Energy consumption prediction and energy-saving control method adopting deep neural network

Abstract

The invention is suitable for the technical field of energy consumption control, and provides an energy consumption prediction and energy-saving control method adopting a deep neural network. According to the method, virtual training data is generated by constructing a machine room digital twinborn model including a three-dimensional geometric model and a CFD simulation model, and the electric energy use efficiency and the overheating condition are predicted in combination with a deep neural network optimized by an attention mechanism. And the operation strategy system generates an optimal control suggestion by predicting the energy-saving effect and the safety risk after the air conditioner is turned off or the temperature is increased. The visual interface integrates real-time parameters, historical data and three-dimensional airflow organization dynamic display, and assists an administrator in decision making. The method breaks through the limitation of real data, improves the prediction precision, reduces the PUE on the premise of ensuring that the operation number of air conditioners is greater than or equal to 3, and achieves the efficient and energy-saving management of a machine room. Experiments show that the PUE can be reduced by about 4.3%, and the cabinet overheating probability can be reduced by 60%.

Description

Technical Field

[0001] The present invention belongs to the technical field of energy consumption control, and in particular relates to an energy consumption prediction and energy-saving control method using a deep neural network. Background Art

[0002] As data centers expand, energy consumption becomes increasingly prominent. Traditional methods for controlling data center energy consumption rely on precise sensor data and complex control algorithms, making them incapable of handling the nonlinear and dynamically changing environment. For example, traditional building temperature control systems experience significant declines in prediction accuracy and reliability when exposed to complex weather conditions or fluctuating equipment loads, leading to high power usage effectiveness (PUE).

[0003] In recent years, neural network-based control methods have been introduced to data center energy conservation, but they rely on large amounts of training data. However, for safety reasons, real data centers are reluctant to provide abnormal data beyond safe limits, resulting in insufficient model generalization and prone to overfitting. Furthermore, existing technologies lack sufficient integration of visual analysis of airflow distribution in computer rooms, making it difficult for administrators to intuitively understand equipment operating status, limiting decision-making efficiency.

[0004] Because data center energy control systems are complex, nonlinear systems, traditional methods struggle to achieve efficient control. However, neural network-based control methods, due to their high degree of automation, adaptability, and scalability, have been widely adopted in data center energy-saving applications. Optimization systems incorporating neural networks can autonomously learn from complex environments and make optimal energy-saving decisions. However, training artificial intelligence requires a large amount of sample data. However, in real data centers, data samples beyond safe limits are often not available for security reasons. Summary of the Invention

[0005] The purpose of the present invention is to provide an energy consumption prediction and energy-saving control method using a deep neural network, aiming to solve the technical problems existing in the prior art identified in the background technology.

[0006] The present invention is implemented as follows: a method for energy consumption prediction and energy-saving control using a deep neural network, the method comprising:

[0007] Establish a digital twin model of the target computer room, including a 3D geometric model and a computational fluid dynamics simulation model. Generate a simulation data set containing power efficiency, overheating probability, and equipment operating parameters through computational fluid dynamics simulation.

[0008] The simulation data set is preprocessed, abnormal samples are removed, and the data is normalized and divided into a training set, a validation set, and a test set; the training data includes power efficiency, cabinet temperature, and air conditioning operating parameters under different working conditions;

[0009] Construct a deep neural network including an attention layer, train it using the training set and the validation set, and output a prediction model for power usage efficiency and overheating probability.

[0010] Based on the current operating parameters of the computer room, use the deep neural network to predict the power usage efficiency and overheating probability after adjusting the air conditioner operating parameters, and generate energy-saving operation suggestions based on the prediction results.

[0011] Real-time display the operating status, prediction results, and operation suggestions of the computer room through a visualization interface, and provide interactive control.

[0012] As a further aspect of the present invention, the computational fluid dynamics simulation includes dynamic simulations of the air flow distribution, temperature field, and heat load distribution in the computer room, and the simulation data set covers normal and abnormal operating conditions.

[0013] As a further aspect of the present invention, the deep neural network is a five-layer fully connected network, each layer contains 50 nodes, and the activation function uses ReLU.

[0014] As a further aspect of the present invention, the training using the training set and the validation set adopts the momentum stochastic gradient descent optimization algorithm, sets the maximum number of iterations to 3000 times, allows the number of validation failures to be 8 times, and the learning rate to be 0.01.

[0015] As a further aspect of the present invention, the visualization interface is constructed based on the Unity3D platform.

[0016] As a further aspect of the present invention, the logic for generating operation suggestions is: preferentially select the solution that reduces the power usage efficiency the most and does not cause overheating among turning off the air conditioner or increasing the temperature set point, and the number of operating air conditioners is not less than 3.

[0017] The beneficial effects of the present invention are:

[0018] The present invention uses digital twin technology to realize the creation of a real-time and dynamic virtual model of a physical entity in a digital way. It uses information such as physical models, sensor data, and historical operation data to achieve an image mapping of the working process of the physical prototype in the virtual space. It can not only comprehensively and completely map the geometric features, physical behaviors, and operating rules of the physical entity, but also improve and optimize the physical entity through simulation data analysis. Using digital twin to establish a simulation model of the data center can provide a wider range and more training data for the neural network to obtain a more accurate and stable model. Description of the Drawings

[0019] Figure 1 It is a system development flow chart;

[0020] Figure 2 It is a flowchart for data preprocessing and DNN training;

[0021] Figure 3 It is a flowchart for generating logical judgments for operation suggestions;

[0022] Figure 4 It is a schematic diagram of the Unity3D visualization interface. Specific implementation manners

[0023] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0024] The present invention relates to a method for predicting and controlling energy consumption in a computer room space using a deep neural network. By integrating the best operation suggestions and displaying them on the platform to assist the administrator in making decisions, the purpose of reducing the PUE (Power Usage Effectiveness) and improving the energy utilization rate is achieved. First, a model of a real computer room is established and CFD (Computational Fluid Dynamics) simulation is performed, and the simulation data is used for training a constructed DNN (Deep Neural Network) to predict the PUE and overheating conditions. In the judgment logic of the operation suggestion system, first, the PUE and overheating conditions after turning off one air conditioner and raising the air conditioner temperature set point by one degree are predicted, and the one with the most reduction in PUE and no overheating is selected as the operation control strategy. Finally, the control strategy and prediction data are displayed in the interface constructed in the Unity3D environment to assist the administrator in more efficient management of the system.

[0025] Embodiment:

[0026] A small data center computer room is adopted, which contains four rows of cabinets and four precision air conditioners. The development process of the system is as Figure 1 shown. First, a three-dimensional model 001 of the computer room and a CFD simulation model 002 are established. The simulation data is used to train a DNN deep neural network model 003 to predict the PUE and overheating conditions of the computer room space, and the generated operation suggestions are integrated into Unity3D 001 for display.

[0027] A 5-layer deep neural network is adopted, with 50 nodes in each layer, and the ReLU (Rectified Linear Unit) function is used as the activation function. As Figure 2 shown, the data preprocessing and network training processes include:

[0028] As Figure 2As shown in Figs. 101 - 104, training data is obtained through CFD simulation and then training begins. In terms of data preprocessing, first, samples with complete failure of the computer room in the dataset are removed. Then, the dataset is divided into a training set, a validation set, and a test set according to the ratio of 70%, 15%, and 15%. The data is normalized to the interval [0, 1] using the MinMax (minimum - maximum normalization) method.

[0029] In the network training part, as Figure 2 shown in Figs. 105 - 109, first, the weight parameters in the network are randomly initialized, and the iteration step s and the validation failure count v are initialized to 0. Then, the training set and the validation set are randomly shuffled. On the training set, an attention layer is introduced, and the output of the network, that is, the PUE and the overheating situation, is calculated through forward propagation. Then, according to the calculation result of the loss function, backpropagation is performed to update the network parameters.

[0030] Subsequently, as, as Figure 2 shown in Figs. 113 - 118, the output of the network, that is, the PUE and the overheating situation, is calculated on the validation set, and the loss is calculated. It is judged whether the loss on the validation set is greater than or equal to the minimum loss before the end of training. If so, the validation failure count v is incremented by 1, and then it is judged whether the validation failure count v is greater than the allowable failure count P. If so, the training ends.

[0031] As Figure 2 shown in Figs. 110 - 118, if the validation set loss is less than the minimum loss, the iteration step s is incremented by 1. At this time, the error is calculated. Here, RMSE (root - mean - square error) is selected, and it is judged whether the error is less than the training target and whether the iteration step s is greater than the maximum iteration number E. If so, the training ends; if not, the next round of training begins.

[0032] In an example according to the present invention, the training of the network adopts the SGDM (Stochastic Gradient Descent with Momentum) method. The maximum iteration number E is set to 3000; the maximum allowable failure number is set to 8; the learning rate is set to 0.01; the learning rate decay factor is set to 0.8; the learning rate decay period is set to 100; the validation frequency is set to 50.

[0033] After completing the training of the deep neural network, a set of optimal operation strategy recommendation systems is established. The air - conditioner temperature setpoint and the number of operating air - conditioners are selected as adjustable variables. The judgment process for making recommendations is based on the following criteria:

[0034] (1) The PUE should be as small as possible after adjustment according to the recommendation;

[0035] (2) There should be no overheating of the cabinets after adjustment according to the recommendation;

[0036] (3) For safety reasons, the number of operating air - conditioners should not be less than three.

[0037] For example:

[0038] Air conditioner shutdown judgment:

[0039] If the PUE decreases and the overheating probability < 5%, then select to shut down one air conditioner.

[0040] Temperature increase judgment:

[0041] If it is not feasible to shut down the air conditioner (e.g., the number of units = 3), or overheating occurs after shutdown, then judge whether the PUE decreases and the overheating probability < 5% after the temperature is increased.

[0042] Priority: The temperature increase strategy takes precedence over shutting down the air conditioner (if both meet the conditions, select the strategy with a greater decrease in PUE).

[0043] The input of the system is the current allowable operating conditions of the computer room and the PUE. The PUE and overheating probability after shutting down one air conditioner or increasing the air conditioner temperature set point by one degree are predicted through the DNN network trained above. The output is the operation suggestion, including three types: "Shut down one air conditioner", "Increase the air conditioner temperature set point by one degree", and "The current is the best operating state".

[0044] Such as Figure 3 shown in 201 - 211, its judgment logic is as follows:

[0045] First, input the current operating parameters and PUE, and predict the PUE and overheating conditions after shutting down one air conditioner and increasing the air conditioner temperature set point by one degree. Judge whether the PUE after shutting down the air conditioner is less than the current PUE, and whether shutting down the air conditioner causes overheating. When judging whether shutting down the air conditioner causes overheating, since the safety criterion limits the number of air conditioners in operation to be more than 3, it is also necessary to judge whether the current number of operating air conditioners is greater than 3. If shutting down the air conditioner cannot save energy or causes overheating after shutdown, then judge the relationship between the PUE after increasing the air conditioner temperature set point and the current PUE, and whether it overheats. Select the final output as "Maintain the current state" or "Increase the air conditioner temperature by one degree" according to the judgment result. If shutting down the air conditioner can save energy and does not cause overheating, then judge the relationship between the PUE after increasing the air conditioner temperature set point and the PUE after shutting down the air conditioner, and whether it overheats. Select the final output as "Increase the air conditioner temperature by one degree" or "Shut down one air conditioner" according to the judgment result.

[0046] Such as Figure 4As shown, the interactive interface in the present invention consists of three parts: display of computer room operating parameters, generation of operating suggestions, and air conditioning control. In terms of the display of computer room operating parameters, cabinet temperature and air conditioner sensible cooling power are displayed on the device in the form of color. Other parameters, such as the values of cabinet temperature and air conditioner sensible cooling power, PUE and indoor and outdoor temperatures, historical temperatures, etc. are displayed on three panels. Operation suggestions can be obtained by clicking on the panel. Air conditioning control can be achieved through the slider bar on the control panel. At the same time, a lens movement function is designed to enable administrators to better observe the airflow organization of the computer room space from different angles. Specific embodiment:

[0048] 1. Build a 3D model of the computer room based on the actual layout (including 4 rows of cabinets and 4 precision air conditioners), and restore the geometric shape and spatial distribution of equipment such as cabinets, air conditioners, wiring racks, and floor air vents.

[0049] Model parameters include:

[0050] Cabinet dimensions (height × width × depth): 2m × 0.6m × 1.2m;

[0051] Air supply volume: 5000m 3 / h;

[0052] Floor static pressure box height: 0.3m.

[0053] 2. Perform fluid mechanics simulation, with boundary conditions including:

[0054] Cabinet heat load: Single cabinet power 3kW~5kW (obeying normal distribution);

[0055] Air supply temperature: 18℃~25℃;

[0056] Ambient temperature: 15℃~35℃ (simulating working conditions in different seasons).

[0057] Simulation output data includes:

[0058] PUE value: calculates the ratio of total energy consumption to IT equipment energy consumption;

[0059] Cabinet overheating probability: the percentage of cabinets with temperatures exceeding 35°C;

[0060] Air conditioning operating parameters: start / stop status, temperature set point, sensible cooling power;

[0061] Airflow organization data: velocity field, temperature field, and pressure field distribution.

[0062] The simulation covers normal and abnormal operating conditions (such as air conditioner failure and cabinet overload), generating a simulation data set containing 10,000 sets of samples.

[0063] 3. Excluding abnormal samples with complete failure of the computer room (such as all air conditioners shutting down), there are 9,500 groups of valid data left;

[0064] Divide the training set, validation set, and test set according to 70% (6,650 groups), 15% (1,425 groups), and 15% (1,425 groups);

[0065] Perform min-max normalization on the input features (such as temperature and air conditioner power).

[0066] 4. Establish a deep neural network model:

[0067] Network structure:

[0068] Input layer: Receive 12-dimensional features (number of air conditioners, temperature set point, cabinet temperature, etc.);

[0069] Attention layer: Enhance the contribution of key features (such as air conditioner supply air temperature) through a weight matrix;

[0070] Hidden layer: 5 fully connected layers, with 50 nodes in each layer, and the activation function is ReLU;

[0071] Output layer: 2 nodes, predicting PUE and overheating probability respectively.

[0072] Training parameters:

[0073] Optimizer: SGDM (Stochastic Gradient Descent with Momentum), momentum factor 0.9;

[0074] Learning rate: Initial value 0.01, decreasing by 0.8 times every 100 iterations;

[0075] Loss function: Root Mean Square Error (RMSE);

[0076] Maximum number of iterations: 3000;

[0077] Early stopping mechanism: If the validation set loss has not improved for 8 consecutive times, terminate the training.

[0078] Training process:

[0079] Calculate the predicted value through forward propagation and update the weights through backward propagation;

[0080] Test the model performance on the validation set every 50 iterations;

[0081] The final RMSE of the test set is 0.023, and the prediction accuracy of the overheating probability is 96.5%.

[0082] 5. Generate policy suggestions

[0083] Input the current operating parameters: number of air conditioners (4 units), temperature set point (20°C), cabinet temperature (28°C - 32°C), PUE (1.6);

[0084] Use the trained DNN to predict the PUE and overheating probability of the following two adjustment plans:

[0085] Plan A: Turn off 1 air conditioner (3 units remaining);

[0086] Plan B: Increase the air conditioner temperature set point by 1°C (to 21°C).

[0087] Priority 1: If the PUE of Plan A < the current PUE, and the overheating probability < 5%, and the number of air conditioners ≥ 3, then it is recommended to turn off the air conditioner;

[0088] Priority 2: If the PUE of Plan B < the current PUE, and the overheating probability < 5%, then it is recommended to increase the temperature set point;

[0089] Priority 3: If neither is satisfied, then maintain the current state.

[0090] Example results:

[0091] Predicted PUE of Plan A = 1.55, overheating probability = 3%;

[0092] Predicted PUE of Plan B = 1.58, overheating probability = 1%;

[0093] The final recommendation is to choose Plan A (turn off 1 air conditioner).

[0094] 6. Effect verification

[0095] PUE reduction effect: In the continuous 30-day test, the average PUE decreased from 1.62 to 1.52, saving about 6.2% in energy;

[0096] Overheating warning: The system successfully predicted 3 potential overheating events and adjusted the air conditioner strategy in advance to avoid failures;

[0097] Administrator feedback: The visualization interface significantly improved the efficiency of controlling the computer room status, and the adoption rate of operation suggestions exceeded 90%.

[0098] It should be understood that although the steps in the flowcharts of the embodiments of the present invention are shown in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear description in this article, there is no strict order limit for the execution of these steps, and these steps can be executed in other orders. Moreover, at least a part of the steps in each embodiment may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same moment, but can be executed at different moments. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed alternately or in turn with at least a part of other steps or sub-steps or stages of other steps.

[0099] Those of ordinary skill in the art can understand that all or part of the processes of implementing the methods in the above embodiments can be completed by instructing relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the embodiments provided in the present application can include non-volatile and / or volatile memories. Non-volatile memories can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memories can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0100] The technical features of the above-described embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered to be within the scope described in this specification.

[0101] The above-described embodiments merely represent several implementation manners of the present invention. The description thereof is relatively specific and detailed, but it should not be construed as a limitation to the scope of the patent of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several variations and improvements can still be made, and these all fall within the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the appended claims.

[0102] The above is only the preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for energy consumption prediction and energy-saving control using a deep neural network, characterized in that, The method includes: Establishing a digital twin model of the target computer room, including establishing a three-dimensional geometric model of the computer room and a computational fluid dynamics simulation model, and generating a simulation data set containing power usage efficiency, overheating probability, and equipment operation parameters through computational fluid dynamics simulation; Preprocessing the simulation data set, removing abnormal samples, normalizing, and then dividing it into a training set, a validation set, and a test set; the training data includes power usage efficiency, cabinet temperature, and air conditioner operation parameters under different working conditions; Constructing a deep neural network containing an attention layer, training it using the training set and the validation set, and outputting a prediction model for power usage efficiency and overheating probability; Based on the current computer room operation parameters, using the deep neural network to predict the power usage efficiency and overheating probability after adjusting the air conditioner operation parameters, and generating energy-saving operation suggestions based on the prediction results; Real-time displaying the computer room operation status, prediction results, and operation suggestions through a visualization interface, and providing interactive control; 2. The method according to claim 1, characterized in that, The computational fluid dynamics simulation includes dynamic simulation of the air flow distribution, temperature field, and heat load distribution in the computer room, and the simulation data set covers normal and abnormal operation conditions; 3. The method according to claim 1, wherein The deep neural network is a five-layer fully connected network, with 50 nodes in each layer, and the activation function uses ReLU; 4. The method according to claim 1, wherein The training using the training set and the validation set adopts the momentum stochastic gradient descent optimization algorithm, sets the maximum number of iterations to 3000 times, allows the number of validation failures to be 8 times, and the learning rate to be 0.01; 5. The method according to claim 1, characterized in that, The visualization interface is constructed based on the Unity3D platform; 6. The method according to claim 1, characterized in that, The logic for generating operation suggestions is: preferentially select the scheme that reduces the power usage efficiency the most and does not cause overheating among turning off the air conditioner and increasing the temperature set point, and the number of operating air conditioners is not less than 3.

Images