 1, the executable control instruction set includes chiller unit operation setpoint instructions, chilled water pump frequency conversion instructions, cooling water pump frequency conversion instructions, cooling tower fan control instructions, and valve opening adjustment instructions. The chiller unit operation setpoint instructions are generated by writing the chiller unit supply water temperature setpoint from the optimal collaborative energy-saving strategy into the chiller unit control register. The chilled water pump frequency conversion instructions are generated by writing the chilled water pump frequency from the optimal collaborative energy-saving strategy into the chilled water pump frequency converter frequency register. The cooling water pump frequency conversion instructions are generated by writing the cooling water pump frequency from the optimal collaborative energy-saving strategy into the cooling water pump frequency converter frequency register. The cooling tower fan control instructions are generated by writing the cooling tower fan speed setpoint from the optimal collaborative energy-saving strategy into the cooling tower fan control register and generating start-stop logic based on the speed threshold. The valve opening adjustment instructions are generated by converting the hydraulic distribution result calculated by the mirror-calibrated digital twin model under the optimal collaborative energy-saving strategy into the valve opening write value.

[0050] Example 2

[0051] During the continuous operation of the central air conditioning system in a large commercial building, the system continuously collects multimodal operation measurement data at a uniform time step. Within a certain continuous operation cycle, the system recorded that the electrical power of the chiller unit CH-02 showed a slow upward trend in multiple time windows, while the temperature difference between the inlet and outlet water on the cooling water side decreased.

[0052] During the data acquisition phase, the system continuously records the following operational samples: Within a continuous 5-minute time step, the chilled water mass flow rate is 185 kg / s, the inlet water temperature on the evaporator side of the chiller is 285.4 K, and the outlet water temperature on the evaporator side of the chiller is 279.8 K. Based on the cooling capacity output channel, the instantaneous cooling capacity on the evaporator side of the chiller is calculated to be 4366 kW; at the same time, the electrical power of the chiller is 712 kW.

[0053] During the subsequent 30-minute time window, the cooling water mass flow rate was 210 kg / s, the cooling water inlet temperature was 302.1 K, and the cooling water outlet temperature was 306.8 K. Based on the heat discharge channel, the instantaneous heat dissipation on the cooling water side was calculated to be 4128 kW.

[0054] The digital twin baseline model predicts the power of the CH-02 chiller unit to be 668kW, the power of the cooling water pump to be 198kW, and the power of the cooling tower fan to be 156kW under the same set of input variables. The actual measured values ​​of the physical central air conditioning system are 712kW, 213kW, and 168kW, respectively.

[0055] The system constructs a time series window, forming a sequence sample tensor of length 6, and extracts multi-scale embedding feature vectors. For this operating interval, a total of 864 sets of multi-scale embedding feature samples are generated.

[0056] During the collaborative risk modeling phase, the system applied a small perturbation (1%) to the cooling tower fan speed setpoint based on the digital twin baseline model, adjusting the CH-02's electrical power. The calculated change in CH-02's electrical power was 5.6 kW, with a power sensitivity of 5.6 kW / (1% of the setpoint). Combining the current cooling tower fan setpoint of 0.82 (dimensionless) with CH-02's electrical power of 712 kW, a power coupling coefficient of 0.0064 was obtained through proportional conversion.

[0057] Using the same method, the power coupling coefficient between CH-02 and the cooling water pump was calculated to be 0.0091, and the coefficient between CH-02 and the chilled water pump was 0.0048. The system then summed these power coupling coefficients pairwise to form a matrix of equipment cooperative coupling strength.

[0058] For CH-02, the system sums the absolute values ​​of the power coupling coefficients with other devices and adds one to obtain a collaborative risk amplification factor of 1.0203.

[0059] The system calculates the energy consumption per unit of cooling capacity. At the current moment, the electrical power of CH-02 is 712kW, and the instantaneous cooling capacity on the evaporator side of the chiller unit is 4366kW. Therefore, the energy consumption per unit of cooling capacity is 0.163. The reference energy consumption per unit of cooling capacity during the digital twin baseline model calibration stage is 0.148, with a relative deviation of 10.1%. Multiplying the relative deviation by the collaborative risk amplification factor, the collaborative weighted energy efficiency deviation is obtained as 0.103.

[0060] The system presets the threshold for the collaborative weighted energy efficiency decay event to be 0.095. When the collaborative weighted energy efficiency deviation exceeds the threshold for three consecutive time windows, a decay flag is triggered. In actual operation, the collaborative weighted energy efficiency deviation of CH-02 in four consecutive time windows were 0.098, 0.104, 0.109, and 0.111, respectively. The system assigns the collaborative decay event indicator variable of the corresponding sample to one.

[0061] In the improved random survival forest training sample construction phase, a total of 21,600 samples were collected, of which 3,180 were co-decay event samples, accounting for 14.7%. The system calculated the sample sampling probability by weighted summation of the co-decay event indicator variable of each sample across all devices and the corresponding device's co-risk amplification coefficient. The sampling probability of a high-risk sample was 0.000126, while the sampling probability of a normal sample was 0.000041, representing an increase in sampling probability of approximately 3.07 times.

[0062] After training with a collaborative weighted random survival forest model, the system predicts the current operating characteristics of CH-02 and finds that the collaborative survival function shows a rapid downward trend in the next 48 hours. The remaining survival time of the collaborative energy efficiency model is calculated to be 31.6 hours, while the traditional random survival forest model predicts a remaining time of 47.2 hours.

[0063] The system receives the mirror calibration trigger signal, collects energy consumption data for the most recent 24 time indices within the calibration interval, and constructs a physical-virtual residual vector.

[0064] At a certain time index, the actual energy consumption observation vector of the physical central air conditioning system is: [712kW, 213kW, 168kW, 4366kW, 4128kW], and the energy consumption vector predicted by the digital twin baseline model is: [668kW, 198kW, 156kW, 4212kW, 4010kW]. The calculated physical-virtual residual vector is: [44kW, 15kW, 12kW, 154kW, 118kW].

[0065] After constructing the residual matrix under 24 consecutive time indices, the system performs vectorized parameter inversion. After 6 iterations, the equivalent heat transfer coefficient on the evaporator side is adjusted from 285 kW / K to 263 kW / K, the equivalent heat transfer coefficient on the condenser side is adjusted from 312 kW / K to 294 kW / K, and the cooling tower heat transfer correction coefficient is adjusted from 0.98 to 0.93.

[0066] After calibration, the predicted power under the same operating conditions is 705kW, and the error between the actual power of 712kW and the power of ...

[0067] In the risk perception model prediction and control optimization phase, the system performs rolling optimization based on 12 steps in the prediction time domain.

[0068] Under the traditional method, the controlled variables are: chiller supply water temperature setpoint 279.5K, chilled water pump frequency 48Hz, cooling water pump frequency 50Hz, and cooling tower fan speed 85%. The total system power consumption is 2126kW.

[0069] The optimized control variables obtained by the method of this invention are: chiller supply water temperature setpoint 281.2K, chilled water pump frequency 44Hz, cooling water pump frequency 46Hz, and cooling tower fan speed 79%. The total power consumption of the system is reduced to 1974kW.

[0070] During a 30-day comparative test, the traditional method consumed an average of 104,600 kWh per day, while the method of this invention consumed an average of 95,820 kWh per day, resulting in an average daily power saving of 8,780 kWh and a power saving rate of approximately 8.39%.

[0071] Within the same operating period, the traditional method calculates the high-risk operating time as 96 hours, while the method of this invention calculates it as 63 hours, reducing the high-risk operating time by 34.4%. Furthermore, regarding the prediction accuracy of the random survival forest, the traditional model has a mean absolute error of 5.1 hours in survival time prediction on the validation set, while the collaborative weighted model of this invention has an error of 3.6 hours, representing a 29.4% reduction in prediction error.

[0072] 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.

Claims

1. An energy-saving operation method for a central air conditioning system based on multi-device collaborative control, characterized in that, include: Establish a device topology information database and store it as multimodal operation measurement data; Based on multimodal operation measurement data and equipment topology information, a digital twin baseline model is constructed according to the thermodynamic mechanism equation and the energy conservation equation; The multimodal operation measurement data is processed by time series structuring, and multi-scale embedding features are extracted through a spatiotemporal representation learning network to generate a random survival forest input feature set; Based on the operating condition where energy efficiency drops to a preset performance threshold, the energy efficiency degradation event of the equipment is marked, and the improved random survival forest algorithm is used to train the equipment energy efficiency survival probability prediction model, and the collaborative hazard function is output. The equipment energy efficiency survival probability prediction model is used to infer the latest multimodal operation measurement data, continuously generate equipment cooperative hazard function sequences and construct dynamic cooperative hazard function band index sequences. The dynamic cooperative hazard function band index sequences are compared with preset dynamic cooperative hazard function band thresholds. When any equipment index exceeds the threshold or shows a monotonically increasing trend within the sliding time window, a mirror calibration trigger signal is generated. After receiving the mirror calibration trigger signal, the physical-virtual residual vector between the energy consumption predicted by the digital twin baseline model and the actual energy consumption of the physical central air conditioning system is calculated, the digital twin baseline model is updated, and the digital twin model after mirror calibration is obtained. The mirror-calibrated digital twin model is imported into the risk perception model predictive control optimization module. Global collaborative optimization of multiple central air conditioning devices is performed in the prediction time domain to obtain the optimal collaborative energy-saving strategy. The optimal collaborative energy-saving strategy is decomposed into an executable control instruction set and sent to the programmable logic controller or distributed control system of the physical central air conditioning system for execution.

2. The energy-saving operation method for a central air conditioning system based on multi-device collaborative control according to claim 1, characterized in that, The establishment of the equipment topology information database and its storage as multimodal operation measurement data includes: Construct a multi-modal operation measurement data acquisition system for central air conditioning systems to acquire in real time temperature, pressure, flow, power, valve opening parameters, external meteorological parameters, and building load parameters of chillers, chilled water pumps, cooling water pumps, and cooling towers.

3. The energy-saving operation method for a central air conditioning system based on multi-device collaborative control according to claim 1, characterized in that, The construction of the digital twin baseline model based on the thermodynamic mechanism equation and the energy conservation equation includes: Based on the equipment topology information database, the connection relationships and loop directions between the chiller unit, chilled water pump, cooling water pump and cooling tower are determined, the set of equipment nodes and the set of connection edges are constructed, and the system topology diagram is formed. An energy conservation sub-model on the evaporator side of the chiller unit is constructed based on the system topology diagram as the cooling output channel of the digital twin baseline model; An energy conservation sub-model on the cooling water side is constructed based on the system topology diagram as the heat emission channel of the digital twin baseline model; A power sub-model of pump equipment is constructed based on the system topology diagram as the hydraulic execution channel of the digital twin baseline model; A cooling tower heat exchange sub-model is constructed based on the system topology diagram as an environmental coupling channel for the digital twin baseline model; The cooling output channel, heat discharge channel, hydraulic actuation channel, and environmental coupling channel are coupled according to the system topology diagram to form a set of state equations for the digital twin baseline model. The initial parameters of the set of state equations are calibrated based on the set of input variables to obtain the digital twin baseline model.

4. The energy-saving operation method for a central air conditioning system based on multi-device collaborative control according to claim 1, characterized in that, The process of performing time-series structuring on multimodal operational measurement data and extracting multi-scale embedding features through a spatiotemporal representation learning network includes: A unified time step is determined based on multimodal operation measurement data, and the multimodal operation measurement data is time-aligned and resampled according to the unified time step to obtain the set of input variables corresponding to the time index; Organize the set of input variables into column vectors of dimension d, and form a matrix of input variable sets; The input variable set matrix is ​​subjected to time series structuring processing. The length of the structuring window is set and a window sequence tensor with the time index as the endpoint is constructed. Perform normalization mapping on the window sequence tensor to generate a normalized window sequence tensor; Based on the standardized window sequence tensor, a standardized differential window sequence tensor is constructed to characterize the trend of changes in the operating state. The system topology graph is represented as a topological adjacency matrix; The standardized window sequence tensor, the standardized difference window sequence tensor, and the topological adjacency matrix are input into the spatiotemporal representation learning network. The spatiotemporal representation learning network performs temporal modeling and topological constraint aggregation on the input and extracts the embedded feature vector. The embedded feature vectors of each time scale branch are concatenated to form a multi-scale embedded feature vector; Constructing a random survival forest input feature set based on multi-scale embedded feature vectors; The input feature set of the random survival forest consists of multi-scale embedded feature vectors corresponding to each time index and their one-to-one correspondence with the time index of the time series of the input variable set. The input feature set of the random survival forest is then output to the device energy efficiency survival probability prediction model.

5. The energy-saving operation method for a central air conditioning system based on multi-device collaborative control according to claim 1, characterized in that, The process involves marking equipment energy efficiency degradation events based on operating conditions where energy efficiency drops to a preset performance threshold, and training an improved random survival forest algorithm to obtain an equipment energy efficiency survival probability prediction model, including: Based on the digital twin baseline model, the power sensitivity of the power of any device a to the control setpoint of any device b is calculated. The power sensitivity is proportionally converted to the control setpoint of device b and the power of device a to obtain the power coupling coefficient between device a and device b at the same time step. The power coupling coefficients between all pairs of devices constitute the device cooperative coupling strength matrix. Based on the equipment cooperative coupling strength matrix, the absolute values ​​of the power coupling coefficients between equipment j and all other equipment are summed and a constant 1 is added to obtain the cooperative risk amplification coefficient of equipment j. Based on the digital twin baseline model, the power of device j is divided by the instantaneous cooling capacity of the chiller evaporator to obtain the energy consumption index per unit cooling capacity. Based on the reference unit cooling energy consumption index obtained by device j during the digital twin baseline model calibration stage, the relative deviation between the unit cooling energy consumption index and the reference unit cooling energy consumption index is calculated, and the relative deviation is multiplied by the collaborative risk amplification coefficient of device j to obtain the collaborative weighted energy efficiency deviation. When the deviation of the collaborative weighted energy efficiency is greater than or equal to the threshold of the collaborative weighted energy efficiency decay event, sample i is marked as the collaborative energy efficiency decay event of device j, and the collaborative decay event indicator variable of the corresponding sample is assigned a value of one; otherwise, the collaborative decay event indicator variable is assigned a value of zero. When constructing training samples for a random survival forest, the sampling probability of the samples is calculated based on the co-decay event indicator variable. Based on the sampling probability, the input feature set of the random survival forest is stratified and sampled to obtain a training sample subset. The training sample subset is then used to train the collaborative weighted random survival forest model, thus forming the collaborative weighted random survival forest model. After training, the cooperative weighted random survival forest model outputs a cooperative survival function and a cooperative hazard function after inputting multi-scale embedded feature vectors.

6. The energy-saving operation method for a central air conditioning system based on multi-device collaborative control according to claim 5, characterized in that, The method of training a collaboratively weighted random survival forest model using a subset of training samples includes: During the node splitting process of each tree model, the samples belonging to the current node are weighted according to the collaborative risk amplification coefficient of device j, and the difference between the number of samples that actually experience collaborative decay events and the expected number of events under the assumption of no decay is weighted and accumulated to obtain the collaborative risk weighted log-rank statistic. The feature that maximizes the absolute value of the collaborative risk weighted log-rank statistic is selected as the optimal splitting feature to form a collaborative weighted random survival forest model.

7. The energy-saving operation method for a central air conditioning system based on multi-device collaborative control according to claim 1, characterized in that, The calculation of the physical-virtual residual vector between the energy consumption predicted by the digital twin baseline model and the actual energy consumption of the physical central air conditioning system, and the updating of the digital twin baseline model, includes: After receiving the mirror calibration trigger signal, the actual energy consumption observation vector of the physical central air conditioning system is read from the multimodal operation measurement data under a unified time step. At the same time, the digital twin baseline model is called to calculate the predicted energy consumption vector under the same time index and the same set of input variables. By subtracting the actual energy consumption observation vector from the predicted energy consumption vector, the physical-virtual residual vector is obtained, and the physical-virtual residual vectors corresponding to all time indices within the calibration interval are stacked in chronological order to form a residual matrix; The parameters used to characterize equipment performance degradation and heat transfer deviation in the digital twin baseline model are arranged in a fixed order to form a thermodynamic parameter vector; The sum of squares of the physical-virtual residual vectors corresponding to each time index in the residual matrix is ​​obtained by summing the squares of the residuals for the corresponding time index. The sum of squares of the residuals for all time indices within the calibration interval is accumulated to obtain the vectorized parameter inversion objective function. In the digital twin baseline model, a small perturbation is applied to each parameter in the thermodynamic parameter vector, the change in each component of the predicted energy consumption vector is recorded, and the ratio of the change to the corresponding small perturbation is used as the partial derivative approximation of the predicted energy consumption vector with respect to the thermodynamic parameter vector. All partial derivative approximations are arranged in the order of components to form the Jacobian matrix. The thermodynamic parameter increment vector is calculated based on the residual matrix and the Jacobian matrix, and the thermodynamic parameter vector is updated to obtain the calibration parameter vector; The calibration parameter vector is written back to the cooling output channel, heat discharge channel, hydraulic execution channel, and environmental coupling channel of the digital twin baseline model. This updates the evaporator-side equivalent heat transfer coefficient, condenser-side equivalent heat transfer coefficient, chilled water pump efficiency coefficient, cooling water pump efficiency coefficient, and cooling tower heat transfer correction coefficient used in the calculations within each channel to the corresponding components in the calibration parameter vector, thus obtaining the mirror-calibrated digital twin model.

8. The energy-saving operation method for a central air conditioning system based on multi-device collaborative control according to claim 1, characterized in that, The process of importing the mirror-calibrated digital twin model into the risk perception model prediction and control optimization module, and performing global collaborative optimization of multiple central air conditioning devices in the prediction time domain, includes: In the prediction time domain, the total power consumption of the system is calculated based on the mirror-calibrated digital twin model; Based on the collaborative hazard function output by the collaborative weighted random survival forest model, the energy efficiency life loss cost of equipment is calculated in the prediction time domain. By weighted summing and accumulating the total system power consumption and equipment energy efficiency lifespan depreciation cost at each prediction time point in the prediction time domain, the objective function for predictive control optimization of the risk perception model is obtained. During the optimization process, operational constraints are set, including cooling balance constraints and equipment operating boundary constraints. Under the conditions of satisfying the cooling balance constraint and the equipment operation boundary constraint, the global collaborative optimization of the control variable sequence in the prediction time domain is carried out. By simulating different control variable sequences in the digital twin model after mirror calibration, the corresponding risk perception model predictive control optimization objective function value is calculated, and the control variable sequence that minimizes the risk perception model predictive control optimization objective function value is selected as the optimal collaborative energy saving strategy. The optimal collaborative energy-saving strategy is decomposed into a set of executable control instructions, which are then sent to the programmable logic controller or distributed control system of the physical central air conditioning system for execution via a communication interface.

Images