Multi-model fusion collaborative energy-saving optimization control method and system based on knowledge graph

Abstract

The invention provides a multi-model fusion collaborative energy-saving optimization control method and system based on a knowledge graph, and is applied to the technical field of industrial equipment optimization control. The method comprises the steps that target knowledge graph information and real-time data of the cooling capacity requirement of a target air conditioning system are obtained, wherein the target knowledge graph information is used for representing entities, attributes and corresponding relations matched with the target air conditioning system; feature extraction processing is conducted on the real-time data information of the cooling capacity requirement of the target air conditioning system, and real-time environment parameter features and equipment running state features are generated; the real-time environment parameter features and the equipment operation state features are processed based on a target load prediction model, and cooling capacity demand information and load prediction information in the target time period are generated; and the target knowledge graph information, the cooling capacity demand information of the target time period and the load prediction information are processed based on the target DQN model, global optimization control strategy information is generated, and the global optimization control strategy information is used for adjusting system parameters of the target air conditioning system.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial equipment optimization control, and in particular to a multi-model fusion collaborative energy-saving optimization control method and system based on a knowledge graph. Background Art

[0002] Industrial equipment such as central air conditioners and air compressor stations are significant energy consumers, and their operating efficiency directly impacts a company's production costs and carbon emissions. Traditional equipment control strategies, which rely primarily on preset rules and operator experience, struggle to adapt to complex and changing operating environments and load demands, leading to energy waste and equipment wear and tear. Existing AI optimization methods mostly rely on single data-driven models and lack effective utilization of domain knowledge. This results in poor model interpretability and generalization, making it difficult to achieve ideal results in practical applications.

[0003] Current industrial equipment optimization control technology mainly relies on traditional control methods (such as PID control, fixed rule base) or single AI algorithm, and has the following significant defects: Traditional control methods are only designed for a single device or fixed working conditions, and cannot achieve multi-device coordination and global energy efficiency optimization. For example, in central air-conditioning systems, chillers and water pumps often adopt independent control strategies, resulting in a high proportion of transmission and distribution energy consumption. The rule base needs to be manually formulated and maintained (such as start and stop thresholds, temperature set points), which makes it difficult to cope with dynamic load changes and equipment aging problems, and lacks self-correction capabilities when rules conflict. It is unable to effectively handle complex scenarios with nonlinearity and multi-variable coupling (such as sudden changes in outdoor temperature and humidity, and partial load operation of equipment).

[0004] Furthermore, traditional AI algorithms rely heavily on trial-and-error training in pure reinforcement learning (RL), which can lead to ineffective exploration (e.g., frequent startups and shutdowns of units), slow convergence, and safety risks. They fail to integrate domain knowledge, resulting in strategies that violate actual operational constraints (e.g., cooling water overtemperature, current overload), and fail to conform to physical laws. They also struggle to balance conflicting objectives such as energy consumption, comfort, and equipment lifespan (e.g., excessive energy conservation can cause indoor temperature and humidity to exceed standards). They also fail to account for equipment performance degradation, leading to gradual control strategy failure.

[0005] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present disclosure, and therefore may include information that does not constitute prior art known to ordinary technicians in the field. Summary of the Invention

[0006] The present application aims to provide a knowledge graph-based multi-model fusion collaborative energy-saving optimization control method and system, which, at least to a certain extent, overcomes the problems of existing technologies. This method obtains target knowledge graph information and real-time data on cooling demand of the air-conditioning system. Feature extraction is performed on the real-time data to obtain real-time environmental parameters and equipment operating status characteristics. Using a training sample set containing historical environmental and equipment operating characteristics, an LSTM model is trained to predict future cooling demand. A knowledge graph is constructed, incorporating domain knowledge about equipment entities, attributes, and relationships. The knowledge graph information, load forecast results, and other relevant features are combined and input into a DQN model. The DQN model is trained using experience replay, a Q-learning algorithm, and a reward function to generate a preliminary control strategy. After optimization and adjustment, a global optimal control strategy is obtained, which is used to adjust air-conditioning system parameters. Using a second training sample set generated by the system based on the optimization strategy, the DQN model is iteratively updated, achieving self-learning and self-iterative optimization. The entire process forms a closed-loop control system of "monitoring-optimization-execution-feedback," enabling continuous optimization of the strategy based on changes in the environment and equipment status.

[0007] Other features and advantages of the present application will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0008] According to one aspect of the present application, a multi-model fusion collaborative energy-saving optimization control method based on a knowledge graph is provided, including: obtaining target knowledge graph information and real-time data of cooling demand of a target air-conditioning system, wherein the target knowledge graph information is used to characterize entities, attributes and corresponding relationships that match the target air-conditioning system; performing feature extraction processing on the real-time data information of cooling demand of the target air-conditioning system to generate real-time environmental parameter features and equipment operation status features; processing the real-time environmental parameter features and equipment operation status features based on a target load prediction model to generate cooling demand information and load prediction information for a target time period; processing the target knowledge graph information, cooling demand information and load prediction information for a target time period based on a target DQN model to generate global optimization control strategy information, wherein the global optimization control strategy information is used to adjust system parameters of the target air-conditioning system.

[0009] Another aspect of the present application is a multi-model fusion collaborative energy-saving optimization control device based on knowledge graph, characterized in that it includes: a data collection module for acquiring target knowledge graph information and real-time data of cooling demand of a target air-conditioning system, wherein the target knowledge graph information is used to characterize entities, attributes and corresponding relationships that match the target air-conditioning system; a data processing module for performing feature extraction and processing on the real-time data information of cooling demand of the target air-conditioning system to generate real-time environmental parameter features and equipment operation status features; processing the real-time environmental parameter features and equipment operation status features based on a target load prediction model to generate cooling demand information and load prediction information for a target time period; processing the target knowledge graph information, cooling demand information and load prediction information for a target time period based on a target DQN model to generate global optimization control strategy information, wherein the global optimization control strategy information is used to adjust system parameters of the target air-conditioning system.

[0010] According to another aspect of the present application, an electronic device is characterized in that it includes: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the above-mentioned knowledge graph-based multi-model fusion collaborative energy-saving optimization control method by executing the executable instructions.

[0011] According to another aspect of the present application, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a second processor, the multi-model fusion collaborative energy-saving optimization control method based on the knowledge graph is implemented.

[0012] The present application provides a multi-model fusion collaborative energy-saving optimization control method and system based on a knowledge graph. The server obtains target knowledge graph information and real-time data on the cooling demand of the air-conditioning system. Feature extraction is performed on the real-time data to obtain real-time environmental parameters and equipment operating status features. The LSTM model is trained using a training sample set containing historical environmental and equipment operating characteristics to predict future cooling demand. At the same time, a knowledge graph is constructed, incorporating field knowledge such as equipment entities, attributes, and relationships. The knowledge graph information, load forecast results, and other relevant features are combined and input into the DQN model. The DQN model generates a preliminary control strategy through experience replay, Q-learning algorithm, and reward function training. After optimization and adjustment, a global optimization control strategy is obtained for adjusting the parameters of the air-conditioning system. The DQN model is iteratively updated using the second training sample set generated by the system based on the optimization strategy to achieve self-learning and self-iterative optimization, forming a "monitoring-optimization-execution-feedback" closed-loop control, and continuously optimizing the strategy according to changes in the environment and equipment status.

[0013] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A flowchart of a multi-model fusion collaborative energy-saving optimization control method based on a knowledge graph is shown in an embodiment of the present application;

[0015] Figure 2 A structural schematic diagram of a multi-model fusion collaborative energy-saving optimization control device based on knowledge graph provided by an embodiment of the present application is shown. DETAILED DESCRIPTION

[0016] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0017] In one embodiment, the present application also proposes a multi-model fusion collaborative energy-saving optimization control method and system based on knowledge graph. Figure 1 The following schematically shows a flow chart of a multi-model fusion collaborative energy-saving optimization control method based on a knowledge graph according to an embodiment of the present application. Figure 1 As shown, the method is applied to the server and includes:

[0018] S101, obtain target knowledge graph information and real-time data on the cooling demand of the target air-conditioning system.

[0019] In one embodiment, the target knowledge graph information is a representation of the entities, attributes and corresponding relationships that match the target air-conditioning system. When constructing the knowledge graph, relevant information of the central air-conditioning system needs to be collected from multiple data sources such as equipment nameplates, schematics, equipment manuals, industry standards, expert experience, etc. Through this information, key entities are identified, including equipment entities (such as chillers, refrigeration pumps, cooling pumps, cooling towers, sensor nodes, etc.) and non-equipment entities (such as environmental parameters, historical strategies, etc.). Static or dynamic attributes are assigned to each entity to describe its physical characteristics or operating status. The rated power and rated cooling capacity of the chiller are static attributes, while the current load rate, operating status, etc. are dynamic attributes. At the same time, it is also necessary to describe the logical dependencies, physical couplings or control constraints between entities, such as physical connection relationships, operational coupling relationships, environmental impact relationships and optimization constraints. Afterwards, the entities, attributes, and relationships are encoded into machine-understandable triples, such as <chiller_1#, rated power, 550kW>, and efficient storage and query solutions such as the Neo4j graph database and the InfluxDB time series database are selected to facilitate the subsequent use of the knowledge graph.

[0020] Real-time data on the cooling demand of the target air-conditioning system: This real-time data is crucial to the entire energy-saving optimization control method. This data is collected from various data collected during the operation of the target air-conditioning system, including environmental parameters such as indoor temperature fluctuations, outdoor temperature extremes, indoor relative humidity change rate, outdoor absolute humidity, real-time load demand, and weather conditions, as well as equipment operating status data such as chiller, pump, and cooling tower operating characteristics. After acquiring this real-time data, feature extraction is performed. First, relevant features are extracted separately. Then, environmental-related features such as indoor temperature fluctuations and outdoor temperature extremes are integrated into real-time environmental parameter features. Equipment-related features such as chiller and pump operating characteristics are integrated into equipment operating status features. This processed feature data serves as an important basis for subsequent load forecasting and optimization control strategy generation.

[0021] The target knowledge graph information provides domain knowledge support for the entire energy-saving optimization control. It clarifies the attributes, relationships, and operational constraints of each entity in the system. The real-time data on the cooling demand of the target air-conditioning system reflects the current actual operating status of the system. Based on the target load prediction model, the cooling demand information and load forecast information for the target time period can be generated by utilizing the real-time environmental parameter characteristics and equipment operating status characteristics. Combined with the target knowledge graph information, it is processed through the target DQN model to ultimately generate global optimization control strategy information, realize the adjustment of the system parameters of the target air-conditioning system, and achieve the purpose of energy-saving optimization control. In this process, the target knowledge graph information and the real-time data on the cooling demand of the target air-conditioning system work together and are indispensable to ensure the efficient, energy-saving, and safe operation of the air-conditioning system.

[0022] S102 , performing feature extraction processing on the real-time data information of the cooling demand of the target air-conditioning system to generate real-time environmental parameter features and equipment operation status features.

[0023] In one embodiment, the real-time data information of the cooling demand of the target air-conditioning system is subjected to feature extraction processing to generate indoor temperature fluctuation amplitude characteristics, outdoor temperature extreme value characteristics, indoor relative humidity change rate characteristics, outdoor absolute humidity characteristics, real-time load demand characteristics, weather condition characteristics, as well as chiller operation characteristics, water pump operation characteristics, and cooling tower operation characteristics. The indoor temperature fluctuation amplitude characteristics reflect the stability of the indoor temperature. For example, in the air-conditioning system of a shopping mall, the indoor temperature fluctuates between 24°C and 26°C over a period of time, and its fluctuation amplitude is calculated to be ±1°C. The outdoor temperature extreme value characteristics reflect the extreme conditions of the outdoor temperature. In summer, the outdoor temperature in a certain area can reach as high as 38°C, which is the outdoor temperature extreme value characteristic for that period. The indoor relative humidity change rate characteristics describe the speed of change of indoor humidity. If the indoor relative humidity rises from 50% to 60% within one hour, its change rate can be calculated. The outdoor absolute humidity characteristic is the actual content of water vapor in the outdoor air, which is measured by professional equipment. For example, at a certain moment, the outdoor absolute humidity is 15g / m 3 Real-time load demand characteristics are determined based on the current air conditioning system's operating data and actual demand. For example, during peak office hours, the air conditioning system's real-time load demand is 500kW to meet indoor cooling needs. Weather condition characteristics include sunny, cloudy, and rainy weather types, which can be obtained from meteorological data or related sensors. For example, if the weather today is sunny,

[0024] Chiller operation characteristics cover multiple operating parameters of the chiller, such as the chiller's current load rate and operating status. A chiller with a rated power of 550kW, a current load rate of 70%, and an operating status of "operating" constitute the chiller operation characteristics. Water pump operation characteristics include the water pump's current operating frequency, real-time efficiency, etc. A certain refrigeration pump has a rated frequency of 50Hz, a current operating frequency of 45Hz, and a real-time efficiency of 80%. These data are the water pump operation characteristics. Cooling tower operation characteristics include the cooling tower's fan frequency, inlet and outlet water temperature difference, etc. For example, if the current frequency of a cooling tower fan is 35Hz and the inlet and outlet water temperature difference is 5°C, these are some of the cooling tower's operating characteristics.

[0025] The previously extracted indoor temperature fluctuation amplitude features, outdoor temperature extreme value features, indoor relative humidity change rate features, outdoor absolute humidity features, real-time load demand features, and weather condition features are integrated to form real-time environmental parameter features. The current cooling consumption value feature is extracted from the real-time cooling demand data of the target air conditioning system. This feature reflects the current cooling demand intensity of the system, such as the current hourly cooling usage, and is a key indicator for measuring the system's real-time cooling demand.

[0026] In a specific scenario, the indoor temperature fluctuation amplitude characteristic reflects the changes in indoor temperature over a certain period of time, reflecting the temperature stability. In a hospital ward area, the air conditioning system is set to 22°C. Within a certain hour, the indoor temperature monitored by the temperature sensor fluctuates between 21.5°C and 22.5°C. In this case, the indoor temperature fluctuation amplitude is ±0.5°C. This feature is very critical for places with high temperature requirements (such as hospitals and precision laboratories). Excessive temperature fluctuations may affect patient recovery or experimental accuracy.

[0027] The outdoor temperature extremes feature refers to the highest or lowest outdoor temperature reached during a specific time period. On a certain summer day in a city, the outdoor temperature reached 38°C in the afternoon; this was the outdoor temperature extreme (highest) for the day. In winter, the outdoor temperature can reach as low as -5°C. Air conditioning systems need to adjust cooling or heating power based on outdoor temperature extremes. When the outdoor temperature is extremely high, the air conditioner needs to increase cooling capacity to maintain a comfortable indoor temperature; when the temperature is extremely low, the heating capacity needs to be increased. Understanding this feature helps optimize the energy efficiency of the air conditioning system and avoid excessive cooling or heating.

[0028] The indoor relative humidity change rate characteristic indicates how quickly the indoor relative humidity changes over time. In a conference room, the indoor relative humidity is 50% before the meeting begins, and one hour after the meeting, the relative humidity rises to 60%. By calculation, it can be obtained that the indoor relative humidity change rate is (60%-50%) ÷ 1 hour = 10% / hour. Indoor relative humidity that is too high or too low will affect human comfort and may also cause damage to some equipment. Based on this characteristic, the air-conditioning system can perform dehumidification or humidification operations in a timely manner to maintain appropriate indoor humidity. The outdoor absolute humidity characteristic refers to the mass of water vapor contained in a unit volume of air. In the summer of a coastal city, at 2 o'clock one afternoon, the outdoor absolute humidity was measured by professional instruments to be 20g / m 3 Outdoor absolute humidity affects the cooling and dehumidification load of the air conditioning system. When the outdoor absolute humidity is high, the air conditioner needs to perform more dehumidification work while cooling to ensure the comfort and quality of the indoor air.

[0029] The real-time load demand characteristic represents the amount of cooling or heating required by the current air-conditioning system to meet indoor environmental requirements. In a large shopping mall, during peak business hours, due to factors such as dense crowds and full lighting equipment, the real-time load demand of the air-conditioning system may reach 1000kW. The real-time load demand characteristic is dynamically changing and is affected by multiple factors such as indoor and outdoor environment, personnel activities, and equipment use. Accurately obtaining this characteristic will help the air-conditioning system adjust the operating power in a timely manner to avoid energy waste. In the "A Multi-Model Fusion Collaborative Energy-Saving Optimization Control Method Based on Knowledge Graph - Draft", the analysis of real-time load demand characteristics and combination with other data provide a basis for optimizing the control strategy.

[0030] Weather condition characteristics include different weather types, such as sunny, cloudy, rainy, and snowy. On a particular day, the local meteorological department reported a sunny weather condition. Different weather conditions have a significant impact on the load of the air conditioning system. On sunny days, sunlight increases indoor heat gain, requiring the air conditioning system to provide more cooling. On rainy days, the outdoor temperature may drop, reducing the air conditioning load accordingly. Weather condition characteristics can help the air conditioning system adjust its operating mode in advance to achieve energy-saving optimization. Chiller operating characteristics include various chiller operating parameters and status information. For example, a centrifugal chiller with a rated power of 800 kW and a current load factor of 75% means that the actual output power is 800 kW x 75% = 600 kW. The chiller's operating status can be "operating," "standby," or "faulty." At this time, the chiller is in the "operating" state and has accumulated 10,000 operating hours. These parameters reflect the chiller's operating status and performance and are important for evaluating chiller efficiency, predicting faults, and optimizing control.

[0031] The operating characteristics of a water pump mainly involve parameters such as the operating frequency, head, flow rate, and efficiency of the water pump. The rated frequency of a chilled water pump is 50Hz, and the current operating frequency is 42Hz, which indicates the current speed of the water pump. Its rated head is 30m, and the actual head under the current working conditions is 25m, and the rated flow rate is 150m. 3 / h, the actual flow rate is 130m 3 / h, with a real-time efficiency of 82%. The operating characteristics of the water pump directly affect the water flow and transportation energy consumption of the air conditioning system. By monitoring these characteristics, the operating status of the water pump can be adjusted to improve the overall operating efficiency of the system. Cooling tower operating characteristics include the cooling tower fan frequency, inlet and outlet water temperature difference, and heat dissipation efficiency. The rated frequency of a cooling tower fan is 50Hz, and the current operating frequency is 38Hz, which determines the cooling tower's heat dissipation capacity. Its cooling water inlet temperature is 35°C, the outlet temperature is 30°C, and the inlet and outlet water temperature difference is 5°C, which reflects the cooling tower's heat dissipation effect. Under current operating conditions, the cooling tower's heat dissipation efficiency is 350kW, which reflects the amount of heat dissipated by the cooling tower per unit time. These characteristics are critical to ensuring the normal operation of the chiller and the overall performance of the system.

[0032] The indoor temperature fluctuation amplitude characteristics, outdoor temperature extreme value characteristics, indoor relative humidity change rate characteristics, outdoor absolute humidity characteristics, real-time load demand characteristics and weather condition characteristics extracted above are integrated to form real-time environmental parameter characteristics. In a specific scenario, the indoor temperature fluctuation amplitude is ±1°C, the outdoor temperature extreme value is 38°C, the indoor relative humidity change rate is 10% per hour in the past hour, and the outdoor absolute humidity is 15g / m 3, the real-time load demand is 500kW, and the weather condition is sunny. These characteristics together constitute the real-time environmental parameter characteristics at that moment. These comprehensive environmental parameter characteristics can more comprehensively reflect the current environmental conditions of the air-conditioning system, providing richer and more accurate information for subsequent load forecasting and control strategy formulation. The chiller operating characteristics, water pump operating characteristics, and cooling tower operating characteristics are integrated to obtain the equipment operating status characteristics. Taking a simple air-conditioning system as an example, the chiller's current load rate is 70% and the operating status is "operating"; the water pump's current operating frequency is 45Hz and the real-time efficiency is 80%; the cooling tower fan's current frequency is 35Hz and the inlet and outlet water temperature difference is 5°C. By integrating these chiller, water pump, and cooling tower operating characteristics, the air-conditioning system equipment operating status characteristics are formed, which describes the equipment's operating status as a whole and provides a basis for evaluating equipment operating efficiency and formulating optimization strategies.

[0033] S103: Processing the real-time environmental parameter characteristics and the equipment operation status characteristics based on the target load prediction model to generate cooling demand information and load prediction information for the target time period.

[0034] In one embodiment, a training sample set and a preset initial load forecasting model are obtained. The training sample set is used to characterize the historical environmental parameter characteristics and historical equipment operation characteristics of the target air-conditioning system. The training sample set includes the historical environmental parameter characteristics and historical equipment operation characteristics of the target air-conditioning system. Taking the central air-conditioning system of a large shopping mall as an example, hourly environmental parameters such as indoor temperature fluctuation amplitude, outdoor temperature extremes, indoor relative humidity change rate, outdoor absolute humidity, real-time load demand, and weather conditions over the past year are collected. Furthermore, equipment operation data such as the chiller load rate and operating status, water pump operating frequency and real-time efficiency, and cooling tower fan frequency and inlet and outlet water temperature difference are collected to form a training sample set. The preset initial load forecasting model can use a long short-term memory network (LSTM) model. LSTM is a special recurrent neural network (RNN) that can effectively handle long-term dependencies in time series data. Its basic structure includes an input layer, an LSTM layer, and an output layer. The input layer receives input data. The LSTM layer selectively memorizes and updates information through the collaborative work of a forget gate, an input gate, and an output gate. The output layer makes predictions based on the output of the LSTM layer. In this shopping mall case, it is assumed that the number of LSTM units of the initial setting LSTM model is 64, the learning rate is 0.001, and the batch size is 32.

[0035] The training sample set is cleaned, feature extracted, and normalized to generate preprocessed feature data. Data cleaning primarily involves processing missing and outliers in the training sample set. For example, if outdoor temperature data at a certain moment is missing, linear interpolation can be used to fill in the missing values. If the load rate of a chiller is abnormally high and it is found to be caused by a sensor failure, it can be eliminated based on domain knowledge or statistical methods. Feature extraction is the process of extracting useful information from the raw data. For example, time features such as year, month, day, hour, and day of the week are extracted from the timestamp, temperature and humidity are combined into a temperature and humidity index (THI), and the lagged values of historical cooling data are added as features. The processed feature data is normalized using the Min-Max normalization method to map the data to the [0,1] interval.

[0036] The preprocessed feature data is transformed to generate a time series data set. Data cleaning mainly deals with missing values and outliers in the training sample set. If the outdoor temperature data at a certain moment is missing, linear interpolation can be used to fill it in; if the load rate of a certain chiller is abnormally high and it is found to be caused by a sensor failure, it can be eliminated based on domain knowledge or statistical methods. Feature extraction is to extract useful information from the raw data, such as extracting time features such as year, month, day, hour, and day of the week from the timestamp, combining temperature and humidity into a temperature and humidity index (THI), and adding the lagged value of historical cooling data as a feature. The processed feature data is normalized using the Min-Max normalization method to map the data to the [0,1] interval.

[0037] The time series dataset is processed to generate the load forecasting model's prediction parameters. Using the time series dataset, a training algorithm is used to generate the load forecasting model's prediction parameters. During LSTM model training, the model continuously adjusts weights and biases using a backpropagation algorithm to minimize the error between the predicted and actual values. For example, the mean squared error (MSE) is used as the loss function. During training, the parameters are continuously optimized iteratively to gradually reduce the loss function value, thereby obtaining an optimal set of prediction parameter vectors.

[0038] Based on the load forecasting model's prediction parameter vector, the pre-set initial load forecasting model is optimized and trained to generate a trained load forecasting model. During training, the model uses the input time series data to learn long-term dependencies within the data through the LSTM layer. The fully connected layer then maps the LSTM layer outputs to the target value space and outputs the forecast results. Over multiple rounds of training, the model continuously adjusts its parameters to improve forecast accuracy. In this shopping mall case, after 100 rounds of training, the model's fit to historical data gradually improved, and the forecast error continued to decrease.

[0039] Evaluate and adjust the trained load forecasting model to generate a target load forecasting model. For example, for an air conditioning system, if the evaluation reveals large prediction errors and high Mean Sequential Error (MSE) values during high summer temperatures, model parameters may need to be adjusted. For example, increasing the number of LSTM units from 64 to 128 to enhance the model's learning ability or reducing the learning rate from 0.001 to 0.0001 to ensure more stable model training may be necessary. Alternatively, adjust the batch size to observe changes in model performance. Consider increasing the amount of training data, collecting more historical data, and enriching the model's learning samples. If the model overfits, employ early stopping (monitoring model performance on a validation set and stopping training when the validation set error no longer decreases). Alternatively, employ dropout (randomly dropping neurons during training) to prevent overfitting. After a series of adjustments and re-evaluations, a target load prediction model with better accuracy and generalization ability was obtained, which enables it to more accurately predict the cooling demand and load change trends of the target air-conditioning system under different conditions, providing a reliable basis for subsequent optimization control strategies.

[0040] In another embodiment, real-time environmental parameter characteristics and equipment operating status characteristics are processed based on a target load forecasting model to generate initial cooling demand and load forecast information. The target load forecasting model utilizes an LSTM (Long Short-Term Memory) model, a variant of a recurrent neural network (RNN) specifically designed to handle long-term dependencies in time series data. The LSTM model structure comprises an input layer, an LSTM layer, and an output layer. The input layer receives real-time environmental parameter characteristics (e.g., integrated features such as indoor temperature fluctuation amplitude, outdoor temperature extremes, indoor relative humidity change rate, outdoor absolute humidity, real-time load demand, and weather conditions) and equipment operating status characteristics (integrated features such as chiller operating characteristics, water pump operating characteristics, and cooling tower operating characteristics). In the LSTM layer, the forget gate, input gate, and output gate work together to selectively memorize and update information, effectively capturing long-term dependencies in the time series. The output layer, based on the output of the LSTM layer and mapped by a fully connected layer, outputs predicted values for future cooling demand and load, i.e., the initial cooling demand and load forecast information. Assume that in a shopping mall's air conditioning system, the LSTM model has 128 LSTM units, a learning rate of 0.001, and a batch size of 32. Real-time environmental parameter characteristics and equipment operating status characteristics from the current moment and a previous period (e.g., the past 24 hours) are input into the model. The model outputs initial cooling demand and load forecast information for the next hour. For example, the cooling demand for the next hour is predicted to be 800 kW, and the load forecast information indicates an upward load trend.

[0041] Initial cooling demand and load forecast information is categorized and organized to generate cooling demand and load forecast grouping information for different scenarios. Initial cooling demand and load forecast information is categorized based on different scenario factors, such as season, time period, and indoor occupant activity. In a shopping mall scenario, this can be divided into weekday daytime business hours, weekday nighttime non-business hours, weekend business hours, and weekend non-business hours. Initial cooling demand and load forecast information is organized according to these scenario categories to generate cooling demand and load forecast grouping information for different scenarios. For example, during the weekday daytime business hours, a collection of initial cooling demand and load forecast information is compiled for multiple time points within that period; corresponding collections are also compiled for weekend business hours.

[0042] Perform feature extraction on the group information to generate feature data sets for different groups. For each group information, further extract more representative features to form feature data sets for different groups. Extract features such as slope and peak value from the changing trends of cooling demand and load forecast data; combine with time information to extract time-related features, such as whether it is a peak electricity consumption period and the time interval since the last load change. For groups during the daytime business hours on weekdays, extract the changing slope of cooling demand during this period, and find that the rising slope of cooling demand is larger between 10 am and 11 am; also extract the peak information of load forecast during this period, etc. These features together constitute the feature data set for this group.

[0043] The characteristic datasets for different groups are processed to obtain target cooling demand and load forecasts. A decision tree algorithm is used to construct and train decision trees based on the individual features in the characteristic dataset to predict target cooling demand and load forecasts. For the grouped characteristic dataset for weekend business hours, the decision tree model uses previously extracted cooling demand change slopes and time-related features to produce more accurate cooling demand and load forecasts for that period. For example, if the cooling demand at a certain point during the weekend business hours is predicted to be 900kW, the load will continue to rise over the next two hours.

[0044] Based on the target cooling demand and load forecast results, the characteristic data sets of different groups are processed to generate attribute information for the cooling demand and load forecast corresponding to each group. This attribute information includes the reliability assessment of the cooling demand (e.g., high, medium, or low reliability) and the stability assessment of load fluctuations (stable, relatively stable, or unstable). If the cooling demand forecast results for a group are highly consistent with historical data and closely match the current environment and equipment operating status, their reliability is assessed as "high." If the load forecast results fluctuate significantly, their stability is judged as "unstable" based on relevant characteristics.

[0045] Based on the attribute information of cooling demand and load forecasts corresponding to different groups, cooling demand and load forecast information for the target time period are generated. Attribute information for each group, including weekday daytime business hours, nighttime non-business hours, weekend business hours, and weekend non-business hours, is aggregated and analyzed. If a comprehensive assessment is made over the next 24 hours, cooling demand during weekday daytime business hours is high and reliable, cooling demand during weekend business hours is lower but load fluctuations are erratic, and cooling demand during nighttime non-business hours is relatively low, cooling demand and load forecast information for the target time period (the next 24 hours) are generated. For example, it is determined that overall cooling demand over the next 24 hours will first increase and then decrease, reaching a peak between 10:00 and 14:00 on weekdays, with an estimated maximum cooling demand of 1000kW. Cooling demand during nighttime non-business hours remains relatively low, at approximately 200kW. Furthermore, load forecast information indicates that load fluctuations are significant during weekend business hours due to personnel turnover and uncertain equipment usage. Therefore, equipment scheduling and energy reserves are necessary to address potential load fluctuations.

[0046] S104: Based on the target DQN model, the target knowledge graph information, the cooling demand information in the target time period, and the load forecast information are processed to generate global optimization control strategy information.

[0047] In one embodiment, historical data related to the cooling demand of the target air-conditioning system, knowledge graph information related to the target air-conditioning system, and a preset DQN initial model are obtained. Hourly cooling demand data, outdoor temperature and humidity, indoor temperature, time information, and equipment operating status data from the past year are collected as historical data related to the cooling demand of the target air-conditioning system. At the same time, knowledge graph information of the air-conditioning system is obtained from channels such as equipment manuals, industry standards, and expert experience, including equipment entities (such as chillers, refrigeration pumps, etc.), attributes (rated power, current load rate, etc.), and relationships between them (physical connections, operational coupling relationships, etc.). A preset DQN initial model is used. DQN (DeepQ-Network) is a model in deep reinforcement learning that combines deep learning and Q learning algorithms. It mainly includes a neural network for approximating the Q-value function. The network structure typically has an input layer, a hidden layer, and an output layer. In this example, the number of neurons in the input layer is determined according to the dimension of the state representation information. The hidden layer is set to two layers, with 128 and 64 neurons in each layer, respectively. The number of neurons in the output layer is consistent with the dimension of the action space.

[0048] Historical data related to the cooling demand of the target air-conditioning system is cleaned, features are extracted, and normalized to generate time features, temperature and humidity index features, and lag features. Lag features are the lag values of historical cooling data. The collected historical data is cleaned and checked for missing and outliers. If outdoor temperature data at a certain moment is missing, linear interpolation can be used to fill in the gaps. If the load rate of a chiller is abnormally high and is found to be caused by sensor failure, it can be removed based on domain knowledge or statistical methods. Time features such as year, month, day, hour, and day of the week are extracted from the historical data. Temperature and humidity are combined into a temperature and humidity index (THI) feature. The lag values of historical cooling data are added as lag features, for example, cooling data from the previous hour and two hours are added as lag features for the current moment. The processed feature data is normalized using the Min-Max normalization method, mapping the data to the [0, 1] interval.

[0049] Graph embedding methods (such as Node2Vec and TransE) are used to convert entities and relationships in the knowledge graph into low-dimensional vectors. Taking a chiller as an example, graph embedding methods transform its relationships with other device entities and its own attribute information into a low-dimensional vector representation, which retains the key information in the knowledge graph. Assuming the Node2Vec method is used, with parameters p = 1, q = 1, a step size of 10, and a number of walks of 80, a 64-dimensional graph embedding vector is generated for subsequent integration with other features. Processed device state features (such as the current chiller load rate and water pump operating frequency), environmental parameter features (such as indoor and outdoor temperature and humidity), load forecast results (generated by the target load forecasting model, such as predicting cooling demand for the next hour), time features, temperature and humidity index features, and hysteresis features are combined with the graph embedding vector to form the state representation information of the DQN model. At a certain moment, the state representation information may include information such as the current load rate of the chiller is 70%, the outdoor temperature is 30℃, and the predicted increase in cooling demand in the next hour. This information is combined into a multidimensional vector as the input state of the DQN model.

[0050] Obtain the action space information and reward function of the DQN model. The action space information is used to represent the operating parameters of the central air conditioning system controlled by the preset DQN initial model. The reward function includes energy-saving rewards, task completion rewards, and constraint penalties. The action space information is used to represent the operating parameters of the central air conditioning system controlled by the preset DQN initial model, including the number of chillers on (discrete values such as 0, 1, or 2), chilled water outlet temperature (generally a continuous value between 7°C and 12°C), cooling water return temperature (a continuous value between 32°C and 37°C), chilled water supply and return pressure differential (a continuous value between 0.1MPa and 0.3MPa), chilled water supply and return temperature differential (a continuous value between 3°C and 5°C), cooling water temperature differential (a continuous value between 5°C and 7°C), chilled water pump frequency (a continuous value between 30Hz and 50Hz), cooling water pump frequency (a continuous value between 30Hz and 50Hz), and cooling tower frequency (a continuous value between 30Hz and 50Hz). In the air conditioning system of an office building, during the peak period in summer, two chillers may be turned on according to actual conditions, and the chilled water outlet temperature may be set to 8°C to meet the cooling demand.

[0051] The reward function includes energy-saving rewards, task completion rewards, and constraint penalties, guiding the DQN model to learn energy-saving optimization strategies. Energy-saving rewards are awarded based on the degree of reduction in system energy consumption, for example, 10 points for every 1kW reduction in energy consumption, incentivizing the model to reduce energy consumption. Task completion rewards are linked to the degree to which cooling demand is met. If the cooling demand is met at 90% or above, 100 points are awarded. For every 10% reduction in the satisfaction rate, the reward score decreases by 20 points, encouraging the model to meet indoor cooling demand as much as possible. Constraint penalties target actions that violate the constraints in the knowledge graph. For example, if the chilled water temperature falls below 7°C, 50 points will be deducted for each violation. Points will also be deducted if equipment operating parameters exceed the safe range to ensure safe and stable system operation.

[0052] The preset DQN initial model is trained using experience replay, a preset reward function, and a Q-learning algorithm to generate a target DQN model. The experience replay mechanism stores data such as the state, action, reward, and next state generated during the model's interaction with the environment in an experience replay buffer. As training progresses, the buffer accumulates data. During training, batches of data are randomly sampled from the buffer to break temporal correlations between the data and improve training stability and efficiency. For example, during the initial training phase, the experience replay buffer accumulates operating data for the air conditioning system at different times. During each training session, 128 samples are randomly sampled from this data for learning.

[0053] The preset reward function provides a clear goal-oriented approach for model learning. In the Q-learning algorithm, the model maximizes long-term cumulative rewards by continuously trying different actions. The Q-value function is used to estimate the value of taking a specific action in a certain state. Its update formula is: , where is the Q-value of taking the action in the current state; is the learning rate, which controls the step size of each update and is usually set to 0.1; is the reward obtained after performing the action; is the discount factor, which measures the importance of future rewards and is generally between 0 and 1, such as 0.95; is the maximum Q-value of all possible actions in the next state.

[0054] During training, the model selects actions based on its current state. Initially, the model may randomly select actions. As training progresses, it gradually selects the action with the highest expected reward based on its Q-value. This balance is achieved using the ε-greedy strategy. For example, initially setting ε to 0.9 means there's a 90% chance of randomly selecting an action and a 10% chance of selecting the action with the highest Q-value. As training progresses, ε is gradually reduced, making the model more inclined to select the optimal action. After numerous training iterations, the pre-set initial DQN model continuously adjusts network parameters, gradually learning the optimal control strategy and ultimately generating a target DQN model. This model accurately outputs optimized control parameters based on input state information, achieving efficient and energy-efficient control of the central air conditioning system. For example, the trained target DQN model can intelligently adjust parameters such as the number of chillers activated and the frequency of water pumps to meet cooling demand in an office building during different hours, thereby meeting cooling requirements while reducing energy consumption.

[0055] In another embodiment, the target knowledge graph information, the cooling demand information of the target time period, and the load forecast information are processed based on the target DQN model to generate preliminary control strategy information. On a certain working day in summer, the target load forecast model predicts that the cooling demand of the shopping mall will increase significantly from 14:00 to 16:00 on that day. The target DQN model receives this information and refers to the entity attributes (such as the rated power of the chiller, the rated flow of the chiller, etc.) and relationships (such as physical connection relationships and operational coupling relationships) of equipment such as chillers, refrigeration pumps, and cooling towers in the knowledge graph. The preliminary decision is to start an additional chiller before 14:00, appropriately increase the frequency of the chilled water pump to increase the water flow, and preliminarily set the chilled water outlet temperature to 8°C. This series of decisions constitutes the preliminary control strategy information.

[0056] Based on the preliminary control strategy information, the decision parameter vector within the target DQN model is processed to generate a policy optimization adjustment vector. After the preliminary control strategy information is determined, the DQN model analyzes the rationality of the preliminary decision based on the current state and expected reward. If the initial chilled water outlet temperature of 8°C is found to meet the cooling demand, the energy consumption is high and does not meet the energy conservation target, the model will adjust the decision parameter vector. For example, by appropriately raising the chilled water outlet temperature to 9°C and fine-tuning the chilled water pump and cooling water pump frequencies, the model aims to reduce energy consumption while still meeting the cooling demand. These adjustments constitute the policy optimization adjustment vector.

[0057] The strategy optimization adjustment vector is parsed and converted to generate global optimization control strategy information. Once the strategy optimization adjustment vector is determined, it is parsed into specific equipment control instructions. For example, information such as the adjusted chilled water outlet temperature, the number of chillers turned on, and the frequencies of each water pump and cooling tower are converted into actual executable control signals. It was ultimately determined that at 2:00 PM, two chillers would be turned on, the chilled water outlet temperature would be set to 9°C, the chilled water pump frequency would be adjusted to 45Hz, the cooling water pump frequency would be adjusted to 42Hz, and the cooling tower fan frequency would be adjusted to 38Hz. These precise control parameters constitute the global optimization control strategy information, which is used to adjust the system parameters of the mall's air conditioning system to achieve efficient and energy-saving operation.

[0058] In another embodiment of the present application, the target knowledge graph information, the cooling demand information in the target time period, and the load forecast information are processed based on the target DQN model to generate global optimization control strategy information, further comprising:

[0059] Acquire a second training sample set, wherein the second training sample set is generated by the target air-conditioning system based on global optimization control strategy information;

[0060] The target DQN model is iteratively updated based on the second training sample set to generate a self-learning and self-iteratively optimized DQN model, and the self-learning and self-iteratively optimized DQN model is used as the latest target DQN model.

[0061] In one implementation, within a knowledge graph-based multi-model fusion collaborative energy-saving optimization control system, obtaining a second training sample set and iteratively updating the target DQN model are key steps in improving model performance and achieving continuous optimization. This process enables the model to better adapt to the complex and changing operating environment of air-conditioning systems and continuously optimize control strategies. The second training sample set is derived from data generated by the target air-conditioning system based on information from the global optimization control strategy. For example, after implementing the global optimization control strategy for a central air-conditioning system, the system continuously collects operational data. This data includes operating parameters for each device under the new strategy, such as the number of chillers in operation, the temperature and flow of chilled and cooling water, and the operating frequency of water pumps and cooling towers. It also includes environmental parameters such as indoor and outdoor temperature, humidity, and real-time load demand, as well as system energy consumption data and cooling capacity supply. During a specific time period, the global optimization control strategy was implemented, with two chillers in operation and the chilled water outlet temperature set to 9°C. Under this strategy, hourly data on device operating parameters, environmental parameters, and energy consumption were recorded. These data together constitute the second training sample set. These data reflect the actual operating status of the system under the optimization strategy. Compared with the initial training sample set, they can better reflect the effects and system changes after the implementation of the optimization strategy.

[0062] The target DQN model is iteratively updated using the rich new data from the second training sample set. During the update process, the model structure remains unchanged: a neural network consisting of an input layer, hidden layers, and an output layer. The input layer receives state representation information, the hidden layer extracts and processes features, and the output layer outputs control actions. Methods similar to those used to train the initial DQN model are employed, such as experience replay, the Q-learning algorithm, and the ε-greedy strategy. Experience replay stores the state, action, reward, and next-state data from the new sample set in the experience replay buffer. Training is performed using random sampling to break data correlation and improve training stability and efficiency. During training, the Q-value function is updated using the Q-learning algorithm. The newly acquired system state at a specific moment (such as chiller load factor, outdoor temperature, and predicted cooling demand) is used as input. The current model selects an action (such as adjusting the chilled water pump frequency). After executing the action, the reward (e.g., based on energy consumption reduction and cooling demand satisfaction) and the next state are stored in the experience replay buffer. Training is then performed using random sampling to update the model's Q-value function, gradually adapting the model to the new operating data.

[0063] After multiple rounds of iterative training, the target DQN model continuously adjusts network parameters and learns new control strategies. As training progresses, the model gradually adapts to new operating environments and changes in equipment status, continuously improving its ability to handle complex operating conditions and achieving self-learning and self-iterative optimization. In a shopping mall scenario, after multiple iterative updates, the model can more accurately adjust equipment operating parameters to accommodate changes in cooling demand during different time periods and under varying indoor and outdoor environmental conditions. For example, during peak weekend business hours, the new target DQN model can more effectively control the number of chillers activated and the chilled water outlet temperature, further reducing energy consumption and improving overall system efficiency while meeting cooling demand. This self-learning and self-iteratively optimized DQN model is used as the latest target DQN model for subsequent control of the central air conditioning system. In subsequent operations, the model continuously outputs more optimized control strategies based on real-time system status monitoring, achieving more efficient and energy-efficient control of the air conditioning system and improving overall system performance.

[0064] In this application, the server obtains the target knowledge graph information and the real-time data of the cooling demand of the air-conditioning system. Feature extraction is performed on the real-time data to obtain the real-time environmental parameters and equipment operating status features. The LSTM model is trained using a training sample set containing historical environmental and equipment operating characteristics to predict future cooling demand. At the same time, a knowledge graph is constructed to incorporate field knowledge such as equipment entities, attributes, and relationships. The knowledge graph information, load forecast results, and other relevant features are combined and input into the DQN model. The DQN model generates a preliminary control strategy through experience replay, Q-learning algorithm, and reward function training. After optimization and adjustment, a global optimization control strategy is obtained, which is used to adjust the parameters of the air-conditioning system.

[0065] Furthermore, the system utilizes a second training sample set generated based on the optimization strategy to iteratively update the DQN model, enabling self-learning and self-iterative optimization. This entire process forms a closed-loop control system of "monitoring-optimization-execution-feedback," continuously optimizing strategies based on environmental and device status changes. This system overcomes the shortcomings of traditional control methods and single AI algorithms, offering significant advantages in energy conservation, equipment operating efficiency, adaptability, user comfort, intelligence, and safety. It also offers excellent scalability and is suitable for energy-saving optimization control of a wide range of industrial equipment.

[0066] In one embodiment, Figure 2 As shown, the present application also provides a multi-model fusion collaborative energy-saving optimization control device based on knowledge graph, including:

[0067] The data collection module 201 is used to obtain target knowledge graph information and real-time data on the cooling demand of the target air-conditioning system, wherein the target knowledge graph information is used to represent entities, attributes, and corresponding relationships that match the target air-conditioning system;

[0068] The data processing module 202 is used to perform feature extraction processing on the real-time data information of the cooling demand of the target air-conditioning system to generate real-time environmental parameter features and equipment operating status features; process the real-time environmental parameter features and equipment operating status features based on the target load prediction model to generate cooling demand information and load prediction information for the target time period; process the target knowledge graph information, cooling demand information and load prediction information for the target time period based on the target DQN model to generate global optimization control strategy information, wherein the global optimization control strategy information is used to adjust the system parameters of the target air-conditioning system.

[0069] Each embodiment in this application is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the evaluation of the multi-model fusion collaborative energy-saving optimization control method based on knowledge graph, electronic device, electronic device, and readable storage medium embodiment, since it is basically similar to the above-mentioned multi-model fusion collaborative energy-saving optimization control method embodiment based on knowledge graph, the description is relatively simple, and the relevant parts can be referred to the partial description of the above-mentioned multi-model fusion collaborative energy-saving optimization control method embodiment based on knowledge graph.

[0070] Although the present application is disclosed as above, the present application is not limited thereto. Any person skilled in the art may make various changes and modifications without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be based on the scope defined by the claims.

Claims

1. A multi-model fusion collaborative energy-saving optimization control method based on knowledge graph, characterized in that: include: Obtain target knowledge graph information and real-time data on the cooling demand of the target air-conditioning system. The target knowledge graph information is used to represent entities, attributes, and corresponding relationships that match the target air-conditioning system. Perform feature extraction and processing on the real-time data information of the target air-conditioning system's cooling demand to generate real-time environmental parameter features and equipment operating status features; Based on the target load prediction model, the real-time environmental parameter characteristics and equipment operating status characteristics are processed to generate cooling demand information and load forecast information for the target time period; Based on the target DQN model, the target knowledge graph information, the cooling demand information in the target time period and the load forecast information are processed to generate global optimization control strategy information, wherein the global optimization control strategy information is used to adjust the system parameters of the target air-conditioning system.

2. The method according to claim 1, wherein Perform feature extraction and processing on the real-time data information of the target air-conditioning system's cooling demand to generate real-time environmental parameter features and equipment operating status features, including: Perform feature extraction and processing on the real-time data information of the target air-conditioning system's cooling demand to generate indoor temperature fluctuation amplitude characteristics, outdoor temperature extreme value characteristics, indoor relative humidity change rate characteristics, outdoor absolute humidity characteristics, real-time load demand characteristics, weather condition characteristics, as well as chiller operation characteristics, water pump operation characteristics, and cooling tower operation characteristics; Integrate and process indoor temperature fluctuation amplitude characteristics, outdoor temperature extreme value characteristics, indoor relative humidity change rate characteristics, outdoor absolute humidity characteristics, real-time load demand characteristics, and weather condition characteristics to generate real-time environmental parameter characteristics; The chiller operation characteristics, water pump operation characteristics, and cooling tower operation characteristics are integrated and processed to generate equipment operation status characteristics.

3. The method according to claim 1, wherein Obtain the target load forecast model, including: Obtaining a training sample set and a preset load forecasting initial model, wherein the training sample set is used to characterize historical environmental parameter characteristics and historical equipment operation characteristics of the target air-conditioning system; Perform data cleaning, feature extraction and normalization on the training sample set to generate pre-processed feature data; Perform data transformation on the preprocessed feature data to generate a time series data set; Process the time series data set to generate the prediction parameters of the load forecasting model; Performing optimization training on a preset load forecasting initial model based on a load forecasting model prediction parameter vector to generate a trained load forecasting model; The trained load forecasting model is evaluated and adjusted to generate a target load forecasting model.

4. The method according to claim 1, wherein Based on the target load forecasting model, the real-time environmental parameter characteristics and equipment operating status characteristics are processed to generate cooling demand information and load forecast information for the target time period, including: Based on the target load prediction model, the real-time environmental parameter characteristics and equipment operating status characteristics are processed to generate initial cooling demand and load prediction information; Classify and organize the initial cooling demand and load forecast information to generate cooling demand and load forecast grouping information for different scenarios; Perform feature extraction on the group information to generate feature data sets for different groups; Process the characteristic data sets of different groups to obtain the target cooling demand and load forecast results; Based on the target cooling demand and load forecast results, the characteristic data sets of different groups are processed to generate attribute information of cooling demand and load forecast corresponding to different groups; Based on the attribute information of cooling demand and load forecast corresponding to different groups, cooling demand information and load forecast information for the target time period are generated.

5. The method according to claim 4, wherein Get the target DQN model, including: Obtain historical data related to the cooling demand of the target air-conditioning system, knowledge graph information related to the target air-conditioning system, and a preset DQN initial model; Perform data cleaning, feature extraction, and normalization on historical data related to the cooling demand of the target air-conditioning system to generate time features, temperature and humidity index features, and hysteresis features. The hysteresis feature is the hysteresis value of the historical cooling data. Based on the graph embedding method, the entities and relationships in the knowledge graph are converted into graph embedding vectors, where the graph embedding vectors are low-dimensional vectors; The device state features, environmental parameter features, load forecast results, time features, temperature and humidity index features, and hysteresis features are combined with the graph embedding vector to generate the state representation information of the DQN model, where the load forecast results are generated by the target load forecast model. Obtain the action space information and reward function of the DQN model. The action space information is used to represent the operating parameters of the central air conditioning system controlled by the preset DQN initial model. The reward function includes energy-saving rewards, task completion rewards, and constraint penalties. The preset DQN initial model is trained based on experience replay, preset reward function and Q-learning algorithm to generate the target DQN model.

6. The method according to claim 5, characterized in that Based on the target DQN model, the target knowledge graph information, cooling demand information in the target time period, and load forecast information are processed to generate global optimization control strategy information, including: Based on the target DQN model, the target knowledge graph information, the cooling demand information in the target time period, and the load forecast information are processed to generate preliminary control strategy information; Based on the preliminary control strategy information, the decision parameter vector inside the target DQN model is processed to generate a strategy optimization adjustment vector; The strategy optimization adjustment vector is parsed and converted to generate global optimization control strategy information.

7. The method according to claim 1, characterized in that Based on the target DQN model, the target knowledge graph information, the cooling demand information in the target time period, and the load forecast information are processed to generate global optimization control strategy information, which also includes: Acquire a second training sample set, wherein the second training sample set is generated by the target air-conditioning system based on global optimization control strategy information; The target DQN model is iteratively updated based on the second training sample set to generate a self-learning and self-iteratively optimized DQN model, and the self-learning and self-iteratively optimized DQN model is used as the latest target DQN model.

8. A multi-model fusion collaborative energy-saving optimization control device based on knowledge graph, characterized in that: The device comprises: A data collection module is used to obtain target knowledge graph information and real-time data on the cooling demand of the target air-conditioning system. The target knowledge graph information is used to represent the entities, attributes, and corresponding relationships that match the target air-conditioning system. The data processing module is used to extract features from the real-time data information of the cooling demand of the target air-conditioning system to generate real-time environmental parameter features and equipment operating status features; process the real-time environmental parameter features and equipment operating status features based on the target load prediction model to generate cooling demand information and load prediction information for the target time period; process the target knowledge graph information, cooling demand information and load prediction information for the target time period based on the target DQN model to generate global optimization control strategy information, wherein the global optimization control strategy information is used to adjust the system parameters of the target air-conditioning system.

9. An electronic device, characterized in that: include: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the knowledge graph-based multi-model fusion collaborative energy-saving optimization control method according to any one of claims 1 to 7 by executing the executable instructions.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the second processor, the multi-model fusion collaborative energy-saving optimization control method based on knowledge graph is implemented.

Images