Cold source system energy consumption prediction and state diagnosis method and device, electronic equipment and storage medium
By using a physically based interpretable graph attention network, combined with multi-source operating data and dynamic adjacency matrix of the cold source system, the problems of accuracy and interpretability of energy consumption prediction and status diagnosis of the cold source system are solved, and accurate prediction of energy consumption and effective diagnosis of operating status of the cold source system are achieved.
Patent Information
- Application Number
- CN202511820312.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-04
- Publication Date
- 2026-02-17
AI Technical Summary
Existing methods for predicting energy consumption and diagnosing the condition of cold source systems suffer from problems such as a lack of physical causal interpretation in data-driven methods, parameter drift in mechanistic models and high computational costs, and a lack of dynamic adaptability in graph learning methods, making it difficult to achieve accurate prediction and real-time optimization.
A physical mechanism-based interpretable graph attention network is adopted. By acquiring multi-source operating data of the cold source system, reference values are determined, and a dynamic adjacency matrix is constructed using the physical residuals. The physical mechanism-embedded interpretable graph attention network is then used for energy consumption prediction and state diagnosis. The prediction and diagnosis are performed by combining the current dynamic adjacency matrix and multi-source operating data.
It enables accurate prediction of energy consumption and effective diagnosis of operating status of cold source systems, outputs interpretable information, improves user experience, and provides support for energy efficiency optimization and fault early warning.
Smart Images
Figure CN121543439A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, electronic device, and storage medium for predicting energy consumption and diagnosing the condition of a cold source system. Background Technology
[0002] Central air conditioning and other cooling systems are key energy-consuming components in data centers and large public buildings. Improving the energy efficiency of cooling systems has become an essential path for enterprises to reduce operating costs and achieve green transformation.
[0003] Existing methods for predicting energy consumption and diagnosing the condition of cold source systems can be mainly divided into three categories: pure data-driven methods, pure mechanism model methods, and graph learning methods.
[0004] Purely data-driven methods primarily rely on various machine learning and deep learning algorithms to achieve prediction and optimization by learning from historical data. However, these models tend to learn statistical correlations between variables rather than physical causal relationships, resulting in insufficient interpretability. For example, an increase in pump frequency may physically indicate an increase in system load, not always ineffective energy consumption, but the model often fails to distinguish this causal logic. Moreover, these methods heavily depend on large-scale, high-quality historical data and have poor transferability across different cooling source systems. While reinforcement learning methods can optimize parameters through self-learning, they require a large amount of interactive data and need to be retrained for new cooling source system applications.
[0005] Pure mechanistic modeling methods primarily construct mechanistic models based on physical laws and prior knowledge of the system. However, mechanistic models require precise acquisition of equipment parameters such as heat transfer coefficients, pipe resistance, and compressor efficiency curves. These parameters drift with equipment aging and changes in operating conditions, making it difficult to maintain long-term validity. Furthermore, high-precision mechanistic models require substantial computational resources, which is insufficient for real-time optimization needs.
[0006] Graph learning methods often employ fixed adjacency matrices or construct topologies based on k-nearest neighbors, failing to reflect the impact of dynamic changes in operating conditions on coupling relationships. Traditional clustering methods cannot capture evolving characteristic relationships. They also lack physical semantics and dynamic adaptability. The meanings of nodes and edges are mostly statistical correlations, lacking physical interpretation. Summary of the Invention
[0007] This invention provides a method, apparatus, electronic device, and storage medium for predicting energy consumption and diagnosing the condition of a cold source system, in order to overcome the deficiencies in related technologies.
[0008] This invention provides a method for predicting energy consumption and diagnosing the condition of a cold source system, comprising: Obtain the current multi-source operating data of the cooling system; Based on the current multi-source operating data, determine the reference values of each physical parameter in the current multi-source operating data under the current operating conditions of the cold source system; Based on the physical mechanism of the cold source system, the reference values of each physical parameter are applied to calculate the current mechanism residual between each physical parameter, and based on the current mechanism residual, the current dynamic adjacency matrix of each physical parameter is constructed. Based on the current multi-source operating data and the current dynamic adjacency matrix, an interpretable graph attention network with embedded mechanism is applied to determine the predictive diagnostic results of the cold source system. The predicted diagnostic results include the current energy consumption prediction value of the cold source system, the current equipment status, and interpretability information.
[0009] According to the present invention, a method for predicting energy consumption and diagnosing the condition of a cold source system includes determining the predicted diagnostic result of the cold source system by applying a mechanism-embedded interpretable graph attention network based on the current multi-source operating data and the current dynamic adjacency matrix. Obtain a multi-source running data sequence containing the current multi-source running data for a preset duration; The multi-source running data sequence is input into the feature extraction model to obtain the current multi-source running features output by the feature extraction model; The current multi-source operating features and the current dynamic adjacency matrix are input into the interpretable graph attention network to obtain the prediction and diagnosis results output by the interpretable graph attention network.
[0010] According to the present invention, a method for predicting energy consumption and diagnosing the condition of a cold source system is provided, wherein the interpretable graph attention network is specifically used for: Based on the current multi-source operation characteristics and the current dynamic adjacency matrix, a graph attention mechanism is applied to calculate the attention coefficient between each node and its neighboring nodes in the current dynamic adjacency matrix. Based on the attention coefficient, the attention weight and the final update characteristics of each node in the current dynamic adjacency matrix are determined. Based on the final update features, the contribution weight of each node in the current dynamic adjacency matrix to the current energy consumption prediction value is determined, and the aggregation features are determined based on the contribution weight of each node in the current dynamic adjacency matrix to the current energy consumption prediction value. Based on the aggregation features, the current energy consumption prediction value and the current equipment status of the cold source system are determined; the interpretability information is determined based on at least one of the contribution weights of each node in the current dynamic adjacency matrix to the current energy consumption prediction value, the attention weights, the final update features, and the values of each edge in the current dynamic adjacency matrix.
[0011] According to the present invention, a method for predicting energy consumption and diagnosing the condition of a cold source system includes calculating the attention coefficient between each node and its neighboring nodes in the current dynamic adjacency matrix, comprising: For any node in the current dynamic adjacency matrix and the neighboring nodes of any node, determine the linear transformation result of the feature vector of any node; Based on the linear transformation result, the edge values between any node and its neighboring nodes, and the current mechanism residuals corresponding to the edge values, the attention coefficient between any node and its neighboring nodes is calculated.
[0012] According to the present invention, a method for predicting energy consumption and diagnosing the condition of a cold source system includes constructing a current dynamic adjacency matrix for each physical parameter based on the current mechanism residual, comprising: Convert the current mechanism residuals into weights; Based on the physical parameters, each node in the current dynamic adjacency matrix is determined, and the weight is used as the edge value between the corresponding nodes in the current dynamic adjacency matrix to construct the current dynamic adjacency matrix.
[0013] According to the present invention, a method for predicting energy consumption and diagnosing the condition of a cold source system is provided, wherein the interpretability information includes at least one of the following: the importance of each node in the current dynamic adjacency matrix to the current energy consumption prediction value, the influence strength between each node in the current dynamic adjacency matrix, and the change of each edge value.
[0014] According to the present invention, a method for predicting energy consumption and diagnosing the condition of a cold source system includes determining reference values for each physical parameter in the current multi-source operating data under the current operating condition of the cold source system, based on the current multi-source operating data. Based on the current multi-source operating data, determine the current load rate and current outdoor wet-bulb temperature of the cold source system, and based on the current load rate and current outdoor wet-bulb temperature, determine the current operating condition; Based on the mapping relationship between the offline determined working conditions and reference values, the reference values of each physical parameter under the current working condition are determined.
[0015] According to the present invention, a method for predicting energy consumption and diagnosing the condition of a cold source system is provided, wherein the interpretable graph attention network is updated online based on the following steps: Real-time determination of the energy consumption error of the current energy consumption prediction value; If the energy consumption error exceeds a preset threshold within a specified time period, the interpretable graph attention network is fine-tuned based on the historical energy consumption prediction value of the cold source system within the most recent preset time period and the actual energy consumption measurement value at the corresponding time.
[0016] According to the present invention, a method for predicting and diagnosing the energy consumption of a cold source system includes fine-tuning an interpretable graph attention network based on historical energy consumption predictions of the cold source system within a recent preset time period and actual energy consumption measurements at corresponding times. Obtain the historical mechanism residuals within the preset time period, and construct a mechanism residual penalty term based on the historical mechanism residuals; Calculate the historical energy consumption error between the historical energy consumption prediction value and the actual energy consumption measurement value, and calculate the fine-tuning loss based on the historical energy consumption error and the mechanism residual penalty term; The interpretable graph attention network is fine-tuned based on the fine-tuning loss.
[0017] The present invention also provides a device for predicting and diagnosing the energy consumption of a cold source system, comprising: The data acquisition module is used to acquire the current multi-source operating data of the cold source system; The reference value determination module is used to determine the reference values of each physical parameter in the current multi-source operating data under the current operating conditions of the cold source system, based on the current multi-source operating data. The adjacency matrix construction module is used to calculate the current mechanism residual between each physical parameter based on the physical mechanism of the cold source system, apply the reference values of each physical parameter, and construct the current dynamic adjacency matrix of each physical parameter based on the current mechanism residual. The predictive diagnostic module is used to determine the predictive diagnostic results of the cold source system based on the current multi-source operating data and the current dynamic adjacency matrix, by applying a mechanism-embedded interpretable graph attention network; wherein the predictive diagnostic results include the current energy consumption prediction value, interpretability information, and the current equipment status of the cold source system.
[0018] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the energy consumption prediction and status diagnosis method for a cold source system as described above.
[0019] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the energy consumption prediction and status diagnosis method for a cold source system as described above.
[0020] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the energy consumption prediction and status diagnosis method for a cold source system as described above.
[0021] The present invention provides a method, apparatus, electronic device, and storage medium for predicting energy consumption and diagnosing the condition of a cold source system. First, it acquires the current multi-source operating data of the cold source system. Then, using this data, it determines the reference values of each physical parameter in the current multi-source operating data under the current operating condition of the cold source system. Next, utilizing the physical mechanism of the cold source system, it calculates the current mechanism residuals between each physical parameter using the reference values, and constructs the current dynamic adjacency matrix for each physical parameter based on these residuals. Finally, combining the current multi-source operating data and the current dynamic adjacency matrix, it applies an interpretable graph attention network embedded with the mechanism to determine the predicted diagnostic result of the cold source system. This method considers the physical mechanism of the cold source system, combines the reference values of each physical parameter under the current operating condition, determines the current mechanism residuals, and then dynamically constructs the current dynamic adjacency matrix. This allows the current dynamic adjacency matrix to match the current operating condition, enabling the interpretable graph attention network to capture dynamic changes in the operating condition, thus achieving accurate prediction of energy consumption and effective diagnosis of the operating condition of the cold source system. Moreover, the predictive diagnostic results output by the interpretable graph attention network also include interpretability information, making the predictive diagnostic results physically interpretable, improving the user experience, and providing comprehensive support for energy efficiency optimization and fault early warning of cold source systems. Attached Figure Description
[0022] To more clearly illustrate the technical solutions in this invention or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0023] Figure 1 This is a flowchart illustrating the energy consumption prediction and status diagnosis method for cold source systems provided by the present invention.
[0024] Figure 2 This is a schematic diagram of the structure of the energy consumption prediction and status diagnosis device for cold source systems provided by the present invention.
[0025] Figure 3 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation
[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.
[0027] In existing technologies, purely data-driven methods suffer from problems such as black-box nature, insufficient interpretability, spurious correlations due to lack of physical logic, data dependence, and weak generalization ability. Although models such as LSTM and random forests can achieve high prediction accuracy, their decision-making processes are opaque and lack physical interpretability. For example, while AI adaptive adjustment methods for central air conditioning can achieve energy-saving control, they do not solve the problem of model interpretability. Recent research indicates that the closed data sources and black-box nature of traditional AI methods hinder their large-scale adoption in industries. Purely data-driven methods tend to learn statistical correlations between variables rather than physical causal relationships. For example, an increase in pump frequency may physically mean an increase in system load, not always ineffective energy consumption, but data-driven models often cannot distinguish this causal logic. Furthermore, these methods heavily rely on large-scale, high-quality historical data and have poor transferability between different systems. While data center cold source reinforcement learning methods can optimize parameters through self-learning, they require a large amount of interactive data and need to be retrained for application in new systems.
[0028] The limitations of pure mechanistic models lie in their complexity, parameter drift, computational cost, and poor real-time performance. Mechanistic models require precise acquisition of equipment parameters such as heat transfer coefficients, pipe resistance, and compressor efficiency curves. These parameters drift with equipment aging and changes in operating conditions, making it difficult to maintain long-term validity. Research on ice storage cooling system optimization indicates that physical models are highly complex and computationally time-consuming, making them unsuitable for online control. Furthermore, high-precision simulation models require substantial computational resources, making them insufficient for real-time optimization needs. For example, while dynamic physical simulation of the cooling source system in electronic cleanrooms is implemented on Dymola, its high computational cost hinders online applications.
[0029] Existing graph learning methods often employ fixed adjacency matrices or construct topologies based on k-nearest neighbors, failing to reflect the impact of dynamic changes in operating conditions and equipment degradation on coupling relationships. Furthermore, they suffer from limitations in physical semantics and dynamic adaptability. The meanings of nodes and edges are mostly statistical correlations, lacking physical interpretation and remaining a "black box" problem. While multi-agent protection schemes employ hierarchical cooperation, the physical meaning of interactions between agents is not clearly explained. They also struggle to adaptively adjust to changes in system operating status, equipment performance degradation, and external environment. Interpretable dynamic clustering research emphasizes that traditional methods cannot simultaneously optimize temporal continuity and context alignment constraints.
[0030] In summary, existing technologies lack a method for predicting energy consumption and diagnosing the condition of cold source systems that integrates the advantages of physical mechanisms and data-driven approaches, possesses dynamic graph structure learning capabilities, and is physically interpretable. Therefore, this invention provides a method for predicting energy consumption and diagnosing the condition of cold source systems, aiming to address the aforementioned key technical bottlenecks and promote the development of energy efficiency management for cold source systems to a higher level.
[0031] Figure 1 This is a flowchart illustrating a method for predicting and diagnosing the energy consumption of a cold source system provided in an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes: S1, obtain the current multi-source operation data of the cold source system; S2, Based on the current multi-source operating data, determine the reference values of each physical parameter in the current multi-source operating data under the current operating conditions of the cold source system; S3. Based on the physical mechanism of the cold source system, the reference values of each physical parameter are applied to calculate the current mechanism residual between each physical parameter, and based on the current mechanism residual, the current dynamic adjacency matrix of each physical parameter is constructed. S4. Based on the current multi-source operating data and the current dynamic adjacency matrix, apply the mechanism-embedded interpretable graph attention network to determine the predictive diagnosis result of the cold source system. The predicted diagnostic results include the current energy consumption prediction value of the cold source system, the current equipment status, and interpretability information.
[0032] Specifically, the energy consumption prediction and status diagnosis method for a cold source system provided in this embodiment of the invention is executed by a cold source system energy consumption prediction and status diagnosis device, which can be configured on a central data platform. The central data platform can be a local computer or a cloud computer, and the local computer can be a PC, without specific limitations here.
[0033] First, execute step S1 to obtain the current multi-source operating data of the cold source system. The cold source system can be a system with refrigeration function, such as a central air conditioning system. The cold source system may include multiple devices and the piping connecting each device. The devices in the cold source system may include chiller units, chilled water pumps, cooling water pumps, cooling tower fans, frequency converters, wet bulbs, etc., and the piping may include the chilled water main pipe, the cooling water main pipe, and key branch pipes, etc.
[0034] Current multi-source operational data refers to data synchronously collected at a frequency of no less than once per minute by each sensor in the sensor network at the current moment. The sensors in the sensor network are installed at different locations within the cold source system, and the types of sensors can include temperature sensors, power transmitters, and electromagnetic flow meters, etc.
[0035] Temperature sensors were installed at the chilled water supply inlet, chilled water return inlet, cooling water supply inlet, cooling water return inlet of each chiller unit, and in the outdoor environment to measure the chilled water supply temperature. Chilled water return temperature Cooling water supply temperature Cooling water return temperature and outdoor wet-bulb temperature .
[0036] By installing power transmitters at the power input terminals of each chiller unit, chilled water pump, and cooling water pump, the power of each chiller unit was measured. chilled water pump power Cooling water pump power .
[0037] Electromagnetic flow meters were installed on the chilled water main, cooling water main, and key branch lines to measure the chilled water flow rate. and cooling water flow rate .
[0038] In addition, the sensor network can also collect the start-up and shutdown status of chillers, water pumps, cooling tower fans, and frequency signals of inverters from the building automation system.
[0039] The sensor network will connect all the collected data to the central data platform via OPC UA or Modbus protocol. The central data platform will then add a unified timestamp to all the data to obtain the initial multi-source operational data.
[0040] To ensure data quality and the accuracy of subsequent processing results, the initial multi-source running data can be preprocessed to obtain the current multi-source running data, thus ensuring data quality. Preprocessing operations can include data cleaning, data filtering, and data standardization. Data cleaning can set thresholds based on physical feasibility to automatically identify and remove obvious outliers, such as... Higher than the chilled water supply temperature threshold chilled water flow rate Negative value, chiller unit power The data is outside the range specified on the chiller unit's nameplate. For transient data loss caused by sensor transient failures, linear interpolation is used to fill in the gaps.
[0041] Data filtering refers to the smoothing of all continuous variables, except for equipment status signals, by applying a first-order low-pass filter or a moving average window to eliminate measurement noise.
[0042] Data standardization refers to normalizing all continuous variables within the interval [0,1] to eliminate the influence of dimensions. The standardization calculation formula is: , and The theoretical or actual minimum and maximum values of this variable are determined based on the device's nameplate parameters and historical operating data.
[0043] The physical parameters in the current multi-source operation data can include chilled water supply temperature. Chilled water return temperature Cooling water supply temperature Cooling water return temperature Outdoor wet-bulb temperature chiller unit power chilled water pump power Cooling water pump power chilled water flow rate Cooling water flow rate and the cooling load of the cold source system .
[0044] Then, step S2 is executed to determine the reference values of each physical parameter in the current multi-source operating data under the current operating conditions of the cold source system, using the current multi-source operating data.
[0045] Different operating conditions of the cooling source system can be determined by the cooling source system's load rate (PLR) and the outdoor wet-bulb temperature. Characterization is performed. The load factor (PLR) of the cooling source system at time t can be determined by the ratio of the cooling load of the cooling source system at time t to the design cooling load, i.e.: ; in, It is the design cooling load of the cold source system, in kW. It is the cooling load of the cold source system at time t.
[0046] The cooling load of the cold source system at time t can be calculated using the following formula: ; in, It is the density of chilled water, usually taken as... . It is the specific heat capacity of water. For time t , For time t , For time t .
[0047] Therefore, the current load rate and the current outdoor wet-bulb temperature can be used to determine the current operating conditions.
[0048] Subsequently, the reference values of each physical parameter under the current operating condition can be determined by utilizing the mapping relationship between the offline determined operating conditions and the reference values.
[0049] This mapping relationship can be constructed offline using pre-processed multi-source operating data sequences for a specified time period, corresponding to various operating conditions, and reference values of various physical parameters determined from the multi-source operating data sequences for the specified time period through expert experience. Here, the reference values refer to the data of various physical parameters corresponding to the high energy efficiency indicators of the cooling source system. For multi-source operating data sequences for a specified time period, only efficient and healthy periods are selected, i.e., alarm / fault windows need to be removed; transition segments within 15-20 minutes of start-up / shutdown / switching need to be removed. The energy efficiency indicators of the cooling source system can be determined by taking values from positions such as the high percentile of COP or the low percentile of PUE in the data center.
[0050] For the current operating conditions, the reference values of each physical parameter can be determined by looking up the above mapping relationship, and a reference vector can be constructed: ; in, As the reference vector, This is a reference value for the chilled water supply temperature. This is a reference value for the temperature difference between chilled water and return water. This is a reference value for the cooling water supply temperature. This is a reference value for the cooling water return temperature difference. This is a reference value for chilled water flow rate. This is a reference value for cooling water flow rate. This is a reference value for compressor power. This is a reference value for the power of the chilled water pump. This is a reference value for the power of the cooling water pump. This is a reference value for cooling load. This is a reference value for outdoor wet-bulb temperature.
[0051] At the same time, in order to avoid different working conditions and There are abrupt jumps at the boundaries of the bins. A weighted average of the reference values for each physical parameter is taken to obtain a smoothed reference value: ; in, To smooth the reference value, It is the reference value for the i-th physical parameter, and the weight. Using a Gaussian kernel: ; in, and They are and The corresponding kernel bandwidth. The smaller the kernel bandwidth, the more it emphasizes nearest neighbors; the larger the kernel bandwidth, the smoother the surface, but the easier it is to over-smooth out details.
[0052] The mapping relationship can be presented in the form of a working condition lookup table. For example, Table 1 shows a partial working condition lookup table for a chiller unit in a cold source system.
[0053] Table 1. Partial Operating Condition Lookup Table
[0054] Subsequently, by utilizing the physical mechanisms of the cold source system and combining reference values from current multi-source operating data, the mechanistic residuals between various physical parameters can be calculated. The physical mechanisms of the cold source system can include chilled water pump energy balance, chilled water pump similarity law, cooling capacity balance, cooling water pump similarity law, cooling tower approach temperature, and unit efficiency consistency. Therefore, the mechanistic residuals can include chilled water pump residuals, chilled water pump residuals, cooling capacity residuals, cooling water pump residuals, cooling tower approach temperature residuals, and unit efficiency consistency residuals.
[0055] Frozen side energy residual It can be calculated using the following formula: ; Chilled water pump residual It can be calculated using the following formula: ; The residual of cooling capacity measurement can be calculated using the following formula: ; The residual of the cooling water pump can be calculated using the following formula: ; The temperature residual of the cooling tower can be calculated using the following formula: ; The unit efficiency consistency residual can be calculated using the following formula: .
[0056] Next, step S3 is executed to construct the dynamic adjacency matrix of the current multi-source running data using the mechanistic residuals. The construction strategy can be to use each physical parameter as a node in the current dynamic adjacency matrix and the mechanistic residuals as the edge values between corresponding nodes in the current dynamic adjacency matrix, thus obtaining the current dynamic adjacency matrix. The current dynamic adjacency matrix is an n×n matrix, where n is the number of physical parameters.
[0057] It is understandable that the current dynamic adjacency matrix is a condition-adaptive dynamic graph structure.
[0058] Finally, in step S4, the current multi-source running features and dynamic adjacency matrix of the current multi-source running data can be input into the Physics-Guided Interpretable Graph Attention Network (PGI-GAT) to obtain the prediction and diagnosis results output by the Physics-Guided Interpretable Graph Attention Network.
[0059] The attention mechanism of interpretable graph attention networks can assign different importance weights to neighboring nodes, which is highly consistent with the strength of physical relationships expressed by dynamic edge weights, thus enabling a more refined capture of the system's internal state. Here, the predictive diagnostic results include current energy consumption predictions, interpretability information, and the current equipment status of the cooling system. The current equipment status can include alarm and normal states. The predictive diagnostic results can be packaged into a comprehensive report for output.
[0060] The energy consumption prediction and status diagnosis method for a cold source system provided in this embodiment of the invention first acquires the current multi-source operating data of the cold source system; then, using the current multi-source operating data, it determines the reference values of each physical parameter in the current multi-source operating data under the current operating condition of the cold source system; subsequently, using the physical mechanism of the cold source system, it calculates the current mechanism residual between each physical parameter by applying the reference values of each physical parameter, and constructs the current dynamic adjacency matrix of each physical parameter based on the current mechanism residual; finally, it combines the current multi-source operating data and the current dynamic adjacency matrix, and applies an interpretable graph attention network embedded in the mechanism to determine the prediction and diagnosis result of the cold source system. This method considers the physical mechanism of the cold source system, combines the reference values of each physical parameter under the current operating condition, determines the current mechanism residual, and then dynamically constructs the current dynamic adjacency matrix. This allows the current dynamic adjacency matrix to match the current operating condition, enabling the interpretable graph attention network to capture dynamic changes in the operating condition, thereby achieving accurate prediction of energy consumption and effective diagnosis of the operating status of the cold source system. Moreover, the predictive diagnostic results output by the interpretable graph attention network also include interpretability information, making the predictive diagnostic results physically interpretable, improving the user experience, and providing comprehensive support for energy efficiency optimization and fault early warning of cold source systems.
[0061] Based on the above embodiments, the step of determining the predictive diagnostic results of the cold source system by applying a mechanism-embedded interpretable graph attention network based on the current multi-source operating data and the current dynamic adjacency matrix includes: Obtain a multi-source running data sequence containing the current multi-source running data for a preset duration; The multi-source running data sequence is input into the feature extraction model to obtain the current multi-source running features output by the feature extraction model; The current multi-source operating features and the current dynamic adjacency matrix are input into the interpretable graph attention network to obtain the prediction and diagnosis results output by the interpretable graph attention network.
[0062] Specifically, when determining the predictive diagnostic results of the cold source system, a multi-source operation data sequence containing current multi-source operation data for a preset duration can be obtained first. This preset duration can be set as needed and is not specifically limited here. The multi-source operation data sequence can be determined using a window with a width equal to the preset duration.
[0063] Subsequently, the multi-source runtime data sequence is input into the feature extraction model to obtain the current multi-source runtime features output by the feature extraction model. The feature extraction model can be a Gated Recurrent Unit (GRU). The current multi-source runtime features can be represented as: ; in, Given the current characteristics of multi-source operation, For multi-source running data sequences, w is the preset duration.
[0064] Finally, the current multi-source operating features and the current dynamic adjacency matrix are input into the interpretable graph attention network. The interpretable graph attention network uses the graph attention mechanism to jointly analyze the current multi-source operating features and the current dynamic adjacency matrix, and obtains and outputs the prediction and diagnosis results.
[0065] In this embodiment of the invention, by determining the current multi-source operating features and using them together with the current dynamic adjacency matrix as input to the interpretable graph attention network, the computational load of the interpretable graph attention network can be reduced, thus facilitating improved prediction and diagnosis efficiency.
[0066] Based on the above embodiments, the interpretable graph attention network is specifically used for: Based on the current multi-source operation characteristics and the current dynamic adjacency matrix, a graph attention mechanism is applied to calculate the attention coefficient between each node and its neighboring nodes in the current dynamic adjacency matrix. Based on the attention coefficient, the attention weight and the final update characteristics of each node in the current dynamic adjacency matrix are determined. Based on the final update features, the contribution weight of each node in the current dynamic adjacency matrix to the current energy consumption prediction value is determined, and the aggregation features are determined based on the contribution weight of each node in the current dynamic adjacency matrix to the current energy consumption prediction value. Based on the aggregation features, the current energy consumption prediction value and the current equipment status of the cold source system are determined; the interpretability information is determined based on at least one of the contribution weights of each node in the current dynamic adjacency matrix to the current energy consumption prediction value, the attention weights, the final update features, and the values of each edge in the current dynamic adjacency matrix.
[0067] Specifically, an interpretable graph attention network can include multiple attention layers, each performing the same operations. For each node i in the current dynamic adjacency matrix, the information aggregation and feature update process of each attention layer includes: The graph attention mechanism is applied to calculate the attention coefficient between each node and its neighbors in the current dynamic adjacency matrix. The attention coefficient is the similarity value before normalization, and can be the exponential part of a Gaussian kernel. Subsequently, the attention coefficient is normalized to obtain the attention weights.
[0068] ; in, The attention weights between node i and its neighbor node j at time t are calculated by the h-th attention head of attention layer l. The attention coefficient between node i at time t and its neighbor node j is calculated by the h-th attention head of attention layer l. The attention coefficient between node i at time t and its neighbor node k is calculated by the h-th attention head of attention layer l. Let be the number of neighboring nodes of node i at time t, which is shared across all attention layers. .
[0069] By weighting and summing the inputs of the current attention layer using attention weights, we can obtain the updated features of the current attention layer's output, which include: , , ; in, The updated features are the output of the h-th attention head of attention layer l. Here, || represents the activation function, || represents the concatenation operation, and H represents the total number of attention heads in attention layer l. It is the aggregated representation of the h-th attention head pair of node i at time t in attention layer l. Let i be the feature vector of node i at time t, which is input to attention layer l. Let h be the linear transformation weight of the h-th attention head in attention layer l. Let h be the linear transformation result of the h-th attention head of attention layer l corresponding to node i.
[0070] The feature vector of node i at time t output by the last attention layer is the final updated feature of node i in the current dynamic adjacency matrix.
[0071] Understandably, each attention layer can also perform a small, learnable scaling of the current dynamic adjacency matrix to accommodate data adaptation.
[0072] Subsequently, the output layer of the interpretable graph attention network uses the final updated features to determine the output. The output layer can obtain contribution weights through fully connected layers. : ; Here, n represents the number of each physical parameter, which is also the number of nodes. The final updated feature of node i at time t is the output of attention layer L. The final updated feature of node k at time t is the output of attention layer L. These are the output layer weights. This contribution weight... The weights contribute to the global calculation and are used to explain which physical parameters dominate the energy consumption of the cold source system at that moment.
[0073] Aggregated features output by the output layer It can be represented as: .
[0074] Finally, the aggregated features are passed through two linear layers to obtain the current energy consumption prediction value and the current equipment status of the cold source system.
[0075] In the above process, the contribution weights of each node in the current dynamic adjacency matrix to the current energy consumption prediction value, the attention weights, the final update features, and the values of each edge in the current dynamic adjacency matrix can all be used to determine interpretability information.
[0076] Based on the above embodiments, the interpretability information includes at least one of the following: the importance of each node in the current dynamic adjacency matrix to the current energy consumption prediction value, the influence strength between each node in the current dynamic adjacency matrix, and the change in the value of each edge.
[0077] Specifically, the contribution weights of each node in the current dynamic adjacency matrix to the current energy consumption prediction value, as well as the final update features, can be used to determine the importance of each node to the current energy consumption prediction value and to rank them by importance. For example, the L2 norm of the contribution weights of each node, or the final update features of each node, can be used to characterize the importance of each node to the current energy consumption prediction value.
[0078] For example, regarding changes in the values of each edge, if the edge weight between the "cooling water pump power" node and the "cooling tower outlet water temperature" node is consistently significantly lower than the historical normal level, the diagnosis suggests that "the heat exchange efficiency of the cooling tower may have decreased and needs to be checked."
[0079] Based on the above embodiments, calculating the attention coefficient between each node and its neighboring nodes in the current dynamic adjacency matrix includes: For any node in the current dynamic adjacency matrix and the neighboring nodes of any node, determine the linear transformation result of the feature vector of any node; Based on the linear transformation result, the edge values between any node and its neighboring nodes, and the current mechanism residuals corresponding to the edge values, the attention coefficient between any node and its neighboring nodes is calculated.
[0080] Specifically, when calculating the attention coefficient, the linear transformation result of the feature vector of each node, the edge values between each node and its neighbors, and the current mechanism residual corresponding to the edge values can be combined to calculate the attention coefficient.
[0081] ; in, For feature similarity terms, For mechanism residual bias, This is the adjacent prior bias. The attention coefficients between node i and its neighbor node j at time t are calculated by the h-th attention head of attention layer l, and LeakyReLU is the activation function. , and All are learnable parameters. Let h be the linear transformation result of the h-th attention head of attention layer l corresponding to its neighbor node j. Based on the current mechanism residuals corresponding to node i and its neighbor node j The additive bias obtained from the transformation This is a priori bias. It is a constant. The values are assigned to the edges corresponding to node i and its neighbor node j in the current dynamic adjacency matrix. The multiplicative integration corresponding to "data likelihood × mechanism prior" is highly interpretable.
[0082] Based on the above embodiments, the step of constructing the current dynamic adjacency matrix of each physical parameter based on the current mechanism residual includes: Convert the current mechanism residuals into weights; Based on the physical parameters, each node in the current dynamic adjacency matrix is determined, and the weight is used as the edge value between the corresponding nodes in the current dynamic adjacency matrix to construct the current dynamic adjacency matrix.
[0083] Specifically, when constructing the current dynamic adjacency matrix, the current mechanism residuals can first be converted into weights. The formula for calculating the weights is: ; in, Let i be the weights of its neighbor j. Let i be the current mechanism residual corresponding to node i and its neighbor node j. Let be the tolerance scale for the edge between node i and its neighbor node j. This is the sensitivity factor.
[0084] Subsequently, each physical parameter can be used as a node in the current dynamic adjacency matrix, and the weights of node i and its neighbor j calculated above can be used as the edge values between node i and its neighbor j in the current dynamic adjacency matrix to construct the current dynamic adjacency matrix.
[0085] Assume the current multi-source running data is as follows: The reference values for each physical parameter are as follows: .
[0086] and Take 0.05, Taking 1.0, the current dynamic adjacency matrix can be represented as: .
[0087] The order of the physical parameters corresponding to each row or column of the current dynamic adjacency matrix is as follows: .
[0088] Based on the above embodiments, the interpretable graph attention network is updated online according to the following steps: Real-time determination of the energy consumption error of the current energy consumption prediction value; If the energy consumption error exceeds a preset threshold within a specified time period, the interpretable graph attention network is fine-tuned based on the historical energy consumption prediction value of the cold source system within the most recent preset time period and the actual energy consumption measurement value at the corresponding time.
[0089] Specifically, in this embodiment of the invention, the interpretable graph attention network can be updated online. During the online update process, the energy consumption error of the current energy consumption prediction value can be determined in real time, that is, the difference between the current energy consumption prediction value and the current actual energy consumption measurement value.
[0090] If the energy consumption error exceeds the preset threshold within a specified time period, the historical energy consumption error between the historical energy consumption prediction value of the cold source system in the most recent preset time period and the actual energy consumption measurement value at the corresponding time is used to fine-tune the interpretable graph attention network, so as to efficiently adapt to the new characteristics of the cold source system and ensure the long-term accuracy of prediction and diagnosis.
[0091] The specified duration can be set as needed, for example, it can be set to 24 hours. The preset time period can also be set as needed, for example, it can be set to the last 30 days.
[0092] Based on the above embodiments, the fine-tuning of the interpretable graph attention network based on the historical energy consumption prediction values of the cold source system within the most recent preset time period and the actual energy consumption measurement values at the corresponding times includes: Obtain the historical mechanism residuals within the preset time period, and construct a mechanism residual penalty term based on the historical mechanism residuals; Calculate the historical energy consumption error between the historical energy consumption prediction value and the actual energy consumption measurement value, and calculate the fine-tuning loss based on the historical energy consumption error and the mechanism residual penalty term; The interpretable graph attention network is fine-tuned based on the fine-tuning loss.
[0093] Specifically, when fine-tuning the interpretable graph attention network, historical mechanism residuals within a preset time period can be obtained first. These residuals, combined with a norm, are then used to construct a mechanism residual penalty term. Simultaneously, the historical energy consumption error between predicted and actual energy consumption values is calculated. The fine-tuning loss is then calculated by combining the historical energy consumption error and the mechanism residual penalty term. This fine-tuning loss is then used to iteratively train the interpretable graph attention network until the fine-tuning loss converges or a specified number of iterations is reached.
[0094] In this embodiment of the invention, a mechanistic residual penalty term is introduced when calculating the fine-tuning loss, which can ensure that the prediction results maintain physical consistency such as energy conservation and pump similarity law. like Figure 2As shown, based on the above embodiments, this embodiment of the invention provides a cold source system energy consumption prediction and status diagnosis device, including: Data acquisition module 21 is used to acquire the current multi-source operating data of the cold source system; The reference value determination module 22 is used to determine the reference values of each physical parameter in the current multi-source operating data under the current operating conditions of the cold source system based on the current multi-source operating data. The adjacency matrix construction module 23 is used to calculate the current mechanism residual between each physical parameter based on the physical mechanism of the cold source system, apply the reference values of each physical parameter, and construct the current dynamic adjacency matrix of each physical parameter based on the current mechanism residual. The prediction and diagnosis module 24 is used to determine the prediction and diagnosis results of the cold source system based on the current multi-source operating data and the current dynamic adjacency matrix, by applying a mechanism-embedded interpretable graph attention network; wherein the prediction and diagnosis results include the current energy consumption prediction value, interpretability information, and the current equipment status of the cold source system.
[0095] Based on the above embodiments, the cold source system energy consumption prediction and status diagnosis device provided in this embodiment of the invention, wherein the prediction and diagnosis module is specifically used for: Obtain a multi-source running data sequence containing the current multi-source running data for a preset duration; The multi-source running data sequence is input into the feature extraction model to obtain the current multi-source running features output by the feature extraction model; The current multi-source operating features and the current dynamic adjacency matrix are input into the interpretable graph attention network to obtain the prediction and diagnosis results output by the interpretable graph attention network.
[0096] Based on the above embodiments, the cold source system energy consumption prediction and status diagnosis device provided in this embodiment of the invention, wherein the interpretable graph attention network is specifically used for: Based on the current multi-source operation characteristics and the current dynamic adjacency matrix, a graph attention mechanism is applied to calculate the attention coefficient between each node and its neighboring nodes in the current dynamic adjacency matrix. Based on the attention coefficient, the attention weight and the final update characteristics of each node in the current dynamic adjacency matrix are determined. Based on the final update features, the contribution weight of each node in the current dynamic adjacency matrix to the current energy consumption prediction value is determined, and the aggregation features are determined based on the contribution weight of each node in the current dynamic adjacency matrix to the current energy consumption prediction value. Based on the aggregation features, the current energy consumption prediction value and the current equipment status of the cold source system are determined; the interpretability information is determined based on at least one of the contribution weights of each node in the current dynamic adjacency matrix to the current energy consumption prediction value, the attention weights, the final update features, and the values of each edge in the current dynamic adjacency matrix.
[0097] Based on the above embodiments, the energy consumption prediction and status diagnosis device for a cold source system provided in this embodiment of the invention includes, in calculating the attention coefficient between each node and its neighboring nodes in the current dynamic adjacency matrix, the following steps: For any node in the current dynamic adjacency matrix and the neighboring nodes of any node, determine the linear transformation result of the feature vector of any node; Based on the linear transformation result, the edge values between any node and its neighboring nodes, and the current mechanism residuals corresponding to the edge values, the attention coefficient between any node and its neighboring nodes is calculated.
[0098] Based on the above embodiments, the adjacency matrix construction module of the cold source system energy consumption prediction and status diagnosis device provided in this embodiment of the invention is specifically used for: Convert the current mechanism residuals into weights; Based on the physical parameters, each node in the current dynamic adjacency matrix is determined, and the weight is used as the edge value between the corresponding nodes in the current dynamic adjacency matrix to construct the current dynamic adjacency matrix.
[0099] Based on the above embodiments, the cold source system energy consumption prediction and status diagnosis device provided in the embodiments of the present invention includes at least one of the following: the importance of each node in the current dynamic adjacency matrix to the current energy consumption prediction value, the influence strength between each node in the current dynamic adjacency matrix, and the change of each edge value.
[0100] Based on the above embodiments, the energy consumption prediction and status diagnosis device for a cold source system provided in this embodiment of the invention, wherein the reference value determination module is specifically used for: Based on the current multi-source operating data, determine the current load rate and current outdoor wet-bulb temperature of the cold source system, and based on the current load rate and current outdoor wet-bulb temperature, determine the current operating condition; Based on the mapping relationship between the offline determined working conditions and reference values, the reference values of each physical parameter under the current working condition are determined.
[0101] Based on the above embodiments, the cold source system energy consumption prediction and status diagnosis device provided in this embodiment of the invention further includes an online update module, used for: Real-time determination of the energy consumption error of the current energy consumption prediction value; If the energy consumption error exceeds a preset threshold within a specified time period, the interpretable graph attention network is fine-tuned based on the historical energy consumption prediction value of the cold source system within the most recent preset time period and the actual energy consumption measurement value at the corresponding time.
[0102] Based on the above embodiments, the online update module of the cold source system energy consumption prediction and status diagnosis device provided in this embodiment of the invention is specifically used for: Obtain the historical mechanism residuals within the preset time period, and construct a mechanism residual penalty term based on the historical mechanism residuals; Calculate the historical energy consumption error between the historical energy consumption prediction value and the actual energy consumption measurement value, and calculate the fine-tuning loss based on the historical energy consumption error and the mechanism residual penalty term; The interpretable graph attention network is fine-tuned based on the fine-tuning loss.
[0103] Specifically, the functions of each module in the cold source system energy consumption prediction and status diagnosis device provided in this embodiment of the invention correspond one-to-one with the operation flow of each step in the above method-like embodiments, and the achieved effects are also the same. For details, please refer to the above embodiments, and this will not be repeated in this embodiment of the invention.
[0104] Figure 3 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 3 As shown, the electronic device may include a processor 810, a communications interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communications interface 820, and the memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute the energy consumption prediction and status diagnosis methods for the cold source system provided in the above embodiments.
[0105] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to related technologies, or a 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 several 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 described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0106] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the cold source system energy consumption prediction and status diagnosis method provided in the above embodiments.
[0107] In another aspect, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the energy consumption prediction and status diagnosis method for the cold source system provided in the above embodiments. This computer-readable storage medium can be either a non-transitory computer-readable storage medium or a transient computer-readable storage medium, and is not specifically limited herein.
[0108] The device embodiments described above are merely illustrative. The units described 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 modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.
[0109] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of software products. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.
[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for predicting energy consumption and diagnosing a state of a cold source system, characterized by, The method comprises: obtaining current multi-source operation data of a cold source system; based on the current multi-source operation data, determining reference values of each physical parameter in the current multi-source operation data under the current working condition of the cold source system; based on the physical mechanism of the cold source system, applying the reference values of each physical parameter to calculate the current mechanism residual between each physical parameter, and based on the current mechanism residual, constructing a current dynamic adjacency matrix of each physical parameter; based on the current multi-source operation data and the current dynamic adjacency matrix, applying a mechanism-embedded interpretable graph attention network to determine a predicted diagnosis result of the cold source system; wherein the predicted diagnosis result comprises a current energy consumption prediction value, a current device state and an interpretability information of the cold source system.
2. The method of claim 1, wherein The method comprises: obtaining a multi-source operation data sequence containing the current multi-source operation data for a preset time length; inputting the multi-source operation data sequence into a feature extraction model to obtain current multi-source operation features output by the feature extraction model; inputting the current multi-source operation features and the current dynamic adjacency matrix into the interpretable graph attention network to obtain the predicted diagnosis result output by the interpretable graph attention network.
3. The method of claim 2, wherein The interpretable graph attention network is specifically used for: based on the current multi-source operation features and the current dynamic adjacency matrix, applying a graph attention mechanism to calculate an attention coefficient between each node and a neighbor node in the current dynamic adjacency matrix, and based on the attention coefficient, determining an attention weight and a final updated feature of each node in the current dynamic adjacency matrix; based on the final updated feature, determining a contribution weight of each node in the current dynamic adjacency matrix to the current energy consumption prediction value, and based on the contribution weight of each node in the current dynamic adjacency matrix to the current energy consumption prediction value, determining an aggregated feature; based on the aggregated feature, determining the current energy consumption prediction value and the current device state of the cold source system; the interpretability information is determined based on at least one of the contribution weight of each node in the current dynamic adjacency matrix to the current energy consumption prediction value, the attention weight, the final updated feature, and the edge value in the current dynamic adjacency matrix.
4. The method of claim 3, wherein the energy consumption of the cold source system is predicted and diagnosed based on the state of the cold source system. The method comprises: for any node and neighbor node in the current dynamic adjacency matrix, determining a linear transformation result of a feature vector of the any node; based on the linear transformation result, an edge value between the any node and the neighbor node, and the current mechanism residual corresponding to the edge value, calculating an attention coefficient between the any node and the neighbor node.
5. The method of claim 1, wherein the method further comprises: The method comprises: converting the current mechanism residual into a weight; The current dynamic adjacency matrix is constructed based on each of the physical parameters, and the weight is taken as the value of an edge between corresponding nodes in the current dynamic adjacency matrix.
6. The method of claim 1, wherein The interpretability information includes at least one of the importance of each node in the current dynamic adjacency matrix to the current energy consumption prediction value, the influence strength between each node in the current dynamic adjacency matrix, and the change of the value of each edge.
7. The method of claim 1-6, wherein The reference values of each physical parameter in the current multi-source operation data under the current operating condition of the cold source system are determined based on the current multi-source operation data, including: The current load rate and the current outdoor wet-bulb temperature of the cold source system are determined based on the current multi-source operation data, and the current operating condition is determined based on the current load rate and the current outdoor wet-bulb temperature. The reference values of each of the physical parameters under the current operating condition are determined based on the mapping relationship between the offline determined operating conditions and reference values.
8. The method of claim 1-6, wherein, The interpretable graph attention network is updated online based on the following steps: The energy consumption error of the current energy consumption prediction value is determined in real time; If the energy consumption error exceeds the preset threshold within a specified time period, the interpretable graph attention network is fine-tuned based on the historical energy consumption prediction values of the cold source system and the actual energy consumption measurement values at the corresponding time within the recent preset time period.
9. The method for predicting energy consumption and diagnosing the condition of a cold source system according to claim 8, characterized in that, The interpretable graph attention network is fine-tuned based on the historical energy consumption prediction values of the cold source system and the actual energy consumption measurement values at the corresponding time within the recent preset time period, including: The historical mechanism residual errors within the preset time period are obtained, and a mechanism residual error penalty term is constructed based on the historical mechanism residual errors; The historical energy consumption error between the historical energy consumption prediction values and the actual energy consumption measurement values is calculated, and a fine-tuning loss is calculated based on the historical energy consumption error and the mechanism residual error penalty term; The interpretable graph attention network is fine-tuned based on the fine-tuning loss.
10. A cold source system energy consumption prediction and state diagnosis device, characterized by, It includes: A data acquisition module for acquiring current multi-source operation data of a cold source system; A reference value determination module for determining reference values of each physical parameter in the current multi-source operation data under the current operating condition of the cold source system based on the current multi-source operation data; An adjacency matrix construction module for calculating the current mechanism residual error between each of the physical parameters based on the reference values of each of the physical parameters and the physical mechanism of the cold source system, and constructing a current dynamic adjacency matrix of each of the physical parameters based on the current mechanism residual error; A prediction and diagnosis module for determining a prediction and diagnosis result of the cold source system by applying a mechanism-embedded interpretable graph attention network based on the current multi-source operation data and the current dynamic adjacency matrix, wherein the prediction and diagnosis result includes a current energy consumption prediction value, interpretability information, and a current device state of the cold source system.
11. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the cold source system energy consumption prediction and state diagnosis method of any one of claims 1-9.
12. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by a processor to implement the cold source system energy consumption prediction and state diagnosis method according to any one of claims 1-9.
13. A computer program product comprising a computer program, characterized in that, The computer program is executed by a processor to implement the cold source system energy consumption prediction and state diagnosis method according to any one of claims 1-9.