Intelligent energy-saving control method based on deep learning

Abstract

The invention discloses an intelligent energy-saving control method based on deep learning, and the method comprises the steps: obtaining the load data and environment data of a machine room, building a PUE prediction model, training the model through the obtained load data and environment data, deploying the trained PUE prediction model to a target device, and carrying out the prediction of the machine room. And obtaining an optimal parameter combination through the PUE prediction model in combination with the genetic algorithm model, and implementing the optimal parameter combination. Meanwhile, according to the energy-saving control method, a service guarantee model can be added, and the performability of the genetic algorithm model for generating the parameter combination in the process of iterating the optimal parameter combination is guaranteed through the service guarantee model. Through verification, the intelligent energy-saving control technology based on deep learning provided by the invention can greatly improve the energy use efficiency of the data center machine room.

Description

Technical Field

[0001] The present invention relates to an intelligent energy-saving control method, and particularly to an intelligent energy-saving control method based on deep learning. Background Art

[0002] With the rapid development of information technology, as the core facility for information storage and processing, the energy efficiency of data centers has attracted increasing attention. One of the main metrics for data center energy efficiency is the Power Usage Effectiveness (PUE), which is defined as the ratio of the total energy consumption of the data center to the energy consumption of IT equipment. Ideally, the closer the PUE value is to 1, the less energy is consumed by non-IT equipment (such as cooling and power distribution systems), and the higher the energy efficiency level of the data center.

[0003] However, in actual operation, the PUE value of a data center is often affected by various factors, including the environmental parameters of the computer room (such as temperature, humidity, air flow), the operating status of equipment (such as CPU usage rate, memory occupancy rate), and load changes, etc. The changes of these factors are complex and uncertain, making the energy efficiency management of data centers a challenge.

[0004] Traditional energy efficiency management methods mainly rely on manual experience and regular maintenance. This method is not only inefficient but also difficult to cope with rapidly changing loads and environmental conditions. With the development of big data and artificial intelligence technologies, data-driven energy efficiency optimization methods have gradually become a research hotspot. Among them, deep learning technology shows great potential in energy efficiency prediction and optimization due to its powerful data processing and prediction capabilities. Therefore, providing an intelligent energy-saving control system based on deep learning has become an urgent problem to be solved in the industry. Summary of the Invention

[0005] In view of the above deficiencies in the current energy efficiency control of data centers, the present invention provides an intelligent energy-saving control system based on deep learning, which can accurately control the temperature of the data center computer room to effectively reduce the PUE value and improve the energy usage efficiency of the data center computer room.

[0006] To achieve the above object, one aspect of the present invention provides an intelligent energy-saving control method based on deep learning, including: Obtain the load data and environmental data of the computer room; Use the obtained load data and environmental data to construct and train a PUE prediction model; Deploy the trained PUE prediction model to the target device; Obtain the optimal parameter combination through the PUE prediction model and in combination with a genetic algorithm model; Implement the optimal parameter combination.

[0007] According to one aspect of the present invention, the intelligent energy-saving control method based on deep learning further includes: performing feature engineering processing on the acquired load data and environmental data to enhance the representativeness of the basic features in the data.

[0008] According to one aspect of the present invention, the feature engineering processing of the load data and environmental data includes: Selecting data with a high correlation with PUE as basic features; Performing sliding window statistics on the basic features to calculate the maximum value, minimum value, mean value, and variance of the feature within the number of steps of historical data and using them as statistical features; Adding the statistical features as a part of the data features to the data set.

[0009] According to one aspect of the present invention, the obtaining of the optimal parameter combination through the genetic algorithm model and the PUE prediction model includes: Searching for sample points in the historical load data and environmental data that are similar to the current load data and environmental data; Randomly generating an initial parameter combination according to the value range of the adjustable parameters in the sample points and adding the initial parameter combination to the candidate set; Calculating the PUE value of each parameter combination in the candidate set through the PUE prediction model; Selecting parameter combinations from the candidate set according to the principle of giving priority to smaller PUE values and adopting the roulette wheel strategy to perform crossover operation and mutation operation to generate offspring parameter combinations; Calculating the PUE value of the offspring parameter combination through the PUE prediction model. When the PUE value is less than the maximum PUE value of the current candidate set, adding the offspring parameter combination to the candidate set and deleting the parameter combination with the largest PUE value in the candidate set; Continuously evolving and iterating until a parameter combination with a PUE value less than the actual PUE value of the current computer room or the number of iterations reaches the set value is found.

[0010] According to one aspect of the present invention, the intelligent energy-saving control method based on deep learning further includes a service guarantee model. Before adding the initial parameter combination and the offspring parameter combination to the candidate set, they both need to be verified by the service guarantee model. If the verification fails, they will not be added to the candidate set.

[0011] According to one aspect of the present invention, the process of the service guarantee model verifying the parameter combination includes; Setting the alarm temperature of the computer room; Inputting the parameter combination into the service guarantee model to predict the temperature of the computer room; If the predicted temperature of the computer room is lower than the alarm temperature of the computer room, the verification passes; if the predicted temperature of the computer room exceeds the alarm temperature of the computer room, the verification fails.

[0012] According to one aspect of the present invention, the process of the crossover operation includes randomly selecting intersection points and exchanging a part of the adjustable variables of two parent vectors to generate offspring.

[0013] According to one aspect of the present invention, the process of the mutation operation includes randomly selecting an adjustable variable in the offspring and generating a random value that satisfies the value range to replace the current value of the adjustable variable to generate a new offspring.

[0014] According to one aspect of the present invention, the threshold of the mutation operation is set to 0.05.

[0015] According to one aspect of the present invention, the implementation of the optimal parameter combination includes: adopting an automatic or manual method to control the computer room equipment to perform adjustment operations. If the adjustment operation is to put the refrigeration equipment into sleep, the number of refrigeration equipment put into sleep each time does not exceed 1; if the adjustment operation is to wake up the refrigeration equipment, there is no upper limit on the number of refrigeration equipment woken up each time.

[0016] Advantages of the implementation of the present invention: First, collect the load data and environmental data of the computer room and perform feature engineering processing on them. Then, construct a PUE prediction model and train the PUE prediction model with the data after feature engineering processing. The trained PUE prediction model can predict the PUE value of the computer room according to the parameter combination. Through the genetic algorithm model and combined with the service guarantee model, the optimal parameter combination of the data center computer room can be iteratively obtained and the parameter combination can be made implementable. Finally, implement parameter adjustment according to the optimal parameter combination. After verification, the intelligent energy-saving control method based on deep learning proposed in this application can greatly improve the energy use efficiency of the data center computer room. Brief Description of the Drawings

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0018] Figure 1 It is a schematic flowchart of the intelligent energy-saving control method based on deep learning according to the present invention; Figure 2 It is a schematic diagram of the modeling process of the deep learning PUE prediction module according to the present invention; Figure 3 It is a schematic diagram of the process of screening the optimal parameter combination by the genetic algorithm model according to the present invention; Figure 4 Schematic diagram of the process of generating the initial parameter combination candidate set for the genetic algorithm model described in the present invention; Figure 5 Schematic diagram of the process of performing crossover operation and mutation operation on the genetic algorithm model described in the present invention; Figure 6 Schematic diagram of the feature engineering processing flow described in the present invention; Figure 7 Schematic diagram of the modeling process of the deep learning PUE prediction module described in the present invention after adding feature engineering processing; Figure 8 Schematic diagram of the process of verifying the parameter combination of the service guarantee model described in the present invention; Figure 9 Schematic diagram of the process of screening the optimal parameter combination by combining the genetic algorithm model and the service guarantee model described in the present invention; Figure 10 Schematic diagram of the process of generating the initial parameter combination candidate set by combining the genetic algorithm model and the service guarantee model described in the present invention. Detailed implementation manners

[0019] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0020] Before further elaborating on the present invention, the nouns and terms involved in the embodiments of the present invention are described. The nouns and terms involved in the embodiments of the present invention are applicable to the following explanations: PUE (Power Usage Effectiveness) is the power utilization efficiency, which is an index for evaluating the energy efficiency of a data center. PUE is the ratio of all the energy consumed by the data center to the energy used by the IT load. The formula is: PUE = total energy consumption of data center equipment / energy consumption of IT equipment. It can be seen from the formula that the value range of PUE is from 1 to infinity. The closer the PUE value is to 1, the higher the energy utilization efficiency and the higher the degree of greening of the data center. On the contrary, the larger the PUE value, the higher the energy consumption outside the IT equipment in the data center and the lower the energy utilization efficiency.

[0021] LSTM (Long Short-Term Memory) is a special type of Recurrent Neural Network (RNN) architecture designed to address the problem of vanishing or exploding gradients that traditional RNNs often encounter when dealing with long sequence data. LSTM controls the flow of information by introducing self-recurrent connections and some carefully designed "gates" (input gate, forget gate, and output gate), enabling it to better capture long-term dependencies in time series.

[0022] MSE (Mean Squared Error), namely the mean squared error, is a commonly used metric to measure the degree of difference between the predicted values and the true values of a model.

[0023] The first embodiment of the present invention: Figure 1 Shows a schematic flowchart of an intelligent energy-saving control method based on deep learning in an embodiment of the present invention, including the following steps: Step S21: Obtain the load data and environmental data of the computer room.

[0024] Among them, the load data includes the CPU usage rate, memory usage rate, and energy consumption data of electronic devices in the data center computer room. The environmental data includes temperature, humidity, air flow, etc., and the collected data is saved as time series data according to the time series.

[0025] It can be understood that to ensure the timeliness of the data, the obtained data is the load data and environmental data of the data center computer room for the past 18 days, and this data is used as the training data, validation data, and test data of the PUE prediction model.

[0026] Step S22: Build a PUE prediction model and train the model with the obtained load data and environmental data.

[0027] In this embodiment, the PUE prediction model is constructed using an LSTM neural network to avoid the problems of vanishing and exploding gradients in the RNN neural network.

[0028] The LSTM neural network model includes memory units and a gating mechanism. The memory unit is the core part of the LSTM, used to store and transmit long-term information. The LSTM uses three main gating mechanisms to determine which information needs to be retained, updated, or output. Specifically, it includes three key gates: the input gate, the forget gate, and the output gate. These gates are similar to "filters", controlling the flow of information and determining which information should be retained, forgotten, or output, thereby precisely controlling and managing the information.

[0029] The following part is a brief description of the structure and algorithm design logic of the LSTM neural network model: Forget Gate: Determines which information to discard from the cell state. It generates a value between 0 and 1 through a sigmoid function, indicating the retention degree of each state value. The closer the value is to 1, the more important the information is and should be retained; the closer it is to 0, the more the information can be forgotten.

[0030] The role of the forget gate is to determine based on the current input and the previous hidden state which information to retain and which to discard. The calculation formula is: .

[0031] where is the output of the forget gate, is the sigmoid activation function. The sigmoid outputs values between [0, 1], so is a vector with elements in the range [0, 1]. This vector is multiplied by the information in the cell state before time step t. A value of 0 means the information at that position is completely discarded, 1 means the information is completely retained, and values between 0 and 1 mean the information is partially retained. is the weight output of the forget gate, is the bias term of the forget gate.

[0032] Input Gate: Consists of two parts. A sigmoid layer determines which values will be updated, and a tanh layer generates a new candidate value vector. The outputs of the sigmoid layer and the tanh layer of the input gate are multiplied to obtain the updated candidate values.

[0033] The input gate determines which information from the current time step's input should be input into the cell state . This part requires two calculations. The first calculation is for the importance calculation of the input information, and the formula is:

[0034] where is the importance vector of the input information. It also uses the sigmoid activation function to determine which of the current input information should be input into the cell state . The second calculation is for the input information, and the formula is , where is the candidate input information, containing the new information to be input into the unit state. Among them, and are the weight matrices respectively, and are the bias terms respectively.

[0035] Output gate: It determines the hidden state at the current moment, that is, the final output. Based on the current cell state and input, it calculates a vector between 0 and 1 through the sigmoid function, and then multiplies this vector by the cell state processed by the tanh function to obtain the final hidden state.

[0036] The output gate determines the content of the final output. Based on the current cell state, the output result is divided into two parts, and the calculation formulas are respectively: 、 。

[0037] Among them, is the output of the output gate, which determines which parts of the cell state are used as the output at this time step. is the final output at this time step, which is used to predict the PUE value in the future for a period of time and as part of the input for the next time step. is the weight matrix of the output gate, is the bias of the output gate.

[0038] Memory cell update: The new memory cell is calculated by combining the outputs of the forget gate and the input gate, that is, retaining a part of the cell state at the previous moment (determined by the forget gate) and adding a part of the current candidate cell state (determined by the input gate).

[0039] According to the outputs of the forget gate and the input gate, the memory cell is updated to , and the formula is: , where represents the memory cell at the current moment.

[0040] Error backpropagation: After passing through a fully connected layer, we get and the label Calculate the MSE loss and backpropagate the loss to the weight matrix to correct the matrix parameters. The formula is: 。

[0041] In this embodiment, the LSTM neural network model is not specifically limited. As shown in combination with Figure 2 , the training process of the PUE prediction model is as follows: Step S22a: Set the initialization parameters of the model. The parameters include the number of historical data steps (look_back), the number of steps to predict the future (T), the number of training epochs (epochs), the number of network layers (num_layer), the number of features (num_features), the learning rate (learn_rate), and the batch size (batch_size).

[0042] Step S22b: Data preprocessing, including data cleaning, denoising, outlier detection, and 0-1 normalization to accelerate the model training speed and eliminate the impact of data dimensions on model training.

[0043] Step S22c: Divide the dataset. Divide the model training data into a training set, a validation set, and a test set in chronological order.

[0044] In this embodiment, the collected load data and environmental data for the past 18 days are divided into a training set, a validation set, and a test set. To ensure the periodic characteristics of the data. Specifically, the data for the two weeks from the 18th day to the 4th day in history is used as the training set, the data from the 4th day to the 2nd day in history is used as the validation set, and the data from the 2nd day to the current day is used as the test set.

[0045] Step S22d: Model training. Use the training set data processed by feature engineering as features, and use the PUE value for a period of time in the future at the time point of the training data as the label to train the deep learning prediction model. In this embodiment, the PUE value 20 minutes in the future at the time point of the training data is used as the label.

[0046] During the model training process, use the validation set data to verify the model. If the validation set loss does not meet the expectation, adjust the relevant parameters and continue training. When the validation set loss meets the expectation, stop training and export the model.

[0047] The validation set loss is the loss calculated on the validation dataset. The validation dataset is not used to update the model parameters during the training process and is only used to evaluate the model performance. By comparing the training set loss and the validation set loss, it can be determined whether the model is overfitting or underfitting, and then the generalization ability of the model can be evaluated.

[0048] After the validation set loss meets the expectation, use the test set data to test the exported model. When the test set loss meets the expectation, stop the model training. If it does not meet the expectation, adjust the parameters and continue the model training until it meets the expectation.

[0049] The test set loss is the loss calculated on the test dataset. The test dataset is also a dataset independent of the training dataset and the validation dataset, and it is not used during the model development and validation phases. The main purpose of the test set loss is to provide an unbiased estimate of the final performance of the model to evaluate the performance of the model in actual applications.

[0050] During the model training process, the losses of the validation set and the test set are calculated using a loss function. For a neural network of the regression type, commonly used loss functions include Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE), etc. In this embodiment, Mean Squared Error (MSE) is used as the loss function for model training.

[0051] Step S23: Deploy the trained PUE prediction model to the target device.

[0052] Step S24: Deploy the genetic algorithm model to the target device.

[0053] The genetic algorithm model will continuously iterate through the genetic algorithm and combine with the PUE prediction model to finally output the optimal parameter combination of the computer room equipment. The following part will introduce the iteration process of the genetic algorithm model in detail.

[0054] Step S25: Obtain the optimal parameter combination through the genetic algorithm model and the PUE prediction model.

[0055] Figure 3 The specific process of screening the optimal parameter combination through the genetic algorithm model is shown as follows: Step S25a: Search for sample points in the historical data that are similar to the current load data and environmental data, and randomly generate an initial parameter combination candidate set according to the value range of the adjustable parameters in the sample points.

[0056] Among them, the adjustable parameters include the starting quantity of the refrigeration equipment, the outlet air temperature, the fan speed, etc.

[0057] Step S25b: As shown in Figure 4 Calculate the PUE value of each parameter combination in the candidate set through the PUE prediction model.

[0058] Step S25c: As shown in Figure 5 According to the principle of giving priority to the smaller PUE value and adopting the roulette wheel strategy, select parameter combinations from the candidate set for crossover operation and mutation operation to generate new offspring parameter combinations.

[0059] It should be noted that the crossover operation and the mutation operation are only limited to the adjustable variable part of the individual. The process of the crossover operation includes randomly selecting the intersection point and exchanging a part of the adjustable variables of the two parent vectors to generate offspring. The process of the mutation operation includes randomly selecting an adjustable variable in the offspring and generating a random value that satisfies the value range to replace the current value to generate a new offspring.

[0060] The calculation processes of the crossover operation and the mutation operation will be briefly introduced in the following part: Cross operation: Suppose there are parent 1 and parent 2. The total length of the individual vector is L, where the length of the adjustable variables is L1 and it is located in the first L1 positions of the vector, and the length of the non-adjustable variables is L2, which is located in the last L2 positions of the vector. It satisfies L1 + L2 = L. The algorithm randomly generates a value S in the range of [0, L1]. Taking S as the intersection point, the first S part of parent 1 is given to parent 2 to generate offspring 2, and the first S part of parent 2 is given to parent 1 to generate offspring 1.

[0061] Mutation operation: For each offspring, generate a random number in the range of [0, 1] and set a threshold. When the random number is less than the threshold, the offspring will mutate. Select one from the adjustable variables and generate a random value to replace the current value. The generation of the random value should satisfy the value range of the adjustable parameters.

[0062] In this embodiment, the threshold of the mutation operation is set to 0.05 to ensure a low probability of mutation and avoid excessive destruction of excellent genes during the iteration process.

[0063] Step S25d: Calculate the PUE value of the offspring parameter combination through the PUE prediction model. When the PUE value of the offspring parameter combination is less than the maximum PUE of the current candidate set, add it to the candidate set and delete the parameter combination with the maximum PUE value in the candidate set. This ensures that a more optimal parameter combination can be added to the candidate set and keeps the total number of parameter combinations in the candidate set stable.

[0064] Step S25e: Continuously evolve and iterate until the PUE value of the parameter combination is less than the actual PUE value of the current computer room or the number of iterations reaches the set value.

[0065] It can be understood that if no better offspring can be found after reaching the preset number of iterations, the data center computer room remains unchanged with the original settings.

[0066] Step S26: Implement the optimal parameter combination.

[0067] The optimal parameter combination generated through the iteration of the genetic algorithm model can be implemented in an automatic or manual way to execute the optimization result. Of course, it can also be manually executed after manual evaluation. Specifically, set the working parameters of the relevant equipment in the data center computer room according to the optimized parameters, such as the number of refrigeration equipment turned on, temperature, and fan speed.

[0068] It can be understood that to ensure that the environment of the data center computer room does not experience violent fluctuations during the implementation of the optimized parameters, when implementing the optimized parameters, if it is necessary to put the refrigeration equipment into sleep mode, the number of refrigeration equipment put into sleep mode each time does not exceed 1, so as to ensure that the temperature of the computer room environment can change smoothly. If it is necessary to wake up the refrigeration equipment, there is no upper limit on the number of refrigeration equipment woken up each time, so as to ensure that the temperature of the computer room environment can drop rapidly to avoid temperature alarms.

[0069] The second embodiment of the present invention: In order to improve the generalization ability of the PUE prediction model, in this embodiment, feature engineering processing is performed on the training data of the PUE prediction model. As can be seen from the above part, the training data here includes the collected load data and environmental data.

[0070] Figure 6 The specific process of performing feature engineering processing on the training data is shown, including: Step S31: Screen the data with high correlation with PUE as the basic features. In actual operation, a threshold can be set as the judgment criterion, and the data with a correlation with PUE higher than this threshold is used as the basic feature data.

[0071] Step S32: Perform sliding window statistical processing on the selected basic features, and calculate the maximum value, minimum value, mean value, and variance of the basic features within the number of historical data steps as statistical features.

[0072] Step S33: Add the statistical features as part of the data features to the data set.

[0073] Performing feature engineering processing on the training data of the PUE prediction model can enhance the representativeness of the basic features in the data, and the model trained with the data after feature engineering processing will have better generalization ability.

[0074] Figure 7 The training process of the PUE prediction model after adding feature processing is shown. In this embodiment, the input data of the LSTM model at each time step t includes environmental data and load data , and these data are all the data after feature engineering processing.

[0075] The third embodiment of the present invention: In order to ensure that the parameter combination iterated by the genetic algorithm model meets the implementation conditions, that is, implementing this parameter combination will not cause the temperature in the data center computer room to exceed the threshold and trigger a temperature alarm, in this embodiment, a service guarantee model is added on the basis of the first embodiment or the second embodiment to verify whether the parameter combination meets the requirements. The specific solution is as follows: Construct a service guarantee model and train this model. The training data can use the historical parameter data collected in the computer room. The parameter data includes the outlet air temperature of the refrigeration equipment, fan speed, air humidity, and computer room load, etc. The service guarantee model can be constructed using a CNN neural network or an LSTM neural network. The construction and training process of it is basically the same as that of the PUE prediction model introduced above, and will not be elaborated here.

[0076] Figure 8It shows the verification process of the service guarantee model for parameter combinations. As can be seen from the figure, its verification process includes: Step S41: Set the alarm temperature of the computer room. The alarm temperature here can be set according to the actual situation of the data center computer room.

[0077] Step S42: Input the parameter combination into the service guarantee model to predict the temperature of the computer room.

[0078] Step S43: Compare the predicted temperature with the alarm temperature set for the computer room. If the predicted temperature is lower than the alarm temperature, the verification passes; if the predicted temperature exceeds the alarm temperature of the computer room, the verification fails. In this embodiment, the alarm temperature of the computer room is set to 27°C.

[0079] Deploy the trained service guarantee model to the target device, and use the service guarantee model to verify whether the parameter combination meets the requirements during the iteration process of the genetic algorithm model. Figure 9 It shows the process of the genetic algorithm model combined with the service guarantee model to screen the optimal parameter combination. This process is basically the same as the process in the first embodiment, but the following two steps are different. Specifically: Adjust step S25b in the first embodiment to: Step S25b: Combine Figure 10 As shown, calculate the PUE value of each parameter combination in the candidate set through the PUE prediction model, and verify whether all the parameter combinations in the candidate set meet the requirements through the service guarantee model. If they do not meet the requirements, delete the parameter combination from the candidate set.

[0080] Adjust step S25d in the first embodiment to: Step S25d: Calculate the PUE value of the offspring through the PUE prediction model, and verify the offspring. When the PUE value of the offspring is less than the maximum PUE of the current candidate set population and meets the requirements of the service guarantee model, add it to the candidate set population, and delete the parameter combination with the largest PUE value in the candidate set, so as to ensure that a better parameter combination can be added to the candidate set and keep the total number of parameter combinations in the candidate set stable.

[0081] Advantages of the implementation of the present invention: The present application provides an intelligent energy-saving control method based on deep learning. First, the load data and environmental data of the computer room are collected and feature engineering is performed on them. Then, a PUE prediction model is constructed and the PUE prediction model is trained through the data after feature engineering. The trained PUE prediction model can predict the PUE value of the computer room according to the parameter combination. Through the genetic algorithm model and combined with the business guarantee model, the optimal parameter combination of the data center computer room can be iterated and the parameter combination can be implemented. Finally, the parameter adjustment is implemented with the optimal parameter combination. After verification, the intelligent energy-saving control method based on deep learning proposed in the present application can greatly improve the energy efficiency of the data center computer room. Therefore, the technical solution of the present invention effectively overcomes the various shortcomings of the prior art and has a high industrial utilization value.

[0082] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with the art within the technical scope disclosed in the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. An intelligent energy-saving control method based on deep learning, characterized in that: include: Obtain the load data and environmental data of the computer room; Use the acquired load data and environmental data to build and train a PUE prediction model; Deploy the trained PUE prediction model to the target device; The optimal parameter combination is obtained by combining the PUE prediction model with the genetic algorithm model; Implement the optimal parameter combination.

2. The intelligent energy-saving control method based on deep learning according to claim 1 is characterized in that: The intelligent energy-saving control method based on deep learning also includes: performing feature engineering processing on the acquired load data and environmental data to enhance the representativeness of basic features in the data.

3. The intelligent energy-saving control method based on deep learning according to claim 2 is characterized in that: The load data and environment data are processed by feature engineering and include: Filter data with high correlation with PUE as basic features; Perform sliding window statistics on the basic features to calculate the maximum value, minimum value, mean value and variance of the features within the historical data steps and use them as statistical features; The statistical features are added to the data set as part of the data features.

4. The intelligent energy-saving control method based on deep learning according to claim 1 is characterized in that: The obtaining of the optimal parameter combination by using the genetic algorithm model and the PUE prediction model includes: Find sample points close to current load data and environmental data from historical load data and environmental data; Randomly generate an initial parameter combination according to the value range of the adjustable parameter in the sample point and add the initial parameter combination to a candidate set; Calculate the PUE value of each parameter combination in the candidate set by using a PUE prediction model; According to the principle of small PUE value priority and using a roulette wheel strategy, a parameter combination is selected from the candidate set to perform a crossover operation and a mutation operation to generate a child parameter combination; Calculate the PUE value of the child parameter combination by using a PUE prediction model, and when the PUE value is less than the maximum PUE value of the current candidate set, add the child parameter combination to the candidate set and delete the parameter combination with the maximum PUE value in the candidate set; Continue to evolve and iterate until the PUE value of the parameter combination is found to be less than the actual PUE value of the current computer room or the number of iterations reaches the set value.

5. The intelligent energy-saving control method based on deep learning according to claim 4 is characterized in that: The intelligent energy-saving control method based on deep learning also includes a business assurance model. The first-generation parameter combination and the child-generation parameter combination need to be verified by the business assurance model before being added to the candidate set. If the verification fails, they will not be added to the candidate set.

6. The intelligent energy-saving control method based on deep learning according to claim 5 is characterized in that: The process of verifying the parameter combination of the business assurance model includes: Set the alarm temperature of the equipment room; The parameter combination is input into the business assurance model to predict the temperature of the computer room; If the predicted temperature of the computer room is lower than the alarm temperature of the computer room, the verification is passed; if the predicted temperature of the computer room exceeds the alarm temperature of the computer room, the verification is failed.

7. The intelligent energy-saving control method based on deep learning according to claim 4 is characterized in that: The crossover operation process includes randomly selecting a crossover point and exchanging a part of adjustable variables of two parent vectors to generate a child.

8. The intelligent energy-saving control method based on deep learning according to claim 4 is characterized in that: The mutation operation process includes randomly selecting an adjustable variable in the offspring, generating a random value that satisfies a value range to replace the current value of the adjustable variable, and thus generating a new offspring.

9. The intelligent energy-saving control method based on deep learning according to claim 4 is characterized in that: The threshold of the mutation operation is set to 0.

05.

10. The intelligent energy-saving control method based on deep learning according to any one of claims 1 to 9, characterized in that: The implementation of the optimal parameter combination includes: automatically or manually controlling the equipment in the computer room to perform adjustment operations. If the adjustment operation is to hibernate the refrigeration equipment, no more than one refrigeration equipment is hibernated each time; if the adjustment operation is to wake up the refrigeration equipment, there is no upper limit on the number of refrigeration equipment woken up each time.

Images