A super-low ring temperature carbon dioxide-based combined supply energy control system

Abstract

The application relates to the technical field of energy control, and provides a two-supply energy control system based on ultra-low ambient temperature carbon dioxide, which comprises a main circulation unit, an auxiliary circulation unit, a heat exchange waterway module and a controller; the main circulation adopts a CO2 heating loop, and the auxiliary circulation adopts an R134a condensation supercooling loop; the controller is based on a thermal topology to construct a directed graph structure, adopts a three-branch convolution mechanism to model an energy transmission path, combines interval-contrast loss functions to optimize control embedding representation, generates accurate control instructions to adjust components such as compressors and throttling valves; in terms of fault diagnosis, a time parameterized convolution-symmetrical GRU model is constructed, a Legendre polynomial is used to model a time convolution kernel, a hidden state path is reconstructed by combining a Verlet integral mechanism, and accurate identification of abnormal trends in multi-source data is realized; the system improves the energy supply stability and intelligent level in an extremely low temperature environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of energy control, and particularly relates to a dual-supply energy control system based on ultra-low ambient temperature carbon dioxide. BACKGROUND

[0002] The cold and heat combined supply system based on natural working medium carbon dioxide gradually becomes an important direction for upgrading building energy systems. CO2 is particularly suitable for the ultra-low ambient temperature heating demand in cold regions due to its non-toxic, non-polluting, extremely low GWP, excellent critical thermodynamic performance and other characteristics. In the prior art, some systems have introduced auxiliary working medium such as R134a to construct a double-cycle structure to improve the heat pump efficiency under low temperature conditions. However, there are still many technical difficulties to be broken through. Firstly, the heat exchange process between the double cycles has strong dynamic coupling, and the current system generally lacks effective coordinated control mechanism, which is easy to cause energy redundancy or uneven heating. Secondly, although some studies attempt to introduce fuzzy control, expert system or reinforcement learning method to adjust the operating state, due to the complex thermodynamic topology structure, dynamic change of energy transmission path, the traditional intelligent control method still has limitations in modeling expression ability, path dependence capture and multi-scale energy coupling feature extraction, and cannot realize fine scheduling of key control quantities such as compressor frequency, throttle valve opening degree and heat exchanger switching path. Thirdly, for system faults, the existing method mainly relies on the alarm mechanism based on static rules or shallow machine learning model for state judgment, which is difficult to adapt to high-dimensional time sequence feature modeling under asynchronous acquisition of multi-source sensors, leading to untimely identification of early faults such as compressor performance degradation, throttle valve jamming and heat exchanger blockage, and response lag, which has great operation risk. SUMMARY

[0003] The purpose of the present application is to provide a kind of based on ultra-low ambient temperature carbon dioxide's double supply energy control system, to solve the existing CO2 Energy supply system in extreme low temperature environment heating performance instability, double cycle system control coordination is poor, traditional intelligent control model generalization ability is insufficient and fault response lag etc., realize efficient collaborative control and early fault intelligent diagnosis in heating process;System innovatively introduces the thermal topology directed graph modeling method, abstracts each heat exchange and refrigeration component as graph node, constructs directed edge relationship in combination with refrigerant flow direction, energy propagation dependence on edge, entry edge and self-loop path is modeled respectively through three branch convolution mechanism, capture the thermodynamic synergy characteristics of multi-hop path in deep graph convolution structure, and through interval-contrast loss optimization graph embedding expression, the adaptability of control strategy to system dynamic structure change is enhanced;In the aspect of fault diagnosis, construct time parameterized convolution-symmetric GRU model, adopt Legendre polynomial to model time function for convolution kernel, extract multi-scale feature trend;Combined with Verlet integral mechanism to reconstruct GRU hidden state update path, realize the time symmetry and high stability of fault identification model;The above scheme not only improves the energy supply stability and regulation precision in extreme low temperature environment, also enhances the response ability and prediction performance of system to early operation abnormality.

[0004] The present application provides a kind of based on ultra-low ambient temperature carbon dioxide's double supply energy control system, the system includes: main cycle unit, auxiliary cycle unit, heat exchange waterway module and controller;

[0005] Main cycle unit, acquisition main cycle operating state data;

[0006] Auxiliary cycle unit, acquisition auxiliary cycle operating state data;

[0007] Heat exchange waterway module, including heating outlet water line and domestic hot water outlet water line;Acquisition heat exchange waterway operating state data;

[0008] Controller, including energy supply control module and fault diagnosis module;

[0009] Energy supply control module, according to the physical connection relationship of main cycle unit and auxiliary cycle unit, constructs system thermal topology graph;Adopt directed graph contrast energy supply control model to process system thermal topology graph, obtain main cycle operating parameter and auxiliary cycle parameter, generate energy supply control instruction;Directed graph contrast energy supply control model includes directed graph convolutional neural network;

[0010] Fault diagnosis module, combine heat exchange waterway operating state data, main cycle operating state data and auxiliary cycle operating state data, form multidimensional time series data;Adopt time parameterized convolution-symmetric GRU model to process multidimensional time series data for fault identification.

[0011] Further, the process of generating energy supply control instructions by using a directed graph to process the system thermal topology graph compared with the energy supply control model includes the following steps:

[0012] Step S1: Collect the thermal state data of the current nodes and edges of the system thermal topology graph, and construct an initial graph feature representation; the initial graph feature representation includes node features and edge features;

[0013] Step S2: Process the initial graph feature representation using a directed graph convolutional neural network to generate a high-quality embedding representation;

[0014] Step S3: Based on the high-quality embedding representation, generate the main loop operating parameters and auxiliary loop parameters through a multi-layer perception structure; the operating parameters are issued as energy supply control instructions to each execution component.

[0015] Further, step S2 includes the following steps:

[0016] Step S21: Based on the directional connection relationship between nodes in the initial graph feature representation, construct an in-edge adjacency matrix and an out-edge adjacency matrix respectively, which are used to distinguish the energy receiving direction and energy releasing direction in the heat transfer process, and obtain a directional adjacency matrix; combine the node features and the directional adjacency matrix to construct a directed graph structure;

[0017] Step S22: For the directed graph structure, a three-branch propagation mechanism is used to perform convolutional feature propagation on the out-edge direction, in-edge direction and self-loop path of each node, respectively, to extract their energy output and receiving dependency in the heat flow process; a direction weight factor is introduced during the propagation process to assign a learnable weighting coefficient to the out-edge direction, in-edge direction and self-loop path; the above three-branch propagation mechanism is recursively unfolded in a directed graph neural network structure containing 5 layers, and through each layer, the features carried by different propagation paths (including out-edge path, in-edge path and node self-loop path) are extracted layer by layer to capture the topological dependency between cross-direction and multi-hop paths, and finally obtain a set of multi-path global embedding vectors;

[0018] Step S23: Differentiate the embedding distance between positive and negative path pairs in the set of multi-path global embedding vectors to construct an interval-contrast loss function; through interval-contrast loss function, the set of multi-path global embedding vectors is learned by contrast, and by maximizing the consistency and discriminability of different path embeddings, a high-quality embedding representation is generated.

[0019] Further, the process of processing multi-dimensional time series data for fault identification using a time parameterized convolution-symmetric GRU model includes the following steps:

[0020] Step B1: define a convolution kernel group for capturing change trends at different time scales, introduce a time continuity weight function for each convolution kernel weight of the convolution kernel group, model the convolution kernel weight as a time function form, and use Legendre polynomials for low-dimensional parameterization to obtain a time parameterized convolution kernel; use the time parameterized convolution kernel to extract initial features of the multi-dimensional time series data to generate multi-scale feature data;

[0021] Step B2: process the multi-scale feature data through a self-attention fusion mechanism to highlight the feature channels most sensitive to fault identification at different time scales to generate fused multi-scale time series features;

[0022] Step B3: input the fused multi-scale time series features into a symmetric BiGRU variant model to extract forward and backward dependencies in the time series data, locate the fault components, including compressor efficiency decline, throttling valve jam, heat exchanger blockage and sensor failure, and perform fault identification.

[0023] The beneficial effects achieved by the above scheme are as follows:

[0024] The present application introduces a thermal topology directed graph modeling and a three-branch convolution propagation mechanism, which first maps the complex thermal structure of the ultra-low ring temperature CO2 dual supply energy system into a directional graph structure, fully expresses the direction dependence and path diversity in the energy transmission process on the basis of node-edge collaborative modeling, and effectively solves the technical bottlenecks of topology relationship ambiguity and control strategy rigidity in the operation and scheduling of multi-node multi-source coupled systems compared with the existing control methods which do not fully express the system structure, realizes accurate mapping from physical structure to control space, and greatly improves the dynamic regulation and control capability of the system under load fluctuation and temperature disturbance, laying a precise foundation for the control strategy generation of the present application.

[0025] Further, the directed graph contrast energy control mechanism proposed by the present application learns the difference between path embedding representations by introducing interval-contrast loss function, so that the control model can actively identify the semantic similarity and scheduling priority between different energy paths in the embedding space, significantly enhancing the discrimination ability and instruction generation accuracy of the control system under complex thermal state; compared with the generalization barrier of traditional neural networks in control expression ability, the present application shows adaptability and migratability to multi-scenario thermal topology in actual deployment, can accurately adjust the frequency of the compressor, the opening of the throttling valve and the switching valve of the heat exchanger and other key executive components, and significantly improves the control accuracy and collaborative stability in the heating process.

[0026] In terms of fault identification, the time parameterized convolution-symmetric GRU model is innovatively constructed, the Legendre polynomial parameterized convolution kernel is used to extract the trend change characteristics of multiple time scales, the time symmetry evolution of the hidden state is realized by combining the Verlet integral mechanism, and the sensitivity and stability of the system to the fault state are greatly improved. The mechanism is suitable for the scene of asynchronous acquisition of multi-source data and mutation signal identification in the complex operating environment of the present application, and can realize accurate identification and early warning response to early faults such as compressor performance attenuation, throttling valve jamming, heat exchanger blockage, etc. It provides a robust and reliable operation safety protection mechanism for the system, greatly enhancing the continuous operation capability and maintenance response efficiency of the present application under extremely low temperature. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 A structure diagram of a dual-supply energy control system based on ultra-low ambient temperature carbon dioxide is provided for the present application.

[0028] Figure 2 A multi-dimensional time series abnormal trend graph is provided for Example Seven.

[0029] Figure 2 In the figure, the abscissa represents the sampling time point, and the sampling period is 30 seconds, and the sampling points are:

[0030] 38 hours 00 minutes (38h:00m), 38 hours 00 minutes 30 seconds (38h:00.5m), 38 hours 01 minutes (38h:01m) and 38 hours 01 minutes 30 seconds (38h:01.5m);

[0031] The ordinate represents the parameter value of the three key operating parameters, which are respectively: the blue line represents the compressor frequency (Hz); the orange line represents the pressure before the throttle valve (MPa); and the green line represents the outlet temperature of the subcooler (℃);

[0032] The yellow shaded area represents an abnormal fluctuation window. DETAILED DESCRIPTION

[0033] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0034] In Example One, according to Figure 1 The present application provides a dual-supply energy control system based on ultra-low ambient temperature carbon dioxide, which comprises: a main circulation unit, an auxiliary circulation unit, a heat exchange waterway module and a controller.

[0035] The main circulation unit, including compressor 1, air cooler 2, air cooler 1, subcooler, throttle valve 1 and evaporator, constitutes a main CO2 heating circuit; collect main circulation operation state data;

[0036] The main CO2 heating circuit path: compressor 1->air cooler 2->air cooler 1->subcooler->throttle valve 1->evaporator->compressor 1;

[0037] The auxiliary circulation unit, including compressor 2, condenser, throttle valve 2 and subcooler, constitutes an R134a circulation circuit for condensation and subcooling; collect auxiliary circulation operation state data;

[0038] The R134a circulation circuit path: compressor 2->condenser->throttle valve 2->subcooler->compressor 2;

[0039] The heat exchange waterway module, including heating outlet water pipeline and domestic hot water outlet water pipeline; collect heat exchange waterway operation state data;

[0040] The heating outlet water pipeline path: heating return water->temperature sensor->condenser->air cooler 1->circulating water pump 1->heating water supply;

[0041] The domestic hot water outlet water pipeline path: hot water return water->temperature sensor->air cooler 2->circulating water pump 2->hot water supply;

[0042] The heating return water and the domestic hot water return water are set as independent pipelines to avoid cross contamination;

[0043] The controller, including energy supply control module and fault diagnosis module;

[0044] The energy supply control module, according to the physical connection relationship of the main circulation unit and the auxiliary circulation unit, constructs a system thermal topology graph; adopts a directed graph to compare the energy supply control model to process the system thermal topology graph, obtains main circulation operation parameters and auxiliary circulation parameters, and generates energy supply control instructions; the directed graph comparison energy supply control model includes a directed graph convolutional neural network;

[0045] Constructing a system thermal topology graph:

[0046] Modeling each component (including compressor 1, air cooler 2, air cooler 1, subcooler, throttle valve 1 and evaporator; and including compressor 2, condenser, throttle valve 2 and subcooler of the auxiliary circulation unit) as a node of a graph; model the connection (refrigerant flow direction) between each component as a directed edge, the direction of the edge being consistent with the actual heat transfer path; assign each edge with current state variable characteristics, including temperature difference, pressure difference, flow rate and subcooling degree;

[0047] The fault diagnosis module combines heat exchange waterway operation state data, main circulation operation state data and auxiliary circulation operation state data to form multi-dimensional time sequence data; a time parameterized convolution-symmetric GRU model is used to process the multi-dimensional time sequence data for fault identification; the time parameterized convolution-symmetric GRU model is constructed by introducing a time parameterized convolution kernel based on Legendre expansion, a self-attention feature fusion mechanism and a symmetric GRU structure constructed based on a Verlet integral method; and the time parameterized convolution-symmetric GRU model includes a symmetric BiGRU variant model.

[0048] In example two, based on example one, in this example, the process of generating energy supply control instructions by using a directed graph to process the system thermal topology graph, specifically includes the following steps:

[0049] Step S1: Collect the thermal state data of the current nodes and edges of the system thermal topology graph, and construct an initial graph feature representation; the initial graph feature representation includes node features and edge features; the node features include compressor outlet temperature, condensing pressure, subcooler inlet / outlet temperature, evaporation temperature and subcooling degree; and the edge features include temperature difference, pressure difference and refrigerant flow;

[0050] Step S2: Process the initial graph feature representation by using a directed graph convolutional neural network to generate a high-quality embedding representation;

[0051] Step S3: Based on the high-quality embedding representation, generate main circulation operation parameters and auxiliary circulation parameters through a multi-layer perception structure; and issue the operation parameters as energy supply control instructions to each execution component:

[0052] Control the operating frequency of the compressor, the opening degree of the throttle valve and the bypass valve of the heat exchanger;

[0053] Collect temperature, pressure and power data in real time to update the node state.

[0054] In example three, based on example one, in this example, the process of generating energy supply control instructions by processing the system thermal topology graph, specifically includes the following steps:

[0055] Step R1: Collect the thermal state data of the current nodes and edges of the system thermal topology graph, and construct an initial graph feature representation; the initial graph feature representation includes node features and edge features; the node features include compressor outlet temperature, condensing pressure, subcooler inlet / outlet temperature, evaporation temperature and subcooling degree; and the edge features include temperature difference, pressure difference and refrigerant flow;

[0056] Step R2: Process the initial graph feature representation by using a traditional graph convolutional neural network to generate a high-quality embedding representation;

[0057] Step R3: generating the main cycle operation parameters and the auxiliary cycle parameters through a multi-layer perception structure based on the high-quality embedded representation; and issuing the operation parameters as energy supply control instructions to each execution component:

[0058] Controlling the compressor operation frequency, the throttle opening, and the heat exchanger bypass valve;

[0059] Collecting temperature, pressure, and power data in real time in a closed loop to update the node state.

[0060] In this embodiment based on embodiment two, step S2 specifically includes the following steps:

[0061] Step S21: based on the directional connection relationship between nodes in the initial graph feature representation, constructing an in-edge adjacency matrix and an out-edge adjacency matrix respectively, for distinguishing the energy receiving direction and the energy releasing direction in the heat transfer process, to obtain a directional adjacency matrix; and constructing a directed graph structure in combination with the node features and the directional adjacency matrix;

[0062] Step S22: for the directed graph structure, adopting a three-branch propagation mechanism to perform convolutional feature propagation on the out-edge direction, the in-edge direction, and the self-loop path of each node respectively, to extract the energy output and receiving dependency relationship in the heat flow process; introducing a directional weight factor in the propagation process to assign a learnable weighting coefficient to the out-edge direction, the in-edge direction, and the self-loop path; the above three-branch propagation mechanism is recursively unfolded in a directed graph neural network structure containing 5 layers, through each layer to perform layer-by-layer feature fusion and abstract extraction on the features carried by different propagation paths (including the out-edge path, the in-edge path, and the node self-loop path), to further capture the topological dependency relationship between cross-direction and multi-hop paths, and finally to obtain a set of multi-path global embedding vectors, with the formula as follows:

[0063] The formula of the node representation update process in the three-branch propagation mechanism:

[0064] ;

[0065] Wherein, represents the update representation of node in the th layer, represents an activation function, , and represent the learnable weight factors of the three directions, represents the out-edge neighbor set of node , represents the in-edge neighbor set of node ; represents the degree of node , represents the node The degree, Indicates the first Layer nodes The representation of, Indicates the first Layer nodes The representation of; Indicates the first Weight matrix of the outgoing edge direction, Indicates the first Weight matrix for the inbound edge direction; The feature transformation matrix representing the self-loop path (does not change with the number of layers);

[0066] A self-loop path is an edge that points from a node to itself, i.e., from the node... Departure, and return to the node. ;

[0067] Step S23: Construct a margin-contrast loss function by differentially modeling the embedding distances between positive and negative path pairs in the multi-path global embedding vector set; perform contrastive learning on the multi-path global embedding vector set using the margin-contrast loss function to maximize the consistency of positive sample path pairs and minimize the similarity of negative sample path pairs, generating high-quality embedding representations; the margin-contrast loss function consists of two parts: the sum of squared Euclidean distances between positive sample path pairs and the sum of squared margin differences between negative sample path pairs, specifically in the form of:

[0068] ;

[0069] in, This represents the interval-contrast loss function. This represents the set of positive sample path pairs, that is, path pairs that are semantically / structurally similar. This represents the set of negative sample path pairs, that is, path pairs that are semantically / structurally dissimilar. , Indicates the first In the layer (i.e., the last layer), the first , The global embedding vector of each path; Indicates the first The first in the layer The path embedding vector and the first path embedding vector The square of the Euclidean distance between the path embedding vectors; Indicates the first The first in the layer The path embedding vector and the first path embedding vector Euclidean distance between path embedding vectors; This represents the preset margin threshold, used to control the minimum distance between negative samples; The cut-off function only produces a loss when the distance between negative sample pairs is less than the margin.

[0070] Embodiment five, based on embodiment four, in this embodiment, the process of using time parameterized convolution-symmetric GRU model to process multi-dimensional time series data for fault identification, specifically includes the following steps:

[0071] Step B1: define a convolution kernel group for capturing trends at different time scales, introduce a time continuity weight function for each convolution kernel weight of the convolution kernel group, model the convolution kernel weight as a time function form, and use Legendre polynomials for low-dimensional parameterization to obtain a time parameterized convolution kernel; use the time parameterized convolution kernel to extract initial features of multi-dimensional time series data, and generate multi-scale feature data;

[0072] The parameterization formula for expressing the convolution kernel weight function as a time function and expanding it with Legendre polynomials is as follows:

[0073] ;

[0074] Wherein, represents the size of the receptive field, represents the time step, represents the parameterized weight vector of the th receptive field convolution kernel at time ; represents the basis function index (indicating the use of the th Legendre polynomial), represents the maximum expansion order, represents the weight of the th convolution kernel on the th Legendre polynomial, represents the th Legendre polynomial;

[0075] Step B2: process the multi-scale feature data through a self-attention fusion mechanism to highlight the feature channels that are most sensitive to fault identification at different time scales, and generate fused multi-scale time series features;

[0076] Step B3: input the fused multi-scale time series features into the symmetric BiGRU variant model to extract the forward and backward dependencies in the time series data, locate the fault components including compressor efficiency decline, throttling valve sticking, heat exchanger blockage and sensor failure, and perform fault identification; the symmetric BiGRU variant model is constructed as follows: a BiGRU model is established, the hidden state updating method in the BiGRU model is replaced by introducing the Verlet integration mechanism to obtain the symmetric BiGRU variant model.

[0077] The symmetric BiGRU variant model updates the hidden state by the second-order difference form of Verlet integration method, so that the numerical path of the network remains consistent in the process of time forward and backward propagation, and the stability and reversibility in long time series modeling are improved.

[0078] The Verlet integration is a numerical integration method specially used for solving second-order differential equations.

[0079] The formula for updating the hidden state by the second-order difference form of Verlet integration method is:

[0080]

[0081] Among them, represents the hidden state at the next time, represents the hidden state at the current time, represents the hidden state at the last time, represents the square term of the time step; represents the current input feature, i.e. the fused multi-scale time series features; represents a nonlinear transformation function;

[0082] Compared with the traditional BiGRU hidden state updating formula:

[0083] , which belongs to the first order;

[0084] The second-order difference form of Verlet integration method realizes the prediction of the next state by introducing the combination relationship of the historical hidden state and the current state and combining the current driving term , so as to construct a neural state evolution mechanism with time symmetry and numerical stability.

[0085] In this embodiment, the process of processing multi-dimensional time series data for fault identification includes the following steps:

[0086] ​Step E1: Define a group of convolutional kernels to capture the changing trends at different time scales. For each convolutional kernel weight in the group, introduce a time-continuous weight function to model the convolutional kernel weights as a time function and use Legendre polynomials for low-dimensional parameterization to obtain time-parameterized convolutional kernels. Use the time-parameterized convolutional kernels to extract the initial features of multi-dimensional time series data and generate multi-scale feature data.

[0087] Step E2: Process multi-scale feature data through a self-attention fusion mechanism to highlight the feature channels most sensitive to fault identification at different time scales and generate fused multi-scale temporal features;

[0088] Step E3: Input the fused multi-scale time series features into the BiGRU model, extract the forward and backward dependencies in the time series data, locate the faulty components, including compressor efficiency decline, throttle valve sticking, heat exchanger blockage and sensor failure, and perform fault identification.

[0089] Example 7, according to Figure 2 This embodiment is based on Embodiment 5. In this embodiment, the present invention provides a dual-power supply control system based on ultra-low ambient temperature carbon dioxide. The system includes: a main circulation unit, an auxiliary circulation unit, a hot water exchange circuit module, and a controller.

[0090] The main circulation unit, including compressor 1, air cooler 2, air cooler 1, subcooler, expansion valve 1 and evaporator, constitutes the main CO2 heating circuit; collects main circulation operation status data;

[0091] The auxiliary circulation unit includes compressor 2, condenser, throttle valve 2 and subcooler, forming an R134a circulation loop for condensation and subcooling; it collects auxiliary circulation operation status data.

[0092] The hot water exchange circuit module includes heating water outlet pipes and domestic hot water outlet pipes; it collects operating status data of the hot water exchange circuit.

[0093] The heating return water and domestic hot water return water are set up with independent pipes to avoid cross-contamination;

[0094] The controller includes a power supply control module and a fault diagnosis module;

[0095] The power supply control module constructs a system thermal topology diagram based on the physical connection relationship between the main circulation unit and the auxiliary circulation unit; it processes the system thermal topology diagram using a directed graph comparison power supply control model to obtain the main circulation operating parameters and auxiliary circulation parameters, and generates power supply control commands; the directed graph comparison power supply control model includes a directed graph convolutional neural network;

[0096] System thermodynamic topology: node features include: compressor outlet temperature (81.5℃), condenser inlet / outlet temperature difference (18.6℃→60.1℃), subcooler subcooling degree (8.9℃);

[0097] Edge features include: refrigerant flow rate (main cycle CO2: 18.4 L / min, auxiliary cycle R134a: 9.7 L / min), pressure difference (0.92 MPa);

[0098] Main cycle operating parameters and auxiliary cycle parameters:

[0099] Compressor 1 frequency adjustment to 52 Hz;

[0100] Throttle valve 1 opening contraction to 52%;

[0101] Switching gas cooler 1 bypass valve to partial opening state (opening 35%);

[0102] CO2 evaporation temperature increased to -5.1℃, subcooler outlet temperature increased to 62.3℃, heating return water temperature reached 43.7℃, meeting the indoor constant temperature target;

[0103] Fault diagnosis module, combined with heat exchange waterway operating state data, main cycle operating state data and auxiliary cycle operating state data, forms multi-dimensional time series data; adopts time parameterization convolution-symmetric GRU model to process multi-dimensional time series data for fault identification;

[0104] Generation Figure 2 Multi-dimensional time series anomaly trend graph during the 38th hour of operation (actual sampling time length 2 minutes);

[0105] The abscissa is the sampling time point, the sampling period is 30 seconds, and the sequence is:

[0106] 38 hours 00 minutes (38h:00m), 38 hours 00 minutes 30 seconds (38h:00.5m), 38 hours 01 minute (38h:01m) and 38 hours 01 minute 30 seconds (38h:01.5m);

[0107] The ordinate represents the parameter value of the three key operating parameters, respectively: the blue line represents the compressor frequency (Hz); the orange line represents the pressure before the throttle valve (MPa); the green line represents the subcooler outlet temperature (℃);

[0108] The yellow shaded area represents the abnormal fluctuation window.

[0109] The above describes the present application and its embodiments, which are not limited, and the drawings only show one of the embodiments of the present application, and the actual structure is not limited thereto; in general, if a person skilled in the art is inspired thereby, without departing from the purpose of the present application, without creative design, similar structure and embodiments to the technical solution, which should belong to the protection scope of the present application.

Claims

1. A kind of based on ultra-low ring temperature carbon dioxide's two supply energy control system, including main circulation unit, auxiliary circulation unit and heat exchange waterway module, the main circulation unit, acquisition main circulation operating state data;The auxiliary circulation unit, acquisition auxiliary circulation operating state data;The heat exchange waterway module, acquisition heat exchange waterway operating state data;Its characterized in that: The system further comprises a controller; The controller comprises a power supply control module and a fault diagnosis module; The power supply control module constructs a system thermal topology graph according to the physical connection relationship between the main circulation unit and the auxiliary circulation unit; The directed graph contrast power supply control model processes the system thermal topology graph to generate power supply control instructions; The fault diagnosis module combines the main circulation operation state data, the auxiliary circulation operation state data and the heat exchange waterway operation state data to form multi-dimensional time series data; and adopts a time parameterized convolution-symmetric GRU model to process the multi-dimensional time series data for fault identification; The directed graph contrast power supply control model comprises a directed graph convolutional neural network; The process of generating power supply control instructions by using the directed graph contrast power supply control model specifically comprises the following steps: Step S1: Collecting thermal state data of current nodes and edges of the system thermal topology graph to construct initial graph feature representation; the initial graph feature representation comprises node features and edge features; Step S2: Processing the initial graph feature representation by using the directed graph convolutional neural network to generate high-quality embedding representation; Step S3: Generating power supply control instructions based on the high-quality embedding representation.

2. The control system for combined heat and power based on ultra-low-temperature carbon dioxide according to claim 1, characterized in that: The time parameterized convolution-symmetric GRU model comprises a symmetric BiGRU variant model.

3. The control system for combined heat and power based on ultra-low temperature carbon dioxide according to claim 1, characterized in that: Step S2 specifically comprises the following steps: Step S21: Constructing an in-edge adjacency matrix and an out-edge adjacency matrix based on the initial graph feature representation to obtain directional adjacency matrices; and constructing a directed graph structure by combining the node features and the directional adjacency matrices; Step S22: Adopting a three-branch propagation mechanism to perform convolutional feature propagation on the out-edge direction, the in-edge direction and the self-loop path of each node, and introducing a direction weight factor in the propagation process to capture the topological dependency relationship between cross-direction and multi-hop paths and obtain a set of multi-path global embedding vectors; Step S23: Constructing an interval-contrast loss function to perform contrast learning on the set of multi-path global embedding vectors by using the interval-contrast loss function to generate high-quality embedding representation.

4. The control system for combined heat and power based on ultra-low-temperature carbon dioxide according to claim 3, characterized in that: The interval-contrast loss function is constructed in the following manner: the embedding distance between positive and negative path pairs in the set of multi-path global embedding vectors is differentiated to obtain the interval-contrast loss function.

5. The control system for combined heat and power based on ultra-low-temperature carbon dioxide according to claim 2, characterized in that: The process of adopting the time parameterized convolution-symmetric GRU model to process the multi-dimensional time series data for fault identification specifically comprises the following steps: Step B1: Defining a convolution kernel group, introducing a time continuity weight function for each convolution kernel weight in the convolution kernel group, modeling the convolution kernel weight as a time function, and using a Legendre polynomial for low-dimensional parameterization to obtain a time parameterized convolution kernel; using the time parameterized convolution kernel to extract initial features of the multi-dimensional time series data to generate multi-scale feature data; Step B2: Processing the multi-scale feature data by using a self-attention fusion mechanism to generate fused multi-scale time series features; Step B3: Inputting the fused multi-scale time series features into the symmetric BiGRU variant model to locate the fault component and perform fault identification.

6. The combined heat and power control system based on ultra-low ambient temperature carbon dioxide according to claim 5, characterized in that: The symmetric BiGRU variant model is constructed in the following manner: a BiGRU model is established, a Verlet integral mechanism is introduced to replace the hidden state updating manner in the BiGRU model, and a symmetric BiGRU variant model is obtained.

Images