Method, device, equipment, medium and product for checking load data of HVAC system

Abstract

The embodiment of the invention discloses a method, device and equipment for checking load data of an HVAC system, a medium and a product. The method comprises the following steps: acquiring a first time sequence of load data and a second time sequence of load influence data of the HVAC system; dividing the first time sequence and the second time sequence into a plurality of time windows; determining a first causal graph of each time window in a meta-learning causal discovery mode; determining a second causal graph of each time window in a non-element learning causal discovery mode; and based on the graph distance between the first causal graph and the second causal graph of each time window, determining the time window with collective anomaly in the load data from the plurality of time windows. The first causal graph considers the similarity between different time windows, the second causal graph does not consider the similarity between different time windows, and the time window of the collective anomaly can be determined based on the graph distance between the first causal graph and the second causal graph, so that accurate detection of the collective anomaly is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of Heating, Ventilation, and Air Conditioning (HVAC), and particularly to a method, apparatus, device, medium, and product for verifying load data of an HVAC system. Background Art

[0002] The HVAC system plays a crucial role in the building environment and can be used to regulate indoor temperature, humidity, air quality, and air flow distribution to provide a comfortable indoor environment. The load of the HVAC system refers to the heating, cooling, or ventilation capacity required for a building to maintain the indoor environment (temperature, humidity, air quality, etc.) at the design requirements under specific conditions, and so on. Accurately calculating the load data of the HVAC system is a key step in designing an efficient, energy-saving, and reliable system.

[0003] The types of outliers in load data usually include point outliers, contextual outliers, and collective outliers. During the process of collecting load data, some unmonitored problems are often encountered, such as sudden equipment shutdowns (e.g., repairs, normal maintenance) and interruptions in data collection. These problems may cause a large number of collective outliers to appear in the load data, and these collective outliers seriously interfere with tasks associated with the load data (e.g., constructing training data for training a load prediction model using the load data, etc.). Summary of the Invention

[0004] Embodiments of the present invention provide a method, apparatus, device, medium, and product for verifying load data of an HVAC system.

[0005] A method for verifying load data of an HVAC system includes:

[0006] Obtaining a first time series of load data of the HVAC system and a second time series of load impact data;

[0007] Dividing the first time series and the second time series into multiple time windows;

[0008] Determining a first causal graph for each time window in a meta-learning causal discovery manner;

[0009] Determining a second causal graph for each time window in a non-meta-learning causal discovery manner;

[0010] Determine the time window in which the load data shows a collective anomaly from the multiple time windows based on the graph distance between the first causal graph and the second causal graph for each time window.

[0011] It can be seen that the first causal graph (determined by the meta-learning causal discovery method) takes into account the similarity between different time windows, while the second causal graph (not determined by the meta-learning causal discovery method) does not take into account the similarity between different time windows. Based on many judgment methods such as the difference between the graph distance between the first causal graph and the second causal graph for each time window and the mean of all graph distances of all time windows, the time window in which the load data shows a collective anomaly can be determined from multiple time windows, achieving accurate detection of collective anomalies.

[0012] In one embodiment, the determining the time window in which the load data shows a collective anomaly from the multiple time windows based on the graph distance between the first causal graph and the second causal graph for each time window includes:

[0013] Determine the median and median absolute deviation of the multiple graph distances of the multiple time windows;

[0014] Based on the median, the median absolute deviation, and the graph distance of each time window, determine the robust-scaled graph distance of each time window;

[0015] When the robust-scaled graph distance is greater than a predetermined threshold, determine that the load data in the time window shows a collective anomaly.

[0016] It can be seen that based on the comparison process between the robust-scaled graph distance and the threshold, the time window in which a collective anomaly occurs can be quickly determined.

[0017] In one embodiment, the load impact data includes the production plan data and weather data of the HVAC system;

[0018] The splitting the first time series and the second time series into multiple time windows includes:

[0019] Taking the load data at the first moment and the production plan data at the second moment as the first input, and the load data at the second moment as the first output, determine the first hidden state from the first input to the first output; wherein the first moment is before the second moment;

[0020] Taking the weather data at the first moment as the second input and the weather data at the second moment as the second output, determine the second hidden state from the second input to the second output;

[0021] Split the first hidden state and the second hidden state into the multiple time windows.

[0022] Therefore, by comprehensively considering the impact of production plan data and weather data on load data and capturing the complex relationships of the impact through hidden states, the accuracy of collective anomaly detection can be improved.

[0023] In one embodiment, the weather data includes multiple dimensions; wherein the dimensions include at least one of the following: temperature; humidity; air pressure; wind; precipitation; radiation; visibility; cloud cover; air quality index.

[0024] It can be seen that by considering multi-dimensional weather, the accuracy of collective anomaly detection can be further improved.

[0025] In one embodiment, the load impact data includes weather data; the dividing of the first time series and the second time series into a plurality of time windows includes:

[0026] Using the load data at the first moment as the first input and the load data at the second moment as the first output, determining a first hidden state from the first input to the first output; wherein the first moment is before the second moment;

[0027] Using the weather data at the first moment as the second input and the weather data at the second moment as the second output, determining a second hidden state from the second input to the second output;

[0028] Dividing the first hidden state and the second hidden state into the plurality of time windows.

[0029] Therefore, considering the impact of weather data on load data and capturing the complex relationships of the impact through hidden states, the accuracy of collective anomaly detection can be improved.

[0030] In one embodiment, the load impact data includes production plan data of the HVAC system; the dividing of the first time series and the second time series into a plurality of time windows includes:

[0031] Using the load data at the first moment and the production plan data at the second moment as the first input and the load data at the second moment as the first output, determining a first hidden state from the first input to the first output; wherein the first moment is before the second moment;

[0032] Dividing the first hidden state into the plurality of time windows.

[0033] Therefore, considering the impact of production plan data on load data and capturing the complex relationships of the impact through hidden states, the accuracy of collective anomaly detection can be improved.

[0034] In one embodiment, it includes:

[0035] Remove the time windows in which the load data shows collective anomalies from the multiple time windows;

[0036] Generate training data based on the load data in the remaining time windows, where the training data is used to train a deep learning model suitable for detecting anomalies in the load data of the HVAC system.

[0037] Therefore, by removing the time windows with collective anomalies, the accuracy of the training data for training a deep learning model to detect anomalies in the load data of the HVAC system can be improved.

[0038] A device for inspecting the load data of an HVAC system, comprising:

[0039] An acquisition module for acquiring a first time series of the load data of the HVAC system and a second time series of load impact data;

[0040] A segmentation module for segmenting the first time series and the second time series into multiple time windows;

[0041] A first determination module for determining a first causal graph for each time window in a meta-learning causal discovery manner;

[0042] A second determination module for determining a second causal graph for each time window in a non-meta-learning causal discovery manner;

[0043] A third determination module for determining the time windows in which the load data shows collective anomalies from the multiple time windows based on the graph distance between the first causal graph and the second causal graph of each time window.

[0044] It can be seen that the first causal graph (determined in a meta-learning causal discovery manner) takes into account the similarity between different time windows, while the second causal graph (not determined in a meta-learning causal discovery manner) does not take into account the similarity between different time windows. Based on many judgment methods such as the difference between the graph distance between the first causal graph and the second causal graph of each time window and the mean of all graph distances of all time windows, the time windows in which the load data shows collective anomalies can be determined from multiple time windows, achieving accurate detection of collective anomalies.

[0045] In one embodiment, the third determination module is configured to determine the median and median absolute deviation of the multiple graph distances of the multiple time windows; determine the robustly scaled graph distance of each time window based on the median, the median absolute deviation, and the graph distance of each time window; and determine that the load data in the time window shows collective anomalies when the robustly scaled graph distance is greater than a predetermined threshold.

[0046] It can be seen that based on the comparison process between the graph distance after robust scaling and the threshold, the time window of collective anomalies can be quickly determined.

[0047] In one embodiment, the load impact data includes production plan data and weather data of the HVAC system;

[0048] The segmentation module is configured to take the load data at the first moment and the production plan data at the second moment as the first input, take the load data at the second moment as the first output, and determine the first hidden state from the first input to the first output; wherein the first moment is before the second moment; take the weather data at the first moment as the second input, take the weather data at the second moment as the second output, and determine the second hidden state from the second input to the second output; and segment the first hidden state and the second hidden state into the multiple time windows.

[0049] Therefore, by comprehensively considering the influence of production plan data and weather data on load data, the accuracy of collective anomaly detection can be improved.

[0050] An electronic device includes a processor and a memory;

[0051] The memory stores an application program executable by the processor, which is used to cause the processor to execute the method for inspecting the load data of the HVAC system as described in any one of the above.

[0052] A computer-readable storage medium stores computer instructions, and when the computer instructions are executed by a processor, the method for inspecting the load data of the HVAC system as described in any one of the above is implemented.

[0053] A computer program product includes a computer program, and when the computer program is executed by a processor, the method for inspecting the load data of the HVAC system as described in any one of the above is implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The following will make the above and other features and advantages of the present invention clearer to those of ordinary skill in the art by referring to the preferred embodiments of the present invention described in detail with reference to the drawings, in which:

[0055] Figure 1 is a schematic diagram of point anomalies in load data.

[0056] Figure 2 is a schematic diagram of context anomalies in load data.

[0057] Figure 3 is a schematic diagram of collective anomalies in load data.

[0058] Figure 4It is a flowchart of a method for inspecting load data of an HVAC system according to an embodiment of the present invention.

[0059] Figure 5A It is a first schematic diagram for determining a hidden state according to an embodiment of the present invention.

[0060] Figure 5B It is a second schematic diagram for determining a hidden state according to an embodiment of the present invention.

[0061] Figure 6 It is a schematic diagram of a causal graph for determining a time window in a meta - learning causal discovery manner according to an embodiment of the present invention.

[0062] Figure 7 It is a structural diagram of a device for inspecting load data of an HVAC system according to an embodiment of the present invention.

[0063] Figure 8 It is a schematic structural diagram of an electronic device according to an embodiment of the present invention.

[0064] Among them, the reference numerals are as follows:

[0065]

[0066] Detailed Embodiments

[0067] To make the objectives, technical solutions, and advantages of the present invention clearer, the following examples are given to further elaborate on the present invention in detail. In this patent application, nouns and pronouns related to people are not limited to a specific gender.

[0068] For the sake of brevity and intuitiveness in description, the following elaborates on the solutions of the present invention by describing several representative embodiments. A large number of details in the embodiments are only used to help understand the solutions of the present invention. However, it is obvious that the technical solutions of the present invention can be implemented without being limited to these details. To avoid unnecessarily obscuring the solutions of the present invention, some embodiments are not described in detail but only present a framework. Hereinafter, "including" means "including but not limited to", and "according to..." means "at least according to..., but not limited to only according to...". Due to the language habits of Chinese, when the quantity of a component is not specifically indicated hereinafter, it means that the component can be one or more, or can be understood as at least one.

[0069] In an HVAC system, load data is a key input for designing and optimizing the system. For example, load data can include: (1) heating load data; (2) cooling load data; (3) ventilation load data; (4) output power of a solar photovoltaic (PV) system, etc. Load data plays an important role in HVAC systems and broader building energy management. It can be used for various purposes, including but not limited to system design, operation optimization, energy conservation analysis, equipment selection, fault diagnosis, and predictive maintenance, etc.

[0070] The following is a demonstration of specific application scenarios of load data:

[0071] (I) System design

[0072] (1.1) Equipment selection: Accurate selection: By analyzing load data, determine the appropriate capacity of each device (such as boilers, chillers, fans, pumps, etc.) in the HVAC system. For example, select a boiler or chiller with an appropriate power according to the peak heating and cooling loads to ensure that the equipment can meet the maximum load demand during operation and will not waste resources due to excessive redundancy. Avoid energy waste and cost increase caused by oversized equipment selection. For example, if the cooling load is inaccurately calculated and an oversized chiller is selected, it may cause the chiller to operate at a low load for most of the time, resulting in reduced efficiency and increased energy consumption. (1.2) System layout planning: Determine the reasonable pipe and duct sizes based on load data to ensure that the system can efficiently transport heat or cold during operation and reduce energy losses. Arrange the installation locations of equipment reasonably according to the load distribution to improve the overall performance of the system. For example, installing the chiller near the load center can reduce energy losses during the transportation of chilled water.

[0073] (II) Operation optimization

[0074] Adjust equipment operation parameters according to load changes: Automatically adjust the operation parameters of the HVAC system, such as temperature set points, fan speeds, pump flows, etc., through real-time monitoring of load data to achieve energy-saving operation. Coordinated operation optimization of multiple devices: In a complex HVAC system, there are usually multiple devices (such as boilers, chillers, heat pumps, etc.). By analyzing load data, coordinated operation optimization of these devices can be achieved to improve the overall efficiency of the system. For example, in winter, when the load is low, the heat pump can be preferentially used for heating, and when the load is high, the boiler can be started to assist in heating to achieve reasonable utilization of energy. Optimize energy distribution: According to load data and energy prices, reasonably allocate the usage ratios of different energies (such as electricity, gas, solar energy, etc.) to reduce operating costs. For example, during periods of low electricity prices, electric-driven devices can be preferentially used; when the gas price is low, the usage ratio of gas equipment can be increased.

[0075] (III) Energy Saving Analysis

[0076] Energy Consumption Assessment and Analysis: By analyzing the load data and combining the operating parameters and energy consumption data of the equipment, the energy consumption of the HVAC system can be accurately calculated. For example, based on the refrigeration load and the operating time, power, etc. of the chiller, the power consumption of the chiller can be calculated; based on the heating load and the operating time, gas consumption, etc. of the boiler, the gas consumption of the boiler can be calculated.

[0077] (IV) Equipment Maintenance

[0078] Equipment Selection Verification: Verify whether the equipment meets the load requirements: By analyzing the load data during actual operation, verify whether the installed equipment can meet the load requirements of the system. If it is found that the equipment cannot meet the load requirements during certain periods, equipment upgrade or increase in equipment capacity may need to be considered. Evaluate equipment performance: According to the load data and the operating data of the equipment, evaluate the actual performance of the equipment and determine whether the equipment is operating under the best working conditions. For example, by analyzing the operating efficiency of the chiller at different loads, determine whether it needs maintenance or optimization of operating parameters. Fault Diagnosis and Predictive Maintenance: By analyzing the abnormal changes in the load data and the equipment operating data, discover equipment faults in a timely manner. For example, if the refrigeration load of a certain chiller suddenly drops, it may be caused by problems such as refrigerant leakage or compressor failure. Discovering and dealing with it in a timely manner can prevent the expansion of the fault.

[0079] (V) By using the load data, an equipment fault prediction model can also be established to predict in advance the possible faults of the equipment and reasonably arrange the maintenance plan. For example, by analyzing the operating data of the equipment at different loads, predict the service life and maintenance cycle of the equipment, prepare maintenance materials and personnel in advance, and reduce the equipment downtime. Example: A prediction model based on artificial intelligence can be used to accurately predict the refrigeration load under different weather conditions and the electrical load including fans and photovoltaic power generation. The accuracy of the prediction is closely related to the quality of the training data used. Therefore, it is crucial to thoroughly clean and optimize the training data before starting the training process of the prediction model.

[0080] The above exemplary description shows the typical uses of the load data. Those skilled in the art can realize that such a description is only exemplary and is not used to limit the protection scope of the embodiments of the present invention.

[0081] The load data of the HVAC system is time series data, and the types of outliers in it can generally include:

[0082] (1) Point Anomaly: These anomalies are individual data points that deviate significantly from the expected pattern or distribution of the time series.

[0083] Figure 1 It is a schematic diagram of a point anomaly in the load data. In Figure 1 , a point anomaly occurs at point 10.

[0084] (2) Contextual anomaly: A contextual anomaly occurs when the behavior of a data point is considered anomalous in a specific context or condition, but may not be anomalous in a different context or condition.

[0085] Figure 2 It is a schematic diagram of a contextual anomaly in the load data. In Figure 2 , a contextual anomaly occurs at point 11.

[0086] (3) Collective anomaly: A collective anomaly refers to a group or subset of data points that collectively exhibit anomalous behavior, although individual data points may not be considered anomalous.

[0087] Figure 3 It is a schematic diagram of a collective anomaly in the load data. In Figure 3 , a collective anomaly occurs in interval 12.

[0088] Detecting collective anomalies in the time series of HVAC load data is generally more challenging than detecting single-point anomalies, mainly because of the complexity and context-dependence of collective anomalies. A collective anomaly means that a group of data points exhibits anomalous behavior as a whole, while individual data points themselves may not be anomalous. Collective anomalies often depend on local time series patterns. For example, under normal circumstances, the time series may exhibit a certain periodicity or trend, while collective anomalies may manifest as a sudden change in this pattern, and detecting such a change requires an in-depth understanding of the context of the time series. Moreover, the detection of collective anomalies usually requires considering background information, including historical data of the time series, seasonal factors, trend changes, etc. Without this background information, it is difficult to determine whether a group of data points is anomalous. In addition, collective anomalies usually also have complex time dependencies. Time series data is usually dynamically changing, and collective anomalies may exhibit different characteristics in different time periods.

[0089] Therefore, detecting collective anomalies in the time series of HVAC load data is a complex task, and the main difficulties lie in its context-dependence, complex time-dependence, data noise and uncertainty, lack of labeled data, subjectivity of anomaly definition, and computational complexity. These factors are intertwined, making collective anomaly detection more challenging than single-point anomaly detection.

[0090] In an embodiment of the present invention, based on the degree of isolation of the gap between the causal graph determined by the meta-learning causal discovery method and the causal graph determined by the non-meta-learning causal discovery method within a certain time window among all gaps of this type in all time windows, it is determined whether a collective anomaly occurs in this time window, so as to implement a technical solution suitable for detecting collective anomalies in the load data of the HVAC system.

[0091] Figure 4 is a flowchart of a method for detecting load data of an HVAC system according to an embodiment of the present invention. As Figure 4 shown, the method includes:

[0092] Step 101: Obtain the first time series of the load data of the HVAC system and the second time series of the load impact data.

[0093] For example, a first time series of the load data of the HVAC system can be collected by using sensors (such as temperature sensors, humidity sensors, flow sensors, pressure sensors, electricity meters, and power sensors, etc.), or the first time series of the load data can be obtained from a data acquisition system or an energy consumption monitoring platform.

[0094] The load of the HVAC system is comprehensively affected by various factors, which can be roughly divided into internal factors and external factors. Internal factors can include building characteristics, personnel and equipment, operation strategies and controls, etc., and external factors can include meteorological conditions, geographical and building environments, social and economic factors, etc. For example, the load impact data affecting the load data of the HVAC system can include: weather data and production plan data (such as maintenance time, normal maintenance time, holidays, etc.), etc. Here, a second time series of the load impact data is also obtained, such as obtaining weather data from a weather station and obtaining production plan data from production planning software, etc. Preferably, the first time series and the second time series are aligned in time.

[0095] Step 102: Divide the first time series and the second time series into multiple time windows.

[0096] Here, the first time series and the second time series are divided according to time windows, for example, every half hour, every hour or every day, into a time window. Each time window contains the HVAC load data and load impact data within this time period.

[0097] Step 103: Determine the first causal graph of each time window in a meta-learning causal discovery manner.

[0098] Here, a meta-learning model suitable for causal discovery can be selected, such as a model based on a graph neural network (GNN) or a model based on Bayesian methods. Then, based on the meta-learning causal discovery method, the first causal graph for each time window is determined. The first causal graph for each time window characterizes the influence relationship of the load impact data within that time window on the load data within that time window.

[0099] Specifically, the process of determining the first causal graph may include:

[0100] First, a meta-learning model is trained. The training process includes: preparing the training data of multiple time windows as the training set, where the training data of each time window includes the training data of HVAC load and the load impact training data; training the meta-learning model based on the training set to learn the causal relationship patterns in different time windows.

[0101] Then, using the trained meta-learning model, causal relationship inference is performed on the data in each time window segmented in step 102. The meta-learning model can output the causal relationship scores between variables to represent the possibility of a causal relationship between variables. According to the causal relationship scores, high-confidence causal relationships are screened out. According to the screened high-confidence causal relationships, the structure of the first causal graph is constructed. Among them: The first causal graph is a directed graph, the nodes represent variables (such as load data, load impact data, etc.), and the edges represent the causal relationships between variables, and the weights of the edges can be determined by the causal relationship scores output by the meta-learning model.

[0102] Step 104: Determine the second causal graph for each time window in a non-meta-learning causal discovery manner.

[0103] Here, for each time window, its second causal graph is determined in a non-meta-learning causal discovery manner. In non-meta-learning causal discovery methods, determining the causal graph usually involves a series of classical causal inference techniques. The non-meta-learning causal discovery method mainly infers the causal relationships between variables by analyzing information such as statistical dependencies, time orders, and intervention effects in the data. For example, the non-meta-learning causal discovery method may include:

[0104] (1) Constraint-Based Methods (such as the Peter-Clark algorithm).

[0105] (2) FunctionalModel-Based Methods (such as causal Bayesian networks or causal function models, etc.).

[0106] (3) Intervention-Based Methods (such as Causal Inference Experiments or Causal Effect Estimation, etc.);

[0107] (4) Time-Series Methods (such as Granger Causality Test) or Dynamic Causal Models, etc.

[0108] Step 105: Determine the time window in which the load data shows a collective anomaly from multiple time windows based on the graph distance between the first causal graph and the second causal graph for each time window.

[0109] Here, calculate the graph distance between the first causal graph and the second causal graph for each time window, and then determine the time window with a collective anomaly from all time windows based on the graph distances of all time windows. Among them: The graph distance of the time window with a collective anomaly has a large gap from the mean of all graph distances (for example, greater than a predetermined threshold).

[0110] It can be seen that the first causal graph (determined by the meta-learning causal discovery method) takes into account the similarity between different time windows, while the second causal graph (not determined by the meta-learning causal discovery method) does not take into account the similarity between different time windows. Based on many judgment methods such as the difference between the graph distance between the first causal graph and the second causal graph for each time window and the mean of all graph distances of all time windows, the time window in which the load data shows a collective anomaly can be determined from multiple time windows, realizing accurate detection of collective anomalies.

[0111] In one implementation, determining the time window in which the load data shows a collective anomaly from multiple time windows based on the graph distance between the first causal graph and the second causal graph for each time window includes: determining the median and median absolute deviation of the multiple graph distances of the multiple time windows; determining the robustly scaled graph distance for each time window based on the median, median absolute deviation, and the graph distance of each time window; when the robustly scaled graph distance of at least one of the multiple time windows is greater than a predetermined threshold, determining that the load data in the at least one time window shows a collective anomaly.

[0112] It can be seen that based on the comparison process between the robustly scaled graph distance and the threshold, the time window with a collective anomaly can be quickly determined.

[0113] Here, taking the median and the median absolute deviation as an example, an instance of determining the time window for the occurrence of collective anomalies is described. Those skilled in the art can realize that such a description is only exemplary and is not used to limit the protection scope of the embodiments of the present invention.

[0114] In one embodiment, the load impact data includes production plan data and weather data of the HVAC system; step 102 specifically includes: taking the load data at the first moment and the production plan data at the second moment as the first input, and the load data at the second moment as the first output, to determine the first hidden state from the first input to the first output; wherein the first moment is before the second moment; taking the weather data at the first moment as the second input, and the weather data at the second moment as the second output, to determine the second hidden state from the second input to the second output; and dividing the first hidden state and the second hidden state into multiple time windows. Therefore, by comprehensively considering the impact of production plan data and weather data on the load data, and capturing the complex relationship of the impact through the hidden state, the accuracy of collective anomaly detection can be improved.

[0115] In one embodiment, the weather data includes multiple dimensions; wherein the dimensions include at least one of the following: temperature; humidity; air pressure; wind; precipitation; radiation; visibility; cloud cover; air quality index. It can be seen that by considering multi-dimensional weather, the accuracy of anomaly detection is further improved.

[0116] In one embodiment, the load impact data includes weather data; step 102 specifically includes: taking the load data at the first moment as the first input, and the load data at the second moment as the first output, to determine the first hidden state from the first input to the first output; wherein the first moment is before the second moment; taking the weather data at the first moment as the second input, and the weather data at the second moment as the second output, to determine the second hidden state from the second input to the second output; and dividing the first hidden state and the second hidden state into multiple time windows. Therefore, by considering the impact of weather data on the load data, and capturing the complex relationship of the impact through the hidden state, the accuracy of anomaly detection for several days can be improved.

[0117] In one embodiment, the load impact data includes production plan data of the HVAC system; step 102 specifically includes: taking the load data at the first moment and the production plan data at the second moment as the first input, and the load data at the second moment as the first output, to determine the first hidden state from the first input to the first output; wherein the first moment is before the second moment; and dividing the first hidden state into multiple time windows. Therefore, by considering the impact of production plan data on the load data, and capturing the complex relationship of the impact through the hidden state, the accuracy of collective anomaly detection can be improved.

[0118] In one embodiment, the method includes: removing time windows in which load data exhibit collective anomalies from a plurality of time windows; and generating training data based on the load data in the remaining time windows, wherein the training data is used to train a deep learning model suitable for detecting anomalies in load data of an HVAC system.

[0119] Therefore, by removing the time windows where collective anomalies occur, the accuracy of the training data used to train the deep learning model for detecting anomalies in the load data of the HVAC system can be improved.

[0120] After removing the time windows where the load data exhibits collective anomalies from multiple time windows, various tasks can be performed based on the load data in the remaining time windows. For example, tasks such as equipment selection, system layout planning, operation optimization, energy consumption assessment and analysis, and equipment maintenance can be performed, although this is not a limitation in the embodiments of the present invention.

[0121] Figure 5A is a first schematic diagram of determining a hidden state according to an embodiment of the present invention. Figure 5A In the paper, a schematic diagram of the structure of the recurrent neural network (RNN) is shown.

[0122] exist Figure 5A In this paper, taking high temperature cooling load as an example, the time series of the load is expressed as Load impact data includes production plan data, where the time series of production plan data (e.g., whether it is a holiday, whether there is maintenance, production quantity, etc.) is

[0123] The input layer includes: load data y at time (t-2) i,t-2 And the production plan data x at time (t-1) i,t-1 ; Load data y at time (t-1) i,t-1 And the production plan data x at time t i,t ; Load data y at time t i,t-1 And the production plan data x at time (t+1) i,t+1 .

[0124] Hidden Layers include: the hidden state h at time (t-1) i,t-1 , the hidden state h at time t i,t and the hidden state h at time (t+1) i,t+1 , which represent the hidden states of the RNN at different time steps. The hidden state is a key concept in RNN, which carries the information of the previous time step in the sequence and passes it to the next time step.

[0125] The output layer includes: the load data y at time (t-1) i,t-1 ; the load data y at time t i,t ; the load data y at time (t+1) i,t+1 .

[0126] The input layer receives the input data at the current time step, while the output layer generates the prediction results of the model. The input layer and the output layer are respectively connected to the hidden layer, indicating how the input data affects the hidden state and how the hidden state affects the output. The characteristic of the RNN is its recurrent connection, which allows the network to consider previous information when processing sequential data. This structure enables the RNN to capture the temporal dependencies in the sequence. In the RNN, the same weights are shared across all time steps, which means that the network uses the same parameters to process data at each time step, helping to reduce the number of parameters of the model and enabling the model to generalize to different sequence lengths. Based on the hidden layer, a time series of the hidden state can be provided.

[0127] Figure 5B is the first schematic diagram for determining the hidden state according to an embodiment of the present invention. In Figure 5B , taking the high-temperature refrigeration load as an example, the time series of the load is represented as The load impact data includes weather data, and the time series of production plan data (such as whether it is a holiday, whether there is maintenance, production quantity, etc.) is

[0128] The input layer includes: the weather data w at time (t-2) i,t-2 ; the weather data w at time (t-1) i,t-1 and the weather data w at time t i,t .

[0129] The hidden layer includes: the hidden state h marked at time (t-1) i,t-1 , the hidden state h at time t i,,t and the hidden state h at time (t+1) i,,t+1 , respectively representing the hidden states of the RNN at different time steps. The hidden state is a key concept in the RNN. It carries the information of the previous time steps in the sequence and passes it to the next time step.

[0130] The output layer includes: the weather data w at time (t-1) i,t-1 ; the weather data w at time t i,t ; the weather data w at time (t+1) i,t+1 .

[0131] Figure 6 is a schematic diagram of the causal graph for determining the time window in the way of meta-learning causal discovery according to an embodiment of the present invention. Wherein: the load impact data includes multiple dimensions (inFigure 6 in which, the weather data and production plan data have a dimension of 4, such as temperature, humidity, air pressure, and wind).

[0132] The time series 61 of the first hidden state is a data sequence of hidden states generated in the manner as Figure 5A shown (where: the input layer includes the load data at the second moment and the production plan data at the first moment, and the output layer includes the load data at the first moment, where the second moment is before the first moment).

[0133] The time series 62 - 65 of the second hidden state are respectively data sequences of hidden states generated in the manner as Figure 5B shown. Among them: the input layer includes weather data with 4 dimensions (all weather data at the second moment), and 4 time series 62 - 65 can be generated, while the output layer includes the weather data at the first moment, where the second moment is before the first moment).

[0134] Align the time series 61 of the first hidden state with the time series 62 - 65 of the second hidden state in time. Then, divide the time series 61 and the time series 62 - 65 into time windows of a predetermined size, obtaining N time windows (preferably with the same window size), which are respectively time windows 31 - 3N. Each of the time windows 31 - 3N contains the time series 61 and the time series 62 - 65 within that time window.

[0135] Input the time windows 31 - 3N into the meta - learning causal discovery module 40 (for example, a trained meta - learning model), and causal graphs 51 - 5N corresponding to the time windows 31 - 3N can be obtained. In each of the causal graphs 51 - 5N, there are 5 nodes, respectively representing the time series 61, the time series 62 - 65; the edges between the nodes represent the causal relationships between the nodes. The causal graphs 51 - 5N generated by the meta - learning causal discovery module 40 respectively consider the similarities between the time windows 31 - 3N. The causal graphs 51 - 5N can form a graph matrix

[0136] Generate causal graphs for each of the time windows 31 - 3N in a non - meta - learning causal discovery manner to form a graph matrix

[0137] Calculate the graph structure matrix the distance between each causal graph in and the corresponding causal graph in the graph structure matrix , denoted as d s . Among them

[0138] If the absolute value of the robustly scaled distance K for each time window is greater than a predetermined threshold (for example, 2.5), then the load data for that time window will be marked as a collective outlier. Among them: MED is the median function; MAD is the median absolute deviation function.

[0139] In Figure 6 , the exemplary description shows that the load impact data includes weather data and production plan data. In fact, the load impact data may only include weather data without production plan data, or the load impact data may only include production plan data without weather data. The embodiments of the present invention have no such limitation.

[0140] It can be seen that in the embodiments of the present invention, the time series data is divided into different windows, and the meta-learning causal discovery method of the time series is used to obtain the stable structure embedded in the long time series data. Compared with learning the graph structure between individual multivariate time series within each window, the structure learned using the meta-learning technique is less affected by collective outliers. By comparing the causal graph structures obtained under two different conditions, the collective outliers will be automatically marked. In addition, by combining the reconstruction of the time series data, point outliers and context outliers in the time series data can also be detected. There are currently some methods for detecting outliers in time series data using time graphs. However, these methods usually require a period of clean data without outliers to train the model, and then detect outliers in new data based on the prediction error. The main differences between the embodiments of the present invention and these methods are as follows: First, the embodiments of the present invention use a causal graph instead of an undirected graph model; second, the embodiments of the present invention are completely unsupervised and allow direct training on time series data containing outliers.

[0141] The embodiments of the present invention also propose a device for inspecting the load data of the HVAC system. Figure 7 is a structural diagram of a device for inspecting the load data of the HVAC system according to the embodiments of the present invention. As Figure 7 shown, the device 200 for inspecting the load data of the HVAC system includes: an acquisition module 201 for acquiring the first time series of the load data of the HVAC system and the second time series of the load impact data; a segmentation module 202 for segmenting the first time series and the second time series into multiple time windows; a first determination module 203 for determining the first causal graph of each time window in a meta-learning causal discovery manner; a second determination module 204 for determining the second causal graph of each time window in a non-meta-learning causal discovery manner; and a third determination module 205 for determining the time window in which the load data has collective outliers from the multiple time windows based on the graph distance between the first causal graph and the second causal graph of each time window.

[0142] In one embodiment, the third determination module 205 is configured to determine the median and the median absolute deviation of the graph distances of multiple time windows; determine the robustly scaled graph distance of each time window based on the median, the median absolute deviation, and the graph distance of each time window; and when the robustly scaled graph distance is greater than a predetermined threshold, determine that the load data in the time window has a collective anomaly.

[0143] In one embodiment, the load impact data includes production plan data and weather data of the HVAC system; the segmentation module 202 is configured to determine a first hidden state from a first input to a first output with the load data at a first moment and the production plan data at a second moment as the first input and the load data at the second moment as the first output, where the first moment is before the second moment; determine a second hidden state from a second input to a second output with the weather data at the first moment as the second input and the weather data at the second moment as the second output; and segment the first hidden state and the second hidden state into multiple time windows.

[0144] An embodiment of the present invention also provides an electronic device having a processor-memory architecture. Figure 8 is a schematic structural diagram of an electronic device according to an embodiment of the present invention. As Figure 8 shown, the electronic device 300 includes a processor 301, a memory 302, and a computer program stored on the memory 302 and executable on the processor 301. When the computer program is executed by the processor 301, it implements the method for checking the load data of the HVAC system as described above. Among them, the memory 302 can be specifically implemented as various storage media such as an electrically erasable programmable read-only memory (EEPROM), a flash memory, and a programmable read-only memory (PROM). The processor 301 can be implemented as including one or more central processing units or one or more field programmable gate arrays, where the field programmable gate array integrates one or more central processing unit cores. Specifically, the central processing unit or the central processing unit core can be implemented as a CPU, an MCU, a DSP, etc.

[0145] It should be noted that not all steps and modules in the above-mentioned processes and structural diagrams are necessary, and some steps or modules can be ignored according to actual needs. The execution order of the steps is not fixed and can be adjusted according to needs. The division of each module is only for the convenience of description by functional division. In actual implementation, one module can be implemented by multiple modules, and the functions of multiple modules can also be implemented by the same module. These modules can be located in the same device or in different devices.

[0146] The hardware modules in each embodiment can be implemented mechanically or electronically. For example, a hardware module can include specially designed permanent circuits or logic devices (such as dedicated processors, such as FPGAs or ASICs) for performing specific operations. A hardware module can also include programmable logic devices or circuits (such as including general-purpose processors or other programmable processors) temporarily configured by software for performing specific operations. As for whether to specifically adopt a mechanical approach, or dedicated permanent circuits, or temporarily configured circuits (such as configured by software) to implement the hardware module, it can be determined based on cost and time considerations.

[0147] The above are only the preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

[0148] In this patent application, nouns and pronouns related to people are not limited to a specific gender.

Claims

1. A method for checking the load data of an HVAC system, characterized in that, Including: Obtaining (101) a first time series of load data and a second time series of load impact data of the HVAC system; Dividing (102) the first time series and the second time series into a plurality of time windows; Determining (103) a first causal graph for each time window in a meta-learning causal discovery manner; Determining (104) a second causal graph for each time window in a non-meta-learning causal discovery manner; Based on the graph distance between the first causal graph and the second causal graph of each time window, determining (105) the time window in which collective anomalies occur in the load data from the plurality of time windows.

2. The method according to claim 1, wherein The determining (105) the time window in which collective anomalies occur in the load data from the plurality of time windows based on the graph distance between the first causal graph and the second causal graph of each time window includes: Determining the median and median absolute deviation of the plurality of graph distances of the plurality of time windows; Based on the median, the median absolute deviation, and the graph distance of each time window, determining the robustly scaled graph distance of each time window; When the robustly scaled graph distance is greater than a predetermined threshold, determining that collective anomalies occur in the load data in the time window.

3. The method according to claim 1, characterized in that The load impact data includes production plan data and weather data of the HVAC system; The dividing (102) the first time series and the second time series into a plurality of time windows includes: Using the load data at the first moment and the production plan data at the second moment as the first input, and the load data at the second moment as the first output, determining a first hidden state from the first input to the first output; wherein the first moment is before the second moment; Using the weather data at the first moment as the second input, and the weather data at the second moment as the second output, determining a second hidden state from the second input to the second output; Dividing the first hidden state and the second hidden state into the plurality of time windows.

4. The method according to claim 3, wherein The weather data includes multiple dimensions; wherein the dimensions include at least one of the following: temperature; humidity; air pressure; wind; precipitation; radiation; visibility; cloud cover; air quality index.

5. The method according to claim 1, characterized in that The load impact data includes weather data; the dividing (102) the first time series and the second time series into a plurality of time windows includes: Using the load data at the first moment as the first input, and the load data at the second moment as the first output, determining a first hidden state from the first input to the first output; wherein the first moment is before the second moment; Using the weather data at the first moment as the second input, and the weather data at the second moment as the second output, determining a second hidden state from the second input to the second output; Dividing the first hidden state and the second hidden state into the plurality of time windows.

6. The method according to claim 1, characterized in that The load impact data includes production plan data of the HVAC system; the dividing (102) the first time series and the second time series into a plurality of time windows includes: Using the load data at the first moment and the production plan data at the second moment as the first input, and the load data at the second moment as the first output, determine the first hidden state from the first input to the first output; wherein the first moment is before the second moment; Divide the first hidden state into the plurality of time windows.

7. The method according to any one of claims 1-6, characterized in that, Comprising: From the plurality of time windows, remove the time windows in which the load data has a collective anomaly; Based on the load data in the remaining time windows, generate training data for training a deep learning model suitable for detecting anomalies in the load data of the HVAC system.

8. A device for checking the load data of an HVAC system, characterized in that, Comprising: An acquisition module (201) for acquiring a first time series of load data of the HVAC system and a second time series of load impact data; A segmentation module (202) for dividing the first time series and the second time series into a plurality of time windows; A first determination module (203) for determining a first causal graph for each time window in a meta-learning causal discovery manner; A second determination module (204) for determining a second causal graph for each time window in a non-meta-learning causal discovery manner; A third determination module (205) for determining the time windows in which the load data has a collective anomaly from the plurality of time windows based on the graph distance between the first causal graph and the second causal graph for each time window.

9. The apparatus according to claim 8, wherein The third determination module (205) is configured to determine the median and the median absolute deviation of the plurality of graph distances of the plurality of time windows; based on the median, the median absolute deviation, and the graph distance of each time window, determine the robustly scaled graph distance of each time window; when the robustly scaled graph distance is greater than a predetermined threshold, determine that the load data in the time window has a collective anomaly.

10. The device according to claim 8, characterized in that, The load impact data includes the production plan data and the weather data of the HVAC system; The segmentation module (202) is configured to use the load data at the first moment and the production plan data at the second moment as the first input, and the load data at the second moment as the first output, to determine the first hidden state from the first input to the first output; wherein the first moment is before the second moment; using the weather data at the first moment as the second input, and the weather data at the second moment as the second output, determine the second hidden state from the second input to the second output; divide the first hidden state and the second hidden state into the plurality of time windows.

11. An electronic device (300), characterized in that, Comprising a processor (301) and a memory (302); The memory (302) stores an application program executable by the processor (301) for causing the processor (301) to execute the method for inspecting the load data of the HVAC system according to any one of claims 1 to 7.

12. A computer-readable storage medium having computer instructions stored thereon, characterized in that, When the computer instructions are executed by the processor, the method for inspecting the load data of the HVAC system according to any one of claims 1 to 7 is implemented.

13. A computer program product, characterized in that, Comprising a computer program which, when executed by a processor, implements the method of verifying load data of an HVAC system according to any one of claims 1 to 7.

Images