Building energy-saving data mining and analysis processing system based on artificial intelligence

Abstract

The invention belongs to the technical field of computer data processing, and discloses a building energy-saving data mining and analysis processing system based on artificial intelligence, and the system comprises an energy efficiency index calculation and feature engineering module which is used for calculating energy efficiency feature data representing the physical efficiency of equipment; the system is provided with a machine learning model and an expert system model in parallel, the machine learning model combines energy efficiency characteristic data and historical energy consumption data to generate an operation reference interval, the expert system model is constrained by rules to prohibit reading of the historical energy consumption data, and a theoretical minimum energy consumption value is derived based on inherent physical parameters of equipment. According to the method, the real-time energy consumption is doubly compared with the double references, so that overfitting of a data driving model to bad historical data is avoided, and systematic decoupling and recognition of operation and maintenance abnormal waste and structural waste are realized.

Description

TECHNICAL FIELD

[0001] The application relates to an artificial intelligence-based building energy-saving data mining and analysis processing system and belongs to the technical field of computer data processing. BACKGROUND

[0002] In current computer system applications based on specific computing models, an important direction is to use a machine learning model to process massive multi-dimensional building energy consumption data, so as to replace simple threshold values or fixed rules relied on by a traditional data statistical system, and further mine a complex nonlinear relationship among energy consumption, outdoor temperature, building occupancy and other factors, and establish a dynamic energy consumption benchmark model. However, in practice, it is found that a data-driven model has an overfitting problem for historical data, and the reason is that the result of the machine learning model is strongly dependent on the quality of the training sample when the historical energy consumption data used for model training is the result of the joint action of long-term bad operation habits or unreasonable device configuration, and the dynamic benchmark generated by the model trained also tends to reproduce this high energy consumption state, which makes it difficult for the system to identify the structural waste hidden in the historical normal operation appearance.

[0003] Even if the field tries to introduce more dimensional external feature data such as weather and work history to optimize the model, the target variable fitted is still the total energy consumption value that lacks a clear physical meaning and represents the historical result. The computing model lacks necessary physical constraint information, for example, it cannot judge in the training whether the real operation efficiency of the historical energy consumption value corresponding to a specific external condition is in a reasonable range, resulting in that the data mining process of the model tends to reproduce the historical behavior rather than follow the physical law. The prior art has not fundamentally solved this data dependence problem even if it has turned to a more complex artificial intelligence-based control level. For example, the Chinese patent application CN111025895A discloses an artificial intelligence-based building energy-saving control system, which proposes a data-driven control model. The artificial intelligence module learns historical data (including building use, external weather and energy station operation state) to find the optimal settings and control instructions of the energy station device. However, this technical idea fundamentally still belongs to fitting of historical data, and the so-called optimal result is strongly dependent on the training sample. It cannot avoid the pollution caused by the bad operation habits or unreasonable device configuration contained in the historical data itself. The model lacks an objective reference benchmark based on physical laws independent of historical data, so its optimization result is at most a reproduction of the history optimal, and it cannot judge the real gap between the history optimal and the physical optimal, which makes it also difficult to identify and eliminate the structural waste.

[0004] Therefore, how to design a new data processing system and method, through the combination of specific calculation models, introduce constraint information representing physical operation efficiency into the machine learning model in the data processing stage, so that it has the ability to analyze the rationality of historical data in the training process, and then overcome the excessive fitting of bad historical data, realize the effective identification of structural waste, become the technical problem to be solved by the present application. SUMMARY

[0005] The application provides an artificial intelligence-based building energy-saving data mining and analysis processing system, which mainly aims to solve the problem that existing calculation models cannot identify structural waste due to excessive fitting of historical data when processing building energy consumption data.

[0006] To achieve the above purpose, the application provides an artificial intelligence-based building energy-saving data mining and analysis processing system, which comprises: An energy efficiency index calculation and feature engineering module is configured to calculate a set of energy efficiency feature data representing the physical operation efficiency of the equipment based on the collected raw data and according to a preset energy efficiency calculation formula; An AI energy consumption prediction model module, which is a data-driven model based on machine learning, is configured to read historical energy consumption data in the storage and combine the energy efficiency feature data as training input, and after model training and fitting, generate a running reference energy consumption interval representing historical running habits and including upper and lower limit values; An energy consumption data deviation analysis module, which is built-in a knowledge-driven model based on an expert system, is configured to be pre-designed with calculation rules to prohibit reading historical energy consumption data; instead, it retrieves the device inherent physical parameters stored in the knowledge base and combines the real-time received dynamic boundary condition data to deduce the theoretical minimum energy consumption value representing the physical law constraint according to the thermodynamic formula; the energy consumption data deviation analysis module is also configured to perform a double comparison calculation rule when receiving real-time energy consumption data, which includes: comparing the real-time energy consumption data with the upper limit value of the running reference energy consumption interval for the first time, calculating the first deviation, which is defined as operation and maintenance abnormal waste; and at the same time, comparing the real-time energy consumption data with the theoretical minimum energy consumption value for the second time, calculating the difference between the real-time energy consumption data and the theoretical minimum energy consumption value, which is defined as structural waste.

[0007] Preferably, the energy efficiency index calculation and feature engineering module is configured to calculate energy efficiency feature data, which includes at least one of total cooling capacity, total heat dissipation, cold station EER, cold unit COP, refrigeration pump delivery coefficient and cooling pump delivery coefficient.

[0008] Preferably, the AI ​​energy consumption prediction model module is configured to read historical energy consumption data, which includes one or more of the following: historical energy consumption data of this project, energy consumption data of similar projects in the same group, and external reference energy consumption data; the AI ​​energy consumption prediction model module is also configured to start training and generate a predicted energy consumption dataset covering future periods after receiving an instruction to execute the algorithm prediction.

[0009] Preferably, the system also includes an energy consumption target value decomposition and calculation module, which is configured to: after receiving an overall energy consumption target value, treat it as a data sum to be processed; and call the predicted energy consumption dataset generated by the AI ​​energy consumption prediction model module, which contains the predicted daily energy consumption ratio; the energy consumption target value decomposition and calculation module automatically splits the overall energy consumption target value into each day within the target period according to the daily energy consumption ratio.

[0010] Preferably, the energy consumption data deviation analysis module is configured to retrieve the inherent physical parameters of the equipment stored in the knowledge base. The inherent physical parameters of the equipment include at least one of the rated power read from the equipment nameplate, the COP efficiency curve provided by the manufacturer, and the pipeline resistance coefficient.

[0011] Preferably, the energy consumption data deviation analysis module is configured to combine real-time received dynamic boundary condition data, which includes outdoor temperature data measured by sensors and actual total indoor cooling load demand data.

[0012] Preferably, the energy consumption data deviation analysis module is configured to execute a dual comparison calculation rule. The calculation of the first deviation in the dual comparison calculation rule specifically includes: when the real-time energy consumption data is higher than the upper limit of the operating reference energy consumption range. At times, computational maintenance is exceptionally wasteful. = - ,in This is real-time energy consumption data.

[0013] Preferably, the energy consumption data deviation analysis module is also configured to mark the real-time energy consumption data as too low when the real-time energy consumption data is lower than the lower limit of the operating reference energy consumption range.

[0014] Preferably, the system also includes a data visualization module, which is configured to: receive the operation and maintenance anomaly waste and structural waste output by the energy consumption data deviation analysis module; and on a data display interface, divide the height of the real-time energy consumption data into at least two segments for visualization display, where the first segment represents the theoretical minimum energy consumption value and the second segment represents structural waste.

[0015] Preferably, the data visualization module is also configured to: when the operational and maintenance anomaly waste is greater than zero, overlay a third segment on the height of the real-time energy consumption data, the third segment representing the operational and maintenance anomaly waste; and configure different visualization labels for the theoretical minimum energy consumption value, structural waste and operational and maintenance anomaly waste.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. In the data processing flow, the energy efficiency index calculation and feature engineering module is first run to process the raw collected data into a set of energy efficiency feature data that characterize the operating efficiency of the equipment, such as the COP of the chiller unit or the EER of the chiller plant, according to the preset physical formula. These feature data with definite physical meaning are used as the training input of the machine learning model, so that when the subsequent AI energy consumption prediction model module establishes a dynamic reference interval, the data fitting basis is no longer the isolated historical total energy consumption, but the intrinsic relationship between the external energy consumption environment and the physical efficiency of the equipment. This avoids the model blindly fitting historical bad operating habits from the data source level.

[0017] 2. Since the training process of the AI ​​energy consumption prediction model already includes deep learning of energy efficiency characteristic data, the dynamic operating reference energy consumption range it generates inherently contains the expectation of reasonable physical efficiency under specific operating conditions. Therefore, when the energy consumption data deviation analysis module compares the real-time energy consumption data with this reference range, the data analysis results output by the system can not only identify sudden maintenance waste that deviates from historical habits, but also reveal the structural waste hidden in the regular energy consumption fluctuations caused by the decline in equipment physical efficiency (such as COP).

[0018] 3. The data processing method adopted by this system combines rule-based energy efficiency index calculation (corresponding to data preprocessing in expert systems) with data-driven machine learning models (corresponding to pattern discovery in machine learning), forming a specific combination of computational models. The former provides the latter with a high-dimensional analytical perspective with physical constraints, while the latter compensates for the former's inability to handle complex nonlinear relationships, making the final output data analysis insights possess both the rigor of physical rules and the dynamic adaptability of data mining. The energy consumption target value decomposition calculation module in the system does not base its calculation on historical statistical averages, but directly calls the prediction dataset generated by the AI ​​energy consumption prediction model, which already considers future environmental characteristics. After receiving a total energy consumption target value, the system splits the data according to the future daily energy consumption ratio calculated by the AI ​​model, making the target value decomposition process itself a prediction-based data processing application. Attached Figure Description

[0019] Fig. 1 This is a diagram of the dual-reference data processing and waste decoupling architecture of the system of this invention; Fig. 2 This is a diagram showing the relationship between the theoretical minimum energy consumption and structural waste of the energy consumption components of this invention. Fig. 3 This is a diagram illustrating the causes of two types of energy waste in operation and maintenance, and structural issues, as presented in this invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments. It should be understood that the described embodiments are only some embodiments of this invention, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0021] This invention discloses an AI-based building energy conservation data mining and analysis system. Within a specific computational model framework, it constructs a computer system comprising multiple collaborative data processing modules. The system architecture includes an energy efficiency index calculation and feature engineering module as a data preprocessing unit, responsible for transforming raw collected data into energy efficiency characteristic data with physical meaning. It also concurrently sets up a machine learning-based data-driven model, namely an AI energy consumption prediction model module, and an expert system-based knowledge-driven model, namely an energy consumption data deviation analysis module. The AI ​​energy consumption prediction model module is trained based on energy efficiency characteristic data and historical energy consumption data to fit and generate an operational reference energy consumption range representing historical operating habits. The energy consumption data deviation analysis module is then used to... The constraint prohibits reading historical energy consumption data. Instead, based on fixed equipment physical parameters and real-time boundary conditions, the theoretical minimum energy consumption value is derived forward. Ultimately, the system achieves systematic decoupling and analysis of energy waste from different causes by doubly comparing real-time energy consumption data with two drastically different benchmarks. In the practice of building energy consumption data processing, there is a lack of direct, physically meaningful correlation between raw sensor data and total system energy consumption, making it difficult for data-driven models to capture the true operating efficiency of equipment. To address this issue, the energy efficiency index calculation and feature engineering module in this system is configured as a front-end data processing unit. Its processor receives a set of raw data streams with clearly defined ranges and precisions from the data acquisition middleware, such as supply and return water temperature sensors from the chilled water network (precision 0.1). The module collects data from the flow meter (2% accuracy) and the power meter (0.5 accuracy) of the chiller and water pump. Based on preset energy efficiency calculation formulas stored in its internal memory, such as thermodynamic formulas, it calculates the total cooling capacity and total heat dissipation, or compares the calculated cooling capacity with the corresponding measured power consumption of the equipment to calculate the chiller COP, chiller station EER, or chilled pump delivery coefficient, which characterize the physical operating efficiency of the equipment. The output of this module is a structured energy efficiency feature dataset, which will be used as a high-dimensional feature input and called by the downstream AI energy consumption prediction model module. This allows the subsequent machine learning model to fit the data based on the intrinsic relationship between energy consumption and equipment physical efficiency, rather than the isolated total energy consumption value.

[0022] Given that traditional fixed threshold alarms cannot adapt to the nonlinear dynamic characteristics of building energy consumption changing with the external environment, this system introduces an AI energy consumption prediction model module. This module is implemented as a data-driven model based on machine learning. The module's processor is configured to read two types of data from memory as training inputs upon receiving training instructions: historical energy consumption data from the current or similar projects, and energy efficiency feature data generated by the energy efficiency index calculation and feature engineering module. By using energy efficiency feature data (such as COP and EER) as training features, the model obtains physical efficiency constraints when fitting historical energy consumption data. After model training, for example, through ensemble learning algorithms, the module's processor is configured to generate a dynamic operating reference energy consumption interval that includes upper and lower limits. This interval is determined, for example, by calculating the 95% confidence interval of the predicted value, and its upper limit is defined as... The lower limit is defined as This interval represents the historical operating habits learned by the system based on historical data and serves as the first benchmark in the dual comparison. However, machine learning models that rely solely on historical data fitting have limited recognition capabilities due to the quality of historical data, making it difficult to identify long-standing structural waste in the data. To circumvent this limitation, this system includes a parallel energy consumption data deviation analysis module. This module is constructed as a knowledge-driven model based on an expert system. The processor of this module is constrained by preset calculation rules to prohibit reading any historical energy consumption data from the memory. This configuration aims to make its derivation process independent of historical operating data. Alternatively, the processor of this module is configured to retrieve specific knowledge... The knowledge base stores the inherent physical parameters of the equipment obtained from equipment nameplates or manufacturer manuals, such as the manufacturer's COP efficiency curves for chillers at different load rates and cooling water temperatures, the rated power of water pumps, and the fluid resistance coefficient of the piping network. During system operation, the inference engine of this expert system model receives dynamic boundary condition data (i.e., facts) in real time, such as outdoor temperature data measured by sensors and actual total indoor cooling load demand data. It then calls upon the physical parameters (i.e., rules) in the knowledge base and performs forward derivation based on physical formulas such as the first law of thermodynamics to calculate the theoretical minimum energy consumption required to meet the current load demand under the current boundary conditions, representing the constraints of physical laws. This value constitutes the second benchmark in the double alignment.

[0023] Furthermore, the core calculation procedure of the energy consumption data deviation analysis module lies in the fact that when its processor receives real-time energy consumption data... Then, immediately execute the double comparison calculation rule; the double comparison calculation rule includes: the processor performs the first comparison, and... Compared with the upper limit of the operating reference energy consumption range obtained from the AI ​​energy consumption prediction model module When comparing, Higher than At that time, the first deviation is calculated, that is ,Should Defined and flagged as operational anomalies and waste by the system, these represent sudden wastes that deviate from historical operational habits; simultaneously, the processor performs a second comparison, which... Compared with the theoretical minimum energy consumption value derived by this module itself. Compare them and calculate the difference between them, that is This difference is defined and marked as structural waste by the system, representing the portion of real-time energy consumption that exceeds the theoretical minimum energy consumption. Through this composite deviation calculation based on a dual-benchmark model, the system achieves systematic decoupling of operational anomaly waste and structural waste at the data analysis level. It should be noted that the processor of this module is also configured to, when real-time energy consumption data is monitored... Below the lower limit of the operating reference energy consumption range When this happens, real-time energy consumption data will be marked as too low to alert maintenance personnel to the risk of sensor malfunction or loss of comfort.

[0024] To further utilize the data analysis results generated by the aforementioned specific computational model, this system may also include an energy consumption target value decomposition calculation module. Logically, this module is configured to, upon receiving an overall energy consumption target value (e.g., a monthly total target), not use the historical average for the same period for decomposition, but instead call the AI ​​energy consumption prediction model module to obtain the predicted energy consumption dataset generated by the model, which already considers future environmental characteristics. Based on the future daily energy consumption proportions contained in this predicted dataset, it automatically decomposes the overall energy consumption target value into each day within the target period, making the target decomposition process itself a data processing application based on AI prediction. Furthermore, the system also includes a data visualization module, whose processor is configured to receive the output from the energy consumption data deviation analysis module. , Data on structural waste and operational anomaly waste, etc., are displayed on a single data interface, representing real-time energy consumption data. The height of the visual primitives (such as bar charts) is divided into at least two segments in the vertical direction for visualization, where the height of the first segment (e.g., the bottom) represents the theoretical minimum energy consumption value. The second segment superimposed on top represents structural waste. When the waste caused by abnormal operation and maintenance is greater than zero, a third segment representing abnormal operation and maintenance waste is superimposed on top, and different visualization labels (such as green, yellow, and red) are configured for each segment, so as to present the different causes of waste decoupled from the model to the operation and maintenance personnel.

[0025] Example 1: In an energy-saving data analysis scenario of a large commercial building that has been operating stably for many years, the building's operation and maintenance team has consistent operating habits, and its historical energy consumption data exhibits periodicity. When an analysis system relying solely on a data-driven model (such as a machine learning model) is deployed, the model reads long-term historical energy consumption data for training and fits this historical energy consumption state. The resulting dynamic energy consumption benchmark interval reproduces this state, causing the system to fail to identify energy-saving opportunities, even though the building's overall energy cost is higher than the reference level of similar buildings. The AI-based building energy-saving data mining and analysis system of this invention is applied to this scenario. The system first activates the energy efficiency index calculation and feature engineering module, calculating energy efficiency characteristic data, including chiller COP and chiller EER, based on the collected raw data. Subsequently, the AI ​​energy consumption prediction model module within the system reads historical energy consumption data and combines it with the energy efficiency characteristic data for training, generating an operating reference energy consumption interval representing the building's historical operating habits, with an upper limit of [value missing]. At the same time, the system starts the energy consumption data deviation analysis module in parallel. This module is constrained by rules to prohibit reading historical energy consumption data, and instead retrieves the inherent physical parameters of the equipment stored in its knowledge base, specifically including the COP efficiency curve and pipeline resistance coefficient provided by the chiller manufacturer.

[0026] When the system receives a set of real-time operating conditions, such as an outdoor temperature of 32 degrees Celsius... Furthermore, when the actual total indoor cooling load demand is 6000kW, the inference engine of the energy consumption data deviation analysis module, based on thermodynamic formulas and COP efficiency curves, positively derives a theoretical minimum energy consumption value that must be consumed under the current dynamic boundary conditions. For example, calculated The power consumption is 1000kW; at this moment, the system receives real-time energy consumption data. The value is 1500kW, which falls within the operating reference energy consumption range generated by the AI ​​energy consumption prediction model module. Below (For example (1550kW); the energy consumption data deviation analysis module then executes its dual comparison calculation rule; the first comparison output result of this dual comparison calculation rule is: Not higher than Therefore, the calculated operational and maintenance anomalies are wasteful. The result is equal to 0; while the second comparison of this rule calculates... and The difference between 1500kW and 1000kW is calculated, and the output result is that the structural waste is equal to 500kW. Finally, the system presents two decoupled analysis results to the management through the data visualization module: the waste due to operation and maintenance anomalies is 0, and the structural waste is 500kW.

[0027] Example 2: This example constructs a computer-based simulation test platform for a building central air conditioning system to verify the analytical capability of the dual-benchmark model of the present invention in decoupling energy waste of different natures. The platform is built based on a thermodynamic model and pre-sets the inherent physical parameters (rated power, COP efficiency curve, and pipe network resistance coefficient) of key equipment such as chillers, chilled water pumps, and cooling pumps. This simulation platform can generate two types of datasets: the first type is long-term historical energy consumption data, generated by simulating a long-standing, high-energy-consumption operating habit (e.g., chilled water pumps always operating at 100% power frequency); the second type... This data is short-cycle real-time energy consumption data. Based on simulating the aforementioned poor operating habits, a sudden operational anomaly is injected at a specific time period, such as simulating a cooling tower fan stopping for 2 hours due to a malfunction. The experiment sets up two analysis groups: a control group and the sample group of this invention. Both groups load the same long-cycle historical energy consumption data generated by the aforementioned simulation platform as training data and receive the same short-cycle real-time energy consumption data as the data to be analyzed. The control group only uses one AI energy consumption prediction model module (model) as described in the specific implementation. After training, this model generates an operating reference energy consumption range representing historical operating habits (its upper limit is...). The sample group of this invention utilizes the complete system of this invention, namely, the parallel operation of the AI ​​energy consumption prediction model module and an energy consumption data deviation analysis module. This energy consumption data deviation analysis module is constrained by rules to prohibit reading historical energy consumption data, but its knowledge base is pre-loaded with the inherent physical parameters of the simulation platform's equipment. During the experiment, both systems process the input real-time energy consumption data stream in parallel, while the processor in the control group only processes the real-time energy consumption data... Its generated Compare to calculate operational anomaly waste. The processor of the present invention's sample group executes a dual comparison calculation rule, that is, in the calculation... Meanwhile, its built-in expert system model also receives dynamic boundary conditions from the simulation platform in real time (such as real-time load and outdoor wet-bulb temperature), and calls upon the inherent physical parameters in the knowledge base to derive the theoretical minimum energy consumption value. And then calculate structural waste ( Data from two typical experimental time points were selected for analysis, and the results are shown in Table 1.

[0028] Table 1: Comparison of analysis results of the two calculation models.

[0029] As can be seen from the data in Table 1, at time point 1 (normal operation, including only historically inherent poor operating habits), the control group processor will... (1250kW) and (1300kW) comparison, calculation A value of 0 indicates normal operation; no detection was detected. (1250kW) and There is a 450kW energy consumption difference between (800kW); the sample processor of this invention performs a dual comparison and calculates... Both are 0, but according to the definition, the structural waste (1250kW-800kW) is calculated to be 450kW; at point 2 (fault injection, which includes both historical bad habits and sudden anomalies). Increased to 1600kW, control group processor calculation The power output was 250kW (1600kW-1350kW), which was deemed abnormal; the sample processor of this invention calculated the power output based on the dual comparison calculation rules. (Operational and maintenance anomalies waste) is also 250kW, and structural waste (i.e.) is also calculated. The power consumption is 750kW (1600kW-850kW). The test data objectively presents the combination of calculation models of the sample group of the present invention, which can simultaneously quantify the abnormal waste of operation and maintenance and the structural waste, while the single model of the control group can only identify the former.

[0030] To more clearly compare the differences between the dual-benchmark model combination used in this invention and the prior art, the experimental process and results of the control group in the aforementioned Example 2 are described independently as Comparative Example 1.

[0031] Comparative Example 1: This comparative example uses a computer-based building central air conditioning system simulation test platform that is completely identical to that of Example 2. It uses the same long-term historical energy consumption dataset and short-term real-time energy consumption data generated by this platform. The long-term historical energy consumption data is generated by simulating a long-term high-energy-consumption operating habit (e.g., chilled water pumps always operating at 100% power frequency). The short-term real-time energy consumption data is used to simulate regular operating periods containing historical poor operating habits, as well as fault injection periods with additional sudden operational anomalies. This comparative example uses a conventional data analysis system containing only an AI energy consumption prediction model module (model). This system corresponds to the control group described in Example 2. This conventional data analysis system is configured to: read the long-term historical energy consumption data containing the aforementioned poor operating habits as training data; after model training and fitting, generate an operating reference energy consumption range representing the historical operating habits; its upper limit is defined as... This conventional data analysis system, because it lacks the energy consumption data deviation analysis module (model) described in this invention in its system architecture, has limited calculation rules and cannot retrieve the inherent physical parameters of the equipment stored in the knowledge base (such as the manufacturer's COP efficiency curve and pipeline resistance coefficient). Therefore, this system does not have the capability to derive the theoretical minimum energy consumption value based on physical laws. The system's processor was configured to execute only a single comparison rule during trial operation, specifically: upon receiving the short-cycle real-time energy consumption data. Then, it is compared with the historical baseline upper limit value generated by the aforementioned model. Comparison. If and only if Higher than At times, system computation and maintenance are exceptionally wasteful. Using the same test time point data as Table 1 in Example 2, the analysis results of this comparative example are as follows.

[0032] At time 1 (10:00, normal operation, including only historically inherent poor operating habits), the system received the real-time energy consumption. The power is 1250kW, while the historical benchmark upper limit generated by the system during training is... It is 1300kW, because (1250kW) is not higher than (1300kW), the calculation result of the system processor is kW, judged as normal operation; at time point 2 (15:00, fault injection, including both historical bad habits and sudden anomalies), real-time energy consumption The power output rose to 1600kW due to a sudden anomaly, which corresponds to the historical upper limit. It is 1350kW, because (1600kW) higher than (1350kW), system processor calculation kW, judged as abnormal; experimental results show that the analytical capability of the single data-driven model representing the conventional technical path used in this comparative example is limited by the quality of the training samples. Because the historical energy consumption data relied upon for model training itself contains structural waste, the operating baseline generated by the model fitting is... (At time 1, the power consumption was 1300kW) This high energy consumption state was also reproduced, causing it to deviate significantly from the theoretical minimum energy consumption value constrained by the laws of physics under this operating condition. (800kW), which means that at point 1, although the conventional system can be determined to be operating in accordance with historical habits, it completely loses the ability to identify the structural waste of up to 450kW (1250kW-800kW) hidden beneath the normal appearance.

[0033] Example 3: This example combines Figs. 1 to 3 A description of an AI-based building energy conservation data mining and analysis system, such as... Fig. 1 As shown, the system starts by receiving raw data and first processes it through the energy efficiency index calculation and feature engineering module. This module calculates energy efficiency characteristic data, such as COP and EER, which characterize physical efficiency. This energy efficiency characteristic data, along with the externally input historical energy consumption data, is read by the AI ​​energy consumption prediction model module (machine learning model) to train and generate a benchmark A representing historical operating habits: the reference energy consumption range. The system also sets up a parallel energy consumption data deviation analysis module (expert system model). This module is constrained by a rule that prohibits reading historical energy consumption data and instead retrieves the inherent physical parameters of the equipment in the knowledge base. It combines dynamic boundary condition data with real-time energy consumption data and introduces the aforementioned benchmark A to perform a double comparison, thereby decoupling waste. This module finally outputs two decoupled analysis results: operation and maintenance abnormal waste and structural waste. These two results are called by the downstream data visualization module, which divides the real-time energy consumption into multiple segments for visualization.

[0034] like Fig. 2 As shown, this chart uses cooling load demand (kW) as the horizontal axis and energy consumption (kW) as the vertical axis. The chart uses three curves of different shapes to illustrate the theoretical minimum energy consumption, actual energy consumption, and structural waste. All three curves show that within the indicated cooling load demand range of 1000kW to 10000kW, the three energy consumption components increase synchronously with the load increase. In terms of numerical relationships and relative positions, the chart intuitively reveals that actual energy consumption is always the highest value, while structural waste quantifies the difference between actual energy consumption and the theoretical minimum energy consumption; that is, under this specific operating condition, actual energy consumption equals the sum of the theoretical minimum energy consumption and structural waste. Fig. 3 As shown in the figure, building energy waste is considered the final result, and its main causes are summarized into four categories: The first category is operational anomaly waste, which is specifically manifested in personnel operation errors, deviations from historical operating habits, or sudden operational anomalies; the second category is structural waste, the root of which can be traced back to long-term poor operating habits, unreasonable equipment configuration, or decreased physical efficiency of equipment; the third category is equipment and physical constraints, which relates to whether the system has not reached the theoretical minimum energy consumption, whether the actual efficiency of the equipment has not reached the optimal level, and the impact of the pipeline resistance coefficient; the fourth category is data and model limitations, specifically referring to problems in existing technologies such as overfitting to poor historical data, lack of physical constraints in the model, and poor quality of training samples.

[0035] Example 4: In the actual deployment and calibration process of the system of the present invention, in order to ensure that the operating reference energy consumption range generated by the AI ​​energy consumption prediction model module and the theoretical minimum energy consumption value derived by the energy consumption data deviation analysis module have engineering reproducibility and computational determinism, a series of standardized model construction and parameter calibration procedures need to be executed. This procedure aims to eliminate the ambiguity in model selection, training process and key parameter setting. For the construction of the AI ​​energy consumption prediction model module, its initial state definition includes obtaining historical hourly energy consumption data (including total power consumption, etc.) of the target object (e.g., the central air conditioning system of a commercial building) for at least 12 consecutive months. The system acquires cooling load data, concurrent outdoor meteorological data (dry-bulb temperature, wet-bulb temperature), and work calendar characteristics (whether it is a weekday or a holiday). It also obtains energy efficiency characteristic data for the corresponding time period, such as the chiller station's EER and chiller unit's COP, calculated by the energy efficiency index calculation and feature engineering modules. The system's functional specifications require a computing platform capable of performing machine learning model training, such as a server equipped with a processor with a main frequency of at least 2.0 GHz and more than 8 GB of memory. The specific model training process includes: First, data preprocessing, cleaning historical data, handling missing and outlier values, and converting time... The process involves four main steps: First, one-hot encoding of categorical features such as time stamps and work calendars. Second, feature selection, which combines preprocessed historical energy consumption data, meteorological data, calendar features, and energy efficiency features (such as COP and EER) into a feature set. Third, model selection and training, using the Gradient Boosting Regressor (GBR) algorithm as the machine learning model. The dataset is divided into training and validation sets at 80% and 20% respectively. The GBR model is trained using the training set, and key hyperparameters (such as the number of trees, learning rate, and maximum depth) are optimized using grid search and 5-fold cross-validation to minimize the root mean square error (RMSE) on the validation set. Fourth, baseline interval generation, using the trained GBR model to predict energy consumption for future periods, obtains a predicted value sequence, and calculates the upper limit of the dynamic reference energy consumption interval based on the standard deviation of the predicted value sequence or the distribution of historical prediction errors (e.g., taking the 95th percentile). and lower limit value .

[0036] Theoretical minimum energy consumption value in the energy consumption data deviation analysis module The core of this calculation lies in the construction of the expert system's knowledge base and the determinism of its reasoning logic. The operational process for constructing the knowledge base includes: First, equipment parameter input. This involves structurally storing the nameplate parameters (rated power, rated flow rate, rated head) of key energy-consuming equipment in the central air conditioning system, such as centrifugal chillers, chilled water pumps, cooling water pumps, and cooling tower fans, as well as performance curves provided by the manufacturers. For example, the two-dimensional performance curve of chiller COP changing with load rate and cooling water inlet temperature, and the curve of pump efficiency changing with flow rate, are stored in the database. For instance, the COP curve can be stored as a lookup table indexed by load rate and cooling water temperature. Second, system network characteristic calibration. This involves obtaining the resistance coefficients of the chilled water and cooling water network systems through on-site testing or fluid simulation. And store it in the knowledge base; The computational reasoning procedure is solidified into a series of sequential calculation steps based on physical laws: assuming that the currently received dynamic boundary condition data is the total indoor cooling load demand. and outdoor wet-bulb temperature The design temperature difference between chilled water supply and return is: The design temperature difference between the cooling water supply and return water is: The first step is to calculate the required chilled water flow rate. ,in The specific heat capacity of water, The density of water is given; the second step is to calculate the theoretical minimum power of the chilled water pump. ,in This is the resistance coefficient of the chilled water pipe network. The third step is to determine the pump efficiency at the corresponding flow rate obtained from the knowledge base; based on the outdoor wet-bulb temperature. Based on the approximation of the cooling tower design, determine the cooling water supply temperature. The fourth step is to query the current load from the knowledge base. and cooling water inlet temperature The corresponding theoretical maximum coefficient of performance of the chiller unit Fifth step: Calculate the theoretical minimum power consumption of the chiller unit. Similarly, calculate the cooling water flow rate. Theoretical minimum power of cooling water pump and the theoretical minimum power of cooling tower fans (Based on heat dissipation) and (Calculations are performed); ultimately, the theoretical minimum energy consumption value is determined to be the sum of the theoretical minimum power of each major energy-consuming component, i.e. .

[0037] Waste due to abnormal operation and maintenance The determination threshold, for example, as mentioned in the specific implementation method. or Its parameters or The determination of energy consumption also needs to follow standardized calibration procedures; one feasible procedure is to select a period of historical operating data in which maintenance personnel have confirmed that there are no known faults or abnormal operations (e.g., a continuous month), and calculate the real-time energy consumption during that period. Compared to the upper limit of model predictions during the same period Positive deviation sequence (considering only) (in the case of...); then, calculate the standard deviation of the positive deviation sequence. ;parameter It can be set to a predetermined multiple of the standard deviation, for example, setting... This setting is based on statistical principles, intending to consider deviations falling outside three standard deviations as low-probability events, i.e., potential operational anomalies; or, the parameter It can be configured according to management needs, for example, setting This means that real-time energy consumption is allowed to fluctuate by 10% above the historical baseline upper limit; similarly, there is a threshold for determining excessively low energy consumption. or Its parameters or It can also be determined by analyzing historical negative deviation data or by based on the minimum energy consumption level that is not allowed in actual operation.

[0038] Example 5: In the engineering application deployment phase of the system of the present invention, a correspondence is established between the inherent physical parameters of the equipment stored in the knowledge base of the energy consumption data deviation analysis module and the actual system characteristics, especially the pipeline resistance coefficient. A pre-deployment calibration procedure must be performed. The initial state of this procedure is defined as the central air conditioning system being installed and operating stably. The enabling environment includes flow meters, pressure sensors (accuracy class not less than 0.5%), and data recording equipment. The calibration process involves selecting several stable operating points covering commonly used flow ranges during system commissioning and operation. By adjusting the pump frequency or valve opening, the flow rate is controlled at 30%, 50%, 70%, and 90% of the rated flow rate. At each stable operating point, the total flow rate of the corresponding pipe network (chilled water or cooling water pipe network) is recorded simultaneously. and the pressure difference at both ends of the pipeline network Based on the fluid resistance formula Using the recorded multiple sets ( , The actual resistance coefficient of this specific pipeline system is calculated by fitting the data points using the least squares method. and the calculated actual The values ​​are updated in the knowledge base of the energy consumption data deviation analysis module, replacing the initial design values. For the COP efficiency curve of the chiller unit, several typical load points can be selected during the system commissioning phase for actual performance testing, and the test results can be compared with the curve provided by the manufacturer. If there is a deviation, the COP lookup table in the knowledge base can be corrected or a deviation model can be established to adjust the theoretical minimum energy consumption value. The basis for the calculation.

[0039] To maintain the applicability of the reference energy consumption range output by the AI ​​energy consumption prediction model module over time, the system is configured with a timeliness guarantee and reconstruction mechanism. This mechanism is set to be triggered periodically (every month) or when the model's predictive performance declines beyond a preset threshold (the root mean square error (RMSE) of prediction increases by more than 15% compared to the baseline value for a consecutive week). Upon triggering, the system automatically performs a reconstruction step, updating the data window by selecting validated historical energy consumption data from the most recent period (the past 12 months), corresponding meteorological data, calendar features, and data generated by the energy efficiency index calculation and feature engineering module. The energy efficiency characteristic data is used to form a new training dataset. Then, the model is retrained using the model training operation process described in Example 3. The AI ​​energy consumption prediction model module (GBR model) is retrained using the updated training dataset to obtain an updated prediction model. Finally, model validation and switching are performed. The performance of the updated model and the currently running model is compared using a reserved recent validation dataset (data from the past month). Only when the RMSE of the updated model on the validation set is lower than that of the current model will the system deploy the updated model as an online running model to generate reference energy consumption intervals for subsequent periods. and This mechanism enables AI energy consumption prediction models to adapt to changes in building operating characteristics or the external environment through data updates and model iterations.

[0040] Example 6: During the initial deployment and long-term operation and maintenance of the system of this invention, in order to handle sensor data noise and ensure the stability of the input of the AI ​​energy consumption prediction model module, a data preprocessing procedure needs to be executed. This procedure applies to the raw sensor data stream, such as energy consumption, load, and temperature, which serves as the input for model training and real-time prediction. The operation process includes: First, setting a sliding time window, the size of which is determined according to the data sampling frequency, for example, 15 minutes for minute-level data, and calculating the mean and standard deviation of each sensor data within the window; Second, applying... The criteria (or other statistical outlier detection methods, such as IQR) identify and mark data points within the window that exceed the mean plus or minus 3 standard deviations as potential outliers; the third step is to fill and replace the marked outliers with linear interpolation or moving averages based on nearby time points to generate a cleaned data stream for use by the AI ​​energy consumption prediction model module.

[0041] To ensure that abnormal operation and maintenance waste Threshold parameters for judgment (such as parameters) The multiplier factor 3 in the system can adapt to changes in building operation characteristics. The system is also configured with a threshold parameter adaptive adjustment procedure; this procedure is set to be executed periodically (e.g., quarterly), and its operation process includes: First, retrieving all cases of operational anomalies and waste identified by the system in the past cycle. Event logs and their corresponding deviations The first step involves numerical analysis of the deviation values. This analysis correlates these deviation values ​​with actual operational errors not caused by equipment failures confirmed by maintenance personnel within the cycle. The result is a statistical analysis of the false alarm rate (normal fluctuations are judged as abnormal) and the missed alarm rate (real anomalies are not detected) under the current threshold settings. The second step involves adjusting the multiplier factor according to a pre-defined strategy. For example, if the false alarm rate exceeds 5%, the multiplier factor is appropriately increased (e.g., from 3 to 3.5). If the missed alarm rate exceeds 10%, the multiplier factor is appropriately decreased (e.g., from 3 to 2.5). The adjusted multiplier factor is then applied to subsequent calculations for determining operational anomalies and waste. This procedure allows for dynamic adjustment of the sensitivity of anomaly detection.

[0042] In addition, to address the issue of the energy consumption data deviation analysis module in calculating the theoretical minimum energy consumption value Key sensor data that may be encountered at that time (such as total indoor cooling load demand) or outdoor wet-bulb temperature In the event of temporary data loss or data quality verification failure, the system has a built-in set of data completion and calculation degradation procedures; when real-time monitoring detects calculation errors... When necessary sensor data becomes invalid, the system first attempts to use the data value from the previous valid moment for a short timescale (e.g., the missing time is no more than 5 minutes) as a substitute for calculation. If the data remains invalid for a longer period (e.g., more than 5 minutes), the system then queries the historical database for the average load or temperature under the corresponding conditions, based on the current time, day of the week, and season information, as a substitute input. The procedure estimates the sensor data and simultaneously alerts maintenance personnel to abnormal sensor data; this ensures that even in cases where some sensor data is abnormal, The calculations can still be performed in a degraded but continuous manner.

[0043] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0044] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A building energy conservation data mining and analysis system based on artificial intelligence, characterized in that, The system includes: An energy efficiency index calculation and feature engineering module is configured to: calculate a set of energy efficiency feature data characterizing the physical operating efficiency of the equipment based on the collected raw data and according to the preset energy efficiency calculation formula; An AI energy consumption prediction model module is a data-driven model based on machine learning. The AI ​​energy consumption prediction model module is configured to: read historical energy consumption data in the memory and combine it with energy efficiency characteristic data as training input. After model training and fitting, it generates an operating reference energy consumption range that represents historical operating habits and includes upper and lower limits. An energy consumption data deviation analysis module is included. This module incorporates a knowledge-driven model based on an expert system. It is configured to: be constrained by preset calculation rules to prohibit reading historical energy consumption data; instead, retrieve the device's inherent physical parameters stored in the knowledge base and combine them with real-time received dynamic boundary condition data to derive the theoretical minimum energy consumption value representing the constraints of physical laws based on thermodynamic formulas; and be configured to: upon receiving real-time energy consumption data, execute a dual comparison calculation rule including: performing a first comparison between the real-time energy consumption data and the upper limit of the operating reference energy consumption range to calculate the first deviation, which is defined as abnormal operational waste; and simultaneously performing a second comparison between the real-time energy consumption data and the theoretical minimum energy consumption value to calculate the difference between the two values, which is defined as structural waste.

2. The artificial intelligence-based building energy-saving data mining and analysis system according to claim 1, characterized in that, The Energy Efficiency Index Calculation and Characteristic Engineering module is configured to calculate energy efficiency characteristic data, which includes at least one of the following: total cooling capacity, total heat dissipation, chiller plant EER, chiller unit COP, chiller pump delivery coefficient, and cooling pump delivery coefficient.

3. The artificial intelligence-based building energy-saving data mining and analysis system according to claim 1, characterized in that, The AI ​​energy consumption prediction model module is configured to read historical energy consumption data, which includes one or more of the following: historical energy consumption data of this project, energy consumption data of similar projects in the same group, and external reference energy consumption data. The AI ​​energy consumption prediction model module is also configured to start training and generate a predicted energy consumption dataset covering future periods after receiving an instruction to execute the algorithm prediction.

4. The artificial intelligence-based building energy-saving data mining and analysis system according to claim 3, characterized in that, The system also includes an energy consumption target value decomposition and calculation module, which is configured to: after receiving an overall energy consumption target value, treat it as a data sum to be processed; and call the predicted energy consumption dataset generated by the AI ​​energy consumption prediction model module, which contains the predicted daily energy consumption ratio. The energy consumption target value decomposition calculation module automatically breaks down the overall energy consumption target value into each day within the target period based on the daily energy consumption ratio.

5. The artificial intelligence-based building energy-saving data mining and analysis system according to claim 1, characterized in that, The energy consumption data deviation analysis module is configured to retrieve the inherent physical parameters of the equipment stored in the knowledge base. The inherent physical parameters of the equipment include at least one of the rated power read from the equipment nameplate, the COP efficiency curve provided by the manufacturer, and the pipeline resistance coefficient.

6. The artificial intelligence-based building energy-saving data mining and analysis system according to claim 1, characterized in that, The energy consumption data deviation analysis module is configured to combine real-time received dynamic boundary condition data, which includes outdoor temperature data measured by sensors and actual total indoor cooling load demand data.

7. The artificial intelligence-based building energy-saving data mining and analysis system according to claim 1, characterized in that, The energy consumption data deviation analysis module is configured to execute a dual comparison calculation rule. The calculation of the first deviation in the dual comparison calculation rule specifically includes: when the real-time energy consumption data is higher than the upper limit of the operating reference energy consumption range. At times, computational maintenance is exceptionally wasteful. = - ,in This is real-time energy consumption data.

8. The artificial intelligence-based building energy-saving data mining and analysis system according to claim 1, characterized in that, The energy consumption data deviation analysis module is also configured to mark real-time energy consumption data as too low when the real-time energy consumption data is lower than the lower limit of the operating reference energy consumption range.

9. The artificial intelligence-based building energy-saving data mining and analysis system according to claim 1, characterized in that, The system also includes a data visualization module, which is configured to receive the operational anomaly waste and structural waste output by the energy consumption data deviation analysis module; and to visualize the real-time energy consumption data by dividing the height into at least two segments on a data display interface, where the first segment represents the theoretical minimum energy consumption value and the second segment represents structural waste.

10. A building energy conservation data mining and analysis system based on artificial intelligence according to claim 9, characterized in that, The data visualization module is also configured to: when the operational waste is greater than zero, overlay a third segment on the height of the real-time energy consumption data, the third segment representing the operational waste; and configure different visualization labels for the theoretical minimum energy consumption value, structural waste and operational waste.

Images