Air conditioning system power response characterization method based on thermal inertia guidance and related devices

Abstract

Embodiments of the present application provide a kind of air conditioning system power response characterization method based on thermal inertia guide and related equipment, including obtaining the multidimensional operation data of each equipment in central air-conditioning refrigeration machine room, and based on the equivalent thermal inertia time scale of preset different time length position, the operation data of each dimension is converted into the thermal inertia state corresponding to each time length position, and node feature is constructed based on thermal inertia state and operation data;Based on the working condition division of multidimensional operation data, the current operating mode label is obtained, and the corresponding adjacency relationship and convolution parameter are determined based on the current operating mode label;Based on the neural network model using adjacency relationship and convolution parameter, node feature is analyzed, and system total power and each equipment power are obtained.The method can improve the accuracy of power response characterization.

Description

Technical Field

[0001] This application relates to the field of energy control technology, and in particular to a method for characterizing the power response of an air conditioning system based on thermal inertia guidance, and related equipment. Background Technology

[0002] Water-cooled central air conditioning chiller plants are a common type of cooling system in commercial, office, and public buildings. These systems typically consist of chillers, chilled water pumps, cooling water pumps, cooling towers, chilled water piping, and terminal heat exchange equipment. The total system power is subject to complex dynamic coupling with variables such as chilled water supply and return temperatures, chilled water flow rates, cooling water temperatures, outdoor weather conditions, equipment start-up and shutdown status, and pump and fan frequencies. Accurately establishing the mapping relationship from cooling demand and operating status to equipment power—i.e., chilled-electricity mapping—is a crucial foundation for assessing building flexible energy consumption potential, forecasting air conditioning loads, analyzing demand response, and optimizing energy-saving control.

[0003] Cold power mapping can analyze operating data using pre-trained models to obtain device power and improve the efficiency of power response characterization. However, the current cold power mapping scheme does not have high accuracy in power response characterization. Summary of the Invention

[0004] The main purpose of this application is to propose a power response characterization method and related equipment for air conditioning systems based on thermal inertia guidance, in order to improve the accuracy of power response characterization.

[0005] To achieve the above objectives, this application proposes a method for characterizing the power response of an air conditioning system based on thermal inertia guidance, including... The system acquires multi-dimensional operating data of various equipment in the central air conditioning refrigeration room, and converts the operating data of each dimension into the thermal inertia state corresponding to each time period based on the equivalent thermal inertia time scale of different preset durations. It also constructs node features based on the thermal inertia state and operating data. Based on multidimensional operational data, the operating conditions are divided to obtain the current operating mode label, and the corresponding adjacency relationship and convolution parameters are determined based on the current operating mode label; Based on a neural network model employing adjacency relationships and convolutional parameters, node features are analyzed to obtain the total system power and the power of each device.

[0006] Optionally, in one implementation, based on preset equivalent thermal inertia time scales for different duration levels, the operational data for each dimension is converted into a thermal inertia state corresponding to each duration level, including: Obtain the equivalent thermal inertia time scale for different preset durations and the inertia decay coefficient corresponding to each duration. The thermal inertia state of the operational data for each dimension is determined based on the inertia decay coefficient at the corresponding time level.

[0007] Optionally, in one implementation, operating conditions are divided based on multi-dimensional operating data to obtain the current operating mode label, including: Cluster the runtime data at consecutive sampling times to obtain multiple clusters at each sampling time; Based on the pre-trained temporal state model, the characteristics of each cluster at each sampling time are analyzed to obtain the temporal working mode label corresponding to each cluster at each sampling time that conforms to the temporal change. Determine the current operating mode label corresponding to the current time from the time sequence operating mode labels corresponding to multiple sampling times.

[0008] Optionally, in one implementation, the training process of the time-series state model includes: Obtain historical operation data distributed over time, and cluster the historical operation data to obtain each historical cluster distributed over time, as well as the reference operation mode label corresponding to each cluster. Based on the temporal state model, the characteristics of each historical cluster are analyzed to obtain the prediction working mode label corresponding to each cluster. The parameters of the time-series state model are updated based on the difference between the reference operating mode label and the predicted operating mode label.

[0009] Optionally, in one implementation, based on a neural network model employing adjacency relationships and convolutional parameters, node features are analyzed to obtain the total system power and the power of each device, including: Based on the neural network model, the node features are graph convolutioned according to the adjacency relationship and convolution parameters to obtain the node embedding matrix; Based on the feature mapping of the node embedding matrix using a multilayer perceptron, the total power of the system and the power of each device are obtained.

[0010] Optionally, in one implementation, a node embedding matrix is ​​obtained by performing graph convolution on node features based on adjacency relationships and convolution parameters using a neural network model, including: Based on the neural network model, the parameters corresponding to different running mode labels are used to perform graph convolution on the node features to obtain the initial node embedding matrix corresponding to each running mode label; The fusion weights are determined based on the current operating mode labels, and the initial node embedding matrices corresponding to each operating mode label are fused based on the fusion weights to obtain the node embedding matrices.

[0011] Optionally, in one embodiment, after converting the operational data of each dimension into the thermal inertia state corresponding to each duration based on the preset equivalent thermal inertia time scales for different durations, the method further includes: The deviation value of each thermal inertia state is determined based on the preset first-order inertia recursive relationship, and the thermal inertia state corresponding to each time period is adjusted based on the deviation value.

[0012] Another aspect of this application provides a power response characterization device for an air conditioning system based on thermal inertia guidance, comprising: The conversion unit is used to acquire multi-dimensional operating data of various equipment in the central air conditioning refrigeration room, and based on the preset equivalent thermal inertia time scales of different duration levels, convert the operating data of each dimension into the thermal inertia state corresponding to each duration level, and construct node features based on the thermal inertia state and operating data. The determination unit is used to divide the working conditions based on multidimensional operating data, obtain the current operating mode label, and determine the corresponding adjacency relationship and convolution parameters based on the current operating mode label; The analysis unit is used to analyze node features based on a neural network model employing adjacency relationships and convolutional parameters to obtain the total system power and the power of each device.

[0013] Another aspect of this application provides an electronic device, comprising: Memory, transceiver, processor, and bus system; The memory is used to store programs; The processor is used to execute programs in memory, including methods for performing the aspects mentioned above; Bus systems are used to connect memory and processor to enable communication between them.

[0014] Another aspect of this application provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the methods described above.

[0015] As can be seen from the above technical solutions, the embodiments of this application have the following advantages: By pre-setting equivalent thermal inertia time scales of different durations, the multi-dimensional operating data of each device in the central air conditioning chiller room are transformed into corresponding thermal inertia states at each time scale. These thermal inertia states are then combined with the original operating data to construct node features. Subsequently, operating conditions are classified based on the multi-dimensional operating data to obtain the current operating mode label. Based on this label, corresponding adjacency relationships and convolution parameters are matched. The matched neural network performs graph convolution analysis on the node features and outputs the total system power and the power of each device. This approach can solidify the time-series lag characteristics caused by heat storage in equipment and pipelines into the input features, and allows the network to perform calculations using parameters that conform to the current interaction patterns of the devices under different operating modes, improving the accuracy of power response characterization. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the architecture of the application of the power response characterization method for air conditioning systems based on thermal inertia provided in the embodiments of this application; Figure 2 This is a flowchart illustrating the power response characterization method for an air conditioning system based on thermal inertia guidance provided in an embodiment of this application. Figure 3 This is a schematic flowchart of the power response characterization method provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of the cold electro-mapping system provided in the embodiments of this application; Figure 5 This is a schematic diagram of the structure of the power response characterization device for an air conditioning system based on thermal inertia guidance provided in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0018] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0019] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments.

[0020] In the embodiments of this application, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0021] Furthermore, to better illustrate this application, numerous specific details are provided in the following detailed embodiments. Those skilled in the art should understand that this application can be implemented without certain specific details. In some instances, methods, means, components, and circuits well-known to those skilled in the art have not been described in detail in order to highlight the main points of this application.

[0022] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0023] Water-cooled central air conditioning chiller plants are widely used in commercial, office, and public buildings, and are the mainstream system for building cooling. The system mainly consists of chiller units, chilled water pumps, cooling water pumps, cooling towers, chilled water piping networks, and terminal heat exchange equipment. The overall operating power of the chiller plant is affected by many factors, including chilled water supply and return temperatures, chilled water flow rate, cooling water temperature, outdoor weather conditions, equipment start-up and shutdown status, and pump and fan operating frequencies. These variables have complex and dynamically changing coupling relationships. Accurately constructing a mapping relationship between cooling demand, equipment operating status, and equipment power consumption—that is, a chiller-electricity mapping model—is an indispensable foundation for conducting building flexible energy consumption potential assessments, air conditioning load forecasting, demand response analysis, and energy-saving optimization control of air conditioning systems.

[0024] However, water-cooled chiller plants face three unavoidable technical challenges in constructing a chilled water-electricity mapping model. First, the chilled water network exhibits significant thermal inertia; temperature changes at the chiller outlet do not immediately reflect on the terminal or return water temperature measurement points. The temperature signal requires several minutes to tens of minutes of transmission delay and amplitude attenuation to be transmitted, and this time lag continuously changes with real-time adjustments to chilled water flow and terminal valve openings, increasing the modeling difficulty due to dynamic time delay characteristics. Second, chiller plant daily operation involves various operating conditions, with significant differences in the strength of coupling between equipment, the dominant dynamic changes in the system, and the degree of operational nonlinearity under different operating modes. A single fixed model cannot adapt to all operating conditions. Finally, the number of monitoring points deployed in actual engineering sites is limited, and key demand-side data such as indoor temperature changes, personnel activity, and terminal valve openings are often incomplete, making it difficult to directly deduce the actual power consumption of the chiller plant from the terminal cooling demand side.

[0025] Currently, the mainstream technical approaches related to chiller-electricity mapping modeling in the industry can be divided into three categories, each with obvious limitations. The first category is gray box models built based on physical mechanisms. These methods use the performance curves of chillers, pumps, and fans, combined with the hydraulic and thermodynamic equations of the pipe network and the heat exchange mechanism of the cooling tower to build a power calculation framework. Some studies further incorporate a one-dimensional heat conduction model of the pipeline, using the method of characteristics and the finite difference method to solve the dynamic evolution of water temperature inside the pipeline to analyze the demand response characteristics of the chiller plant. Other studies use the recursive least squares algorithm to update model parameters online to achieve adaptive adjustment of the model, serving operational optimization control. These models have intuitive physical meanings and highly interpretable output results, but once the model structure is built, it is difficult to adjust flexibly. The modeling is highly dependent on detailed information such as the pipe network topology, pipe length and diameter, insulation performance parameters, and equipment factory performance parameters. Different project deployments require a large amount of manpower for parameter calibration, and it cannot well adapt to the time-varying transmission lag problem caused by the continuous speed regulation of variable frequency pumps.

[0026] The second type is a pure data-driven black-box model that relies entirely on operational data for training. This type typically employs deep learning architectures such as Long Short-Term Memory (LSTM) networks, convolutional neural networks, graph neural networks, and temporal convolution combined with dynamic graph convolution. It directly fits the nonlinear mapping between input operational variables and system output power using massive amounts of historical operational data. This modeling approach eliminates the need to manually construct complex physical equations and has a strong ability to fit complex nonlinear relationships. However, the internal computational logic of the model is difficult to interpret. It does not explicitly incorporate the evolution law of pipeline thermal inertia and the system's energy conservation mechanism into the model structure. It can only rely on massive amounts of historical data to allow the model to autonomously learn signal lag patterns, resulting in low data sample utilization efficiency. Furthermore, the overall model structure is fixed and cannot adaptively adjust the computational logic according to changes in operating modes, typically leading to a significant increase in power estimation errors during operating condition transitions.

[0027] The third type is a relatively simple hybrid modeling scheme. Typically, lag features are artificially constructed only during the model input stage. For example, load, supply and return water temperature, flow rate, and power data at certain historical moments are selected, or the time series is expanded by a sliding window and fed into algorithms such as fully connected networks, random forests, and gradient boosting trees for training. This type of method has a low threshold for deployment and is easy to promote in engineering applications. However, it essentially relies on experience to complete feature processing, and the lag time is mostly a fixed value. It does not characterize the thermal inertia of the pipeline network from the physical essence of convection transport and heat dissipation. Therefore, when the chilled water flow rate changes or the operating conditions change abruptly, the model cannot accurately reproduce the real dynamic response process of the chiller plant, and the power prediction accuracy will decline significantly.

[0028] Based on this, embodiments of this application provide a power response characterization method for air conditioning systems based on thermal inertia guidance to solve this technical problem.

[0029] System architecture and scenario description used in the embodiments of this application: Figure 1 This is a system architecture diagram applied to the thermal inertia-guided power response characterization method for air conditioning systems according to embodiments of this application. It includes a terminal 140, an Internet connection 130, a gateway 120, a server 110, etc.

[0030] Terminal 140 includes various forms of devices with display screens, such as desktop computers, laptops, PDAs (personal digital assistants), mobile phones, in-vehicle terminals, home theater terminals, dedicated terminals, intelligent voice interaction devices, smart home appliances, or aircraft. Furthermore, it can be a single device or a collection of multiple devices. Terminal 140 can communicate with the Internet 130 via wired or wireless means to exchange data.

[0031] Server 110 refers to a computer system that can provide certain services to terminal 140. Compared to ordinary terminal 140, server 110 has higher requirements in terms of stability, security, and performance. Server 110 can be a single high-performance computer in a network platform, a cluster of multiple high-performance computers, a portion of a single high-performance computer (e.g., a virtual machine), or a combination of portions of multiple high-performance computers (e.g., virtual machines).

[0032] Gateway 120, also known as an internetwork connector or protocol converter, is a computer system or device that acts as a translator, enabling network interconnection at the transport layer. It bridges the gap between two systems using different communication protocols, data formats, languages, or even completely different architectures. Gateways can also provide filtering and security functions. Messages sent from terminal 140 to server 110 are forwarded to the corresponding server 110 via gateway 120. Messages sent from server 110 to terminal 140 are also forwarded to the corresponding terminal 140 via gateway 120.

[0033] The power response characterization method for air conditioning systems based on thermal inertia guidance provided in this application embodiment can be implemented separately in terminal 140, separately in server 110, or partially in terminal 140 and partially in server 110.

[0034] When the power response characterization method of the air conditioning system based on thermal inertia guidance provided in this application embodiment is implemented alone in the terminal 140, the terminal 140 acquires multi-dimensional operating data of each device in the central air conditioning refrigeration room, and converts the operating data of each dimension into the thermal inertia state corresponding to each time period based on the preset equivalent thermal inertia time scale of different duration levels, and constructs node features based on the thermal inertia state and operating data; the terminal 140 performs working condition division based on the multi-dimensional operating data to obtain the current operating mode label, and determines the corresponding adjacency relationship and convolution parameters based on the current operating mode label; the terminal 140 analyzes the node features based on the neural network model using adjacency relationship and convolution parameters to obtain the total power of the system and the power of each device.

[0035] The following description, in conjunction with the accompanying drawings, illustrates the power response characterization method for air conditioning systems based on thermal inertia guidance provided in this application. The execution entity of this method is a terminal device, specifically implemented by the terminal device running the various computer programs described above. Of course, based on the understanding of the following text, it is not difficult to see that the power response characterization method for air conditioning systems based on thermal inertia guidance provided in this application can also be implemented collaboratively by the terminal device and the server.

[0036] Please see Figure 2 ,like Figure 2 The diagram shown is a flowchart illustrating a power response characterization method for an air conditioning system based on thermal inertia guidance provided in this embodiment. The method includes: Step 201: Obtain multi-dimensional operating data of each device in the central air conditioning refrigeration room, and based on the preset equivalent thermal inertia time scale of different duration levels, convert the operating data of each dimension into the thermal inertia state corresponding to each duration level, and construct node features based on the thermal inertia state and operating data.

[0037] Multidimensional operational data can be various operational monitoring values ​​collected in real time by various sensors at the water-cooled chiller station, such as operational data of chiller units, chilled water pumps, cooling water pumps, cooling towers, indoor temperature and humidity, and outdoor weather.

[0038] The equivalent thermal inertia time scale can be a physical parameter used to quantify the lag time experienced by a water temperature / heat signal from a change to the point where it has an impact. Different time ranges can refer to the pre-selection of multiple sets of equivalent time constants (such as 15min, 30min, and 60min) corresponding to short-term disturbance response, medium heat transfer inertia, and long-term thermal buffering effect, respectively. For each key operational observation variable, a multi-scale thermal inertia state is generated according to each time constant to comprehensively characterize the multi-level dynamic lag characteristics brought about by the water-side circulation, heat transfer, and heat storage of the chiller.

[0039] Thermal inertia state can be a derived state quantity that incorporates the accumulated heat storage effect over a period of time into the current data. It reflects the state after the lag and decay characteristics of heat transfer. That is, for each piece of original operating data, inertial recursion calculation is performed according to each equivalent thermal inertia duration to obtain the heat storage state of the variable at different lag scales, supplementing the missing time-series lag information in the original instantaneous data. The original operating monitoring values ​​corresponding to each device and the thermal inertia state values ​​calculated at each duration are vectorized and concatenated to obtain the node features carried by each node in the graph when modeling the graph neural network.

[0040] Step 202: Divide the operating conditions based on the multidimensional operating data to obtain the current operating mode label, and determine the corresponding adjacency relationship and convolution parameters based on the current operating mode label.

[0041] Operating condition classification can be based on the distribution characteristics of equipment operating parameters, dividing the continuously changing operating states of the cooling plant into several typical operating conditions to distinguish different equipment linkage operation modes. The operating mode described in the current system can be indicated by the current operating mode label.

[0042] The adjacency relationship can be a topological matrix in a graph structure used to describe whether there is an association between nodes and the strength of that association. The size of the matrix elements represents the energy coupling strength between devices. In this embodiment, a separate set of adjacency matrices can be trained for each type of operating mode. The coupling strength between devices varies under different operating conditions, and the adjacency matrix can reflect the energy transfer and interaction characteristics of devices under that mode.

[0043] The convolution parameters can be weight matrices used in graph convolutional networks for node feature transformation and information aggregation, and are used to update node embeddings layer by layer. That is, each type of operating mode is configured with exclusive convolution weights to adapt to the nonlinear mapping rules under that condition.

[0044] Step 203: Based on the neural network model using adjacency relationships and convolution parameters, analyze the node features to obtain the total system power and the power of each device.

[0045] A neural network model can be a deep learning network consisting of multiple layers of graph convolutional layers and output regression layers. It can aggregate node association information based on graph topology and output predicted values.

[0046] The total system power can be the total power consumption of all electrical equipment in the entire chiller plant; the individual power consumption of each piece of equipment can be the power consumption of each chiller unit, chilled water pump, cooling water pump, and cooling tower fan. The node features are input into a pattern-adaptive graph neural network that has been matched with the current operating condition's adjacency relationships and convolution parameters. Through multi-layer graph convolution operations, the energy coupling relationships between different devices are aggregated to generate the overall total power of the chiller plant system, as well as the individual power consumption values ​​for each piece of equipment, such as chillers, various pumps, and cooling towers.

[0047] The power response characterization method for air conditioning systems based on thermal inertia guidance provided in this application converts the multi-dimensional operating data of each device in the central air conditioning chiller room into corresponding thermal inertia states at each time scale by pre-setting equivalent thermal inertia time scales of different durations. The thermal inertia states and the original operating data are then combined to construct node features. Subsequently, operating conditions are divided based on the multi-dimensional operating data to obtain the current operating mode label. Based on this label, corresponding adjacency relationships and convolution parameters are matched. The matched neural network performs graph convolution analysis on the node features and outputs the total system power and the power of each device. This method can solidify the time-series lag characteristics caused by heat storage in equipment and pipelines into the input features, and allows the network to perform calculations using parameters that conform to the current device interaction patterns under different operating modes, thus improving the accuracy of power response characterization.

[0048] In one feasible implementation, based on preset equivalent thermal inertia time scales for different duration levels, the operational data for each dimension is converted into a thermal inertia state corresponding to each duration level, including: Obtain the equivalent thermal inertia time scale for different preset durations and the inertia decay coefficient corresponding to each duration. The thermal inertia state of the operational data for each dimension is determined based on the inertia decay coefficient at the corresponding time level.

[0049] In this implementation, the inertial decay coefficient corresponding to each duration level can be a coefficient used to weigh the weight of the thermal inertia state at the previous moment against the weight of the current instantaneous observation value; the closer the inertial decay coefficient is to 1, the higher the proportion of historical heat storage state is retained, the smaller the impact of the current instantaneous data, and the stronger the hysteresis effect.

[0050] The system retrieves multiple sets of equivalent thermal inertia time scale durations that were pre-configured in the early stages of modeling. Combined with the data sampling interval, it calculates the inertia decay coefficient corresponding to each duration using an exponential formula. Subsequently, for the runtime sequence data of each dimension collected from the computer room, it iterates the calculation time-by-time using the corresponding duration inertia decay coefficient to solve the thermal inertia state sequence corresponding to each runtime dimension at each duration.

[0051] By first presetting multiple thermal inertia time scales and then converting them into corresponding inertia decay coefficients for recursive calculations, the dynamics of heat storage lag at different levels in the short, medium, and long term of the cold station can be characterized simultaneously in a standardized mathematical way. This eliminates the need to calibrate each section of the pipeline parameters on-site or dynamically adjust the lag time according to real-time flow, thus reducing the data requirements for project implementation.

[0052] In one feasible implementation, operating conditions are divided based on multi-dimensional operating data to obtain the current operating mode label, including: Cluster the runtime data at consecutive sampling times to obtain multiple clusters at each sampling time; Based on the pre-trained temporal state model, the characteristics of each cluster at each sampling time are analyzed to obtain the temporal working mode label corresponding to each cluster at each sampling time that conforms to the temporal change. Determine the current operating mode label corresponding to the current time from the time sequence operating mode labels corresponding to multiple sampling times.

[0053] In this implementation, clustering can automatically divide samples into several groups based on the similarity of their numerical distributions, with samples within each group having similar characteristics and significant differences between groups. Multidimensional operational data (water temperature, flow rate, frequency, start / stop status, etc.) collected at each sampling time are used as sample feature vectors for static clustering. This initially divides different operational categories from the perspective of single-time-time operational features. Specifically, the multidimensional operational vector for each sampling time is extracted, input into a clustering algorithm (such as K-means), and the operational data at each time time is assigned to the nearest cluster, resulting in the preliminary static clustering grouping result for that time moment.

[0054] Clusters can be groups of samples from clustering outputs, with the same cluster representing samples with similar operational characteristics.

[0055] A time-series state model can be a probabilistic model specifically designed for processing time series. It assumes that the system state transitions smoothly over time without random abrupt changes, and can perform time-series smoothing corrections to static clustering results. The model can be pre-trained using historical cold-station operating time-series data to learn the transition patterns before and after normal operating conditions. Using static cluster features as observation input, the output is an operating condition state that takes into account temporal continuity.

[0056] The time-series working mode label can be introduced on the basis of static clustering to introduce time-series constraints, avoid misjudgment of working conditions caused by data fluctuations at a single moment, and output candidate working condition labels after time-series smoothing for each sampling moment.

[0057] Understandably, the process involves first collecting multidimensional operational data from multiple consecutive sampling times in the computer room, then using a clustering algorithm to statically group the operational features at each time point to obtain a preliminary cluster corresponding to each sampling time point. Subsequently, a time-series state model pre-trained based on historical time-series data is invoked, using the cluster features at each time point as the observation input. Based on the time-series transition law of the operating conditions, the defects of static clustering that are prone to jumps are corrected, and the time-series working mode label corresponding to each sampling time point that meets the requirements of operational sequence continuity is obtained. Finally, the final current operating mode label for the current sampling time point is determined.

[0058] First, clustering is used to mine the inherent operating condition grouping features in the operational data. Then, a pre-trained time series state model is used to introduce time continuity constraints, so that the operating condition judgment fits the actual law of smooth transition of the cold station's operating state. Finally, the output current operating mode label is more stable and the operating condition classification is more reliable.

[0059] In one feasible implementation, the training process of the time-series state model includes: Obtain historical operation data distributed over time, and cluster the historical operation data to obtain each historical cluster distributed over time, as well as the reference operation mode label corresponding to each cluster. Based on the temporal state model, the characteristics of each historical cluster are analyzed to obtain the prediction working mode label corresponding to each cluster. The parameters of the time-series state model are updated based on the difference between the reference operating mode label and the predicted operating mode label.

[0060] In this implementation, the historical operational data distributed in time sequence can refer to the historical monitoring dataset of the computer room recorded in chronological order of sampling time and continuous in time dimension, which includes multi-dimensional operational variables such as temperature, flow rate, equipment frequency, and start / stop status.

[0061] Historical clusters can be grouped by dividing the multidimensional running vectors at each historical sampling time using a clustering algorithm. Clusters with similar running characteristics have similar characteristics.

[0062] The reference operating mode label can be an operating condition category that can be used as a training ground value, obtained from historical operating experience, manual operating condition annotation, or statistics from operating over a longer period.

[0063] The predicted working mode label can be a working condition category number that the model autonomously infers based on the cluster features of the input. In other words, it is the working condition prediction result given by the model for historical samples under the current parameters.

[0064] Understandably, the process begins by retrieving historical multidimensional operational data stored in the computer room in chronological order. Clustering is then performed on the operational data from all historical sampling points to obtain historical clusters arranged sequentially along the time series. Simultaneously, a corresponding reference operational mode label is matched to each cluster as the training ground value. Subsequently, the features of each historical cluster are input into the current time-series state model to be trained. The model, combined with the time-series state transition rules, derives and outputs the predicted operating mode label corresponding to each cluster. Next, the deviation between the predicted label and the pre-set reference label is calculated. This deviation is used as the optimization target to iteratively adjust various parameters within the time-series state model. After multiple rounds of training iterations, the model can learn the true time-series transition patterns of the cold storage operation, completing the pre-training of the time-series state model.

[0065] By using time-series ordered historical data for training, the model can not only learn the correspondence between single-moment operating characteristics and operating conditions, but also grasp the inherent transition law of the smooth switching of cooling station operating conditions over time, reducing power estimation fluctuations caused by misjudgment of operating conditions.

[0066] In one feasible implementation, based on a neural network model employing adjacency relationships and convolutional parameters, node features are analyzed to obtain the total system power and the power of each device, including: Based on the neural network model, the node features are graph convolutioned according to the adjacency relationship and convolution parameters to obtain the node embedding matrix; Based on the feature mapping of the node embedding matrix using a multilayer perceptron, the total power of the system and the power of each device are obtained.

[0067] In this implementation, graph convolution can be an operation in a graph neural network that aggregates neighbor node information and updates its own features based on the adjacency relationships between nodes. It relies on convolution parameters to complete feature transformation and uncover the interrelationships between nodes. Under the constraints of the adjacency relationships and convolution parameters corresponding to the current operating mode, multi-layer information aggregation operations are performed on the node features constructed above, integrating the coupling effects between nodes from different devices, states, and environments.

[0068] The node embedding matrix can be generated by aggregating neighborhood information and transforming weights for each original node feature after graph convolution operation, producing a deeper feature with higher dimension and containing node interaction information. All the deep features of the nodes are arranged in rows and columns to form the node embedding matrix.

[0069] A multilayer perceptron can be a feedforward neural network composed of multiple layers of fully connected neurons. It can perform nonlinear mapping on high-dimensional features to achieve regression or classification outputs. Here, features can be further extracted from the node embedding matrix to output the final power value.

[0070] Feature mapping can be a nonlinear transformation process in which a multilayer perceptron converts high-dimensional intermediate features into target output quantities, transforms the node embedding containing device coupling information into quantifiable power consumption values, and simultaneously separates the overall output and the power results of a single device into two categories.

[0071] Understandably, the node features obtained by integrating the original operating data and the multi-level thermal inertia state are input into a neural network model that has been matched with the corresponding adjacency relationships and convolution parameters according to the current operating mode label. The energy coupling relationship between each node is aggregated according to the adjacency matrix of the current mode, and the corresponding convolution parameters are used to complete the layer-by-layer transformation and information extraction of the node features, generating a node embedding matrix containing deep interaction information. Then, the node embedding matrix is ​​input into a pre-trained multilayer perceptron, and the mapping relationship between deep features and power consumption is mined through multilayer nonlinear feature mapping. Finally, the total system power of the central air conditioning chiller and the sub-equipment power corresponding to each device are output simultaneously.

[0072] By leveraging adjacency relationships adapted to the current operating mode to perform graph convolution operations, we can accurately capture the differentiated energy transfer and mutual influence patterns between devices under different operating conditions. Subsequently, we can improve the accuracy of the cold power mapping results by embedding nodes using a multilayer perceptron and performing nonlinear regression mapping.

[0073] In one feasible implementation, a node embedding matrix is ​​obtained by performing graph convolution on node features based on adjacency relationships and convolution parameters using a neural network model, including: Based on the neural network model, the parameters corresponding to different running mode labels are used to perform graph convolution on the node features to obtain the initial node embedding matrix corresponding to each running mode label; The fusion weights are determined based on the current operating mode labels, and the initial node embedding matrices corresponding to each operating mode label are fused based on the fusion weights to obtain the node embedding matrices.

[0074] In this implementation, the parameters corresponding to different operating mode labels can be learnable network parameters such as adjacency matrices and convolution weights obtained separately from each type of operating condition during graph convolution operations. Different operating conditions correspond to different coupling rules and feature transformation rules. Graph convolution is performed using the exclusive parameters for each type of operating mode, and each mode generates an independent initial node embedding matrix, representing the interaction features of device nodes under the assumed system operating condition.

[0075] Fusion weights can be based on the operating mode label identified at the current moment, and different operating modes can be initially embedded with fusion weights. The current operating mode has a higher weight, and adjacent transitional operating conditions are assigned lower weights to smooth the operating condition switching process.

[0076] Understandably, when performing graph convolution operations, the adjacency relationships and convolution parameters of all pre-trained neural networks corresponding to different operating mode labels are first retrieved. Each set of mode-specific parameters is used to perform graph convolution operations on the same set of input node features, thereby generating the initial node embedding matrix corresponding to each type of operating mode. Then, based on the current operating mode label obtained from the online working condition classification, the fusion weight corresponding to each operating mode is calculated. Next, the fusion weight is used to perform a weighted fusion operation on all initial node embedding matrices, and the weighted sum is used to obtain the final fused node embedding matrix.

[0077] First, the parameters of all operating modes are used to generate corresponding initial node embeddings. Then, the fusion weights are dynamically configured and multi-mode feature fusion is performed based on the current operating mode label. On the one hand, when the operating conditions are stable, the features of the current dominant mode can make the main contribution and accurately match the coupling characteristics of the current device. On the other hand, during the transition phase of operating conditions, the feature information of different operating modes before and after can be smoothly connected, the output oscillation caused by slight misjudgment of operating condition labels is weakened, and the continuity and stability of node embedding features during the transition phase of operating conditions are improved.

[0078] In one feasible implementation, after converting the operational data of each dimension into the thermal inertia state corresponding to each duration based on the preset equivalent thermal inertia time scales for different durations, the method further includes: The deviation value of each thermal inertia state is determined based on the preset first-order inertia recursive relationship, and the thermal inertia state corresponding to each time period is adjusted based on the deviation value.

[0079] In this implementation, the first-order inertial recursive relationship can be a fixed mathematical formula that recursively derives the current thermal inertial state from the previous thermal inertial state and the current instantaneous observation value. This formula is used to determine whether the currently calculated thermal inertial state conforms to the inherent evolution law of the system's heat storage change.

[0080] After initially generating the thermal inertia state corresponding to each operating dimension at each time interval by using the equivalent thermal inertia time scale of different duration levels through basic first-order inertia calculations, the deviation value of the preliminary thermal inertia state relative to the theoretical inertia evolution result is calculated moment by moment using the preset first-order inertia recursive relationship as the physical verification basis. Then, the preliminary thermal inertia state is corrected and adjusted according to the magnitude of the calculated deviation, and unreasonable state abrupt changes are filtered out so that the temporal change of the thermal inertia state at each time interval after correction always conforms to the physical law of the heat storage lag on the water side of the chiller.

[0081] Understandably, first-order inertial recursion has a time-accumulation characteristic. Once the original observation data at a certain moment is incorrect, this error will continue to affect the thermal inertia state at multiple subsequent sampling moments along the recursive calculation, causing false and violent fluctuations in the heat storage state curve, which contradicts the true physical characteristics of the chilled water pipe network where heat is transferred slowly and does not change instantaneously. Therefore, using a preset first-order inertial recursive relationship as a physical benchmark, the deviation value of the preliminary thermal inertia state relative to the theoretical evolution trajectory is calculated moment by moment. As long as the deviation exceeds a preset reasonable threshold, the thermal inertia state at the corresponding time level is corrected according to the magnitude of the deviation, cutting off the spread of abnormal measurement values ​​in the time-series recursive chain, so that each thermal inertia state conforms to the dynamic change law of heat storage and release in the pipe network, making the extraction of coupling information embedded in the nodes more accurate, and improving the anti-interference capability of the cold electrical mapping model in complex sensing environments on site.

[0082] In one example, the power response representation based on a thermally inert guided graph neural network can be described as follows.

[0083] First, collect operational data. This data can be categorized into three types: The first type is essential data, including chilled water supply temperature, chilled water return temperature, chilled water flow rate, start / stop status of each chiller unit, outdoor dry-bulb temperature, relative humidity, total system power or total chiller plant power, and basic configuration parameters such as equipment type, measurement point correspondence, rated cooling capacity, and rated power. The second type is recommended data, including chiller unit power, chilled water pump power, cooling water pump power, cooling tower power, cooling water inlet and outlet temperatures, cooling water flow rate, pump or fan frequency, and chilled water supply setpoints. The third type is auxiliary configuration data, including system process relationships, necessary operating boundary parameters, and other field information that can be used to assist in judging operating conditions.

[0084] In the data preprocessing stage, multi-source data can be aligned according to a unified timestamp, with a preferred sampling interval of 1 minute, and resampled to a 5-minute average. For continuous variables with missing times less than 30 minutes, linear interpolation is used for repair; for continuous segments with missing times greater than or equal to 30 minutes, they are directly removed from the training samples. Z-score normalization is applied to continuous variables, while discrete variables such as equipment start / stop status are kept in 0 / 1 form. One-hot vector encoding is used for subsequent operating condition mode labels, i.e., the position corresponding to the current mode is set to 1, and the other positions are set to 0. This yields a unified time-based input feature matrix.

[0085] Let the first The node feature matrix at each sampling time is:

[0086] in, This represents the number of devices or pipeline nodes in the diagram. The input feature dimensions for each node. Node features may include the temperature, flow rate, start / stop status, frequency, power, and related meteorological variables of the device corresponding to the node.

[0087] Secondly, determine the thermal inertia state corresponding to the running data, and construct node features based on the thermal inertia state and the running data.

[0088] The key equipment, system state variables, and thermal inertia states in a water-cooled central air conditioning chiller plant are abstracted into a graph structure to represent the coupling relationships between the operation of internal equipment, chilled water side state, cooling water side state, and outdoor environment. The graph structure is defined as follows:

[0089] Where V is the set of nodes and E is the set of edges. Nodes can be set as equipment nodes, state nodes, and environment nodes based on field measurement points and modeling granularity. Equipment nodes may include chillers, chilled water pumps, cooling water pumps, and cooling towers; state nodes may include variables such as cooling load, chilled water supply and return temperatures, chilled water temperature difference, cooling water supply and return temperatures, cooling water temperature difference, equipment frequency feedback, and historical power; environment nodes may include meteorological variables such as outdoor temperature and outdoor humidity. Edges are used to describe energy flow, control relationships, operational dependencies, and thermal inertia transfer relationships between equipment or variables.

[0090] To characterize the dynamic lag and smooth response caused by water-side circulation, equipment heat exchange, and control regulation in the chiller plant, a thermal inertia state can be constructed for key operating variables. Let the observed value of the j-th key operating variable at time t be... The thermal inert state at time constant τ is Then, using a first-order inertial recurrence relation, it can be expressed as:

[0091]

[0092] in, The sampling interval is... The equivalent thermal inertia time constant, This is the corresponding inertial decay coefficient. Multiple candidate values ​​can be set according to the sampling granularity and system response characteristics, such as 15min, 30min and 60min, so as to simultaneously describe short-term disturbances, medium-time-scale responses and long-time-scale thermal buffering effects.

[0093] Specifically, thermal inertia states can be constructed for variables such as cooling load, chilled water temperature difference, cooling water temperature difference, chilled water return temperature, cooling water return temperature, and outdoor temperature, and these states, together with current operating variables, time characteristics, and historical power characteristics, form the model input. This yields the initial node features or global input features at the t-th sampling time:

[0094] in, This represents a feature organization or node mapping function. This represents contextual features such as equipment start-up and shutdown, frequency feedback, time encoding, and operating condition labels. This representation does not limit nodes to completely corresponding to a single physical device; rather, it allows for the integration of equipment, operating variables, and thermal inertia states into the graph neural network based on the granularity of field measurement points.

[0095] Next, the operating conditions are identified based on node characteristics to obtain the current operating mode.

[0096] A time-series operating condition identification method can be adopted by combining K-means clustering (K-Means) with a Hidden Markov Model (HMM). This method uses clustering algorithms to obtain initial operating condition divisions and utilizes the state transition structure of the HMM to perform temporal smoothing of the pattern sequence, thereby reducing pattern jitter caused by instantaneous noise.

[0097] First, for each moment Construction condition identification feature vector .in, This can include the following seven characteristics: system load rate, chilled water supply setpoint change rate, chiller start / stop change indicator, cooling tower fan frequency change rate, system power standard deviation over the past 30 minutes, chilled water supply temperature change rate over the past 30 minutes, and chilled water return temperature change rate over the past 30 minutes. The system load rate can be written as:

[0098] in, The real-time cooling capacity of the refrigeration station is given at time t. The rated cooling capacity of all chiller units in the chiller plant.

[0099] The rate of change of chilled water setpoint is:

[0100] The setpoint for chilled water supply temperature issued by the controller at time t; This is the setpoint for the chilled water supply temperature corresponding to the previous sampling time (t-1).

[0101] The power fluctuation characteristics are:

[0102] in, This represents the number of sampling points corresponding to a 30-minute window. This is the function for calculating standard deviation. This represents the real-time total power consumption of the chiller plant system. Let be the system total power time series from time (tr) to time t for r consecutive sampling points.

[0103] Then, for all historical data Perform Z-score normalization and use K-means clustering to obtain the initial label sequence:

[0104] in, For the number of operating modes, the preferred option is... These correspond to standby or very low load, low load stability, medium-high load stability, chiller loading / unloading transition, cooling tower adjustment transition, and chilled water setpoint step, respectively. K-means clustering was repeated several times, and the result with the smallest sum of squares within each cluster was selected as the initial partition. The mean of each pattern was obtained from the clustering results. Covariance Matrix , which serves as the initial parameter for the observation probability of the Hidden Markov Model.

[0105] In a hidden Markov model, let the hidden state be:

[0106] The observed variable is Initial state distribution and state transition matrix They are respectively recorded as:

[0107] No. The observation probability under this model is defined as a Gaussian distribution:

[0108] The hidden Markov model parameters are trained using the Baum-Welch Algorithm, and then the most probable operating condition sequence is decoded using the Viterbi Algorithm.

[0109] The sequence of hidden operating modes represents sampling times 1 to T; The sequence of observed features represents the sampling time from time 1 to time T; argmax is used to find the hidden state sequence that maximizes the conditional posterior probability. This is the optimal timing mode tag sequence of the decoded output.

[0110] Finally, short mode segments with a duration of less than 3 time steps are post-processed and merged into adjacent dominant modes to remove noise spikes and output stable mode labels. .

[0111] Subsequently, the system's total power and the power of each device are analyzed based on the parameters corresponding to the current operating mode using a pattern-adaptive graph neural network.

[0112] The mode-adaptive graph neural network in this embodiment consists of multiple layers of graph convolutions, such as a graph convolutional network (GCN). It can employ a 3-layer structure with hidden dimensions set to 64, 128, and 64 respectively. Unlike ordinary graph convolutional networks, it can accommodate various operating modes. Maintain an independent adjacency matrix and convolution weights During online forward propagation, the model determines the current pattern label. Dynamically select the corresponding parameters to achieve adaptive operation.

[0113] In graph convolution computation, adjacency relations are used to describe the coupling strength between the nodes mentioned above. For the m-th operating mode, the adjacency matrix can be represented by the initial adjacency matrix, the mode-related scaling term, and the trainable bias term:

[0114] in, This is the initial adjacency matrix determined based on system process relationships, variable coupling relationships, and operational relevance; the symbol ⊙ represents the element-wise multiplication of the matrix, i.e., the Hadamard product. Let be the trainable scaling matrix for the m-th operating mode; Let be the trainable bias matrix for the m-th operating mode. This structure enables the model to learn differentiated edge weight distributions based on different operating conditions while preserving the basic coupling relationships of the cooling station, thereby expressing the dynamic differences under operating conditions such as stable operation, equipment loading and unloading, cooling tower adjustment, and setpoint changes.

[0115] For the The first mode, the second Layered graph convolution is calculated as follows:

[0116] in:

[0117]

[0118] The initial node feature matrix, It is the identity matrix. For the activation function, a Rectified Linear Unit (ReLU) function is preferred, and the output layer can use linear activation.

[0119] Mode scheduling is performed using a hard handover method, i.e.:

[0120] Among them, only the current mode The corresponding adjacency matrix and convolution weights are involved in the calculation. Hard switching helps maintain the interpretability of parameters across different modes, allowing engineers to view the edge weight distribution in each mode separately. To reduce potential prediction jumps at mode boundaries, embodiments of this application can also provide a soft transition alternative. During the inference phase, the graph convolution outputs of all modes can be calculated simultaneously. The fusion weights are calculated using a lightweight gating network.

[0121]

[0122] in, This is the one-hot vector representation of the current pattern label. This is the global pooling vector from the previous time step. It is a gated multilayer perceptron. Soft transitions can be used during the inference phase to smooth the output, as an optional supplement to hard-switching scheduling.

[0123] After the final graph convolution layer, the node embedding matrix is ​​obtained:

[0124] in, To hide the feature dimension, this embodiment performs global average pooling on the node dimension to obtain a global feature vector:

[0125] Then, the total system power is output through two independent multilayer sensing machines. Power distribution to each device :

[0126]

[0127] in, Used to output the total power of the chiller plant system The weights are shared across all nodes and used to output the power components corresponding to each device or node. When some nodes lack measured power component labels, the power component loss is calculated only for nodes with labels.

[0128] It is understood that the embodiments of this application may also introduce physical constraint loss during the training process, and jointly optimize it with the main task loss to improve the model's adherence to dynamic response and energy relationship.

[0129] First, based on the energy balance on the chilled water side, the system cooling capacity can be calculated from the chilled water flow rate and the supply and return water temperature difference:

[0130] in, This is the volumetric flow rate of chilled water. This refers to the chilled water return temperature. The temperature of the chilled water supply. Energy conservation loss is defined as:

[0131] in, The coefficient of performance (COP) of the chiller unit can be output from an auxiliary model or the measured COP can be used as a soft target. This represents the chiller power output by the model.

[0132] Secondly, regarding the construction process of the thermal inertial state, the thermal inertial state obtained through internal model updates or external recursion should be consistent with the first-order inertial recursion relationship. The thermal inertial consistency constraint is defined as follows:

[0133] Where j represents the key operating variable involved in thermal inertia modeling, and τ represents the set of equivalent thermal inertia time constants. When the thermal inertia state is pre-calculated as an external feature, this relationship serves as a feature construction rule; when the thermal inertia state is updated within the network, this relationship can serve as a training constraint in joint optimization.

[0134] Furthermore, to prevent the adaptive adjacency relationship from deviating excessively from the basic operational coupling relationship of the cold station, an adjacency matrix regularization loss is introduced:

[0135] in, It is the Frobenius norm. Let m be the adjacency matrix for the m-th mode. This is the initial adjacency matrix. This term is used to encourage the learned edge weights to be adjusted around the basic coupling relationships, while allowing different operating modes to form differentiated connection strengths.

[0136] The total loss function is defined as:

[0137] in, The mean square error of the total system power. The mean square error of the power of the sub-devices. Loss due to energy conservation For thermal inertia consistency loss, For regularization loss, , , and This is a weighting coefficient; the specific value can be adjusted based on the quality of the project data and the completeness of the tags.

[0138] In one example, a flowchart illustrating the power response characterization method of this application can be referred to. Figure 3 As shown, it includes a training phase and an inference phase.

[0139] The training phase includes the following steps: First, offline completion of hot inertia state construction rules, initial adjacency relation construction, and convolution parameter initialization for each pattern graph. Second, historical data is divided into pattern labels, and then the pattern recognition model is trained based on historical data and pattern labels. For example, historical system operation data is accessed, data preprocessing is performed to remove abnormal samples and complete feature standardization, initial pattern division of historical operating conditions is carried out using the K-means clustering algorithm, and the Hidden Markov Model is trained based on the clustering results. At the same time, temporal pattern labels are smoothed to reduce label jump problems caused by instantaneous disturbances. Third, the data is divided into training, validation, and test sets according to time order, and an adaptive moment estimation optimizer (Adam) is used to train the pattern adaptive graph neural network. The initial learning rate can be set according to the sample size and model complexity, and can be gradually decayed with each training epoch. Among them, an early stopping strategy can be adopted, where training is stopped when the total power loss of the validation set system does not decrease for several consecutive training epochs, and the model parameters with the best performance on the validation set are rolled back.

[0140] Online inference phase: First, real-time acquisition of operational data within the current timeframe and historical windows is used to construct the current hot inertia state. Second, calculation of operating condition identification features is performed, and the current pattern label is output through a trained pattern recognition model. Third, based on the pattern label, the corresponding adjacency matrix and convolution weights are selected, and forward propagation of the pattern adaptive graph neural network is executed to output the total system power and the power distribution of each device.

[0141] In one example, this application embodiment also provides a cold electrical mapping system, the structural schematic diagram of which can be referred to... Figure 4 As shown, the system's overall architecture includes a data acquisition module, a thermal inertia state construction module, a pattern recognition module, a pattern adaptive graph neural network inference engine, and an output interface module. It also includes a field sensing layer, a model and parameter library, and upper-layer applications, forming a complete operational chain. The field layer deploys temperature sensors, flow sensors, and power measurement points to collect equipment status and frequency feedback, as well as meteorological data. Data sources for the field layer include various front-end sensing points and can also be accessed via SCADA / BAS / databases to retrieve existing operational information. The data acquisition module directly connects to the field sensors and the Supervisory Control and Data Acquisition (SCADA) system to acquire real-time temperature, flow, power, equipment status, and meteorological data, completing the aggregation of raw operational information.

[0142] The raw data is first processed by a data preprocessing module to reduce noise, clean, and standardize it. The processed data is then input into a thermal inertia state construction module. This module can call upon the thermal inertia state construction rules stored in the model and parameter library, and perform time-series recursive calculations based on current operating variables and historical states to generate multi-timescale thermal inertia states. This depicts the load delay variation caused by the heat storage lag of buildings and heat exchange equipment. The generated thermal inertia states, along with the preprocessed feature data, are then sent to a pattern recognition module. This module operates continuously online, reading pattern recognition model parameters from the model and parameter library, continuously identifying operating conditions, and continuously outputting the current operating condition mode label, thus achieving dynamic online determination of the operating mode.

[0143] The pattern-adaptive graph neural network inference engine loads the pre-trained GNN model parameters from the model and parameter library, receives the pattern labels output by the pattern recognition module, and then dynamically schedules the corresponding graph convolutional structures to perform inference calculations based on the obtained pattern labels. It can perform system and device power prediction calculations through either hard switching of a single pattern or soft transition fusion of multiple patterns. The prediction results output by the inference engine are fed into the output interface module, which outputs the prediction results in a standard format. Supported interface types include the Open Platform Unified Architecture (OPC UA), the Representational State Transfer Application Programming Interface (REST API), and database write interfaces, enabling flexible integration with different types of upper-layer application systems.

[0144] Data distributed outward through the output interface module will be sent to the upper-layer application module. The upper-layer application specifically includes four types of business functions: flexible energy consumption assessment, demand response analysis, energy-saving optimization control, and operation anomaly diagnosis. Based on the prediction results output by the model, it can complete actual business tasks such as load adjustability potential calculation, demand response strategy formulation, equipment energy-saving control command issuance, and equipment operation fault diagnosis.

[0145] The above describes a method for characterizing the power response of an air conditioning system based on thermal inertia. The following describes the apparatus for implementing this method.

[0146] Please see Figure 5 ,like Figure 5 The diagram shown is a schematic representation of a power response characterization device for an air conditioning system based on thermal inertia guidance, according to an embodiment of this application. The device 500 includes: The conversion unit 501 is used to acquire multi-dimensional operating data of various equipment in the central air conditioning refrigeration room, and based on the preset equivalent thermal inertia time scale of different duration levels, convert the operating data of each dimension into the thermal inertia state corresponding to each duration level, and construct node features based on the thermal inertia state and operating data. The determination unit 502 is used to divide the working conditions based on multidimensional operating data, obtain the current operating mode label, and determine the corresponding adjacency relationship and convolution parameters based on the current operating mode label; Analysis unit 503 is used to analyze node features based on a neural network model employing adjacency relationships and convolution parameters to obtain the total system power and the power of each device.

[0147] Optionally, in one embodiment, the conversion unit 501 is specifically used for: Obtain the equivalent thermal inertia time scale for different preset durations and the inertia decay coefficient corresponding to each duration. The thermal inertia state of the operational data for each dimension is determined based on the inertia decay coefficient at the corresponding time level.

[0148] Optionally, in one embodiment, the determining unit 502 is specifically used for: Cluster the runtime data at consecutive sampling times to obtain multiple clusters at each sampling time; Based on the pre-trained temporal state model, the characteristics of each cluster at each sampling time are analyzed to obtain the temporal working mode label corresponding to each cluster at each sampling time that conforms to the temporal change. Determine the current operating mode label corresponding to the current time from the time sequence operating mode labels corresponding to multiple sampling times.

[0149] Optionally, in one embodiment, the determining unit 502 is further configured to: Obtain historical operation data distributed over time, and cluster the historical operation data to obtain each historical cluster distributed over time, as well as the reference operation mode label corresponding to each cluster. Based on the temporal state model, the characteristics of each historical cluster are analyzed to obtain the prediction working mode label corresponding to each cluster. The parameters of the time-series state model are updated based on the difference between the reference operating mode label and the predicted operating mode label.

[0150] Optionally, in one embodiment, the analysis unit 503 is specifically used for: Based on the neural network model, the node features are graph convolutioned according to the adjacency relationship and convolution parameters to obtain the node embedding matrix; Based on the feature mapping of the node embedding matrix using a multilayer perceptron, the total power of the system and the power of each device are obtained.

[0151] Optionally, in one embodiment, the analysis unit 503 is specifically used for: Based on the neural network model, the parameters corresponding to different running mode labels are used to perform graph convolution on the node features to obtain the initial node embedding matrix corresponding to each running mode label; The fusion weights are determined based on the current operating mode labels, and the initial node embedding matrices corresponding to each operating mode label are fused based on the fusion weights to obtain the node embedding matrices.

[0152] Optionally, in one embodiment, the conversion unit 501 is further configured to: The deviation value of each thermal inertia state is determined based on the preset first-order inertia recursive relationship, and the thermal inertia state corresponding to each time period is adjusted based on the deviation value.

[0153] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method for characterizing the power response of an air conditioning system based on thermal inertia. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0154] Please see Figure 6 , Figure 6 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes: The processor 601 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory 602 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 602 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 602 and is called and executed by the processor 601 to implement the thermal inertia-guided power response characterization method for air conditioning systems according to the embodiments of this application. The input / output interface 603 is used to implement information input and output; The communication interface 604 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 605 transmits information between various components of the device (e.g., processor 601, memory 602, input / output interface 603, and communication interface 604); The processor 601, memory 602, input / output interface 603, and communication interface 604 are connected to each other within the device via bus 605.

[0155] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for characterizing the power response of an air conditioning system based on thermal inertia.

[0156] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0157] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0158] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0159] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0160] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0161] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0162] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0163] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0164] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0165] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0166] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0167] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A method for characterizing the power response of an air conditioning system based on thermal inertia guidance, characterized in that, include: The system acquires multi-dimensional operating data of each device in the central air conditioning refrigeration room, and converts the operating data of each dimension into the thermal inertia state corresponding to each time period based on the preset equivalent thermal inertia time scale of different duration levels. Node features are then constructed based on the thermal inertia state and the operating data. Based on the multidimensional operating data, the operating conditions are divided to obtain the current operating mode label, and the corresponding adjacency relationship and convolution parameters are determined based on the current operating mode label; Based on the neural network model employing the adjacency relationship and the convolution parameters, the node features are analyzed to obtain the total system power and the power of each device.

2. The method according to claim 1, characterized in that, The equivalent thermal inertia time scale based on preset different duration levels converts the operational data of each dimension into a thermal inertia state corresponding to each duration level, including: Obtain the equivalent thermal inertia time scale for different preset durations and the inertia decay coefficient corresponding to each duration. The thermal inertia state of the running data in each dimension at the corresponding time level is determined based on the inertia decay coefficient.

3. The method according to claim 1, characterized in that, The process of classifying operating conditions based on the multidimensional operating data to obtain the current operating mode label includes: The running data at consecutive sampling times are clustered to obtain multiple clusters at each sampling time; Based on the pre-trained temporal state model, the characteristics of each cluster at each sampling time are analyzed to obtain the temporal working mode label corresponding to each cluster at each sampling time that conforms to the temporal change. Determine the current operating mode label corresponding to the current time from the time sequence operating mode labels corresponding to the multiple sampling times.

4. The method according to claim 3, characterized in that, The training process of the time-series state model includes: Acquire historical operation data distributed in time sequence, and cluster the historical operation data to obtain each historical cluster distributed in time sequence, and a reference operation mode label corresponding to each cluster. The characteristics of each historical cluster are analyzed based on the time-series state model to obtain the prediction working mode label corresponding to each cluster. The parameters of the time-series state model are updated based on the difference between the reference operating mode label and the predicted operating mode label.

5. The method according to claim 1, characterized in that, The neural network model, based on the adjacency relationship and the convolution parameters, analyzes the node features to obtain the total system power and the power of each device, including: Based on the neural network model, graph convolution is performed on the node features according to the adjacency relationship and the convolution parameters to obtain the node embedding matrix; Based on the feature mapping of the node embedding matrix using a multilayer perceptron, the total system power and the power of each device are obtained.

6. The method according to claim 5, characterized in that, The method of performing graph convolution on the node features according to the adjacency relationship and the convolution parameters based on the neural network model to obtain the node embedding matrix includes: Based on the neural network model, the node features are graph convolutional with parameters corresponding to different running mode labels to obtain the initial node embedding matrix corresponding to each running mode label. The fusion weight is determined based on the current operating mode label, and the initial node embedding matrix corresponding to each operating mode label is fused based on the fusion weight to obtain the node embedding matrix.

7. The method according to claim 1, characterized in that, The method, which converts the operational data of each dimension into the thermal inertia state corresponding to each time period based on the preset equivalent thermal inertia time scales of different durations, further includes: The deviation value of each thermal inertia state is determined based on a preset first-order inertia recursive relationship, and the thermal inertia state corresponding to each time period is adjusted based on the deviation value.

8. A power response characterization device for an air conditioning system based on thermal inertia guidance, characterized in that, include: The conversion unit is used to acquire multi-dimensional operating data of various equipment in the central air conditioning refrigeration room, and convert the operating data of each dimension into the thermal inertia state corresponding to each time period based on the preset equivalent thermal inertia time scale of different duration levels, and construct node features based on the thermal inertia state and the operating data. The determining unit is used to divide the working conditions based on the multidimensional operating data, obtain the current operating mode label, and determine the corresponding adjacency relationship and convolution parameters based on the current operating mode label; The analysis unit is used to analyze the node features based on the neural network model using the adjacency relationship and the convolution parameters to obtain the total power of the system and the power of each device.

9. An electronic device, characterized in that, include: Memory, transceiver, processor, and bus system; The memory is used to store programs; The processor is configured to execute a program in the memory, including performing the method as described in any one of claims 1 to 7; The bus system is used to connect the memory and the processor to enable communication between the memory and the processor.

10. A computer-readable storage medium, characterized in that, Includes instructions that, when run on a computer, cause the computer to perform the method as described in any one of claims 1 to 7.

Images