Energy efficiency integrated improvement method and system

Abstract

The invention provides an energy efficiency integrated improvement method and system. The method comprises the steps of obtaining energy consumption data of a high-energy-consumption system; preprocessing the obtained energy consumption data based on a Hadoop technology to form an energy consumption data sample; generating a synthetic data sample based on the deep generation network of Transform and the energy consumption data sample; constructing a high-energy-consumption system environment model based on the generated synthetic data sample; training a reinforcement learning network based on the constructed high-energy-consumption system environment model to guide the working state adjustment of the frequency conversion equipment; and deploying the trained reinforcement learning network into an actual high-energy-consumption system, monitoring the state of the high-energy-consumption system in real time, and adjusting the working state of the frequency conversion equipment. According to the method, Hadoop is used for collecting and processing energy consumption data of a high-energy-consumption unit; the features of the data are enhanced on the basis of a deep generation network of Transform; a learning network is constructed and combined with a physical frequency conversion technology to adjust the power of high-energy-consumption equipment, and the energy efficiency of a high-energy-consumption system is improved.

Description

Technical Field

[0001] The present invention belongs to the field of electric power, and in particular relates to an integrated energy efficiency improvement method and system. Background Art

[0002] In today's rapidly developing social context, the development model of "high energy consumption, high pollution, and huge energy consumption" has become a common problem faced by countries around the world; this trend has not only brought unprecedented pressure to the environment, but also posed a severe test to the sustainability of natural resources; in the face of such challenges, saving energy, reducing carbon emissions, and developing a low-carbon economy have become important ways to achieve sustainable economic development and protect the ecological environment. Therefore, it is particularly important to improve the energy efficiency of high-energy-consuming industries such as electricity, steel, construction, and data centers; it is not only a key way to achieve high-quality development, reduce carbon emissions, and promote green economic transformation, but also has great significance for protecting the environment and optimizing the energy structure; energy efficiency improvements in these high-energy-consuming industries can not only significantly reduce energy consumption and carbon emissions, but also drive the green transformation and upgrading of the entire industry.

[0003] Nowadays, how to improve the energy efficiency of high-energy-consuming units through technological innovation and management innovation is one of the current hot research topics; taking high-energy-consuming units as the research object, the technical paths and strategic measures to improve the energy efficiency of high-energy-consuming units are comprehensively applied to methods such as deep learning, reinforcement learning, and computer technology. At present, Wang Yueming, Dong Jie, et al. "Energy Efficiency Improvement Method of Hot and Cold Drinking Machine Based on Multi-layer Water Distribution Plate Structure" [J]. Journal of Refrigeration, 2023, 44(01): 159-166 summarized the energy efficiency improvement methods in high-energy-consuming units from the physical level; Yi Tao "Research and Application of Electric Power Energy Efficiency Improvement Method Based on Demand-side Management" [D]. Fuzhou: Fuzhou University, 2019 proposed a method to improve the power energy efficiency based on demand-side management.

[0004] In terms of data processing methods, with the surge in data volume and the increase in complexity, traditional data processing methods can no longer meet the needs; the combination of big data processing technology and machine learning provides an effective solution for processing multi-dimensional and highly complex data. At present, Zhang Lina, Yan Tao, Jiang Kaizhong, Jiang Xinying, et al. "Unbalanced data processing method based on Gaussian mixture clustering sampling" [J]. Computer Applications and Software, 2023, 40(12): 305-311 and Li Dan "Research on oil-based mud electrical imaging logging data processing method based on machine learning" [D]. Xi'an: Xi'an Shiyou University, 2024 used machine learning methods to process multivariate or unbalanced data and achieved good results. In terms of the use of big data technology, Cai Xinlei, Cui Yanlin, Fu Weihao, et al. "Design of distributed massive power data processing method" [J]. Information Technology, 2023, 47(02): 62-68, Zhang Mingxu "Research on data mining and short-term power load forecasting based on Hadoop" [D]. Nanchang: Nanchang University, 2022, Guo Hao, Wang Qi, Song Congcong, et al. "Real-time processing method of optical intersection data based on distributed processing platform" [J]. Infrared, 2023, 44(11): 36-41, all used distributed technology to process multi-dimensional complex data, and achieved certain results in their respective experiments.

[0005] The application of generative adversarial networks and reinforcement learning in the field of power systems has attracted widespread attention. Shao Zhenguo, Zhang Chengsheng, Chen Feixiong, et al. "A Review of Generative Adversarial Networks and Their Applications in Power Systems" [J]. Proceedings of the CSEE, 2023, 43(03): 987-1004. A review of various applications of generative adversarial networks in power systems, including data generation and anomaly detection. Lu Danhong, Fan Wenyao, Yang Ting, et al. "Anomaly Detection of Power Load Data Based on Generative Adversarial Transformer" [J]. Electric Power Engineering Technology, 2024, 43(01): 157-164. A power load data anomaly detection method based on generative adversarial Transformer was proposed, demonstrating the potential of generative adversarial networks in improving data security and reliability. In addition, Tan Shanwen proposed the use of reinforcement learning to achieve intelligent frequency regulation of power systems; Xu Hongsheng, Lu Jixiang, Yang Zhihong, et al. "Incentive Demand Response Decision Optimization Model Based on Deep Reinforcement Learning" [J]. Automation of Electric Power Systems, 2021, 45(14): 97-103 discussed the demand response decision optimization model based on deep reinforcement learning, providing a new perspective and solution for the intelligent operation of power systems. The existing technical methods use distributed technology collection, and the amount of data collected is small, so a flexible data processing framework is needed that can effectively process the current scale of data and process larger scale data in the future, while reducing energy consumption as much as possible.

[0006] The physical level energy efficiency improvement method proposed by Wang Yueming and Dong Jie, and the power energy efficiency improvement method based on demand side management proposed by Yi Tao, although they can effectively adjust the balance between power demand and supply, lack the ability to respond to real-time changes in the system. With the dynamic changes of external conditions and demand, this method may not be able to adjust in time, thus affecting the optimization of energy efficiency. Zhang Lina, Yan Tao, Li Dan and others used machine learning methods to process multivariate and unbalanced data. Although they achieved certain results, these methods often face performance bottlenecks when processing extremely large-scale data. Cai Xinlei, Zhang Mingxu, Guo Hao and others used distributed technology to process multidimensional complex data. Although they solved the problem of large data volume, the existing distributed framework still has shortcomings in real-time and energy efficiency. Although the application of generative adversarial networks and reinforcement learning in power systems has shown potential, existing applications are often isolated and lack close integration with large-scale data processing technology. As a result, in practical applications, these methods may be difficult to effectively process massive real-time data, thus affecting the training and reasoning effects of the model. Therefore, there is an urgent need for an energy efficiency improvement method that dynamically adjusts the control strategy according to the real-time operating status of the system and the external environment to optimize the system energy efficiency. Summary of the invention

[0007] In order to solve the problem of difficulty in processing large-scale complex energy consumption data and the problem of improving energy efficiency of high energy consumption units in the prior art, the present invention provides an integrated energy efficiency improvement method.

[0008] The present invention adopts the following technical solution.

[0009] An integrated energy efficiency improvement method, comprising:

[0010] Obtain energy consumption data for high energy consumption systems;

[0011] Preprocessing the energy consumption data based on Hadoop's multi-dimensional measurement data parallel processing technology to form energy consumption data samples;

[0012] Generate synthetic data samples based on Transformer's deep generative network and energy consumption data samples;

[0013] Build environmental models for high-energy consumption systems based on synthetic data samples;

[0014] Based on the high-energy consumption system environment model, a reinforcement learning network is trained to guide the working state adjustment of the frequency conversion equipment; the trained reinforcement learning network is deployed to the actual high-energy consumption system to monitor the high-energy consumption system state in real time and adjust the working state of the frequency conversion equipment.

[0015] The present invention further includes the following preferred embodiments:

[0016] The energy consumption data of the high energy consumption unit includes at least one of a timestamp, a device ID, power consumption, peak power and valley power, a power factor, and voltage data.

[0017] The preprocessing of energy consumption data includes: using the MapReduce framework to process the energy consumption data in parallel; distributing the energy consumption data after parallel processing in the form of data blocks; and distributing large-scale energy consumption data using HDFS.

[0018] The establishment of the high-energy consumption system environment model includes: establishing the state space of the high-energy consumption system; establishing the action space of the high-energy consumption system; and designing the reward function of the high-energy consumption system environment model.

[0019] The establishment of the state space of the high energy consumption system includes:

[0020] Min-Max is used to normalize the state space data, and the normalized data is integrated into a state vector to represent the state of the system at a certain moment. The state vector is expressed as:

[0021] S t =[E t , E t-1 , T t , N t , f t ]

[0022] Among them, the state space data includes the current energy consumption data E t , historical energy consumption data E t-1 , temperature T t , current speed of the water pump N t , system frequency f t .

[0023] The establishment of the action space of the high-energy consumption system includes constructing the action-value function according to the following formula:

[0024] Q(s,a)←Q(s,a)+α[r+γ·max a′ Q(s',a')-Q(s,a)]

[0025] Among them, Q(s, a) represents the value of action a in the current state s of the system; α represents the learning rate; r represents the immediate reward; γ represents the discount factor; max a′ Q(s', a') represents the maximum value of all possible actions of the system in the next state s'.

[0026] Reward function R of the high energy consumption system environment model t The design is shown in the following formula:

[0027]

[0028] Among them, E t-1 and E t Respectively represent the system energy consumption at the previous moment and the system energy consumption at the current moment; N t-1 and N t They represent the pump speed at the previous moment and the pump speed at the current moment respectively; T t and T represent the current ambient temperature and the target temperature respectively; α represents the reward weight for reducing system energy consumption, β represents the penalty weight for the change of pump speed, and γ represents the reward weight for temperature control.

[0029] Based on the high-energy consumption system environment model, the reinforcement learning network is trained to guide the working state adjustment of the frequency conversion equipment, including:

[0030] Step 5.1: In the high energy consumption system environment model, use the reinforcement learning Q-Learning algorithm to train the Q function model;

[0031] Step 5.2: Update the Q table according to the Q function;

[0032] Step 5.3: Update the Q value in the Q table by introducing a loss function combined with a regularization term;

[0033] Step 5.4: Update the Q value by gradient descent method;

[0034] Step 5.5: The frequency conversion device obtains all possible actions from the Q table according to the current system state, and selects the action with the largest Q value as the current optimal action.

[0035] In step 5.3, the formula of the loss function L is as follows:

[0036]

[0037] Among them, Q(s, a) represents the value of action a in the current state s of the system, λ represents the regularization coefficient, and Q target (s, a) is the target Q value.

[0038] In step 5.3, the calculation formula of Q value is as follows:

[0039] Q target (s, a) = r + γ·max a′ Q(s', a')

[0040] Among them, Q target (s, a) represents the target value of action a in the current state s, r represents the immediate reward, γ represents the discount factor, max a′ Q(s', a') represents the maximum value of all possible actions of the system in the next state s'.

[0041] In step 5.4, updating the Q value by gradient descent includes updating the Q value according to the following formula:

[0042]

[0043] Among them, α represents the learning rate, Q(s, a) represents the value of action a in the current state s of the system, and L represents the loss function.

[0044] At the same time, the present invention also provides an energy efficiency integrated improvement system, comprising:

[0045] A data acquisition module, used to acquire energy consumption data of high energy consumption systems;

[0046] The preprocessing module preprocesses the acquired energy consumption data based on Hadoop's multi-dimensional measurement data parallel processing technology to form energy consumption data samples;

[0047] The sample data synthesis module generates synthetic data samples based on the Transformer's deep generation network and the constructed energy consumption data samples;

[0048] The environmental model building module builds the environmental model of the high-energy consumption system based on the generated synthetic data samples;

[0049] The reinforcement learning network training module trains the reinforcement learning network based on the constructed high-energy consumption system environment model to guide the working state adjustment of the frequency conversion equipment;

[0050] The equipment status adjustment module is used to deploy the trained reinforcement learning network to the actual high-energy consumption system, monitor the status of the high-energy consumption system in real time and adjust the working status of the frequency conversion equipment.

[0051] Based on the constructed high-energy consumption system environment model, the reinforcement learning network is trained to guide the working state adjustment of the frequency conversion equipment, including:

[0052] In the high energy consumption system environment model, the reinforcement learning Q-Learning algorithm is used to train the Q function model;

[0053] According to the Q function, update the Q table;

[0054] Update the Q value in the Q table by introducing a loss function combined with a regularization term;

[0055] Update the Q value by gradient descent;

[0056] The frequency conversion device obtains all possible actions from the Q table according to the current system state and selects the action with the largest Q value as the current optimal action.

[0057] The formula of the loss function L is as follows:

[0058]

[0059] Among them, Q(s, a) represents the value of action a in the current state s of the system, λ represents the regularization coefficient, and Q target (s, a) is the target Q value.

[0060] The calculation formula of Q value is as follows:

[0061] Q target (s, a) = r + γ·max a′ Q(s', a')

[0062] Among them, Q target (s, a) represents the target value of action a in the current state s, r represents the immediate reward, γ represents the discount factor, max a′ Q(s', a') represents the maximum value of all possible actions of the system in the next state s'.

[0063] Updating the Q value by gradient descent includes updating the Q value according to the following formula:

[0064]

[0065] Among them, α represents the learning rate, Q(s, a) represents the value of action a in the current state s of the system, and L represents the loss function.

[0066] A terminal comprises a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of any one of the methods.

[0067] A computer-readable storage medium stores a computer program, which implements the steps of any one of the methods when executed by a processor.

[0068] The beneficial effect of the present invention is that, compared with the prior art, the present invention uses Hadoop's distributed big data collection and processing capabilities to obtain energy consumption data of high-energy-consuming units scattered in various production links; designs a deep generation network based on Transformer to further enhance the characteristics of the collected data; and finally constructs an adaptive reinforcement learning network combined with physical frequency conversion technology to adjust the power of high-energy-consuming electrical appliances, providing a new idea and solution for improving the energy efficiency of high-energy-consuming units. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is a technical framework diagram of an integrated energy efficiency improvement method of the present invention;

[0070] Figure 2 is a schematic diagram of a variable frequency drive system in an embodiment of the present invention;

[0071] Figure 3 is a workflow diagram of the AI ​​module in an embodiment of the present invention;

[0072] FIG4( a ) is a performance comparison diagram of stand-alone processing and Hadoop processing in an embodiment of the present invention;

[0073] FIG4( b ) is a schematic diagram of the time taken by Hadoop to process large-scale energy consumption data in an embodiment of the present invention;

[0074] FIG5( a ) is a graph showing the loss value variation trend of the deep generative network generator and the discriminator in an embodiment of the present invention;

[0075] FIG5( b ) is a graph showing a trend of similarity between data generated by a deep generative network generator and source data in an embodiment of the present invention;

[0076] FIG6( a ) is a graph showing a trend of changes in the loss value of a reinforcement learning network according to an embodiment of the present invention;

[0077] FIG6( b ) is a graph showing a trend of reward value changes of a reinforcement learning network in an embodiment of the present invention. DETAILED DESCRIPTION

[0078] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The embodiments described in this application are only embodiments of a part of the present invention, rather than all embodiments. Based on the spirit of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the protection scope of the present invention.

[0079] The present invention provides an integrated energy efficiency improvement method, such as Figure 1 The figure shows a technical framework diagram of an integrated energy efficiency improvement method of the present invention, the method comprising the following steps:

[0080] Obtain energy consumption data for high energy consumption systems.

[0081] In one embodiment of the present invention, the energy consumption data of the high energy consumption unit includes a timestamp, a device ID, power consumption, peak and valley power, power factor, and voltage data.

[0082] The energy consumption data is preprocessed based on Hadoop's multi-dimensional measurement data parallel processing technology to form energy consumption data samples.

[0083] In one embodiment of the present invention, preprocessing the energy consumption data of the high energy consumption system includes: using the MapReduce framework to process the energy consumption data in parallel; distributing the energy consumption data after parallel processing in the form of data blocks; and distributing large-scale energy consumption data using HDFS.

[0084] This embodiment uses the Hadoop framework for data processing; Hadoop, as a mature big data processing platform, can effectively manage and process a large amount of energy consumption data; in order to adapt to the needs of different business scenarios, taking into account the characteristics of the data and its timeliness requirements, an integrated collection solution covering various types of energy data is proposed, aiming to optimize the management and utilization of data streams by integrating the collection process of various types of energy data, thereby improving the efficiency and economy of energy use; on the basis of energy consumption monitoring and management, the present invention specially designs a set of multi-dimensional measurement data parallel processing technology based on Hadoop; uses Sqoop to extract energy consumption data and imports it into Hadoop, uses the MapReduce framework to process the data in parallel, stores the data on HDFS, and performs distributed storage in the form of data blocks, and uses the high throughput characteristics of HDFS to quickly distribute large-scale data, together forming a comprehensive integrated energy data collection architecture.

[0085] In this architecture, as an important part of the power grid system, the data center is particularly critical for collecting and analyzing multi-dimensional information such as energy consumption data, environmental parameters, and user behavior patterns; these data are essential for understanding the energy usage of data centers, identifying energy conservation and emission reduction potential, and optimizing power grid management strategies. Based on this, the present invention uses efficient data processing technology to process these multi-dimensional data sets to ensure the stable operation of the power grid and the efficient use of energy; Hadoop, as a mature big data processing framework, can effectively store, process and analyze a large amount of multi-dimensional data from data centers; its basic steps include:

[0086] (1) Input: In the power grid monitoring system, the energy consumption data of the data center and cooling system are usually collected by smart meters in various areas, which record the power usage of each area; the energy consumption data includes timestamp, device ID, power consumption, peak and valley power, power factor, and voltage data. In the input stage, the energy consumption data is divided into manageable small blocks (256MB) and imported into the Hadoop Distributed File System (HDFS) for distributed processing.

[0087] (2) Mappers: In the Mapper stage, each Mapper is responsible for processing the data blocks assigned to it by HDFS; a Mapper is assigned to the data of a region within a certain time range; the role of the Mapper is to convert the original power consumption data into a set of intermediate key-value pairs. The Mapper converts the read data into a string and divides it into arrays according to the data type. For example, the read data is 1001 (DeviceID), 2024-08-01T12:00:00 (Timestamp), 150.5 (PowerConsumption), 180.0 (PeakPower), 130.0 (ValleyPower), 0.95 (PowerFactor), 230.0 (Voltage); first convert it into a string, "1001, 2024-08-01T12:00:00, 150.5, 180.0, 130.0, 0.95, 230.0"; then use commas to divide it into arrays of size 7.

[0088] The device ID and timestamp in each array are used as keys, and other fields are used as values; this process is the first step in data conversion.

[0089] (3) Intermediate data: After processing the data, the Mapper will output a series of intermediate key-value pairs, which will then be transmitted to the Reducers; the intermediate key-value pairs will be sorted and merged according to the key to ensure that all values ​​​​with the same key are sent to the same Reducer.

[0090] (4) Reducers: Reducers receive intermediate results from Mapper output. Once Reducers receive these intermediate values, they merge and process them to produce the final output results. Reducers receive key-value pairs from Mappers and only output one record for each unique key, thereby achieving data deduplication. When the received data has missing values, combined with deep generative networks, generative adversarial networks can be used to generate missing data samples during the data cleaning process, thereby helping to complete the data set and improve data quality and integrity.

[0091] (5) Output: The output of the final Reducer forms the final result of the MapReduce task; the final result includes the average power consumption in each time period, peak and valley power consumption, power factor, and voltage fluctuation.

[0092] Generate synthetic data samples based on Transformer's deep generative network and energy consumption data samples.

[0093] In one embodiment of the present invention, when Hadoop and a deep generative network are processing large-scale energy consumption data, the deep generative network can be used to generate synthetic data samples with representative features, wherein the representative features include time series features and abnormal features.

[0094] The time series feature means that the generated samples can reflect the periodicity in the energy consumption data, including the difference in energy consumption between weekdays and weekends, and the energy consumption trends between summer and winter.

[0095] Energy consumption is higher on weekdays, especially during the peak period at noon, and energy consumption is lower on weekends than on weekdays. When training the model, samples representing peak energy consumption at 12 noon on weekdays and samples representing low energy consumption at 12 noon on weekends are generated. In this way, the model can identify changes in energy consumption at different times.

[0096] In summer, due to the high temperature, the pump system operates intensively, and the power consumption increases significantly; in winter, the power consumption decreases significantly, especially in the early morning when the temperature is the lowest. Generate samples with high energy consumption in the summer afternoon and samples with high energy consumption in the winter morning. Through these samples, the model can learn the changes in energy consumption in different seasons.

[0097] The time series features include the difference in energy consumption between weekdays and weekends and the energy consumption trends in four seasons; the abnormal features indicate that the generated samples include some abnormal situations or extreme conditions, including energy consumption trends during extreme weather and extreme high temperatures, to enhance the robustness of the model in dealing with abnormal or extreme events.

[0098] Extreme weather includes hail, strong winds, tornadoes, thunderstorms, and tropical cyclones. On extremely hot days, the temperature reaches over 40°C, causing the pump to run at full capacity and consume more energy than usual. This extreme weather is relatively rare, so there is less relevant data. Generate abnormally high energy consumption data under extreme high temperature conditions. Through this type of data, the model can learn how to manage energy consumption under extreme conditions.

[0099] The synthetic data samples can be used to enhance the data set to improve the performance of the machine learning model. By combining with Hadoop, deep generative networks can be applied to large-scale energy consumption data processing to generate more data with richer features, providing better support for model training. Hadoop can be used for cleaning and preprocessing large-scale data, deduplication of data, and filling of missing values. Deep generative networks can be used to generate representative normal data samples, while abnormal data is different from them.

[0100] Build an environmental model for high-energy consumption systems based on synthetic data samples.

[0101] In one embodiment of the present invention, establishing a high-energy consumption system environment model includes: establishing a state space of the high-energy consumption system; establishing an action space of the high-energy consumption system; and designing a reward function of the high-energy consumption system environment model.

[0102] In this embodiment, the establishment of the state space of the high energy consumption system includes: using Min-Max to normalize the state space data, and then integrating the normalized data into a state vector to represent the state of the system at a certain moment. The state vector is expressed as:

[0103] S t =[E t , E t-1 , T t , N t , f t ] (1)

[0104] Among them, the state space data includes the current energy consumption data E t , historical energy consumption data E t-1 , temperature T t , current speed of the water pump N t , system frequency f t The temperature range is set to 20-80 degrees Celsius, the water pump speed is set to 500-3000r / min, the system frequency is set to 30-60Hz, the maximum energy consumption of the system is the highest limit state of the above parameters, and the minimum energy consumption is the lowest limit state of the above parameters.

[0105] In this embodiment, the establishment of the action space includes: discretizing the frequency adjustment and the speed adjustment into fixed step sizes, and constructing a discrete action space A as shown below: t =[Δf, ΔN], where:

[0106] Δf∈{-5,-4,-3,-2,-1,0,+1,+2,+3,+4,+5}Hz(2)ΔN∈{-500,-400,-300,-200,-100,0,+100,+200,+300,+400,+500}r / min (3)

[0108] When constructing the action space, the frequency adjustment Δf and the speed adjustment ΔN are used as control variables, the system frequency is between 30-60 Hz, and the pump speed is between 500-3000 r / min.

[0109] In one embodiment of the present invention, the reward function R of the high energy consumption system environment model t The design is shown in the following formula:

[0110]

[0111] Among them, E t-1 and E t Represents the system energy consumption at the previous moment and the system energy consumption at the current moment respectively. The more the energy consumption is reduced, the higher the reward; N t-1 and N t They represent the pump speed at the previous moment and the pump speed at the current moment respectively. If the pump speed changes too much, more energy will be consumed. Therefore, the smaller the change in pump speed, the lower the penalty. t and T represent the current ambient temperature and target temperature respectively. The closer the current ambient temperature is to the target temperature, the higher the reward. α represents the reward weight for reducing system energy consumption, β represents the penalty weight for changing the speed of the water pump, and γ represents the reward weight for temperature control.

[0112] In this embodiment, in order to avoid the problem of gradient explosion or gradient disappearance in the reinforcement learning algorithm and prevent numerical instability caused by excessive or too small reward values, the reward value is set between 10 and 20, and the reward function is updated according to the following formula:

[0113]

[0114] The reinforcement learning network is trained based on the high-energy consumption system environment model to guide the adjustment of the working state of the frequency conversion equipment.

[0115] When designing a chilled water pump system, the equipment is configured based on the maximum load demand to ensure that the system can cope with the demand during peak hours; however, actual operation often reveals that this design approach will lead to significant energy waste, because the system does not need to work at full load during most of the operating time; this over-design phenomenon, especially during non-peak hours, leads to inefficient use of energy, increased operating costs, and unnecessary burden on the environment.

[0116] In contrast, pump systems using variable frequency drives (VFDs) provide an efficient way to solve this problem; variable frequency drive pumps can adjust their operating speed in real time based on the actual needs of the system, ensuring that the pump only works at the required output; this intelligent regulation not only reduces energy waste, but also significantly improves the overall energy efficiency of the system; using a variable speed drive (VFD) to control the chilled water pump means that under partial load conditions, the pump can operate at optimal efficiency, meeting system needs while minimizing energy consumption.

[0117] The principle of variable frequency drive system is as follows Figure 2As shown in the figure, the operation interface allows the operator to set the operating parameters of the motor, such as the speed, and provides control functions such as starting, stopping and monitoring. The frequency converter is used to convert the standard sinusoidal AC power input into an adjustable frequency power supply. The variable frequency power output by the frequency converter enables the AC motor to run at the adjusted speed, thereby achieving precise control of the water pump. Ultimately, the motor converts electrical energy into mechanical energy to drive the connected mechanical equipment.

[0118] Using a VFD to control speed not only puts less stress on the pump during startup and shutdown, increasing the life of the components, but also greatly reduces the energy consumption of the motor. The power required to run the pump motor and its speed have the following relationship:

[0119] P∝N 3 (6)

[0120] Formula (3) explains the effect of the pump speed on its power. P represents the power of the pump in kW, and N represents the speed (r / min). The power is proportional to the cube of the speed.

[0121]

[0122] From formula (4), it can be seen that when N1 = 1000r / min, N2 = 500r / min, the speed is reduced by 50%, and the power is reduced by about 87.5%; when N1 = 1000r / min, N2 = 900r / min, the speed is only reduced by 10%, but the power is reduced by about 27.1%. Therefore, even a small speed reduction will lead to a significant reduction in energy consumption.

[0123] In one embodiment of the present invention, in step 5, a reinforcement learning network is trained based on the high-energy consumption system environment model constructed in step 4 to guide the working state adjustment of the frequency conversion equipment, including: Step 5.1: In the high-energy consumption system environment model, a reinforcement learning Q-Learning algorithm is used to train a Q function model, including initializing a Q table Q(s, a), where s is a state and a is an action.

[0124] Step 5.2: Update the Q table according to the Q function. The traditional Q function update formula is as follows:

[0125] Q(s,a)←Q(s,a)+α[r+γ·max a′ Q(s', a')-Q(s, a)] (8)

[0126] Among them, Q(s, a) represents the value of action a under the current state s of the system; α is the learning rate, which determines the magnitude of the system state update; r is the immediate reward; γ is the discount factor, which is used to calculate the present value of future rewards; max a′Q(s', a') is the maximum value of all possible actions of the system in the next state s'.

[0127] Step 5.3: Update the Q value by introducing the mean square error (MSE) loss function combined with the regularization term. The loss function formula is as follows:

[0128]

[0129] Where Q(s, a) is the current value; λ is the regularization coefficient, which is used to control the absolute value of the Q value to prevent the Q value from being too large or too small to avoid overfitting. target (s, a) is the target Q value, and the calculation formula is:

[0130] Q target (s, a) = r + γ·max a′ Q(s',a') (10)

[0131] r is the immediate reward; γ is the discount factor used to calculate the present value of future rewards, max a′ Q(s', a') is the maximum value of all possible actions of the system in the next state s'.

[0132] Step 5.4: Update the Q value by gradient descent, the formula is as follows:

[0133]

[0134] Where α is the learning rate, which is automatically adjusted using the Adama optimization algorithm. Q(s, a) represents the value of action a in the current state s of the system, and L represents the loss function.

[0135] During the training process, the initial value of the Q table is set to zero, and then the Q value is updated using the gradient descent method. Select the ∈-greedy strategy, select a random action with probability ∈, select the currently known best action with probability 1-∈, and generate a random number between 0 and 1. If the random number is less than ∈, select a random action; otherwise, select the currently known best action.

[0136] In the initial stage of training, ∈ is initialized to 1, the decay rate of ∈ is 0.99, and the minimum value of ∈ is 0.01. The environmental model of the high-energy consumption system is used to train the reinforcement learning network to train the working state of the variable frequency equipment. In the early stage, random actions interact with the environment, perform actions, observe state changes, and obtain rewards. After multiple rounds of training, the value of ∈ is 0.01. The model will have a very low probability of randomly selecting actions, and most of the time it will select the optimal action with the largest known Q value.

[0137] Step 5.5: Adjustment of the working state of the equipment. After the training is completed, the Q-Learning algorithm generates a Q table, in which each entry Q(s, a) represents the expected cumulative reward value of taking action a under a given state s. The steps for adjusting the working state of the equipment are as follows: At each moment, the frequency conversion device obtains all possible actions a from the Q table according to the current system state s, including frequency adjustment Δf and pump speed adjustment ΔN, and selects the action with the largest Q value as the current optimal action. In this embodiment, assuming that the current observed state is s, action a, and Q value is Q(s, a), after executing action a, the system transfers to a new state s'. According to the reward function R t Calculate the instant reward r when s transfers s′, and query the Q table to find that the state has the largest Q value max a′ Q(s', a'), calculate Q according to the formula target (s, a), and then calculate the loss value L. If the L value is too large, it means that there is still a lot of room for improvement in the strategy. Use gradient descent to update the Q value. By repeatedly calculating the loss value and updating the Q value, the model gradually learns how to choose a better strategy, thereby improving the overall performance.

[0138] The trained reinforcement learning network is deployed to the actual high-energy consumption system to monitor the status of the high-energy consumption system in real time and adjust the working status of the frequency conversion equipment.

[0139] In one embodiment of the present invention, the trained reinforcement learning network is deployed in an actual high-energy consumption system, and real-time monitoring of the high-energy consumption system state and adjustment of the working state of the frequency conversion device include:

[0140] Continuously collect status data of actual high-energy consumption systems and feed the status data back to the reinforcement learning network;

[0141] According to the real-time feedback information, the reward function weight of the reinforcement learning network is adaptively adjusted through the Adam optimization algorithm.

[0142] In this embodiment, the parameters of the Q function are continuously adjusted according to formula (8) based on the real-time feedback information.

[0143] Among them, the learning rate α is adaptively adjusted through the Adam optimization algorithm; the discount factor γ can be increased if the long-term reward is important, and the γ value can be reduced if the immediate reward is more critical; the ∈ value decays by 0.99 each iteration, and the minimum decay is 0.01.

[0144] In this embodiment, the weights α, β, and γ of the reward function described in formula (5) are appropriately adjusted according to environmental factors. For example, when the temperature is high in summer, since the temperature itself is high, the proportion of temperature γ in the reward function should be reduced to adapt to the dynamic changes of the system and continuously optimize energy utilization efficiency.

[0145] In one embodiment of the present invention, a Hadoop-AI fusion intelligent frequency conversion system is constructed based on reinforcement learning network and frequency conversion technology, aiming to improve the energy efficiency of data centers and other high-energy-consuming systems. On the one hand, by integrating deep generative networks and Hadoop big data processing capabilities, efficient management and analysis of energy consumption data is achieved; on the other hand, the power of equipment is intelligently controlled by combining reinforcement learning network and frequency conversion technology, thereby optimizing equipment operation efficiency and achieving the purpose of energy saving and emission reduction. The overall process is as follows: Figure 3 shown.

[0146] Energy consumption data including timestamp, device ID, power consumption, peak and valley power, power factor, and voltage data are collected through smart meters and input into the system. The energy consumption data is initially processed by a multi-layer perceptron (MLP) to extract key features. After feature extraction, the data is further processed by a Transformer encoder to generate data representations with time series features and abnormal features. These feature data generate synthetic data samples in a deep generative network to enhance the model training effect.

[0147] The data after feature extraction is stored in Hadoop's distributed file system (HDFS). After the data enters Hadoop, the system uses MapReduce to preprocess and analyze the data.

[0148] State observation: The reinforcement learning network observes the current system state from the environmental model. Strategy selection: Based on the current state and historical data, the optimal action is selected. Model update: According to the changes in the feedback energy consumption data after actual execution, the reinforcement learning network adjusts and updates the strategy. Value estimation optimization: The model optimizes its value function by updating the Q value and gradually approaches the optimal strategy. Dynamic environment adaptation: The reinforcement learning network ultimately outputs optimized strategies and executes these strategies through the intelligent agent. In the actual energy consumption system, the intelligent agent performs specific actions based on the strategy to adjust the frequency or speed of the water pump. The system continuously adapts to the dynamic environment through observation, updating and adjustment to maximize energy efficiency.

[0149] In one embodiment of the present invention, in order to evaluate the performance of Hadoop distributed system for data collection, it is compared with traditional single-machine collection. Data sets of different sizes are used for experiments, including 0.2GB, 0.5GB, 1GB, 3GB, 5GB and 20GB. These data sets represent energy consumption readings collected from a data center, including detailed timestamps and consumption. The processing performance of the two methods is shown in Figure 4.

[0150] As can be observed from Figure 4(a), for small-scale data sets, a single-machine system may be efficient enough and may even outperform a distributed system because a single-machine system does not involve network communication and data distribution between nodes. However, as the amount of data increases, the resources of a single-machine system (such as CPU, memory, and storage) may become a bottleneck, resulting in performance degradation. In this data center energy consumption management system, the amount of data is usually very large. When the amount of data grows, the Hadoop distributed system can increase the processing capacity by simply adding nodes. This scalability is very important for processing the growing data in the power grid. Distributed processing can effectively reduce processing time, especially when performing complex data analysis and pattern recognition tasks, its performance is far better than that of a single-machine system. Hadoop's distributed architecture enables it to split large amounts of data into smaller blocks and process them in parallel on multiple machines in a cluster. The advantages of this architecture are particularly evident when processing large-scale energy consumption data. No matter how large the amount of data is, Hadoop can coordinate multiple nodes to share the computing tasks, thereby ensuring the stability of the processing speed. As shown in Figure 4(b), even when the amount of data continues to increase, the processing speed does not drop significantly. By using Hadoop's efficient processing capabilities to analyze energy consumption data in different time periods, the operation strategy of the chilled water pump can be effectively adjusted, and accurate energy consumption analysis can be provided for frequency conversion technology, making energy consumption optimization more data-driven and accurate. This sophisticated energy consumption analysis helps to achieve more flexible energy management, reduce energy waste, and improve energy efficiency, thereby providing strong support for enterprises to save costs and reduce environmental impact.

[0151] From Figure 5(a) and Figure 5(b), we can see that after multiple rounds of training, the loss values ​​of the generator and discriminator of the deep generative network gradually decrease to a stable state, and the generated synthetic data are highly similar to the real data. These synthetic data can be used to supplement the training data set to make the data set richer and more diverse. In the energy consumption optimization scenario, these high-quality synthetic energy consumption data can be used to train more accurate and robust energy consumption prediction models. Moreover, the synthetic data has diversity in features. This diversity enables the model to learn more potential energy consumption patterns and change laws. In the energy consumption system, by generating energy consumption data under different loads and operating conditions, it can help the system better adapt to different working environments and demand changes.

[0152] In the embodiment, a reinforcement learning network is first used for training to evaluate the energy-saving effect of variable frequency technology on the chilled water pump system of the data center. The changes in the loss value and reward value of the reinforcement learning network during the training process are recorded. In Figure 6(a), it is observed that the loss value gradually decreases with the increase of training rounds, which indicates that the reinforcement learning network gradually converges during the training process and learns a better strategy. In Figure 6(b), it can be seen that with the increase of training rounds, the reward value gradually increases, which indicates that the reinforcement learning network gradually improves its performance in the environment during the learning process and obtains more positive feedback.

[0154] By training the reinforcement learning network, an intelligent agent with good performance is obtained. Next, the intelligent agent is applied to the cold water pump system of the data center to evaluate its energy saving effect. Specifically, this embodiment compares the monthly energy consumption in 2022 (without frequency conversion technology) and 2023 (with intelligent frequency conversion technology), and records the data shown in Table 2.

[0155] Table 2 Electricity consumption of a data center in 2022 and 2023 (unit: kWh)

[0156] January February March April May June 2022 584555 548118 445321 431479 465949 594312 2023 428654 403525 331675 317223 363627 446388 Energy saving 155901 144593 113646 114256 102322 147924 Energy saving percentage 26.67％ 26.38％ 25.52％ 26.48％ 21.96％ 24.89％ July August September October November December 2022 587678 590902 414660 445658 416259 619009 2023 456038 457240 327291 347123 321768 446615 Energy saving 155901 144593 113646 114256 102322 147924 Energy saving percentage 22.40％ 22.62％ 21.07％ 22.11％ 22.70％ 27.85％

[0157] In this example, the energy consumption data generally shows seasonal fluctuations, especially in summer and winter, due to the frequent use of the cooling and heating system, which has relatively high energy consumption. However, after applying the frequency conversion technology to the chilled water pump system of the data center in 2023, the energy consumption data shows a significant reduction every month, with an average energy saving percentage of 20%-30%, reaching a maximum of 27.85%. This finding confirms the effectiveness of frequency conversion technology in energy saving, especially in seasons with high energy consumption.

[0158] By analyzing the real collected data, the energy-saving potential of frequency conversion technology in data center buildings has been confirmed. Especially in situations where large HVAC systems need to be adjusted, frequency conversion technology can adjust system operation according to actual needs, thereby effectively reducing unnecessary energy consumption. The application of this technology is of great significance for improving energy efficiency, reducing operating costs and reducing environmental impact.

[0159] This embodiment uses the Hadoop framework and deep generative network to process large-scale energy consumption data. The traditional stand-alone data processing and the distributed data processing method based on Hadoop are compared. The experimental results show that as the amount of data increases, Hadoop's processing performance shows significant advantages over traditional methods, especially when processing large-scale data sets, its processing speed and efficiency are far superior to the stand-alone processing method. This discovery not only proves Hadoop's ability in big data processing, but also provides strong support for the efficient management and analysis of energy data. The deep generative network has many advantages in processing large-scale energy consumption data. It can learn the potential distribution of data and generate new data samples similar to the original data. This helps to discover potential patterns and laws in the data, provide more information for subsequent analysis and decision-making, and provide important support for the efficient management and analysis of energy data.

[0160] In addition, this embodiment demonstrates the energy-saving effect of intelligent power control by comparing and analyzing the energy consumption data of the data center chilled water pump system in 2022 and 2023. The results show that after the application of reinforcement learning network and frequency conversion technology, the monthly energy consumption is reduced by an average of 20% to 30% compared with when it is not applied, especially in the peak energy consumption periods of summer and winter. This result not only verifies the energy-saving potential of reinforcement learning network and frequency conversion technology in practical applications, but also demonstrates its wide application prospects in high-energy-consuming industries such as data centers.

[0161] In the future, with the further development and optimization of technology, Hadoop-AI intelligent frequency conversion technology will play an important role in a wider range of fields, especially in promoting energy transformation and improving energy efficiency in high-energy-consuming industries, which has great theoretical significance and practical value. The technical solution of the present invention not only provides a new way to achieve sustainable economic development and protect the ecological environment, but also provides an important reference for energy efficiency improvement and green transformation in other high-energy-consuming industries, which will help promote the progress of global energy management and energy conservation and emission reduction.

[0162] At the same time, the present invention also realizes an integrated energy efficiency improvement system, including a data acquisition module, a preprocessing module, a sample data synthesis module, an environmental model construction module, a reinforcement learning network training module and an equipment state adjustment module;

[0163] A data acquisition module, used to acquire energy consumption data of high energy consumption systems;

[0164] The preprocessing module preprocesses the acquired energy consumption data based on Hadoop's multi-dimensional measurement data parallel processing technology to form energy consumption data samples;

[0165] The sample data synthesis module generates synthetic data samples based on the Transformer's deep generation network and the constructed energy consumption data samples;

[0166] The environmental model building module builds the environmental model of the high-energy consumption system based on the generated synthetic data samples;

[0167] The reinforcement learning network training module builds a high-energy consumption system environment model to train the reinforcement learning network to guide the working state adjustment of the frequency conversion equipment, including:

[0168] In the high energy consumption system environment model, the reinforcement learning Q-Learning algorithm is used to train the Q function model;

[0169] According to the Q function, update the Q table;

[0170] Update the Q value in the Q table by introducing a loss function combined with a regularization term;

[0171] The formula of the loss function L is as follows:

[0172]

[0173] Among them, Q(s, a) represents the value of action a in the current state s of the system, λ represents the regularization coefficient, and Q target (s, a) is the target Q value;

[0174] The calculation formula of Q value is as follows:

[0175] Q target (s, a) = r + γ·max a′ Q(s', a')

[0176] Among them, Q target (s, a) represents the target value of action a in the current state s, r represents the immediate reward, γ represents the discount factor, max a′ Q(s', a') represents the maximum value of all possible actions of the system in the next state s';

[0177] Update the Q value by gradient descent;

[0178] Updating the Q value by gradient descent includes updating the Q value according to the following formula:

[0179]

[0180] Among them, α represents the learning rate, Q(s, a) represents the value of action a in the current state s of the system, and L represents the loss function.

[0181] The frequency conversion device obtains all possible actions from the Q table according to the current system state and selects the action with the largest Q value as the current optimal action.

[0182] The equipment status adjustment module is used to deploy the trained reinforcement learning network to the actual high-energy consumption system, monitor the status of the high-energy consumption system in real time and adjust the working status of the frequency conversion equipment.

[0183] The present disclosure may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0184] A computer-readable storage medium may be a tangible device that can hold and store instructions used by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples of computer-readable storage media (a non-exhaustive list) include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure in a groove on which instructions are stored, and any suitable combination of the foregoing. As used herein, a computer-readable storage medium is not to be interpreted as a transient signal per se, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse through a fiber optic cable), or an electrical signal transmitted through a wire.

[0185] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device.

[0186] The computer program instructions for performing the operation of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages, such as Smalltalk, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. Computer-readable program instructions may be executed completely on a user's computer, partially on a user's computer, as an independent software package, partially on a user's computer, partially on a remote computer, or completely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect via the Internet). In some embodiments, an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), may be customized by utilizing the state information of the computer-readable program instructions, and the electronic circuit may execute the computer-readable program instructions, thereby realizing various aspects of the present disclosure.

[0187] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An integrated energy efficiency improvement method, characterized in that: include: Obtain energy consumption data for high energy consumption systems; Preprocessing the energy consumption data based on Hadoop's multi-dimensional measurement data parallel processing technology to form energy consumption data samples; Generate synthetic data samples based on Transformer's deep generative network and constructed energy consumption data samples; Based on synthetic data samples, build an environmental model for high-energy consumption systems; Based on the high-energy consumption system environment model, a reinforcement learning network is trained to guide the working state adjustment of the frequency conversion equipment; the trained reinforcement learning network is deployed to the actual high-energy consumption system to monitor the high-energy consumption system state in real time and adjust the working state of the frequency conversion equipment.

2. The method for improving energy efficiency according to claim 1, characterized in that: The energy consumption data of the high energy consumption unit includes at least one of a timestamp, a device ID, power consumption, peak power and valley power, a power factor, and voltage data.

3. The method for improving energy efficiency according to claim 1, characterized in that: The preprocessing of the acquired energy consumption data includes: Use MapReduce framework to process energy consumption data in parallel; Distributed storage of energy consumption data after parallel processing in the form of data blocks; Distribute large-scale energy consumption data using HDFS.

4. The method for improving energy efficiency in an integrated manner according to claim 1, characterized in that: The establishment of the high energy consumption system environmental model includes: Establish the state space of high energy consumption systems; Establishing the action space of energy-intensive systems; and reward functions for designing environmental models for energy-intensive systems.

5. The method for improving energy efficiency in an integrated manner according to claim 4, characterized in that: The establishment of the state space of the high energy consumption system includes: Min-Max is used to normalize the state space data, and the normalized data is integrated into a state vector to represent the state of the system at a certain moment. The state vector is expressed as: S t =[E t ,E t-1 ,T t ,N t ,f t ] Among them, the state space data includes current energy consumption data Et, historical energy consumption data Et -1 , temperature Tt, current speed of the water pump Nt, and system frequency ft.

6. The method for improving energy efficiency in an integrated manner according to claim 4, characterized in that: The establishment of the action space of the high-energy consumption system includes constructing the action-value function according to the following formula: Q(s,a)←Q(s,a)+α[r+γ·maxaQ(s′,a')-Q(s,a)] Among them, Q(s, a) represents the value of action a in the current state s of the system; α represents the learning rate; r represents the immediate reward; γ represents the discount factor; max a 'Q(s', a') represents the maximum value of all possible actions of the system in the next state s'.

7. The method for improving energy efficiency in an integrated manner according to claim 4, characterized in that: The design of the reward function Rt of the high energy consumption system environment model is shown in the following formula: Among them, Et -1 and Et represent the system energy consumption at the previous moment and the system energy consumption at the current moment respectively; Nt -1 and Nt represent the pump speed at the previous moment and the pump speed at the current moment respectively; Tt and T represent the current ambient temperature and the target temperature respectively; α represents the reward weight for reducing system energy consumption, β represents the penalty weight for changing the pump speed, and γ represents the reward weight for temperature control.

8. The method for improving energy efficiency in an integrated manner according to claim 1, characterized in that: In step 5, a reinforcement learning network is trained based on the high energy consumption system environment model constructed in step 4 to guide the working state adjustment of the frequency conversion equipment, including: Step 5.1: In the high energy consumption system environment model, use the reinforcement learning Q-Learning algorithm to train the Q function model; Step 5.2: Update the Q table according to the Q function; Step 5.3: Update the Q value in the Q table by introducing a loss function combined with a regularization term; Step 5.4: Update the Q value by gradient descent method; Step 5.5: The frequency conversion device obtains all possible actions from the Q table according to the current system state, and selects the action with the largest Q value as the current optimal action.

9. The method for improving energy efficiency in an integrated manner according to claim 8, characterized in that: In step 5.3, the formula of the loss function L is as follows: Among them, Q(s, a) represents the value of action a in the current state s of the system, λ represents the regularization coefficient, and Qtarget(s, a) is the target Q value.

10. The method for improving energy efficiency in an integrated manner according to claim 8, characterized in that: In step 5.3, the calculation formula of Q value is as follows: Qtarget(s,a)=r+γ·maxa′Q(s′,a) Among them, Qtarget(s, a) represents the target value of action a in the current state s, r represents the immediate reward, γ represents the discount factor, max a 'Q(s', a') represents the maximum value of all possible actions of the system in the next state s'.

11. An integrated energy efficiency improvement method according to claim 8, 9 or 10, characterized in that: In step 5.4, updating the Q value by gradient descent includes updating the Q value according to the following formula: Among them, α represents the learning rate, Q(s, a) represents the value of action a in the current state s of the system, and L represents the loss function.

12. An integrated energy efficiency improvement system, characterized in that: Said A data acquisition module, used to acquire energy consumption data of high energy consumption systems; The preprocessing module preprocesses the acquired energy consumption data based on Hadoop's multi-dimensional measurement data parallel processing technology to form energy consumption data samples; The sample data synthesis module generates synthetic data samples based on the Transformer's deep generation network and the constructed energy consumption data samples; The environmental model building module builds the environmental model of the high-energy consumption system based on the generated synthetic data samples; The reinforcement learning network training module trains the reinforcement learning network based on the constructed high-energy consumption system environment model to guide the working state adjustment of the frequency conversion equipment; The equipment status adjustment module is used to deploy the trained reinforcement learning network to the actual high-energy consumption system, monitor the status of the high-energy consumption system in real time and adjust the working status of the frequency conversion equipment.

13. The energy efficiency integrated improvement system according to claim 12, characterized in that: Based on the constructed high-energy consumption system environment model, the reinforcement learning network is trained to guide the working state adjustment of the frequency conversion equipment, including: In the high energy consumption system environment model, the reinforcement learning Q-Learning algorithm is used to train the Q function model; According to the Q function, update the Q table; Update the Q value in the Q table by introducing a loss function combined with a regularization term; Update the Q value by gradient descent; The frequency conversion device obtains all possible actions from the Q table according to the current system state and selects the action with the largest Q value as the current optimal action.

14. The energy efficiency integrated improvement system according to claim 13, characterized in that: The formula of the loss function L is as follows: Among them, Q(s, a) represents the value of action a in the current state s of the system, λ represents the regularization coefficient, and Qtarget(s, a) is the target Q value.

15. The energy efficiency integrated improvement system according to claim 13, characterized in that: The calculation formula of Q value is as follows: Qtarget(s,a)=r+γ·maxaQ(s′,a) Among them, Qtarget(s, a) represents the target value of action a in the current state s, r represents the immediate reward, γ represents the discount factor, max a 'Q(s', a') represents the maximum value of all possible actions of the system in the next state s'.

16. The energy efficiency integrated improvement system according to claim 13, characterized in that: Updating the Q value by gradient descent includes updating the Q value according to the following formula: Among them, α represents the learning rate, Q(s, a) represents the value of action a in the current state s of the system, and L represents the loss function.

17. A terminal comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method according to any one of claims 1-11.

18. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 11 are implemented.

Images